content
stringlengths 1
15.9M
|
|---|
\section{\label{sec:level1}Introduction}
In 1925, by generalizing the statistics of photons to a system of non-interacting and massive bosons \cite{2,3}, which had been developed by Bose \cite{1}, Einstein concluded that a finite fraction of the particles fill the same single-particle quantum state, below a critical temperature. This fascinating state of matter is now known as a Bose-Einstein condensate (BEC). After the experimental realization of Bose-Einstein condensation in dilute gases of rubidium \cite{4}, sodium \cite{5} and lithium \cite{6}, ultracold Bose gases became versatile and robust systems for probing different properties of quantum systems such as quantum turbulence \cite{Yukalov}, quantum information processing \cite{bloch} and entanglement \cite{tasgin}.
Superradiance originates from the famous argument by Dicke in 1954 \cite{dicke1} which is collective interaction of an ensemble of atoms with a single mode of radiation. Thereafter, collective interactions with fields have been a major subject of research over the past several years. In an experimental analysis in 2005 \cite{ottl}, a combined system with an ultracold atomic ensemble located in an optical cavity was reported and a few years later, strong coupling of a BEC to a cavity was realized \cite{brennecke,colombe}. Placing the BEC inside a cavity would offer several advantages. First, we can prepare the oscillator in its ground state without thermally activated excitations \cite{brennecke}. Second, it dramatically enhances the atom-light interaction by collective coupling to the same light mode \cite{colombe,black,gupta}. And third, decoherence effects associated with inhomogeneities and with atomic motions can be suppressed for a BEC \cite{sherson}. After this experimental realization, there has been a surge of interest in investigation of this system which leads to many interesting developments including cavity enhanced superradiance of a BEC \cite{salma}, self-organization of atoms \cite{nagy}, quantum information \cite{bohi} and superluminal light propagation \cite{kazemi}, to mention a few.
OB, predicted by Sz\"{o}ke and coworkers in 1969 \cite{szoke}, is a phenomenon in which optical resonators containing saturable absorbers have nonlinear characteristics and can exhibit hysteresis. It has been extensively demonstrated both experimentally and theoretically in many systems such as superconducting quantum circuits \cite{hamedi}, quantum dots \cite{mahmoudi2,hamedi2,mahmoudi} and quantum wells \cite{mahmoudi3,sete}. Recently, ultracold atom groups at Berkeley \cite{gupta} and ETH Z\"{u}rich \cite{brennecke} have found OB in systems comprising of vapours of ultracold atoms trapped in optical cavities. Unlike an optomechanical system consisting of an empty cavity with a movable end mirror, the photon numbers in the cavity with a BEC are usually low and even below unity which is desirable for applications such as optical communication and quantum computation.
On the other hand, because of the practical applications of the OB such as all-optical switches, optical transistors and optical memory elements controlling the OB seems to be crucial for the prominent applications. However, from an experimental point of view due to lack of controllability, it has limited applications. In recent years, some attempts have appeared both theoretically and experimentally in order to control the OB. Joshi \textit{et al.} \cite{joshi}, manipulated the OB by changing the intensity and frequency detuning of the coupling field. Almost a year later, Chang and colleagues \cite{chang} showed controllable shift of the threshold points of the OB induced by two suitable tuned fields. By controlling this shift, all-optical flip-flop and storage of optical pulse signals were implemented. More recently, the OB of an ultracold atomic ensemble located in an optical cavity was investigated and it was shown that a transverse pumping field could be used to control the bistable behaviour \cite{yang}. In another study, the OB was investigated for BECs of atoms in a driven optical cavity with Kerr medium and it was shown that the OB can be controlled by adjusting Kerr interaction between the photons \cite{kerr}. In this paper, we investigate effects of the system parameters, i.e., decay rate of the cavity photons, coupling strength between the cavity and the BEC as well as the pump-atom detuning, on the OB in an optomechanical system with a BEC. Controlling the OB is done by simply changing the corresponding parameters without necessity of another field or Kerr medium which can be much more efficient than the previous ones.
This paper is organized as follows. In section 2, we introduce the Hamiltonian of the system consisting of a BEC trapped in a cavity in the discrete-mode approximation. Then, in section 3, we focus on the conditions for bistability and controllability of the OB characteristics via adjusting properly the corresponding parameters of the system, which may lead to promising applications such as all-optical switching. We also investigate the stability of the different branches of the OB curve by applying Routh-Hurwitz criterion and find the minimum value of the threshold intensity. Moreover, we investigate the dynamic properties by introducing an effective potential for the system.
\section{Formalism of the Optomechanical System}
The system under consideration is depicted schematically in Fig. 1. We consider a cigar-shaped BEC of $N$ $\rm{{}^{87}}$Rb atoms trapped in an optical ultrahigh-finesse Fabry-Perot cavity. The coupled BEC-cavity system is driven by a pump laser with frequency $\omega_{p}$ and amplitude $\vert E_{p} \vert=\sqrt{\frac{2\kappa P_{p}}{\hbar\omega_{p}}}$ which is a function of laser power $P_{p}$ and decay rate of the cavity photons $\kappa$. In the large detuning between the pump laser frequency and the atomic resonance, $\triangle_a=\omega_{p}-\omega_{a}$, we can eliminate the excited state of the atoms and the Hamiltonian describing the coupled BEC-cavity system reads \cite{maschler}
\begin{eqnarray}\label{eq1}
\hat{H}_{eff}\ &=& \int \hat{\Psi}^{\dagger}({x}) \{\frac{-\hbar^{2}}{2 m} \frac{d^{2}}{dx^{2}}
+ V_{ext}({x})+\hbar U_{0} \cos^{2}(kx) \hat{a}^{\dagger} \hat{a} \}\hat{\Psi}({x})dx \nonumber\\
&+&\hat{H}_{A-A} + \hbar \Delta \hat{a}^{\dagger} \hat{a}+i \hbar E_{p} (\hat{a}^{\dagger}- \hat{a}).\
\end{eqnarray}
\begin{figure}
\hspace{+0.25cm}
\centerline{\includegraphics[width=8 cm]{1b.eps}}
\caption{\small Sketch of a BEC-cavity system where the cavity is driven by a pump-laser field along the cavity axis.}
\label{figure 1}
\end{figure}
Here $\hat{a}^{\dagger} (\hat{a}) $ denotes the creation (annihilation) operator of the cavity photons with mode function $\cos(kx)$ and wave vector $k=\frac{2\pi}{\lambda}$. The atomic field operator for the creation of an atom in the ground state with mass $m$ at position $x$ is given by $\hat{\Psi}^{\dagger}({x})$. As seen in Eq. (\ref{eq1}), the atom-cavity photon interaction provides a dynamic and quantized potential lattice $U_0 \cos^{2}(kx) \hat{a}^{\dagger} \hat{a}$ for the atoms. Also, $U_{0}=\frac{g_{0}^2}{\Delta_a}$ denotes the potential depth for a single photon with the atom-photon coupling constant $g_0$. Cavity frequency is detuned from the pump laser field by $\triangle=\omega_{c}-\omega_{p}$. Moreover, $V_{ext}(x)$ and $\hat{H}_{A-A}$ are the external trapping potential and atom-atom interaction, respectively.
Applying the discrete-mode approximation \cite{zhang}, we can expand the atomic field operator as
\begin{equation}
\hat{\Psi}({x})= \varphi_{0} \hat{c}_{0}+ \varphi_{2}\hat{c}_{2}.
\end{equation}
Where $\varphi_{0}=1$ and $\varphi_{2}=\sqrt{2}\cos(2kx)$ are two spatial modes and $\hat{c}_{0}$ and $\hat{c}_{2}$ denote the corresponding annihilation operators. For weakly interacting atoms in a shallow external trapping potential, after applying the Bogoliubov approximation and in a rotating frame, the Hamiltonian describing the coupled BEC-cavity reduces to
\begin{equation}\label{eq3}
\hat{H} = 4 \hbar\ \omega_{rec} \hat{c}_{2}^{\dagger} \hat{c}_{2} + \hbar \Delta_{c} \hat{a}^{\dagger} \hat{a} +\hbar g (\hat{c}_{2}^{\dagger} +\hat{c}_{2} ) \hat{a}^{\dagger} \hat{a} + i \hbar E_{p} (\hat{a}^{\dagger}- \hat{a}).
\end{equation}
Here, $\hbar \omega_{rec}=\frac{\hbar^2 k^2}{2 m}$ is the recoil energy. Also, $\Delta_c=\Delta+\frac{NU_{0}}{2} $ and $g=\frac{U_0}{2} \sqrt{\frac{N}{2}}$ stand for the shifted cavity-pump detuning and the effective coupling strength between the cavity and the BEC, respectively. Moreover, the last term in the Hamiltonian, describes the classical pump light input.
From Eq. (\ref{eq3}), Langevin equations for the optomechanical system are given by
\begin{equation}
\frac{d}{dt} \hat{a}({t})=\\
(- i \Delta_c - \frac{\kappa}{2})\ \hat{a}({t}) - i g \sqrt{2} \ \hat{X}({t}) \ \hat{a}({t})+E_{p},
\end{equation}
\begin{equation}\label{eq5}
\frac{d}{dt} \hat{X} ({t}) = 4\omega_{rec} \hat{P}({t}),
\end{equation}
\begin{equation}\label{eq6}
\frac{d}{dt} \hat{P}({t})=- 4 \omega_{rec}\ \hat{X}({t})-\gamma_{m}\hat{P}({t}) - \sqrt{2} \ g \ \hat{a}^{\dagger}({t})\ \hat{a}({t}).
\end{equation}
We have introduced the position and momentum operators of the BEC, $\hat{X}=\frac{\hat{c}_{2}^{\dagger} +\hat{c}_{2}}{\sqrt{2}}$ and $\hat{P}=\frac{\hat{c}_{2}-\hat{c}_{2}^{\dagger} }{ i\sqrt{2}}$ as well as the damping rate of the atomic excited state $\gamma_{m}$.
Assuming $\hat{a}({t})=\bar{a}+\delta\hat{a}({t})$, $ \hat{X}({t})=\bar{x}+\delta\hat{X}({t})$ and $ \hat{P}({t})=\delta\hat{P}({t})$, we first find the self-consistent steady state solutions
\begin{equation}\label{eq7}
\bar{a} = \frac{E_{p}}{ i\Delta_c +i\ g \ \sqrt{2} \ \bar{x}+\frac{\kappa}{2}},
\end{equation}
\begin{equation}\label{eq8}
\bar{x} = \frac{- g \sqrt{2} |\bar{a}|^{2}}{4\ \omega_{rec}},
\phantom{\hspace{1.5cm}}
\end{equation}
here
$\bar{a}$ and $\bar{x}$ stand for the interactivity field and the mechanical position. Moreover, $\delta \hat{a}$ and $\delta \hat{X}$ are the fluctuations around these steady states. Combining Eqs. (\ref{eq7}) and (\ref{eq8}), leads
\begin{equation}\label{eq9}
n=\frac{E_{p}^2}{(\frac{\kappa}{2})^2+(\Delta_c-\frac{g^2}{2 \omega_{m}} n)^2},
\end{equation}
where $\omega_{m}=4 \omega_{rec}$ and $n=|\bar{a}|^{2}$ is the number of photons.
Rearranging Eq. (\ref{eq9}), the resulting equation is
\begin{equation}\label{eq10}
n((\frac{\kappa}{2})^2+(\Delta_c-\frac{g^2}{2 \omega_{m}} n )^2)=E_{p}^2,
\end{equation}
Setting Eq. (\ref{eq10}) to zero, gives a third-order polynomial equation of the form
\begin{equation}\label{eq14}
n^3+a_2n^2+a_1n+a_0=0,
\end{equation}
where
\begin{equation}
b=\frac{g^4}{4 \omega ^2_{m}}, \ \ a_2=\frac{-4 \Delta_{c} \omega_m}{g^2},\ \ a_1=\frac{\frac{ \kappa^2}{4}+ \Delta ^2 _{c}}{b}, \ \ a_0=\frac{-E_{p}^2}{b}.
\end{equation}
By noting the fact that the steady state photon numbers exhibit OB for a certain range of cavity-pump detuning by increasing the amplitude of the pump field, in the following we are going to calculate the critical values of cavity-pump detuning and pump amplitude. So the system first becomes bistable at a single value of the detuning, denoted by $\Delta_{cr}$ and the critical pump amplitude, at which the bistability at $\Delta_{cr}$ occurs, is denoted by $E_{cr}$.
In order to calculate the critical values, let us have a look at Eq. (\ref{eq14}). The critical points of the third-order polynomial equation are obtained by setting derivative of the equation equal to zero and are given by $n=\frac{-a_2\pm\sqrt{a_2^2-3a_1}}{3}$. In the case where $\sqrt{a_2^2-3a_1}$ is positive, the function has a local minimum and a local maximum. Equating $\sqrt{a_2^2-3a_1}$ to zero so that the function has only a critical point, the critical cavity-pump detuning is given by
\begin{equation}
\Delta_{cr}=\frac{\sqrt{3}}{2}\kappa,
\end{equation}
and we define the critical pump amplitude, at which the bistability at $\Delta_{cr}$ occurs, as
\begin{equation}
E_{cr}=\sqrt{\frac{\omega_m \kappa^3}{6\sqrt{3}\ g^2}}.
\end{equation}
In the following, by using the ansatz and retaining only first order terms in the small quantities $\delta \hat{a},\delta \hat{X}$, $ \delta \hat{a}^{\dagger}$ and $\delta \hat{P}$, one gets the linearized quantum Langevin equations for the fluctuation operators
\begin{equation}\label{6}
\frac{d}{dt}\delta \hat{a}({t})=\\
(- i \bar{\Delta}_c - \frac{\kappa}{2})\delta \hat{a}({t}) - i g \sqrt{2} \ \bar{a} \ \delta\hat{X}({t}) ,
\phantom{\hspace{0.75cm}}
\end{equation}
\begin{equation}\label{7}
\frac{d}{dt} \delta\hat{X}({t})= \omega_{m} \delta\hat{P}({t}),
\end{equation}
\begin{equation}\label{8}
\frac{d}{dt} \delta\hat{P}({t})=- \omega_{m}\delta\hat{X}({t}) - \gamma_{m}\delta\hat{P}({t})- \sqrt{2} g (\bar{a}\ \delta\hat{a}^{\dagger}({t}) + \bar{a} ^{*}\delta\hat{a}({t})),
\end{equation}
where $\bar{\Delta}_c=\Delta_c+\ g \ \sqrt{2} \ \bar{x} $.
\section{Results and discussion}
\begin{figure*}
\centerline{\includegraphics[width=12cm]{2b.eps}}
\caption{\small steady state photon numbers as a function of the shifted cavity-pump detuning for the four different amplitude of the pump field, $E_{p}=0$ (a), $E_{p}=0.5 $ MHz (b), $E_{p}=0.83 $ MHz (c) and $E_{p}=1.5 $ MHz (d). Other used parameters are $\kappa=2\pi\times 1.3$ KHz, $ N=1.2\times10^5$, $\Delta_a=2\pi\times 32 $ GHz, $g_0=2\pi\times 10.9 $ MHz and $\omega_{m}=2\pi\times 15.2 $ KHz .}
\end{figure*}
In the begining of this section, we present the results based on the numerical solutions of Eq. (\ref{eq9}) and focus on the conditions for bistability.
\begin{figure}[!ht]
\centerline{\includegraphics[width=12cm]{3b.eps}}
\caption{\small S-shaped curve for the steady state photon numbers as a function of the amplitude of the pump for $\Delta_{c}= 7 $ MHz (a) and $\Delta_{c}= 11 $ MHz (b). Other parameters are the same as in Fig. 2.}
\end{figure}
In Fig. 2, we show how the photon numbers depend on the shifted cavity-pump detuning for given amplitudes of the pump $E_{p}$. Experimentally used parameters are $\kappa=2\pi\times 1.3 $ MHz, $ N=1.2\times10^5$, $\Delta_a=2\pi\times 32$ GHz, $g_0=2\pi\times 10.9$ MHz and $\omega_{m}=2\pi\times 15.2$ KHz \cite{brennecke}. It can be seen from Eq. (\ref{eq9}) that when $E_{p}$ is sufficiently small, the photon numbers are also small, so we can neglect $ n^2$ and we have a symmetrical lorentzian curve centred at $\Delta_c=0$. In Fig. 2(b), it is shown that by increasing the amplitude of the pump field, the lorentzian curve is largely unchanged; however, the curve becomes more and more asymmetric and its only maximum moves to the right. But as $E_{p}$ reaches a critical value, the nature of the curve changes so that it has an infinite slope at $\Delta=\Delta_{cr}$, according to Fig. 2(c). For amplitudes beyond the critical value, Eq. (\ref{eq9}) has three real roots, corresponding to the branch CD in Fig. 2(d) where the multiple solutions indicate the bistable behaviour.
The bistable behaviour can also be seen from the OB hysteresis curve, shown in Fig. 3. For the detuning above the critical value by increasing the amplitude of the pump field system goes to the bistability branch, according to Fig. 3(b). Consider that amplitude increases from zero gradually; the photon numbers initially follow the lower stable branch. When the amplitude increases to the first bistable point, it jumps to the upper stable branch and if the amplitude is increased further, continues to follow that branch. Now if we start decreasing the amplitude, the steady state photon numbers remain on the upper stable branch at first; however, by decreasing the amplitude even further, when it reaches the second bistable point, it jumps down to the lower stable branch and continues to decrease along that branch for further decrease of the amplitude. We Note that placing the BEC inside an optical cavity allows for OB at extremely low cavity photon numbers; however, an empty optomechanical system does not exhibit the behaviour.
In the following, we are mainly focused on controllability of the OB.
\begin{figure}[!hb]
\centerline{\includegraphics[width=8cm]{4b.eps}}
\caption{\small Steady state photon numbers as a function of the amplitude of the pump for different decay rate of the cavity photons $ \kappa=2 \pi \times 0.5$ MHz (solid line), $\kappa=2\pi \times 1.5$ MHz (dashed line) and $\kappa=2\pi \times 3$ MHz (dotted line). Other parameters are the same as in Fig. 3(b).}
\end{figure}
First, we are going to examine the effects of the decay rate of the cavity photons on the OB. Figure 4 depicts steady state photon numbers as a function of amplitude of the pump for various decay rates. Here, we can observe that by increasing the decay rate, width of the bistability curve decreases. In addition, the OB threshold increase by increasing the value of the decay rate. By further increasing the decay rate, the OB may be eliminated. In order to clarify the behaviour of the OB curve with variation of the decay rate, we present the Fig. 5 considering the same parameters used for Fig. 3(b). As one expects, decay rate of the cavity photons pushes the OB threshold to higher values.
\begin{figure}
\centerline{\includegraphics[width=8cm]{5b.eps}}
\caption{\small Threshold intensity as a function of $ \kappa $. Other parameters are the same as in Fig. 3(b).}
\end{figure}
In what follows, we will investigate the effects of the coupling strength between the cavity and the BEC on the OB threshold and the width of the bistability curve. In refrence \cite{colombe}, the authors have shown the possibility of controlling the coupling strength by locating the BEC anywhere within the cavity, so we can introduce another control parameter. Steady state photon numbers as a function of the amplitude of the pump and for different coupling strengths are plotted in Fig. 6. Other parameters are the same as in Fig. 3(b). As is seen, increasing the coupling strength, i.e., $g_{0}$, causes the system to become bistable at lower amplitudes of the pump. It also decreases the width of the bistability curve. Therefore, it is considered as another control parameter which can provide the possibility of realizing a controllable optical switch. Moreover, the coupling strength allows for the OB at extremely low photon numbers as well as fast and low-threshold optical switches.
Then we show that optical behaviour is sensitive to the detuning between the pump laser field and the atomic resonance.
\begin{figure}[!hb]
\centerline{\includegraphics[width=8cm]{6b.eps}}
\caption{\small Steady state photon numbers as a function of the amplitude of the pump for three different coupling strength $ g_{0}=2\pi\times 8 $ MHz (solid line), $g_{0}=2\pi\times 10 $ MHz (dashed line) and $g_{0}=2\pi\times 12 $ MHz (dotted line). The value of other parameters is the same as those in Fig. 3(b).}
\end{figure}
Figure 7 depicts steady state photon numbers versus the amplitude of the pump for $\Delta_a =2 \pi\times 30$ GHz (solid line), $\Delta_a =2 \pi\times 40 $ GHz (dotted line) and $ \Delta_a =2 \pi\times 50 $ GHz (dashed line). According to the figure, decreasing the pump-atom detuning causes the system to become bistable at lower amplitudes and also decrease the width of the bistability curve. This behavior opens the possibility of the optical switching by varying the pump laser field frequency.
As the OB threshold intensity is a function of the corresponding parameters, we minimize it over the range of experimental data \cite{brennecke,colombe} and find the minimum threshold intensity of the OB, $E_{p_{min}}=1 $ MHz and the corresponding laser power $P_{p_{min}}=3\times 10^{-14} $W.
\begin{figure}
\centering
\includegraphics[width=8cm]{7b.eps}
\caption{\small Steady state photon numbers as a function of the amplitude of the pump for $\Delta_a =2 \pi\times 30 $ GHz (solid line), $\Delta_a =2 \pi\times 40$ GHz (dashed line) and $ \Delta_a =2 \pi\times 50 $ GHz (dotted line). Other parameters are the same as in Fig. 3(b).}
\end{figure}
In the following, we are going to study the stability of the different branches of the OB curve. Applying the Routh-Hurwitz criterion to the Eqs. (\ref{6}), (\ref{7}) and (\ref{8}), we can get the stability conditions
\begin{equation}
\begin{split}
s_1 &= \kappa \gamma_{m} \lbrace [\frac{\kappa^2}{4} +(\omega_{m}-\bar{\Delta}_c)^2][ \frac{\kappa^{2}}{4}+(\omega_{m}+\bar{\Delta}_c)^2]\\
& + \gamma_{m} [ (\gamma_{m}+\kappa) (\frac{\kappa^{2}}{4}+\bar{\Delta}_c^{2})+\kappa \omega^{2}_{m}] \rbrace +4 \,g^2 \,n\,\omega_{m} \,\bar{\Delta}_c (\gamma_{m}+\kappa)^2>0,
\end{split}
\end{equation}
\begin{equation}\label{eq17}
s_2=\omega_{m} (\frac{\kappa^{2}}{4}+\bar{\Delta}_{c}^{2})-4\,g^2\, n\, \bar{\Delta}_c >0.\phantom{\hspace{1.25cm}}
\end{equation}
For the positive detuning, $\bar{\Delta}_c>0$, the first condition is always satisfied. In Fig. 8, we see that for the middle branch, $s_2$ is negative and so the steady states on the region are unstable. It is worth mentioning that the condition is consistent with the fact that the branches with negative slope are always unstable \cite{ikeda}.
\begin{figure*}[!hb]
\centerline{\includegraphics[width=12cm]{8b.eps}}
\caption{\small (a) Behavior of steady state photon numbers in response to the pump field. (b) shows the normalized second stability condition $s_2$ versus the photon numbers. Parameters values are the same as in Fig. 3(b), with except that $g_{0}$ is replaced by the coupling strength of Fig. 6(a). }
\end{figure*}
In what follows, we will try to obtain the physical interpretation of the condition. Note that experimentally \cite{brennecke,colombe}, the decay rate of the cavity photons is often two orders of magnitude faster than the mechanical motion of the condensate, so we safely assume that the cavity field follows the condensate adiabatically. Setting $\frac{d}{dt}\delta\hat{a} =0$ in Eq. (\ref{6}), one gets:
\begin{equation}
\delta{a({t})=- \frac{ig\sqrt{2}\ \bar{a} \ \delta{X({t})}} {i\bar{\Delta}_{c}+\frac{\kappa}{2}}},
\phantom{\hspace{4.75cm}}
\end{equation}
\begin{equation}
\delta{a^{*}({t})=\frac{ig\sqrt{2} \ \bar{a}^{*} \ \delta{X({t})}} {-i\bar{\Delta}_{c}+\frac{\kappa}{2}}},
\phantom{\hspace{4.5cm}}
\end{equation}
\begin{equation}\label{9}
\frac{d}{dt} \delta\hat{P}({t})=-(\omega_{m}+\frac{4 g^{2}n \bar{\Delta}_{c}}{\bar{\Delta}_{c}^{2}+\frac{\kappa^{2}}{4}}) \delta\hat{X}({t})- \gamma_{m}\delta\hat{P}({t}),
\phantom{\hspace{1cm}}
\end{equation}
\begin{equation}\label{19}
\frac{d^{2}}{dt^{2}} \delta\hat{X}({t}) + \gamma_{m}\frac{d}{dt}\delta\hat{X}({t}) + (\omega^{2}_{m}-\frac{4 g^{2}\ n \omega_{m} \bar{\Delta}_{c}}{\bar{\Delta}_{c}^{2}+\frac{\kappa^{2}}{4}}) \delta\hat{X}({t})=0.
\end{equation}
It can be seen that the coefficient of $\delta\hat{X} $ in third term, corresponds to a restoring force. On the other hand, from the second condition, Eq. (\ref{eq17}), we can see that the mechanical oscillator is stable if the restoring force is positive. In the middle branch, the third coefficient in the Eq. (\ref{19}) is negative and so it is no longer a restoring force leading to the unstable behaviour. The phenomenon can be also viewed as optical spring effect \cite{genes,aldana}. It is worth to mention that except here, we do not make an adiabatic approximation.
Considering that OB can be well understood in the framework of the simple double-well potential model, so we proceed to obtaining the effective potential. Putting the interactivity field into Eqs. (\ref{eq5}) and (\ref{eq6}), we get the evolution of the displacement for the BEC
\begin{equation}
\frac{d^{2}}{dt^{2}}X({t}) = -\omega^{2}_{m}X({t})-\sqrt{2}\ g \ \omega_{m} n,
\end{equation}
and the potential is given by
\begin{equation}
V(X)= \dfrac{1}{2} \omega^{2}_{m} X^{2} +\frac{E^2_{p}\, \arctan(\frac{\Delta_c + \sqrt{2} \, g\, X}{\sqrt{1/4} \kappa})}{\sqrt{4} \kappa}.
\end{equation}
As is seen, we have a sum of the harmonic potential and inverse tangent terms. The latter may lead to two minima and the double-well like potential.
\begin{figure}
\centerline{\includegraphics[width=8cm]{9b.eps}}
\caption{\small The effective potential for the different pump amplitude. The values are $E_{p}=200 $ KHz (solid line), $E_{p}=980 $ KHz (dashed line) and $E_{p}=2 $ MHz (dotted line). The top and the bottom curve have only one minimum, while the middle one has two minimum and one maximum. The inset shows a zoom-in of the middle curve. Parameters values are the same as in Fig. 3(b).}
\end{figure}
Next, we investigate the dynamic properties by the effective potential which is plotted in Fig. 9 as a function of the displacement and for different pump amplitudes. The potential with $E_{p}=200 $ KHz is shown as the solid line. It gets the form of the harmonic potential and has a single minimum close to the origin, corresponding to the stable lower branch. The dashed line corresponds to the case $E_{p}=980 $ KHz which lies in the OB region.
So, the potential is a double-well with three extrema, two minimum and one maximum. The first minimum located in the vicinity of the origin, corresponding to the lower branch of OB and the other one, is far away from the origin and corresponds to the upper branch. The maximum also represents the unstable branch. For $E_{p}=2 $ MHz, the inverse tangent term gives just a shift to the potential and so it has a new minimum at a larger displacement corresponding to the purely upper stable branch.
\section{Conclusion}
We investigated the OB in a system consisting of a BEC inside an optical cavity. It was first shown that how the bistability develops in this cavity optomechanical system for a certain range of cavity-pump detunings and pump amplitudes. Further, it was demonstrated that one can control the OB threshold and the width of the bistability curve by means of the decay rate of the cavity photons, coupling strength between the cavity and the BEC as well as the pump-atom detuning. To the best of our knowledge, our work is the first of its kind to investigate the effects of the system parameters on the optical behaviour. Another additional advantage is that extremely low cavity photon numbers exhibit bistable behaviour which is absent for previous systems. Also, the system parameters allow for a considerable reduction of the threshold and this phenomenon has potential applications in optical switches at very low intensities. Moreover, width of the bistability curve may be altered by changing the parameters allowing all-optical flip-flop and storage of optical pulse signals. We then analyzed the stability of the different branches of the OB curve by applying the Routh-Hurwitz criterion as well as offering a physical interpretation for these conditions. Finally, we derived an effective potential in order to have an intuitive picture.
\newline
\bibliographystyle{}
|
\section{Introduction}
\subsection{Background}
Many problems in number theory can be recast in the language of homogeneous dynamics. Let $G$ be a Lie group and $\Gamma$ be a lattice in $G$, i.e. a discrete subgroup of finite covolume. Take a sequence $\{g_i\}$ in $G$ and a probability measure $\mu$ on $G/\Gamma$ which is supported on a smooth submanifold of $G/\Gamma$. The following question was raised by Margulis in \cite{Mar02}:
\begin{question}[Margulis]
What is the distribution of $g_i\mu$ in $G/\Gamma$ when $g_i$ tends to infinity in $G$?
\end{question}
In 1993, Duke, Rudnick and Sarnak \cite{DRS93} studied the case where $\mu$ is a finite invariant measure supported on a symmetric subgroup orbit, and applied it to obtain asymptotic estimates for the number of integral points of bounded norm on affine symmetric varieties. At the same time, Eskin and McMullen \cite{EM93} gave a simpler proof using the mixing property of geodesic flows. It was later generalized by Eskin, Mozes and Shah \cite{EMS96} to the case where $\mu$ is a finite invariant measure supported on a reductive group orbit, and applied it to count integral matrices of bounded norm with a given characteristic polynomial. Later Gorodnik and Oh \cite{GO11} worked in the Adelic setting, and gave an asymptotic formula for the number rational points of bounded height on homogeneous varieties. \par
In another direction, the dynamical behavior of translates of a submanifold of expanding horospherical subgroups in $\SL_n(\R)/\SL_n(\Z)$ is closely related to metric Diophantine approximation. In 1998, Kleinbock and Margulis \cite{KM98} proved extremality of a non-degenerate submanifold in $\R^n$, and their proof was based on quantitative non-divergence of translates of the submanifold by semisimple elements. Their work was later extended from $\R^n$ to the space $M_{m\times n}(\R)$ of $m\times n$ real matrices (see e.g. \cite{KMW10}\cite{BKM15}\cite{ABRS18}).\par
While quantitative non-divergence results are useful in the study of extremality, equidistribution results can be applied to study the improvability of Dirichlet's theorem. In 2008, Kleinbock and Weiss \cite{KW08} first explored improvability in the language of homogeneous dynamics, based on earlier observations by Dani \cite{Dan84} as well as Kleinbock and Margulis \cite{KM98}. Later Shah \cite{Sha09Invention} obtained a strengthened result for analytic curves in $\R^n$ by showing equidistribution of expanding translates of curves in $\SL_{n+1}(\R)/\SL_{n+1}(\Z)$ by singular diagonal elements $a(t)=\mathrm{diag}(t^{n},t^{-1},\cdots, t^{-1})$. This work has also been generalized to $m\times n$ matrices in a recent preprint \cite{SY16} by Shah and Lei Yang, where they considered the case $G=\SL_{m+n}(\R)$ and $a(t)=\mathrm{diag}(t^n,\cdots,t^n,t^{-m},\cdots,t^{-m})$. We shall discuss this subject in more details in \Cref{subsect:Grassmannian_Dirichlet}. \par
It is also worth considering the case $G=\SO(n,1)$, as there are interesting applications to hyperbolic geometry. See Shah's works \cite{Sha09Duke1}\cite{Sha09Duke2} and later generalizations by Lei Yang \cite{Yan16Israel}\cite{Yan17}. We shall provide more details in \Cref{subsect:equidistribution}. \\
Motivated by the previous works, we are interested in the following equidistribution problem, which was proposed by Shah in ICM 2010 \cite{Sha10PICM}. Let $G=\mathbf{G}(\R)$ be a semisimple connected real algebraic group of non-compact type, and let $L$ be a Lie group containing $G$. Let $\Lambda$ be a lattice in $L$. Let $\{a(t)\}_{t\in\R^{\times}}$ be a multiplicative one-parameter subgroup of $G$, i.e. we have a homomorphism of real algebraic group $a \colon \mathbb{G}_m \rightarrow \mathbf{G}$. Suppose we have a bounded piece of an analytic curve on $G$ given by $\phi\colon I=[a,b]\rightarrow G$, and we fix a point $x_0$ on $L/\Lambda$ such that $Gx_0$ is dense in $L/\Lambda$. Let $\lambda_{\phi}$ denote the measure on $L/\Lambda$ which is the parametric measure supported on the orbit $\phi(I)x_0$, that is, $\lambda_{\phi}$ is the pushforward of the Lebesgue measure. When does $a(t)\lambda_{\phi}$ converge to the Haar measure on $L/\Lambda$ with respect to the weak-* topology, as $t$ tends to infinity? \par
In \cite{Sha10PICM}, Shah found natural algebraic obstructions to equidistribution, and asked if those are the only obstructions. In this article, we give an affirmative answer to Shah's question. This generalizes previous results on $G=\SO(n,1)$ \cite{Sha09Duke1}\cite{Yan16Israel}, $G=\SO(n,1)^k$ \cite{Yan17}, as well as $G=\SL_n(\R)$ and $a(t)$ being singular \cite{Sha09Invention}\cite{SY16}. We also apply the equidistribution result to show that for almost every point on a ``non-degenerate'' analytic curve in the space of $m\times n$ real matrices, the Dirichlet's theorem cannot be improved. This sharpens a result of Shah and Yang \cite{SY16}. \par
We remark that our method also applies to analytic submanifolds. For convenience, we restrict our discussions to curves.
\subsection{Non-escape of mass to infinity}\label{subsect:non_escape_of_mass}
Let $G=\mathbf{G}(\R)$ be a semisimple connected real algebraic group of non-compact type, and let $L$ be a Lie group containing $G$. Let $\Lambda$ be a lattice in $L$. Let $\{a(t)\}_{t\in\R^{\times}}$ be a multiplicative one-parameter subgroup of $G$ with non-trivial projection on each simple factor of $G$. There is a parabolic subgroup $P=P(a)$ of $G$ associated with $a(t)$:
\begin{equation}
P := \{ g \in G \colon \lim_{t\to\infty} a(t)ga(t)^{-1} \text{ exists in $G$}\}.
\end{equation}
Suppose we have a bounded piece of an analytic curve on $G$ given by $\phi:I=[a,b]\rightarrow G$, and we fix a point $x_0$ on $L/\Lambda$ such that the orbit $Gx_0$ is dense in $L/\Lambda$. Let $\lambda_{\phi}$ denote the parametric measure on $L/\Lambda$. If we expect the translated measures to get equidistributed, it is necessary that there is no escape of mass to infinity. \par
Let us first consider the special case $G=L=\SL_{m+n}(\R)$, $\Lambda=\SL_{m+n}(\Z)$ and $a(t)=\mathrm{diag}(t^n,\cdots,t^n,t^{-m},\cdots,t^{-m})$. In \cite{ABRS18}, Aka, Breuillard, Rosenzweig and de Saxc\'e defined a family of algebraic sets called \emph{constraining pencils} (see \cite[Definition 1.1]{ABRS18}), and used it to describe the obstruction to quantitative non-divergence. They remarked that constraining pencils give rise to certain Schubert varieties in Grassmannians.\par
Inspired by their work, we define the notion of \emph{unstable Schubert varieties}\footnote{The name comes from the notion of stability in geometric invariant theory, and should not be confused with unstable manifolds for a diffeomorphism.} (see \Cref{def:Schubert_variety}) with respect to $a(t)$ for general partial flag variety $G/P$, which naturally generalizes the notion of constraining pencils. This enables us to describe obstructions to non-divergence in general case. \par
Now we project our curve $\phi$ onto $G/P$. Consider
\begin{align}\label{eq:phi_tilde}
\widetilde{\phi}\colon&[a,b]\longrightarrow G/P \nonumber \\
&s \longmapsto \phi(s)^{-1}P.
\end{align}
We are taking inverse here simply because we would like to quotient $P$ on the right, which is the case in most literatures.\par
We are ready to state our first main theorem on non-escape of mass.
\begin{theorem}[Non-escape of mass]\label{thm:non_divergence_introduction}
Let $\phi:I=[a,b]\rightarrow G$ be an analytic curve such that the image of $\widetilde{\phi}$ is not contained in any unstable Schubert variety of $G/P$ with respect to $a(t)$. Then for any $\epsilon > 0$, there exists a compact subset $K$ of $L/\Lambda$ such that for any $t > 1$, we have
\begin{equation}
\frac{1}{b-a}\left\vert\left\{ s\in[a,b]\colon a(t)\phi(s)x_0\in K \right\} \right\vert > 1 - \epsilon.
\end{equation}
\end{theorem}
To prove \Cref{thm:non_divergence_introduction}, we consider a certain finite dimensional representation $V$ of $G$ (see \Cref{def:V}), and show that the corresponding curve in $V$ cannot be uniformly contracted to the origin. The key ingredient is the following theorem, which is the main technical contribution of this article.
\begin{theorem}[Linear stability]\label{thm:linear_stability}
Let $\rho\colon G\rightarrow \GL(V)$ be any finite-dimensional linear representation of $G$, with a norm $\lVert\cdot\rVert$ on $V$. Suppose that the image of $\widetilde{\phi}$ is not contained in any unstable Schubert variety of $G/P$ with respect to $a(t)$. Then there exists a constant $C>0$ such that for any $t>1$ and any $v\in V$, one has
\begin{equation}\label{eq:linear_stablility}
\sup_{s\in [a,b]}\lVert a(t)\phi(s)v \rVert \geq C\lVert v \rVert.
\end{equation}
\end{theorem}
\Cref{thm:linear_stability} is of independent interest, as it is also applicable to obtain quantitative non-divergence results (see e.g. \cite{Shi15}). Compared to the previous works on special cases of the theorem, the novel part of our proof is that we use a result in geometric invariant theory, which is Kempf's numerical criterion \cite[Theorem 4.2]{Kem78}. \par
Geometric invariant theory was first developed by Mumford to construct quotient varieties in algebraic geometry; its connections to dynamics have been found in recent years. Kapovich, Leeb and Porti \cite[Section 7.4]{KLP18} explored the relation with geometric invariant theory for groups of type $A_1^n$. In a recent preprint \cite{Kha15}, Khayutin utilized geometric invariant theory to study the double quotient of a reductive group by a torus. In \cite[Section 6]{RS16}, Richard and Shah applied \cite[Lemma 1.1(b)]{Kem78} to deal with focusing, which also came from the study of geometric invariant theory. \par
\Cref{thm:linear_stability} is proved in \Cref{sect:linear_stability}, and \Cref{thm:non_divergence_introduction} is proved in \Cref{sect:nondivergence}.
\subsection{Equidistribution of translated measures}\label{subsect:equidistribution}
Let the notations be as in \Cref{subsect:non_escape_of_mass}, and suppose that the image of $\widetilde{\phi}\colon s\mapsto\phi(s)^{-1}P$ is not contained in any unstable Schubert variety of $G/P$ with respect to $a(t)$ (see \Cref{def:Schubert_variety}). Due to \Cref{thm:non_divergence_introduction}, for any sequence $t_i \to\infty$, the sequence of translated measures $a(t)\lambda_{\phi}$ is tight, i.e. any weak-* limit is a probability measure on $L/\Lambda$. If one can further show that any limit measure is the Haar measure on $L/\Lambda$, then the translated measure $a(t)\lambda_{\phi}$ gets equidistributed as $t\to\infty$. In order to achieve this, one needs to exclude a larger family of obstructions. \par
In a sequence of papers \cite{Sha09Duke1}\cite{Sha09Duke2}\cite{Sha09Invention}, Shah initiated the study of the curve equidistribution problem with several important special cases. For example, when $G=\SL_{n+1}(\R)$ and $a(t)=\mathrm{diag}(t^{n},t^{-1},\cdots,t^{-1})$, the obstructions to equidistribution come from linear subspaces of $\R\mathbb{P}^n$, which are exactly the unstable Schubert varieties with respect to $a(t)$.\par
Another interesting case is when $G=\SO(n,1)$ and $\{a(t)\}$ being the geodesic flow on the unit tangent bundle $T^1(\mathbb{H}^n)$ of the hyperbolic space $\mathbb{H}^n\cong \SO(n,1)/\SO(n)$. The visual boundary of $\mathbb{H}^n$ has the identification
\begin{equation}
\partial\mathbb{H}^n\cong\mathbb{S}^{n-1}\cong G/P.
\end{equation}
Shah found that the obstructions to equidistribution comes from proper subspheres $\mathbb{S}^{m-1}$ of $\mathbb{S}^{n-1}$ ($m<n$). However, since the real rank of $G$ is one, the proper Schubert varieties of $G/P$ are just single points. Therefore, these obstructions are not given by Schubert varieties. Nonetheless, the subspheres are still natural geometric objects, as they are closed orbits of the subgroups $\SO(m,1)\subset\SO(n,1)$, which correspond to totally geodesic submanifolds $\mathbb{H}^m\subset\mathbb{H}^n$. \par
Motivated by these results, Shah \cite{Sha10PICM} found the following algebraic obstruction to equidistribution in the general setting. Suppose that $F$ is a proper subgroup of $L$ containing $\{a(t)\}$, and $g\in G$ is an element such that the orbit $Fgx_0$ is closed and carries a finite $F$-invariant measure. Suppose that $\phi(I)\subset P(F\cap G)g$. Then for any sequence $t_i\to\infty$, it follows that any weak-* limit of probability measures $a(t_i)\lambda_{\phi}$ is a direct integral of measures which are supported on closed sets of the form $bFgx_0$, where $b\in P$. Such limiting measures are concentrated on strictly lower dimensional submanifolds of $L/\Lambda$. Shah also asked if these are the only obstructions. \par
We now state our main theorem on equidistribution, which answers Shah's question affirmatively. Recall that $x_0$ is an element in $L/\Lambda$ such that $Gx_0$ is dense in $L/\Lambda$. Let $\widetilde{\phi}$ be as in \eqref{eq:phi_tilde}. For the definition of \emph{unstable Schubert variety}, see \Cref{def:Schubert_variety}.
\begin{theorem}\label{thm:equidistribution0}
Let $\phi\colon I=[a,b]\rightarrow G$ be an analytic curve such that the following two conditions hold:
\begin{enumerate}[(a)]
\item The image of $\widetilde{\phi}$ is not contained in any unstable Schubert variety of $G/P$ with respect to $a(t)$;
\item For any $g\in G$ and any proper algebraic subgroup $F$ of $L$ containing $\{a(t)\}$ such that $Fgx_0$ is closed and admits a finite $F$-invariant measure, the image of $\phi$ is not contained in $P(F\cap G)g$.
\end{enumerate}
Then for any $f\in C_c(L/\Lambda)$, we have
\begin{equation}\label{eq:equidistribution0}
\lim_{t\to\infty}\frac{1}{b-a}\int_{a}^{b}f(a(t)\phi(s)x_0)\,\mathrm{d}s = \int_{L/\Lambda}f\,\mathrm{d}\mu_{L/\Lambda},
\end{equation}
where $\mu_{L/\Lambda}$ is the $L$-invariant probability measure on $L/\Lambda$.
\end{theorem}
\begin{remark}
In \Cref{thm:equidistribution0}, if we assume $(a)$ holds, then by the above discussion we know that \eqref{eq:equidistribution0} holds if and only if $(b)$ holds. In this sense, our result is sharp.
\end{remark}
One can even require $F\cap G$ to be reductive if we replace the family of unstable Schubert varieties with the slightly larger family of \emph{weakly unstable Schubert varieties} (see \Cref{def:Schubert_variety}).
\begin{theorem}\label{thm:equidistribution1}
Let $\phi\colon I=[a,b]\rightarrow G$ be an analytic curve such that the following two conditions hold:
\begin{enumerate}[(A)]
\item The image of $\widetilde{\phi}$ is not contained in any weakly unstable Schubert variety of $G/P$ with respect to $a(t)$;
\item For any $g\in G$ and any proper algebraic subgroup $F$ of $L$ containing $\{a(t)\}$ such that $Fgx_0$ is closed and admits a finite $F$-invariant measure and that $F\cap G$ is reductive, the image of $\phi$ is not contained in $P(F\cap G)g$.
\end{enumerate}
Then for any $f\in C_c(L/\Lambda)$, we have
\begin{equation}\label{eq:equidistribution1}
\lim_{t\to\infty}\frac{1}{b-a}\int_{a}^{b}f(a(t)\phi(s)x_0)\,\mathrm{d}s = \int_{L/\Lambda}f\,\mathrm{d}\mu_{L/\Lambda},
\end{equation}
where $\mu_{L/\Lambda}$ is the $L$-invariant probability measure on $L/\Lambda$.
\end{theorem}
If a reductive subgroup $H$ contains $\{a(t)\}$, then $P_H=P\cap H$ is a parabolic subgroup of $H$ associated with $a(t)$, and $HP/P$ is homeomorphic to $H/P_H$. Hence we give the following definition.
\begin{definition}[Partial flag subvariety]\label{def:partial_flag_subvariety}
A \emph{partial flag subvariety of $G/P$ with respect to $a(t)$} is a subvariety of the form $gHP/P$, where $g$ is an element in $G$, and $H$ is a reductive subgroup of $G$ containing $\{a(t)\}$.
\end{definition}
In view of \Cref{def:partial_flag_subvariety}, \Cref{thm:equidistribution1} shows that the obstructions consist of two families of geometric objects: weakly unstable Schubert varieties and partial flag subvarieties.\par
\Cref{thm:equidistribution0} and \Cref{thm:equidistribution1} are proved in \Cref{sect:focusing}.
\subsection{Grassmannians and Dirichlet's approximation theorem on matrices}\label{subsect:Grassmannian_Dirichlet}
In this section, we give an application of our equidistribution result to simultaneous Diophantine approximation.\par
In 1842, Dirichlet proved a theorem on simultaneous approximation of a matrix of real numbers (DT): Given any two positive integers $m$ and $n$, a matrix $\Psi\in M_{m\times n}(\R)$, and $N>0$, there exist integral vectors $\mathbf{p}\in\Z^n\backslash\{\mathbf{0}\}$ and $\mathbf{q}\in\Z^m$ such that
\begin{equation}
\lVert\mathbf{p}\rVert\leq N^m \quad \text{and} \quad \lVert\Psi\mathbf{p}-\mathbf{q}\rVert \leq N^{-n},
\end{equation}
where $\lVert\cdot\rVert$ denotes the supremum norm, that is, $\lVert x\rVert=\max_{1\leq i\leq k}\lvert x_i\rvert$ for any $\mathbf{x}=(x_1,x_2,\cdots,x_k)\in\R^k.$
Given $0<\mu<1$. After Davenport and Schmidt \cite{DS6970}, we say that $\Psi\in\mathrm{M}_{m\times n}(\R)$ is $\mathrm{DT}_\mu$-improvable if for all sufficiently large $N>0$, there exists nonzero integer vectors $\mathbf{p}\in\Z^n$ and $\mathbf{q}\in\Z^m$ such that
\begin{equation}
\lVert\mathbf{p}\rVert\leq \mu N^m \quad \text{and} \quad \lVert\Psi\mathbf{p}-\mathbf{q}\rVert \leq \mu N^{-n}.
\end{equation}
We say that $\Psi$ is not DT-improvable, if for any $0<\mu<1$, $\Psi$ is not $\mathrm{DT}_\mu$-improvable.\par
In \cite{DS6970}, it was proved that Dirichlet's theorem cannot be improved for Lebesgue almost every $m \times n$ real matrix. In \cite{DS70}, they also proved that Dirichlet's theorem cannot be $(1/4)$-improved for almost every point on the curve $\phi(s)=(s,s^2)$ in $\mathbb{R}^2$. This result was generalized by Baker \cite{Bak78} for almost all points on smooth curves in $\mathbb{R}^2$, and by Bugeaud \cite{Bug02} for almost every point on the curve $\phi(s)=(s, s^2, \cdots, s^k)$ in $\mathbb{R}^k$; in each case the result holds for some small value $0<\mu \leq \epsilon$, where $\epsilon$ depends on the curve.\par
Kleinbock and Weiss \cite{KW08} recast the problem in the language of homogeneous dynamics, and obtained $\epsilon$-improvable results for general measures. Later Shah \cite{Sha09Invention} studied the case $m=1$, and showed that if an analytic curve in $\R^n$ is not contained in any proper affine subspace, then almost every point on the curve is not DT-improvable. Lei Yang \cite{Yan16Proc} studied the case $m=n$, and proved an analogous result for square matrices. These results have been generalized to \emph{supergeneric} curves in $M_{m\times n}(\R)$ in the recent preprint \cite{SY16}, where an inductive algorithm was introduced to define \emph{generic} and \emph{supergeneric} curves. \par
In the meantime, Aka, Breuillard, Rosenzweig and de Saxc\'e \cite{ABRS18} worked on extremality of an analytic submanifold of $M_{m\times n}(\R)$, and found a sharp condition for extremality in terms of a certain family of algebraic sets called \emph{constraining pencils} (see \cite[Definition 1.1]{ABRS18}).\par
Based on \cite{SY16}, and combined with ideas from \cite{Sha09Duke1}\cite{ABRS18}, we replace supergeneric condition by a natural geometric condition, and obtain a sharper result. \par
We first make some preparations. Let $\mathrm{Gr}(m,m+n)$ denote the real Grassmannian variety of $m$-dimensional linear subspaces of $\R^{m+n}$.
\begin{definition}[pencil; c.f. \cite{ABRS18} Definition 1.1]\label{def:pencil}
Given a real vector space $W \subsetneq \R^{m+n}$, and an integer $r\leq m$, we define the pencil $\mathfrak{P}_{W,r}$ to be the set
\begin{equation}
\{V\in\mathrm{Gr}(m,m+n)\colon \mathrm{dim}(V\cap W)\geq r\}.
\end{equation}
\end{definition}
We call $\mathfrak{P}_{W,r}$ a \emph{constraining pencil} if
\begin{equation}\label{eq:constraining_pencil}
\frac{\dim W}{r} < \frac{m+n}{m};
\end{equation}
we call $\mathfrak{P}_{W,r}$ a \emph{weakly constraining pencil} if
\begin{equation}\label{eq:weakly_constraining_pencil}
\frac{\dim W}{r} \leq \frac{m+n}{m}.
\end{equation}
We say that the pencil $\mathfrak{P}_{W,r}$ is \emph{rational} if $W$ is rational, i.e. $W$ admits a basis in $\Q^{m+n}$.
\begin{remark}\label{rem:pencils}
\begin{enumerate}[(1)]
\item If $m$ and $n$ are coprime, then $\frac{m+n}{m}$ is an irreducible fraction, and it follows that \eqref{eq:weakly_constraining_pencil} and \eqref{eq:constraining_pencil} are equivalent. Therefore weakly constraining pencils coincide with constraining pencils in this case.
\item If $m=1$, then (weakly) constraining pencils are proper linear subspaces of $\R\mathbb{P}^n$.
\end{enumerate}
\end{remark}
To avoid confusions, we explain the relationship between our pencils and the pencils in \cite{ABRS18}. Given $W \subsetneq \R^{m+n}$ and $0<r<m$, in \cite{ABRS18} a pencil $\mathcal{P}_{W,r}$ is defined to be an algebraic subset of $M_{m\times(m+n)}(\R)$. More precisely,
\begin{equation}
\mathcal{P}_{W,r} = \left\{ x\in M_{m\times(m+n)}(\R)\colon \dim(xW)\leq r \right\}.
\end{equation}
And a pencil $\mathcal{P}_{W,r}$ is called constraining if
\begin{equation}
\frac{\dim W}{r} > \frac{m+n}{m}.
\end{equation}
Let $x$ be a full rank $m\times(m+n)$ real matrix. For any subspace $E\subset\R^{m+n}$, let $E^{\vee}\subset(\R^{m+n})^{\ast}$ denote the set of linear functionals on $\R^{m+n}$ which vanish on $E$. Then $\dim(xW)\leq r$ if and only if $\dim\left((\ker x)^{\vee}\cap W^{\vee} \right) \geq m-r$. Hence
\begin{equation}
x\in\mathcal{P}_{W,r}\quad\iff\quad (\ker x)^{\vee}\in\mathfrak{P}_{W^{\vee},m-r}.
\end{equation}
Moreover, since $\dim W^{\vee} = m+n-\dim W$, we have
\begin{equation}
\frac{\dim W}{r} > \frac{m+n}{m} \quad \iff \quad \frac{\dim W^{\vee}}{m-r} < \frac{m+n}{m}.
\end{equation}
As explained in \cite[Section 4]{ABRS18}, we don't lose any essential information when passing to kernels. Therefore, our constraining pencils are dual to the constraining pencils in \cite{ABRS18}. We modified the definition to fit into our framework of Schubert varieties. See \Cref{def:Schubert_variety} and \Cref{thm:pencil_equal_Schubert} for more details. \\
To any $\Psi\in M_{m\times n}(\R)$, we attach an $m$-dimensional subspace $V_{\Psi}\subset\R^{m+n}$ which is spanned by the row vectors of the full rank $m\times(m+n)$ matrix
\begin{equation}
\begin{bmatrix}
I_{m\times m}\vert \Psi
\end{bmatrix}.
\end{equation}
\par
Let $\varphi\colon[a,b]\rightarrow M_{m\times n}(\R)$ be an analytic curve. It induces a curve on $\mathrm{Gr}(m,m+n)$ by
\begin{align*}
\Phi\colon[a,b]&\longrightarrow \mathrm{Gr}(m,m+n)\\
s &\longmapsto V_{\varphi(s)}.
\end{align*}
We identify $\mathrm{Gr}(m,m+n)$ with $G/P$, where $G=\SL_{m+n}(\R)$ and $P=P(a)$ is the parabolic subgroup associated with $a(t)=\mathrm{diag}(t^n,\cdots,t^n,t^{-m},\cdots,t^{-m})$. Hence it makes sense to talk about partial flag subvarieties of $\mathrm{Gr}(m,m+n)$. (See \Cref{def:partial_flag_subvariety}.) \par
Now we are ready for our main theorem on DT-improvability.
\begin{theorem}[DT-improvability]\label{thm:Dirichlet}
Let $\varphi\colon[a,b]\rightarrow M_{m\times n}(\R)$ be an analytic curve. Suppose that both of the following hold:
\begin{enumerate}[(A)]
\item The image of $\Phi$ is not contained in any weakly constraining pencil;
\item The image of $\Phi$ is not contained in any proper partial flag subvariety of the Grassmannian variety $\mathrm{Gr}(m,m+n)$ with respect to $a(t)$.
\end{enumerate}
Then for Lebesgue almost every $s\in [a,b]$, $\varphi(s)$ is not DT-improvable.
\end{theorem}
\Cref{thm:Dirichlet} follows from \Cref{thm:equidistribution1} and \Cref{thm:pencil_equal_Schubert} via Dani's correspondence, as explained in \cite{KW08}\cite{Sha09Invention}\cite{Yan16Proc}\cite{SY16}. The proof also shows that for Lebesgue almost every $s\in [a,b]$, $\varphi(s)$ is not DT-improvable along $\mathcal{N}$ (see \cite{Sha10}), where $\mathcal{N}$ is any infinite set of positive integers.
\subsection{Organization of the paper}
In \Cref{sect:linear_stability}, we review the concept of Kempf's one-parameter subgroup, and use Kempf's numerical criterion to prove linear stability. \par
In \Cref{sect:nondivergence}, we review the $(C,\alpha)$-good property defined by Kleinbock and Margulis, and apply linearization technique combined with linear stability to prove non-divergence of translated measures. \par
In \Cref{sect:unipotent_invariance}, we apply the idea of twisting due to Shah, and prove a general result on unipotent invariance. \par
In \Cref{sect:focusing}, we use Ratner's theorem on unipotent flows and Dani-Margulis linearization technique to study the dynamical behavior of trajectories near singular sets, and obtain equidistribution results.\par
In \Cref{sect:Grassmannian_Schubert}, we study the special case of Grassmannians, and use Young diagrams to give a combinatorial description of constraining and weakly constraining pencils. \par
\subsection*{Acknowledgments}
I would like to express my deep gratitude to my advisor Nimish Shah for suggesting this problem, for generously sharing his ideas, and for numerous helpful discussions and constant encouragement. I would like to thank Manfred Einsiedler and Alex Eskin for drawing \cite{Kha15} and \cite{KLP18} to my attention. I would like to thank David Anderson for many helpful discussions and his course on equivariant cohomology, where I learned a lot about Schubert varieties. Thanks are due to Menny Aka, Jayadev Athreya, Asaf Katz, Shi Wang, Barak Weiss and Runlin Zhang for helpful discussions and suggestions.\par
Special thanks to Osama Khalil and Dmitry Kleinbock for discussions which helped me to find out a mistake in an earlier version of this paper. \par
I would like to thank my wife, Yushu Hu, for her unconditional support.
\section{Linear stability and Kempf's one-parameter subgroups}\label{sect:linear_stability}
Let $G=\mathbf{G}(\R)$ be a semisimple connected real algebraic group. If $\delta\colon \mathbb{G}_m \rightarrow \mathbf{G}$ is a homomorphism of real algebraic groups, we call $\delta$ a \emph{multiplicative one-parameter subgroup} of $G$. We associate a parabolic subgroup with $\delta$ as:
\begin{equation}
P(\delta) := \{ g \in G \colon \lim_{t\to\infty} \delta(t)g\delta(t)^{-1} \text{ exists in $G$}\},
\end{equation}
Let $\Gamma(G)$ be the set of the multiplicative one-parameter subgroups of $G$. Following Kempf \cite{Kem78}, we define the \emph{Killing length} of a multiplicative one-parameter subgroup $\delta$ by the equation
\begin{equation}
2\| \delta \|^2 = \mathrm{Trace}[(\mathrm{ad}(\delta_*\mathrm{d}/\mathrm{d}t))^2],
\end{equation}
and it follows from the invariance of the Killing form that the Killing length is $G$-invariant. \par
Now fix a multiplicative one-parameter subgroup $a$ of $G$. We choose and fix a maximal $\R$-split torus $T$ of $G$ containing $\{a(t)\}$. Let $\Gamma(T)$ be the set of the multiplicative one-parameter subgroups of $T$, and $X(T)$ be the set of characters of $T$. We define a pairing as following: if $\chi\in X(T)$ and $\delta\in\Gamma(T)$, $\langle \chi,\delta \rangle$ is the integer which occurs in the formula $\chi(\delta(t)) = t^{\langle \chi,\delta \rangle}$. Let $(\cdot,\cdot)$ denote the positive definite bilinear form on $\Gamma(T)$ such that $(\delta, \delta) = \| \delta \|^2$. \par
By a suitable choice of positive roots $R^+$, we may assume that $a$ is a dominant cocharacter of in $T$. Recall that the set $\Gamma^+(T)$ of dominant cocharacters of $T$ is defined by:
\begin{equation}
\Gamma^+(T) = \{ \delta \in \Gamma(T) \colon \langle \delta,\alpha \rangle \geq 0,\, \forall \alpha\in R^+ \}.
\end{equation}
Let $B$ be the corresponding minimal parabolic subgroup of $G$ whose Lie algebra consists of all the non-positive root spaces.
\par
Let $P=P(a)$ be the parabolic subgroup associated with $a$. Let $W^P$ denote set of minimal length coset representatives of the quotient $W/W_{P}$, where $W = N_G(T)/Z_G(T)$ and $W_{P} = N_{P}(T)/Z_{P}(T)$ are Weyl groups of $G$ and $P$. Then $W$ acts on $\Gamma(T)$ by conjugation: $w\cdot\delta = w\delta w^{-1}$. Denote $\delta^w=w\cdot\delta$. We take the Bruhat order on $W^P$ such that $w' \leq w$ if and only if the closure of the Schubert cell $BwP$ contains $Bw'P$. We note that the Bruhat order coincides with the folding order defined in \cite{KLP18} (See \cite[Remark 3.8]{KLP18}). \par
\begin{definition}[Schubert variety]\label{def:Schubert_variety}
Given an element $w\in W^P$, the standard Schubert variety $X_w$ is the Zariski closure of the Schubert cell $BwP$. A \emph{Schubert variety} is a subvariety of $G/P$ of the form $gX_w$, where $g\in G$ and $w\in W^P$. \par
We say that a Schubert variety $gX_w$ is \textbf{unstable} with respect to $a(t)$ if there exists $\delta\in\Gamma^{+}(T)$ such that $(\delta, a^w) > 0$. We say that $gX_w$ is \textbf{weakly unstable} with respect to $a(t)$ if there exists non-trivial $\delta\in\Gamma^{+}(T)$ such that $(\delta, a^w) \geq 0$. \par
For short, we will just say unstable or weakly unstable Schubert variety if $a(t)$ is clear in the context.
\end{definition} \par
\begin{remark}
In this article, when we project from $G$ to $G/P$, we always take the following map
\begin{align}
\pi_P\colon&G\longrightarrow G/P \nonumber \\
&g \longmapsto g^{-1}P.
\end{align}
When we write $BwP$, we treat it as a subvariety of $G/P$; while $Pw^{-1}B$ is treated as a subset of $G$.
\end{remark}
For $\delta\in \Gamma^{+}(T)$, define the subset $W^+(\delta, a)$ of $W^P$ as
\begin{equation}\label{eq:W_plus}
W^+(\delta, a) = \{ w \in W^P \colon (\delta, a^w) > 0 \},
\end{equation}
and we define $W^{-}(\delta, a), W^{0+}(\delta, a)$ and $W^{0-}(\delta, a)$ similarly, with $<$, $\geq$ and $\leq$ in place of $>$ in \eqref{eq:W_plus} respectively. We note that $W^{+}(\delta,a)$ is a ``metric thickening'' as defined in \cite[Section 3.4]{KLP18}.
\begin{lemma}\label{lem:union_of_Schubert}
\begin{enumerate}[(a)]
\item Let $w'\leq w$ be elements in $W^P$, and $\delta\in\Gamma^{+}(T)$. Then one has $(\delta, a^{w'})\geq (\delta, a^w)$.
\item $\bigsqcup_{w\in W^+(\delta, a)}BwP$ is a finite union of unstable Schubert subvarieties of $G/P$.
\item $\bigsqcup_{w\in W^{0+}(\delta, a)}BwP$ is a finite union of weakly unstable Schubert subvarieties of $G/P$.
\end{enumerate}
\end{lemma}
\begin{proof}
Both (b) and (c) follow from (a). For a proof of (a), see e.g. \cite[Lemma 3.4]{KLP18}.
\end{proof}
Let $\rho: G\rightarrow \GL(V)$ be any finite dimensional linear representation of $G$. Let us recall some notions from geometric invariant theory (see e.g. \cite{MFK94} for more details). A nonzero vector $v$ is called \emph{unstable} if the closure of the $G$-orbit $Gv$ contains the origin. $v$ is called \emph{semistable} if it is not unstable. For any $v\in V\backslash\{0\}$ and $\delta\in\Gamma(G)$, by \cite[Lemma 1.2]{Kem78} we can write $v=\sum v_i$ where $\delta(t)v_i = t^iv_i$. Define the numerical function $m(v,\delta)$ to be the maximal\footnote[1]{It is ``minimal'' in Kempf's original definition. Since we are taking limit as $t$ tends to $\infty$ instead of $0$, our numerical function is actually opposite to Kempf's.} $i$ such that $v_i\neq 0$.\par
By a theorem of Kempf (see \cite[Theorem 4.2]{Kem78}), the function $m(v, \delta)/\|\delta\|$ has a negative minimum value $B_v$ on the set of non-trivial multiplicative one-parameter subgroups $\delta$. Let $\Lambda(v)$ denote the set of primitive multiplicative one-parameter subgroup $\delta$ such that $m(v, \delta) = B_v\cdot\|\delta\|$. Kempf \cite[Theorem 4.2]{Kem78} shows that the parabolic subgroup $P(\delta)$ does not depend on the choice of $\delta\in\Lambda(v)$, which is denoted by $P(x)$. Moreover, $\Lambda(v)$ is a principal homogeneous space under conjugation by the unipotent radical of $P(x)$. In particular, for any $\delta$ in $\Lambda(v)$ and $b$ in $P(x)$, we know that $b\delta b^{-1}$ is also contained in $\Lambda(v)$. \\
For $v\in V\backslash\{0\}$, define
\begin{equation}\label{eq:definiton_of_zero_unstable}
G(v, V^{-}(a)) = \{ g\in G \colon gv \in V^{-}(a) \},
\end{equation}
where
\begin{equation}
V^{-}(a) = \{ v\in V \colon \lim_{t\to\infty}a(t)v = 0 \}.
\end{equation}
As noted in \cite[Section 3.3]{Kha15}, though the limits in \cite{Kem78} are defined algebraically, they coincide with limits in the Hausdorff topology induced from the usual topology on $\R$, by \cite[Lemma 1.2]{Kem78}. \par
Now we proceed to the main result of this section.
\begin{proposition}\label{prop:basic_lemma}
For any $v\in V\backslash\{0\}$, there exits $\delta_0\in\Gamma^{+}(T)$ and $g_0\in G$ such that
\begin{equation}\label{eq:basic_inclusion}
G(v,V^{-}(a)) \subset \bigsqcup_{w\in W^+(\delta_0, a)}Pw^{-1}Bg_0^{-1}.
\end{equation}
\end{proposition}
\begin{proof}
By definition we have the following identities because of $G$-equivariance:
\begin{equation}
G(gv, V^{-}(a)) = G(v, V^{-}(a))g^{-1},\quad \forall g\in G;
\end{equation}
\begin{equation}\label{eq:G_equivarianc_Lambda}
\Lambda(gv) = g\Lambda(v)g^{-1},\quad \forall g\in G.
\end{equation} \par
If $v$ is semistable, then $G(v, V^{-}(a))$ is empty, and the conclusion trivially holds. From now on we assume that $v$ is unstable, and thus $\Lambda(v)$ is non-empty. Take $\delta_1\in\Lambda(v)$, then there exists $g_0\in G$ and $\delta_0\in\Gamma^{+}(T)$ such that $g_0^{-1}\delta_1g_0 = \delta_0$. It follows from \eqref{eq:G_equivarianc_Lambda} that $\delta_0\in\Lambda(g_0^{-1}v)$.\par
We prove by contradiction. Suppose that \eqref{eq:basic_inclusion} does not hold. Considering the Bruhat decomposition
\begin{equation}
G = \bigsqcup_{w\in W^P} Pw^{-1}B,
\end{equation}
we can take $g\in G(g_0^{-1}v, V^{-}(a))$ such that it can be written as
\begin{equation}
g = pw^{-1}b, \text{ where } p\in P,\, w\in W^{0-}(\delta_0, a),\, b\in B.
\end{equation}
Write $v' = bg_0^{-1}v$. In view of \eqref{eq:G_equivarianc_Lambda}, by \cite[Theorem 4.2(3)]{Kem78} we have $\Lambda(g_0^{-1}v)=\Lambda(v')$. Hence $\delta_0$ is an element in $\Lambda(v')$. \par
We also have $v'\in V^{-}(a^w)$. Indeed, $gg_0^{-1}v\in V^-(a)$ implies that $pw^{-1}v'\in V^{-}(a)$. Since $V^{-}(a)$ is $P$-invariant, we know that $w^{-1}v'\in V^{-}(a)$. Hence $v'\in V^{-}(a^w)$. \par
Take a large integer $N$, we define $\delta_{N} = N\delta_0 + a^w$. We claim that for a sufficiently large $N$, one has
\begin{equation}\label{eq:destablize_faster}
\frac{m(v', \delta_{N})}{\| \delta_{N} \|} < \frac{m(v', \delta_0)}{\| \delta_0 \|},
\end{equation}
and this will contradict the fact that $\delta_0\in\Lambda(v')$.\par
To prove the claim, consider the weight space decomposition $V = \bigoplus V_{\chi}$, where $T$ acts on $V_{\chi}$ by multiplication via the character $\chi$ of $T$. It suffices to prove that for any $\chi$ such that the projection of $v'$ on $V_{\chi}$ is nonzero, one has
\begin{equation}\label{eq:higher_speed}
\frac{\langle \chi, \delta_{N} \rangle}{\|\delta_{N} \|} <
\frac{\langle \chi, \delta_0 \rangle}{\| \delta_0 \|}.
\end{equation}
To prove \eqref{eq:higher_speed}, we define an auxiliary function:
\begin{equation}
\begin{split}
f(s) &= \frac{\langle \chi, \delta_0 + s\cdot a^w \rangle ^ 2}{\| \delta_0 + s \cdot a^w \| ^ 2}\\
&= \frac{\langle \chi , \delta_0 \rangle ^ 2 + 2s\langle \chi, \delta_0 \rangle\langle \chi, a^w \rangle + s^2 \langle \chi, a^w \rangle ^2}
{(\delta_0, \delta_0) + 2s(\delta_0, a^w) + s^2(a^w, a^w)}
\end{split}
\end{equation}
Compute its derivative at 0:
\begin{equation}
f'(0) = \frac{2\langle \chi,\delta_0\rangle\langle\chi,a^w\rangle(\delta_0, \delta_0) - 2(\delta_0, a^w)\langle\chi, \delta_0\rangle^2}
{(\delta_0, \delta_0)^2}
\end{equation}
Since $v'\in V^{-}(a^w)$, we know that $\langle \chi, a^w \rangle < 0$. Since $\delta_0\in\Lambda(v')$, we know that $\langle \chi, \delta_0 \rangle < 0$. Also by the choice of $w$ we know that $(\delta_0, a^w) \leq 0$. Combining the above one gets $f'(0) > 0$. Hence for $N$ large we have
\begin{equation}\label{eq:temp14}
f(1/N) > f(0),
\end{equation}
and \eqref{eq:higher_speed} follows because each side of \eqref{eq:temp14} is the square of each side of \eqref{eq:higher_speed}. Therefore \eqref{eq:destablize_faster} holds, contradicting the fact that $\delta_0\in\Lambda(v')$.
\end{proof}
Now we are ready to prove \Cref{thm:linear_stability}.
\begin{proof}[Proof of \Cref{thm:linear_stability}]
We prove by contradiction. Suppose that for all $C>0$, there exist $t$ and $v$ such that \eqref{eq:linear_stablility} does not hold. We take a sequence $C_i\to 0$. Then after passing to a subsequence we can find $t_i\to\infty$ and a sequence $(v_i)_{i\in\N}$ in $V$ such that
\begin{equation}
\sup_{s\in [a,b]}\lVert a(t_i)\phi(s)v_i \rVert < C_i\lVert v_i \rVert.
\end{equation}
Without loss of generality we may assume that $\lVert v_i \rVert = 1$. Then after passing to a subsequence, we may assume that $v_i \to v_0$. Hence we have
\begin{equation}
\sup_{s\in [a,b]}\lVert a(t_i)\phi(s)v_0 \rVert \stackrel{t_i\to\infty}{\longrightarrow} 0.
\end{equation}
Therefore $\phi(s)v_0$ is contained in $V^{-}(a)$ for all $s\in[a,b]$, and it follows that the image of $\phi$ is contained in $G(v_0, V^{-}(a))$. (See \eqref{eq:definiton_of_zero_unstable}.) By \Cref{lem:union_of_Schubert}(b) and \Cref{prop:basic_lemma}, the image of $G(v_0, V^{-}(a))$ under $\pi_P$ in $G/P$ is a finite union of unstable Schubert varieties. But $\phi$ is analytic, which implies that the image of $\widetilde{\phi}$ is contained in one single unstable Schubert variety. This contradict our assumption on $\phi$.
\end{proof}
\Cref{prop:basic_lemma} and \Cref{thm:linear_stability} will play a central role in proving the non-divergence of translated measures. To handle non-focusing, one needs a slightly generalized version, motivated by the work of Richard and Shah \cite[Section 6]{RS16}. We need the following result due to Kempf.
\begin{lemma}[\cite{Kem78} Lemma 1.1(b)]\label{lem:Kempf_reduction}
Let $G$ be a connected reductive algebraic group over a field $k$, and $X$ be any affine $G$-scheme. If $S$ is a closed $G$-subscheme of $X$, then there is a $G$-equivariant morphism $f:X\rightarrow W$, where $W$ is a representation of $G$, such that $S$ is the scheme-theoretic inverse image $f^{-1}(0)$ of the reduced closed subscheme of $W$ supported by zero.
\end{lemma} \par
In view of Kempf's \Cref{lem:Kempf_reduction}, the following is a corollary of \Cref{prop:basic_lemma}.
\begin{corollary}\label{cor:basic_lemma1}
Let the notations be as in the beginning of this section. Let $S$ be the real points of any $G$-subscheme of $V$. For any $v\in V$, define the following subset of $G$:
\begin{equation}\label{eq:cor_basic_lemma1}
G(v, S,a) = \{ g\in G \colon \lim_{t\to\infty}a(t)gv\in S \}.
\end{equation}
Then for any $v\in V\backslash S$, there exists $\delta_0\in\Gamma^{+}(T)$ and $g_0\in G$ such that
\begin{equation}
G(v, S,a) \subset \bigsqcup_{w\in W^+(\delta_0, a)}Pw^{-1}Bg_0^{-1}.
\end{equation}
\end{corollary}
\begin{proof}
By \Cref{lem:Kempf_reduction}, there exist a $G$-equivariant morphism $f:V\rightarrow W$ where $f^{-1}(0) = S$. Hence it follows from the definition that
\begin{equation}
G(v, S, a) \subset G(f(v), W^{-}(a)).
\end{equation}
Now it remains to apply \Cref{prop:basic_lemma} for $W$ and $f(v)$.
\end{proof}
Now we present the following variance of \Cref{prop:basic_lemma}.
\begin{proposition}\label{prop:basic_lemma2}
Let $v\in V$ such that the $G$-orbit $Gv$ is not closed. Define
\begin{equation}
G(v, V^{0-}(a)) = \{ g\in G : gv\in V^{0-}(a) \},
\end{equation}
where
\begin{equation}
V^{0-}(a) = \{ v\in V : \lim_{t\to \infty} a(t)v \text{ exists} \}.
\end{equation}
Then there exists $\delta_0\in\Gamma^{+}(T)$ and $g_0\in G$ such that
\begin{equation}\label{eq:basic_lemma2}
G(v, V^{0-}(a)) \subset \bigsqcup_{w\in W^{0+}(\delta_0, a)}Pw^{-1}Bg_0^{-1}.
\end{equation}
\end{proposition}
\begin{proof}
Let $S=\partial(Gv)$. Since any $G$-orbit is open in its closure, we know that $S$ is closed and $G$-invariant. By \Cref{lem:Kempf_reduction}, there exists a $G$-equivariant morphism $f:V\rightarrow W$ where $f^{-1}(0) = S$. Notice that $f(v)$ is unstable in $W$. We claim that
\begin{equation}\label{eq:temp10}
G(f(v),W^{0-}(a))\subset \bigsqcup_{w\in W^{0+}(\delta_0, a)}Pw^{-1}Bg_0^{-1}.
\end{equation}
\par
To prove the claim, we argue with $W$ and $f(v)$ in exactly the same way as in the proof of \Cref{prop:basic_lemma}. The only difference is the following. When showing $f'(0)>0$, one needs $\langle\chi, a^w\rangle < 0$ and $(\delta_0, a^w)\leq 0$ there; but here one has $\langle\chi, a^w\rangle \leq 0$ and $(\delta_0, a^w)< 0$, which also implies that $f'(0)>0$. Hence \eqref{eq:temp10} holds. \par
Finally, since $f$ is $G$-equivariant, we have $f(V^{0-})\subset W^{0-}$. Hence
\begin{equation}
G(v, V^{0-}(a)) \subset G(f(v),W^{0-}(a)).
\end{equation}
Therefore \eqref{eq:basic_lemma2} holds.
\end{proof}
\section{Non-divergence of the limiting distribution}\label{sect:nondivergence}
Let $G=\mathbf{G}(\R)$ be a connected semisimple real algebraic group, and $L$ be a real Lie group containing $G$. Let $\{a(t)\}_{t\in\R^{\times}}$ be a multiplicative one-parameter subgroup of $G$ with non-trivial projection on each simple factor of $G$. Let $P=P(a)$ be the parabolic subgroup of $G$ whose real points consists of the elements $g\in G$ such that the limit $\lim_{t\to \infty}a(t)ga(t)^{-1}$ exists. Let $\phi \colon I = [a,b] \rightarrow G$ be an analytic map, and let $\pi_P \colon G \rightarrow G/P$ be the projection which maps $g$ to $g^{-1}P$. Then $\widetilde{\phi} = \pi_P\circ\phi$ is an analytic curve on $G/P$. In this section we assume that the image of $\widetilde{\phi}$ is not contained in any unstable Schubert variety of $G/P$ with respect to $a(t)$. \par
Let $x_0 = l\Lambda \in L/\Lambda$. We will assume that the orbit of $x_0$ under $G$ is dense in $L/\Lambda$; that is $\overline{Gx_0} = L/\Lambda$. Let $t_i\to \infty$ be any sequence in $\R_{>0}$. Let $\mu_i$ be the parametric measure supported on $a(t_i)\phi(I)x_0$, that is, for any compactly supported function $f \in C_c(L/\Lambda)$ one has
\begin{equation}\label{eq:mu_i}
\int_{L/\Lambda}f\,d\mu_i = \frac{1}{|I|}\int_{I}f(a(t_i)\phi(s)x_0)\,\mathrm{d}s.
\end{equation}
\begin{theorem}\label{thm:non_divergence}
Given $\epsilon > 0$ there exists a compact set $\mathcal{F} \subset L/\Lambda$ such that $\mu_i(\mathcal{F}) \geq 1-\epsilon$ for all large $i\in\N$.
\end{theorem}
This theorem will be proved via linearization technique combined with \Cref{thm:linear_stability}. We follow \cite[Section 3]{Sha09Invention} closely, as most of the arguments there work not only for $G=\SL_n(\R)$ but also for general $G$.
\begin{definition}\label{def:V}
Let $\mathfrak{l}$ denote the Lie algebra of $L$, and denote $d = \mathrm{dim}\,L$. We define
\begin{equation*}
V = \bigoplus_{i=1}^d \bigwedge\nolimits^{\!i} \mathfrak{l},
\end{equation*}
and let $L$ act on $V$ via $\bigoplus_{i=1}^{d}\bigwedge\nolimits^{\!i}\mathrm{Ad}(L)$. This defines a linear representation of $L$ (and of $G$ by restriction):
\begin{equation*}
L \rightarrow \GL(V).
\end{equation*}
\end{definition}
The following theorem due to Kleinbock and Margulis is the basic tool to prove that there is no escape of mass to infinity:
\begin{theorem}[see \cite{Dan84}, \cite{KM98} and \cite{Sha09Duke1}]\label{thm:nondivergence_criterion}
Fix a norm $\lVert\cdot\rVert$ on $V$. There exist finitely many vectors $v_1, v_2, \cdots, v_r \in V$ such that for each $i = 1, 2, \cdots, r$, the orbit $\Lambda v_i$ is discrete, and moreover, the following holds: for any $\epsilon > 0$ and $R > 0$, there exists a compact set $\mathcal{F} \subset L/\Lambda$ such that for any $t > 0$ and any subinterval $J \subset I$, one of the following holds:
\begin{enumerate}[(I)]
\item There exist $\gamma \in \Lambda$ and $j \in \{ 1, \cdots, r \}$ such that
\begin{equation*}
\sup_{s\in J}\lVert a(t)\phi(s)l\gamma v_j\rVert < R;
\end{equation*}
\item
\begin{equation*}
\lvert \{ s\in J \colon a(t)\phi(s)x_0 \in K \} \rvert \geq (1-\epsilon)\lvert J \rvert.
\end{equation*}
\end{enumerate}
\end{theorem}
The key ingredient of the proof, as explained in \cite[Section 3.2]{Sha09Invention} and \cite[Section 2.1]{Sha09Duke1}, is the following growth property called the $(C, \alpha)$-good property, which is due to \cite[Proposition 3.4]{KM98}. Following Kleinbock and Margulis, we say that a function $f\colon I \rightarrow \R$ is $(C, \alpha)$-good if for any subinterval $J \subset I$ and any $\epsilon > 0$, the following holds:
\begin{equation*}
\lvert \{ s\in J\colon \lvert f(s) \rvert < \epsilon \} \rvert
\leq C\left(\frac{\epsilon}{\sup_{s\in J}\lvert f(s) \rvert}\right)^{\alpha}\lvert J \rvert.
\end{equation*} \\
Now we are ready to prove the main result of this section.
\begin{proof}[Proof of \Cref{thm:non_divergence}]
Take any $\epsilon > 0$. Take a sequence $R_k \to 0$ as $k \to \infty$. For each $k \in \N$, let $\mathcal{F}_k \subset L/\Lambda$ be a compact set as determined by \Cref{thm:nondivergence_criterion} for these $\epsilon$ and $R_k$. If the theorem fails to hold, then for each $k \in \N$ we have $\mu_i(\mathcal{F}_k) > 1 - \epsilon$ for infinitely many $i \in \N$. Therefore after passing to a subsequence of $\{ \mu_i \}$, we may assume that $\mu_i(\mathcal{F}_i) < 1-\epsilon$ for all $i$. Then by \Cref{thm:nondivergence_criterion}, after passing to a subsequence, we may assume that there exists $v_0$ and $\gamma_i\in\Lambda$ such that
\begin{equation*}
\sup_{s\in I} \lVert a(t_i)\phi(s)l\gamma_iv_0 \rVert \leq R_i\stackrel{i\to\infty}{\longrightarrow} 0.
\end{equation*}
Since $\Lambda\cdot v_0$ is discrete, there exists $r_0 > 0$ such that $\lVert l\gamma_iv_0 \rVert \geq r_0$ for each i. We put $v_i = l\gamma_iv_0/\lVert l\gamma_iv_0 \rVert$. Then $v_i \rightarrow v \in V$ and $\lVert v \rVert = 1$. Therefore
\begin{equation}
\sup_{s\in I} \lVert a(t_i)\phi(s)v_i \rVert \leq R_i/r_0\stackrel{i\to\infty}{\longrightarrow} 0.
\end{equation}
Then it follows that
\begin{equation}\label{eq:uniform_shrinking}
\sup_{s\in I} \lVert a(t_i)\phi(s)v \rVert \stackrel{i\to\infty}{\longrightarrow} 0.
\end{equation}
This contradict \Cref{thm:linear_stability}.
\end{proof}
As a consequence of \Cref{thm:non_divergence}, we deduce the following:
\begin{corollary}\label{cor:weak_limit_exists}
After passing to a subsequence, $\mu_i\to\mu$ in the space of probability measures on $L/\Lambda$ with respect to the weak-* topology.
\end{corollary} \par
We note that \Cref{thm:non_divergence_introduction} follows from \Cref{thm:non_divergence}.
\section{Invariance under a unipotent flow}\label{sect:unipotent_invariance}
Let $G = \mathbf{G}(\R)$ be a connected semisimple real algebraic group, and $\{a(t)\}_{t\in\R^{\times}}$ be a multiplicative one-parameter subgroup of $G$ with non-trivial projection on each simple factor of $G$. Define
\begin{equation}
P = \{g\in G \colon \lim_{t\to\infty}a(t)ga(t)^{-1}\text{ exists}\}.
\end{equation}
Let $\mathfrak{X}$ be a locally compact second countable Hausdorff topological space, with a continuous $G$-action. Let $\phi\colon I = [a, b]\rightarrow G$ be an analytic curve, whose projection under $g\mapsto g^{-1}P$ on $G/P$ is non-trivial. Let $\mathfrak{g}$ denote the Lie algebra of $G$. \par
Since the exponential map $\exp\colon\mathfrak{g}\rightarrow G$ is a local homeomorphism, we can take a sufficiently small $\eta>0$ such that for any $s\in I$ and $0<\xi<\eta$, there exists $\Psi(s,\xi)$ in $\mathfrak{g}$ such that
\begin{equation}\label{eq:def_Psi}
\phi(s+\xi)\phi(s)^{-1}= \exp \Psi(s,\xi).
\end{equation}
Moreover, $\Psi$ is an analytic map in both $s$ and $\xi$.
\begin{lemma}\label{lem:converge_to_one_direction}
There exists $m>0$ and a nilpotent element $Y_s$ in $\mathfrak{g}$ such that for all but finitely many $s\in I$,
\begin{equation}\label{eq:converge_to_one_direction}
\mathrm{Ad}\,a(t)\,\Psi(s,t^{-m})\rightarrow Y_s,\quad t\to\infty.
\end{equation}
Moreover, one can assume that the map $s\rightarrow Y_s$ is analytic, and the convergence is uniform in $s$.
\end{lemma}
\begin{proof}
Since $\Psi$ is an analytic map in both $s$ and $\xi$, we can write
\begin{equation}\label{eq:temp11}
\Psi(s,\xi) = \sum_{i=1}^{\infty}\xi^i\psi_i(s),
\end{equation}
where $\psi_i\colon I\rightarrow\mathfrak{g}$ is analytic for each $i$. \par
Notice that $\mathrm{Ad}\,a(t)$ is semisimple and acts on the finite dimensional vector space $\mathfrak{g}$, then for each $i$ there exist $m_i\in\Z$ such that
\begin{equation}\label{eq:temp12}
\mathrm{Ad}\,a(t)\psi_i(s) = \sum_{j\leq m_i}t^j \psi_{i,j}(s),
\end{equation}
where $\psi_{i,j}(s)$ is analytic in $s$, and $\psi_{i,m_i}(s)\neq 0$ for all but finitely many $s\in I$. Since the projection of $\phi$ on $G/P$ is non-trivial, there exists $i$ such that $m_i>0$.\par
Combining \eqref{eq:temp11}\eqref{eq:temp12}, we get
\begin{equation}\label{eq:temp13}
\mathrm{Ad}\,a(t)\, \Psi(s,\xi) = \sum_{i=0}^{\infty}\sum_{j\leq m_i}t^j\xi^i\psi_{i,j}(s).
\end{equation} \par
Now set $m=\max_{i\geq 1} \{m_i/i\}$. Since $m_i$ are all eigenvalues of $\mathrm{Ad}\,a(t)$, they are uniformly bounded from above. Hence we know that $m$ exists and $m>0$. Denote $I=\{i\geq 1\colon m_i/i = m \}$, and we see that $I$ is a finite set. We set
\begin{equation}
Y_s = \sum_{i\in I}\psi_{i,m_i}(s).
\end{equation}
Since the eigenvalues of $\mathrm{Ad}\,a(t)$ acting on $Y_s$ are all positive, $Y_s$ is nilpotent. \par
In view of \eqref{eq:temp13},
\begin{equation}
\mathrm{Ad}\,a(t)\, \Psi(s,t^{-m}) = Y_s + \sum_{j-im<0}t^{j-im}\psi_{i,j}(s),
\end{equation}
and \eqref{eq:converge_to_one_direction} follows.
\end{proof}
We could then twist $Y_s$ into one direction due to the following lemma.
\begin{lemma}
There are only finitely many $G$-conjugacy classes of the nilpotent elements in the Lie algebra $\mathfrak{g}$ of $G$.
\end{lemma}
\begin{proof}
This result has been proved for groups over the complex numbers $\C$ (see \cite{Ric67}). Let $X$ be any non-zero nilpotent element in $\mathfrak{g}$. Now it remains to show that there are only finitely many $\mathbf{G}(\R)$-orbits in the real points of $\mathbf{G}(\C)\cdot X$. Let $\mathbf{H}$ be the stabilizer of $X$ in $\mathbf{G}$. Then $\mathbf{H}$ is an algebraic group defined over $\R$. It is well known that the $\mathbf{G}(\R)$-orbits in $(\mathbf{G}/\mathbf{H})(\R)$ are parametrized by the Galois cohomology $H^1(\mathrm{Gal}(\C/\R),\mathbf{H}(\C))$. Then the statement of the lemma follows from the finiteness of $H^1(\mathrm{Gal}(\C/\R),\mathbf{H}(\C))$, which is guaranteed by \cite[Theorem 6.14]{PR94}.
\end{proof}
Since there are only finitely many conjugacy classes of nilpotent elements in $\mathfrak{g}$, up to at most finitely many points we may assume that all the $Y_s$ are in the same conjugacy class. Hence there exists $w_0$ in $\mathfrak{g}$, and $\delta(s)$ in $G$ which is also analytic in $s$, such that for all but finitely many $s\in I$ one has
\begin{equation}
\mathrm{Ad}(\delta(s))\cdot Y_s = w_0.
\end{equation}
Define the unipotent one-parameter subgroup of $G$ as
\begin{equation}
W = \{\exp(tw_0)\colon t\in\R\}.
\end{equation}
\par
Let $(t_i)_{i\in\N}$ be a sequence in $\R$ such that $t_i\to\infty$ as $i\to\infty$. Let $x_i\to x$ a convergent sequence in $\mathfrak{X}$. For each $i\in\N$, let $\lambda_i$ be the probability measure on $\mathfrak{X}$ such that
\begin{equation}\label{eq:def_lambda_i}
\int_{\mathfrak{X}} f\,\mathrm{d}\lambda_i = \frac{1}{\lvert I \rvert}\int_{s\in I}f(\delta(s) a(t_i)\phi(s)x_i)\,\mathrm{d}s,\quad\forall f\in \mathrm{C}_c(\mathfrak{X}).
\end{equation}
The following theorem is the main result of this section. The new idea here due to Nimish Shah is that we can actually twist the curve after translating by $a(t)$.
\begin{theorem}\label{thm:unipotent_invariance}
Suppose that $\lambda_i\to\lambda$ in the space of finite measures on $\mathfrak{X}$ with respect to the weak-* topology, then $\lambda$ is invariant under $W$.
\end{theorem}
\begin{proof}
Given $f\in\mathrm{C}_c(\mathfrak{X})$ and $\epsilon > 0$. Since $f$ is uniformly continuous, there exists a neighborhood $\Omega$ of the neutral element in $G$ such that
\begin{equation}\label{eq:uniform_continuity}
\lvert f(\omega y) - f(y) \rvert < \epsilon,\quad\forall \omega\in\Omega,\,\forall y\in \mathfrak{X}.
\end{equation}
Define
\begin{equation}
\Omega' = \bigcap_{s\in I} \delta(s)^{-1}\Omega\delta(s),
\end{equation}
and $\Omega'$ is non-empty and open because $\{\delta(s)\}_{s\in I}$ is compact.\par
By \Cref{lem:converge_to_one_direction}, there exists $T>0$ such that for all $t>T$ and for all but finitely many $s\in I$, there exists $\omega_{t,s}\in\Omega'$ such that
\begin{equation}
a(t)\exp\Psi(s,t^{-m})a(t)^{-1} = \omega_{t,s} \exp Y_s.
\end{equation}
Take $\xi_i=t_i^{-m}$. In view of \eqref{eq:def_Psi}, for $i$ large enough we have
\begin{equation}
\phi(s+\xi_i) = \exp\Psi(s,\xi_i)\phi(s).
\end{equation}
Hence there exists $i_0\in\N$ such that for all $i>i_0$,
\begin{equation}
\begin{split}
\delta(s)a(t_i)\phi(s+\xi_i) &= \delta(s)a(t_i)\exp\Psi(s,\xi_i)\phi(s)\\
&= \delta(s)\omega_{t_i,s}\exp Y_sa(t_i)\phi(s)\\
&=\left(\delta(s)\omega_{t_i,s}\delta(s)^{-1}\right)\delta(s)\exp Y_sa(t_i)\phi(s)\\
&=\left(\delta(s)\omega_{t_i,s}\delta(s)^{-1}\right)(\exp w_0)\delta(s)a(t_i)\phi(s)\\
&\in \Omega(\exp w_0)\delta(s)a(t_i)\phi(s).
\end{split}
\end{equation}
By \eqref{eq:uniform_continuity} we know that for all but finitely many $s\in I$,
\begin{equation}
\lvert f((\exp w_0)\delta(s)a(t_i)\phi(s)x_i) - f(\delta(s)a(t_i)\phi(s+\xi_i)x_i) \rvert < \epsilon.
\end{equation}
It follows that for all $i>i_0$,
\begin{equation}
\left\vert \frac{1}{\lvert I \rvert}\int_{I}f((\exp w_0)\delta(s)a(t_i)\phi(s)x_i)\,\mathrm{d}s - \frac{1}{\lvert I \rvert}\int_{I}f(\delta(s)a(t_i)\phi(s+\xi_i)x_i)\,\mathrm{d}s \right\vert < \epsilon.
\end{equation}
On the other hand, since $f$ is bounded on $\mathfrak{X}$, there exists $i_1\in\N$ such that for all $i > i_1$,
\begin{equation}
\left\vert\frac{1}{\lvert I \rvert}\int_{I}f(\delta(s)a(t_i)\phi(s+\xi_i)x_i)\,\mathrm{d}s - \frac{1}{\lvert I \rvert}\int_{I}f(\delta(s)a(t_i)\phi(s)x_i)\,\mathrm{d}s \right\vert < \epsilon.
\end{equation}
Combining the above two equations we get
\begin{equation}
\left\vert \frac{1}{\lvert I \rvert}\int_{I}f((\exp w_0)\delta(s)a(t_i)\phi(s)x_i)\,\mathrm{d}s - \frac{1}{\lvert I \rvert}\int_{I}f(\delta(s)a(t_i)\phi(s)x_i)\,\mathrm{d}s \right\vert < 2\epsilon.
\end{equation}
Therefore, for $i$ large enough we have
\begin{equation}
\left\vert\int_{\mathfrak{X}} f((\exp w_0)\cdot x)\,\mathrm{d}\lambda_i - \int_{\mathfrak{X}} f(x)\,\mathrm{d}\lambda_i \right\vert < 2\epsilon.
\end{equation}
Taking $i\to\infty$,
\begin{equation}
\left\vert\int_{\mathfrak{X}} f((\exp w_0)\cdot x)\,\mathrm{d}\lambda - \int_{\mathfrak{X}} f(x)\,\mathrm{d}\lambda \right\vert \leq 2\epsilon.
\end{equation}
Since $\epsilon$ is arbitrary, we conclude that $\lambda$ is $\exp w_0$-invariant. \par
If we replace $w_0$ with any scalar multiple of $w_0$, the above arguments still work. Hence $\lambda$ is invariant under $W = \{\exp(tw_0)\colon t\in\R\}$.
\end{proof}
\section{Dynamical behavior of translated trajectories near singular sets}\label{sect:focusing}
Let notations be as in \Cref{sect:nondivergence}. Recall that the image of $\widetilde{\phi}$ is not contained in any unstable Schubert varieties of $G/P$ with respect to $a(t)$. Let $\{\lambda_i:i\in\N\}$ be the sequence of probability measures on $L/\Lambda$ as define in \eqref{eq:def_lambda_i}, where we take $\mathfrak{X} = L/\Lambda$ and $x_i=x_0$. Due to \Cref{thm:non_divergence}, by passing to a subsequence we assume that $\lambda_i\to\lambda$ as $i\to\infty$, where $\lambda$ is a probability measure on $L/\Lambda$. By \Cref{thm:unipotent_invariance}, $\lambda$ is invariant under a unipotent subgroup $W$. We would like to describe the limit measure $\lambda$ using the description of ergodic invariant measures for unipotent flows on homogeneous spaces due to Ratner \cite{Rat91}. We follow the treatment in \cite[Section 4]{Sha09Duke1}. \par
\subsection{Ratner's theorem and linearization technique}
Let $\pi \colon L \rightarrow L/\Lambda$ denote the natural quotient map. Let $\mathcal{H}$ denote the collection of closed connected subgroups $H$ of $L$ such that $H\cap\Lambda$ is a lattice in $H$, and suppose that a unique unipotent one-parameter subgroup of $H$ acts ergodically with respect to the $H$-invariant probability measure on $H/H\cap\Lambda$. Then $\mathcal{H}$ is a countable collection (see \cite{Rat91}).\par
For a closed connected subgroup $H$ of $L$, define
\begin{equation}
N(H,W) = \{g\in L\colon g^{-1}Wg \subset H \}.
\end{equation}
Now, suppose that $H\in\mathcal{H}$. We define the associated singular set
\begin{equation}
S(H,W) = \bigcup_{\stackrel {F\in\mathcal{H}}{F\subsetneq H}}N(F,W).
\end{equation}
Note that $N(H,W)N_L(H)=N(H,W)$. By \cite[Proposition 2.1, Lemma 2.4]{MS95},
\begin{equation}
N(H,W)\cap N(H,W)\gamma \subset S(H,W), \; \forall \gamma\in\Lambda\backslash N_G(H).
\end{equation}
By Ratner's theorem \cite[Theorem 1]{Rat91}, as explained in \cite[Theorem 2.2]{MS95}, we have the following.
\begin{theorem}[Ratner]
Given the $W$-invariant probability measure $\lambda$ on $L/\Lambda$, there exists $H\in\mathcal{H}$ such that
\begin{equation}\label{eq:positive_limit_measure_on_singular_set}
\lambda(\pi(N(H,W)))>0 \quad \text{and} \quad \lambda(\pi(S(H,W)))=0.
\end{equation}
Moreover, almost every $W$-ergodic component of $\lambda$ on $\pi(N(H,W))$ is a measure of the form $g\mu_H$, where $g\in N(H,W)\backslash S(H,W)$ and $\mu_H$ is a finite $H$-invariant measure on $\pi(H)\cong H/H\cap\Lambda$. In particular if $H$ is a normal subgroup of $L$ then $\lambda$ is $H$-invariant.
\end{theorem}\par
Let $V$ be as in \Cref{sect:nondivergence}. Let $d = \dim H$, and fix $p_H \in \bigwedge^d\mathfrak{h}\backslash \{0\}$. Due to \cite[Theorem 3.4]{DM93}, the orbit $\Lambda p_H$ is a discrete subset of $V$. We note that for any $g\in N_L(H)$, $gp_H = \det(\mathrm{Ad}\,g\vert_\mathfrak{h})p_H$. Hence the stabilizer of $p_H$ in $L$ equals
\begin{equation}
N_L^1(H) := \{ g\in N_L(H)\colon \det(\mathrm{Ad}\,g\vert_\mathfrak{h})=1\}.
\end{equation}
Recall that $\mathrm{Lie}(W)=\R w_0$. Let
\begin{equation}
\mathcal{A} = \{ v\in V \colon v\wedge w_0 = 0 \},
\end{equation}
where $V$ is defined in \Cref{def:V}. Then $\mathcal{A}$ is a linear subspace of $V$. We observe that
\begin{equation}
N(H,W) = \{ g\in L\colon g\cdot p_H \in \mathcal{A} \}.
\end{equation}\par
Recall that $x_0=l\Lambda\in L/\Lambda$. Using the fact that $\phi$ is analytic, we obtain the following consequence of the linearization technique and $(C,\alpha)$-good property (see \cite{Sha09Duke1}\cite{Sha09Invention}\cite{Sha10}).
\begin{proposition}\label{prop:nonfocusing_creterion}
Let $C$ be a compact subset of $N(H,W)\backslash S(H,W)$. Given $\epsilon > 0$, there exists a compact set $\mathcal{D}\subset\mathcal{A}$ such that, given a relatively compact neighborhood $\Phi$ of $\mathcal{D}$ in $V$, there exists a neighborhood $\mathcal{O}$ of $\pi(C)$ in $L/\Lambda$ such that for any $t\in\R$ and subinterval $J\subset I$, one of the following statements holds:
\begin{enumerate}[(I)]
\item $\lvert \{ s\in J \colon \delta(s)a(t)\phi(s)x_0 \in \mathcal{O} \} \rvert \leq \epsilon \lvert J \rvert$.
\item There exists $\gamma\in\Lambda$ such that $\delta(s)a(t)\phi(s)l\gamma p_H \in \Phi$ for all $s\in J$.
\end{enumerate}
\end{proposition}\par
\subsection{Algebraic consequences of positive limit measure on singular sets}
Recall the definition of $\lambda_i$ in \eqref{eq:def_lambda_i}, where we take $\mathfrak{X} = L/\Lambda$ and $x_i=x_0$. After passing to a subsequence, $\lambda_i\to\lambda$ in the space of probability measures on $L/\Lambda$, and by \Cref{thm:non_divergence} and \Cref{thm:unipotent_invariance}, we know that there exists $H\in\mathcal{H}$ such that
\begin{equation}
\lambda(\pi(N(H,W)\backslash S(H,W))>0.
\end{equation}\par
In this section, we use \Cref{prop:nonfocusing_creterion} and \Cref{thm:linear_stability} to obtain the following algebraic consequence, which is an analogue of \cite[Proposition 4.8]{Sha09Duke1}.
\begin{proposition}\label{prop:algebraic_consequence}
Let $l\in L$ such that $x_0=l\Lambda$. Suppose $\lambda_i\to\lambda$, then there exists $\gamma\in\Lambda$ such that
\begin{equation}\label{eq:algebraic_consequence}
\phi(s)l\gamma p_H \in V^{0-}(a),\quad \forall s\in I.
\end{equation}
\end{proposition}
\begin{proof}
By \eqref{eq:positive_limit_measure_on_singular_set} there exists a compact subset $C\subset N(H,W)\backslash S(H,W)$ and a constant $c_0 > 0$ such that $\lambda(\pi(C))>c_0$. We fix $0<\epsilon < c_0$, and apply \Cref{prop:nonfocusing_creterion} to obtain $\mathcal{D}$. We choose any relatively compact neighborhood $\Phi$ of $\mathcal{D}$, and obtain an $\mathcal{O}$ such that either (I) or (II) holds.\par
Since $\lambda_i\to\lambda$, there exists $i_0\in\N$ such that for all $i>i_0$, (I) does not hold. Therefore (II) holds for all $i>i_0$. In other words, there exists a sequence $\{\gamma_i \}$ in $\Lambda$ and a subinterval $J\subset I$ such that
\begin{equation}
\delta(s)a(t_i)\phi(s)l\gamma_ip_H\in \Phi,\quad \forall i>i_0,\,\forall s\in J.
\end{equation}\par
By \Cref{thm:linear_stability}, we know that $\{\gamma_ip_H \}$ is bounded. Hence after passing to a subsequence, we may assume that there exists $\gamma\in\Lambda$ such that $\gamma_ip_H = \gamma p_H$ holds for all $i$. It follows that $a(t_i)\phi(s)l\gamma p_H$ remains bounded in $V$. This concludes the proof.
\end{proof}
Next we are able to obtain more algebraic information from \Cref{prop:algebraic_consequence}. First we show that the limiting process actually happens inside the $G$-orbit $G\cdot l\gamma p_H$.
\begin{proposition}\label{prop:limit_on_G_orbit}
Let the notations be as in \Cref{prop:algebraic_consequence}. Then for all but finitely many $s\in I=[a,b]$, there exists $\xi(s)\in P$ such that
\begin{equation}\label{eq:limit_on_G_orbit}
\lim_{t\to\infty}a(t)\phi(s)l\gamma p_H = \xi(s)\phi(s)l\gamma p_H.
\end{equation}
\end{proposition}
\begin{proof}
Denote $v=l\gamma p_H$. According to \Cref{prop:algebraic_consequence}, the limit on the left-hand side of \eqref{eq:limit_on_G_orbit} exists. We claim that the limit actually lies in the $G$-orbit $Gv$ for all but finitely many $s\in I$. \par
Consider the boundary $S=\partial(Gv)=\overline{Gv}\backslash Gv$. If $S$ is empty then the claim holds automatically. Now suppose that $S$ is non-empty, and that there exist infinitely many $s\in I$ such that $\lim_{t\to\infty}a(t)\phi(s)v$ is contained in $S$. Since $\phi$ is analytic, we have that for any $s\in I$, $\lim_{t\to\infty}a(t)\phi(s)v$ is contained in $S$. Hence in view of \eqref{eq:cor_basic_lemma1},
\begin{equation}\label{eq:temp8}
\phi(s)\in G(v,S,a),\quad\forall s\in J.
\end{equation}
Moreover, by \Cref{cor:basic_lemma1} there exists $\delta_0\in\Gamma^{+}(T)$ and $g_0\in G$ such that
\begin{equation}\label{eq:temp9}
G(v, S,a) \subset \bigsqcup_{w\in W^+(\delta_0, a)}Pw^{-1}Bg_0^{-1}.
\end{equation}
By \eqref{eq:temp8},\eqref{eq:temp9} and \Cref{lem:union_of_Schubert}(b), the image of $\widetilde{\phi}$ is contained in an unstable Schubert variety with respect to $a(t)$, which contradicts our assumption. \par
Hence for all but finitely many $s\in I$, there exists $\eta(s)\in G$ such that
\begin{equation}\label{eq:temp15}
\lim_{t\to\infty}a(t)\phi(s)v = \eta(s)\phi(s)v.
\end{equation}
Now fix any $s$ such that \eqref{eq:temp15} holds. Take $t_0>0$, and set $w=a(t_0)\phi(s)v$. Then
\begin{equation}\label{eq:temp18}
\lim_{t\to\infty}a(t)w = \eta(s)a(t_0)^{-1}w.
\end{equation}
By taking $t_0$ large enough, we may assume that $\eta(s)a(t_0)^{-1}$ is contained in a small neighborhood of the neutral element in $G$. Let $F$ denote the stabilizer of $\eta(s)a(t_0)^{-1}w=\eta(s)\phi(s)v$ in $G$, and let $\mathfrak{f}$ be the Lie algebra of $F$. It is easy to see that $F$ contains $\{a(t)\}$.\par
Now the Lie algebra $\mathfrak{f}$ of $F$ is $\mathrm{Ad}\,a(t)$-invariant, and thus we have the following decomposition as a consequence of $a(t)$ being semisimple:
\begin{equation}\label{eq:temp16}
\mathfrak{g} = \mathfrak{f}^{\perp} \oplus \mathfrak{f},
\end{equation}
where $\mathfrak{f}^{\perp}$ is an $\mathrm{Ad}\,a(t)$-invariant subspace of $\mathfrak{g}$. \par
On the other hand, according to the eigenvalues of $\mathrm{Ad}\,a(t)$, we can decompose $\mathfrak{g}$ into
\begin{equation}\label{eq:temp17}
\mathfrak{g} = \mathfrak{g}^-\oplus \mathfrak{g}^0 \oplus \mathfrak{g}^+.
\end{equation} \par
Combining the above two decompositions \eqref{eq:temp16}\eqref{eq:temp17}, we get
\begin{equation}
\mathfrak{g} = \mathfrak{g}^-\oplus \mathfrak{g}^0 \oplus (\mathfrak{g}^+\cap\mathfrak{f}^{\perp})\oplus (\mathfrak{g}^+\cap\mathfrak{f}).
\end{equation}
Hence there exist $X_s^{0-}\in \mathfrak{g}^0\oplus \mathfrak{g}^-$ and $X_s^+\in \mathfrak{g}^+\cap\mathfrak{f}^{\perp}$ such that
\begin{equation}
a(t_0)\eta(s)^{-1} \in \exp X_s^{0-}\exp X_s^{+}F.
\end{equation}
By \eqref{eq:temp18}, we have that $X_s^{+}=0$. Hence
\begin{equation}
a(t_0)\eta(s)^{-1}\in \exp X_s^{0-}F.
\end{equation}
Set $\xi(s)=\exp(-X_s^{0-})a(t_0)$, and one can verify that \eqref{eq:limit_on_G_orbit} holds.
\end{proof}
If we consider the slightly larger family of \emph{weakly} unstable Schubert varieties, and further assume that the image of $\widetilde{\phi}$ is not contained in any weakly unstable Schubert variety, then we could obtain the following refinement of \Cref{prop:limit_on_G_orbit}.
\begin{proposition}\label{prop:closed_orbit_reductive_stablizer}
In the situation of \Cref{prop:algebraic_consequence}, further assume that the image of $\widetilde{\phi}$ is not contained in any weakly unstable Schubert variety of $G/P$ with respect to $a(t)$. Then the orbit $G\cdot l\gamma p_H$ is closed, and the stabilizer of $l\gamma p_H$ in $G$ is reductive.
\end{proposition}
\begin{proof}
Write $v=l\gamma p_H$. Suppose that $Gv$ is not closed, then the boundary $S=\partial(Gv)$ is non-empty. By \Cref{prop:basic_lemma2} there exists $\delta_0\in\Gamma^{+}(T)$ and $g_0\in G$ such that
\begin{equation}\label{eq:temp1}
G(v, V^{0-}(a)) \subset \bigsqcup_{w\in W^{0+}(\delta_0, a)}Pw^{-1}Bg_0^{-1}.
\end{equation}
Also by \eqref{eq:algebraic_consequence} we know
\begin{equation}\label{eq:temp2}
\phi(s) \in G(v, V^{0-}(a)),\quad \forall s\in I.
\end{equation}
By \eqref{eq:temp1}, \eqref{eq:temp2} and \Cref{lem:union_of_Schubert}(c), the image of $\widetilde{\phi}$ is contained in a weakly unstable Schubert variety, which contradicts our assumption on $\phi$. \par
Therefore $Gv$ is closed, i.e. $G\cdot l\gamma p_H$ is closed. By Matsushima's criterion, the stabilizer of $l\gamma p_H$ in $G$ is reductive.
\end{proof}
The following proposition describes the obstructions to equidistribution. (C.f. \cite[Theorem 6.1]{SY16}.)
\begin{proposition}\label{prop:algebraic_obstruction}
Suppose that the image of $\widetilde{\phi}$ is not contained in any unstable Schubert variety of $G/P$ with respect of $a(t)$, and that $\lambda_i\to\lambda$. Then there exists $g\in G$ and an algebraic subgroup $F$ of $L$ containing $\{a(t)\}$ such that $Fgl\Lambda$ is closed and admits a finite $F$-invariant measure, and that
\begin{equation}\label{eq:stuck_in_F}
\phi(s) \in P(F\cap G)g,\quad\forall s\in I.
\end{equation} \par
Furthermore, if the image of $\widetilde{\phi}$ is not contained in any weakly unstable Schubert variety, then we can choose $F$ such that $F\cap G$ is reductive.
\end{proposition}
\begin{proof}
Let $\xi(s)$ be defined as in \Cref{prop:limit_on_G_orbit}. Since the right hand side of \eqref{eq:stuck_in_F} is left $a(t)$-invariant, without loss of generality we may replace $\phi(s)$ with $a(t_0)\phi(s)$ for some large $t_0>0$, and assume that $\xi(s)$ lies in a small neighborhood of $e$ in $G$, for all $s\in I$. Hence we may take $\xi(s)\in P$. \par
Fix any $s_0\in I$. Let $g = \xi(s_0)\phi(s_0)$ and $v=l\gamma p_H$. We set $F= \mathrm{Stab}_L(gv)=gl\gamma N_L^1(H)\gamma^{-1}l^{-1}g^{-1}$. By \Cref{prop:limit_on_G_orbit} we have $\{a(t)\}\subset F$. Since $\Lambda\cdot p_H$ discrete, $N_L^1(H)\cdot\Lambda$ is closed. Hence $Fgl\Lambda$ is also closed. \par
Now the Lie algebra $\mathfrak{f}$ of $F$ is $\mathrm{Ad}\,a(t)$-invariant, and thus we have the following decomposition as a consequence of $a(t)$ being semisimple:
\begin{equation}\label{eq:temp6}
\mathfrak{g} = \mathfrak{f}^{\perp} \oplus \mathfrak{f},
\end{equation}
where $\mathfrak{f}^{\perp}$ is an $\mathrm{Ad}\,a(t)$-invariant subspace of $\mathfrak{g}$. \par
On the other hand, according to the eigenvalues of $\mathrm{Ad}\,a(t)$, we can decompose $\mathfrak{g}$ into
\begin{equation}\label{eq:temp7}
\mathfrak{g} = \mathfrak{g}^-\oplus \mathfrak{g}^0 \oplus \mathfrak{g}^+.
\end{equation} \par
Combining the above two decompositions \eqref{eq:temp6}\eqref{eq:temp7}, we get
\begin{equation}
\mathfrak{g} = \mathfrak{g}^-\oplus \mathfrak{g}^0 \oplus (\mathfrak{g}^+\cap\mathfrak{f}^{\perp})\oplus (\mathfrak{g}^+\cap\mathfrak{f}).
\end{equation}
Hence for all $s$ near $s_0$, there exist $X_s^{0-}\in \mathfrak{g}^0\oplus \mathfrak{g}^-$ and $X_s^+\in \mathfrak{g}^+\cap\mathfrak{f}^{\perp}$ such that
\begin{equation}\label{eq:temp3}
\xi(s_0)\phi(s)g^{-1} \in \exp X_s^{0-}\exp X_s^+F.
\end{equation} \par
Since $a(t_i)\phi(s)v$ converge in $V$ as $i\to\infty$, by \Cref{prop:limit_on_G_orbit} we know that $a(t_i)\phi(s)g^{-1}F$ converge in $G/F$ as $i\to\infty$. It follows that
\begin{equation}\label{eq:temp4}
X_s^+ = 0,\quad \forall s\in I.
\end{equation} \par
Since $X_s^{0-}\in \mathfrak{g}^0\oplus \mathfrak{g}^-$, we have
\begin{equation}\label{eq:temp5}
\exp X_s^{0-}\in P.
\end{equation} \par
Combining \eqref{eq:temp3}\eqref{eq:temp4}\eqref{eq:temp5} we get
\begin{equation}
\phi(s) \in PFg,
\end{equation}
for all $s\in I$. This implies \eqref{eq:stuck_in_F}. Moreover, by \cite[Theorem 2.3]{Sha91}, there exists a subgroup $F_1$ of $F$ containing all $\mathrm{Ad}$-unipotent one-parameter subgroups of $L$ contained in $F$ such that $F_1gl\Lambda$ admits a finite $F_1$-invariant measure. Since $F$ contains $\{a(t)\}$, $P$ contains the central torus of $F$. Hence $PFg=PF_1g$, and we may replace $F$ by $F_1$. \par
If we further assume that the image of $\widetilde{\phi}$ is not contained in any weakly unstable Schubert variety, then by \Cref{prop:closed_orbit_reductive_stablizer} we know that the stabilizer of $l\gamma p_H$ in $G$ is reductive, i.e. $g^{-1}Fg\cap G$ is reductive. Hence $F\cap G$ is also reductive.
\end{proof}
\subsection{Lifting of obstructions and proof of equidistribution results}
In this section, we show that the conditions in \Cref{thm:equidistribution1} are preserved under projections. This enables us to use induction to prove the equidistribution results.
\begin{lemma}\label{lem:functoriality}
Let $G$ be a connected semisimple real algebraic group, and $p\colon G\rightarrow\overline{G}$ be a surjective homomorphism. Let $a(t)$ be a multiplicative one-parameter subgroup of $G$, and $\overline{a(t)}$ be its image in $\overline{G}$. Suppose that $\overline{a(t)}$ is non-trivial. Define (weakly) unstable Schubert varieties and partial flag subvarieties of $\overline{G}/\overline{P}$ with respect to $\overline{a(t)}$, $\overline{T}$ and $\overline{B}$. Then the preimage of any unstable (resp. weakly unstable) Schubert subvariety of $\overline{G}/\overline{P}$ with respect to $\overline{a(t)}$ is an unstable (resp. weakly unstable) Schubert subvariety of $G/P$ with respect to $a(t)$.
\end{lemma}
\begin{proof}
Let $X_{\overline{w}}$ be an unstable Schubert subvariety of $\overline{G}/\overline{P}$, where $\overline{w}\in W^{\overline{P}}$ such that $(\overline{\delta},\overline{a}^{\overline{w}})\geq 0$ for some $\overline{\delta}\in\Gamma^{+}(\overline{T})$. Let $G_1$ denote the kernel of $p$, and we have $W_G=W_{G_1}\times W_{\overline{G}}$. Let $w_0$ denote the unique maximal element in $W^{P_1}$. Then the preimage of $X_{\overline{w}}$ is $X_{(w_0,\overline{w})}$. Now it remains to check instability. We note that the Killing form on $\mathfrak{g}$ is the sum of the Killing forms on $\mathfrak{g}_1$ and $\overline{\mathfrak{g}}$. Hence we consider the lifted multiplicative one-parameter subgroup $(e,\overline{\delta})\in\Gamma^{+}(T)$, and use it to check that $X_{(w_0,\overline{w})}$ is unstable. \par
The same proof also works for weakly unstable Schubert varieties.
\end{proof}
We now proceed to the equidistribution results. Recall that $l\in L$ such that $x_0=l\Lambda$, and $\lambda_i$ are probability measures on $L/\Lambda$ as defined in \eqref{eq:def_lambda_i}.
\begin{proposition}\label{prop:twisted_limit_measure_Haar}
Let $\phi$ be an analytic curve on $G$ such that the following two conditions hold:
\begin{enumerate}[(a)]
\item the image of $\widetilde{\phi}$ is not contained in any unstable Schubert variety of $G/P$ with respect to $a(t)$;
\item For any $g\in G$ and any proper algebraic subgroup $F$ of $L$ containing $\{a(t)\}$ such that $Fgx_0$ is closed and admits a finite $F$-invariant measure, the image of $\phi$ is not contained in $P(F\cap G)g$.
\end{enumerate}
Suppose that $\lambda_i\to\lambda$ in the weak-* topology, then $\lambda$ is the unique $L$-invariant probability measure on $L/\Lambda$.
\end{proposition}
\begin{proof}
By \Cref{prop:algebraic_obstruction}, there exists an algebraic subgroup $F$ of $L$ such that \eqref{eq:stuck_in_F} holds. Then condition (b) implies that $F=G$, and thus $G$ fixes $l\gamma p_H$. Arguing as in the proof of \cite[Theorem 5.6]{Sha09Invention}, we know that $L=N_L^1(H)$, i.e. $H$ is normal in $L$. \par
Now we can prove the theorem by induction on the number of simple factors in $L$. If $L$ is simple, then we have $H=L$, and $\lambda$ is $H=L$-invariant. For the inductive step, we consider the natural quotient map $p\colon L\rightarrow L/H$. For any subset $E\subset L$, let $\overline{E}$ denote its image under the quotient map. By \Cref{lem:functoriality}, $\overline{\phi(I)}$ is not contained in any unstable Schubert variety with respect to $\overline{a(t)}$. Hence $\overline{\phi}$ still satisfies condition (a). One can also verify that $\overline{\phi}$ still satisfies condition (b). Indeed, if the image of $\overline{\phi}$ is contained in $\overline{P}(F_0\cap\overline{G})\overline{g}$ for some $F_0\subsetneq\overline{L}$ such that $F_0\overline{g}\overline{x_0}$ is closed, then the image of $\phi$ is contained in $P(p^{-1}(F_0)\cap G)g$ and $p^{-1}(F_0)gx_0$ is also closed. \par
Now both conditions still hold for the projected curve $\overline{\phi}$. By inductive hypothesis we know that the projected measure $\overline{\lambda}$ is the $L/H$-invariant measure on $L/H\Lambda$. In addition, we already know that $\lambda$ is $H$-invariant. Therefore $\lambda$ is $L$-invariant.
\end{proof}
\begin{corollary}\label{cor:limit_measure_Haar}
Let $\phi$ be an analytic curve satisfying (a) and (b) in \Cref{prop:twisted_limit_measure_Haar}. Let $\mu_i$ be the probability measure on $L/\Lambda$ as defined in \eqref{eq:mu_i}. Suppose that $\mu_i\to\mu$ with respect to the weak-* topology, then $\mu$ is the unique $L$-invariant probability measure on $L/\Lambda$.
\end{corollary}
\begin{proof}
The deduction of \Cref{cor:limit_measure_Haar} from \Cref{prop:twisted_limit_measure_Haar} is analogous to the proof of \cite[Corollary 5.7]{Sha09Invention}.
\end{proof}
Parallel to \Cref{prop:twisted_limit_measure_Haar} and \Cref{cor:limit_measure_Haar}, the following results could be proved with the same arguments.
\begin{proposition}\label{prop:twisted_limit_measure_Haar1}
Let $\phi$ be an analytic curve on $G$ such that the following two conditions hold:
\begin{enumerate}[(A)]
\item the image of $\widetilde{\phi}$ is not contained in any weakly unstable Schubert variety of $G/P$ with respect to $a(t)$;
\item For any $g\in G$ and any proper algebraic subgroup $F$ of $L$ containing $\{a(t)\}$ such that $Fgx_0$ is closed and admits a finite $F$-invariant measure and that $F\cap G$ is reductive, the image of $\phi$ is not contained in $P(F\cap G)g$.
\end{enumerate}
Suppose that $\lambda_i\to\lambda$ in the weak-* topology, then $\lambda$ is the unique $L$-invariant probability measure on $L/\Lambda$.
\end{proposition}
\begin{corollary}\label{cor:limit_measure_Haar1}
Let $\phi$ be an analytic curve satisfying (A) and (B) in \Cref{prop:twisted_limit_measure_Haar1}. Let $\mu_i$ be the probability measure on $L/\Lambda$ as defined in \eqref{eq:mu_i}. Suppose that $\mu_i\to\mu$ with respect to the weak-* topology, then $\mu$ is the unique $L$-invariant probability measure on $L/\Lambda$.
\end{corollary}
Now we are ready to prove the main theorems in \Cref{subsect:equidistribution}.
\begin{proof}[Proof of \Cref{thm:equidistribution0}]
If \eqref{eq:equidistribution0} fails to hold, then there exist $\epsilon>0$ and a sequence $t_i\to\infty$ such that for each $i$,
\begin{equation}
\left\vert \frac{1}{b-a}\int_{a}^{b}f(a(t_i)\phi(s)x_0)\,\mathrm{d}s - \int_{L/\Lambda}f\,\mathrm{d}\mu_{L/\Lambda}
\right\vert \geq \epsilon.
\end{equation}
In view of \eqref{eq:mu_i} and \Cref{cor:weak_limit_exists}, this statement contradicts \Cref{cor:limit_measure_Haar}.
\end{proof}
\begin{proof}[Proof of \Cref{thm:equidistribution1}]
If \eqref{eq:equidistribution1} fails to hold, then there exist $\epsilon>0$ and a sequence $t_i\to\infty$ such that for each $i$,
\begin{equation}
\left\vert \frac{1}{b-a}\int_{a}^{b}f(a(t_i)\phi(s)x_0)\,\mathrm{d}s - \int_{L/\Lambda}f\,\mathrm{d}\mu_{L/\Lambda}
\right\vert \geq \epsilon.
\end{equation}
In view of \eqref{eq:mu_i} and \Cref{cor:weak_limit_exists}, this statement contradicts \Cref{cor:limit_measure_Haar1}.
\end{proof}
\section{Grassmannians and Schubert varieties}\label{sect:Grassmannian_Schubert}
In this section we consider the special case where $G = L = \SL_{m+n}(\R)$, and $\Lambda = \SL_{m+n}(\Z)$. Define
\[
a(t) =
\begin{bmatrix}
t^nI_m & \\
& t^{-m}I_n
\end{bmatrix}.
\]
Then $\{a(t)\}$ is a multiplicative one-parameter subgroup of $G$. In this section, all the unstable and weakly unstable Schubert varieties are with respect to this $a(t)$. Let $P$ be the parabolic subgroup associated with $\{a(t)\}$. We have
\begin{equation}
P = \left\{
\begin{bmatrix}
A & \mathbf{0}\\
C & D
\end{bmatrix}\in\SL_{m+n}(\R) \colon
A\in M_{m\times m}(\R),\,C\in M_{n\times m}(\R),\,D\in M_{n\times n}(\R)
\right\}.
\end{equation}
Hence the partial flag variety $G/P$ coincide with $\mathrm{Gr}(m,m+n)$, the Grassmannian of $m$-dimensional subspaces of $\R^{m+n}$. It is an irreducible projective variety of dimension $mn$.\par
\subsection{Schubert cells and Schubert varieties}\label{subsect:Schubert}
Let $B$ be the Borel subgroup of lower triangular matrices in $G$, and $T$ the group of diagonal matrices in $G$. The Weyl group $W=N_G(T)/Z_G(T)$ is isomorphic to $S_{m+n}$, the permutation group on $m+n$ elements. The Weyl group $W_P$ of $P$ is isomorphic to $S_{m}\times S_{n}$, and the set $W^P$ of minimal length coset representatives of $W/W_P$ consists of the permutations $w=(w_1,\cdots,w_{m+n})$ such that $w_1<\cdots<w_m$ and $w_{m+1}<\cdots<w_{m+n}$. We identify $w$ in $W^P$ with the subset $I_w=\{w_1, \cdots, w_m \}$ of $\{1,2,\cdots, m+n \}$. The cosets $wP$ are exactly the $T$-fixed points of $G/P$. The Schubert cell $C_w$ is by definition $BwP$, and the Schubert variety $X_w$ is defined to be $\overline{BwP}$, the closure of $C_w$ in $G/P$. For $w,w'\in W^P$, $w'\in X_w$ if and only if $w'\leq w$ in the Bruhat order. We note that the Bruhat order here is the order on the tuples $(w_1,\cdots ,w_m)$ given by
\[
(w_i)\leq (v_i) \iff w_i\leq v_i,\forall 1\leq i\leq m.
\]
The dimension of $X_w$ is given by $l(w)$, which equals $\sum_{k=1}^{m} (w_k-k)$.\par
The definitions above coincide with the classical definitions. For $1\leq k \leq m+n$, let $F_k$ be the standard $k$-dimensional subspace of $\R^{m+n}$ spanned by $\{e_1,\cdots, e_k \}$. We have the complete flag of subspaces
\begin{equation}
0=F_0\subset F_1\subset F_2\cdots\subset F_{m+n-1}\subset F_{m+n}=\R^{m+n}.
\end{equation}
For an $m$-dimensional subspace $V\in\mathrm{Gr}(m,m+n)$ of $\R^{m+n}$, consider the intersections of the subspace with the flag:
\begin{equation}
0\subset (F_1\cap V)\subset (F_2\cap V)\cdots\subset (F_{m+n-1}\cap V)\subset W.
\end{equation}
For $w\in W^P$, we have a tuple $(w_1,\cdots, w_m)$, and the Schubert cell $C_w$ has the following description:
\begin{equation}
C_w = \left\{V\in\mathrm{Gr}(m,m+n)\colon\dim(V\cap F_{w_{k}})=k;\;\dim(V\cap F_l)<k,\,\forall l<w_k \right\}.
\end{equation}
In other words, the tuple $(w_1,\cdots, w_m)$ gives the indices where the dimension jumps.\par
Similarly, the Schubert variety $X_w$ has the following description:
\begin{equation}
X_w = \left\{V\in\mathrm{Gr}(m,m+n)\colon\dim(V\cap F_{w_{k}})\geq k,\; 1\leq k\leq m \right\}.
\end{equation}\par
Now it is easy to see that
\begin{equation}
X_w = \bigsqcup_{w'\leq w}C_{w'}.
\end{equation}
Hence the Schubert cells give a stratification of the Grassmannian variety.\par
\begin{example}\label{ex:Schubert}
\begin{enumerate}[(1)]
\item For $m = 1$, the Grassmannian $\mathrm{Gr}(1,n)$ is just the projective space $\R\mathbb{P}^n$, and the Schubert varieties form a flag of linear subspaces $X_0\subset X_1\subset\cdots\subset X_n$, where $X_j\cong \R\mathbb{P}^{j}$.
\item For $m=n=2$ one gets the following poset of Schubert varieties in $\mathrm{Gr}(2,4)$:
\begin{equation}
\begin{tikzcd}
&X_{34} \arrow[d, dash] \\
&X_{24} \arrow[dr, dash] \\
X_{14} \arrow[ur, dash] & &X_{23} \arrow[dl, dash] \\
&X_{13} \arrow[ul, dash] \\
&X_{12} \arrow[u, dash]
\end{tikzcd}
\end{equation}
where $X_{12}$ is one single point, and $X_{34}$ is $\mathrm{Gr}(2,4)$.
\end{enumerate}
\end{example}
\subsection{Pencils}
The main goal of this section is to show that maximal (weakly) constraining pencils coincide with maximal (weakly) unstable Schubert varieties in the Grassmannian case, and hence the latter is a natural generalization to all partial flag varieties.\par
Given a real vector space $W \subsetneq \R^{m+n}$, and an integer $r\leq m$, we recall from \Cref{def:pencil} that the pencil $\mathfrak{P}_{W,r}$ is the set
\[
\{V\in\mathrm{Gr}(m,m+n)\colon \mathrm{dim}(V\cap W)\geq r\}.
\]
Denote $d=\dim W$. Let $w\in W^P$ be the element such that $(w_1, \cdots, w_m)$ is the tuple
\[
(d-r+1, \cdots, d,r+1,\cdots, m).
\]
One can verify that the pencil $\mathfrak{P}_{W,r}$ is the Schubert variety $gX_w$, where $g$ is an element in $\SL_{m+n}(\R)$ such that $W = g\cdot F_d$. The pencil is called constraining (resp. weakly constraining) if the inequality \eqref{eq:constraining_pencil} (resp. \eqref{eq:weakly_constraining_pencil}) holds. \par
On the other hand, we recall that the Schubert variety $X_w$ is unstable (resp. weakly unstable) if there exists a non-trivial multiplicative one-parameter subgroup $\delta$ in $\Gamma^+(T)$ such that $(\delta, a^w)>0$ (resp. $\geq 0$). Let $\Delta$ be the element in the Lie algebra $\mathfrak{t}$ of $T$ such that $\delta(t) = \exp(\log t\cdot\Delta)$. Then $\Delta$ could be written as $\mathrm{diag}(t_1,t_2,\cdots, t_{m+n})$, where $t_1\geq t_2 \geq\cdots\geq t_{m+n}$ and $\sum t_i=0$. Hence in the case of Grassmannian we have the following criterion of stability.
\begin{lemma}\label{lem:Schubert_stability_criterion}
Let $w$ be an element in $W^P$, then the corresponding Schubert variety $X_w$ is unstable (resp. weakly unstable) if and only if the following system is soluble:
\begin{gather}
t_1\geq\cdots\geq t_k>0\geq t_{k+1}\geq\cdots\geq t_{m+n} \label{eq:order_t} \\
\sum_{i=1}^{m+n} t_i=0 \label{eq:sum_equals_zero} \\
\sum_{j=1}^{m} t_{w_j} > 0 \;(\text{resp. }\sum_{j=1}^{m} t_{w_j} \geq 0) \label{eq:combinatorial_inequality}
\end{gather}
\end{lemma}
\begin{example}[$m=n=2$]
We continue with \Cref{ex:Schubert}(2). If $w=(14)$, then we can take $t_1 = 3, t_2=t_3=t_4=-1$, which gives $t_1+t_4>0$. Hence by \Cref{lem:Schubert_stability_criterion} we have $X_{14}$ is unstable. Similarly we can show that $X_{23}$ is unstable by taking $t_1=t_2=t_3=1, t_4=-3$.\par
When $w=(24)$, $t_2+t_4 \geq 0$ is soluble as we can take $t_1=t_2=1,t_3=t_4=-1$. However, $t_2+t_4>0$ is insoluble. Indeed, suppose $t_2+t_4>0$, then $t_1+t_3\geq t_2+t_4>0$, and it follows that $t_1+t_2+t_3+t_4>0$, which contradicts \eqref{eq:sum_equals_zero}. Therefore we conclude that $X_{24}$ is weakly unstable but not unstable.
\end{example}
Now we are ready for the main results of this section.
\begin{proposition}\label{prop:pencil_in_Schubert}
Every constraining (resp. weakly constraining) pencil is an unstable (resp. weakly unstable) Schubert variety of $\mathrm{Gr}(m,m+n)$.
\end{proposition}
\begin{proof}
Let $\mathfrak{P}_{W,r}$ be a constraining pencil, and thus by definition we have
\begin{equation}\label{eq:constraining_pencil_recall}
\frac{d}{r} < \frac{m+n}{m},
\end{equation}
where $d = \dim W$. Then $\mathfrak{P}_{W,r}=gX_w$, where $g\in G$ and $w\in W^P$ such that
\begin{equation}
(w_1,\cdots, w_m) = (d-r+1,\cdots,d,n+r+1,\cdots,m+n).
\end{equation}
Now set $t_1 = \cdots = t_d = m + n - d$ and $t_{d+1} = \cdots = t_{m+n} = -d$. It is clear that \eqref{eq:order_t} and \eqref{eq:sum_equals_zero} are satisfied. Moreover,
\begin{equation}
\begin{split}
\sum_{j=1}^{m} t_{w_j} &= r(m+n-d) - (m-r)d\\
&= r(m+n) - md\\
&= mr\left(\frac{m+n}{m}-\frac{d}{r}\right)\\
&> 0.
\end{split}
\end{equation}
Hence \eqref{eq:combinatorial_inequality} also holds. Therefore, by \Cref{lem:Schubert_stability_criterion} we conclude that $\mathfrak{P}_{W,r}$ is an unstable Schubert variety. The same proof also works for weakly constraining pencils.
\end{proof}
\begin{proposition}\label{prop:Schubert_in_pencil}
Every unstable (resp. weakly unstable) Schubert variety of $\mathrm{Gr}(m,m+n)$ is contained in a constraining (resp. weakly constraining) pencil.
\end{proposition}
\begin{proof}
Let $X_w$ be an unstable Schubert variety and consider the set $I_w=\{w_1,\cdots, w_m\}$. Let $J_w$ be the subset of $I_w$ consisting of the elements with jump, that is, $w_k$ is contained in $J_w$ if and only if $w_{k+1}-w_{k}>1$. Here we set $w_{m+1}=0$. Notice that for any $w_k\in J_w$, if we set $W=F_{w_k}$ and $r=k$, then $X_w$ is contained in the pencil $\mathfrak{P}_{W,r}$. Now it suffices to show that there exists $w_k\in J_w$ such that
\begin{equation}\label{eq:slope_inequality}
\frac{w_k}{k} < \frac{m+n}{m}.
\end{equation}
Actually, the function $k\mapsto{w_k}/{k}$ achieves its minimum at some $k$ such that $w_k\in J_w$. Hence it suffices to prove the following claim.
\begin{claim}
There exists $1\leq k \leq m$ such that \eqref{eq:slope_inequality} holds.
\end{claim}
\par
We prove the claim by contradiction. Suppose that for any $1\leq k \leq m$ we have
\begin{equation}\label{eq:contradiction_slope_inequality}
\frac{w_k}{k} \geq \frac{m+n}{m}.
\end{equation}\par
For $1\leq i\leq m+n$, consider the auxiliary function
\begin{equation}
g(i) =
\begin{cases}
-m & i\notin I_w;\\
n & i\in I_w.
\end{cases}
\end{equation}
For any $1\leq i< m+n$, let $w_k$ be the largest element in $I_w$ such that $w_k\leq i$ (and set $w_k=0$ if $i<w_1$). As a consequence of \eqref{eq:contradiction_slope_inequality}, we have
\begin{equation}
\begin{split}
\sum_{j=1}^{i}g(i) &\leq \sum_{j=1}^{w_k}g(i)\\
&= -m(w_k-k) + nk\\
&=(m+n)k-mw_k\\
&\leq 0.\quad\text{By }\eqref{eq:contradiction_slope_inequality}
\end{split}
\end{equation}
It is also clear that
\begin{equation}
\sum_{j=1}^{m+n}g(i) = 0.
\end{equation}\par
Since $X_w$ is unstable, we may find $t_1,\cdots,t_{m+n}$ satisfying \eqref{eq:order_t}\eqref{eq:sum_equals_zero}\eqref{eq:combinatorial_inequality}. Denote
\begin{gather}
A = \sum_{i\in I_w}t_i;\\
B = \sum_{i\notin I_w}t_i.
\end{gather}
Then $A>0$ and $A+B=0$ by \eqref{eq:sum_equals_zero}\eqref{eq:combinatorial_inequality}. Hence $B<0$, and $nA-mB>0$.\par
On the other hand, summation by parts leads to
\begin{equation}
\begin{split}
nA-mB &= n\sum_{i\in I_w}t_i - m\sum_{i\notin I_w}t_i\\
&= \sum_{i=1}^{m+n}g(i)t_i\\
&= \sum_{i=1}^{m+n-1}\left[(t_i-t_{i+1})\sum_{j=1}^{i}g(j)\right]+t_{m+n}\sum_{j=1}^{m+n}g(j)\\
&= \sum_{i=1}^{m+n-1}\left[(t_i-t_{i+1})\sum_{j=1}^{i}g(j)\right]\\
&\leq 0.
\end{split}
\end{equation}
This is a contradiction.\par
Therefore we have proved the claim, and thus $\mathfrak{P}_{W,r}$ is a constraining pencil containing the Schubert variety $X_w$. The same proof works for weakly unstable Schubert varieties.
\end{proof}
Combining \Cref{prop:pencil_in_Schubert} and \Cref{prop:Schubert_in_pencil}, we conclude the following.
\begin{theorem}\label{thm:pencil_equal_Schubert}
Let $E$ be any subset of $\mathrm{Gr}(m,m+n)\cong G/P$. Then $E$ is contained in an unstable (resp. weakly unstable) Schubert variety with respect to $a(t)$ if and only if $E$ is contained in a constraining (resp. weakly constraining) pencil.
\end{theorem}
\subsection{Young diagrams}
In this section, we will give a combinatorial description of pencils and (weakly) constraining pencils, using Young diagrams. This will enable us to quickly see whether a Schubert variety is a pencil, and whether a pencil is (weakly) constraining. The readers are referred to Fulton's book \cite{Ful97} for more details.\par
A \emph{partition} is a sequence of integers $\lambda=(\lambda_1,\cdots,\lambda_m)$ such that $\lambda_1\geq\cdots\geq\lambda_m\geq 0$. Let $\Pi_{m,n}$ denote the set of partitions such that $\lambda_1 \leq n$. A \emph{Young diagram} is a set of boxes arranged in a left justified array, such that the row lengths weakly decrease from top to bottom. To any partition $\lambda$ we associate the Young diagram $D_{\lambda}$ whose $i$-th row contains $\lambda_i$ boxes. An \emph{outside corner} of the Young diagram $D_{\lambda}$ is a box in $D_{\lambda}$ such that removing the box we still get a Young diagram.
\begin{example}\label{ex:Young_diagram}
Let $m = 3$, $n = 5$, and $\lambda=(4,3,1)\in\Pi_{m,n}$. The Young diagram $D_{\lambda}$ fits inside an $m\times n$ rectangle.
\[
\ytableausetup{nosmalltableaux}
\ytableausetup{centertableaux}
\ytableaushort
{\none\none\none\bullet, \none\none\bullet,\bullet}
*{5,5,5}
*[*(yellow)]{4,3,1}
\]
There are three outside corners, which are marked with a dot in the diagram.
\end{example}\par
Given $\lambda\in\Pi_{m,n}$, the associated Schubert variety $X_{\lambda}\subset\mathrm{Gr}(m,m+n)$ is defined by the conditions
\begin{equation}
\dim(V\cap F_{n+i-\lambda_i}) \geq i,\quad 1\leq i\leq m.
\end{equation}
Actually we only need outside corners to define $X_{\lambda}$; the pairs $(i,\lambda_i)$ which are not outside corners are redundant. (See \cite[Exercise 9.4.18]{Ful97}.) Therefore, we have the following lemma.
\begin{lemma}
Given $\lambda\in\Pi_{m,n}$, the Schubert variety $X_\lambda$ is a pencil if and only if the Young diagram $D_{\lambda}$ has only one outside corner.
\end{lemma}
The Schubert variety given by \Cref{ex:Young_diagram} is not a pencil, as the Young diagram has three outside corners. However, every Schubert variety can be written as an intersection of pencils.\\
One can also recognize constraining and weakly constraining pencils with the help of Young diagrams.\par
For an $m\times n$ rectangle, we draw the diagonal connecting the northeast and the southwest of the rectangle. A \emph{node} is a vertex of a box. We call a node \emph{unstable} if it is lying below the diagonal, and \emph{weakly unstable} if it is lying on or below the diagonal. See \Cref{fig:nodes} for an example.\par
Now we can reformulate the definition of constraining and weakly constraining pencils.
\begin{lemma}\label{lem:pencil_criterion}
A pencil $X_{\lambda}$ is constraining (resp. weakly constraining) if and only if the bottom-right vertex of the outside corner of $D_{\lambda}$ is an unstable (resp. weakly unstable) node.
\end{lemma}
\begin{figure}
\centering
\[
\begin{tikzpicture}
\draw [help lines] (0,0) grid (3,3);
\draw [dashed] (0,0) -- (3,3);
\node[circle,draw=black, fill=white, inner sep=0pt,minimum size=5pt] (b) at (0,0) {};
\node[circle,draw=black, fill=white, inner sep=0pt,minimum size=5pt] (b) at (1,1) {};
\node[circle,draw=black, fill=white, inner sep=0pt,minimum size=5pt] (b) at (2,2) {};
\node[circle,draw=black, fill=white, inner sep=0pt,minimum size=5pt] (b) at (3,3) {};
\fill [black] (1,0) circle (3pt);
\fill [black] (2,0) circle (3pt);
\fill [black] (3,0) circle (3pt);
\fill [black] (2,1) circle (3pt);
\fill [black] (3,1) circle (3pt);
\fill [black] (3,2) circle (3pt);
\end{tikzpicture}
\]
\caption{Unstable and weakly unstable nodes in a $3\times3$ rectangle. The black nodes are unstable, while the white nodes are weakly unstable but not unstable.} \label{fig:nodes}
\end{figure}
\begin{example}
Let $m=2$ and $n=3$. By \Cref{lem:pencil_criterion} there are $5$ constraining pencils: $X_{12},X_{15},X_{23},X_{25}$ and $X_{34}$. Among those $X_{25}$ and $X_{34}$ are the maximal ones, and they give the obstruction to non-divergence.
\[
\ytableausetup{nosmalltableaux}
\ytableausetup{centertableaux}
\ytableaushort
{\none\bullet, \none}
*{3,3}
*[*(yellow)]{2,0}
\quad\quad
\ytableausetup{nosmalltableaux}
\ytableausetup{centertableaux}
\ytableaushort
{\none, \bullet}
*{3,3}
*[*(yellow)]{1,1}
\]
\end{example}
As noted in \Cref{rem:pencils}, the weakly constraining pencils coincide with the constraining pencils in the case that $m$ and $n$ are coprime. This also follows from the simple observation that there are no nodes lying on the diagonal of $D_{\lambda}$.
\bibliographystyle{alpha}
|
\section{Introduction}
\label{intro}
The problem of optimizing a function over the set of efficient solutions of a multiobjective optimization problem has attracted the attention of researchers in the last decades due to the wide range of applications of this kind of problems. Needless to say that most (if not all) of the real-world problems, in different fields (manufacturing, logistic, finance, ...), have sometimes to deal with several conflicting goals. Several techniques have been designed to solve the different types of problems under this framework (see e.g., \cite{ehrgott,gandibleux,steuer}). However, in general, the solution set of these problems is large, which is not reasonable from the decision maker viewpoint. Hence, optimizing a function over the set of solutions of a multiobjective function becomes a very useful tool to select the ``best'' solution among that set (see \cite{shigeno} for a recent real-world application of this problem). But if multiobjective optimization is difficult itself, optimizing over the unknown set of efficient solution is a challenge in the mathematical programming community.
The continuous linear case have been widely studied (see \cite{philip,benson,ecker1,sayin,yamamoto}, among others). However, there is a lack of literature analyzing the optimization of a linear function over the efficient set of a multiobjective discrete problem, and relatively few algorithms have been proposed to solve them (see\cite{abbas,mebrek,jorge,boland}).
As mentioned above, there are many applications of optimizing a function over the set of efficient solutions of a multiobjective problem. In this paper we analyze one of them arising in a more abstract field: the Algebraic Theory of Numbers. In particular, we study an algebraic invariant that appears when trying to identify the structure of a Numerical Semigroup. A Numerical Semigroup is a subset of the set of nonnegative integer numbers, which represent all the nonnegative integer combinations of certain set of integers. The study of these objects allows to understand the insights of some interesting Diophantine Equations, which are closely related with knapsack polytopes. We study here the $\omega$-invariant of a Numerical Semigroup, an invariant proposed in \cite{geroldinger}, and that has paid a lot of attention in the last years from researchers form Computational Algebra (see\cite{halter}\cite{blanco}\cite{omidali}, among others).
In \cite{bvictor}, the author formulates the problem of computing the $\omega$-invariant as a problem of optimizing a linear function over the set of efficient solutions of a discrete multiobjective problem. Hence, a methodology based on \cite{jorge} was applied to solve the problems. The main advantage of this approach with respect to those proposed before is that since it uses mathematical programming tools it allows one to avoid the complete enumeration of the discrete feasible set.
In this paper, we propose a different strategy for solving the problem in which we exploit the structure of the problem by computing new bounds for the problem that allows to strengthen the search of the solution. This simple idea allows us to speed up the method and solve the problem in smaller times that those consumed by the one in \cite{bvictor}.
The paper is organized in five sections. In Section \ref{notation}, we briefly recall some preliminary notions on Multiobjective Optimization and Numerical Semigroups, needed for the development in the rest of the paper. Section \ref{description} is devoted to describe our approach for computing the $\omega$-invariant. In Section \ref{num-exp} we report the results of some computational experiments. Finally, in Section \ref{conclusion} we give some concluding remarks and perspectives.
\section {Preliminaries}\label{notation}
\subsection{Optimization over the Efficient set}
We are given a finite set of linear objective functions induced by the $p$ vectors $c_1, \ldots, c_p \in \mathbb{R}^n$, and a polyhedral feasible set, $P$, defined by means of linear equations $P=\{x\in \mathbb{R}_+^n: Ax = b\}$, where $A\in\mathbb{R}^{m\times n}$ and $b \in \mathbb{R}^m$. A multiobjective linear integer programming problem is written as:
\begin{align}
\max & \;\; C x := (c_1(x), \ldots, c_p(x))\label{molip}\tag{${\rm MOLIP}$}\\
\mbox{s.t. }& x\in P \cap \mathbb{Z}_+^n. \nonumber
\end{align}
Solving \eqref{molip} means finding the set of the so-called \textit{efficient} solutions. A feasible solution $x^*\in P \cap \mathbb{Z}_+^n$ is said efficient solution of $\eqref{molip}$ if and only if there is no feasible solution $x \in D $ such that $C x \geq C x^{*}$ with $C x \neq C x^{*}$. Otherwise, $x^{*}$ is not efficient and $C x^{*}$ is said to be dominated by $C x$. The set of efficient solutions for the problem above will be denoted by $\mathrm{Eff}(C,P)$.
Despite finding the whole set of efficient solutions is hard since the problem is \#P-hard~\cite{EG02}, one can check whether a given feasible solution $\bar x \in P \cap \mathbb{Z}^n_+$ is an efficient solution or not solving an integer linear programming problem. Ecker and Kouada \cite{ecker} stated that one can solve the following problem:
\begin{align}
\Psi(\bar x) = \max & \;\;\sum\limits_{i=1}^{p}\psi_{i}\nonumber\\
\mbox{s.t. }& c_j x+\psi_j= c_j \tilde{x}, \forall j=1, \ldots, p,\label{EK}\tag{$\rm{EK(}\tilde{x}\rm{)}$}\\
&x \in P \cap \mathbb{Z}_+^n, \nonumber\\
&\psi_{j} \geqslant 0, \forall j=1, \ldots, p.\nonumber
\end{align}
Then, $\bar {x}$ is an efficient solution if and only if $\Psi(\bar x) = 0$. Furthermore, in case $\bar{x}$ is not efficient, the optimal $x$-variables of \eqref{EK} provide an efficient solution dominating $\bar{x}$.
In this paper, we analyze beyond multiobjective optimization, since we analyze the problem of optimizing a linear function over the set of efficient solutions of a multiobjective linear integer programming problem. The problem can be formulated as follows:
\begin{align}
\max & \; \phi(x) := d x\nonumber\\
\mbox{s.t. }&x\in \mathrm{Eff}(C,P). \label{Main}\tag{${\rm OES}$}
\end{align}
where $d =(d_1, \ldots, d_n) \in \mathbb{R}^n$.
A possible relaxation of the problem above consists of relaxing the requirements of efficiency for $x$:
\begin{align}
\max , & \;\phi(x) \\
\mbox{s.t. }& x\in P \cap \mathbb{Z}^n_+. \label{OESR}\tag{${\rm OES}_R$}
\end{align}
\subsection{The $\omega$-invariant}
In what follows we describe the problem analyzed in this paper. It is defined over the algebraic structure of numerical semigroups. Hence, we first describe the basic elements under this theory and the notions needed to state the problem of computing the $\omega$-primality of a numerical semigroup.
$S\subseteq \mathbb{Z}_+$ is said a numerical semigroup if: (1) $0 \in S$; (2) $\mathbb{Z}_+ \backslash S$ is finite; and (3) $S$ is closed for the sum ($x + y \in S$ for all $x, y \in S$). A set $\{n_{1},\ldots, n_{p}\} \subset S$ with $\gcd = (n_{1}, \ldots, n_{p})= 1$ is system of generators of $S$ if $S =\left\{s \in \mathbb{Z}_+: s = x_{1}n_{1} + ...+ x_{p} n_{p}, x_1, \ldots, x_p \in \mathbb{Z}_+ \right\}$. The set $\{n_{1}, \ldots, n_{p}\}$ is a minimal system of generators if no proper subset of it generates $S$. It is not difficult to see that every numerical semigroup has a unique minimal generating system \cite{rosales}. If $\{n_1, \ldots, n_p\}$ is a minimal system of generators of $S$, we denote $S=\langle n_1, \ldots, n_p\rangle$.
One interesting question when analyzing numerical semigroups is the study of decompositions of the elements of the semigroup with respect to the minimal generating system, that is, its set of factorizations. Let $S=\langle n_1, \ldots, n_p\rangle$ be a numerical semigroup, and $s\in S$. The possible factorizations of $s$ are the different ways of obtaining $s$ as an integer linear combination of the minimal system of generators, i.e., the solutions of the following system of diophantine equations:
\begin{align*}
s= \sum_{i=1}^p n_i x_i,
\end{align*}
with $x_1, \ldots, x_p \in \mathbb{Z}_+$.
The solutions of the above system is denoted by $F(s)$, the set of factorizations of $s$ in $S$.
In the analysis of factorization, Gerolding and Hassler \cite{gerol,geroldin} introduced the notion of $\omega$-primality which allows to measure how far is an element of the semigroup of being uniquely factorized. A lot of attention has been paid in the recent years to the computation and analysis of such an algebraic invariant (see~\cite{anderson,ander,gero,you}, among others). Observe this measures is closely related with some integer programming models, since unique factorizations will provide single feasible solutions to knapsack problems.
\begin{definition}[Omega Invariant]
Let $S = \langle n_{1},\cdots, n_{p}\rangle$ be a numerical semigroup, and $s \in S$. The $\omega$-invariant of $s$ in $S$, $\omega(S, s)$, is the smallest nonnegative integer $K$, such that
for each finite set of elements $\{s_1, \ldots, s_n\} \subseteq S$ that verifies $\sum_{i \in S_K} s_i - s \in S$, there exists $\Omega \subseteq \{1,\ldots,n\}$ with cardinality $K$ and such that $\sum_{i \in \Omega} s_i - s \in S$. The global $\omega$-invariant of the semigrouo $S$ is defined as $\omega(S)= \:\max , \{ \omega(S, n_{i})\: : \: i= 1,\cdots, p \}$.
\end{definition}
With such an invariant, $s$ is prime, i.e. uniquely factorized if $\omega(S,s)=1$. Otherwise, if $\omega(S,s)>1$, then $\omega(S)>1$ and the semigroup is said $\omega(S)$-prime. In \cite{blanco}, is is proved that:
$$
\omega(S, s) = \max \left\{ \displaystyle \sum_{i=1}^{p} x_{i} \: : \: x \in {\rm Minimals} \{F(s+s'): s' \in S\} \right\}.
$$
(here, ${\rm Minimals}\{Q\}$ stands for the set of component-wise minimal elements of $Q \subset\mathbb{R}^p$.)
This equation states that one can compute the $\omega$-invariant by searching among the set of minimal elements of certain set.
The interesting point here is the connection between the $\omega$-invariant and mathematical programming. In particular, from an algebraic result in \cite{blanco}, in \cite{bvictor} the author formulates the problem as a problem with the shape of \eqref{Main}, i.e. of optimizing a linear function over the efficient set of a multiobjective linear integer programming problem:
\begin{lemma}[\cite{bvictor}]
Let $S=\langle n_1, \ldots, n_p\rangle$ and $s \in S$. Then,
\begin{align} \omega(S,n_j)= \max & \sum_{i=1}^n x_i\\
\mbox{s.t. }& (x, y) \in \mathrm{Eff}(C,P)
\end{align}
where $C=(Id_{p},0_p)$, and $P=\{x\in \mathbb{R}_+^n: Ax = b\}$ with $A=(n_1,\ldots,n_p,-n_1,\ldots,-n_p)$ and $b=n_j$ (here, $Id_p$ stands for the $p\times p$ identity matrix and $0_p$ is the $p\times p$ matrix with all its entries zero.
\end{lemma}
The aim of this paper is to efficiently compute the $\omega$-invariant of a numerical semigroup.
\section{A solution scheme for the problem}\label{description}
\noindent In this section we propose a new mathematical programming approach for computing the $\omega$-invariant of a numerical semigroup. We show that our new algorithm provides the $\omega(S)$ without an explicit enumeration of all efficient factorizations in a finite number of iterations. In \cite{bvictor} the author proposed the first mathematical optimization model and resolution approach for the problem. However, a major drawback of the resolution presented in \cite{bvictor} is that it does not fully exploit the specific nature of the problem. Here, we propose an improved algorithm that incorporates information about the algebraic structure of the problem.
Let $S= \langle n_{1},...,n_{p} \rangle$ be a numerical semigroup. The initial step consists of generating a subset of $p-1$ efficient solutions for the multiobjective problem, which is based on the following result whose proof is straightforward.
\begin{lemma}
Let $j\in\{1, \ldots,p\}$. Then, for $k \in \{1,\ldots,p\}\backslash\{j\}$, the solution of the following integer programming problem:
\begin{align}
\overline x_k^j = \min & \;\; x_{k}\nonumber\\
\mbox{s.t. } & n_{k} x_{k} -\displaystyle \sum_{i=1}^{p} n_{i}y_{i}= n_{j}\label{p0}\tag{${\rm UBound}_{jk}$}\\
&x_k \in \mathbb{Z}_+, y \in \mathbb{Z}_+^{p}\nonumber
\end{align}
is in $\mathrm{Eff}(C,P)$.
\end{lemma}
The above result allows us to provide an initial lower bound for $\omega(S,n_j)$: $UB_j := \max_{k\neq j}, \left\{ \sum_{i=1}^{p}\overline{x}_{i}^{k} \right\}$, with associated efficient solution that we denote by $\underline{x}^j$. Consequently, we have also an upper bound for $\omega(S)$, $UB=\displaystyle\max_{j=1,\ldots,p} UB_j$.
Next, in order to find another solution not dominated by $ \underline{x}^{j}$ we use the Sylva and Crema's cut \cite{sylva2004,sylva2007}, which consists on solving the following integer linear programming problem:
\begin{align}
\max& \sum_{i=1}^{p} x_{i} \label{psj}\tag{${\rm PS}(\overline{x}_{j})$}\\
\mbox{s.t. } & x_{i} \leq k_{i}(\underline{x}_{i}^{j}-1)- M_{i}(k_{i}-1)&i=1, \ldots, p,\label{c1}\\
&\displaystyle \sum_{i=1}^{p}n_{i}x_{i} -\displaystyle \sum_{i=1}^{p} n_{i} y_{i} = n_{j},\nonumber\\
&x_{j}= 0,\nonumber\\
&\sum_{i=1}^{p} k_{i} \geq 1,\label{c2}\\
&x, y \in \mathbb{Z}_+^{p}, k \in \{0,1\}^{p}.\nonumber
\end{align}
\noindent where $M_{i}$ is an upper bound for the value of the $x_i$ variable in the problem, for $i=1, \ldots, p$, which was obtained when solving \eqref{p0} for $k=i$.
Observe that by \eqref{c1} the binary $k$-variables indicate whether in the new solution the value of $x_i$ strictly increases (being $k_i=1$) or not (being $k_i=0$). Hence, by constraint \eqref{c2}, the obtained solution at least increases the value of one of the values, being then non dominated by $\underline{x}^j$. The solution of \eqref{psj} provides a lower bound for $\omega(S,n_j)$, that will be denoted by $LB_j$.
Once a solution is obtained, one can solve \eqref{EK} to check whether the new solution is efficient, and otherwise it provides an efficient solution dominating it. Then, the process is repeated to generate different efficient solutions for the problem. To assure that an already efficient solution is again generated, at the $\ell$-th iteration we incorporate to \eqref{psj}, the following sets of constraints, instead of \eqref{c1}:
\begin{equation}\label{newc}
x_{i} \leq k_{i}^{s}(\widehat{x}_{i}^{r}-1)- M_{i}(k_{i}^{r}-1), \forall i=1,\ldots, p, r=1,...\ell,
\end{equation}
and require that $k_i^s \in \{0,1\}$, for $i=1,\ldots, p, r=1,...\ell$, where $x^1, \ldots, x^\ell \in \mathbb{Z}_+^p$ are the efficient solutions generated up to this iteration. Running the above problem allows us to update the lower bound if the obtained objective function is greater.
At this point, different situations are possible after solving \eqref{psj} at a certain iteration of the procedure (with the extra constraints \eqref{newc}): (1) A new solution is obtained and $LB_j<UB_j$. In that case, a new iteration is performed; (2) \eqref{psj} with the extra constraints \eqref{newc} is infeasible. In such a case, the complete exploration have been performed, and the best solution found up that iteration is the optimal one; and (3) $LB_j=UB_j$. In that case, since no improvement can be done on the objective function, the solution has been reached and $\omega(S,n_j)$ coincides with the computed bounds.
We show in Algorithm \ref{alg} a pseudo-code summarizing the different steps of our approach.
\begin{algorithm}[htp]
\normalsize
$\star$ \textbf{Input Data}: $S=\langle n_{1},\cdots ,n_{p}\rangle$ and $j \in \{1,\ldots,p\}$, $LB_j=0$, $\ell=1$.
$\star$ \textbf{Computation of Initial Lower Bound.}\\
Solve \eqref{p0} for $k=1, \ldots, p$ ($k\neq j$): $UB_j$ and $\overline{x}^j$.
$\star$ \textbf{Iterations:}
Solve \eqref{psj} (for $\ell>1$ with extra constraints \eqref{newc}).\\
\If{Feasible and $ \sum_{i=1}^{p}\overline{x}^{\ell}_{i} < UB_j $ (Situation A)}
{
$UB_j =\sum_{i=1}^{p}\overline{x}^{\ell}_{i}$,
$x^{\ell+1}= \overline{x}^{\ell}$,\\
$\ell = \ell+1$,\\
$LB_j = \max\{ LB_j, \sum_{i=1}^{p} x_{i}^{\ell}\}$.\\
\textbf{Solve \eqref{EK} and repeat.}
}
\If{\eqref{psj} is unfeasible (Situation B) or $LB_j \geq \sum_{i=1}^{p}\overline{x}^{\ell}_{i}$ (Situation C)}
{\textbf{Stop.}}
$\star$ \textbf{Output}: $UB_j=\omega(S,n_j)$\\
\caption{Finding the omega invariant of numerical semigroup.\label{alg}}
\end{algorithm}
\begin{proposition}
Algorithm \ref{alg} converges in a finite number of iterations.
\end{proposition}
\begin{proof}
Since the cardinality of the admissible region $\overline{D}$ is finite (the decision variables are integers and bounded) and at each iteration, the current domain $\widetilde{D}$ is being gradually reduced with $|\widetilde{D}|< |\overline{D}|$ until it becomes empty $(|\widetilde{D}|=0)$. This indicates that the method process terminates in a finite number of iterations.
Let now analyze the convergence. We suppose the existence of a feasible solution $\overline{x} \in EF(P(n_j))$; $\overline{\omega}(S,n_j)= \sum_{i=1}^{p}\overline{x}_i$ such that $\overline{\omega}(S,n_j)> \omega^{\star}(S,n_j)$. At the terminal stage of the above algorithm, say iteration $k$, we obtain\\
$\omega_{inf} =\omega^{\star}(S,n_j) =\sum_{i=1}^{p}x^{\star}_{i} =\sum_{i=1}^{p}x^{k}_{i} \geq \sum_{i=1}^{p}x^{\ell}_{i}, , \forall \ell \in \{1, \cdots, k-1\} $.\\
Therefore, $\overline{\omega}(S,n_j)> \omega^{\star}(S,n_j) \geq \omega^{\ell}(S,n_j),, \forall \ell \in \{1, \cdots,k-1\}$;
particularly, $\overline{\omega}(S,n_j)> \omega^{\star}(S,n_j) \geq \overline{\omega}(S,n_j)$. Therefore, $\omega^{\star}(S,n_j)=\overline{\omega}(S,n_j)$.
\end{proof}
\begin{example}\label{ex}
Let us consider the numerical semigroup $S=\langle 6,10,14,27\rangle$, and let us compute $\omega(S,27)$. First, solving \eqref{p0} for $k=1,2,3$, we get that $LB_4=9$ with $x^0=(9,0,0,0)$. Then, we solve \eqref{psj} to get a solution not dominated by $x^0$, obtaining $x^{1}= (8,6,5,0)'$ and $y^{1}= (1,1,0,5)$, with an optimal value $UB_4= 19$. Since $UB_4> LB_4$, we solve \eqref{EK} to get an efficient solution which dominates $x^1$. We obtain $\widehat{x}^{1}=(0,4,1,0)$ and $\widehat{y}^{1}=(0,0,0,1)$, with an optimal value of $5<9$, so the lower bound is not updated. The rest of the iterations are summarized in Table \ref{table:ex}, where one can observe that after $8$ iterations one get that $LB_4=UB_4=10$, being then $\omega(S,27)=10$.
\ref{table:ex}.
\begin{table}[h]
\caption{Iterations of the algorithm to compute $\omega(S,27)$ of Example \ref{ex}.}
\label{table:ex}
\begin{center}
\normalsize\begin{tabular}{|ccccc|}\hline
It & \eqref{psj} & \eqref{EK} & $LB_4$ & $UB_4$ \\\hline
1 & $x^{1}=( 8, 6 , 5, 0)$& $\hat{x}^{1}=( 0, 4 , 1, 0)$& 9&19\\
2& $x^{2}=( 8, 3 , 5, 0)$& $\hat{x}^{2}=( 2 , 0 , 3, 0)$& 9&16\\
3& $x^{3}=( 8, 6 , 0, 0)$& $\hat{x}^{3}=( 0 , 6 , 0, 0)$& 9 &14 \\
4& $x^{4}=( 8, 5 , 0, 0)$& $\hat{x}^{4}=( 4 , 3 , 0, 0)$& 9 &13 \\
5&$x^{5}=( 8, 2 , 2, 0)$& $\hat{x}^{5}=( 1 , 2 , 2, 0)$ & 9 &12 \\
6& $x^{6}=( 8, 1 , 2, 0)$& $\hat{x}^{6}=( 5 , 1 , 1, 0)$& 9 &11 \\
7& $x^{7}=( 8, 0 , 2, 0)$& $\hat{x}^{7}=( 6 , 0 , 2, 0)$& 9 &10\\
8& $x^{8}=( 8, 2 , 0, 0)$& $\hat{x}^{8}=( 8 , 2 , 0, 0)$&10&10\\
\hline
\end{tabular}
\end{center}
\end{table}
\end{example}
\section{Numerical experiments}\label{num-exp}
We have run a series of experiments in order to test the efficiency of the proposed method as well as to compare with the approach proposed by Blanco in \cite{bvictor}.
We generate a battery of random instances of different sizes. We use GAP package ``numericalsgps'' \cite{numericalsgps}
to generate numerical semigroups with embedding dimension $p$, with $p$ ranging in $\{ 5,7,10,12,15,17\}$ with integer generators with values running in $[100,2000]$ (calling the functions {\small \texttt{ RandomNumericalSemigroup}} and {\small\texttt{MinimalGeneratingSystemOfNumericalSemigroup}}).
We denote an instance by $(p, S(p))$, where $S(p)= \{n_1,\cdots, n_p\}$. A couple $(p, , S(p))$ produces a multiple objective instance $ (p, m,n)$,
where $p$ is the number of objectives, $m=p+1 $ the number of constraints and $n=2p$ the number of variables. CPLEX solver within MATLAB environment is used to solve these latter. For each instance $(p,S(p))$, $10$ problems are to be solved.
\noindent The method described in section \ref{description} and the one presented in \cite{bvictor} ( referred in table \ref{compute} as Algorithm 1 and Algorithm Blanco respectively)
were implemented in a MATLAB environment and run on a personal computer, Intel (R) Core (TM) i5 CPU 2.5 GHz. CPLEX 12.6 solver is also used to solve linear and integer linear programming problems.
The average CPU times obtained with both approaches are reported in Table \ref{compute}. We report the minimum, maximum and median CPU times when running the $10$ instances for each $p$.
\begin{center}
\begin{table}[h]
\normalsize\centering\begin{tabular}{|l|ll|ll|}\hline
\multirow{2}{*}{$S(p)$} & \multicolumn{2}{c}{Algorithm \ref{alg} CPUTime (s)} & \multicolumn{2}{|c|}{Algorithm \cite{bvictor} CPUTime (s)}\\
& Median & (Max,Min) &Median & (Max,Min) \\\hline
$S(5)$& 2.88 &(0.56, 120.56) & 9.27 & (1.80,16.98)\\
$S(7)$& 7.92 &(1.09, 12.19)& 17.17 & (6.40,31.95)\\
$S(10)$& 48.80& (18.98, 97.80) & 129.97 & (46.36,287.36)\\
$S(12)$& 87.23 &(52.83, 17.17)& 160.96 & (91.10,307.47)\\
$S(15)$& 257.06& (115.09, 513.84)& 357.37 & (169.45,850.22)\\
$S(17)$& 883.49& (229.15, 1721.98)& 2390.11 & (497.18,5287.36)\\
\hline
\end{tabular}
\captionsetup{justification=centering}
\centering\caption{Computational results of both algorithms.\label{compute}}
\end{table}
\end{center}
We can observe from Table \ref{compute} the differences between the proposed modified version of the method and the algorithm that is proposed in \cite{bvictor}. The improvements proposed in this paper clearly outperform the previous methodology proposed for solving the $\omega$-primality problem on the randomly generated instances. The main reason is that the new approach avoids many iterations for each problem \eqref{psj} by using the knowledge about the initial bounds computed for the problem. Hence, the number of efficient solutions explored during the execution of the algorithm is much smaller than those generated in \cite{bvictor}.
\section{Conclusion}\label{conclusion}
In this work, present an improved resolution method for an interesting problem in the Theory of Numbers, which is attracting the researchers in the last years, the $\omega$-invariant of a numerical semigroup. In order to avoid the complete enumeration of a large discrete feasible set, a mathematical programming approach is proposed in \cite{bvictor}. There, the problem was formulated as a problem of optimizing a linear function over the set of efficient solutions of a multiobjective integer linear problem. However, a simple algorithm was proposed, which does not fully exploit the nature of the problem. Here, we apply a similar strategy than that proposed in \cite{bvictor} in which an easy-to-handle lower and upper bound computation of the problem allows to prune the search of the optimal solution of the problem. It allows to avoid many of the steps (related with solving integer programming problems), and then, is reflected on the CPU times needed to solve the problem.
Many extensions are possible under the topic of Mathematical Optimization tools applied to Theory of Numbers, in particular in Numerical Semigroups or in Affine Semigroups. In particular, the formulation and resolution of other algebraic indices, as the catenary degree or the tame degree \cite{blanco} will be the topic of forthcoming papers.
\section*{Acknowledgements}
The third author was partially supported by the projects MTM2016-74983-C2-1-R (MINECO, Spain), PP2016-PIP06 (Universidad
de Granada) and the research group SEJ-534 (Junta de Andaluc\'ia).
|
\section{Introduction}
\label{sec:introduction}
The discovery of the Higgs boson (\PH)~\cite{Aad:2012tfa,Chatrchyan:2012xdj}
with a mass of 125\GeV completes the particle content of the standard model (SM).
Even though the SM yields numerous accurate predictions, there are several open questions,
among them the origin of the $\PH$ mass stability at the
electroweak scale. Various models beyond the SM have been proposed that
stabilise the $\PH$ mass at the measured value;
some examples are Little Higgs~\cite{ArkaniHamed:2002qy,Schmaltz:2002wx,Schmaltz:2005ky} or
Composite Higgs models~\cite{Marzocca:2012zn}, in which additional top quark partners
with masses at the TeV scale are predicted. Since the left- (LH) and right-handed (RH) chiral components
of these particles transform in the same way under the SM electroweak
symmetry group, they are often referred to as ``vector-like quarks'' (VLQs).
In contrast to a fourth chiral quark generation, their impact on the
$\PH$ properties is small, such that VLQs have not been excluded by the measurements of $\PH$
mediated cross sections~\cite{Djouadi:2012ae,VLQHandbook:2013,Khachatryan:2016vau}.
Several searches for VLQs have been performed at the CERN LHC, setting lower exclusion limits on the VLQ
mass \ensuremath{m_\mathrm{{VLQ}}}\xspace~\cite{Chatrchyan:2013uxa,Khachatryan:2015axa,Khachatryan:2015oba,Khachatryan:2016vph,Sirunyan:2017tfc,
Sirunyan:2017jin,Sirunyan:2017usq,Sirunyan:2017pks,Sirunyan:2017ynj,Sirunyan:2018fjh,
Aad:2011yn,Aad:2012bdq,Aad:2014efa,Aad:2015gdg,Aad:2015kqa,Aad:2015mba,Aaboud:2016lwz,Aad:2016qpo,Aad:2016shx,Aaboud:2017qpr,Aaboud:2017zfn,
Aaboud:2018xuw}.
Many of these analyses study the pair production of VLQs via the strong interaction.
In contrast, the analysis presented here searches for the single VLQ production via the weak
interaction, where a hadronic jet is emitted at a low angle with respect to the beam direction.
Furthermore, VLQs with enhanced couplings to the third generation quarks (\ie VLQ \PQB and \ensuremath{X_{5/3}}\ quarks with an electric charge of $1/3$ and $5/3$ respectively) are produced in association with a bottom ($\cPqb$) or top ($\cPqt$) quark, leading to the \PQB{}+\cPqb, \PQB{}+\cPqt, and \ensuremath{X_{5/3}}+\cPqt\ production modes.
While a VLQ \PQB quark could decay into the $\PH\cPqb$, $\PZ\cPqb$, or
$\cPqt\PW$ final state, a VLQ \ensuremath{X_{5/3}}\ quark could only decay into the \cPqt\PW\ final state. This search focuses on the \cPqt\PW\ final state.
In Fig.~\ref{fig:feynman}, two leading-order (LO) Feynman diagrams are shown for the single production of \PQB and \ensuremath{X_{5/3}}\ quarks and their decay into \cPqt\PW. This paper presents the first search of this signature in proton-proton ($\Pp\Pp$) collision data recorded at a centre-of-mass energy of 13 \TeV. Results at $\sqrt{s} = 8 \TeV$ have been obtained by the ATLAS collaboration~\cite{Aad:2015voa}.
\begin{figure}[bh]
\centering
\includegraphics[width=\cmsFeyWidth]{Figure_001-a.pdf}\hspace*{0.08\columnwidth}
\includegraphics[width=\cmsFeyWidth]{Figure_001-b.pdf}
\caption{Leading order Feynman diagrams for the production
of a single vector-like \PQB or $\ensuremath{X_{5/3}}$ quark in association
with a $\cPqb$ (left) or $\cPqt$ (right) and a light-flavour quark,
and the subsequent decay of the VLQ to $\cPqt\PW$.}
\label{fig:feynman}
\end{figure}
In this analysis, final states with a single muon or electron, several hadronic jets, and missing transverse momenta $\ptmiss$ are studied.
Because of the high mass of the VLQ, the $\cPqt$ and $\PW$ can have high Lorentz boosts,
leading to highly collimated decays of the $\PW$ boson, the top quark and non-isolated leptons.
For signal events, the mass of the \PQB and $\ensuremath{X_{5/3}}$ quarks can be reconstructed
using hadronic jets, the lepton, and the $\ptmiss$. The associated \cPqb\ and \cPqt, as well as the leptons originating from their decay,
have much lower transverse momenta $\pt$ and are not considered for the reconstruction or selection.
The dominant SM background processes are top quark pair (\ttbar) production,
$\PW$+jets and $\PZ$+jets production, single $\cPqt$ production, and multijet production via the strong force.
All SM backgrounds contributing to this search are predicted from dedicated control regions
in data, defined through the absence of a forward jet.
This paper is organised as follows: Section~\ref{sec:cmsdec} provides a description of the CMS detector.
Section~\ref{sec:samples} introduces the data set and the simulated events.
This is followed by the event selection in Section~\ref{sec:object}, as well as by
the description of the reconstruction of the VLQ mass in Section~\ref{sec:reco}.
In Section~\ref{sec:background}, a method to estimate the background is discussed.
Systematic uncertainties are detailed in Section~\ref{sec:systematic}.
The final results of the analysis, as well as the statistical
interpretation in terms of exclusion limits, are discussed in Section~\ref{sec:results}.
\section{The CMS detector and physics objects}
\label{sec:cmsdec}
The central feature of the CMS apparatus is a superconducting solenoid of 6\unit{m} internal diameter, providing a magnetic field of 3.8\unit{T}. Within the solenoid volume are a silicon pixel and strip tracker, a lead tungstate crystal electromagnetic calorimeter (ECAL), and a brass and scintillator hadron calorimeter (HCAL), each composed of a barrel and two endcap sections. Forward calorimeters extend the pseudorapidity coverage provided by the barrel and endcap detectors. Muons are detected in gas-ionisation chambers embedded in the steel flux-return yoke outside the solenoid.
{\tolerance=800
The particle-flow event algorithm~\cite{Sirunyan:2017ulk} aims to reconstruct and identify each individual particle with an optimised combination of information from the various elements of the CMS detector. The energy of photons is directly obtained from the ECAL measurement, corrected for zero-suppression effects. The energy of electrons is determined from a combination of the electron momentum at the primary interaction vertex, the energy of the corresponding ECAL cluster, and the energy sum of all bremsstrahlung photons spatially compatible with originating from the electron track~\cite{Khachatryan:2015hwa}. The energy of muons is obtained from the curvature of the corresponding track~\cite{Chatrchyan:2012xi}. The energy of charged hadrons is determined from a combination of their momentum measured in the tracker and the matching ECAL and HCAL energy deposits, corrected for zero-suppression effects and for the response function of the calorimeters to hadronic showers. Finally, the energy of neutral hadrons is obtained from the corresponding corrected ECAL and HCAL energy.
\par}
{\tolerance=800
The reconstructed vertex with the largest value of summed physics-object $\pt^2$ is taken to be the primary $\Pp\Pp$ interaction vertex. The physics objects used are the jets, clustered with the jet finding algorithm~\cite{Cacciari:2008gp,Cacciari:2011ma} with the tracks assigned to the vertex as inputs, and the associated missing transverse momentum, taken as the negative vector sum of the \pt of those jets.
\par}
A more detailed description of the CMS detector, together with a definition of the coordinate system used and the relevant kinematic variables, can be found in Ref.~\cite{Chatrchyan:2008zzk}.
\section{Data and simulated samples}
\label{sec:samples}
In this analysis, $\Pp\Pp$ collision data at a centre-of-mass energy of
$13\TeV$ taken in 2016 by the CMS experiment are analyzed.
The data have been collected with muon and electron triggers~\cite{Khachatryan:2016bia}.
For the muon trigger, a muon candidate with $\pt > 50 \GeV$ is required.
Data events in the electron channel are collected using a logical combination
of two triggers: the first requires an electron candidate
with $\pt > 45\GeV$ and a hadronic jet candidate with
$\pt > 165\GeV$, the second requires an electron candidate with $\pt > 115\GeV$.
In the trigger selection, reconstructed leptons and
jets must be in the central part of the detector, with a pseudorapidity of $\abs{\eta} < 2.4$.
No lepton isolation criteria are applied at the trigger level.
The collected data correspond to an integrated luminosity of 35.9\fbinv~\cite{CMS-PAS-LUM-17-001}.
For the study of dominant SM background processes and for the validation of the background estimation,
simulated samples using Monte Carlo (MC) techniques are used.
The top quark pair production via the strong interaction and single top quark production
in the $t$-channel, and the $\cPqt\PW$ process are generated with the next-to-leading-order (NLO)
generator \POWHEG~\cite{Nason:2004rx,Frixione:2007vw,Alioli:2010xd}
(version v2 is used for the first two and version v1 for the third).
The event generator \MGvATNLO(v2.2.2)~\cite{Alwall:2014hca} at NLO is used for single top quark
production in the $s$-channel. The $\PW$+jets and $\PZ$+jets processes are also simulated
using \MGvATNLO(v2.2.2). The $\PW$+jets events are generated at
NLO, and the FXFX scheme~\cite{Frederix:2012ps} is used to match the parton shower emission.
The $\PZ$+jets events are produced at LO with the MLM parton matching scheme~\cite{Alwall:2007fs}.
The production of quantum chromodynamics (QCD) multijet events has been
simulated at LO using \PYTHIA~\cite{Sjostrand:2014zea}.
All generated events are interfaced with \PYTHIA for the description of the parton
shower and hadronisation.
The parton distribution functions (PDFs) are taken from the NNPDF 3.0~\cite{Ball:2014uwa}
sets, with the precision matching that of the matrix element calculations.
The underlying event tune is CUETP8M1~\cite{Khachatryan:2015pea,Skands2014},
except for the simulation of top quark pairs and single top quark production in the $t$-channel,
which use CUETP8M2T4~\cite{CMS-PAS-TOP-16-021}.
{\tolerance=3000
Signal events are generated at LO using \MGvATNLO for \PQB and $\ensuremath{X_{5/3}}$ with VLQ decay widths
relative to the VLQ mass of $(\Gamma/m)_{\mathrm{VLQ}} = 1$, 10, 20, and 30\%.
The samples with 1\% relative VLQ width are simulated in steps of 100\GeV for masses between 700 and 2000\GeV.
Samples with 10, 20, and 30\% relative VLQ widths are generated in steps of 200\GeV for masses
ranging from 800 to 2000\GeV, using a modified version of the
model proposed in Refs.~\cite{Buchkremer:2013bha,Fuks:2016ftf,Oliveira:2014kla}.
Separate signal samples are generated for the two main production modes, in which VLQs
are produced in association either with a \cPqb\ quark or with a \cPqt\ quark,
viz.\ $\Pp\Pp \to \PQB \cPqb \cPq$ and $\Pp\Pp \to \PQB \cPqt \cPq$.
The theoretical cross sections for VLQ production are calculated
using Refs.~\cite{Campbell:2004ch,Matsedonskyi:2014mna,Carvalho:2018jkq},
where a simplified approach is used to provide a model-independent
interpretation of experimental results for narrow and large mass width scenarios, as already used for the
interpretation of singly produced vector-like \PQT and \PQB quarks~\cite{Sirunyan:2017ynj,Sirunyan:2018fjh}.
The $\textsc{MADSPIN}$ package~\cite{Frixione:2007zp,Artoisenet:2012st}
is used to retain the correct spin correlations of the top quark and $\PW$ boson decay products.
Interference effects between signal and SM processes have been found to be negligible in this analysis.
\par}
All generated events are passed through a \GEANTfour~\cite{AGOSTINELLI2003250}
based detector simulation of the CMS detector. Additional $\Pp\Pp$ interactions originating
from the same bunch crossing (in-time pileup), as well as from the following or
previous bunch crossings (out-of-time pileup) are taken into account in the simulation.
\section{Event selection}
\label{sec:object}
The physics objects used in this analysis are muons, electrons, hadronic jets, $\ptvecmiss$,
and \ensuremath{S_{\mathrm{T},\text{lep}}}\xspace (defined as the scalar sum of the lepton $\pt$ and $\ptmiss$).
For each event, jets are clustered from reconstructed particles using the
infrared and collinear safe anti-\kt algorithm~\cite{Cacciari:2008gp} with a
distance parameter $R=0.4$ (AK4 jet). Additionally, jets with $R=0.8$ (AK8 jet) are also clustered in
every event with the anti-\kt algorithm, which are used for $\cPqt$ and $\PW$ tagging.
The jet clustering is performed with the \FASTJET~\cite{Cacciari:2011ma} package.
Jet momentum is determined as the vectorial sum of all particle momenta in the jet,
and is found from simulation to be within 5--10\% of the true momentum over the whole \pt
spectrum and detector acceptance. Additional $\Pp\Pp$ interactions within the same or
nearby bunch crossings can contribute additional tracks and calorimetric energy depositions
to the jet momentum. To mitigate this effect, tracks identified to be originating from
pileup vertices are discarded, and an offset correction is applied to correct for
remaining contributions. Jet energy corrections are derived from simulation studies
so that the average measured response of jets becomes identical to that of particle level jets. In situ
measurements of the momentum balance in dijet, photon+jet, $\PZ$+jet, and multijet events
are used to account for any residual differences in the jet energy scale in data and simulation.
Additional selection criteria are applied to each jet to remove jets potentially dominated
by anomalous contributions from various subdetector components or reconstruction
failures~\cite{Khachatryan:2016kdb}.
From the corrected and reconstructed AK4 jet$\mathrm{s}$, those are considered that have $\pt > 30 \GeV$ and $\abs{\eta} < 4$,
while AK8 jet$\mathrm{s}$ must have $\pt>170 \GeV$ and $\abs{\eta} < 2.4$.
Events selected in the analysis are required to have one reconstructed
muon or electron with $\pt >55 \GeV$ and
$\abs{\eta} < 2.4$. Electrons and muons are selected using tight quality
criteria with small misidentification probabilities of about 0.1\% for muons
and 1\% for electrons~\cite{Khachatryan:2015hwa, Sirunyan:2018fpa}.
In the electron channel, a AK4 jet\ must have
$\pt>185 \GeV$ and $\abs{\eta} < 2.4$ if the electron has $\pt <120 \GeV$,
reflecting the trigger selection.
Events with more than one muon or electron passing the same tight identification
criteria and having $\pt > 40 \GeV$ and $\abs{\eta} < 2.4$ are discarded.
Selected events contain two AK4 jet$\mathrm{s}$ with $\pt > 50 \GeV$, which are in the central part of the detector with $\abs{\eta} < 2.4$.
Additionally at least one AK8 jet\ is required. For the reconstruction AK4 jet$\mathrm{s}$ are used with $\pt > 30 \GeV$ and $\abs{\eta} < 2.4$,
while the AK4 jet$\mathrm{s}$ emitted close to the beam pipe and employed in the background estimation must fulfill $\pt > 30 \GeV$ and $ 2.4 < \abs{\eta} < 4$.
Because of the high Lorentz boosts of the top quarks and $\PW$ bosons from the heavy VLQ decay,
signal events can have leptons in close vicinity to the jets.
For this reason, standard lepton isolation would reduce the selection
efficiency considerably. Therefore, for the suppression of events originating
from QCD mulitjet processes, either the perpendicular component of the lepton momentum relative to the geometrically closest AK4 jet\ $p_{\mathrm{T,rel}}$, is required to
exceed $40 \GeV$ or the angular distance of the lepton to the jet,
$\Delta R (\ell,\mathrm{jet}) = \sqrt{\smash[b]{(\Delta \eta)^2 + (\Delta \phi)^2}}$, must be larger than 0.4,
where $\phi$ is the azimuthal angle in radians. Furthermore, for selecting an event, the magnitude of
$\ptvecmiss$ has to be greater than 50\GeV in the muon channel and greater than 60\GeV in the electron channel.
This requirement reduces the amount of background from multijet production.
The final selection is based on the variable \ensuremath{S_{\mathrm{T},\text{lep}}}\xspace,
which is required to be larger than 250\GeV in the muon channel and 290\GeV in the electron channel.
Events are separated into categories exploiting the tagging techniques for
boosted top quarks and $\PW$ bosons decaying hadronically, as well as for hadronic jets originating from $\cPqb$ quarks.
Jets with $R=0.8$ are used to identify the hadronic decays of highly boosted top quarks and
$\PW$ bosons~\cite{CMS:2016tvk,CMS-PAS-JME-16-003}.
For top quark jets $\pt>400 \GeV$ is required,
and for $\PW$ boson jets the requirement is $\pt>200 \GeV$.
The ``soft drop'' (SD) declustering and grooming algorithm~\cite{Dasgupta:2013ihk, Larkoski:2014wba}
with $z=0.1$ and $\beta=0$ is employed to identify subjets and to remove soft and
wide-angle radiation.
The groomed jet mass, \ensuremath{m_{\mathrm{SD}}}\xspace, is used to identify top quark and $\PW$ boson candidates.
Tagged top quark candidates ($\cPqt$ tagged) are required to have $105 < \ensuremath{m_{\mathrm{SD}}}\xspace < 220 \GeV$
and one of the subjets must fulfill the loose $\cPqb$ tagging criterion, based on
the combined secondary vertex (CSVv2)~\cite{Sirunyan:2017ezt} algorithm. The loose
criterion is defined to give a 80\% efficiency
of correctly identifying \cPqb\ jets, with a 10\% probability of incorrectly tagging a light quark jet.
Additionally, the jet must have a
N-subjettiness~\cite{Thaler:2010tr,Thaler:2011gf} ratio $\tau_3/\tau_2 < 0.5$ and
its angular distance to the lepton $\Delta R{(\ell,\cPqt\ \mathrm{tag})}$
must be larger than 2.
Identified $\PW$ boson candidates ($\PW$ tag) must have $65 < \ensuremath{m_{\mathrm{SD}}}\xspace < 95 \GeV$.
The medium $\cPqb$ tag criterion is used on AK4 jet$\mathrm{s}$, defined to give a $60\%$ efficiency of correctly identifying \cPqb\ jets,
with a $1\%$ probability of incorrectly tagging a light quark jet.
Selected events are attributed to different mutually exclusive event categories.
Events containing at least one $\cPqt$ tag constitute the first category
(``\cPqt\ tag''). If no $\cPqt$ tag is found,
all events with at least one $\PW$ tag
are grouped into a second category (``\PW\ tag'').
The remaining events are attributed to three further
categories based on the multiplicity of $\cPqb$ tags
found in the event. We distinguish events with at least
two (``${\geq}2$ \cPqb\ tag''),
exactly one (``1 \cPqb\ tag''), and no $\cPqb$ tag (``0 \cPqb\ tag'').
These five categories are built separately in the muon and in the electron
channel leading to a total of ten categories.
\section{Mass reconstruction}
\label{sec:reco}
Hadronic jets, leptons, and $\ptvecmiss$ are used to reconstruct the
mass of the VLQ, denoted \ensuremath{m_\mathrm{reco}}\xspace.
In signal events, the lepton in the final state always originates
from the decay of a $\PW$ boson, either the $\PW$ boson
from the VLQ decay or the $\PW$ boson from the top quark decay.
The neutrino four-momentum can thus be reconstructed from the
components of $\ptvecmiss$, the $\PW$ mass constraint, and the assumption
of mass\-less neutrinos.
In the case when a hadronic jet with a $\cPqt$ tag is found, \ensuremath{m_\mathrm{reco}}\xspace is calculated
from the four-momentum of the $\cPqt$-tagged jet and the four-momentum of the
leptonically decaying $\PW$ boson.
If several hadronic jets with $\cPqt$ tags are present,
the one with the largest angular distance to the reconstructed leptonic
$\PW$ boson decay is used.
Once the $\cPqt$-tagged jet has been selected, all overlapping
AK4 jet\ jets in the event are removed in order to avoid double counting of energy.
For the shown \ensuremath{m_\mathrm{reco}}\xspace distributions these events form the $\cPqt$ tag category.
For events in the other categories the hadronic part of the VLQ decay is reconstructed from
combinations of AK4 jet$\mathrm{s}$ with $\abs{\eta} < 2.4$.
Each possible jet assignment for the decays of the \PW\ boson
and \cPqt\ quark is tested exploiting the following $\chi^2$ quantity
\begin{linenomath}
\ifthenelse{\boolean{cms@external}}
{
\begin{flalign}
\chi^2 &=
\frac{\left(\ensuremath{m_{\cPqt}}\xspace - \ensuremath{\overline{m}_{\cPqt}}\xspace\right)^2}{\sigma^2_{\cPqt}} +
\frac{\left(\ensuremath{m_{\PW}}\xspace - \ensuremath{\overline{m}_{\PW}}\xspace\right)^2}{\sigma^2_{\PW}} \nonumber \\
&+ \frac{\left(\Delta R(\cPqt,\PW)-\pi \right)^2}{\sigma^2_{\Delta R}}+
\frac{\left(\ensuremath{p_{\mathrm{T},\PW}}\xspace / \ensuremath{p_{\mathrm{T},\cPqt}}\xspace - 1\right)^2}{\sigma^2_{\pt}}.
\label{eq:chi2}
\end{flalign}
}
{
\begin{flalign}
\chi^2 =
\frac{\left(\ensuremath{m_{\cPqt}}\xspace - \ensuremath{\overline{m}_{\cPqt}}\xspace\right)^2}{\sigma^2_{\cPqt}} +
\frac{\left(\ensuremath{m_{\PW}}\xspace - \ensuremath{\overline{m}_{\PW}}\xspace\right)^2}{\sigma^2_{\PW}}
+ \frac{\left(\Delta R(\cPqt,\PW)-\pi \right)^2}{\sigma^2_{\Delta R}}+
\frac{\left(\ensuremath{p_{\mathrm{T},\PW}}\xspace / \ensuremath{p_{\mathrm{T},\cPqt}}\xspace - 1\right)^2}{\sigma^2_{\pt}}.
\label{eq:chi2}
\end{flalign}
}
\end{linenomath}
For each event, the jet assignment
with the maximum $\chi^2$ probability is selected.
For the $\chi^2$ quantity
the $\pt$ balance, $\ensuremath{p_{\mathrm{T},\PW}}\xspace / \ensuremath{p_{\mathrm{T},\cPqt}}\xspace$, the angular distance,
$\Delta R(\cPqt,\PW)$, and the reconstructed masses of the top quark candidate
$\ensuremath{m_{\cPqt}}\xspace$ and the $\PW$ boson candidate $\ensuremath{m_{\PW}}\xspace$ are used.
The expected values $\ensuremath{\overline{m}_{\cPqt}}\xspace$ and $\ensuremath{\overline{m}_{\PW}}\xspace$, and their standard
deviations $\sigma_{\cPqt}$ and $\sigma_{\PW}$ are obtained from
simulation for correctly reconstructed events
and it is verified that the values are independent
of the VLQ mass.
Here, correctly reconstructed events are defined by the assignment of jets to
generated $\cPqt$ quarks and $\PW$ bosons, where the generated particles from the
VLQ decay are unambiguously matched within a distance of $\Delta R < 0.4$ to the
reconstructed particles.
It was also verified in simulation that the expected values
of $\Delta R(\cPqt,\PW)$ and the $\pt$ balance are $\pi$ and $1$,
with their standard deviations $\sigma_{\Delta R}$ and $\sigma_{\pt}$.
In order to account for cases where the $\PW$ boson from the VLQ decay
decays into a lepton and neutrino, the $\chi^2$ is calculated for each permutation
with the second term omitted.
Cases where the hadronic decay products of the $\PW$ bosons or the top quark are
reconstructed in a single AK4 jet\ are included
by omitting the first or second term in the calculation
of the $\chi^2$.
\begin{figure}
\centering
\includegraphics[width=\cmsFigWidth]{Figure_002-a.pdf}
\includegraphics[width=\cmsFigWidth]{Figure_002-b.pdf}
\caption{Distributions of $\ensuremath{m_\mathrm{reco}}\xspace$ for the \PQB{}+\cPqb\ production
mode, obtained for simulated events with a muon in the final state, reconstructed with a $\cPqt$ tag
(\cmsLeft) and with the $\chi^2$ method (\cmsRight) for right-handed VLQ couplings and various VLQ masses $m_{\PQB}$. Signal events are shown assuming a production cross section of 1\unit{pb} and a relative VLQ decay width of 1\%.}
\label{fig:Bt_mass_moun_rh}
\end{figure}
The distributions of \ensuremath{m_\mathrm{reco}}\xspace in simulation for the \PQB{}+\cPqb\ production mode with right-handed couplings are shown in Fig.~\ref{fig:Bt_mass_moun_rh}
for events with a muon in the final state. The reconstruction
of events with a $\cPqt$ tag (\cmsLeft) is best suited for high VLQ masses
where the decay products of the top quark are highly boosted, while
the $\chi^2$ method (\cmsRight) yields a stable performance for all VLQ masses,
where the decay products of the $\PW$ boson and top quark are
reconstructed from several jets. Additionally, the latter method
enables the reconstruction of events with a lepton from the top
quark decay chain.
Mass resolutions between 10--15\% are achieved for both reconstruction methods,
with peak values of the $\ensuremath{m_\mathrm{reco}}\xspace$ distributions at the expected values.
The VLQs with left-handed couplings (not shown) have a lower selection
efficiency by 20--25\% because of a smaller lepton $\pt$, on average,
but otherwise features a behaviour similar to VLQs with right-handed couplings.
Distributions obtained for the final states with an electron are similar
to those with a muon.
\section{Background estimation}
\label{sec:background}
The data sample obtained after the selection is then divided into
a signal region with a jet in the forward region of the detector with
$2.4 < \abs{\eta} < 4.0$ and a control region without such a jet.
The distribution of background processes in the signal region
is estimated using the shape of the \ensuremath{m_\mathrm{reco}}\xspace\ distribution in the control region.
Residual differences in the shapes of the \ensuremath{m_\mathrm{reco}}\xspace distributions
between signal and control regions are investigated in each of the
signal categories by using simulated SM events.
Differences can arise from different
background compositions in signal and control regions due to the
presence of a forward jet.
The observed differences are small, with average values of 10\%,
and are corrected for by multiplicative factors applied to the background predictions
in the validation and signal regions. The largest differences are observed for \ensuremath{m_\mathrm{reco}}\xspace
values below 800\GeV, with values no larger than about 20\%.
In order to validate the VLQ mass reconstruction, data are compared to
simulation in the control region. In Fig.~\ref{fig:background_mu}
the distributions of \ensuremath{m_\mathrm{reco}}\xspace are shown
in the muon (upper) and electron (lower) channels for events with
a $\cPqt$ tag (left) and events reconstructed with the $\chi^2$
method (right). The \ttbar and $\cPqt\PW$ standard model processes provide
irreducible backgrounds in the reconstructed VLQ mass distributions, showing good agreement between the data and simulation.
The contribution of signal events in the control region is small and is taken into account by a simultaneous fit to signal and control regions in the statistical extraction of the results.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Figure_003-a.pdf}
\includegraphics[width=0.45\textwidth]{Figure_003-b.pdf}
\includegraphics[width=0.45\textwidth]{Figure_003-c.pdf}
\includegraphics[width=0.45\textwidth]{Figure_003-d.pdf}
\caption{Distributions of \ensuremath{m_\mathrm{reco}}\xspace in data and simulation in the control
region for the muon (upper) and electron (lower) channels
for events reconstructed with a $\cPqt$ tag (left) and with
the $\chi^2$ method (right). The VLQ signal is shown for the \PQB{}+\cPqb\ production mode and right-handed VLQ couplings.
The vertical bars illustrate the statistical
uncertainties on the data, while the shaded area shows the total
uncertainties for the background simulation. The lower panels show the
ratio of data to the background prediction. The dark and light gray bands
correspond to the statistical and total uncertainties,
respectively.
\label{fig:background_mu} }
\end{figure*}
In order to validate the background estimation, a validation region
is constructed from requiring events with reconstruction
$p$-values smaller than 0.08. The $p$-values are calculated as
the probability of obtaining the $\chi^2$ as given by Eq.~\eqref{eq:chi2},
where the number of degrees of freedom of the selected hypothesis
are taken into account. For events with a $\cPqt$ tag, the same $\chi^2$
quantity is evaluated for the selected hypothesis.
The validation region has an order of magnitude fewer events than the
signal region and a negligible amount of signal contamination.
The \ensuremath{m_\mathrm{reco}}\xspace distributions for the two most sensitive categories are
shown in Fig.~\ref{fig:validation_mu} for the muon (upper) and
electron (lower) channels.
The observed number of events is found to be in good
agreement with the predicted number of events from the background estimation
in the validation region, with no statistically significant deviations.
Similar observations are made for the other signal categories.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Figure_004-a.pdf}
\includegraphics[width=0.45\textwidth]{Figure_004-b.pdf}
\includegraphics[width=0.45\textwidth]{Figure_004-c.pdf}
\includegraphics[width=0.45\textwidth]{Figure_004-d.pdf}
\caption{Distributions of \ensuremath{m_\mathrm{reco}}\xspace in the validation region
of the two most sensitive categories in the muon channel (upper)
and electron channel (lower). The lower panels show the difference of data
and background expectations in units of the total (stat. and sys.)
uncertainty on the background estimate.
\label{fig:validation_mu}}
\end{figure*}
\section{Systematic uncertainties}
\label{sec:systematic}
Systematic uncertainties can affect both the overall normalisation
of background components and the shapes of the \ensuremath{m_\mathrm{reco}}\xspace distributions
for signal and background processes.
The main uncertainty in the shape of the \ensuremath{m_\mathrm{reco}}\xspace distribution from the
background estimation based on a control region in data is related to the kinematic difference
between the signal and control regions. Correction factors are applied
to account for this difference, obtained from SM simulations.
These uncertainties have a size of 10\% on average, with maximum values
of 20\% at small values of \ensuremath{m_\mathrm{reco}}\xspace. Compared to these uncertainties, the
effects from uncertainties in the SM simulations are negligible on the background estimation,
as these cancel to a large degree when building the ratios between signal and control regions.
The uncertainties in the overall
normalisation of the background predictions are obtained from a fit
to the data in the signal region.
Uncertainties in the MC simulation are applied to all simulated
signal events. In the following, the systematic uncertainties are summarized.
\begin{itemize}
\item The uncertainty in the integrated luminosity measurement
recorded with the CMS detector in the 2016 run
at $\sqrt{s}=13\TeV$ is 2.5\%~\cite{CMS-PAS-LUM-17-001}.
\item The estimation of pileup effects is based on the total
inelastic cross section. This cross section is determined to
be 69.2\unit{mb}. The uncertainty is taken into account by
varying the total inelastic cross section by 4.6\%~\cite{Chatrchyan:2012nj}.
\item Simulated events are corrected for lepton identification,
trigger, and isolation efficiencies. The corresponding
corrections are applied as functions of $\abs{\eta}$ and $\pt$.
The systematic uncertainties due to these corrections are
taken into account by varying each correction factor within its uncertainty.
\item The scale factors for the jet energy scale
and resolution are determined as functions of $\abs{\eta}$ and $\pt$~\cite{Khachatryan:2016kdb}.
The effect of the uncertainties in these scale factors are considered by varying
the scale factors within their uncertainties. Jets with distance parameters of
0.4 and 0.8 are modified simultaneously. The results of variations for AK4 jet$\mathrm{s}$ are propagated to the measurement of \ptvecmiss.
\item The uncertainties due to the PDFs are evaluated by considering
100 replicas of the NNPDF~3.0 set according to the procedure
described in Ref.~\cite{Butterworth:2015oua}.
The associated PDF uncertainties in the signal acceptance are estimated following the prescription for the LHC~\cite{Butterworth:2015oua}.
\item Uncertainties associated with variations of the factorisation $\mu_\mathrm{f}$ and renormalisation scales $\mu_\mathrm{r}$
are evaluated by varying the respective scales independently, by factors of 0.5 and 2.
\item Corrections for the \cPqb\ tagging efficiencies and
misidentification rates for AK4 jet$\mathrm{s}$, and subjets of AK8 jet$\mathrm{s}$ are applied.
These are measured as a function of the jet $\pt$~\cite{Sirunyan:2017ezt}.
The corresponding uncertainties are taken into account by varying
the corrections within their uncertainties for heavy- and light-flavour
jets separately.
\item An uncertainty on the \cPqt\ tagging efficiency
of $+7$ and $-4\%$ is applied to signal events with a \cPqt\ tag~\cite{CMS-PAS-JME-16-003}.
The uncertainty on the $\PW$ tagging efficiency is determined
from jet mass resolution (JMR) and scale (JMS) uncertainties,
which are added in quadrature. An additional JMR uncertainty is derived from
the differences in the hadronisation and shower models of
\PYTHIA and \HERWIGpp~\cite{Bahr:2008pv}. The uncertainty depends on the \pt
of the \PW\ boson; for VLQs with a mass of 700\GeV it is around 2\% and for
a mass of 1800\GeV it is around 6\%. An uncertainty of 1\% is assigned to the JMS,
as obtained from studies of the jet mass in fully merged hadronic \PW\ boson decays.
\end{itemize}
In Table~\ref{tab:unc_impact}, a summary of the uncertainties
considered for signal events is shown, where the largest uncertainties
come from the jet energy scale and the jet tagging. For the uncertainties
connected to the PDF, $\mu_\mathrm{f}$ and $\mu_\mathrm{r}$ only the signal
acceptance and shape differences are propagated. The uncertainties with the
largest impact on the analysis are the uncertainties associated with
the data-driven background estimation, being more than two times larger than
the jet energy scale uncertainties in the signal.
\begin{table}
\centering
\topcaption{Uncertainties considered for simulated signal
events in the \PQB{}+\cPqb\ production mode ($m_{\PQB{}} = 900 \GeV$) for right-handed VLQ couplings for the \cPqt\ tag and \PW\ tag categories. The uncertainties in the \cPqb\ tag categories are of comparable size to those in the \PW\ tag category.}
\begin{tabular}{l c c c c}
Uncertainty & & $\cPqt$ tag [$\%$] & \PW\ tag [$\%$] \\ \hline
\PW\ tagging & Rate & \NA & 3.3 \\
\cPqt\ tagging & Rate & $^{+7}_{-4}$ & \NA \\
Luminosity & Rate & 2.5 & 2.5 \\
Pileup & Shape & 1--3 & 0.2 \\
Lepton reconstruction & Shape & 2--3 & 2--3 \\
\cPqb\ tagging & Shape & 2.5 & 2.5 \\
Jet energy scale & Shape & 2--6 & 1--5 \\
Jet energy resolution & Shape & 1--2 & 1--2 \\
PDF & Shape & 2--3 & 0.5 \\
$\mu_f$ and $\mu_r$ & Shape & 0.3 & 0.2 \\
\end{tabular}
\label{tab:unc_impact}
\end{table}
\section{Results}
\label{sec:results}
The \ensuremath{m_\mathrm{reco}}\xspace distributions in the ten categories are measured in the signal
and control region, which are defined by the presence or absence of a
forward jet with $\abs{\eta} > 2.4$.
For the background estimate in the signal regions, a simultaneous
binned maximum likelihood fit of both regions is performed
using the \textsc{Theta}~\cite{theta} package.
In these fits, the signal cross section and the background
normalisations in the different signal categories are free parameters.
The shapes of the \ensuremath{m_\mathrm{reco}}\xspace distributions for the SM background in the signal regions are taken
from the corresponding control regions.
Systematic uncertainties are taken into account as additional
nuisance parameters. A common nuisance parameter is used for
uncertainties in the muon and electron channels if a similar effect is
expected on the shape or normalisation of the
\ensuremath{m_\mathrm{reco}}\xspace distribution in both channels similarly.
The nuisance parameters for the shape uncertainties
are taken to be Gaussian distributed. For the uncertainties on the normalisation log-normal prior distributions are assumed.
The measured distributions of \ensuremath{m_\mathrm{reco}}\xspace for the signal
categories are shown in Figs.~\ref{fig:signalregion_mu}
and \ref{fig:signalregion_ele} for the muon and electron channels,
together with the background predictions obtained from the control regions.
The signal \ensuremath{m_\mathrm{reco}}\xspace\ distributions for a vector-like \PQB quark with right-handed
couplings produced in association with a \cPqb\ quark are shown for illustration,
for two different VLQ masses with an assumed production cross section of 1\unit{pb} and
a relative VLQ width of 1\%. No significant deviation from the background expectation is observed in any of the categories.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Figure_005-a.pdf}
\includegraphics[width=0.45\textwidth]{Figure_005-b.pdf}
\includegraphics[width=0.45\textwidth]{Figure_005-c.pdf}
\includegraphics[width=0.45\textwidth]{Figure_005-d.pdf}
\includegraphics[width=0.45\textwidth]{Figure_005-e.pdf}
\caption{Distributions of \ensuremath{m_\mathrm{reco}}\xspace\ measured in the signal region for
events with a jet in the forward direction with $\abs{\eta}>2.4$ in the muon channel.
Shown are the sensitive categories:
\cPqt\ tag (upper left),
\PW\ tag (upper right),
${\geq}2$ \cPqb\ tag (middle left),
1 \cPqb\ tag (middle right) and 0 \cPqb\ tag (lower).
The background prediction is obtained from control regions as
detailed in the main text. The distributions from two example signal samples
for the \PQB{}+\cPqb\ production mode with right-handed VLQ couplings
with a cross section of 1\unit{pb} and a relative width of 1\% are shown for illustration.
\label{fig:signalregion_mu} }
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Figure_006-a.pdf}
\includegraphics[width=0.45\textwidth]{Figure_006-b.pdf}
\includegraphics[width=0.45\textwidth]{Figure_006-c.pdf}
\includegraphics[width=0.45\textwidth]{Figure_006-d.pdf}
\includegraphics[width=0.45\textwidth]{Figure_006-e.pdf}
\caption{Distributions of \ensuremath{m_\mathrm{reco}}\xspace\ measured in the signal region for
events with a jet in the forward direction with $\abs{\eta}>2.4$ in the electron channel.
Shown are the sensitive categories:
\cPqt\ tag(upper left),
\PW\ tag(upper right),
${\geq}2$ \cPqb\ tag (middle left),
1 \cPqb\ tag (middle right) and 0 \cPqb\ tag (lower).
The background prediction is obtained from control regions as
detailed in the main text. The distributions from two example signal samples
for the \PQB{}+\cPqb\ production mode with right-handed VLQ couplings
with a cross section of 1\unit{pb} and a relative VLQ width of 1\% are shown for illustration.
\label{fig:signalregion_ele}}
\end{figure*}
Exclusion limits on the product of the VLQ production cross section and branching fraction
are calculated at 95\% confidence level (\CL) for VLQ masses
between 700 and 2000\GeV by using a Bayesian
statistical method~\cite{theta, bayesbook}.
Pseudo-experiments are performed to extract expected upper
limits under the background-only hypothesis.
For the signal cross section parameter an uniform prior distribution, and
for the nuisance parameters log-normal prior distributions are used.
The nuisance parameters are randomly varied within
their ranges of validity to estimate the 68 and 95\% \CL expected limits.
Correlations between the systematic uncertainties across all
channels are taken into account through a common nuisance parameter.
The statistical uncertainties of the background predictions are treated
as an additional Poisson nuisance parameter in each bin of the
\ensuremath{m_\mathrm{reco}}\xspace\ distribution.
Figure~\ref{fig:limits_Bb} shows the 95\% \CL upper limits on
the product of the cross section and branching fraction for the \PQB{}+\cPqb\
production mode for left- and right-handed VLQ couplings and a relative
VLQ width of 1\% (upper left and upper right), for the left-handed VLQ
couplings and a relative VLQ width of 10\% (lower left), as well as a
comparison of the observed exclusion limits for relative VLQ widths
between 10 and 30\% (lower right).
In Fig.~\ref{fig:limits_combined}, the 95\% \CL upper limits on the product of the cross section and branching fraction for the production
modes \PQB{}+\cPqt\ (upper left) and \ensuremath{X_{5/3}}+\cPqt\ (upper right) and right-handed VLQ couplings are shown.
The figure also shows the \ensuremath{X_{5/3}}+\cPqt\ exclusion limits for left-handed VLQ couplings with a 10\% relative VLQ width (lower left) and
a comparison of the observed exclusion limits for VLQ widths between 10 and 30\% for left-handed couplings (lower right).
The predicted cross sections for variations of the relative VLQ
mass width (dashed lines) are taken from Refs.~\cite{Carvalho:2018jkq,Matsedonskyi:2014mna,Campbell:2004ch}.
For a set of VLQ masses the expected and observed 95\% \CL upper limits for the \PQB{}+\cPqb\ and the \ensuremath{X_{5/3}}+\cPqt\ production modes
are also given in Table~\ref{tab:mass_limits} for VLQs with widths of 1\% and 10\% and left-handed couplings, as well
as for widths of 1\% and right-handed couplings.
The exclusion limits for the \PQB{}+\cPqt\ production mode are similar to those for the \ensuremath{X_{5/3}}+\cPqt\ production mode.
The obtained exclusion limits range from 0.3 to 0.03\unit{pb} for VLQ masses between 700 and 2000\GeV. For VLQs with a relative width of 1\% and purely left-handed
couplings an increase of about 25\% of the 95\% \CL upper limits is observed because of the reduced signal acceptance, in comparison to the right-handed couplings.
The expected limits for VLQ with relative widths of 10--30\% and left-handed couplings only show small differences. Although the predicted cross sections for the SM backgrounds are considerably larger at 13\TeV, similar exclusion limits on the product of cross section and branching fraction are achieved compared to the results obtained at 8\TeV in the more restricted mass range considered in Ref.~\cite{Aad:2015voa}. However, because of the increase of the VLQ signal cross section at 13\TeV, with this analysis, the existence of VLQ \PQB (\ensuremath{X_{5/3}}) quarks with left-handed couplings and a relative width of 10, 20, and 30\% can be excluded for masses below 1490, 1590, and 1660\GeV (920, 1300, and 1450\GeV) respectively. The results represent the most stringent exclusion limits for singly produced VLQ in this channel.
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Figure_007-a.pdf}
\includegraphics[width=0.45\textwidth]{Figure_007-b.pdf}
\includegraphics[width=0.45\textwidth]{Figure_007-c.pdf}
\includegraphics[width=0.45\textwidth]{Figure_007-d.pdf}
\caption{Upper limits at 95\% \CL on the product of the VLQ production cross section and
branching fraction for the \PQB{}+\cPqb\ production mode for a relative VLQ width of 1\% and
left- and right-handed VLQ couplings (upper left and right), for 10\% relative VLQ width
and left-handed VLQ couplings (lower left), and
a comparison of the observed exclusion limits for relative VLQ widths of 10, 20, and 30\%
for left-handed couplings (lower right).
The dashed lines show the theoretical predictions.
\label{fig:limits_Bb} }
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Figure_008-a.pdf}
\includegraphics[width=0.45\textwidth]{Figure_008-b.pdf}
\includegraphics[width=0.45\textwidth]{Figure_008-c.pdf}
\includegraphics[width=0.45\textwidth]{Figure_008-d.pdf}
\caption{Upper limits at 95\% \CL on the product of the VLQ production
cross section and branching fraction for the \PQB{}+\cPqt\ and \ensuremath{X_{5/3}}+\cPqt\ production modes
for right-handed VLQ couplings assuming a relative VLQ width of 1\% (upper left and right),
for the \ensuremath{X_{5/3}}+\cPqt\ production mode with left-handed VLQ couplings and a 10\% relative width (lower left) and a comparison of the observed exclusion
limits for left-handed couplings for relative widths of 10, 20, and 30\% (lower right).
The dashed lines show the theoretical predictions.
\label{fig:limits_combined} }
\end{figure*}
\begin{table*}
\centering
\topcaption{Observed (expected) upper limits at 95\% \CL on the product of the cross section and branching fraction for the \PQB{}+\cPqb\ and \ensuremath{X_{5/3}}+\cPqt\ production modes, for a set of VLQ masses, for VLQs widths of 1\% and 10\%, and for left-handed and right-handed couplings. The exclusion limits for the \PQB{}+\cPqt\ production mode (not shown) are very similar to those for the \ensuremath{X_{5/3}}+\cPqt\ mode.}
\cmsTable{
\begin{tabular}{c c c c c c c}
& \multicolumn{3}{c}{\PQB{}+\cPqb} & \multicolumn{3}{c}{\ensuremath{X_{5/3}}+\cPqt} \\
\ensuremath{m_\mathrm{{VLQ}}}\xspace [TeV] & 1\% LH& 10\% LH& 1\% RH& 1\% LH& 10\% LH& 1\% RH\\ \hline
0.8 & 0.29 (0.36) & 0.27 (0.36) & 0.25 (0.29) & 0.31 (0.27) & 0.32 (0.25) & 0.21 (0.18) \\
1 & 0.29 (0.17) & 0.29 (0.19) & 0.21 (0.12) & 0.25 (0.15) & 0.25 (0.16) & 0.15 (0.10) \\
1.2 & 0.10 (0.10) & 0.11 (0.11) & 0.07 (0.07) & 0.10 (0.09) & 0.10 (0.10) & 0.06 (0.06) \\
1.4 & 0.07 (0.07) & 0.06 (0.08) & 0.03 (0.05) & 0.05 (0.06) & 0.05 (0.07) & 0.03 (0.05) \\
1.6 & 0.05 (0.05) & 0.05 (0.06) & 0.03 (0.04) & 0.04 (0.04) & 0.05 (0.05) & 0.03 (0.03) \\
1.8 & 0.04 (0.04) & 0.05 (0.04) & 0.03 (0.03) & \NA & 0.05 (0.04) & \NA \\
\end{tabular}
}
\label{tab:mass_limits}
\end{table*}
\section{Summary}
\label{sec:summary}
A search for singly produced vector-like quarks decaying into a top quark and a $\PW$
boson has been performed using the 2016 data set recorded by the CMS experiment
at the CERN LHC. The selection is optimised for high vector-like quark masses,
with a single muon or electron, significant missing transverse momentum,
and two jets with high $\pt$ in the final state.
Vector-like quarks in the single production mode can be produced in association with a $\cPqt$ or a $\cPqb$ quark
and a forward jet. The latter feature is used to obtain the background prediction
in the signal regions from data. The mass of the vector-like quark is reconstructed from the hadronic jets,
the missing transverse momentum, and the lepton in the event.
Different decay possibilities of the $\cPqt$ and $\PW$ are considered.
The reach of the search is enhanced by $\cPqt$, $\PW$, and $\cPqb$ tagging methods.
No significant deviation from the standard model prediction is observed.
Upper exclusion limits at 95\% confidence level on the product of the production cross section and
branching fraction range from around 0.3--0.03\unit{pb} for vector-like quark masses between 700 and 2000\GeV.
Depending on the vector-like quark type, coupling, and decay width to \cPqt\PW, mass exclusion limits up to 1660\GeV are obtained.
These represent the most stringent exclusion limits for the single production of vector-like quarks in this channel.
\begin{acknowledgments}
We congratulate our colleagues in the CERN accelerator departments for the excellent performance of the LHC and thank the technical and administrative staffs at CERN and at other CMS institutes for their contributions to the success of the CMS effort. In addition, we gratefully acknowledge the computing centres and personnel of the Worldwide LHC Computing Grid for delivering so effectively the computing infrastructure essential to our analyses. Finally, we acknowledge the enduring support for the construction and operation of the LHC and the CMS detector provided by the following funding agencies: BMWFW and FWF (Austria); FNRS and FWO (Belgium); CNPq, CAPES, FAPERJ, FAPERGS, and FAPESP (Brazil); MES (Bulgaria); CERN; CAS, MoST, and NSFC (China); COLCIENCIAS (Colombia); MSES and CSF (Croatia); RPF (Cyprus); SENESCYT (Ecuador); MoER, ERC IUT, and ERDF (Estonia); Academy of Finland, MEC, and HIP (Finland); CEA and CNRS/IN2P3 (France); BMBF, DFG, and HGF (Germany); GSRT (Greece); NKFIA (Hungary); DAE and DST (India); IPM (Iran); SFI (Ireland); INFN (Italy); MSIP and NRF (Republic of Korea); LAS (Lithuania); MOE and UM (Malaysia); BUAP, CINVESTAV, CONACYT, LNS, SEP, and UASLP-FAI (Mexico); MOS (Montenegro); MBIE (New Zealand); PAEC (Pakistan); MSHE and NSC (Poland); FCT (Portugal); JINR (Dubna); MON, RosAtom, RAS, RFBR, and NRC KI (Russia); MESTD (Serbia); SEIDI, CPAN, PCTI, and FEDER (Spain); MOSTR (Sri Lanka); Swiss Funding Agencies (Switzerland); MST (Taipei); ThEPCenter, IPST, STAR, and NSTDA (Thailand); TUBITAK and TAEK (Turkey); NASU and SFFR (Ukraine); STFC (United Kingdom); DOE and NSF (USA).
\hyphenation{Rachada-pisek} Individuals have received support from the Marie-Curie programme and the European Research Council and Horizon 2020 Grant, contract No. 675440 (European Union); the Leventis Foundation; the A. P. Sloan Foundation; the Alexander von Humboldt Foundation; the Belgian Federal Science Policy Office; the Fonds pour la Formation \`a la Recherche dans l'Industrie et dans l'Agriculture (FRIA-Belgium); the Agentschap voor Innovatie door Wetenschap en Technologie (IWT-Belgium); the F.R.S.-FNRS and FWO (Belgium) under the ``Excellence of Science - EOS" - be.h project n. 30820817; the Ministry of Education, Youth and Sports (MEYS) of the Czech Republic; the Lend\"ulet (``Momentum") Programme and the J\'anos Bolyai Research Scholarship of the Hungarian Academy of Sciences, the New National Excellence Program \'UNKP, the NKFIA research grants 123842, 123959, 124845, 124850 and 125105 (Hungary); the Council of Science and Industrial Research, India; the HOMING PLUS programme of the Foundation for Polish Science, cofinanced from European Union, Regional Development Fund, the Mobility Plus programme of the Ministry of Science and Higher Education, the National Science Center (Poland), contracts Harmonia 2014/14/M/ST2/00428, Opus 2014/13/B/ST2/02543, 2014/15/B/ST2/03998, and 2015/19/B/ST2/02861, Sonata-bis 2012/07/E/ST2/01406; the National Priorities Research Program by Qatar National Research Fund; the Programa Estatal de Fomento de la Investigaci{\'o}n Cient{\'i}fica y T{\'e}cnica de Excelencia Mar\'{\i}a de Maeztu, grant MDM-2015-0509 and the Programa Severo Ochoa del Principado de Asturias; the Thalis and Aristeia programmes cofinanced by EU-ESF and the Greek NSRF; the Rachadapisek Sompot Fund for Postdoctoral Fellowship, Chulalongkorn University and the Chulalongkorn Academic into Its 2nd Century Project Advancement Project (Thailand); the Welch Foundation, contract C-1845; and the Weston Havens Foundation (USA).
\end{acknowledgments}
|
\section{Introduction}
\label{sec: introduction}
Let $\mathbf F$ be a non-archimedean local field and $\mathfrak{o}$ be the ring of integers of $\mathbf F$ such that the residue field $\mathbf k$ is finite of characteristic $p$ and $| \mathbf{k} | =q.$
Let $\mathfrak{p}$ be the unique maximal ideal of $\mathfrak{o}$ and $\varpi$ be a fixed generator of $\mathfrak{p}.$
Let ${\mathbf{G}}$ be a split reductive group scheme defined over $\mathfrak{o}$ and ${\mathbf{G}}(\mathfrak{o})$ be the set of $\mathfrak{o}$-points of ${\mathbf{G}}.$ For $\ell \geq 1,$ let $\mathfrak{o}_{\ell} = \mathfrak{o}/ (\varpi^{\ell}).$ The groups ${\mathbf{G}}(\mathfrak{o})$ are pro-finite therefore every complex continuous finite dimensional irreducible representation of ${\mathbf{G}}(\mathfrak{o})$ factors through the quotient ${\mathbf{G}}(\mathfrak{o}_\ell)$ for some $\ell.$
So it suffices to consider the finite dimensional complex representations of the groups ${\mathbf{G}}(\mathfrak{o}_\ell)$.
Let $\mathbf{U}$ be the unipotent radical of a Borel subgroup $\mathbf{B}$ of ${\mathbf{G}}.$ Fix $\theta : \mathbf{U}(\mathfrak{o}_\ell) \rightarrow \mathbb{C}^{\times}$ a one dimensional representation of $\mathbf{U}(\mathfrak{o}_\ell).$
An irreducible representation $\pi$ of the group ${\mathbf{G}}(\mathfrak{o}_\ell)$ is said to admit a $\theta$-Whittaker model if the dimension of the intertwiner space ${\rm Hom}_{{\mathbf{G}}(\mathfrak{o}_\ell)} \left(\pi, {\rm Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\mathbf{G}}(\mathfrak{o}_\ell)} (\theta) \right)$ is non-trivial and $\pi$ is said to admits at most one $\theta$-Whittaker model if the intertwiner space ${\rm Hom}_{{\mathbf{G}}(\mathfrak{o}_\ell)} \left(\pi, {\rm Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\mathbf{G}}(\mathfrak{o}_\ell)} (\theta) \right)$ has dimension at most one. In view of this, the representation ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\mathbf{G}}(\mathfrak{o}_\ell)} (\theta)$ is multiplicity free if and only if every irreducible representation of ${\mathbf{G}}(\mathfrak{o}_\ell)$ admits at most one $\theta$-Whittaker model. It is natural to ask which irreducible representations of ${\mathbf{G}}(\mathfrak{o}_\ell) $ admit a $\theta$-Whittaker model? More specifically, to determine the dimension of the intertwiner space ${\rm Hom}_{{\mathbf{G}}(\mathfrak{o}_\ell)} \left(\pi, {\rm Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\mathbf{G}}(\mathfrak{o}_\ell)} (\theta)\right)$ for any irreducible representation $\pi$ of ${\mathbf{G}}(\mathfrak{o}_\ell)$.
These questions have already appeared in literature in many contexts. For example, for $\ell = 1$ and ${\mathbf{G}}(\mathbb F_q) = {\rm GL}_n(\mathbb F_q),$ one knows from the work of Gelfand and Graev~\cite{MR0148765, MR0272916} that for a non-degenerate character $\theta$ of $\mathbf{U}(\mathbb F_q),$ the induced representation $\mathrm{Ind}_{\mathrm{U}(\mathbb F_q)}^{{\rm GL}_n(\mathbb F_q)}(\theta)$ (now a days also called Gelfand-Graev representation) is multiplicity free and further every cuspidal representation of ${\rm GL}_n(\mathbb F_q)$ has a $\theta$-Whittaker model for any non-degenerate character $\theta$ of $\mathbf{U}(\mathbb F_q)$. They generalized these results further for finite Chevalley groups. These results turn out to be very important because the cuspidal representations are known to be the building blocks for all irreducible ones in these cases via the Harish-Chandra induction so in a way the essential irreducible representations are captured by the Gelfand-Graev representation. For ${\rm GL}_n(\mathbf F),$ where $\mathbf F$ is a local field, Shalika~\cite[Theorem 1.6, 2.1]{Shalika74} proved that every (smooth) irreducible representation has at most one $\theta$-Whittaker model for non-degenerate $\theta$ and generalized this to split and quasi-split groups as well. The case of local fields has several applications, for example certain integrals involving functions in the Whittaker model gives rise to the local $L$-function \cite[Theorem 2.18]{JL1970} and these local $L$-functions are then used to define global $L$-functions \cite{JL1970, Bump97}. Our focus is on the corresponding questions for the groups ${\mathbf{G}}(\mathfrak{o}_\ell),$ for ${\mathbf{G}} = {\rm GL}_n$ or $ {\rm SL}_n.$ It is to be noted that the construction of all irreducible representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ are not yet known even for $\mathrm{Char}(\mathfrak{o}) = 0$ case, see~\cite{AKOV1, AKOV2} for few results in $\mathrm{Char}(\mathfrak{o}) = 0$ case.
Let ${\mathbf{G}} = {\rm GL}_n$ or ${\rm SL}_n$ and $\mathbf{U}(\mathfrak{o}_{\ell})$ be the group of upper triangular unipotent matrices in ${\mathbf{G}}(\mathfrak{o}_{\ell}).$ A non-trivial one dimensional representation $\varphi : \mathfrak{o}_{\ell} \rightarrow \mathbb{C}^{\times}$ such that $\varphi|_{\varpi^{\ell -1} \mathfrak{o}_{\ell}} \neq 1$ is called a primitive character of $\mathfrak{o}_\ell.$ For any $n-1$ primitive characters $\varphi_1, \varphi_2, \ldots, \varphi_{n-1}$ of $\mathfrak{o}_\ell$, define a one dimensional representation $\theta_{(\varphi_1, \varphi_2, \ldots, \varphi_{n-1})} : \mathbf{U}(\mathfrak{o}_\ell) \rightarrow \mathbb{C}^{\times}$ by
\begin{equation}
\theta_{(\varphi_1,\varphi_2 , \ldots, \varphi_{n-1})}((x_{ij})) := \varphi_1 (x_{12}) \varphi_2 (x_{23}) \cdots \varphi_{n-1}(x_{(n-1)n}).
\end{equation}
Any one dimensional representation of $\mathbf{U}(\mathfrak{o}_\ell)$ of the form $\theta_{(\varphi_1, \varphi_2, \ldots, \varphi_{n-1})}$ corresponding to the primitive characters $\varphi_1, \varphi_2, \ldots, \varphi_{n-1}$ of $\mathfrak{o}_\ell$ will be called a ``non-degenerate character" of $\mathbf{U}(\mathfrak{o}_\ell).$ This is a natural generalization of the well known notion of non-degenerate character of $\mathbf{U}(\mathbb F_q)$ for $\mathbf{U}(\mathfrak{o}_\ell)$. For a non-degenerate character $\theta$ of $\mathbf{U}(\mathfrak{o}_\ell),$ we consider the ${\mathbf{G}}(\mathfrak{o}_\ell)$-representation space $\mathrm{Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\mathbf{G}}(\mathfrak{o}_\ell)}(\theta)$ and prove the following.
\begin{theorem}
\label{thm:main-theorem}
For ${\mathbf{G}} = {\rm GL}_n$ or ${\rm SL}_n,$ the ${\mathbf{G}}(\mathfrak{o}_\ell)$-representation space $\mathrm{Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\mathbf{G}}(\mathfrak{o}_\ell)}(\theta)$ is multiplicity free for every non-degenerate character $\theta$ of $\mathbf{U}(\mathfrak{o}_\ell).$
\end{theorem}
The proof of this theorem is included in Section~\ref{sec:proof-of-multiplicity-one}.
The above questions for the groups ${\rm GL}_n (\mathfrak{o}_{\ell})$ with $\ell \geq 2$ have already appeared in the work of Hill~\cite[Proposition~5.7]{MR1334228} where he proved that the space $\mathrm{Ind}_{\mathbf{U}(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\theta)$ is multiplicity free for even $\ell$ and for $\ell$ odd Hill~\cite[Proposition~5.7]{MR1334228} managed to prove that certain class of regular representations of ${\rm GL}_n(\mathfrak{o}_\ell)$ admits at most one $\theta$-Whittaker model. We briefly recall the definition of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell),$ for more details, see Section~\ref{sec:regular-representation-construction} and also \cite{MR3737836}. We will define and use the term regular representations of ${\rm SL}_n(\mathfrak{o}_\ell)$ only under the conditions $(p, 2) = (p, n) = 1$. In general the definition of regular representations for ${\rm SL}_n(\mathfrak{o}_\ell)$ is slightly complicated than the Definition~\ref{def:regular-over-field} and we will avoid the general one in this article.
Let ${\mathbf{G}}(\mathfrak{o}_\ell)$ be either ${\rm SL}_{n}(\mathfrak{o}_\ell)$ with $(p, 2) = (p, n) = 1$ or ${\rm GL}_{n}(\mathfrak{o}_\ell).$ Let $ \mathfrak{g}$ denote the Lie algebra scheme of ${\mathbf{G}}$. For every positive integer $r,$ let $\textsf{K}_{\ell}^r= \mathrm{Ker}({\mathbf{G}}(\mathfrak{o}_\ell) \mapsto {\mathbf{G}}(\mathfrak{o}/\wp^r))$ be the $r$-th principal congruence subgroup of ${\mathbf{G}}(\mathfrak{o}_\ell).$ There are ${\mathbf{G}}(\mathbf k)$-equivariant isomorphisms $\textsf{K}_\ell^r/\textsf{K}_\ell^{r+1} \cong ( \mathfrak{g}(\mathbf k), +)$ and a ${\mathbf{G}}(\mathbf k)$-equivariant isomorphism $x \mapsto \varphi_x$ between $ \mathfrak{g}(\mathbf k)$ and its Pontryagin dual $ \mathfrak{g}(\mathbf k)^\vee= \mathrm{Hom}_{\mathbb Z}( \mathfrak{g}(\mathbf k), \mathbb C^\times)$ (see \cite[Section~2.3]{MR3737836}).
An element $x \in \mathfrak{g}(\mathbf k)$ is called regular or cyclic if the characteristic polynomial of $x$ is equal to its minimal polynomial. An element $x \in \mathfrak{g}(\mathbf k)$ is called cuspidal if the characteristic polynomial of $x$ is irreducible. A one dimensional representation $ \varphi_x \in \mathfrak{g}(\mathbf k)^\vee$ is called regular (cuspidal) if $x$ is a regular (cuspidal) element.
\begin{definition}\label{def:regular-over-field} [Regular (Cuspidal) representations]
An irreducible representation $\rho$ of ${\mathbf{G}}(\mathfrak{o}_\ell)$ is called regular (cuspidal) if the orbit of its restriction to $\textsf{K}_\ell^{\ell-1} \cong \mathfrak{g}(\mathbf k)$ consists of regular (cuspidal) one dimensional representations.
\end{definition}
It is to be noted that every cuspidal representation of ${\mathbf{G}}(\mathfrak{o}_\ell)$ is regular but the converse is not true for any $n \geq 2$. A construction of all cuspidal representations of ${\rm GL}_n(\mathfrak{o}_\ell)$ was given by Aubert et al.~\cite{MR2607551}.
For odd $p$ and $\ell \geq 2$, the construction of all regular representations for ${\rm SL}_{n}(\mathfrak{o}_{\ell})$ with $p \nmid n$ and for all ${\rm GL}_{n}(\mathfrak{o}_{\ell})$ was obtained in \cite{MR3737836}. For general $p,$ an independent construction of regular representations of ${\rm GL}_{n}(\mathfrak{o}_{\ell})$ is given by Stasinski-Stevens~\cite{MR3743488}.
The regular representations for ''good" characteristic of the residue field, have also been constructed when ${\mathbf{G}}$ is a unitary, orthogonal or symplectic group, see \cite{MR3737836, Shai-regular}. The case of ${\mathbf{G}} = {\rm SL}_n$ for $p \mid n$ is not yet completely known even for $n=2,$ see~\cite{Hassain-Singla-sl2} for few results regarding this. In this article, we complete Hill's results for ${\rm GL}_n(\mathfrak{o}_\ell)$ with $\ell$ odd as well as generalise these to ${\rm SL}_n(\mathfrak{o}_\ell)$ for $p \nmid n.$ In particular, we prove the following.
\begin{theorem}
\label{main-theorem-2}
Let $\ell \geq 2$, and ${\mathbf{G}}(\mathfrak{o}_\ell)$ be either ${\rm SL}_{n}(\mathfrak{o}_\ell)$ with $(p, 2) = (p, n) = 1$ or ${\rm GL}_{n}(\mathfrak{o}_\ell).$ Then the following are true.
\begin{enumerate}
\item An irreducible representation of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ admits a $\theta$-Whittaker model for some non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$ if and only if it is regular.
\item For every non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$, every irreducible representation of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ admits at most one $\theta$-Whittaker model.
\end{enumerate}
\end{theorem}
This theorem follows from a more general result, see Theorem~\ref{thm: existence}, and its proof is included in Section~\ref{sec:proof-of-whittaker-model}. The proofs of these results depend on the description and the construction of the regular representations as given in \cite{MR3737836, MR3743488}, see also Section~\ref{sec:regular-representation-construction} where we recall briefly this construction. The proof of the second part of Theorem~\ref{main-theorem-2} is quite indirect. Usually for a proof of multiplicity one theorem for a $G$-representation $V = \mathrm{Ind}_H^G(\delta)$, where $H $ is a subgroup of $G$ and $\delta$ is a representation of $H$, one expects to use the Gelfand's trick. This trick relies on defining a good involution on $G$ which in turn fixes all the functions $f \in \mathrm{End}_G(V)$, where $f$ is considered as a function on the double cosets $H\backslash G /H$. Therefore to use this trick, one requires a good description of double cosets $\mathbf{U}(\mathfrak{o}_{\ell}) \backslash {\mathbf{G}}(\mathfrak{o}_{\ell}) /\mathbf{U}(\mathfrak{o}_{\ell}).$ In the present case such a description is known only for ${\mathbf{G}} = {\rm GL}_2(\mathfrak{o}_\ell)$ and ${\rm GL}_3(\mathfrak{o}_\ell)$ (see \cite{MR2267575}) and is non-trivial even for ${\rm GL}_3(\mathfrak{o}_\ell)$. The description of double cosets $\mathbf{U}(\mathfrak{o}_{\ell}) \backslash {\mathbf{G}}(\mathfrak{o}_{\ell}) /\mathbf{U}(\mathfrak{o}_{\ell})$ is not known in general and is believed to be quite complicated. Therefore our proof completely avoids the approach using Gelfand's trick and is more towards generalising the ideas of Hill. For the proof, we count the sum of dimensions of few specific regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$. This is achieved by carefully following the steps in the construction of the regular representations. In our case, it turns out that this sum is equal to the dimension of ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\theta).$
As a corollary of the proof of Theorem~\ref{main-theorem-2}, we obtain the following result of Hill\cite[Proposition~5.7]{MR1334228}, see the discussion preceding Theorem~\ref{thm: existence}.
\begin{corollary}
\label{cor:hill-regular-iff-whittaker}
For ${\mathbf{G}}(\mathfrak{o}_\ell) = {\rm GL}_n(\mathfrak{o}_\ell)$ and a non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$, an irreducible representation of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ admits a $\theta$-Whittaker model if and only if it is regular.
\end{corollary}
The above result is not true for ${\rm SL}_n(\mathfrak{o}_\ell)$, see Sections~\ref{sec:examples} and \ref{sec:branching-rules}. We characterize the regular representations of ${\rm SL}_{n}(\mathfrak{o}_\ell)$ for $(p, 2) = (p, n) = 1$ that admit a $\theta$-Whittaker model for every non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$, see Theorem~\ref{prop:special-regular-characterization}. From Theorem~\ref{thm:main-theorem}, we also obtain the following result.
\begin{theorem}
\label{thm:multiplicity-free-induction}
Let $\rho$ be a regular representation of ${\rm GL}_n(\mathfrak{o}_\ell)$ then the restriction of $\rho$ to ${\rm SL}_n(\mathfrak{o}_\ell)$ is multiplicity free.
\end{theorem}
A proof of this result is included in Section~\ref{sec:branching-rules}. This result is quite interesting in view of the fact that although the construction of regular representations of ${\rm GL}_n(\mathfrak{o}_\ell)$ is known in general but very less is known regarding the irreducible representations of ${\rm SL}_n(\mathfrak{o}_\ell)$ for $p \mid n$. The above result may be useful in finding more information regarding the irreducible representations of ${\rm SL}_n(\mathfrak{o}_\ell)$ for $p \mid n$.
For $(p,2) = (p,n) =1$, Theorem~\ref{thm:multiplicity-free-induction} along with the information of dimensions of regular representations of ${\rm GL}_n(\mathfrak{o}_\ell)$ and ${\rm SL}_n(\mathfrak{o}_\ell)$ as given in \cite{MR3737836}, we determine the explicit dimension of the endomorphism algebra $\mathrm{End}_{{\rm SL}_n(\mathfrak{o}_\ell)} \left( \mathrm{Res}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)} (\rho) \right)$ for any regular representation $\rho$ of ${\rm GL}_n(\mathfrak{o}_\ell)$, see Theorem~\ref{thm:branching-numbers}. We believe that the results of Theorems~\ref{thm:multiplicity-free-induction} and ~\ref{thm:branching-numbers} will prove to be helpful in describing the branching rules for the supercuspidal representations of ${\rm SL}_n(\mathbf F)$ as was achieved by Nevins~\cite{MR3008903} for ${\rm SL}_2(\mathbf F)$.
\subsection{Notations} We include here few notations that we use throughout this article. First, it is to be noted that $\mathbf k = \mathfrak{o}_1 = \mathbb F_q,$ we use these interchangeably depending on the context and the convenience. For a group $G,$ the set of all equivalence classes of irreducible representations of $G$ is denoted by $\mathrm{Irr}(G).$ The set of all one dimensional representation of an abelian group $A$ is also denoted by $\widehat{A}.$
For a normal subgroup $N$ of $G$ and an irreducible representation $\varphi$ of $N,$ the set of all $\rho \in \mathrm{Irr}(G) $ such that the restriction $\rho|_N$ of $\rho$ to $N$ has $\varphi$ as a non-trivial constituent is denoted by $\mathrm{Irr}(G \mid \varphi)$ and any element of $\mathrm{Irr}(G \mid \varphi)$ is at times called a representation of $G$ lying above $\varphi$.
\section{Regular elements of $ \mathfrak{g}(\mathfrak{o}_r)$ }
\label{sec:regular-elements}
Throughout this section, we assume that either ${\mathbf{G}} = {\rm SL}_n$ with $(p,n) = (p,2) = 1$ or ${\mathbf{G}} = {\rm GL}_n$. Let $ \mathfrak{g}$ denote the Lie algebra scheme of ${\mathbf{G}}.$ In this section, we collect some facts regarding the regular elements of $ \mathfrak{g}(\mathfrak{o}_r)$ for $r \geq 1$.
An element $x \in \mathfrak{g}(\mathfrak{o}_r) \subseteq M_n(\mathfrak{o}_r)$ is called {\it cyclic} if there exists an element $v \in \mathfrak{o}_r^{\oplus n}$ such that the set $\{v, x(v) , x^2(v), \ldots, x^{n-1} (v) \}$ is a basis of free $\mathfrak{o}_r$-module $\mathfrak{o}_r^{\oplus n}.$ The vector $v$ is called a cyclic vector of $x.$
The following lemma relates the cyclic and regular elements of $M_n(\mathbb F_q)$.
\begin{lemma}
\label{lem:regular-minimal-characteristic}
An element $x \in M_n(\mathfrak{o}_1) = M_n(\mathbb F_q)$ is regular if and only $x$ is cyclic.
\end{lemma}
\begin{proof}
This follows from the definition of regular $x \in M_n(\mathfrak{o}_1)$ (a proof appears in for example, Suprunenko-Tyshkevich~\cite[Theorem~5]{MR0201472}).
\end{proof}
Next, we recall the definition of regular elements of $ \mathfrak{g}(\mathfrak{o}_r)$.
\begin{definition} [Regular element of $ \mathfrak{g}(\mathfrak{o}_r)$]
An element $x \in \mathfrak{g}(\mathfrak{o}_r) \subseteq M_n(\mathfrak{o}_r)$ is called regular if and only if it is cyclic.
\end{definition}
For any $i \leq r,$ there exists a natural projection $\rho_{r, i}: \mathfrak{o}_{r} \rightarrow \mathfrak{o}_i$ which is a ring homomorphism. By applying entry wise, we obtain a projection $\rho_{r,i}: \mathfrak{g}(\mathfrak{o}_{r}) \rightarrow \mathfrak{g}(\mathfrak{o}_i).$ For any $x \in \mathfrak{g}(\mathfrak{o}_{r}),$ the image $\rho_{r,1}(x)$ is denoted by $\bar{x}.$
\begin{lemma}
\label{lem:hill-results-regular} Let $x \in \mathfrak{g}(\mathfrak{o}_r) \subseteq M_n(\mathfrak{o}_r)$ and let $r \geq 1.$ The following are equivalent.
\begin{enumerate}
\item The element $x \in \mathfrak{g}(\mathfrak{o}_r)$ is regular.
\item The projection of $x$ to $ \mathfrak{g}(\mathfrak{o}_i)$ for every $1 \leq i \leq r$ is regular.
\item The centralizer $C_{M_n(\mathfrak{o}_r)}(x) = \{ y \in M_n(\mathfrak{o}_r) \mid xy = yx \}$ of $x$ in $M_n(\mathfrak{o}_r)$ is abelian.
\item $C_{M_n(\mathfrak{o}_r)}(x) = \mathfrak{o}_r[x],$ where $\mathfrak{o}_r[x]$ is a $\mathfrak{o}_r$-subalgebra of $M_n(\mathfrak{o}_r)$ generated by the set $\{ I, x, x^2, \ldots, x^{n-1} \}.$
\end{enumerate}
\end{lemma}
\begin{proof}
A proof of this follows from Hill~\cite[Theorem~3.6, Corollary~3.7]{MR1334228} along with the fact that $x \in \mathfrak{g}(\mathfrak{o}_r)$ is regular if and only if $x$ is regular when viewed as an element of $M_n(\mathfrak{o}_r)$.
\end{proof}
Recall $\mathfrak{o}_1 = \mathbb F_q.$ We use $d_{ \mathfrak{g}}$ to denote the dimension of the $\mathbb F_q$-vector space $ \mathfrak{g}(\mathfrak{o}_1).$ For $\bar{x} \in \mathrm{g}(\mathfrak{o}_1),$ the dimension of the centralizer algebra $C_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}) = \{ \bar{y} \in \mathfrak{g}(\mathfrak{o}_1) \mid \bar{x}\bar{y} = \bar{y}\bar{x} \}$ as $\mathfrak{o}_1$-vector space is denoted by $d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}).$ Note that by Lemmas~\ref{lem:hill-results-regular} and \ref{lem:regular-minimal-characteristic}, for every regular $\bar{x} \in M_n(\mathfrak{o}_1)$ and $\bar{y} \in \sl_n(\mathfrak{o}_1)$ we have $d_{M_n(\mathfrak{o}_1)} (\bar{x}) = n$ and $d_{\sl_n(\mathfrak{o}_1)} (\bar{y}) = n-1.$
\begin{lemma}
Let $x \in \mathfrak{g}(\mathfrak{o}_r)$ be a regular matrix and $C_{ \mathfrak{g}(\mathfrak{o}_r)}(x) = \{ y \in \mathfrak{g}(\mathfrak{o}_r) \mid xy = yx \}$ be the centralizer of $x$ in $ \mathfrak{g}(\mathfrak{o}_r).$ Then $|C_{ \mathfrak{g}(\mathfrak{o}_r)}(x)| = q^{ (d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}) ) r}.$
\end{lemma}
\begin{proof} Both for $ \mathfrak{g} = M_{n}(\mathfrak{o}_{r})$ and $ \mathfrak{g} = \mathrm{sl}_n(\mathfrak{o}_r)$ and regular $x$, the result follows by Lemma~\ref{lem:hill-results-regular}.
\end{proof}
\begin{definition}($a$-regular element) For $a \in \mathfrak{o}_r^\times$, let $x \in M_n(\mathfrak{o}_r)$ be of the following form,
\begin{equation} \label{regular element}
x = \left(
\begin{matrix}
0 & 0 & 0 & \cdots & 0 & x_{1} \\
a & 0 & 0 & \cdots & 0 & x_{2} \\
0 & 1 & 0 & \cdots & 0 & x_{3} \\
\vdots & \vdots & \vdots & & \vdots & \vdots \\
0 & 0 & 0 & \cdots & 1 & x_{n}
\end{matrix}
\right)
\end{equation}
where $x_{i} \in \mathfrak{o}_r$ for $1 \leq i \leq n.$ Then $x \in M_n(\mathfrak{o}_r)$ is regular and is called an $a$-regular element of $M_n(\mathfrak{o}_r)$.
\end{definition}
\begin{lemma}
\label{lem:about-regular-elements} The following are true for the set of regular elements of $ \mathfrak{g}(\mathfrak{o}_r).$
\begin{enumerate}
\item The set of regular elements in $ \mathfrak{g}(\mathfrak{o}_r)$ is invariant under the conjugation action of ${\mathbf{G}}(\mathfrak{o}_r).$
\item Every conjugacy class of regular elements of $ \mathfrak{g}(\mathfrak{o}_r)$ contains an $a$-regular element for some $a \in \mathfrak{o}_r^\times$.
\item For any $a \in \mathfrak{o}_r^\times$, the number of $a$-regular conjugacy classes of $ \mathfrak{g}(\mathfrak{o}_r)$ is $q^{ (d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})) r},$ where $x \in \mathfrak{g}(\mathfrak{o}_r)$ is any regular matrix.
\end{enumerate}
\end{lemma}
\begin{proof} The proof of $(1)$ follows by Lemma~\ref{lem:hill-results-regular}. For ${\rm GL}_n(\mathfrak{o}_r)$ every $a$-regular matrix is conjugate to a $1$-regular element via conjugating with a diagonal matrix $(a, 1, \ldots, 1)$. The rest of the proof of (2) for ${\mathbf{G}} = {\rm GL}_n$ and $ \mathfrak{g} = M_n$ follows from Mcdonald~\cite[p.417-419]{MR769104} (see also Hill~\cite[Remark~3.10]{MR1334228} ). For ${\mathbf{G}} = {\rm SL}_n$, $ \mathfrak{g} = \sl_n$ we first note that by above for $x_1 \in \sl_n(\mathfrak{o}_r),$ there exists $g \in {\rm GL}_n(\mathfrak{o}_r)$ such that $gx_1 g^{-1}$ is $1$-regular. Consider the diagonal matrix $z = (\frac{1}{\det(g)} , 1, \ldots, 1)\in {\rm GL}_n(\mathfrak{o}_r)$ and $g' = zg \in {\rm SL}_n(\mathfrak{o}_r)$ then $g' x_1 (g')^{-1} $ is $\det(g)$-regular. This in particular implies that every similarity class in $\sl_n(\mathfrak{o}_r)$ contains an $a$-regular element for some $a \in \mathfrak{o}_r^\times$. From this (2) and (3) for $ \mathfrak{g} = \sl_n$ follow easily from the corresponding results for $ \mathfrak{g} = M_n.$
It is to be noted that a similarity class in $ \mathfrak{g}(\mathfrak{o}_r)$ may contain both an $a$-regular and a $b$-regular element for $a \neq b$. However for a fixed $a \in \mathfrak{o}_r^\times$, each similarity class contains at most one $a$-regular element.
Last we note that the similarity class of an $a$-regular element is determined by its characteristic polynomial so $(3)$ follows by direct computations from Lemma~\ref{lem:hill-results-regular}.
\end{proof}
\begin{lemma}
\label{lem:stab-unipotent-intersection}
Let $x \in M_n(\mathfrak{o}_r)$ be an $a$-regular matrix. Then $ C_{ \mathfrak{g}(\mathfrak{o}_r)} (x) \cap \mathbf{U}(\mathfrak{o}_r) = C_{{\mathbf{G}}(\mathfrak{o}_r)} (x) \cap \mathbf{U}(\mathfrak{o}_r) = \{ I_n \}$
\end{lemma}
\begin{proof}
We note that for $a$-regular matrix $x,$ for every $1 \leq i \leq n-1,$ there exists $1 \leq j, k \leq n$ such that $j > k$ and $(j,k)^{th}$-entry of $x^i \in \mathfrak{o}_r^\times$. Now result follows by the fact that $ C_{{\mathbf{G}}(\mathfrak{o}_r)} (x) \subseteq C_{ \mathfrak{g}(\mathfrak{o}_r)} (x) \subseteq C_{M_n(\mathfrak{o}_r)} (x) ,$ by Lemma~\ref{lem:hill-results-regular} the set $C_{ \mathfrak{g}(\mathfrak{o}_r)} (x)$ is generated by $\{ I, x, x^2, \ldots, x^{n-1} \}$ as $\mathfrak{o}_r$-algebra and for every $y \in \mathbf{U}(\mathfrak{o}_r),$ the $(j,k)^{th}$ entry of $y$ is zero for $j > k.$
\end{proof}
\begin{lemma}
\label{lem:uppertriangular-regular}
Let $x = (x_{ij}) \in M_n(\mathfrak{o}_r)$ be such that $x_{21} = a \in \mathfrak{o}_r^\times $ and $x_{ij} = 1$ for all $i, j$ such that $i \geq 3$, $i = j + 1$ and $x_{ij} = 0$ for all $i, j$ such that $i-j \geq 2.$ Then $x$ is regular.
\end{lemma}
\begin{proof}
By Lemma~\ref{lem:hill-results-regular}, it is enough to prove that $\bar{x}$ is regular for all above $x.$ For that it is easy to see that $v = (1, 0, 0, \ldots, 0) \in \mathfrak{o}_1^{\oplus n}$ satisfies that $\{ v, \bar{x}v, \bar{x}^2v, \cdots, \bar{x}^{n-1}v \}$ generates the n-dimensional vectors space $\mathfrak{o}_1^{\oplus n}.$
\end{proof}
\section{Construction of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$}
\label{sec:regular-representation-construction}
In this section we follow \cite{MR3737836, MR3743488} to give a brief outline of the construction of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$, where either ${\mathbf{G}} = {\rm SL}_n$ with $(p,n) = (p,2) = 1$ or ${\mathbf{G}} = {\rm GL}_n$.
For $ i \leq \ell$ and the natural projection maps $\rho_{\ell,i}: {\mathbf{G}}(\mathfrak{o}_{\ell}) \rightarrow {\mathbf{G}}(\mathfrak{o}_i),$ as defined in Section~\ref{sec:regular-elements}, let $\textsf{K}_{\ell}^{i} = \ker (\rho_{\ell,i}) $ be the $i$-th congruence subgroups of ${\mathbf{G}}(\mathfrak{o}_\ell).$ It is easy to note that for $i \geq \ell/2,$ the group $\textsf{K}^i_\ell $ is isomorphic to the abelian additive subgroup $ \mathfrak{g}(\mathfrak{o}_{\ell-i})$ of $M_n(\mathfrak{o}_{\ell-i}).$ Let $\varphi: \mathfrak{o}_\ell \rightarrow \mathbb C^\times$ be a fixed primitive one dimensional representation of $\mathfrak{o}_\ell$.
For any $i \leq \ell/2$ and $x \in \mathfrak{g}(\mathfrak{o}_i),$ let $\hat{x} \in \mathfrak{g}(\mathfrak{o}_\ell)$ be an arbitrary lift of $x$ satisfying $\rho_{\ell,i}(\hat{x}) = x.$ Define $\varphi_x: I + \varpi^{\ell-i} \mathfrak{g}(\mathfrak{o}_\ell) \rightarrow \mathbb C^\times$ by
\[
\varphi_x(I + \varpi^{\ell-i} y) = \varphi(\varpi^{\ell-i}\mathrm{tr}(\hat{x}y)),
\]
for all $I + \varpi^{\ell-i} y \in \textsf{K}^{\ell-i}_\ell.$ Then $\varphi_x$ is easily seen to be a well defined one dimensional representation of $\textsf{K}^{\ell-i}_\ell .$ Further it is easy to see that for $i \geq \ell/2$ the following duality of abelian group $\textsf{K}^{i}_\ell$ and $ \mathfrak{g}(\mathfrak{o}_{\ell-i})$ holds.
\begin{equation}
\label{eq:duality}
\mathfrak{g}(\mathfrak{o}_{\ell-i}) \cong \widehat {\textsf{K}^{i}_\ell}\,\,; x \mapsto \varphi_x\,\, \mathrm{where}, \,\, \varphi_x(y) = \varphi(\varpi^{\ell-i}\mathrm{tr}(\hat{x}y)) \,\, \forall \,\, y \in \textsf{K}^{i}_\ell.
\end{equation}
We say a one dimensional representation $\varphi_x \in \widehat{\textsf{K}^{i} _\ell}$ for $i \geq \ell/2$ is regular if and only if $x \in M_n(\mathfrak{o}_{\ell-i})$ is a regular matrix. By Lemma~\ref{lem:hill-results-regular}, for $i \geq \ell/2$ the representation $\varphi_x \in \widehat{\textsf{K}^{i} _\ell}$ is regular if and only if $\varphi_x|_{ \textsf{K}^{\ell-1} _\ell}$ is regular. Recall an irreducible representation $\rho$ of ${\mathbf{G}}(\mathfrak{o}_\ell)$ is called regular if the orbit of its restriction to $\textsf{K}_\ell^{\ell-1} $ consists of one dimensional representations $\varphi_x$ for regular $x$.
Now we summarize the steps of construction of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell) $.
For ${\mathbf{G}}(\mathfrak{o}_\ell) $ with odd $p$ we will outline the method as given in \cite{MR3737836} and we follow \cite{MR3743488} for ${\rm GL}_n(\mathfrak{o}_\ell)$ with $p=2.$ It might be useful to look at Figures~\ref{fig:ell.even} and \ref{fig:ell.odd} while going through the steps of construction. For more details on this see \cite{MR3737836, MR3743488}. The construction in both cases is based on the tools of abstract Clifford theory and orbits. See \cite[Chapter~6]{MR0460423} for general results and their proofs regarding Clifford theory and \cite[Theorem~2.1]{Singla-Pooja} for the precise statements related to Clifford theory that are used in this article. The main point of the construction is that the case of ${\mathbf{G}}(\mathfrak{o}_\ell)$ where either ${\mathbf{G}} = {\rm SL}_n$ with $(p,n) = (p,2) = 1$ or ${\mathbf{G}} = {\rm GL}_n$ are good in the sense that for any regular one dimensional representation $\varphi_x$ of $\widehat{\textsf{K}_\ell^{\lceil \ell/2\rceil}}$, the inertia group $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x) = \{ g \in {\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x^g \cong \varphi_x \} $ is "nice". In all these cases, the required inertia group although is non-abelian but still is a product of an abelian group with a congruence subgroup and therefore all the representations of the required inertia group lying above $\varphi_x$ can be constructed. This combined with the Clifford theory gives the construction of all regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$.
\subsection{Construction of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ for $\ell = 2m$} Let $\varphi_x \in \widehat{\textsf{K}_\ell^m}$ be a regular one dimensional representation of $\textsf{K}_{\ell}^{m}$ for $x \in \mathfrak{g}(\mathfrak{o}_m).$ Then the following are true.
\medskip
\noindent {\bf E.1} Let $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x) = \{ g \in {\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x^g \cong \varphi_x \} $ be the inertia group of $\varphi_x$ in ${\mathbf{G}}(\mathfrak{o}_\ell).$
\begin{enumerate}
\item Then $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x) = \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x}) \textsf{K}_\ell^{m} ,$ where $\tilde{x} \in \mathfrak{o}_\ell$ is any lift of $x$ to $ \mathfrak{g}(\mathfrak{o}_\ell).$
\item Although the inertia group $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x)$ is not abelian but the quotient $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x)/\textsf{K}_\ell^m \cong \mathrm{C}_{{\mathbf{G}}({\cO}_m)}(x)$ is abelian by Lemma~\ref{lem:hill-results-regular}.
\end{enumerate}
\medskip
\noindent {\bf E.2} Let $\delta$ be a one dimensional representation of $\mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})$ extending $\varphi_x|_{\textsf{K}^{m}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})}$. The existence of this $\delta$ follows easily because the group $\mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x}) $ is abelian. Define $\widetilde{\varphi_x}: I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x) \rightarrow \mathbb {C}^\times$ by $\widetilde{\varphi_x}(ab) = \delta(a) \varphi_x(b)$ for all $a \in \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x}) $ and $b \in \textsf{K}_\ell^{m} $. It can be easily shown that $\widetilde{\varphi_x}$ is a well defined one dimensional representation of $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x)$ that extends the representation $\varphi_x$.
\medskip
\noindent {\bf E.3} Let $\rho \in \mathrm{Irr} \left( {\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x \right)$ be a regular representation of ${\mathbf{G}}(\mathfrak{o}_\ell),$ then there exists an extension $\widetilde{\varphi_x}$ of $\varphi_x$ to $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x)$ such that $\rho \cong \mathrm{Ind}_{I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x)}^{{\mathbf{G}}(\mathfrak{o}_\ell)} (\widetilde{\varphi_x}) .$ This follows by ${\bf E.1}$, ${\bf E.2}$ combined with Clifford theory.
\medskip
\noindent {\bf E.4} We have $\left| \mathrm{Irr} ({\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x) \right| = \left|\mathrm{C}_{{\mathbf{G}}({\cO}_m)}(x) \right|.$
\medskip
\noindent {\bf E.5}
Every $\rho \in \mathrm{Irr} \left( {\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x \right)$ has dimension $\dfrac{|{\mathbf{G}}(\mathfrak{o}_\ell)|}{|\mathrm{C}_{{\mathbf{G}}({\cO}_m)}(x)||\textsf{K}_\ell^m|}.$
\begin{figure}[htb!]
\centering
\begin{minipage}{0.45\textwidth}
\centering
\label{fig:ell.even}
\begin{displaymath}
\xymatrix{
& {{\mathbf{G}}({\cO}_\ell)}\ar@{-}[d] & \\
& {I_{{\mathbf{G}}({\cO}_\ell)}(\varphi_x)}\ar@{-}[dr] \ar@{-}[dl] & \\
{\textsf{K}^m_\ell}\ar@{-}[dr] & & \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})\ar@{-}[dl] \\
& {\textsf{K}^{m}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})}\ar@{-}[d] & \\
& {\{1\}}& }
\end{displaymath}
\caption{Even case $\ell=2m$}
\end{minipage}
\begin{minipage}{.5\textwidth}
\centering
\label{fig:ell.odd}
\begin{displaymath}
\xymatrix{
& {{\mathbf{G}}({\cO}_\ell)}\ar@{-}[d] & & \\
& {I_{{\mathbf{G}}({\cO}_\ell)}(\sigma)}\ar@{-}[dd] \ar@{-}[dl] & & \\
{H^m_\ell}\ar@{-}[d]^{q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}({\bar{x}}) }{2}}} & & & \\
{J_x}\ar@{-}[d]^{q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}({\bar{x}}) }{2}}} & \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x}) \ar@{-}[dl] & & \\
R_{{\bar{x}}} \ar@{-}[d] & & & \\
{\textsf{K}^{m+1}_\ell}\ar@{-}[dr] & & &\\
& {\textsf{K}^{m+1}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})}\ar@{-}[d] \ar@{-}[uuu] & & \\
& {\{1\}}& &
}
\end{displaymath}
\caption{Odd case $\ell=2m+1$}
\end{minipage}%
\end{figure}
\subsection{Construction of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ for $\ell = 2m + 1$} The construction for this case is involved as compared to $\ell = 2m$ case. So here we highlight what is important for us and refer the reader to \cite{MR3737836, MR3743488} for more details. Let $\varphi_x \in \widehat{\textsf{K}_\ell^{m+1}}$ be a regular one dimensional representation of $\textsf{K}_{\ell}^{m+1}$ for $x \in \mathfrak{g}(\mathfrak{o}_{\ell-m}).$ Let $\tilde{x} \in \mathfrak{g}(\mathfrak{o}_\ell)$ be a lift of $x$ to $ \mathfrak{g}(\mathfrak{o}_\ell).$ For $p \neq 2,$ either ${\mathbf{G}} = {\rm SL}_n$ with $p \nmid n$ or ${\mathbf{G}} = {\rm GL}_n.$ For $p=2,$ we consider ${\mathbf{G}} = {\rm GL}_n.$
\medskip
\noindent { \bf O.1} Define groups $H^m_\ell$ and $R_{{\bar{x}}}$ as follows.
\[
H^i_\ell = \begin{cases}\textsf{K}_\ell^i, & p \neq 2 \\
\left(\textsf{K}^{1}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})\right) \textsf{K}_\ell^{i}, & p =2.
\end{cases}
\]
\[
R_{{\bar{x}}} = \begin{cases} \left(\textsf{K}^{m}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})\right) \textsf{K}_\ell^{m+1}, & p \neq 2 \\
\left(\textsf{K}^{1}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x})\right) \textsf{K}_\ell^{m+1}, & p =2.
\end{cases}
\]
\medskip
\noindent { \bf O.2}
The following are true for groups $H^i_\ell$ and $R_{{\bar{x}}}.$
\begin{enumerate}
\item The group $R_{{\bar{x}}}$ is a normal subgroup of $H_{\ell}^{m}.$
\item The quotient group $R_{{\bar{x}}}/\textsf{K}_\ell^{m+1} $ is abelian.
\item The one dimensional representation $\varphi_x \in \widehat{K^{m+1}_\ell }$ extends to $R_{{\bar{x}}}.$
\end{enumerate}
\medskip
\noindent { \bf O.3} The extensions of $\varphi_x$ to $R_{{\bar{x}}}$ satisfy the following:
\begin{enumerate}
\item each extension of $\varphi_x$ to $R_{\bar{x}}$ is stable under $\textsf{K}^{m}_\ell.$
\item Each extension of $\varphi_x$ to $R_{\bar{x}}$ determines a non-degenerate bilinear form on $H^{\ell}_m/R_{\bar{x}}$. Then there exists a ``nice'' maximal isotropic space $J_x$ with respect to this bilinear form such that, parallel to the method of construction of ``Heisenberg" groups, every extension of $\varphi_x$ to $R_{\bar{x}}$ determines a unique irreducible representation of $H^m_\ell$ of dimension $q^{\frac{d_{ \mathfrak{g}} - d_{ \mathfrak{g}(\mathfrak{o}_1) }({\bar{x}}) }{2}},$ see \cite[Section~3]{MR3737836} for $p \neq 2$ and \cite[Section~3]{MR3743488} for $p=2$ for more details.
\item Let $\widetilde{\varphi_x}$ be an extension of $\varphi_x$ to $R_{{\bar{x}}}$ and $\sigma \in \mathrm{Irr}(H_\ell^m \mid \varphi_x) $ be unique irreducible representation determined by $\widetilde{\varphi_x}.$ Then,
\[\sigma|_{R_{{\bar{x}}}} = \underbrace{\widetilde{\varphi_x} + \cdots + \widetilde{\varphi_x} }_{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}({\bar{x}}) }{2}-\mathrm{times}}.
\]
\end{enumerate}
\medskip
\noindent { \bf O.4} Let $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\sigma) = \{ g \in {\mathbf{G}}(\mathfrak{o}_\ell) \mid \sigma^g \cong \sigma \}$ be the inertia groups of $\sigma \in \mathrm{Irr}(H_\ell^m \mid \varphi_x).$ Then the following are true:
\begin{enumerate}
\item $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\sigma) = I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x) = \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x}) \textsf{K}_\ell^m,$ where $\tilde{x} \in \mathfrak{g}(\mathfrak{o}_\ell)$ is any lift of $x$ to $ \mathfrak{g}(\mathfrak{o}_\ell).$ This follows by the definition of $R_{\bar{x}}$ along with {\bf O.2}.
\item Every $\sigma \in \mathrm{Irr}(H_\ell^m \mid \varphi_x)$ extends to the inertia group $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\sigma) .$ In particular, every such extension induces irreducibly to ${\mathbf{G}}(\mathfrak{o}_\ell)$ and gives rise to a regular representation of ${\mathbf{G}}(\mathfrak{o}_\ell).$ This is slightly technical part and depends on the structure of groups $R_{\bar{x}}$, $J_x$, $H_\ell^m$ and $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\sigma)$.
\end{enumerate}
\medskip
\noindent { \bf O.5} Combining above all, there exists a bijection $
{\rm Irr}\left( {\mathbf{G}}(\mathfrak{o}_\ell) \mid \sigma \right) \leftrightarrow \mathrm{C}_{{\mathbf{G}}(\mathfrak{o}_{m})}(x).$
\medskip
\noindent { \bf O.6}
Every $\rho \in \mathrm{Irr} \left( {\mathbf{G}}(\mathfrak{o}_\ell) \mid \sigma \right)$ for $\sigma \in \mathrm{Irr}(\textsf{K}_\ell^m \mid \varphi_x)$ has dimension $q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}({\bar{x}}) }{2}}\frac{|{\mathbf{G}}(\mathfrak{o}_\ell)|}{|\mathrm{C}_{{\mathbf{G}}(\mathfrak{o}_{m})}(x)| |\textsf{K}_\ell^m|} .$
\medskip
\noindent { \bf O.7} Altogether, we have the following:
\begin{enumerate}
\item $|\mathrm{Irr} \left( {\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x \right)| = q^{d_{ \mathfrak{g}(\mathfrak{o}_1)}({\bar{x}})} |\mathrm{C}_{{\mathbf{G}}(\mathfrak{o}_{m})}(x)| .$
\item Every $\rho \in \mathrm{Irr} \left( {\mathbf{G}}(\mathfrak{o}_\ell) \mid \varphi_x \right)$ has dimension $q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}({\bar{x}}) }{2}}\frac{|{\mathbf{G}}(\mathfrak{o}_\ell)|}{|\mathrm{C}_{{\mathbf{G}}(\mathfrak{o}_{m})}(x)| |\textsf{K}_\ell^m|} .$
\end{enumerate}
\section{Uniqueness of $\theta$-Whittaker models}
In this section first we prove Theorem~\ref{main-theorem-2} and then use this to prove Theorem~\ref{thm:main-theorem}. Let $\varphi$ be a fixed primitive character of $\mathfrak{o}_\ell.$ For $a \in \mathfrak{o}_{\ell}^\times$, define $\varphi_{a} : \mathfrak{o}_\ell \rightarrow \mathbb{C}^{\times}$ by $\varphi_a(x) = \varphi(ax)$. Then it is easy to see that $\varphi_a$ is a primitive character of $\mathfrak{o}_\ell$ and every primitive character of $\mathfrak{o}_\ell$ is of the form $\varphi_a$ for some $a \in \mathfrak{o}_\ell^\times$.
For $a \in \mathfrak{o}_{\ell}^\times$, define $\theta_{a} : \mathbf{U}(\mathfrak{o}_\ell) \rightarrow \mathbb{C}^{\times}$ by
\begin{equation}
\theta_{a}((x_{ij})) := \varphi (a x_{12} + x_{23} + \cdots + x_{(n-1)n}).
\end{equation}
The subgroup of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ consisting of all diagonal matrices (split torus) acts on $\mathbf{U}(\mathfrak{o}_\ell)$ by conjugation. It is to be noted that every one dimensional non-degenerate representation of $\mathbf{U}(\mathfrak{o}_\ell)$ is equivalent to $\theta_a$ for some $a \in \mathfrak{o}_\ell^\times$ under the conjugation action of the split torus. Therefore without loss of generality, it is enough to prove Theorems~\ref{thm:main-theorem} and \ref{main-theorem-2} for $\theta = \theta_a$ for all $a \in \mathfrak{o}_\ell^\times$ and for any fixed primitive character $\varphi$ of $\mathfrak{o}_\ell$.
\subsection{Proof of Theorem~\ref{main-theorem-2}}
\label{sec:proof-of-whittaker-model}
Throughout this section we assume that either ${\mathbf{G}} = {\rm SL}_n$ with $(p,n) = (p,2) = 1$ or ${\mathbf{G}} = {\rm GL}_n$. For $1 \leq k < \ell,$ let $\mathbf{U}(\varpi^{k} \mathfrak{o}_{\ell})$ be the group of unipotent upper triangular matrices for which the entries in the strictly upper triangular part belong to $\varpi^{k}\mathfrak{o}_{\ell}.$ We assume that $m$ is such that $\ell = 2m$ for even $\ell$ and $\ell = 2m+1$ for odd $\ell$. We also assume that primitive character $\varphi$ of $\mathfrak{o}_\ell$ used in the definition of the construction of regular representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ in Section~\ref{sec:regular-representation-construction} and in the definition of $\theta_a$ is the same one. The following definition plays an important role in what follows.
\begin{definition}\label{def:regular-representation}($\alpha$-regular representation of ${\mathbf{G}}(\mathfrak{o}_\ell)$) A regular representation $\rho$ of $\mathrm{G}(\mathfrak{o}_\ell)$ is called $\alpha$-regular for $\alpha \in \mathfrak{o}_{\lfloor \frac{\ell}{2} \rfloor}$ if there exists an $\alpha$-regular element $x \in \mathfrak{g}(\mathfrak{o}_{\lfloor \frac{\ell}{2} \rfloor})$ such that $(\rho|_{K_\ell^{\lceil \frac{\ell}{2}\rceil}}, \varphi_x) \neq 0$ for $\varphi_x \in \widehat{K_\ell^{\lceil \frac{\ell}{2}\rceil}}$.
\end{definition}
We proceed to state and prove the main theorem of this article. As mentioned above, every non-degenerate character of $\mathbf U(\mathfrak{o}_\ell)$ is conjugate to $\theta_a$ for some $a \in \mathfrak{o}_\ell^\times$ and therefore Theorem~\ref{main-theorem-2} directly follows from Theorem~\ref{thm: existence}. For ${\mathbf{G}}(\mathfrak{o}_\ell) = {\rm GL}_n(\mathfrak{o}_\ell)$ and for any $a \in \mathfrak{o}_\ell^\times$, every regular representation is $\rho_{\ell, m}(a)$ regular. Therefore Corollary~\ref{cor:hill-regular-iff-whittaker} also follows from Theorem~\ref{thm: existence}.
\begin{theorem} \label{thm: existence}
\begin{enumerate}
\item An irreducible representation $\pi$ of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ for $\ell \geq 2$ admits a $\theta_a$-Whittaker model if and only if it is $\rho_{\ell, m }(a)$-regular.
\item An irreducible representation of ${\mathbf{G}}(\mathfrak{o}_\ell)$ for $\ell \geq 2$ admits at most one $\theta_a$-Whittaker model for every $a \in \mathfrak{o}_\ell^\times$.
\end{enumerate}
\end{theorem}
\begin{proof}
First, we prove that for $a \in \mathfrak{o}_\ell^\times$, every $\rho_{\ell, m}(a)$-regular representation admits a $\theta_a$-Whittaker model. Define $N_\ell^m$ as follows.
\[
N_\ell^m = \begin{cases} K_\ell^m, & \ell = 2m \\
H_\ell^m, & \ell = 2m+1,
\end{cases}
\]
where $K_\ell^m$ and $H_\ell^m$ as defined in the Section~\ref{sec:regular-representation-construction}.
Since $\pi$ is a $\rho_{\ell, m }(a)$-regular representation of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ by construction of $\pi$ (see {\bf E.3} and { \bf O.4 }) there exists an irreducible representation $\sigma$ of $N_{\ell}^{m}$ which lies over a regular one dimensional representation $\varphi_{x}$ of $\mathsf{K}_{\ell}^{\lceil \frac{\ell}{2}\rceil}$ for some $\rho_{\ell, m}(a)$-regular element $x$ and $\sigma$ extends to the inertia group $I_{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\sigma)$ written as $\widetilde{\sigma}$ with the property that
\[
\pi \cong {\rm Ind}_{I_{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\sigma)}^{{\mathbf{G}}(\mathfrak{o}_{\ell})} \widetilde{\sigma}.
\]
Write $\mathbf I(\sigma)$ for the inertia group $I_{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\sigma).$ By using Frobenius reciprocity and Mackey theory we get
\[
{\rm Hom}_{{\mathbf{G}}(\mathfrak{o}_{\ell})} \left(\pi, {\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\theta_a) \right) = \bigoplus_{g \in \mathbf I(\sigma) \backslash {\mathbf{G}}(\mathfrak{o}_{\ell}) / \mathbf{U}(\mathfrak{o}_{\ell})} {\rm Hom}_{\mathbf{U}(\mathfrak{o}_{\ell}) \cap (\mathbf I(\sigma))^{g}} \left(\widetilde{\sigma}^{g}, \theta_a \right)
\]
To prove ${\rm Hom}_{{\mathbf{G}}(\mathfrak{o}_{\ell})} \left(\pi, {\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\theta_a) \right) \neq 0$ it is enough to prove that ${\rm Hom}_{\mathbf{U}(\mathfrak{o}_{\ell}) \cap \mathbf I(\sigma)} \left(\widetilde{\sigma}, \theta_a \right) \neq 0$ and this we prove in next few lemmas.
\begin{lemma}
\label{lem:intersection}
For $\ell = 2m$, let $x \in \mathfrak{g}(\mathfrak{o}_{m})$ be a $\rho_{\ell, m}(a)$-regular element. The following are true.
\begin{enumerate}
\item The intersection $\mathbf{U}(\mathfrak{o}_{\ell}) \cap I_{{\mathbf{G}}(\mathfrak{o}_{\ell})} ( \sigma) = \mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}).$
\item We have ${\rm Hom}_{ \mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})} \left( \widetilde{\sigma}, \theta_{a } \right) \neq 0.$
\end{enumerate}
\end{lemma}
\begin{proof}
For $\ell = 2m$, we have $N_\ell^m = \textsf{K}_\ell^m$ and therefore $\sigma = \varphi_x$ in this case. Then $I_{{\mathbf{G}}(\mathfrak{o}_\ell)}(\varphi_x) = \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\tilde{x}) \textsf{K}_\ell^m$ for any lift $\tilde{x}$ of $x$ by {\bf E.1}. it is clear that $\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}) \subseteq \mathbf{U}(\mathfrak{o}_{\ell}) \cap I_{{\mathbf{G}}(\mathfrak{o}_{\ell})} ( \varphi_{x}) ,$ so we proceed to prove the other side inclusion.
The element $x$ is $\rho_{\ell, m }(a)$-regular. Therefore by Lemma~\ref{lem:stab-unipotent-intersection}, we have $\mathbf{U}(\mathfrak{o}_\ell) \cap C_{{\mathbf{G}}(\mathfrak{o}_\ell)} (\tilde{x}) = \{ 1 \} .$ This implies, any $v \in \mathbf{U}(\mathfrak{o}_{\ell}) \cap I_{{\mathbf{G}}(\mathfrak{o}_{\ell})} ( \varphi_{x})$ satisfies $v \in \textsf{K}_\ell^m$ This along with $v \in \mathbf{U}(\mathfrak{o}_\ell)$ gives $v \in \mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}).$ Therefore (1) follows.
The proof of (2) amounts to say that the restriction of $\widetilde{\varphi}_{x}$ to $\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})$ is same as the restriction of $\theta_a$ to $\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}).$
Which is clear since $x$ is $\rho_{\ell, m}$-regular and therefore $\widetilde{\varphi_x}(y) = \theta_a(y)$ for every $y \in \mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}).$
This proves part (2).
\end{proof}
\begin{lemma}
\label{lem:decomposition-of-sigma}
For $\ell = 2m +1 ,$ consider the set $\r = \mathrm{Irr}\left( \mathbf{U}(\varpi^m \mathfrak{o}_\ell ) \mid \theta_a|_{\mathbf{U}(\varpi^{m+1} \mathfrak{o}_\ell ) } \right).$ The following are true.
\begin{enumerate}
\item The set $\r$ consists of one dimensional representations of $\mathbf{U}(\varpi^m \mathfrak{o}_\ell ).$
\item If $x \in \mathfrak{g}(\mathfrak{o}_m)$ is a $\rho_{\ell, m}(a)$-regular element and $\sigma \in \mathrm{Irr} \left( H_\ell^m \mid \varphi_x \right).$ Then,
\[
\sigma|_{\mathbf{U}(\varpi^m \mathfrak{o}_\ell )} \cong \underset{\chi \,\, \in \,\, \r}{\oplus} \,\,\chi.
\]
\end{enumerate}
\end{lemma}
\begin{proof} The one dimensional representation $\theta_a|_{\mathbf{U}(\varpi^{m} \mathfrak{o}_\ell )}$ extends the character $\theta_a|_{\mathbf{U}(\varpi^{m+1} \mathfrak{o}_\ell )}$ to group $\mathbf{U}(\varpi^m \mathfrak{o}_\ell ).$ The quotient group $\frac{\mathbf{U}(\varpi^m \mathfrak{o}_\ell )}{\mathbf{U}(\varpi^{m+1} \mathfrak{o}_\ell ) } $ is abelian, so (1) follows from Clifford theory. For (2), consider the normal subgroup $R_{{\bar{x}}} = {\textsf{K}^{m}_\ell \cap \mathrm{C}_{{\mathbf{G}}({\cO}_\ell)}(\widetilde{x})}\textsf{K}_\ell^{m+1}$ of $\textsf{K}_\ell^m.$ By ${\bf O. 3},$ for $\sigma \in \mathrm{Irr} \left( \textsf{K}_\ell^m \mid \varphi_x \right),$ there exists an extension $\widetilde{\varphi_x}$ of $\varphi_x$ to $R_{{\bar{x}}}$ such that $\sigma|_{R_{{\bar{x}}}} = \underbrace{\widetilde{\varphi_x} + \cdots + \widetilde{\varphi_x} }_{(d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}))/2-\mathrm{times}}$ and $\sigma$ is unique irreducible representation of $\textsf{K}_\ell^m$ lying above $\widetilde{\varphi_x}.$
Consider the hierarchy of groups given in Figure~\ref{figure:decomposition-to-U}. By Lemma~\ref{lem:stab-unipotent-intersection}, we have ${R_{\bar{x}}} \cap \mathbf{U}(\varpi^{m}\mathfrak{o}_\ell ) = \mathbf{U}(\varpi^{m+1}\mathfrak{o}_\ell ).$ Since $x$ is $\rho_{\ell, m}(a)$-regular, we have
\[
\widetilde{\varphi_x}|_{\mathbf{U}(\varpi^{m+1}\mathfrak{o}_\ell )} = {\varphi_x}|_{\mathbf{U}(\varpi^{m+1}\mathfrak{o}_\ell )} = \theta_a|_{\mathbf{U}(\varpi^{m+1}\mathfrak{o}_\ell)}
\]
and $\widetilde{\varphi_x}$ is stable under $\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell ),$ so the map $\widetilde{\varphi_x} \circ\theta_a : (R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) \rightarrow \mathbb C^\times$ defined by $(\widetilde{\varphi_x} \circ \theta_a)(xy) = \widetilde{\varphi_x}(x) \theta_a(y)$ for all $x \in R_{\bar{x}}$ and $y \in \mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )$ is a well defined one dimensional representation of $(R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) .$ Further, the one dimensional representation $\widetilde{\varphi_x} \circ \theta_a$ extends $\widetilde{\varphi_x}.$ The quotient group $(R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell ))/R_{\bar{x}}$ is easily see to be abelian of order $q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}.$ Therefore by Clifford theory, the set $\mathrm{Irr} \left( (R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) \mid \widetilde{\varphi_x} \right) $ consists of exactly $q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}$-many distinct one dimensional representations.
We claim that any $\chi \in \mathrm{Irr} \left( (R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) \mid \widetilde{\varphi_x} \right) $ satisfies $\mathrm{Ind}_{(R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell ))} ^{H_\ell^m} (\chi) \cong \sigma.$ Recall $\sigma$ is a unique irreducible representation of $H_\ell^m$ lying above $\widetilde{\varphi_x}$ and therefore we must have $\langle \sigma, \mathrm{Ind}_{(R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )} ^{H_\ell^m} (\chi) \rangle \neq 0 $ for every $\chi \in \mathrm{Irr} \left( (R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) \mid \widetilde{\varphi_x} \right) $ . Now the claim follows because
\[\dim(\sigma) = \dim(\mathrm{Ind}_{(R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell ))} ^{H_\ell^m} (\chi)) = {q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}},
\]
for every $\chi \in \mathrm{Irr} \left( (R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) \mid \widetilde{\varphi_x} \right) .$
\begin{figure}[htb!]
\label{figure:decomposition-to-U}
\centering
\begin{displaymath}
\xymatrix{
& H_\ell^m\ar@{-}[d]^{q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}} & \\
& (R_{\bar{x}})(\mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )) \ar@{-}[dr]^{q^{d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}} \ar@{-}[dl]^{q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}} & \\
{R_{\bar{x}}}\ar@{-}[dr] & & \mathbf{U}(\varpi^{m}\mathfrak{o}_\ell )\ar@{-}[dl] \\
& \mathbf{U}(\varpi^{m+1}\mathfrak{o}_\ell )\ar@{-}[d] & \\
& {\{1\}}& }
\end{displaymath}
\caption{}
\end{figure}
\end{proof}
\begin{lemma}
For $\ell = 2m+1,$ let $x \in \mathfrak{g}(\mathfrak{o}_m)$ be a $\rho_{\ell, m }(a)$-regular element and $\sigma \in \mathrm{Irr}(H_{\ell}^{m} \mid \varphi_{x}).$
Then the following are true.
\begin{enumerate}
\item The intersection $\mathbf{U}(\mathfrak{o}_{\ell}) \cap I(\sigma) = \mathbf{U}(\varpi^{m} \mathfrak{o}_\ell).$
\item We have, ${\rm Hom}_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})} \left(\widetilde{\sigma}, \theta_a \right) = {\rm Hom}_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})} \left(\sigma, \theta_a \right) \neq 0.$
\end{enumerate}
\end{lemma}
\begin{proof}
For (1), we note that by ${\bf O.4}$, $I(\sigma) = I_{{\mathbf{G}}(\mathfrak{o}_{\ell})} (\varphi_{x}) .$ The rest of the proof is similar to Lemma~\ref{lem:intersection}(1). For (2), note that the dimension of $\sigma$ is $q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}.$
By Lemma~\ref{lem:decomposition-of-sigma}, we have $\sigma|_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})}$ is direct sum of $ {q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}}$ distinct one dimensional representations of $\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}).$
We also know that $\sigma|_{\mathsf{K}_{\ell}^{m +1}}$ is isomorphic to direct sum of $q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}}$ times $\varphi_{x}.$
Therefore all the one dimensional representations of $\sigma|_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})}$ are extensions of $\varphi_{x}|_{\mathbf{U}(\varpi^{m+1} \mathfrak{o}_{\ell})}.$
Note that the index $[\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}) : \mathbf{U}(\varpi^{m +1} \mathfrak{o}_{\ell})] = q^{\frac{d_{ \mathfrak{g}}-d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})}{2}} = q^{\frac{n(n-1)}{2}}$ which is the same as the dimension of $\sigma.$
Thus $\sigma|_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})}$ is direct sum of all the possible extensions of $\varphi_{x}|_{\mathbf{U}(\varpi^{m +1} \mathfrak{o}_{\ell})}$ to the subgroup $\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell}).$
By hypothesis, the element $x$ is $\rho_{\ell, m }(a)$-regular therefore $\theta_a|_{\mathbf{U}(\varpi^{m +1} \mathfrak{o}_{\ell})} = \varphi_{x}|_{\mathbf{U}(\varpi^{m +1} \mathfrak{o}_{\ell})}$. This implies that $\theta_a|_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})}$ is an extension of $\varphi_{x}$ to $\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})$. This combined with the above discussion regarding $\sigma|_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})}$ gives
${\rm Hom}_{\mathbf{U}(\varpi^{m} \mathfrak{o}_{\ell})} (\sigma, \theta_a) \neq 0.$
\end{proof}
This completes the proof of the fact that for $a \in \mathfrak{o}_\ell^\times$ every $\rho_{\ell, m}(a)$ regular representation of ${\mathbf{G}}(\mathfrak{o}_\ell)$ admits a $\theta_a$-Whittaker model. Next, we prove that the sum of the dimensions of all inequivalent $\rho_{\ell, m }(a)$-regular representations of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ is equal to the dimension of the induced representation ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\theta_a).$
\begin{lemma}
Let $\mathcal{R}_a$ be the set of isomorphism classes of $\rho_{\ell, m }(a)$-regular representations of ${\mathbf{G}}(\mathfrak{o}_{\ell})$ and $\bar{x} \in \mathfrak{g}(\mathbb F_q)$ be a regular element. Then the following is true.
\[
\sum_{\pi \in \mathcal{R}_a} \dim(\pi) = \begin{cases} q^{\left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}) \right)m} | {\mathbf{G}}(\mathfrak{o}_m) |, & \mathrm{for} \,\, \ell = 2m \\
q^{ \left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})\right)m} \cdot q^{\frac{ d_{ \mathfrak{g}}+\left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})\right)}{2}} | {\mathbf{G}}(\mathfrak{o}_m) |, & \mathrm{for} \,\, \ell = 2m+1.
\end{cases}
\]
\end{lemma}
\begin{proof}
We prove this result for cases $\ell = 2m$ and $\ell = 2m+ 1$ separately.
For $\ell = 2m,$ by ${\bf E.4}$ and ${\bf E.5},$ the sum of dimensions of all inequivalent $\rho_{\ell, m }(a)$-regular representations lying above the one dimensional representation $\varphi_{x}$ for any $\rho_{\ell, m }(a)$-regular $x \in \mathfrak{g}(\mathfrak{o}_m)$ is the index $[{\mathbf{G}}(\mathfrak{o}_{\ell}) : \mathsf{K}_{\ell}^{m}]= |{\mathbf{G}}(\mathfrak{o}_m) |.$
We recall that, the number of $\rho_{\ell, m }(a)$-conjugacy classes of regular elements in $ \mathfrak{g}(\mathfrak{o}_m)$ is $q^{ \left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}) \right) m}$ (by Lemma \ref{lem:about-regular-elements}). Therefore
$\sum_{\pi \in \mathcal{R}_a} \dim(\pi) $ is $q^{ \left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x}) \right) m} |{\mathbf{G}}(\mathfrak{o}_m) |$ for $\ell = 2m.$
For $\ell = 2m+ 1,$ by ${\bf O.7},$ the sum of the dimensions of all inequivalent $\rho_{\ell, m }(a)$-regular representations lying over the one dimensional representation $\varphi_{x}$ of $\textsf{K}_\ell^{m+1}$ for any $\rho_{\ell, m}(a)$-regular $x \in \mathfrak{g}(\mathfrak{o}_m)$ is $q^{\frac{ d_{ \mathfrak{g}}+\left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})\right)}{2}}|{\mathbf{G}}(\mathfrak{o}_m) |.$ Multiplying this with the number of $\rho_{\ell, m }(a)$-regular conjugacy classes of $ \mathfrak{g}(\mathfrak{o}_m),$ we obtain our result even in this case.
\end{proof}
Next, we proceed to prove that the dimension of ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}( \theta_a)$ is same as $\sum_{\pi \in \mathcal{R}_a} \dim(\pi)$ for all $\ell.$
We note that $| \mathbf{U}(\mathfrak{o}_{\ell}) | = q^{\frac{n(n-1)}{2}\ell} = q^{\frac{d_{ \mathfrak{g}} - \left(d_{ \mathfrak{g}(\mathfrak{o}_1)}(\bar{x})\right)}{2} \ell} $ for all $\ell.$ Consider the short exact sequence, \[
1 \rightarrow \mathsf{K}_{\ell}^{m} \rightarrow {\mathbf{G}}(\mathfrak{o}_{\ell}) \rightarrow {\mathbf{G}}(\mathfrak{o}_m) \rightarrow 1.
\]
Therefore, $| {\mathbf{G}}(\mathfrak{o}_{\ell}) | = |\textsf{K}^m_\ell| | {\mathbf{G}}(\mathfrak{o}_m) | $ and we have the following,
\[
|\textsf{K}^m_\ell| = \begin{cases} q^{(d_{ \mathfrak{g}}) m}, & \mathrm{for}\,\, \ell = 2m\\ q^{d_{ \mathfrak{g}}(m +1)}, & \mathrm{for} \,\, \ell = 2m+ 1.\end{cases}
\]
Since, the dimension of ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\theta)$ is equal to $[{\mathbf{G}}(\mathfrak{o}_{\ell}) : \mathbf{U}(\mathfrak{o}_{\ell})],$ we obtain the result by substituting $|{\mathbf{G}}(\mathfrak{o}_{\ell})|$ and $|\mathbf{U}(\mathfrak{o}_{\ell})|$ from above. We have proved that every $\pi \in \mathcal{R}_a$ is a constituent of ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}(\theta)$ and the dimension of ${\rm Ind}_{\mathbf{U}(\mathfrak{o}_{\ell})}^{{\mathbf{G}}(\mathfrak{o}_{\ell})}( \theta_a)$ is equal to $\sum_{\pi \in \mathcal{R}_a} \dim(\pi)$ for all $\ell \geq 2.$ Both of these facts together complete the proof of Theorem~\ref{thm: existence}.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:main-theorem}}
\label{sec:proof-of-multiplicity-one}
Now we are in a position to complete the proof of Theorem~\ref{thm:main-theorem}. For $\ell \geq 2$ and for the groups ${\mathbf{G}} = {\rm GL}_n$ or ${\mathbf{G}} = {\rm SL}_n$ with $(p, n) = (p,2) = 1,$ Theorem~\ref{thm:main-theorem} follows directly from Theorem~\ref{main-theorem-2}. For ${\mathbf{G}} = {\rm SL}_n$ with $p \mid n$ or $p=2,$ we note that $\mathrm{Ind}_{\mathbf {U}(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\theta) \cong \mathrm{Ind}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}\left( \mathrm{Ind}_{\mathbf {U}(\mathfrak{o}_\ell)}^{{\rm SL}_n(\mathfrak{o}_\ell)}(\theta)\right) $ and $\mathrm{Ind}_{\mathbf {U}(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\theta)$ is multiplicity free for non-degenerate $\theta$ by Theorem~\ref{main-theorem-2}. Therefore $\mathrm{Ind}_{\mathbf {U}(\mathfrak{o}_\ell)}^{{\rm SL}_n(\mathfrak{o}_\ell)}(\theta)$ must be multiplicity free for every non-degenerate character $\theta$ of $\mathfrak{o}_\ell.$
\section{Examples: ${\rm GL}_2$ and ${\rm SL}_2$}
\label{sec:examples}
In this section, we discuss few details regarding the irreducible representations ${\rm GL}_2(\mathfrak{o}_\ell)$ and ${\rm SL}_2(\mathfrak{o}_\ell)$ for $\ell \geq 2$ to bring out the differences in these cases regarding Theorem~\ref{main-theorem-2}.
Throughout this section, we assume that $2 \nmid |\mathfrak{o}/\wp|$.
In the construction of irreducible representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ for ${\mathbf{G}} = {\rm GL}_n$ or ${\rm SL}_n$ one usually focuses on the construction of its primitive irreducible representations. These are precisely the representations that can not be obtained via pull back from the representations of ${\mathbf{G}}(\mathfrak{o}_i)$ for $i < \ell$. Formally, the primitive representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ are defined as below.
\begin{definition}
An irreducible representation $\rho$ of ${\mathbf{G}}(\mathfrak{o}_\ell)$ is called {\rm primitive} if the orbit of its restriction to $\textsf{K}_\ell^{\ell-1} $ does not contain a one dimensional representation $ \varphi_x $ where $x$ is a scalar matrix.
\end{definition} It is to be noted that by definition every regular representation is primitive but converse need not to be true. However it turns out to be true for ${\rm GL}_2(\mathfrak{o}_\ell)$ and ${\rm SL}_2(\mathfrak{o}_\ell)$ due to the fact that any $x \in M_2(\mathbb F_q)$ is either a scalar matrix or a cyclic matrix. Therefore every one dimensional representation $\varphi_x$ of $\textsf{K}_\ell^{\ell -1}$ is such that either $x$ is regular or scalar. In particular, an irreducible representation $\rho$ of ${\rm GL}_2(\mathfrak{o}_\ell)$ or ${\rm SL}_2(\mathfrak{o}_\ell)$ is regular if and only if it is primitive. Let $\mathcal{P}_{{\rm GL}_2(\mathfrak{o}_\ell)}$ and $\mathcal{P}_{{\rm SL}_2(\mathfrak{o}_\ell)}$ be the set of inequivalent primitive irreducible representations of ${\rm GL}_2(\mathfrak{o}_\ell)$ and ${\rm SL}_2(\mathfrak{o}_\ell)$ respectively. These sets are further partitioned into three types as follows. Any irreducible representation $\rho \in \mathcal{P}_{{\rm GL}_2(\mathfrak{o}_\ell)}$ or $\mathcal{P}_{{\rm SL}_2(\mathfrak{o}_\ell)}$ is called (of type) {\it cuspidal}, {\it split non-semisimple} and {\it split semisimple} if $\langle \rho|_{K_\ell^{\ell-1}} , \varphi_x \rangle \neq 0$ implies characteristic polynomial of $x$ is irreducible, splits over $\mathbb F_q$ and has equal roots and splits over $\mathbb F_q$ with distinct roots respectively.
For the groups ${\rm GL}_2(\mathfrak{o}_\ell)$, the information known regarding the dimensions of all elements of $\mathcal{P}_{{\rm GL}_2(\mathfrak{o}_\ell)}$, see \cite[Theorem~1.4]{MR2456275} is given in Table~\ref{GL2-reps}.
\begin{table}[h]
\begin{center}
\begin{tabular}{c|c|c|c}
S.No.& Type of $\rho$ & Number of $ \rho \in \mathcal{P}_{{\rm GL}_2(\mathfrak{o}_\ell)} $ & dimension of $\rho$ \\
\hline
\hline
1. & cuspidal & $\frac{1}{2} (q-1)(q^2-1) q^{2 \ell -3}$ & $q^{\ell-1}(q-1)$ \\
2. & split non-semisimple & $(q-1) q^{2 \ell -2}$ & $(q^2-1)q^{\ell-2}$ \\
3. & split semisimple & $\frac{1}{2} q^{2\ell-3} (q-1)^3 $ & $q^{\ell-1} (q+1)$
\end{tabular}
\end{center}
\caption{ Dimensions of regular representations for ${\rm GL}_2(\mathfrak{o}_\ell)$}
\label{GL2-reps}
\end{table}
So in this case, we obtain the following.
\[
\sum_{\rho \in \mathcal{P}_{{\rm GL}_2(\mathfrak{o}_\ell) }} \dim(\rho) = (q^2 - 1) (q-1) q^{3 \ell -3} = \left|\frac{{\rm GL}_2(\mathfrak{o}_\ell)}{\mathbf U (\mathfrak{o}_\ell)}\right|.
\] In case of ${\rm GL}_2(\mathbb F_q)$, it is well known that an irreducible representation of ${\rm GL}_2(\mathbb F_q)$ admits a $\theta$-Whittaker model for a non-degenerate $\theta$ if and only if it has dimension greater than one and this result by Theorem~\ref{main-theorem-2} and above computations generalize to ${\rm GL}_2(\mathfrak{o}_\ell)$ as follows.
\begin{theorem}
\label{thm:primitive} An irreducible representation of ${\rm GL}_2(\mathfrak{o}_\ell)$ for $\ell \geq 2$ admits a $\theta$-Whittaker model for any non-degenerate character $\theta$ of $\mathfrak{o}_\ell$ if and only if it is primitive.
\end{theorem}
It is easy to see that the above result does not hold for ${\rm GL}_n(\mathfrak{o}_\ell)$ with $\ell \geq 2$ and $n \geq 3$. In comparison of the above result with ${\rm SL}_2(\mathfrak{o}_\ell)$, the numbers and dimensions of elements of $\mathcal{P}_{{\rm SL}_2(\mathfrak{o}_\ell)}$ are as given in Table~\ref{SL2-reps}, see \cite[Section~7]{MR2169043}.
\begin{table}
\begin{center}
\begin{tabular}{c|c|c|c}
S.No.& Type of $\rho$ & Number of $ \rho \in \mathcal{P}_{{\rm SL}_2(\mathfrak{o}_\ell)} $ & dimension of $\rho$ \\
\hline
\hline
1. & cuspidal & $\frac{1}{2} (q^2-1) q^{\ell -2}$ & $q^{\ell-1}(q-1)$ \\
2. & split non-semisimple & $4q^{\ell -1}$ & $\frac{1}{2}(q^2-1)q^{\ell-2}$ \\
3. & split semisimple & $\frac{1}{2} q^{\ell-2} (q-1)^2 $ & $q^{\ell-1} (q+1)$
\end{tabular}
\end{center}
\caption{ Dimensions of regular representations for ${\rm SL}_2(\mathfrak{o}_\ell)$}
\label{SL2-reps}
\end{table}
For group ${\rm SL}_2(\mathfrak{o}_\ell)$, we have the following.
\[
\sum_{\rho \in \mathcal{P}_{{\rm SL}_2(\mathfrak{o}_\ell) }} \dim(\rho) = (q^2-1)(q+1)q^{2 \ell -3} \gneq (q^2-1) q^{2 \ell -4} = \left|\frac{{\rm SL}_2(\mathfrak{o}_\ell)}{\mathbf U (\mathfrak{o}_\ell)}\right|.
\]
This in particular implies that for any non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$, there exists regular(primitive) representations $\rho_1$ and $\rho_2$ of ${\rm SL}_2(\mathfrak{o}_\ell)$ such that $\rho_1$ admits a $\theta$-Whittaker model but $\rho_2$ does not admit a $\theta$-Whittaker model. Further it is to be noted from the dimensions given in Tables~\ref{GL2-reps} and \ref{SL2-reps} that the cuspidal and split semisimple representations of ${\rm GL}_2(\mathfrak{o}_\ell)$ in fact restrict to irreducible representations of ${\rm SL}_2(\mathfrak{o}_\ell)$. This in particular implies the following.
\begin{corollary}
Any cuspidal or split-semisimple regular representation of the group ${\rm SL}_2(\mathfrak{o}_\ell)$ admits a $\theta$-Whittaker model for any non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$.
\end{corollary}
This result also extends to ${\rm SL}_n(\mathfrak{o}_\ell)$ for $(p,2) = (p,n) = 1$ and we discuss this in Section~\ref{sec:branching-rules}, see Corollary~\ref{cor:cuspidal-whittaker}.
\section{Restriction from ${\rm GL}_n(\mathfrak{o}_\ell)$ to ${\rm SL}_n(\mathfrak{o}_\ell)$ }
\label{sec:branching-rules}
In this section, we prove Theorem~\ref{thm:multiplicity-free-induction}. We also obtain the branching rules for the restriction of regular representations of ${\rm GL}_n(\mathfrak{o}_\ell)$ to ${\rm SL}_n(\mathfrak{o}_\ell)$ for $(p,2) = (p,n) = 1$.
\subsection{Proof of Theorem~\ref{thm:multiplicity-free-induction}} Let $\rho$ be a regular representation of ${\rm GL}_n(\mathfrak{o}_\ell)$ and let $\delta$ be a regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ such that $\langle \delta, \rho \rangle_{{\rm SL}_n(\mathfrak{o}_\ell )} \neq 0 $. We need to prove that $\langle \delta, \rho \rangle_{{\rm SL}_n(\mathfrak{o}_\ell )} = 1 $. Let $\sigma$ be a regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ such that $\langle \sigma , \mathrm{Ind}_{\mathbf U(\mathfrak{o}_\ell)}^{{\rm SL}_n(\mathfrak{o}_\ell)}(\theta) \rangle \neq 0$ and $\langle \rho , \mathrm{Ind}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\sigma) \rangle \neq 0$. The representation $\mathrm{Ind}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\sigma)$ is a subrepresentation of $\mathrm{Ind}_{\mathbf U(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\theta)$ and therefore is multiplicity free by Theorem~\ref{thm:main-theorem}. By the fact that $\langle \delta, \rho \rangle_{{\rm SL}_n(\mathfrak{o}_\ell )} \neq 0 $, $\langle \sigma , \rho \rangle_{{\rm SL}_n(\mathfrak{o}_\ell )} \neq 0 $ and Clifford theory, we must have $\delta^g = \sigma $ for some $g \in {\rm GL}_n(\mathfrak{o}_\ell)$. This in particular implies that both ${\rm Ind}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\delta )$ and ${\rm Ind}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\sigma )$ are isomorphic as ${\rm GL}_n(\mathfrak{o}_\ell)$-representations. Therefore ${\rm Ind}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\delta )$ is multiplicity free. By Frobenius reciprocity, we obtain $\langle \delta, \rho \rangle_{{\rm SL}_n(\mathfrak{o}_\ell )} = 1 $.
\subsection{Branching rules}
For this we recall the definition of the type of a regular representation of ${\rm GL}_n(\mathfrak{o}_\ell)$ and ${\rm SL}_n(\mathfrak{o}_\ell)$ as given in \cite{MR3737836}. Now onwards we will always assume $(p, 2) = (p, n) = 1$.
\begin{definition} A matrix $\tau = (\tau_{d,e}) \in M_n(\mathbb Z_{\geq 0})$ is called $n$-typical if $n= \sum_{d, e} de\tau_{d,e}$. The set of $n$-typical matrices is denoted by $\mathcal{A}_n$.
\end{definition}
Let $\mathcal{P}(\mathbf k) $ be the set of all monic irreducible polynomials in $\mathbf k[t]$.
For any $x \in M_n(\mathbf k)$, the characteristic polynomial of $x$ is of the form $\prod_i f_i^{e_i}$ for some finite list of elements $f_i \in \mathcal{P}(\mathbf k) $ with $\sum_i e_i(\deg(f_i)) = n$. Therefore the polynomial $\prod_i f_i^{e_i}$ determines an $n$-typical matrix $\mathbf{m}^x$ with $\mathbf{m}^x_{d,e} = |\{ i \mid \deg(f_i) = d \,\,\, \mathrm{and}\,\,\, e_i = e \}|$. It is well known that the regular elements of $M_n(\mathbf k)$ up to conjugacy are determined uniquely by their characteristic polynomial. By above we have the following map.
\[
\{ \mathrm{ Regular \,\, conjugacy \,\, classes \,\, in \,\, } M_n(\mathbf k) \} \to \mathcal{A}_n\,\,; x \mapsto \mathbf{m}^x
\]
We remark that for $q>n$, the map $x \to \mathbf{m}^x$ is a surjection.
\begin{definition}($\tau$-regular element)
A regular element $x \in M_n(\mathbf k)$ is called $\tau$-regular for $\tau \in \mathcal{A}_n$ if $\mathbf{m}^x = \tau$.
\end{definition}
\begin{definition} ($\tau$-regular representation)
A regular representation $\rho$ of ${\mathbf{G}}(\mathfrak{o}_\ell)$ is said to be $\tau$-regular if restriction of $\rho$ to $K_\ell^{\ell-1}$ contains $\varphi_x$ for a $\tau$-regular element $x$.
\end{definition}
For $\tau \in \mathcal{A}_n$, $r \in \mathbb N$ and for a $\tau$-regular $x$, we define the following.
\[
\begin{array}{ccl}
\mathbf {v} & := & |{\rm GL}_n(\mathbf k)|, \\
\mathbf {u}^{\tau} & := & |C_{{\rm GL}_n(\mathbf k)}(x)|, \\
\upiota (\tau, r) & := & gcd\left( \{ e \mid \exists\,\, d \,\,\mathrm{with}\,\, \tau_{d,e} \neq 0\} \cup \{r \} \right).
\end{array}
\]
\begin{theorem}[\cite{MR3737836}, Theorems B and E]
\label{thm:regular-dimension}
For $\tau \in \mathcal{A}_n$, the dimension of any $\tau$-regular representation $\rho$ of $G(\mathfrak{o}_\ell)$ for $\ell \geq 2$ is given by the following.
\[
\dim(\rho) = \begin{cases} q^{\binom{n}{2} (\ell-2) }\frac{\mathbf{v}}{\mathbf {u}^{\tau} }, & \mathrm{for } \,\, {\mathbf{G}}(\mathfrak{o}_\ell) = {\rm GL}_n(\mathfrak{o}_\ell) \, \, \mathrm{with} \,\, (p,2) = 1, \\ q^{\binom{n}{2} (\ell-2) }\frac{\mathbf{v}}{\upiota(\tau, q-1) \mathbf {u}^{\tau}}, & \mathrm{for} \,\, {\mathbf{G}}(\mathfrak{o}_\ell) = {\rm SL}_n(\mathfrak{o}_\ell) \, \, \mathrm{with} \,\, (p,2) = (p,n) = 1.
\end{cases}
\]
\end{theorem}
The proof of the above result is included in \cite[Thereoms~B, E]{MR3737836}. In the following remark we clarify the differences that arise in the present formulation of above result and that of the corresponding statements in \cite{MR3737836}.
\begin{remark}
\begin{enumerate}
\item Our stated results are for ${\mathbf{G}}(\mathfrak{o}_\ell)$ and the results of \cite{MR3737836} are for ${\mathbf{G}}(\mathfrak{o}_{\ell+1})$. So $\ell$ is modified appropriately.
\item The statement of \cite[Theorem B]{MR3737836} that gives above result for ${\mathbf{G}}(\mathfrak{o}_\ell) = {\rm GL}_n(\mathfrak{o}_\ell)$, has the hypothesis $q>n$. However the remark after the statement of Theorem~B there clarifies that the above result is true even for $q \leq n$. The hypothesis $q > n$ is required there for some other counting arguments which we will never use, so we have omitted the condition $q > n$.
\item The proof of Theorem~\ref{thm:regular-dimension} for ${\mathbf{G}}(\mathfrak{o}_\ell) = {\rm SL}_n(\mathfrak{o}_\ell)$ is part of the proof of \cite[Theorem E]{MR3737836}.
\end{enumerate}
\end{remark}
\begin{corollary}
\label{cor:dimensions-of-regular}
The dimension of any $\tau$-regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ is obtained by dividing the dimension of a $\tau$-regular representation of ${\rm GL}_n(\mathfrak{o}_\ell)$ by $\upiota(\tau, q-1)$.
\end{corollary}
\begin{definition}(special-regular representation of ${\rm SL}_n(\mathfrak{o}_\ell )$) A regular representation $\delta$ of ${\rm SL}_n(\mathfrak{o}_\ell)$ is called special regular if and only if $\delta \cong \mathrm{Res}^{{\rm GL}_n(\mathfrak{o}_\ell)}_{{\rm SL}_n(\mathfrak{o}_\ell)}(\rho)$ for a regular representation $\rho$ of ${\rm GL}_n(\mathfrak{o}_\ell)$.
\end{definition}
By Corollary~\ref{cor:dimensions-of-regular} and the definition of special-regular, we obtain the following.
\begin{lemma}
A $\tau$-regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ is special-regular if and only if $\upiota(\tau, (q-1)) = 1$.
\end{lemma}
\begin{theorem}
\label{prop:special-regular-characterization} A regular representation $\delta$ of ${\rm SL}_n(\mathfrak{o}_\ell)$ admits a $\theta$-Whittaker model for every non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$ if and only if $\delta$ is special-regular.
\end{theorem}
\begin{proof}
Let $\delta$ be a special-regular representations of ${\rm SL}_n(\mathfrak{o}_\ell)$, then $\delta \cong \mathrm{Res}^{{\rm GL}_n(\mathfrak{o}_\ell)}_{{\rm SL}_n(\mathfrak{o}_\ell)}(\rho)$ for a regular representation $\rho$ of ${\rm GL}_n(\mathfrak{o}_\ell)$. Let $\theta$ be a non-degenerate character of $\mathbf U(\mathfrak{o}_\ell)$. Then by Corollary~\ref{cor:hill-regular-iff-whittaker}, we have
\[
\langle \delta , \mathrm{Ind}^{{\rm SL}_n(\mathfrak{o}_\ell)}_{\mathbf U(\mathfrak{o}_\ell)}\theta \rangle = \langle \mathrm{Res}^{{\rm GL}_n(\mathfrak{o}_\ell)}_{{\rm SL}_n(\mathfrak{o}_\ell)}(\rho), \mathrm{Ind}^{{\rm SL}_n(\mathfrak{o}_\ell)}_{\mathbf U(\mathfrak{o}_\ell)}\theta \rangle = \langle \rho, \mathrm{Ind}^{{\rm GL}_n(\mathfrak{o}_\ell)}_{\mathbf U(\mathfrak{o}_\ell)}\theta \rangle = 1.
\]
Conversely, let $\delta$ be a regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ such that $\delta$ is not a special-regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ and admits a $\theta$-Whittaker model for every non-degenerate $\theta$. Then there exists $\delta'$ (possibly equivalent to $\delta$) such that both $\delta$ and $\delta'$ are regular representations of ${\rm SL}_n(\mathfrak{o}_\ell)$ and $\delta \oplus \delta'$ is a subrepresentation of $\mathrm{Res}^{{\rm GL}_n(\mathfrak{o}_\ell)}_{{\rm SL}_n(\mathfrak{o}_\ell)}(\rho)$ for some regular representation $\rho$ of ${\rm GL}_n(\mathfrak{o}_\ell)$. The representation $\delta'$ is a regular representation of ${\rm SL}_n(\mathfrak{o}_\ell)$, therefore by Theorem~\ref{main-theorem-2} there exists a non-degenerate character $\theta$ such that $\delta'$ admits a $\theta$-Whittaker model. Then by our hypothesis,
\[
\langle \rho, \mathrm{Ind}_{\mathbf U(\mathfrak{o}_\ell)} ^{{\rm GL}_n(\mathfrak{o}_\ell)} (\theta) \rangle \geq \langle \delta \oplus \delta' , \mathrm{Ind}_{\mathbf U(\mathfrak{o}_\ell)} ^{{\rm SL}_n(\mathfrak{o}_\ell)}(\theta) \rangle = 2.
\]
This is a contradiction to Theorem~\ref{thm:main-theorem}.
\end{proof}
Recall a regular element $x$ is called cuspidal if the characteristic polynomial of $x$ is irreducible and cuspidal representations of ${\mathbf{G}}(\mathfrak{o}_\ell)$ are the ones that lie above the cuspidal characters.
A regular element $x$ is called split-semisimple if the characteristic polynomial splits over $\mathbf F_q$ and has distinct roots. Parallel to the cuspidal representations, the split semisimple representation of ${\mathbf{G}}(\mathfrak{o}_\ell)$ are defined. Both the cuspidal and split semisimple representations of ${\rm SL}_n(\mathfrak{o}_\ell)$ are special regular. This along with Theorem~\ref{prop:special-regular-characterization} directly gives the following.
\begin{corollary}
\label{cor:cuspidal-whittaker}
Let $\rho$ be a cuspidal or a split semisimple representation of ${\rm SL}_n(\mathfrak{o}_\ell)$ for $(p,2) = (p,n) = 1$. Then $\rho$ admits a $\theta$-Whittaker model for every non-degenerate character $\theta$ of $\mathbf U(\mathfrak{o}_\ell)$.
\end{corollary}
Next, we describe the branching rules for $\mathrm{Res}^{{\rm GL}_n(\mathfrak{o}_\ell)}_{{\rm SL}_n(\mathfrak{o}_\ell)}(\rho)$ for every regular representation $\rho$ of ${\rm GL}_n(\mathfrak{o}_\ell)$.
\begin{theorem}
\label{thm:branching-numbers} For $(p, 2) = (p, n) = 1$. Let $\rho$ be a $\tau$-regular representation of ${\rm GL}_n(\mathfrak{o}_\ell)$ for $\tau \in \mathcal{A}_n$. Then
\[
\mathrm{End}_{{\rm SL}_n(\mathfrak{o}_\ell)} \left( \mathrm{Res}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)} (\rho) \right) \cong \mathbb C^{ \upiota(\tau, q-1) } .
\]
\end{theorem}
\begin{proof}
By Theorem~\ref{thm:multiplicity-free-induction}, the representation $\mathrm{Res}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)} (\rho)$ is multiplicity free. The group ${\rm SL}_n(\mathfrak{o}_\ell)$ is a normal subgroup of ${\rm GL}_n(\mathfrak{o}_\ell)$, therefore by Clifford theory all the constituents of $\mathrm{Res}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)} (\rho)$ must be of the same dimension and also must be $\tau$-regular representations of ${\rm SL}_n(\mathfrak{o}_\ell)$. Therefore by Corollary~\ref{cor:dimensions-of-regular}, the isomorphism of $\mathrm{End}_{{\rm SL}_n(\mathfrak{o}_\ell)} \left( \mathrm{Res}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)} (\rho) \right) $ and $\mathbb C^{ \upiota(\tau, q-1) } $ follows.
\end{proof}
The following is a direct corollary of Theorem~\ref{thm:branching-numbers} by the fact that
$\upiota(\tau, q-1) \leq n$ for every $\tau \in \mathcal{A}_n$.
\begin{corollary}
\label{cor:number-of-constituents}
For $(p, 2) = (p, n) = 1$. Let $\rho$ be a regular representation of ${\rm GL}_n(\mathfrak{o}_\ell)$. Then $\mathrm{Res}_{{\rm SL}_n(\mathfrak{o}_\ell)}^{{\rm GL}_n(\mathfrak{o}_\ell)}(\rho)$ has at most $n$ irreducible constituents.
\end{corollary}
\begin{remark}
\begin{enumerate}
\item We remark that Corollary~\ref{cor:number-of-constituents} can also be obtained by considering the index of the group $Z({\rm GL}_n(\mathfrak{o}_\ell)){\rm SL}_n(\mathfrak{o}_\ell)$ in ${\rm GL}_n(\mathfrak{o}_\ell)$, where $Z({\rm GL}_n(\mathfrak{o}_\ell))$ denotes the centre of the group ${\rm GL}_n(\mathfrak{o}_\ell)$.
\item Both of the above mentioned proofs of Corollary~\ref{cor:number-of-constituents} very much depend on the condition $(p, n) = 1$. It is natural to ask whether this result continues to be true for general $p$ and $n$?
\end{enumerate}
\end{remark}
\noindent {\bf Acknowledgements: } Authors immensely thank Dipendra Prasad for several helpful discussions regarding this paper and are also grateful to Sandeep Varma for his comments that helped in improving the current article. SPP acknowledges the support provided by Inspire Faculty Award [IFA17-MA102], DST, Government of India and PS acknowledges the support of UGC Centre for Advanced Study grant.
|
\section{Introduction}
\paragraph{1.1 Dehn surgery of $3$-manifolds.} In $3$-dimensional topology, Dehn surgery is an operation of modifying a $3$-manifold by cutting off a solid torus and then gluing it back in a different way. The Lickorish-Wallace theorem, which states that every closed connected orientable $3$-manifold can be obtained from the $3$-dimensional sphere by performing finitely many Dehn surgeries, serves as a motivation of the study of Dehn surgeries.
The second step of the surgery, called \textit{Dehn filling}, can be formalized as follows. Let $M$ be a $3$-manifold with toral boundary. Topologically distinct ways of gluing a solid torus to $M$ are parametrized by free homotopy classes of essential simple closed curves on $\partial M$, called \textit{slopes}. For a slope $s$, the Dehn filling $M(s)$ is obtained by attaching a solid torus $S^1\times D^2$ to $\partial M$ such that $\partial D^2$ is mapped to a curve of the slope $s$. The following is a special case of Thurston's hyperbolic Dehn filling theorem.
\begin{theorem}[{\cite[Theorem {[TH1]}]{thurston1983three}}]\label{thm. thurston}
Let $M$ be a compact orientable $3$-manifold with toral boundary such that $M\setminus\partial M$ admits a complete finite-volume hyperbolic structure. Then $M(s)$ is hyperbolic for all but finitely many slopes $s$.
\end{theorem}
\paragraph{1.2 Group theoretic Dehn fillings.} In group theoretic settings, Dehn filling can be generalized as follows. Let $G$ be a group, let $H$ be a subgroup of $G$, and let $N$ be a normal subgroup of $H$. The \textit{group theoretic Dehn filling} associated with the data $(G,H,N)$ is the process of forming the quotient group $G/\ll N \rr$, where $\ll N \rr$ is the normal closure of $N$ in $G$.
Under the assumptions of Theorem \ref{thm. thurston}, let $G=\pi_1(M)$. The natural map $\pi_1(\partial M)\rightarrow \pi_1(M)$ is injective. We think of $\pi_1(\partial M)$ as a subgroup of $\pi_1(M)$ and let $H=\pi_1(\partial M)$. Let $N\lhd H$ be the subgroup generated by the slope $s$. Then $G/\ll N \rr=\pi_1(M(s))$ by the Seifert-van Kampen theorem.
Dehn filling is a fundamental tool in group theory. The solution of the virtually Haken conjecture uses Dehn fillings of hyperbolic groups \cite{agol2013virtual}. For a large number of relatively hyperbolic groups, Dehn fillings are used to prove the Farrell-Jones conjecture \cite{antolin2017farrell} and solve the isomorphism problem \cite{dahmani2018recognizing}. By considering Dehn fillings of hyperbolically embedded subgroups, \cite{dahmani2017hyperbolically} constructs purely pseudo-Anosov normal subgroups of mapping class groups. Other applications of Dehn fillings can be found in \cite{agol2016alternate,groves2016boundaries}.
In group theoretic settings, Thurston's theorem was first generalized by Osin \cite{osin2007peripheral}, and independently by Groves-Manning \cite{groves2008dehn} to Dehn fillings of peripheral subgroups of relatively hyperbolic groups. More recently, Dahmani-Guirardel-Osin \cite{dahmani2017hyperbolically} proved an analog of Thurston's theorem in the more general settings of groups with hyperbolically embedded subgroups (see Theorem \ref{thm. first part of simple DGO} below and the discussion afterwards). We discuss here some examples and refer to Section \ref{subsec.wh} for the definition. We use $H\hookrightarrow_h G$ to indicate that $H$ is a hyperbolically embedded subgroup of $G$.
\begin{example}\label{eg. 1}
If $H$ is a peripheral subgroup of a relatively hyperbolic group $G$, then $H\hookrightarrow_h G$ \cite[Proposition 2.4]{dahmani2017hyperbolically}. For example,
\begin{enumerate}
\item[(a)] if a group $G$ decomposes as a free product $G=A\ast B$, then we have $A\hookrightarrow_h G$ and $B\hookrightarrow_h G$ \cite{osin2006relatively};
\item[(b)] under the assumptions of Theorem \ref{thm. thurston}, we have $\pi_1(\partial M)\hookrightarrow_h\pi_1(M)$ \cite{bowditch2012relatively,farb1998relatively}.
\end{enumerate}
\end{example}
\begin{example}\label{eg. 2}
Let $G$ be a group acting acylindrically on a Gromov hyperbolic space and let $g$ be a loxodromic element of $G$. Then there exists a maximal virtually cyclic subgroup $E(g)\leqslant G$ containing $g$ such that $E(g)\hookrightarrow_h G$ \cite[Corollary 2.9]{dahmani2017hyperbolically}. In particular, if $G$ is a hyperbolic group (resp. the mapping class group of a punctured closed orientable surface \cite[Theorem 2.19]{dahmani2017hyperbolically}, outer automorphism group of a finite rank non-abelian free group \cite[Theorem 2.20]{dahmani2017hyperbolically}) and $g$ is a loxodromic (resp. pseudo-Anosov, fully irreducible) element, then $E(g)\hookrightarrow_h G$.
\end{example}
Other examples of hyperbolically embedded subgroups can be found in \cite{dahmani2017hyperbolically}.
\begin{theorem}[{\cite[Theorem 2.27]{dahmani2017hyperbolically}}]\label{thm. first part of simple DGO}
Let $G$ be a group with a subgroup $H\hookrightarrow_h G$. Then there exists a finite set $\mathcal{F}\subset H\setminus\{1\}$ such that if $N\lhd H$ and $N\cap \mathcal{F}=\emptyset$, then the natural homomorphism $H/N\rightarrow G/\ll N \rr $ maps $H/N$ injectively onto a hyperbolically embedded subgroup of $G/\ll N \rr$.
\end{theorem}
Under the assumptions of Theorem \ref{thm. thurston}, the above theorem, together with some basic facts about relatively hyperbolic groups, implies that $\pi_1(M(s))$ is non-virtually-cyclic and word-hyperbolic for all but finitely many slopes $s$. Thurston's geometrization conjecture, proved by Perelman, implies that this algebraic statement about $\pi_1(M(s))$ is equivalent to the hyperbolicity of $M(s)$. Thus, the above theorem indeed provides a generalization of Theorem \ref{thm. thurston}.
\paragraph{1.3 Motivation: a question on group cohomology.}
Note that in the settings of Thurston's theorem, i.e., if $G=\pi_1(M),H=\pi_1(\partial M),$ and $M(s)$ admits a hyperbolic structure, we have
$$H^{\ast}(G/\ll N \rr;\cdot)\cong H^{\ast}(\pi_1(M(s));\cdot),$$
which can be computed via $M(s)$. Indeed, as $M(s)$ admits a hyperbolic structure, the universal cover of $M(s)$ is $\mathbb{H}^3$, which is contractible, and thus $M(s)$ is a model of $K(G/\ll N \rr,1)$.
However, there are no analogous methods for Dehn fillings of hyperbolically embedded subgroups. The main question motivating our research is the following.
\begin{question}\label{question. main question}
For a group $G$ with a subgroup $H\hookrightarrow_h G$ and a normal subgroup $N\lhd H$, what can be said about $H^{\ast}(G/\ll N \rr;\cdot)$?
\end{question}
In this series of three papers, we answer this question and discuss some applications. The first task is to understand the structure of $\ll N \rr$, which is solved by the present paper. In the second paper \cite{sun2019cohomologyii}, we will combine structural results obtained in this paper and the Lyndon-Hochschild-Serre spectral sequence to compute $H^{\ast}(G/\ll N \rr; \cdot)$. In the third paper \cite{sun2019cohomologyiii}, we will study cohomological properties of $G/\ll N \rr$ and discuss some applications to acylindrically hyperbolic groups.
\section{Main results}
\paragraph{2.1 Cohen-Lyndon type theorems for $\ll N \rr$.}
In general, $\ll N \rr$ does not need to have any particular structure. Nevertheless, it turns out that if $N$ avoids a finite set of bad elements, then $\ll N \rr$ enjoys a nice free product structure. In order to state our main results, we introduce the following terminology.
\begin{definition}
Let $G$ be a group with a subgroup $H\hookrightarrow_h G$. We say that a property $P$ holds \textit{for all sufficiently deep} normal subgroups $N\lhd H$ if there exists a finite set $\mathcal{F}\subset H\setminus\{1\}$ such that $P$ holds for all normal subgroups $N\lhd H$ with $N\cap\mathcal{F}=\emptyset$.
\end{definition}
\begin{definition}\label{simple CLp}
Let $G$ be a group with a subgroup $H$ and let $N\lhd H$. We say that the triple $(G,H,N)$ has the \textit{Cohen-Lyndon property} if there exists a left transversal $T$ of $H\ll N \rr$ in $G$ such that $\ll N \rr$ is the free product of its subgroups $N^t=tNt^{-1}$ for $t\in T$, denoted as
$$\ll N \rr=\prod^{\ast}_{t\in T}N^t.$$
\end{definition}
The latter definition is motivated by the following result \cite[Theorem 4.1]{cohen1963free}, which was later generalized by \cite[Theorem 1.1]{edjvet1987cohen} to free products of locally indicable groups.
\begin{theorem}[Cohen-Lyndon]\label{classical CL}
Let $F$ be a free group and let $C$ be a maximal cyclic subgroup of $F$. Then for all $f\in C\setminus\{1\}$, the triple $(F,C,\langle f \rangle)$ has the Cohen-Lyndon property.
\end{theorem}
Note that $F$ is hyperbolic and $f$ has infinite order. By Example \ref{eg. 2}, we have $C=E(f)\hookrightarrow_h F$ and thus the above theorem fits in the general framework of group theoretic Dehn fillings. For general hyperbolically embedded subgroups, a weak version of the Cohen-Lyndon property is given in \cite[Theorem 2.27]{dahmani2017hyperbolically}.
\begin{theorem}[{Dahmani-Guirardel-Osin}]\label{simple DGO}
Let $G$ be a group with a subgroup $H\hookrightarrow_h G$. Then for all sufficiently deep $N\lhd H$, we have
$$\ll N \rr=\prod^{\ast}_{t\in T}N^t$$
for some subset $T\subset G$.
\end{theorem}
The main difference between Theorems \ref{simple DGO} and \ref{classical CL} is that in Theorem \ref{simple DGO}, $T$ is just some subset of $G$, instead of being a left transversal of $H \ll N \rr$ in $G$. Our result improves Theorem \ref{simple DGO}.
\begin{theorem}\label{simmain}
Suppose that $G$ is a group with a subgroup $H\hookrightarrow_h G$. Then $(G,H,N)$ has the Cohen-Lyndon property for all sufficiently deep $N\lhd H$.
\end{theorem}
In the special case where $G$ and $H$ are finitely generated and $G$ is hyperbolic relative to $H$, Theorem \ref{simmain} is proved in \cite[Theorem 4.8]{groves2016boundaries}. The proofs of \cite[Theorem 7.15]{dahmani2017hyperbolically} and \cite[Theorem 4.8]{groves2016boundaries} use windmills, very rotating families, and spiderwebs. It is possible to prove Theorem \ref{simmain} with these notions. However, the present paper provides a proof of a different flavor, using surgery on van Kampen diagrams and geodesic polygons of Cayley graphs.
\begin{remark}\label{rem. general version of CL}
In fact, we prove Theorem \ref{simmain} in more general settings of a group with a family of \textit{weakly hyperbolically embedded subgroups} (see Definition \ref{hyperbolically embedded}). As an application, we obtain Cohen-Lyndon type theorems for graphs of groups, e.g., amalgamated free products and HNN-extensions (see Corollaries \ref{graph of groups}, \ref{amalproduct}, and \ref{HNN}).
\end{remark}
Combining Theorem \ref{simmain} and Example \ref{eg. 2}, we obtain:
\begin{corollary}
Let $G$ be a group acting acylindrically on a Gromov hyperbolic space, and let $g\in G$ be a loxodromic element. Then $(G,E(g),N)$ has the Cohen-Lyndon property for all sufficiently deep $N\lhd E(g)$.
\end{corollary}
In the case where $G=F$ and $H=C$, we recover Theorem \ref{classical CL} for sufficiently deep (but not all) $\langle f \rangle\lhd C$. In the case where $G$ is a free product of locally indicable groups, by considering the action of $G$ on the corresponding Bass-Serre tree, we recover \cite[Theorem 1.1]{edjvet1987cohen} for sufficiently deep normal subgroups.
\paragraph{2.2 Structure of relative relation modules.} Let $Rel(G,\ll N \rr)$ and $Rel(H,N)$ be the relative relation modules of the exact sequences
$$1\rightarrow \ll N \rr \rightarrow G\rightarrow Q\rightarrow 1$$
and
$$1\rightarrow N\rightarrow H\rightarrow R\rightarrow 1,$$
respectively, i.e. $Rel(G,\ll N \rr)$ (resp. $Rel(H,N)$) is the $\mathbb{Z}Q$-module (resp. $\mathbb{Z}R$-module) whose base set is the abelianization of $\ll N \rr$ (resp. $N$) and the $Q$-action (resp. $R$-action) is induced by conjugation. If $G$ is free, then $Rel(G,\ll N \rr)$ is called a \textit{relation module}. For sufficiently deep $N$, it follows immediately from Theorem \ref{thm. first part of simple DGO} that the natural map identifies $R$ with a subgroup of $Q$. We can then further identify $\mathbb{Z}R$ with a subring of $\mathbb{Z}Q$. Thus, given any $\mathbb{Z}R$-module $A$, it makes sense to talk about the \textit{induced module} of $A$ from $\mathbb{Z}R$ to $\mathbb{Z}Q$, which is denoted by $Ind^Q_R A=\mathbb{Z}Q\bigotimes_{\mathbb{Z}R}A$.
If $G=F$ and $H=C$, Theorem \ref{classical CL} directly implies
$$Rel(F,\ll f \rr)\cong\mathbb{Z}[F/C\ll f \rr]\cong Ind^Q_R \mathbb{Z}\cong Ind^Q_R Rel(C,\langle f \rangle)$$
as $\mathbb{Z}Q$-modules. In general, we have the following corollary to Theorem \ref{simmain}.
\begin{corollary}\label{module}
Let $G$ be a group with a subgroup $H\hookrightarrow_h G$. Then for all sufficiently deep $N\lhd H$, there is an isomorphism of $\mathbb{Z}Q$-modules
\begin{equation}\label{iso}
Rel(G,\ll N \rr)\cong Ind^Q_RRel(H,N).
\end{equation}
\end{corollary}
\begin{remark}
Merely knowing that $\ll N \rr = \prod^{\ast}_{t\in T}N^t$ for some subset $T\subset G$ is not enough to guarantee \eqref{iso}. For example, let $G$ be any abelian group and let $H$ be a proper subgroup of $G$. Then for any subgroup $N$ of $H$, $\ll N \rr = N = \prod^{\ast}_{t\in\{1\}}N^t$. But $Rel(G,\ll N \rr)$ (resp. $Rel(H,N)$) is a $\mathbb{Z}Q$-module (resp. $\mathbb{Z}R$-module) with the trivial $Q$-action (resp. $R$-action) and thus $Rel(G,\ll N \rr)\not\cong Ind^Q_RRel(H,N)$.
\end{remark}
\paragraph{2.3 Organization of the paper.}
The strategy of this paper is to deal with the more general setting where $G$ has a family of weakly hyperbolically embedded subgroups and prove the Cohen-Lyndon property in this case. Theorem \ref{simmain} and Corollary \ref{module} are simple consequences of the general results. Necessary preliminaries are provided in Section \ref{prelim}, which is divided into four subsections. Section \ref{subsec.vk} recalls the definition and basic properties of van Kampen diagrams. Section \ref{subsec.wh} surveys the notion of (weakly) hyperbolically embedded subgroups. The proof of the Cohen-Lyndon property relies heavily on surgeries on van Kampen diagrams and geodesic polygons of Cayley graphs. Section \ref{subsec.ic} is devoted to a concept called isolated components, which is vital to surgeries, and Section \ref{sec.surgery} collects some results about surgeries on van Kampen diagrams. In Section \ref{sec.transversal}, we construct particular transversals (with the aid of Zorn's lemma). Section \ref{proof} states and proves the main theorem of this paper in the general case by using the transversals constructed in Section \ref{sec.transversal}. Finally, Section \ref{sec.rm} analyzes the structure of the relative relation modules of Dehn fillings.
\paragraph{Acknowledgement.} This work was done when I was a graduate student at Vanderbilt University. I would like to thank my supervisor, Professor Denis Osin, for the priceless discussions. This paper would not have been written without his help. I would also like to thank the anonymous referee for the useful comments on an early version of this paper and for showing me an alternative proof of Theorem \ref{simmain} using very rotating families.
\section{Preliminaries}\label{prelim}
This section contains a brief discussion of concepts and tools used to prove the main theorem. Our main reference is \cite{dahmani2017hyperbolically}.
\subsection{Van Kampen diagrams}\label{subsec.vk}
Let $G$ be a group given by the presentation
\begin{equation}\label{pre0}
G=\langle\mathcal{A}\mid\mathcal{R}\rangle,
\end{equation}
where $\mathcal{A}$ is a symmetric set of letters and $\mathcal{R}$ is a symmetric set of words in $\mathcal{A}$ (i.e., for every $w\in\mathcal{R}$, every cyclic shift of $w$ or $w^{-1}$ belongs to $\mathcal{R}$).
A \textit{van Kampen diagram} $\Delta$ over \eqref{pre0} is a finite oriented connected planar $2$-complex with labels on its oriented edges such that
\begin{enumerate}
\item[(a)] each oriented edge of $\Delta$ is labeled by a letter in $\mathcal{A}\cup\{1\}$;
\item[(b)] if an oriented edge $e$ of $\Delta$ has label $a\in\mathcal{A}\cup\{1\}$, then $e^{-1}$ has label $a^{-1}$, where $e^{-1}$ (resp. $a^{-1}$) is the inverse of $e$ (resp. $a$).
\end{enumerate}
Here, $1$ is identified with the empty word over $\mathcal{A}$ and thus $1=1^{-1}$. By convention, the empty word of $\mathcal{A}$ represents the identity of $G$.
Let $p=e_1\cdot\cdot\cdot e_k$ be a path in a van Kampen diagram over \eqref{pre0} or in the Cayley graph $\Gamma(G, \mathcal{A})$. The initial vertex (resp. terminal vertex) of $p$ is denoted as $p^-$ (resp. $p^+$). The \textit{label} of $p$, denoted as $Lab(p)$, is obtained by first concatenating the labels of the edges $e_1,...,e_k$ and then removing all $1$'s, as $1$ is identified with the empty word. Therefore, the label of a path in a van Kampen diagram is a word over $\mathcal{A}$. If $w$ is a word over $\mathcal{A}$, then the notation $Lab(p)\equiv w$ will indicate a letter-by-letter equality between $Lab(p)$ and $w$.
\begin{remark}\label{decompose paths on diagrams}
Suppose that $p$ is a path in a van Kampen diagram over \eqref{pre0} with $Lab(p)\equiv w_1\cdot\cdot\cdot w_k$. Then we can decompose $p$ in the following way: Let $p_{w_1}$ be the maximal subpath of $p$ such that $p^-_{w_1}=p^-$ and $Lab(p_{w_1})\equiv w_1$. For $i=2,...,k$, let $p_{w_i}$ be the maximal subpath of $p$ such that $p^-_{w_i}=p^+_{w_{i-1}}$ and $Lab(p_{w_i})\equiv w_i$.
\end{remark}
Edges labeled by letters from $\mathcal{A}$ are called \textit{essential edges}, while edges labeled by the letter $1$ are called \textit{non-essential edges}. A \textit{face} of $\Delta$ is a $2$-cell of $\Delta$. Let $\Pi$ be a face of $\Delta$, the boundary of $\Pi$ is denoted as $\partial\Pi$. Likewise, the boundary of $\Delta$ is denoted by $\partial\Delta$. Note that if we choose a base point for $\partial\Pi$ (resp. $\partial\Delta$), then $\partial\Pi$ (resp. $\partial\Delta$) becomes a path in $\Delta$. For a word $w$ over $\mathcal{A}$, we use the notation $Lab(\partial\Pi)\equiv w$ (resp. $Lab(\partial \Delta)\equiv w$) to indicate that one can pick a base point to turn $\partial\Pi$ (resp. $\partial\Delta$) into a path $p$ so that $Lab(p)\equiv w$.
\begin{remark}\label{decompose loops on diagrams}
Suppose that $\Delta$ is a diagram with $Lab(\partial\Delta)\equiv w_1\cdot\cdot\cdot w_k$. Then we can decompose $\partial\Delta$ in the following way: Let $p_b$ be vertex of $\partial\Delta$ such that when we use $p_b$ as the base point of $\partial\Delta$, we can turn $\partial\Delta$ into a path $p$ with $Lab(p)\equiv w_1\cdot\cdot\cdot w_k$. And then we use Remark \ref{decompose paths on diagrams} to decompose $p$ and thus decompose $\partial\Delta$.
\end{remark}
Consider the following additional assumption on van Kampen diagrams:
\begin{enumerate}
\item[(c)] For every face $\Pi$ of a van Kampen diagram $\Delta$ over the presentation $\eqref{pre0}$, at least one of the following conditions (c1) and (c2) holds:
\item[(c1)] $Lab(\partial\Pi)$ is equal to an element of $\mathcal{R}$.
\item[(c2)] $\partial\Pi$ either consists entirely of non-essential edges or consists of exactly two essential edges with mutually inverse labels (in addition to non-essential edges).
\end{enumerate}
A face satisfying (c$_2$) is called a \textit{non-essential face}. All other faces are called \textit{essential faces}. The process of adding non-essential faces to a van Kampen diagram is called a \textit{refinement}. Figure \ref{0refine} illustrates a refinement on a van Kampen diagram, where the unlabeled edges are labeled by $1$. The interested readers are referred to \cite{ol2012geometry} for a formal discussion. By using refinements, we can ensure that
\begin{enumerate}
\item[(d)] Every face is homeomorphic to a disc, i.e., its boundary has no self-intersection.
\end{enumerate}
\begin{figure}
{\input{0refine.pdf_tex}}
\caption{A refinement of a van Kampen diagram over the presentation $G=\langle a,b\mid aba^{-1}b^{-1}=1\rangle$}\label{0refine}
\end{figure}
\begin{assumption}
In the sequel, the above assumptions (c) and (d) will be imposed on van Kampen diagrams.
\end{assumption}
The well-known van Kampen lemma states that a word $w$ over $\mathcal{A}$ represents $1$ in $G$ if and only if there is a van Kampen diagram $\Delta$ over \eqref{pre0} such that $\Delta$ is homeomorphic to a disc (such diagrams are called \textit{disk diagrams}), and that $Lab(\partial\Delta)\equiv w$.
\begin{remark}\label{mapofvk}
If a van Kampen diagram $\Delta$ is homeomorphic to a disc, and $O$ is a vertex of $\Delta$, then there exists a unique continuous map $\mu$ from the $1$-skeleton of $\Delta$ to $\Gamma(G,\mathcal{A})$ sending $O$ to the identity vertex, preserving the labels of the essential edges and collapsing non-essential edges to points.
\end{remark}
\subsection{Hyperbolically embedded subgroups and group theoretic Dehn fillings}\label{subsec.wh}
Let $G$ be a group, let $\{H_\lambda\}_{\lambda\in\Lambda}$ be a family of subgroups of $G$, let $X$ be a subset of $G$ such that $G$ is generated by $X$ together with the union of all $H_{\lambda},\lambda\in\Lambda$, and let $\mathcal{H}=\bigsqcup_{\lambda\in\Lambda}H_{\lambda}$. Consider the Cayley graph $\Gamma(G,X\sqcup \mathcal{H})$.
\begin{remark}
It is possible that $X$ and $H_{\lambda},\lambda\in\Lambda$, as subsets of $G$, have non-empty intersections with each other. As a consequence, several letters of $X\sqcup \mathcal{H}$ might represent the same element of $G$. If this is the case, the Cayley graph $\Gamma(G,X\sqcup \mathcal{H})$ will have multiple edges corresponding to those letters.
\end{remark}
\noindent
\textbf{Notation.}
Let $w$ be a word over the alphabet $X\sqcup \mathcal{H}$. The inverse of $w$ is denoted by $w^{-1}$. The \textit{length} of $w$, denoted as $\|w\|$, is the number of letters in $w$. We identify $w$ with the element of $G$ represented by $w$. Thus, if $S$ is a subset of $G$, then we write $w\in S$ to indicate that $w$ represents an element of $S$.
There are two types of equalities for words over $X\sqcup \mathcal{H}$. Given two words $u$ and $v$ over $X\sqcup \mathcal{H}$, the notation $u\equiv v$ indicates the letter-by-letter equality between $u$ and $v$ and the notation $u=_G v$ indicates that $u$ and $v$ represent the same element of $G$.
If $g$ is an element of $G$, then $|g|$ denotes the length of the shortest word in $X\sqcup \mathcal{H}$ representing $g$.
Note that, for each $\lambda\in\Lambda$, the Cayley graph $\Gamma(H_{\lambda},H_{\lambda})$ can be identified as the complete subgraph of $\Gamma(G,X\sqcup \mathcal{H})$ whose vertex set is $H_{\lambda}$, and edges are the ones labeled by letters from $H_{\lambda}$.
\begin{definition}
Fix $\lambda\in\Lambda$. A (combinatorial) path $p$ in $\Gamma(G,X\sqcup \mathcal{H})$ between vertices of $\Gamma(H_{\lambda},H_{\lambda})$ is called $H_{\lambda}$\textit{-admissible} if it does not contain any edge of $\Gamma(H_{\lambda},H_{\lambda})$. Note that a $H_{\lambda}$-admissible path $p$ is allowed to pass through vertices of $\Gamma(H_{\lambda},H_{\lambda})$. For every pair of elements $h,k\in H_{\lambda}$, let $\widehat{d}_{\lambda}(h,k)\in[0,+\infty]$ be the length of a shortest $H_{\lambda}$-admissible path connecting $h,k$. If no such path exists, set $\widehat{d}_{\lambda}(h,k)=+\infty$. The laws of summation on $[0,+\infty)$ extend naturally to $[0,+\infty]$ and it is easy to verify that $\widehat{d}_{\lambda}:H_{\lambda}\times H_{\lambda}\rightarrow [0,+\infty]$ defines a metric on $\Gamma(H_{\lambda},H_{\lambda})$ called the \textit{relative metric on} $\Gamma(H_{\lambda},H_{\lambda})$ \textit{with respect to} $X$.
\end{definition}
\begin{remark}
Let $p$ be a path in $\Gamma(G,X\sqcup \mathcal{H})$ with $Lab(p)\equiv h\in H_{\lambda}$, for some $\lambda\in\Lambda$. For simplicity, we denote $\widehat{d}_{\lambda}(1,h)$ by $\widehat{\ell}_{\lambda}(p)$.
\end{remark}
\begin{definition}\label{hyperbolically embedded}
Let $G$ be a group, let $\{H_\lambda\}_{\lambda\in\Lambda}$ be a family of subgroups of $G$, let $X$ be a subset of $G$, and let $\mathcal{H}=\bigsqcup_{\lambda\in\Lambda}H_{\lambda}$. We say that $\{H_\lambda\}_{\lambda\in\Lambda}$ \textit{weakly hyperbolically embeds into} $(G,X)$ (denoted as $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$) if $G$ is generated by the set $X$ together with the union of all $H_{\lambda},\lambda\in\Lambda$, and the Cayley graph $\Gamma(G,X\sqcup \mathcal{H})$ is a Gromov hyperbolic space.
If the collection $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$ and for each $\lambda\in\Lambda$, the metric space $(H_{\lambda},\widehat{d}_{\lambda})$ is proper, i.e., every ball of finite radius contains only finitely many elements, then we say that $\{H_\lambda\}_{\lambda\in\Lambda}$ \textit{hyperbolically embeds into} $(G,X)$ (denoted as $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_h (G,X)$).
Further, the collection $\{H_\lambda\}_{\lambda\in\Lambda}$ \textit{hyperbolically embeds into} $G$, denoted as $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_h G$, if there exists some subset $X\subset G$ such that $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_h(G,X)$.
\end{definition}
\begin{remark}
Note that if the family $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$ for some subset $X\subset G$ and $Y=X\cup X^{-1}$, then we also have $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,Y)$. In the sequel, we will always assume that the relative generating set $X$ is symmetric, i.e., $X=X^{-1}$.
\end{remark}
\begin{definition}
Suppose $G$ is a group with a family of subgroups $\{H_{\lambda}\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$ for some subset $X\subset G$. For $\lambda\in\Lambda$, let $\widehat{d}_{\lambda}$ be the relative metric on $\Gamma(H_{\lambda},H_{\lambda})$ with respect to $X$. We say that a property $P$ \textit{holds for all sufficiently deep Dehn fillings of} $\{H_{\lambda}\}_{\lambda\in\Lambda}$ (or \textit{for all sufficiently deep} $N_{\lambda}\lhd H_{\lambda},\lambda\in\Lambda,$) if there exists a number $C>0$ such that if $N_{\lambda}\lhd H_{\lambda}$ and $\widehat{d}_{\lambda}(1,n)>C$ for all $n\in N_{\lambda}\setminus\{1\},\lambda\in\Lambda$, then $P$ holds.
\end{definition}
One remarkable property of weakly hyperbolically embedded subgroups is the following group theoretic Dehn filling theorem.
\begin{theorem}[{\cite[Theorem 7.15]{dahmani2017hyperbolically}}]\label{Dehn filling}
Let $G$ be a group with a family of subgroups $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$ for some subset $X\subset G$. Then for all sufficiently deep $N_{\lambda}\lhd H_{\lambda},\lambda\in\Lambda$, we have:
\begin{enumerate}
\item[(a)] For each $\lambda\in\Lambda$, the natural homomorphism $i_{\lambda}:H_{\lambda}/N_{\lambda}\rightarrow G/\ll \mathcal{N} \rr$ is injective (i.e., $H_{\lambda}\cap \ll \mathcal{N} \rr=N_{\lambda}\left.\right)$, where $\mathcal{N}=\bigcup_{\lambda\in\Lambda}N_{\lambda}$.
\item[(b)] $\{i_{\lambda}(H_{\lambda}/N_{\lambda})\}_{\lambda\in\Lambda}\hookrightarrow_{wh}G/\ll \mathcal{N} \rr$.
\item[(c)] There exist subsets $T_{\lambda}\subset G,\lambda\in\Lambda$, such that $\ll \mathcal{N} \rr = \prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}} N^t_{\lambda}$.
\end{enumerate}
\end{theorem}
As mentioned in the introduction, we are going to improve part (c) of the above theorem. To simplify statements, we introduce the following terminologies.
\begin{definition}
Let $G$ be a group with a subgroup $H$. The \textit{collection of left transversals of} $H$ \textit{in} $G$ is denoted by $LT(H,G)$, i.e., every element of $LT(H,G)$ is a left transversal of $H$ in $G$.
\end{definition}
\begin{definition}\label{CLp}
Let $G$ be a group, let $\{H_{\lambda}\}_{\lambda\in\Lambda}$ be a family of subgroups of $G$, and let $N_{\lambda}$ be a normal subgroup of $H_{\lambda}$ for every $\lambda\in\Lambda$. We say that the triple $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ has the \textit{Cohen-Lyndon property} if there exists a left transversal $T_{\lambda}\in LT(H_{\lambda}\ll \mathcal{N} \rr,G)$ for every $\lambda\in\Lambda$ such that $\ll \mathcal{N} \rr = \prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}} N^t_{\lambda}$, where $\mathcal{N}=\bigcup_{\lambda\in\Lambda}N_{\lambda}$.
\end{definition}
\subsection{Isolated components}\label{subsec.ic}
Let us assume, until the end of Section \ref{proof}, that $G$ is a group with a family of subgroups $\{H_\lambda\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$ for some symmetric subset $X\subset G$. For each $\lambda\in\Lambda$, let $\widehat{d}_{\lambda}$ be the relative metric on $\Gamma(H_{\lambda},H_{\lambda})$ with respect to $X$, and let $\mathcal{H}=\bigsqcup_{\lambda\in\Lambda}H_{\lambda}$. The following terminology goes back to \cite{osin2006relatively}.
\begin{definition}\label{connect}
Let $p$ be a path in $\Gamma(G,X\sqcup \mathcal{H})$. Fix $\lambda\in\Lambda$. An $H_{\lambda}$\textit{-subpath} $q$ of $p$ is a nontrivial subpath of $p$ labeled by a word over the alphabet $H_{\lambda}$ (if $p$ is a cycle, we allow $q$ to be a subpath of some cyclic shift of $p$). An $H_{\lambda}$-subpath $q$ of $p$ is called an $H_{\lambda}$-\textit{component} if $q$ is not properly contained in any other $H_{\lambda}$-subpath. Two $H_{\lambda}$-components $q_1,q_2$ of $p$ are called \textit{connected} if there exists a path $c$ in $\Gamma(G,X\sqcup \mathcal{H})$ such that $c$ connects a vertex of $q_1$ to a vertex of $q_2$, and that $Lab(c)$ is a letter from $H_{\lambda}$. An $H_{\lambda}$-component $q$ of $p$ is called \textit{isolated} if it is not connected to any other $H_{\lambda}$-component of $p$.
\end{definition}
The key property of isolated components is that, in a \textit{geodesic polygon} (i.e., a polygon in $\Gamma(G,X\sqcup \mathcal{H})$ with geodesic sides) $p$, the total $\widehat{\ell}$-length of isolated components is uniformly bounded above by a linear function of the number of sides. The following result is proved in \cite[Proposition 4.14]{dahmani2017hyperbolically}, which is a straightforward generalization of \cite[Proposition 3.2]{osin2007peripheral}.
\begin{lemma}[Dahmani-Guirardel-Osin]\label{length}
There exists a positive number $D$ satisfying the following property: Let $p$ be an $n$-gon in $\Gamma(G,X\sqcup \mathcal{H})$ with geodesic sides $p_1,...,p_n$ and let $I$ be a subset of the set of sides of $p$ such that every side $p_i\in I$ is an isolated $H_{\lambda_i}$-component of $p$ for some $\lambda_i\in\Lambda$. Then
$$\sum_{p_i\in I}\widehat{\ell}_{\lambda_i}(p_i)\leqslant Dn.$$
\end{lemma}
\begin{remark}\label{constant}
Theorem \ref{Dehn filling} asserts the existence of a constant $C$ such that if $\widehat{d}_{\lambda}(1,n)\geqslant C$ for every $n\in N_{\lambda}\setminus\{1\}$ and $\lambda\in\Lambda$, then $H_{\lambda}\cap \ll\mathcal{N}\rr=N_{\lambda}$ for all $\lambda\in\Lambda$. In fact, one can let $C=4D$, where $D$ is the constant provided by Lemma \ref{length} (see \cite{dahmani2017hyperbolically}).
\end{remark}
\subsection{Diagram surgery}\label{sec.surgery}
The diagram surgery surveyed in this section was first introduced by Osin in \cite{osin2007peripheral}, where he proved a group theoretic Dehn filling theorem for relatively hyperbolic groups. Later, Dahmani et al. generalized this technique to deal with weakly hyperbolically embedded subgroups \cite{dahmani2017hyperbolically}.
Consider a symmetric set $\mathcal{R}$ of words over the alphabet $X\sqcup \mathcal{H}$ such that $G$ has the presentation
\begin{equation}\label{pre1}
G=\langle X\sqcup \mathcal{H}\mid\mathcal{R}\rangle,
\end{equation}
and that for all $\lambda\in\Lambda$, $\mathcal{R}$ contains all words over the alphabet $H_{\lambda}$ which represent the identity.
Suppose that $N_{\lambda}$ is a normal subgroup of $H_{\lambda}$ for each $\lambda\in\Lambda$. Denote the union of $N_{\lambda},\lambda\in\Lambda$, by $\mathcal{N}$. Killing $\ll \mathcal{N} \rr$ in $G$ is equivalent to adding, to $\mathcal{R}$, all words over $H_{\lambda}$ which represent elements of $N_{\lambda}$, for all $\lambda\in\Lambda$, to form a new presentation
\begin{equation}\label{pre}
Q=G/\ll\mathcal{N}\rr=\langle X\sqcup \mathcal{H},\mathcal{R}\cup \mathcal{S}\rangle,
\end{equation}
where $\mathcal{S}=\bigcup_{\lambda\in\lambda}\mathcal{S}_{\lambda}$ and $\mathcal{S}_{\lambda}$ consists of all words over $H_{\lambda}$ representing elements of $N_{\lambda}$ in $G$.
In the sequel, let $\mathcal{D}$ be the set of all van Kampen diagrams $\Delta$ over \eqref{pre} satisfying the following.
\begin{enumerate}
\item[(D1)] Topologically $\Delta$ is a disc with $k\geqslant 0$ holes. The boundary of $\Delta$ can be decomposed as $\partial \Delta=\partial_{ext}\Delta\cup\partial_{int}\Delta$, where $\partial_{ext}\Delta$ is the boundary of the disc, and $\partial_{int}\Delta$ consists of disjoint cycles (connected components) $c_1,...,c_k$ that bound the holes.
\item[(D2)] For $i=1,...,k$, $c_i$ is labeled by a word from $\mathcal{S}$.
\item[(D3)] Each diagram $\Delta$ is equipped with a \textit{cut system} that is a collection $T=\{t_1,...,t_k\}$ of disjoint paths (\textit{cuts}) $t_1,...,t_k$ in $\Delta$ without self-intersections such that, for $i=1,...,k$, the two endpoints of $t_i$ belong to $\partial\Delta$, and that after cutting $\Delta$ along $t_i$ for all $i = 1,...,k$, one gets a disc van Kampen diagram $\widetilde{\Delta}$ over \eqref{pre1}.
\end{enumerate}
See Figure \ref{findcut} for an illustration of a diagram in $\mathcal{D}$.
\begin{lemma}
A word $w$ over $X\sqcup \mathcal{H}$ represents $1$ in $Q$ if and only if there is a diagram $\Delta\in \mathcal{D}$ such that $Lab(\partial_{ext}\Delta)\equiv w$.
\end{lemma}
\begin{proof}
Let $w$ be a word over $X\sqcup \mathcal{H}$. If there is a diagram $\Delta\in \mathcal{D}$ such that $\partial_{ext}\Delta \equiv w$, by filling the holes of $\Delta$ with faces whose boundaries are labeled by words from $\mathcal{S}$, one creates a disc van Kampen diagram over \eqref{pre1}, whose boundary is labeled by $w$. Conversely, if $w$ represents $1$ in $Q$, then there exists a disc van Kampen diagram $\overline{\Delta}$ over \eqref{pre1} with $Lab(\partial\overline{\Delta})\equiv w$. By removing all faces of $\overline{\Delta}$ labeled by words from $\mathcal{S}$, we obtain a diagram $\Delta'$ satisfying (D1) and (D2). To produce a cut system, choose a vertex $O$ in $\partial_{ext}\Delta'$. Connect $O$ with each component of $\partial_{int}\Delta'$ by a path so that these paths do not cross each other (although they do intersect each other). By passing to a refinement of $\Delta'$, one can separate these paths so that they no longer intersect each other and thus creates a diagram $\Delta$ satisfying (D1), (D2), and (D3) with $Lab(\partial_{ext}\Delta)\equiv w$.
\end{proof}
Figure \ref{findcut} illustrates the last step of the above proof. The left half shows the diagram $\Delta'$ where a dash path and a dot-dash path connect $O$ with two components of $\partial_{int}\Delta'$. By thickening these paths with a refinement, we obtain the right half. The regions shaded by horizontal and vertical lines consist of non-essential faces, while the paths from $O_1$ to $U_1$ and from $O_3$ to $V_2$ form a cut system.
\begin{figure}
{\input{findcut.pdf_tex}}
\caption{How to produce a cut system}\label{findcut}
\end{figure}
Let $\Delta$ be a diagram in $\mathcal{D}$ and let $\widetilde{\Delta}$ be the disc van Kampen diagram resulting from cutting $\Delta$ along its set of cuts. Define $\kappa:\widetilde{\Delta}\rightarrow\Delta$ to be the map that ``sews'' the cuts. Fix an arbitrary vertex $O$ in $\widetilde{\Delta}$ and let $\mu$ be a map sending the $1$-skeleton of $\Delta$ to $\Gamma(G,X\sqcup \mathcal{H})$, as described by Remark \ref{mapofvk}.
\begin{definition}
Let $\Delta_1$ and $\Delta_2$ be two diagrams of $\mathcal{D}$ and let $\Gamma_1$ (resp. $\Gamma_2\left.\right)$ be the subgraph of the $1$-skeleton of $\Delta_1$ (resp. $\Delta_2\left.\right)$ consisting of $\partial\Delta_1$ (resp. $\partial\Delta_2\left.\right)$ and all cuts of $\Delta_1$ (resp. $\Delta_2\left.\right)$. We say that $\Delta_1$ and $\Delta_2$ are \textit{equivalent} if there exists a graph isomorphism $\Gamma_1\rightarrow\Gamma_2$ which preserves labels and orientations of edges, and maps the cuts and boundary of $\Delta_1$ to the cuts and boundary of $\Delta_2$, respectively.
\end{definition}
The following Lemmas \ref{adding path} and \ref{cut} are results from \cite{dahmani2017hyperbolically}, which are straightforward generalizations of results of \cite{osin2007peripheral}. Note that the authors of \cite{dahmani2017hyperbolically} assume that the presentation \eqref{pre1} has a linear relative isoperimetric function, but this assumption is not used in the proofs of those lemmas.
\begin{lemma}[{\cite[Lemma 7.11]{dahmani2017hyperbolically}}]\label{adding path}
Let $a,b$ be two vertices on $\partial\Delta$ and let $\widetilde{a},\widetilde{b}$ be two vertices on $\partial\widetilde{\Delta}$ such that $\kappa(\widetilde{a})=a,\kappa(\widetilde{b})=b$. Then for any path $p$ in $\Gamma(G,X\sqcup \mathcal{H})$ connecting $\mu(\widetilde{a})$ to $\mu(\widetilde{b})$, there is a diagram $\Delta_1\in\mathcal{D}$ with the following properties:
\begin{enumerate}
\item[(a)] $\Delta$ and $\Delta_1$ are equivalent.
\item[(b)] There is a path $q$ in $\Delta_1$ without self-intersections such that (1) $q$ connects $a$ and $b$, (2) $q$ has no common vertices with the cuts of $\Delta_1$ except possibly for $a,b$, and (3) $Lab(q)\equiv Lab(p)$.
\end{enumerate}
\end{lemma}
\begin{definition}
Fix $\lambda\in\Lambda$. An $H_{\lambda}$\textit{-subpath} in $\partial\Delta$ (resp. $\partial\widetilde{\Delta}$) for some $\Delta\in \mathcal{D}$ is a path labeled by a nontrivial word over $H_{\lambda}$. An $H_{\lambda}$-subpath $p$ of $\partial\Delta$ (resp. $\partial\widetilde{\Delta}$) is called an $H_{\lambda}$\textit{-component} if $p$ is not properly contained in any other $H_{\lambda}$-subpath. Two $H_{\lambda}$-components $p,q$ of $\partial\Delta$ are \textit{connected} if there exist $H_{\lambda}$-components $a,b$ in $\partial\widetilde{\Delta}$ such that $\kappa(a)$ (resp. $\kappa(b)$) is a subpath of $p$ (resp. $q$), and that $\mu(a),\mu(b)$ are connected in $\Gamma(G,X\sqcup \mathcal{H})$ (in the sense of Definition \ref{connect}).
\end{definition}
\begin{remark}
The definitions of $H_{\lambda}$-subpaths, $H_{\lambda}$-components, and connected $H_{\lambda}$-components in $\partial\Delta$ for a van Kampen diagram $\Delta\in\mathcal{D}$ or $\partial\widetilde{\Delta}$ do not depend on the pre-chosen vertex $O$.
\end{remark}
\begin{definition}\label{type of diagrams}
The \textit{type} of $\Delta$ is defined by the formula
$$\tau(\Delta)=(k,\sum_{i=1}^k\|Lab(t_i)\|),$$
where $k$ is the number of holes in $\Delta$ and $t_1,...,t_k$ are the cuts. We order the types of diagrams in $\mathcal{D}$ lexicographically: $(k_1,\ell_1)<(k_2,\ell_2)$ if and only if either $k_1<k_2$ or $k_1=k_2$ and $\ell_1<\ell_2$.
\end{definition}
\begin{definition}
A word $w$ over $X\sqcup\mathcal{H}$ is called \textit{geodesic} if it labels a geodesic in $\Gamma(G,X\sqcup\mathcal{H})$.
\end{definition}
\begin{definition}
For any word $w$ over $X\sqcup\mathcal{H}$, let $\mathcal{D}(w)$ be the set of diagrams $\Delta\in\mathcal{D}$ such that $Lab(\partial_{ext}\Delta)\equiv w$.
\end{definition}
\begin{lemma}[{\cite[Lemma 7.17]{dahmani2017hyperbolically} (see also \cite[Lemma 5.2]{osin2007peripheral})}]\label{cut}
Suppose that for every $\lambda\in\Lambda$ and $n\in N_{\lambda}\setminus\{1\}$, we have $\widehat{d}_{\lambda}(1,n)>4D$, where $D$ is the constant given by Lemma \ref{length}. Let $w$ be a geodesic word over $X\sqcup \mathcal{H}$ representing $1$ in $Q$, and let $\Delta$ be a diagram in $\mathcal{D}(w)$ of minimal type. Then there exists $\lambda\in\Lambda$ and a connected component $c$ of $\partial_{int}\Delta$ such that $c$ is connected to an $H_{\lambda}$-component of $\partial_{ext}\Delta$.
\end{lemma}
\section{Construction of the transversals}\label{sec.transversal}
The proof of Theorem \ref{main} relies on constructing a particular left transversal $T_{\lambda}\in LT(H_{\lambda}\ll\mathcal{N}\rr,G)$ for each $\lambda\in\Lambda$. It is convenient to construct a collection $\{T_{\lambda}\}_{\lambda\in\Lambda}$ of sets of words over $X\sqcup \mathcal{H}$ satisfying the following properties (P1) through (P3), and think of $T_{\lambda}$ as a transversal in $LT(H_{\lambda}\ll\mathcal{N}\rr,G)$ (identifying words over $X\sqcup\mathcal{H}$ and the elements of $G$ represented by those words) for $\lambda\in\Lambda$. Recall that $\|w\|$ is the length of $w$ for a word $w$ over $X\sqcup H$, and that $|g|$ denotes the length of a geodesic word over $X\sqcup\mathcal{H}$ representing an element $g\in G$.
\begin{enumerate}
\item[(P1)] [$\{T_{\lambda}\}_{\lambda\in\Lambda}$ is transversal] For each $\lambda\in\Lambda$, $T_{\lambda}\in LT(H_{\lambda}\ll\mathcal{N}\rr,G)$.
\item[(P2)] [$\{T_{\lambda}\}_{\lambda\in\Lambda}$ is geodesic] If $w\in T_{\lambda}$ for some $\lambda\in\Lambda$, and $gH_{\lambda}\ll\mathcal{N}\rr=wH_{\lambda}\ll\mathcal{N}\rr$ for some $g\in G$, then $\|w\|\leqslant |g|$. This implies that, for all $\lambda\in\Lambda$, every $w\in T_{\lambda}$ is a geodesic word over $X\sqcup\mathcal{H}$.
\item[(P3)] [$\{T_{\lambda}\}_{\lambda\in\Lambda}$ is prefix closed] Let $\lambda,\mu\in\Lambda$. If a word $w\in T_{\lambda}$ can be decomposed as $w\equiv uhv$ with $h\in H_{\mu}\setminus\{1\}$ ($u,v$ are allowed to be empty words), then $u\in T_{\mu}$ and $\widehat{d}_{\mu}(1,h)\leqslant \widehat{d}_{\mu}(1,h')$ for all $h'\in hN_{\mu}$.
\end{enumerate}
\begin{lemma}\label{transversal}
There exists a collection $\{T_{\lambda}\}_{\lambda\in\Lambda}$ satisfying (P1), (P2), and (P3).
\end{lemma}
\begin{proof}
Let $\mathcal{W}$ be the poset of collections $\{W_{\lambda}\}_{\lambda\in\Lambda}$ of words satisfying (P2) and (P3), while instead of (P1), we only demand that the words of $W_{\lambda}$ represent a subset of a transversal in $LT(H_{\lambda}\ll\mathcal{N}\rr,G)$ for every $\lambda\in\Lambda$. We order $\mathcal{W}$ by index-wise inclusion, i.e., $\{U_{\lambda}\}_{\lambda\in\Lambda}$ is less than $\{V_{\lambda}\}_{\lambda\in\Lambda}$ if and only if $U_{\lambda}\subset V_{\lambda}$ for every $\lambda\in\Lambda$. $\mathcal{W}$ is non-empty because the collection $\{W_{\lambda}\}_{\lambda\in\Lambda}$ with each $W_{\lambda}$ consisting of only the empty word is a member of $\mathcal{W}$. Moreover, the union of any chain of $\mathcal{W}$ is again a member of $\mathcal{W}$. Therefore, Zorn's lemma implies that $\mathcal{W}$ has a maximal member $\{T_{\lambda}\}_{\lambda\in\Lambda}$. Suppose that $\{T_{\lambda}\}_{\lambda\in\Lambda}$ does not satisfy (P1), i.e., there exist $\lambda_0\in\Lambda$ and $g\in G$ such that no element of the coset $gH_{\lambda_0}M$ is represented by a word in $T_{\lambda_0}$. Without loss of generality, let us assume that if $g'$ is an element of $G$ such that $|g'|<|g|$, then for each $\lambda\in\Lambda$, $g'H_{\lambda}\ll \mathcal{N} \rr\cap T_{\lambda}\neq \emptyset$.
Let $w$ be a geodesic word over $X\sqcup \mathcal{H}$ representing $g$. Consider the collection $\{U_{\lambda}\}_{\lambda\in\Lambda}$ constructed as follows. For every $\lambda\in\Lambda\setminus\{\lambda_0\}$, let $U_{\lambda}=T_{\lambda}$, and construct $U_{\lambda_0}$ by the following manner: If $w$ contains no letter from $\mathcal{H}$, let $U_{\lambda_0}=T_{\lambda_0}\cup\{w\}$. If $w$ contains at least one letter from $\mathcal{H}$, then $w$ can be decomposed as $w\equiv uhv$ such that $h\in H_{\lambda}\setminus\{1\}$ for some $\lambda\in\Lambda$ and $v$ contains no letter from $\mathcal{H}$ ($u,v$ are allowed to be empty words). As $\|u\|<\|w\|=|g|$, there exists a word $u'\in T_{\lambda}$ such that $u'\in uH_{\lambda}\ll \mathcal{N} \rr$. Let $h'$ be an element of $H_{\lambda}$ such that $u\ll \mathcal{N} \rr=u'h'\ll \mathcal{N} \rr$ and let $h''$ be an element of $H_{\lambda}$ such that (a) $h''N_{\lambda}=h'hN_{\lambda}$ and (b) if $k\in h''N_{\lambda}$, then $\widehat{d}_{\lambda}(1,h'')\leqslant\widehat{d}_{\lambda}(1,k)$. Set $U_{\lambda_0}=T_{\lambda_0}\cup\{u'h''v\}$.
It is straight-forward to verify that $\{U_{\lambda}\}_{\lambda\in\Lambda}$ is an element of $\mathcal{W}$. There is a word in $U_{\lambda_0}$ representing an element in $gH_{\lambda_0}\ll \mathcal{N} \rr$, while $T_{\lambda_0}$ has no such words. It follows that $\{U_{\lambda}\}_{\lambda\in\Lambda}$ is strictly greater than $\{T_{\lambda}\}_{\lambda\in\Lambda}$, contradicting the choice of $\{T_{\lambda}\}_{\lambda\in\Lambda}$.
\end{proof}
\section{Main theorem and its proof}\label{proof}
The following Theorem \ref{main} is the main theorem of this paper. We first use it to prove Theorem \ref{simmain}, and then we prove Theorem \ref{main}.
\begin{theorem}\label{main}
Let $G$ be a group with a family of subgroups $\{H_{\lambda}\}_{\lambda\in\Lambda}\hookrightarrow_{wh}(G,X)$ for some $X\subset G$. Then the Cohen-Lyndon property holds for all sufficiently deep Dehn fillings of $\{H_{\lambda}\}_{\lambda\in\Lambda}$.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{simmain}]
By assumption, $H\hookrightarrow_h (G,X)$ for some subset $X\subset G$. Let $\widehat{d}$ be the relative metric on $\Gamma(H,H)$ with respect to $X$. Theorem \ref{main} provides a constant $C$ such that if $N\lhd H$ and $\widehat{d}(n)>C$ for all $n\in N\setminus\{1\}$, then $(G,H,N)$ possesses the Cohen-Lyndon property. As $H\hookrightarrow_h (G,X)$, $\widehat{d}$ is locally finite. In particular,
$$\mathcal{F}=\{h\in H\setminus\{1\}\mid \widehat{d}(h)\leqslant C\}$$
is a finite set. The desired result follows by noting that if $N\lhd H$ and $N\cap \mathcal{F}=\emptyset$, then $(G,H,N)$ possesses the Cohen-Lyndon property.
\end{proof}
Let us prove Theorem \ref{main}. Recall that Lemma \ref{length} provides a number $D>0$ to estimate the total length of isolated components in a geodesic polygon, and that Theorem \ref{Dehn filling} and Remark \ref{constant} implies that if $\widehat{d}_{\lambda}(1,n)\geqslant 4D$ for every $n\in N_{\lambda}\setminus\{1\}$ and $\lambda\in\Lambda$, then $H_{\lambda}\cap \ll\mathcal{N}\rr=N_{\lambda}$ for all $\lambda\in\Lambda$. We assume the following condition.
\begin{enumerate}
\item[(24D)] $\widehat{d}_{\lambda}(1,n)>24D$ for all $n\in N_{\lambda}\setminus\{1\}$ and $\lambda\in\Lambda$.
\end{enumerate}
We prove that (24D) implies the Cohen-Lyndon property of $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$. Let $\{T_{\lambda}\}_{\lambda\in\Lambda}$ be a collection of words over $X\sqcup\mathcal{H}$ satisfying (P1), (P2), and (P3) (by Lemma \ref{transversal}, such a collection exists) and think of each $T_{\lambda}$ as a left transversal in $LT(H_{\lambda}\ll\mathcal{N}\rr,G)$. For every $\lambda\in\Lambda$, we extend $T_{\lambda}$ to a set $T^{ex}_{\lambda}$. Roughly speaking, $T^{ex}_{\lambda}$ is the set of words obtained from $T_{\lambda}$ by replacing letters from $H_{\lambda}$ with other letters from the same coset of $N_{\lambda}$ in $H_{\lambda}$.
\begin{definition}\label{extended T}
For every $\lambda\in\Lambda$, let $T^{ex}_{\lambda}$ be the set of words with the following property: Every word $w\in T^{ex}_{\lambda}$ admits a decomposition $w\equiv w_1h_1\cdot\cdot\cdot w_kh_kw_{k+1}$ ($w_1,...,w_{k+1}$ are allowed to be empty words) such that for every $i\in \{1,...,k\}$, there exists $\lambda_i\in\Lambda$ with the following properties.
\begin{enumerate}
\item[(a)] For $i=1,...,k$, $h_i$ is an element of $H_{\lambda_i}$ ($h_i$ is allowed to equal $1$).
\item[(b)] There exists an element $h'_i\in H_{\lambda_i}\setminus\{1\}$ such that $h'_iN_{\lambda_i}=h_iN_{\lambda_i}$ for $i=1,...,k$, and that the concatenation $w_1h'_1\cdot\cdot\cdot w_kh'_kw_{k+1}$ is a word in $T_{\lambda}$.
\end{enumerate}
\end{definition}
\begin{remark}
If $k=0$ in the above definition, conditions (a) and (b) will be satisfied trivially. Thus, $T_{\lambda}$ is a subset of $T^{ex}_{\lambda}$ for every $\lambda\in\Lambda$.
\end{remark}
\begin{definition}
Let $w$ be a word over $X\sqcup\mathcal{H}$ and let $\lambda\in\Lambda$. If $w\in T^{ex}_{\lambda}$, let $rank_{\lambda}(w)$ be the minimal number $k$ obtained from the decompositions $w\equiv w_1h_1\cdot\cdot\cdot w_kh_kw_{k+1}$ satisfying Definition \ref{extended T}. If $w\not\in T_{\lambda}$, let $rank_{\lambda}(w)=\infty$.
For every word $w$ over $X\sqcup\mathcal{H}$, the \textit{rank} of $w$, denoted as $rank(w)$, is the number $\min_{\lambda\in\Lambda}\{rank_{\lambda}(w)\}$.
\end{definition}
\begin{lemma}\label{u4}
Let $w$ be a word in $T^{ex}_{\lambda}$ for some $\lambda\in\Lambda$. Suppose that $w$ can be decomposed as $w\equiv uhv$ with $h\in H_{\mu}\setminus\{1\}$ for some $\mu\in\Lambda$. Let $h''$ be an element of $H_{\mu}$ such that $h''N_{\mu}=hN_{\mu}$. Then $uh''v\in T^{ex}_{\lambda}$.
\end{lemma}
\begin{proof}
Let $w\equiv w_1h_1\cdot\cdot\cdot w_kh_kw_{k+1}$ be a decomposition satisfying Definition \ref{extended T} and let $h'_1,...,h'_k$ be as in (b) of Definition \ref{extended T}.
Without loss of generality, we may assume that $h=h_i$ for some number $i\in \{1,...,k\}$. Then $uh''v$ can be decomposed as
$$uh''v\equiv w_1h_1\cdot\cdot\cdot w_{i-1}h_{i-1}w_ih''w_{i+1}h_{i+1}w_{i+2}h_{i+2}\cdot\cdot\cdot w_kh_kw_{k+1}.$$
By replacing $h_j$ with $h'_j$ for $j\neq i$ and $h''$ with $h'_i$, we obtain a word in $T_{\lambda}$ and thus $uh''v\in T^{ex}_{\lambda}$.
\end{proof}
\begin{lemma}\label{trunc}
Let $w$ be a word in $T^{ex}_{\lambda}$ for some $\lambda\in\Lambda$ with a decomposition $w\equiv w_1h_1\cdot\cdot\cdot w_kh_kw_{k+1}$ satisfying Definition \ref{extended T}. Then $w_1\in T_{\lambda_1}$.
\end{lemma}
\begin{proof}
Let $h'_1,...,h'_k$ be as in (b) of Definition \ref{extended T}. Note that the word $w_1h'_1\cdot\cdot\cdot w_kh'_kw_{k+1}$ can be decomposed as $$w_1h'_1...w_kh'_kw_{k+1}\equiv w_1h'_1(w_2h'_2\cdot\cdot\cdot w_kh'_kw_{k+1}).$$
By (P3), $w_1\in T_{\lambda_1}$.
\end{proof}
It will be shown that $\ll \mathcal{N} \rr = \prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}}N^t_{\lambda}$. For the moment, let
$$K=\langle N^t_{\lambda},t\in T_{\lambda},\lambda\in\Lambda\rangle\leqslant G.$$
\begin{lemma}\label{inK}
Let $w$ be a word in $\bigcup_{\lambda\in\Lambda}T^{ex}_{\lambda}$, and let $n$ be an element of $N_{\lambda_0}$ for some $\lambda_0\in\Lambda$. Then $wnw^{-1}\in K$.
\end{lemma}
\begin{proof}
Let $\mu$ be an element of $\Lambda$ with $rank(w)=rank_{\mu}(w)$. Thus, $w$ admits a decomposition $w\equiv w_1h_1\cdot\cdot\cdot w_kh_kw_{k+1}$ satisfying Definition \ref{extended T} with $k=rank(w)$. We perform induction on $rank(w)$. If $rank(w)=0$, then $w\in T_{\mu}$ and thus $wnw^{-1}\in K$.
Suppose that, for all $w'\in \bigcup_{\lambda\in\Lambda}T^{ex}_{\lambda}$ with $rank(w')<rank(w)$ and all $n'\in\bigcup_{\lambda\in\Lambda}N_{\lambda}$, we have $w'^{-1}n'w'\in K$. Let $h'_1,...,h'_k$ be as in (b) of Definition \ref{extended T}. Thus, there exists $n_1\in N_{\lambda_1}$ such that $n_1h'_1=h_1$ (note that $N_{\lambda_1}$ is a normal subgroup of $H_{\lambda_1}$). Notice that
$$w=_G (w_1n_1w_1^{-1})(w_1h'_1w_2h_2\cdot\cdot\cdot w_kh_kw_{k+1})$$
and thus
\begin{equation}\label{just use to refer in ink}
wnw^{-1}=_G (w_1n_1w_1^{-1})(w'nw'^{-1})(w_1n_1w_1^{-1})^{-1},
\end{equation}
where $w'\equiv w_1h'_1w_2h_2\cdot\cdot\cdot w_kh_kw_{k+1}$.
By replacing $h_j$ with $h'_j$ for $j=2,...,k$, we can turn $w'$ into a word in $T_{\mu}$. Thus, $w'\in T^{ex}_{\mu}$ and $rank(w')\leqslant k-1 < rank(w)$. It follows from the induction hypothesis that $w'n(w')^{-1}\in K$. By Lemma \ref{trunc}, $w_1\in T_{\lambda_1}$ and thus $w_1n_1w_1^{-1}\in K$. By \eqref{just use to refer in ink}, $wnw^{-1}$ represents a product of elements of $K$.
\end{proof}
For the next two lemmas, recall that $\|w\|$ denotes the length of a word $w$ over $X\sqcup\mathcal{H}$, and that $|g|$ denotes the length of a geodesic word over $X\sqcup \mathcal{H}$ representing an element $g\in G$.
\begin{lemma}\label{bbjs}
Let $\lambda$ be an element of $\Lambda$, $u$ be a word in $T^{ex}_{\lambda}$, $h$ be a letter of $H_{\lambda}\setminus\{1\}$, and $v$ be a word over $X\sqcup\mathcal{H}$ with $\|v\|=\|u\|$. Suppose that every element $m'\in \ll\mathcal{N}\rr$ with $|m'|<2\|u\|+1$ belongs to $K$. If the concatenation $uhv\in \ll\mathcal{N}\rr$, then $uhv\in K$.
\end{lemma}
\begin{proof}
If $uhv$ is not a geodesic word, the desired result will follow from the assumptions trivially. So let us assume that $uhv$ is geodesic. Consider a diagram $\Delta\in\mathcal{D}(w)$ of minimal type (see Definition \ref{type of diagrams}).
We prove Lemma \ref{bbjs} by an induction on the number of holes in $\Delta$. If $\Delta$ has no holes, then it will be a disk van Kampen diagram over \eqref{pre1} with boundary labeled by $uhv$ and thus $uhv$ represents $1\in K$.
Suppose that $\Delta$ has $k\geqslant 1$ holes. By Lemma \ref{cut}, there exists $\mu\in\Lambda$ and a connected component $c$ of $\partial_{int}\Delta$ such that $c$ is connected to an $H_{\mu}$-component of $\partial_{ext}\Delta$. Let $w$ be the label of $c$. Then $w$ is a word over $H_{\mu}$ representing an element $n\in N_{\mu}$. As $Lab(\partial_{ext}\Delta)\equiv uhv$, we can use Remark \ref{decompose loops on diagrams} to decompose $\partial_{ext}\Delta$ as the concatenation $p_up_hp_v$ of three paths $p_u,p_h$, and $p_v$ with $Lab(p_u)\equiv u, Lab(p_h)\equiv h,Lab(p_v)\equiv v$. Depending on where $c$ is connected to, there are three possible cases.
\textit{Case 1:} $c$ is connected to an $H_{\mu}$-component of $p_u$.
In other words, $u$ can be decomposed as $u\equiv u_1h_1u_2$ with $h_1\in H_{\mu}\setminus\{1\}$, and $p_u$ can be decomposed as a concatenation $p_{u_1}p_{h_1}p_{u_2}$ of three paths $p_{u_1},p_{h_1}$, and $p_{u_2}$ such that $Lab(p_{u_1})\equiv u_1, Lab(p_{h_1})\equiv h_1,Lab(p_{u_2})\equiv u_2$ and $c$ is connected to $p_{h_1}$ (see Remark \ref{decompose paths on diagrams}). By Lemma \ref{adding path}, passing to an equivalent diagram if necessary, we may assume that there exists a path $p_{h_2}$ in $\Delta$ with $Lab(p_{h_2})\equiv h_2\in H_{\mu}$, connecting the common vertex of $p_{h_1}$ and $p_{u_1}$ to a vertex of $c$. Note that the conjugate $n_1=h_2nh_2^{-1}\in N_{\mu}$. Let $h_3$ be the letter from $H_{\mu}$ such that $h_3=_Gn_1h_1$. Then
\begin{equation}\label{used in bbjs}
uhv\equiv u_1h_1u_2hv=_G (u_1n^{-1}_1u^{-1}_1)(u_1h_3u_2hv).
\end{equation}
As $h_1\neq 1$, we have $\|u_1\|\leqslant \|u\|-1$ and thus $\|u_1n^{-1}_1u^{-1}_1\|\leqslant 2\|u_1\|-1<2\|u\|+1$. Note that $u_1n^{-1}_1u^{-1}_1\in \ll\mathcal{N}\rr$. By the induction hypothesis, $u_1n^{-1}_1u^{-1}_1\in K$.
Let $u_4\equiv u_1h_3u_2$. Note that $\|u_4\|\leqslant \|u\|$. As $uhv,u_1n^{-1}_1u^{-1}_1\in \ll\mathcal{N}\rr$, it follows from \eqref{used in bbjs} that $u_4hv\in \ll\mathcal{N}\rr$. If $\|u_4\|<\|u\|$, then $\|u_4hv\|<2\|u\|+1$ and thus $u_4hv\in K$, by assumption. So let us assume that $\|u_4\|=\|u\|$. By Lemma \ref{u4}, $u_4\in T^{ex}_{\lambda}$. Let $\Sigma$ be a disc van Kampen diagram over \eqref{pre1} such that
$$Lab(\partial\Sigma)\equiv h_2wh^{-1}_2h_1h^{-1}_3.$$
Cut $\Delta$ along the path $p_{h_2}$ to produce a diagram $\Delta_1\in\mathcal{D}$ with
$$Lab(\partial_{ext}\Delta_1)\equiv u_1h_2wh^{-1}_2h_1u_2hv.$$
Glue $\Sigma$ to $\Delta_1$ by identifying the paths with label $h_2wh^{-1}_2h_1$ (perform refinements if the non-essential edges of the two paths do not match) to construct a diagram $\Delta_2\in\mathcal{D}$ with $$Lab(\partial_{ext}\Delta_2)\equiv u_4hv$$
(see Figure \ref{cut1}). Note that the number of holes in $\Delta_2$ is strictly less than that of $\Delta$. By the induction hypothesis, $u_4hv\in K$. By \eqref{used in bbjs}, $uhv$ is a product of elements of $K$.
\textit{Case 2:} $c$ is connected to an $H_{\mu}$-component of $p_v$.
This case is symmetric to Case 1 and the proof is left to the reader.
\textit{Case 3}: $c$ is connected to $p_h$.
In other words, $\mu=\lambda$ and $h\in H_{\lambda}\setminus\{1\}$. By Lemma \ref{adding path} and passing to an equivalent diagram if necessary, we may assume that there exists a path in $\Delta$, labeled by a letter $h_1\in H_{\lambda}$, connecting the common vertex of $p_h$ and $p_u$ to a vertex of $c$. Note that the conjugate $n_1=h_1nh_1^{-1}\in N_{\lambda}$. Let $h_2$ be a letter from $H_{\lambda}$ such that $h_2=_G n_1h$. Consider the equality
\begin{equation}\label{used in bbjs 1}
uhv=_G (un^{-1}_1u^{-1})(uh_2v).
\end{equation}
As $u\in T^{ex}_{\lambda}$, Lemma \ref{inK} implies that $un^{-1}_1u^{-1}\in K$. An analysis similar to the one in Case 1 (with $uh_2v$ in place of $u_4hv$) shows that $uh_2v\in K$. By \eqref{used in bbjs 1}, $uhv$ is a product of elements of $K$.
\end{proof}
\begin{figure}
{\input{cut1.pdf_tex}}
\caption{An illustration of Case 1 in the proof of Lemma \ref{bbjs}}\label{cut1}
\end{figure}
\begin{definition}
Let $w$ be a word representing an element of $\ll\mathcal{N}\rr$. Define the number $k(w)$ to be the minimal number of holes of a diagram $\Delta\in\mathcal{D}(w)$. The \textit{type} of $w$ is the pair $\tau(w)=(\|w\|,k(w))$. We order the set of types lexicographically (see Definition \ref{type of diagrams}).
\end{definition}
\begin{remark}
If $w$ is a word representing an element of $\ll\mathcal{N}\rr$ and $\Delta$ is a diagram in $\mathcal{D}(w)$ of minimal type, then $\Delta$ necessarily has $k(w)$ holes.
\end{remark}
\begin{proposition}\label{generate}
$\ll\mathcal{N}\rr=K$.
\end{proposition}
\begin{proof}
Clearly, each of the groups $N^t_{\lambda},t\in T_{\lambda},\lambda\in\Lambda,$ is contained in $\ll\mathcal{N}\rr$ and thus $K \leqslant \ll\mathcal{N}\rr$. Let $w$ be a word over $X\sqcup \mathcal{H}$ such that $w\in\ll\mathcal{N}\rr$. Let us show that $w\in K$ by performing induction on the type of $w$. Note that the base case $\|w\|=k(w)=0$ is trivial.
Suppose that, for every word $w'$ over $X\sqcup\mathcal{H}$ with $w'\in \ll\mathcal{N}\rr$, $\tau(w')<\tau(w)$ implies that $w'\in K$. If $w$ is not a geodesic word, the induction hypothesis will imply $w\in K$. Thus, we may assume that $w$ is geodesic. Consider a diagram $\Delta\in\mathcal{D}(w)$ of minimal type.
By Lemma \ref{cut}, there exist $\lambda\in\Lambda$ and a connected component $c$ of $\partial_{int}\Delta$ connected to an $H_{\lambda}$-component of $\partial_{ext}\Delta$. In other words, $w$ can be decomposed as $uhv$ with $h\in H_{\lambda}\setminus\{1\}$ ($u,v$ are allowed to be empty words), and $\partial_{ext}\Delta$ can be decomposed as a concatenation $p_up_hp_v$ of three paths $p_u,p_h$, and $p_v$ such that $Lab(p_u)=u,Lab(p_h)=h,Lab(p_v)=v$ and $c$ is connected to $p_h$ (see Remark \ref{decompose loops on diagrams}). By Lemma \ref{adding path} and passing to an equivalent diagram if necessary, we may assume that there exists a path $p_{h_1}$ in $\Delta$ with $Lab(p_{h_1})\equiv h_1\in H_{\lambda}$, connecting the common vertex of $p_h$ and $p_u$ to a vertex of $c$.
Note that, as $h\neq 1$, at least one of $\|u\|$ and $\|v\|$ is at most $(\|w\|-1)/2$. Without loss of generality, we may assume that $\|v\|\leqslant (\|w\|-1)/2$. The case $\|u\|\leqslant (\|w\|-1)/2$ can be analyzed in almost the same way (or just by considering $w^{-1}$ and reversing every edge of $\Delta$ if one wishes).
Let $w_1\equiv Lab(c)$. Thus, $w_1\in N_{\lambda}$. Let $h_2$ be a letter from $H_{\lambda}$ such that $h_2=_G hh_1nh^{-1}_1$. There exists $t\in T_{\lambda}$ such that $t$ and $v^{-1}$ are in the same left $H_{\lambda}\ll \mathcal{N} \rr$-coset. In other words, there exists $h_3\in H_{\lambda}$ such that $th_3v\in \ll\mathcal{N}\rr$. Let $n_1$ be a letter in $N_{\lambda}$ such that $n_1=_G h_3h_1n^{-1}h^{-1}_1h^{-1}_3$.
Consider the equality
\begin{equation}\label{factor}
w\equiv uhv=_G (uh_2v)(v^{-1}h^{-1}_3t^{-1})(tn_1t^{-1})(th_3v).
\end{equation}
Note that $uh_2v\in \ll\mathcal{N}\rr$, as all other brackets in \eqref{factor} represents elements of $\ll\mathcal{N}\rr$. As in the proof of Lemma \ref{bbjs}, let $\Sigma$ be a disc van Kampen diagram over \eqref{pre1} with $$Lab(\partial\Sigma)\equiv hh_1w_1h^{-1}_1h^{-1}_2.$$
Cut $\Delta$ along $p_{h_1}$ to produce a diagram $\Delta_1\in\mathcal{D}$ with $Lab(\partial_{ext}\Delta_1)\equiv uhh_1w_1h^{-1}_1v$. Glue $\Delta_1$ to $\Sigma$, identifying the paths labeled by $hh_1w_1h^{-1}_1$ (perform refinements if the non-essential edges of the two paths do not match). Denote the resulting diagram by $\Delta_2$. Clearly, $\Delta_2\in\mathcal{D}$ and $Lab(\partial_{ext}\Delta_2)\equiv uh_2v$. Note that the number of holes in $\Delta_2$ is strictly less than that of $\Delta$, and that $\|uh_2v\|\leqslant \|u\|+\|v\|+1=\|uhv\|$, as $uhv$ is a geodesic word. Thus, $\tau(uh_2v)<\tau(w)$ and the induction hypothesis implies $uh_2v\in K$.
Clearly, $tn_1t^{-1}\in K$. Note also that $th_3v\in K$. Indeed, if either $\|t\|<\|v\|$ or $h_3=1$, then $\|th_3v\|<2\|v\|+1=\|w\|$ and the induction hypothesis implies that $th_3v\in K$. If $\|t\|=\|v\|$ and $h_3\neq 1$, then Lemma \ref{bbjs} implies $th_3v\in K$.
As $v^{-1}h^{-1}_3t^{-1}\equiv (th_3v)^{-1}$, we also have $v^{-1}h^{-1}_3t^{-1}\in K$. By \eqref{factor}, $w$ is a product of elements of $K$.
\end{proof}
The cutting process in the proof of Lemma \ref{generate} is exactly the same as the one for Lemma \ref{bbjs}. See Figure \ref{cut1} for an illustration.
The goal of the rest of this section is to prove the following.
\begin{proposition}\label{free}
$\ll\mathcal{N}\rr=\prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}}N^t_{\lambda}$.
\end{proposition}
\begin{proof}
Assume, for the contrary, that there exists a word
\begin{equation}\label{star}
z\equiv \prod_{i=1}^k t_i n_i t_i^{-1}
\end{equation}
representing $1\in G$ such that
\begin{enumerate}
\item[(Z1)] $k\geqslant 2$;
\item[(Z2)] for $i=1,...,k$, there exists $\lambda_i\in\Lambda$ such that $n_i\in N_{\lambda_i}\setminus\{1\}$ and $t_i\in T_{\lambda_i}$;
\item[(Z3)] for $i=1,...,k$, either $\lambda_i\neq\lambda_{i+1}$ or $t_i\not\equiv t_{i+1}$ (subscripts are modulo $k$, i.e., $n_{k+1}=n_1$, $t_0=t_k$, etc.).
\end{enumerate}
Without loss of generality, we may also assume
\begin{enumerate}
\item[(Z4)] $z$ is minimal, i.e., has the minimal $k$ among all other words of the form \eqref{star} representing $1$ in $G$ and satisfying (Z1), (Z2), and (Z3).
\end{enumerate}
The main idea of the proof of Lemma \ref{free} is to show that the existence of such a word $z$ contradicts Lemma \ref{length}. For this purpose, it is convenient to first cyclically permute $z$ and consider the word
$$w\equiv t^{-1}_k(\prod_{i=1}^{k-1} t_i n_i t_i^{-1})t_kn_k.$$
In what follows, subscripts are modulo $k$. Let $p_w$ be the path in $\Gamma(G,X\sqcup \mathcal{H})$ with $Lab(p)\equiv w$ and $p^-=1$. We use $p_{n_i},p_{t^{\pm1}_i}$ to denote subpaths of $p_w$ labeled by $n_i,t^{\pm1}_i$, respectively. More precisely, $p_{n_i}$ (resp. $p_{t_i},p_{t^{-1}_i}$) will denote the path in the Cayley graph $\Gamma(G,X\sqcup\mathcal{H})$ with $Lab(p_{n_i})=n_i$ (resp. $Lab(p_{t_i})=t_i,Lab(p_{t^{-1}_i})=t^{-1}_i$) and $p^-_{n_i}=t^{-1}_k(\prod_{j=1}^{i-1}t_jn_jt^{-1}_j)t_i$ (resp. $p^-_{t_i}=t^{-1}_k(\prod_{j=1}^{i-1}t_jn_jt^{-1}_j),p^-_{t^{-1}_i}=t^{-1}_k(\prod_{j=1}^{i-1}t_jn_jt^{-1}_j)t_in_i$).
Recall that the collection $\{T_{\lambda}\}_{\lambda\in\Lambda}$ satisfies (P1), (P2), and (P3). Note that, for every $\lambda\in\Lambda$ and every word $t\in T_{\lambda}$, the word $t$ does not end with a letter from $H_{\lambda}$, by (P2). It follows that $p_{n_i}$ is an $H_{\lambda_i}$-component of $p_w$ for $i=1,...,k$. Being a cyclic permutation of $z$, the word $w$ represents $1$ in $G$ and thus the terminal vertex of $p_w$ is $1$. Hence, $p_w$ is a geodesic $3k$-gon. As $\widehat{\ell}_{\lambda_i}(p_{n_i})=\widehat{d}_{\lambda_i}(1,n_i)$ for $i=1,...,k$, by Lemma \ref{length} and (24D), there exists some $i\in \{1,...,k\}$ such that $p_{n_i}$ is not an isolated $H_{\lambda_i}$-component of $p_w$.
The rest of the proof is divided into several lemmas. All of them are stated under the assumptions (and using the notations) of Proposition \ref{free}.
\begin{lemma}\label{eliminate}
If $p_{n_i}$ is not an isolated $H_{\lambda_i}$-component of $p_w$ for some $i\in \{1,...,k\}$, then there are only three possibilities:
\begin{enumerate}
\item[(a)] $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, but not connected to any $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$.
\item[(b)] $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$, but not connected to any $H_{\lambda_i}$-component of $p_{t_{i+1}}$.
\item[(c)] $p_{n_i}$ is connected to both an $H_{\lambda_i}$-component of $p_{t_{i+1}}$ and an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Without loss of generality, let us assume that $p_{n_1}$ is not isolated in $p_w$. There are six cases to consider (see Figure \ref{illustrateeliminate} for an illustration).
\textit{Case 1:} $p_{n_1}$ is connected to an $H_{\lambda_1}$-component of either $p_{t_1}$ or $p_{t_1^{-1}}$. In this case, some terminal segment of $t_1$ represents an element of $H_{\lambda_1}$, which contradicts (P2).
\textit{Case 2:} $p_{n_1}$ is connected to either $p_{n_2}$ or $p_{n_k}$. If $p_{n_1}$ is connected to $p_{n_2}$, then $\lambda_1=\lambda_2$, which in turn implies $t_1,t_2\in T_{\lambda_1}$. The assumption that $p_{n_1}$ is connected to $p_{n_2}$ also implies $t^{-1}_1t_2\in H_{\lambda_1}$. By (P1), $t_1\equiv t_2$, which contradicts (Z3) as $\lambda_1=\lambda_2$. The analysis for the subcase where $p_{n_1}$ is connected to $p_{n_k}$ is similar.
\textit{Case 3:} $p_{n_1}$ is connected to $p_{n_i}$ for some $i\in \{3,...,k-1\}$. In other words, there exists $h\in H_{\lambda_1}$ such that the word
\begin{equation*}
u\equiv t_1^{-1}(\prod_{j=2}^{i-1}t_jn_jt_j^{-1})t_ih
\end{equation*}
represents $1$ in $G$. As $\prod_{j=2}^{i-1}t_jn_jt_j^{-1}\in \ll\mathcal{N}\rr\lhd G$, we have $t_1^{-1}t_i\in H_{\lambda_1}\ll\mathcal{N}\rr$. The assumption that $p_{n_1}$ is connected to $p_{n_i}$ also implies $n_1,n_i\in N_{\lambda_1}$ and thus $t_1,t_i\in T_{\lambda_1}$. By (P1), $t_1\equiv t_i$. Thus, the word
\begin{equation*}
u'\equiv t_1ht_1^{-1}(\prod_{j=2}^{i-1}t_jn_jt_j^{-1})
\end{equation*}
is a cyclic permutation of $u$ and represents $1$ in $G$. It follows that $t_1ht_1^{-1}\in\ll\mathcal{N}\rr$. By Theorem \ref{Dehn filling}, Remark \ref{constant}, and Condition (24D), we have $h\in N_{\lambda_1}$. Then the word $t_1ht_1^{-1}(\prod_{j=2}^{i-1}t_jn_jt_j^{-1})$ represents $1$ in $G$, contradicting (Z4).
\textit{Case 4:} $p_{n_1}$ is connected to an $H_{\lambda_1}$-component of $p_{t_i}$ for some $i\in \{3,...,k\}$. Thus, $t_i$ can be decomposed as $t_i\equiv t'_ih't''_i$ with $h'\in H_{\lambda_1}\setminus\{1\}$ and there exists $h\in H_{\lambda_1}$ such that the word
\begin{equation*}
u\equiv t_1^{-1}(\prod_{j=2}^{i-1}t_jn_jt_j^{-1})t'_ih
\end{equation*}
represents $1$ in $G$. By (P3), $t'_i$ belongs to $T_{\lambda_1}$. Arguing as in Case 3, we conclude that the word $t_1ht_1^{-1}(\prod_{j=2}^{i-1}t_jn_jt_j^{-1})$ represents $1$ in $G$, contradicting (Z4).
\textit{Case 5:} $p_{n_1}$ is connected to an $H_{\lambda_1}$-component of $p_{t_i^{-1}}$ for some $i\in \{2,...,k-1\}$. This case can be reduced to Case 4 by considering $w^{-1}$.
Thus, the only possibilities left are (a), (b), and (c).
\end{proof}
\begin{figure}
\begin{center}
\resizebox{0.95\linewidth}{!}{\input{3.pdf_tex}}
\end{center}
\caption{Cases 1 through 6 in the proof of Lemma \ref{eliminate}}\label{illustrateeliminate}
\end{figure}
\begin{lemma}\label{prefix 1}
If $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, then $t_{i+1}$ can be decomposed as $t_{i+1}\equiv uhv$ with $h\in H_{\lambda_i}\setminus\{1\}$ ($u,v$ are allowed to be empty words), $t_i\equiv u$, and $\widehat{d}_{\lambda_i}(1,n_ih)>12D$.
\end{lemma}
\begin{proof}
By Definition \ref{connect}, $t_{i+1}$ can be decomposed as $t_{i+1}\equiv uhv$ with $h\in H_{\lambda_i}\setminus\{1\}$ such that $p_{n_i}$ is connected to the path $p_h$ in $\Gamma(G,X\sqcup\mathcal{H})$ with $Lab(p_h)\equiv h$ and $p^-_h=t^{-1}_k(\prod_{j=1}^i t_jn_jt^{-1}_j)u$. By (P3), $u\in T_{\lambda_i}$. The assumption that $p_{n_i}$ is connected to $p_h$ also implies $t^{-1}_iu\in H_{\lambda_i}$ and thus $t_i\equiv u$, by (P1). Another consequence of (P3) is $$\widehat{d}_{\lambda_i}(1,h)\leqslant\widehat{d}_{\lambda_i}(1,h(h^{-1}n_ih))=\widehat{d}_{\lambda_i}(1,n_ih).$$
Therefore, the triangle inequality implies $$\widehat{d}_{\lambda_i}(1,n_i)\leqslant \widehat{d}_{\lambda_i}(1,n_ih)+\widehat{d}_{\lambda_i}(1,h^{-1})=\widehat{d}_{\lambda_i}(1,n_ih)+\widehat{d}_{\lambda_i}(1,h)\leqslant 2\widehat{d}_{\lambda_i}(1,n_ih)$$
and thus
$$\widehat{d}_{\lambda_i}(1,n_ih)\geqslant\widehat{d}_{\lambda_i}(1,n_i)/2>12D,$$
by (24D).
\end{proof}
The next lemma follows from Lemma \ref{prefix 1} by considering $w^{-1}$.
\begin{lemma}\label{prefix 2}
If $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$, then $t_{i-1}$ can be decomposed as $t_{i-1}\equiv uhv$ with $h\in H_{\lambda_i}\setminus\{1\}$ ($u,v$ are allowed to be empty words), $t_i\equiv u$, and $\widehat{d}_{\lambda_i}(1,h^{-1}n_i)>12D$.
\end{lemma}
\begin{lemma}\label{disconnect}
If $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, then $p_{n_{i+1}}$ is not connected to any $H_{\lambda_{i+1}}$-component of $p_{t^{-1}_i}$. If $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$, then $p_{n_{i-1}}$ is not connected to any $H_{\lambda_{i-1}}$-component of $p_{t_i}$.
\end{lemma}
\begin{proof}
If $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, then $\|t_i\|<\|t_{i+1}\|$ by Lemma \ref{prefix 1}. If, in addition, $p_{n_{i+1}}$ is connected to an $H_{\lambda_{i+1}}$-component of $p_{t^{-1}_i}$, then $\|t_{i+1}\|<\|t_i\|$ by Lemma \ref{prefix 2}, a contradiction.
The second assertion of the Lemma can be proved by considering $w^{-1}$.
\end{proof}
Recall that we assume the existence of a word $z$ satisfying (Z1) through (Z4) and construct $w,p_w$ from $z$. The previous several lemmas reveal some properties of $p_w$ and we are now ready to construct a geodesic polygon $p$ from $p_w$ so that $p$ violates Lemma \ref{length}, and then we can conclude that $z$ does not exist and prove Proposition \ref{free}. The idea is to merge all $H_{\lambda_i}$-components connected to $p_{n_i}$ to form an isolated $H_{\lambda_i}$-component for $i=1,...,k-1$. Of course, one can also merge $p_{n_k}$ with the $H_{\lambda_k}$-components connected to it. We do not perform this merging only because it makes the construction more complicated. Pick elements $h_1,...,h_{k-1}\in \mathcal{H}$ and $g_{1,1},g_{1,2},g_{2,1},g_{2,2},...,g_{k-1,1},g_{k-1,2}\in G$ by the following procedure.
\begin{procedure}\label{pro.}
For $i=1,...,k-1$, perform the following.
\begin{enumerate}
\item[(a)] If $p_{n_i}$ is an isolated $H_{\lambda_i}$-component in $p_w$, let $g_{i,1}\in G$ (resp. $g_{i,2}\in G$) be represented by the word $t^{-1}_k(\prod_{j=1}^{i-1} t_j n_j t_j^{-1})t_i$ (resp. $t^{-1}_k(\prod_{j=1}^{i-1} t_j n_j t_j^{-1})t_in_i$), and let $h_i=n_i$.
\item[(b)] If, in $p_w$, $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, but not connected to any $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$, then by Lemma \ref{prefix 1}, $t_{i+1}$ can be decomposed as $t_{i+1}\equiv u_ih'_iv_i$ with $h'_i\in H_{\lambda_i}\setminus\{1\}$, $t_i\equiv u_i$, and $\widehat{d}_{\lambda_i}(1,n_ih'_i)>12D$. Let $h_i$ be a letter from $H_{\lambda_i}$ such that $h_i=_G n_ih'_i$, and let $g_{i,1}\in G$ (resp. $g_{i,2}\in G$) be represented by the word $t^{-1}_k(\prod_{j=1}^{i-1} t_j n_j t_j^{-1})t_i$ (resp. $t^{-1}_k(\prod_{j=1}^{i-1} t_j n_j t_j^{-1})t_ih_i$).
\item[(c)] If in $p_w$, $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$, but not connected to any $H_{\lambda_i}$-component of $t_{i+1}$, then by Lemma \ref{prefix 2}, $t_{i-1}$ can be decomposed as $t_{i-1}\equiv u_ih'_iv_i$ with $h'_i\in H_{\lambda_i}\setminus\{1\}$, $t_i\equiv u_i$, and $\widehat{d}_{\lambda_i}(1,h'^{-1}_in_i)>12D$. Let $h_i$ be a letter from $H_{\lambda_i}$ such that $h_i=_G h'^{-1}_in_i$, and let $g_{i,1}\in G$ (resp. $g_{i,2}\in G$) be represented by the word $t^{-1}_k(\prod_{j=1}^{i-2} t_j n_j t_j^{-1})t_{i-1}n_{i-1}v^{-1}_i$ (resp. $t^{-1}_k(\prod_{j=1}^{i-2} t_j n_j t_j^{-1})t_{i-1}n_{i-1}v^{-1}_ih_i$).
\item[(d)] If in $p_w$, $p_{n_i}$ is connected to both an $H_{\lambda_i}$-component of $p_{t_{i+1}}$ and an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$, then by Lemmas \ref{prefix 1} and \ref{prefix 2}, $t_{i+1}$ (resp. $t_{i-1}$) can be decomposed as $t_{i+1}\equiv u_ih'_iv_i$ (resp. $t_{i-1}\equiv u'_ih''_iv'_i$) with $h'_i\in H_{\lambda_i}\setminus\{1\}$ (resp. $h''_i\in H_{\lambda_i}\setminus\{1\}$), $t_i\equiv u_i$ (resp. $t_i\equiv u'_i$). Let $h_i$ be a letter from $H_{\lambda_i}$ such that $h_i=_G h''^{-1}_in_ih'_i$, and let $g_{i,1}\in G$ (resp. $g_{i,2}\in G$) be represented by the word $t^{-1}_k(\prod_{j=1}^{i-2} t_j n_j t_j^{-1})t_{i-1}n_{i-1}(v'_i)^{-1}$ (resp. $t^{-1}_k(\prod_{j=1}^{i-2} t_j n_j t_j^{-1})t_{i-1}n_{i-1}(v'_i)^{-1}h_i$).
\end{enumerate}
\end{procedure}
\begin{lemma}\label{order}
$g_{i,1}$ and $g_{i,2}$ are vertices on $p_w$ for $i=1,...,k-1$. Moreover, the order in which $p_w$ visits these vertices is $g_{1,1},g_{1,2},g_{2,1},g_{2,2},...,g_{k-1,1},g_{k-1,2}$.
\end{lemma}
\begin{proof}
The first assertion follows directly from the choices of those vertices. Clearly, the path $p_w$ visits $g_{i,1}$ before visiting $g_{i,2}$ for $i=1,...,k-1$. Thus, the second assertion will be proved once we show that, for all $i,j\in \{1,...,k-1\}$ with $i<j$, the path $p_w$ visits $g_{i,2}$ before visiting $g_{j,1}$.
Suppose, for the contrary, that for some $i,j\in \{1,...,k-1\}$ with $i<j$, the path $p_w$ visits $g_{j,1}$ before visiting $g_{i,2}$. By Lemma \ref{eliminate}, there is only one possibility for this case: $j=i+1$, $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, and $p_{n_{i+1}}$ is connected to an $H_{\lambda_{i+1}}$-component of $p_{t^{-1}_i}$. By Lemma \ref{disconnect}, if $p_{n_i}$ is connected to an $H_{\lambda_i}$-component of $p_{t_{i+1}}$, then $p_{n_{i+1}}$ is not connected to any $H_{\lambda_{i+1}}$-component of $p_{t^{-1}_i}$, a contradiction.
\end{proof}
\begin{lemma}\label{obvious}
For $i=1,...,k-2$, the subpath of $p_w$ from $g_{i,2}$ to $g_{i+1,1}$ consists of at most two geodesic segments.
\end{lemma}
Lemma \ref{obvious} follows immediately from the choices of the vertices $g_{i,1}$ and $g_{i,2}$, $1\leqslant i\leqslant k-1$. We are now ready to construct a geodesic polygon $p$ from $p_w$.
\begin{construction}\label{con.}
For $i=1,...,k-1$, let $p_{h_i}$ the edge of $\Gamma(G,X\sqcup \mathcal{H})$ with $Lab(p_{h_i})=h_i$ and $p^-_{h_i}=g_{i,1}$. Let $p$ be the path in $\Gamma(G,X\sqcup \mathcal{H})$ satisfying: $p^-$ is the identity vertex. $p$ first follows the path of $p_w$ (in the direction of $p_w$) until $p$ visits $g_{1,1}$, and then $p$ travels along $p_{h_1}$ and arrives at $g_{1,2}$. And then $p$ follows the path $p_w$ (in the direction of $p_w$) until $p$ arrives at $g_{2,1}$ (Lemma \ref{order} guarantees that $p$ will arrive at $g_{2,1}$), where $p$ travels along $p_{h_2}$ and then arrives at $g_{2,2}$. The path $p$ continues traveling in this manner until arriving at $g_{k-1,2}$. Finally, $p$ follows the path $p_w$ (in the direction of $p_w$) and comes back to the identity vertex.
\end{construction}
\begin{figure}
{\input{connect.pdf_tex}}
\caption{The construction of $p$}\label{constructp}
\end{figure}
Figure \ref{constructp} illustrates how to construct the geodesic polygon $p$. In Figure \ref{constructp}, the outside boundary with label $t^{-1}_4t_1n_1t^{-1}_1t_2n_2t^{-1}_2t_3n_3t^{-1}_3t_4n_4$ is the geodesic polygon $p_w$. In the outside boundary, $p_{n_2}$ is an isolated $H_{\lambda_2}$-component, $p_{n_1}$ (resp. $p_{n_4}$) is connected to an $H_{\lambda_1}$-component (resp. $H_{\lambda_4}$-component) of $p_{t_2}$ (resp. $p_{t_1}$), and $p_{n_3}$ is connected to both an $H_{\lambda_3}$-component of $p_{t^{-1}_2}$ and an $H_{\lambda_3}$-component of $p_{t_4}$. By Lemma \ref{prefix 1}, $t^{-1}_1$ cancels with a prefix of $t_2$. After this cancellation, $p_{n_1}$ merges with an $H_{\lambda_1}$-component of $p_{t_2}$ to form $p_{h_1}$. Similarly, $p_{n_3}$ merges with both an $H_{\lambda_3}$-component of $p_{t^{-1}_2}$ and an $H_{\lambda_3}$-component of $p_{t_4}$ to form $p_{h_3}$. The merging process does nothing to $n_4$, although $n_4$ is not an isolated $H_{\lambda_4}$-component. Finally, $p_w$ becomes $p$, the boundary of the shaded region.
\begin{remark}
It follows directly from the above construction that $p_{h_i}$ is an isolated $H_{\lambda_i}$-component of $p$ for $i=1,...,k-1$.
\end{remark}
Note that the subpath of $p_w$ from $1$ to $g_{i,1}$ consists of at most $2$ geodesic segments, and the subpath of $p_w$ from $g_{k-1,2}$ to $1$ consists of at most $3$ geodesic segments. Together with Lemma \ref{obvious}, these observations imply that $p$ is a polygon in $\Gamma(G,X\sqcup \mathcal{H})$ with at most $3k$ geodesic sides.
Consider the following partition of $\{1,...,k-1\}=I_1\sqcup I_2$. A number $1\leqslant i\leqslant k-1$ belongs to $I_1$ if in $p_w$, $p_{n_i}$ is connected to both an $H_{\lambda_i}$-component of $p_{t^{-1}_{i-1}}$ and an $H_{\lambda_i}$-component of $p_{t_{i+1}}$. Otherwise, $i$ belongs to $I_2$.
\begin{lemma}
$card(I_1)\leqslant (k-1)/2$.
\end{lemma}
\begin{proof}
First suppose $card(I_1)>k/2$. Then there exists a number $i$ such that both $i$ and $i+1$ belong to $I_1$, contradicting Lemma \ref{disconnect}. Thus, $card(I_1)\leqslant k/2$.
Suppose $card(I_1)=k/2$. Then $k$ is even and $I_1=\{1,3,...,k-3,k-1\}$. For every even number $i\in\{2,4,...,k-2,k\}$, Lemma \ref{disconnect} implies that $p_{n_i}$ is an isolated $H_{\lambda_i}$-component of $p_w$. Note that $\widehat{\ell}_{\lambda_i}(p_{n_i})=\widehat{d}_{\lambda_i}(1,n_i)>24D$ for $i=1,...,k$, by (24D). Therefore, Lemma \ref{length}, applied to the geodesic $3k$-gon $p_w$, yields
$$\dfrac{24Dk}{2}<\widehat{\ell}_{\lambda_2}(p_{n_2})+\widehat{\ell}_{\lambda_4}(p_{n_4})+\cdot\cdot\cdot+\widehat{\ell}_{\lambda_{k-2}}(p_{n_{k-2}})+\widehat{\ell}_{\lambda_k}(p_{n_k})<3kD,$$
a contradiction.
\end{proof}
Thus, $card(I_2)=k-1-card(I_1)\geqslant (k-1)/2$. For each $i\in I_2$, $p_{h_i}$ is an isolated $H_{\lambda_i}$-component of $p$ with $\widehat{\ell}_{\lambda_i}(p_{h_i})=\widehat{d}_{\lambda_i}(1,h_i)>12D$, by Procedure \ref{pro.} and Construction \ref{con.}. Lemma \ref{length}, applied to the geodesic polygon $p$, yields
\begin{equation}\label{estimate}
6D(k-1)=12D(k-1)/2<\sum_{i\in I_2}\widehat{\ell}_{\lambda_i}(p_{h_i})\leqslant 3kD.
\end{equation}
In other words, $k<2$, contradicting (Z1). Proposition \ref{free} is proved.
\end{proof}
Finally, Theorem \ref{main} follows from Proposition \ref{generate} and Proposition \ref{free}.
\begin{remark}
The proof of Theorem \ref{main} implies that if $\{H_{\lambda}\}_{\lambda\in\Lambda}\hookrightarrow_{wh} (G,X)$ for some $X\subset G$, $N_{\lambda}\lhd H_{\lambda}$ for $\lambda\in\Lambda$, and (24D) holds, then for every collection $\{T_{\lambda}\}_{\lambda\in\Lambda}$ satisfying (P1), (P2), and (P3), we have
$$\ll\mathcal{N}\rr=\prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}}N^t.$$
\end{remark}
\begin{remark}
In fact, one can show that if $\{H_{\lambda}\}_{\lambda\in\Lambda}\hookrightarrow_{wh} (G,X)$ for some $X\subset G$, $N_{\lambda}\lhd H_{\lambda}$ for $\lambda\in\Lambda$, and following condition
\begin{enumerate}
\item[(4D)] $\widehat{d}_{\lambda}(1,n)>4D$ for all $n\in N_{\lambda}\setminus\{1\}$ and $\lambda\in\Lambda$
\end{enumerate}
holds, then the triple $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ possesses the Cohen-Lyndon property. For the proof, one needs to merge $p_{n_k}$ with the $H_{\lambda_k}$-components connected to it in the construction of $p$, and sharpen the coarse estimate \eqref{estimate}.
\end{remark}
\section{Relative relation modules}\label{sec.rm}
Let $H$ be a group with a normal subgroup $N$ and $R=H/N$. The \textit{relative relation module} $Rel(H,N)$ \textit{of the exact sequence} $$1\rightarrow N\rightarrow H\rightarrow R\rightarrow 1$$
is the abelianization $\overline{N}=N/[N,N]$ equipped with the $R$-action by conjugation. More precisely, denote by $\overline{n}$ the image of an element $n\in N$ under the quotient map $N\rightarrow \overline{N}$. Then there is an action of $H$ on $\overline{N}$ given by $h\circ \overline{n}=\overline{hnh^{-1}}$ for all $h\in H,\overline{n}\in \overline{N}$. Notice that if $h$ belongs to $N$, then $h\circ \overline{n}=\overline{hnh^{-1}}=\overline{h}\overline{n}\overline{h}^{-1}=\overline{n}$ for all $\overline{n}\in \overline{N}$, as $\overline{h}$ commutes with $\overline{n}$. Hence, the action of $H$ gives rises to an action of $R$, turning $\overline{N}$ into a $\mathbb{Z}R$-module. If $H$ is a free group, then $Rel(H,N)$ is called a \textit{relation module}.
The main goal of this section is to prove Proposition \ref{genmodule}, which, together with Theorem \ref{simmain}, implies Corollary \ref{module}.
\begin{proposition}\label{genmodule}
Suppose that $G$ is a group, that $\{H_{\lambda}\}_{\lambda\in\Lambda}$ is a family of subgroups of $G$, and that $N_{\lambda}$ is a normal subgroup of $H_{\lambda}$ for every $\lambda\in\Lambda$. Let $\mathcal{N}=\bigcup_{\lambda\in\Lambda}N_{\lambda}$, $Q=G/\ll \mathcal{N}\rr$, and $R_{\lambda}=H_{\lambda}/N_{\lambda}$ for $\lambda\in\Lambda$. If $N_{\lambda}\neq\{1\}$ for every $\lambda\in\Lambda$ and the triple $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ has the Cohen-Lyndon property, then
\begin{enumerate}
\item[(a)] for every $\lambda\in\Lambda$, the natural map $R_{\lambda}\rightarrow Q$ is injective (i.e., $H_{\lambda}\cap \ll \mathcal{N} \rr = N_{\lambda}$), identifying $R_{\lambda}$ with a subgroup of $Q$;
\item[(b)] $Rel(G,\ll \mathcal{N} \rr)\cong\bigoplus_{\lambda\in\Lambda}Ind_{R_{\lambda}}^Q Rel(H_{\lambda},N_{\lambda})$ as $\mathbb{Z}Q$-modules.
\end{enumerate}
\end{proposition}
\begin{remark}
If $N_{\lambda_0}=\{1\}$ for some $\lambda_0\in\Lambda$, then we can consider the subset $\Lambda'$ such that $N_{\lambda}\neq \{1\}$ for every $\lambda\in\Lambda'$. Proposition \ref{genmodule} can then be applied to $(G,\{H_{\lambda}\}_{\lambda\in\Lambda'},\{N_{\lambda}\}_{\lambda\in\Lambda'})$.
\end{remark}
Suppose that the assumptions of Proposition \ref{genmodule} are satisfied. Let $T_{\lambda},\lambda\in\Lambda$, be the transversals provided by Definition \ref{CLp}. Fix some $\lambda\in\Lambda$ for the moment. Suppose $h\in H_{\lambda}\cap \ll\mathcal{N}\rr$. Then $h\in N_{\ll\mathcal{N}\rr}(N_{\lambda})$, the normalizer of $N_{\lambda}$ in $\ll\mathcal{N}\rr$. Note that
$$\ll\mathcal{N}\rr=\prod^{\ast}_{\mu\in\Lambda,t\in T_{\mu}}N^t_{\mu}=N_{\lambda}\ast (\prod^{\ast}_{t\in T_{\lambda}\setminus\{1\}}N^t_{\lambda}\ast \prod^{\ast}_{\mu\in\Lambda\setminus\{\lambda\},t\in T_{\mu}}N^t_{\mu})$$
and $N_{\lambda}\neq\{1\}$. Note also the following general fact.
\begin{lemma}
Let $A,B\neq\{1\}$ be groups. Then $N_{A\ast B}(A)=A$.
\end{lemma}
\begin{proof}
Suppose that there exists $a\in A\setminus\{1\}$ and $g\in A\ast B\setminus A$ such that $a^g\in A$. Consider the Bass-Serre tree $Tr$ corresponding to $A\ast B$. The vertex group $A$ fixes a vertex $v$ of $Tr$ and thus $a^g$ fixes $v$. Clearly, the vertex $g\circ v$ is also fixed by $a^g$. As $g\in A\ast B\setminus A$, $g\circ v\neq v$ and thus $a^g$ fixes a non-trivial path between $v$ and $g\circ v$. In particular, $a^g$ fixes an edge of $Tr$ and thus conjugates into the unique edge subgroup $\{1\}$ of $A\ast B$. It follows that $a^g=1$, which is in contradiction with $a\neq 1$.
\end{proof}
Therefore, $N_{\ll\mathcal{N}\rr}(N_{\lambda})=N_{\lambda}$ and $h\in N_{\lambda}$. We conclude:
\begin{lemma}
For every $\lambda\in\Lambda$, $H_{\lambda}\cap \ll \mathcal{N} \rr = N_{\lambda}$.
\end{lemma}
Let us consider the relative relation modules $Rel(G,\ll\mathcal{N}\rr)$ and $Rel(H_{\lambda},N_{\lambda}),\lambda\in\Lambda$. For every $\lambda\in\Lambda$, let $M_{\lambda}$ be the subgroup of $G$ generated by $N^t_{\lambda},t\in T_{\lambda}$. Note that $M_{\lambda}=\prod^{\ast}_{t\in T_{\lambda}}N^t_{\lambda}$ for every $\lambda\in\Lambda$, as $\ll\mathcal{N}\rr=\prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}}N^t_{\lambda}$. Note also that $\ll\mathcal{N}\rr=\prod^{\ast}_{\lambda\in\Lambda}M_{\lambda}$.
For every $\lambda\in\Lambda$, the composition of natural maps $M_{\lambda}\hookrightarrow \ll\mathcal{N}\rr \rightarrow \overline{\ll\mathcal{N}\rr}$ maps $M_{\lambda}$ into the abelian group $\overline{\ll\mathcal{N}\rr}$ and thus factors through
$$i_{\lambda}:\overline{M_{\lambda}}\rightarrow \overline{\ll\mathcal{N}\rr}.$$
The homomorphisms $i_{\lambda},\lambda\in\Lambda,$ extend to an abelian group homomorphism
$$i:\bigoplus_{\lambda\in\Lambda}\overline{M_{\lambda}}\rightarrow \overline{\ll\mathcal{N}\rr}.$$
It is well-known that $i$ is an abelian group isomorphism (for example, see \cite[Problem 4 of Exercise 6.2]{robinson2012course}). Thus, we identify $\overline{M_{\lambda}}$ with its image $i_{\lambda}(\overline{M_{\lambda}})$ for every $\lambda\in\Lambda$ and write
$$Rel(G,\ll\mathcal{N}\rr)=\overline{\ll\mathcal{N}\rr}=\bigoplus_{\lambda\in\Lambda}\overline{M_{\lambda}}.$$
Fix $\lambda\in\Lambda$ for the moment. By the same argument as the one above, we write
$$\overline{M_{\lambda}}=\bigoplus_{t\in T_{\lambda}}\overline{N^t_{\lambda}}.$$
\begin{lemma}\label{permute factors}
$\overline{M_{\lambda}}$ is a $\mathbb{Z}Q$-submodule of $Rel(G,\ll\mathcal{N}\rr)=\bigoplus_{\lambda\in\Lambda}\overline{M_{\lambda}}$. The $Q$-action on $\overline{M_{\lambda}}$ transitively permutes the summands $\overline{N^t_{\lambda}},t\in T_{\lambda},$ and its isotropy group of $\overline{N_{\lambda}}$ is $R_{\lambda}$, i.e., an element $q\in Q$ satisfies $q\circ \overline{n}\in \overline{N_{\lambda}}$ for all $\overline{n}\in \overline{N_{\lambda}}$ if and only if $q\in R_{\lambda}$.
\end{lemma}
\begin{proof}
Fix $t_0\in T_{\lambda}$ and $g\in G$. There exists $t_1\in T_{\lambda},h\in H_{\lambda},$ and $m\in \ll\mathcal{N} \rr$ such that
\begin{equation}\label{find s}
gt_0=t_1hm.
\end{equation}
Consider the summand $\overline{N^{t_0}_{\lambda}}$. For all $n\in N_{\lambda}$,
$$g\circ\overline{t_0nt^{-1}_0}=\overline{gt_0nt^{-1}_0g^{-1}}=\overline{t_1hmnm^{-1}h^{-1}t_1^{-1}}=\overline{t_1hnh^{-1}t_1^{-1}}\in \overline{N^{t_1}_{\lambda}},$$
where the fact that the action of $\ll\mathcal{N}\rr$ acts trivially on $Rel(G,\ll N \rr)$ is used in the second equality. Hence, $g\circ (\overline{N^{t_0}_{\lambda}})\subset \overline{N^{t_1}_{\lambda}}$. As $\overline{M_{\lambda}}=\bigoplus_{t\in T_{\lambda}}\overline{N^t_{\lambda}}$, it follows that $\overline{M_{\lambda}}$ is $G$-invariant and thus $\overline{M_{\lambda}}$ is also $Q$-invariant.
The above paragraph shows that $g$ maps $\overline{N^{t_0}_{\lambda}}$ into $\overline{N^{t_1}_{\lambda}}$. Actually, $g(\overline{N^{t_0}_{\lambda}})=\overline{N^{t_1}_{\lambda}}$. Indeed, given $n\in N_{\lambda}$, we find an element $x$ of $N^{t_0}_{\lambda}$ such that $g\circ \overline{x}=\overline{n^{t_1}}$. Let $x=n^{t_0h^{-1}}$. Note that $n^{h^{-1}}\in N_{\lambda}$, as $N_{\lambda}$ is normal in $H_{\lambda}$. Thus, $x\in N^{t_0}_{\lambda}$. Direct computation shows
$$g\circ\overline{x}=\overline{gxg^{-1}}=\overline{gt_0(h^{-1}nh)t^{-1}_0g^{-1}}=\overline{t_1hm(h^{-1}nh)m^{-1}h^{-1}t^{-1}_1}=\overline{t_1h(h^{-1}nh)h^{-1}t^{-1}_1}=\overline{n^{t_1}},$$
where the fact that the action of $\ll\mathcal{N}\rr$ on $Rel(G,\ll N \rr)$ is trivial is used in the second equality. Hence, $g\circ \overline{x}=\overline{n^{t_1}}$.
As a consequence, $g\circ (\overline{N^{t_0}_{\lambda}})=\overline{N^{t_1}_{\lambda}}$, i.e., the action of $G$ on $\overline{M_{\lambda}}$ permutes the summands $\overline{N^t_{\lambda}},t\in T_{\lambda}$. In fact, this permutation is transitive: Let $t$ be any element of $T_{\lambda}$. We wish to find an element of $G$ which maps $\overline{N^{t_0}_{\lambda}}$ to $\overline{N^t_{\lambda}}$. This can be done by $tt^{-1}_0$:
$$tt^{-1}_0\circ \overline{N^{t_0}_{\lambda}}=\overline{N^t_{\lambda}}.$$
Thus, the action of $G$ on $\overline{M_{\lambda}}$ transitively permutes the summands $\overline{N^t_{\lambda}},t\in T_{\lambda}$. The same is thus true for the action of $Q$ on $\overline{M_{\lambda}}$.
Clearly, for the action of $G$ on $\overline{M_{\lambda}}$, the isotropy group of $\overline{N_{\lambda}}$ contains $H_{\lambda}\ll \mathcal{N} \rr$. Observe that in equation \eqref{find s}, if $t_0=1$ and $g\not\in H_{\lambda}\ll \mathcal{N} \rr$, then $t_1\neq 1$ as $t^{-1}_1g\in H_{\lambda}\ll \mathcal{N} \rr$. It follows that
$$g\circ \overline{N_{\lambda}}=\overline{N^{t_1}_{\lambda}}\neq \overline{N_{\lambda}},$$
i.e., $g$ does not fix $\overline{N_{\lambda}}$ setwise. Therefore, for the action of $G$ on $\overline{M_{\lambda}}$, the isotropy group of $\overline{N_{\lambda}}$ is $H_{\lambda}\ll \mathcal{N} \rr$. As a consequence, for the action of $Q$ on $\overline{M_{\lambda}}$, the isotropy group of $\overline{N_{\lambda}}$ is $R_{\lambda}$.
\end{proof}
Recall that if $\mathcal{O}$ is a ring, $\mathcal{D}$ is a subring of $\mathcal{O}$, and $A$ is a $\mathcal{D}$-module, the \textit{induced module of} $A$ \textit{from} $\mathcal{D}$ \textit{to} $\mathcal{O}$, denoted as $Ind_{\mathcal{D}}^{\mathcal{O}}A$, is the tensor product $\mathcal{O}\bigotimes_{\mathcal{D}}A$. If $\mathcal{O},\mathcal{D}$ are integral group rings, we simplify notations by dropping $\mathbb{Z}$, e.g., we write $Ind^G_H$ instead of $Ind^{\mathbb{Z}G}_{\mathbb{Z}H}$. For $\lambda\in\Lambda$, Lemma \ref{permute factors}, together with the following Proposition \ref{induce}, which is a well-known characterization of induced modules (for example, see \cite[Proposition 5.3 of Chapter III]{brown1982cohomology}), implies that $\overline{M_{\lambda}}\cong Ind_{R_{\lambda}}^QRel(H_{\lambda},N_{\lambda})$ as $\mathbb{Z}Q$-modules.
\begin{proposition}\label{induce}
Let $G$ be a group and $A$ be a $\mathbb{Z}G$-module. Suppose that the underlying abelian group of $A$ is a direct sum $\bigoplus_{i\in I}A_i$ and that the $G$-action transitively permutes the summands. If $H\leqslant G$ is the isotropy group of $A_j$ for some $j\in I$. Then $A_j$ is a $\mathbb{Z}H$-module and $A\cong Ind^G_H A_j$ as $\mathbb{Z}G$-modules.
\end{proposition}
\begin{proof}[Proof of Proposition \ref{genmodule}]
For every $\lambda\in\Lambda$, $\overline{M_{\lambda}}\cong Ind_{R_{\lambda}}^QRel(H_{\lambda},N_{\lambda})$ as $\mathbb{Z}Q$-modules. Thus, $$Rel(G,\ll \mathcal{N} \rr)=\bigoplus_{\lambda\in\Lambda}\overline{M_{\lambda}}\cong \bigoplus_{\lambda\in\Lambda}Ind_{R_{\lambda}}^QRel(H_{\lambda},N_{\lambda})$$ as $\mathbb{Z}Q$-modules.
\end{proof}
\begin{example}
Let $\mathcal{G}$ be a graph of groups, $\pi_1(\mathcal{G})$ be the fundamental group of $\mathcal{G}$, $\{G_v\}_{v\in V\mathcal{G}}$ be the collection of vertex subgroups, and $\{G_e\}_{e\in E\mathcal{G}}$ be the collection of edge subgroups. By \cite[Example 4.12]{dahmani2017hyperbolically}, $\{G_v\}_{v\in V\mathcal{G}}\hookrightarrow_{wh}(\pi_1(\mathcal{G}),X)$ provided that the subset $X\subset \pi_1(\mathcal{G})$ consists of stable letters (i.e., generators corresponding to edges of $\mathcal{G}\setminus T\mathcal{G}$, where $T\mathcal{G}$ is a spanning tree of $\mathcal{G}$), and the corresponding relative metric on a vertex group $G_v$ corresponding to a vertex $v\in V\mathcal{G}$ is bi-Lipschitz equivalent to the word metric with respect to the union of the edge subgroups of $G_v$ corresponding to edges incident to $v$. Thus, we have the following corollary to Theorems \ref{Dehn filling}, \ref{main} and Proposition \ref{genmodule}.
\end{example}
\begin{corollary}\label{graph of groups}
Let $\mathcal{G}$ be a graph of groups, $\pi_1(\mathcal{G})$ be the fundamental group of $\mathcal{G}$, $\{G_v\}_{v\in V\mathcal{G}}$ be the collection of vertex subgroups, and $\{G_e\}_{e\in E\mathcal{G}}$ be the collection of edge subgroups. Suppose that, for every $v\in V\mathcal{G}$, $N_v$ is normal subgroup of $G_v$ with
$$N_v\cap\langle G_e, v\in e \rangle=\emptyset.$$
Then the triple $(G,\{G_v\}_{v\in V\mathcal{G}},\{N_v\}_{v\in V\mathcal{G}})$ has the Cohen-Lyndon property, and $Rel(G,\ll\mathcal{N}\rr)\cong \bigoplus Ind^Q_{Q_v}Rel(G_v,N_v)$ as $\mathbb{Z}Q$-modules, where $\mathcal{N}=\bigcup_{v\in V\mathcal{G}}N_v,Q=G/\ll\mathcal{N}\rr$, and $Q_v=G_v/N_v$ for $v\in V\mathcal{G}$.
\end{corollary}
In particular,
\begin{corollary}\label{amalproduct}
Let $G=A\ast_C B$ be an amalgamated free product. If $N\lhd A$ and $N\cap C=\{1\}$, then $(G,A,N)$ has the Cohen-Lyndon property, and $Rel(G,\ll N \rr)\cong Ind^Q_R Rel(A,N)$ as $\mathbb{Z}Q$-modules, where $Q=G/\ll N \rr$ and $R=A/N$.
\end{corollary}
\begin{corollary}\label{HNN}
Let $G=H\ast_t$ be an HNN-extension with associated subgroups $A,B\leqslant H$. If $N\lhd H$ and $N\cap(A\cup B)=\{1\}$, then $(G,H,N)$ has the Cohen-Lyndon property, and $Rel(G,\ll N \rr)\cong Ind^Q_R Rel(H,N)$ as $\mathbb{Z}Q$-modules, where $Q=G/\ll N \rr$ and $R=H/N$.
\end{corollary}
Alternatively, Corollary \ref{amalproduct} can be deduced from \cite{karrass1970subgroups} and both of Corollaries \ref{amalproduct},\ref{HNN} can be deduced from the Bass-Serre theory.
\bibliographystyle{alpha}
|
\section{\texorpdfstring{$\lambda$}{Lambda}-signatures and \texorpdfstring{$\lambda$}{lambda}-theories}
\begin{definition}[Arity]
A simply-typed first-order arity for a set $T$ of atomic types is a pair
$\alpha\equiv\Parens{\vec{\sigma},\tau}$ of a list of types and a types. We write
$\Ar{T}\eqdef{}T^\star\times T$ for the set of such arities.
\end{definition}
\begin{definition}[Many-typed signature]
Following~\citet{jacobs:1999}, a \emph{many-typed signature}
$\Sigma\equiv\Parens{\Sorts,\Ops}$ is a set of atomic types $\Sorts$ together with an
arity-indexed family of sets of operations $\Ops$, taking each operation $\vartheta$ to $\Ops(\vartheta)\in\Ar{\Sorts}$. More abstractly, the category $\SIGN$ of such signatures
arises as the pullback of the fundamental fibration along the arity
endofunctor.
\[
\begin{tikzcd}[sep=huge]
\SIGN
\arrow[r, "\Ops"]
\arrow[d, -{Triangle[open]}, swap, "\Sorts"]
\arrow[dr, phantom, pos = 0, "\lrcorner"]
& \ARR{\SET}
\arrow[d, -{Triangle[open]}, "\Cod"]
\\
\SET
\arrow[r, swap,"\Ar{\Parens{-}}"]
& \SET
\end{tikzcd}
\]
\end{definition}
From a collection of atomic types $\Sorts$, we generate the type structure of the
$\lambda$-calculus as the least set $\Ty{\Sorts}$ closed under the following
formation rules:
\begin{mathpar}
\inferrule{
\tau\in\Sorts
}{
\tau\in\Ty{\Sorts}
}
\and
\inferrule{
\sigma\in\Ty{\Sorts}
\\
\tau\in\Ty{\Sorts}
}{
{\sigma\times\tau}\in\Ty{\Sorts}
}
\and
\inferrule{
\sigma\in\Ty{\Sorts}
\\
\tau\in\Ty{\Sorts}
}{
{\sigma\to\tau}\in\Ty{\Sorts}
}
\end{mathpar}
\subsection{The clone of a \texorpdfstring{$\lambda$}{lambda}-signature}
From a $\lambda$-signature $\Sigma$ we can freely generate a special family of
sets $\Clone{\Sigma}{\Gamma,\tau}$ called its \emph{clone}, indexed in
$\Parens{\Gamma,\tau}\in\Ar{\Ty{\Sorts[\Sigma]}}$; simultaneously, we define the indexed set
of substitutions $\SbClone{\Sigma}{\Gamma,\Delta}$, indexed in
$\Parens{\Gamma,\Delta}\in\Ty{\Sorts[\Sigma]}^\star\times\Ty{\Sorts[\Sigma]}^\star$. In our
presentation, we choose to use \emph{explicit substitutions} rather than
implicit substitutions, because they are more abstract and allow a formulation
without explicit reference to preterms. Additionally, explicit substitutions
scale up to the metatheory of dependent type theory in a way that the implicit
(admissible) notion of substitution on preterms cannot.
We will write $\Seq{\Sigma}{\Gamma}{t}{\tau}$ and
$\Seq{\Sigma}{\Gamma}{\delta}{\Delta}$ to mean that
$t\in\Clone{\Sigma}{\Gamma,\tau}$ and
$\delta\in\SbClone{\Sigma}{\Gamma,\Delta}$ respectively, and
$\SeqEq{\Sigma}{\Gamma}{s}{t}{\tau}$ and
$\SeqEq{\Sigma}{\Gamma}{\delta_0}{\delta_1}{\Delta}$ to mean that $s$ and $t$
are equal elements of $\Clone{\Sigma}{\Gamma,\tau}$ and $\delta_0$ and
$\delta_1$ are equal elements of $\SbClone{\Sigma}{\Gamma,\Delta}$
respectively. In our notation, we \emph{presuppose} that $s$ and $t$ are
elements of the clone when we state that they are equal.
The clone of $\Sigma$ is defined as a quotient using the following indexed
inductive definition:
\begin{mathpar}
\inferrule[variable]{}{
\Seq{\Sigma}{\Gamma,\tau}{\Var}{\tau}
}
\and
\inferrule[operation]{
\Ops[\Sigma](\vartheta) \equiv (\Gamma,\tau)
}{
\Seq{\Sigma}{\Gamma}{\vartheta}{\tau}
}
\and
\inferrule[subst]{
\Seq{\Sigma}{\Gamma}{\delta}{\Delta}
\\
\Seq{\Sigma}{\Delta}{t}{\tau}
}{
\Seq{\Sigma}{\Gamma}{t[\delta]}{\tau}
}
\\
\inferrule[abstraction]{
\Seq{\Sigma}{\Gamma,\sigma}{t}{\tau}
}{
\Seq{\Sigma}{\Gamma}{
\ObjLam{\sigma}{t}
}{\sigma\to\tau}
}
\and
\inferrule[application]{
\Seq{\Sigma}{\Gamma}{s}{\sigma\to\tau}
\\
\Seq{\Sigma}{\Gamma}{t}{\sigma}
}{
\Seq{\Sigma}{\Gamma}{s(t)}{\tau}
}
\and
\inferrule[pair]{
\Seq{\Sigma}{\Gamma}{s}{\sigma}
\\
\Seq{\Sigma}{\Gamma}{t}{\tau}
}{
\Seq{\Sigma}{\Gamma}{(s,t)}{\sigma\times\tau}
}
\and
\inferrule[proj1]{
\Seq{\Sigma}{\Gamma}{s}{\sigma\times\tau}
}{
\Seq{\Sigma}{\Gamma}{s.1}{\sigma}
}
\and
\inferrule[proj2]{
\Seq{\Sigma}{\Gamma}{s}{\sigma\times\tau}
}{
\Seq{\Sigma}{\Gamma}{s.2}{\tau}
}
\\
\inferrule[sb/idn]{}{
\Seq{\Sigma}{\Gamma}{\SbIdn{\Gamma}}{\Gamma}
}
\and
\inferrule[sb/proj]{}{
\Seq{\Sigma}{\Gamma,\tau}{\SbProj}{\Gamma}
}
\and
\inferrule[sb/ext]{
\Seq{\Sigma}{\Gamma}{\delta}{\Delta}
\\
\Seq{\Sigma}{\Gamma}{t}{\tau}
}{
\Seq{\Sigma}{\Gamma}{\SbExt{\delta}{t}}{\Delta,\tau}
}
\and
\inferrule[sb/cmp]{
\Seq{\Sigma}{\Gamma}{\delta}{\Delta}
\\
\Seq{\Sigma}{\Delta}{\xi}{\Xi}
}{
\Seq{\Sigma}{\Gamma}{\xi\circ\delta}{\Xi}
}
\end{mathpar}
Before we define the equivalence relation by which the clone is quotiented, it
will be useful to define an auxiliary substitution for De Bruijn lifting:
\begin{align*}
\FmtTm{\SbLift{\gamma}} &\eqdef \FmtTm{\SbExt{(\SbProj\circ\gamma)}{\Var}}
\end{align*}
Next, we generate equivalence relations on terms and substitutions from the following rules:
\begin{mathpar}
\inferrule[app/beta]{}{
\SeqEq{\Sigma}{\Gamma}{
\Parens{\ObjLam{\sigma}{t}}\Parens{s}
}{
t[\SbExt{\SbIdn{\Gamma}}{s}]
}{\tau}
}
\and
\inferrule[abs/eta]{}{
\SeqEq{\Sigma}{\Gamma}{t}{
\ObjLam{\sigma}{
(t[\SbProj{}])(\Var)
}
}{\sigma\to\tau}
}
\and
\inferrule[fst/beta]{}{
\SeqEq{\Sigma}{\Gamma}{(s,t).1}{s}{\sigma}
}
\and
\inferrule[snd/beta]{}{
\SeqEq{\Sigma}{\Gamma}{(s,t).2}{t}{\tau}
}
\and
\inferrule[pair/eta]{}{
\SeqEq{\Sigma}{\Gamma}{t}{(t.1,t.2)}{\sigma\times\tau}
}
\and
\inferrule[sb/cmp/idn/l]{}{
\SeqEq{\Sigma}{\Gamma}{\SbIdn{\Delta}\circ\delta}{\delta}{\Delta}
}
\and
\inferrule[sb/cmp/idn/r]{}{
\SeqEq{\Sigma}{\Gamma}{\delta\circ\SbIdn{\Gamma}}{\delta}{\Delta}
}
\and
\inferrule[sb/cmp/assoc]{}{
\SeqEq{\Sigma}{\Gamma}{\gamma\circ(\delta\circ\xi)}{(\gamma\circ\delta)\circ\xi}{\Delta}
}
\and
\inferrule[sb/cmp/proj]{}{
\SeqEq{\Sigma}{\Gamma}{\SbProj\circ\SbExt{\delta}{t}}{\delta}{\Delta}
}
\and
\inferrule[sb/cmp/dot]{}{
\SeqEq{\Sigma}{\Gamma}{\delta\circ(\SbExt{\xi}{t})}{\SbExt{(\delta\circ\xi)}{t[\delta]}}{\Delta}
}
\and
\inferrule[sb/var/idn]{}{
\SeqEq{\Sigma}{\Gamma}{\Var[\SbIdn{\Gamma}]}{\Var}{\tau}
}
\and
\inferrule[sb/var/ext]{}{
\SeqEq{\Sigma}{\Gamma}{\Var[\SbExt{\delta}{t}]}{t}{\tau}
}
\and
\inferrule[sb/abs]{}{
\SeqEq{\Sigma}{\Gamma}{
\Parens{\ObjLam{\sigma}{t}}[\delta]
}{
\ObjLam{\sigma}{
t[\SbLift{\delta}]
}
}{\sigma\to\tau}
}
\and
\inferrule[sb/app]{}{
\SeqEq{\Sigma}{\Gamma}{(t(s))[\delta]}{t[\delta](s[\delta])}{\tau}
}
\and
\inferrule[sb/pair]{}{
\SeqEq{\Sigma}{\Gamma}{(s,t)[\delta]}{(s[\delta],t[\delta])}{\sigma\times\tau}
}
\and
\inferrule[sb/proj1]{}{
\SeqEq{\Sigma}{\Gamma}{t.1[\delta]}{(t[\delta]).1}{\sigma}
}
\and
\inferrule[sb/proj2]{}{
\SeqEq{\Sigma}{\Gamma}{t.2[\delta]}{(t[\delta]).2}{\tau}
}
\end{mathpar}
We omit the congruence cases for brevity. The clone of $\Sigma$ is now defined
as the indexed family of quotients generated by the formation and definitional
equivalence rules given above.
\paragraph{Representation of the quotient}
In these notes, we will not dwell on the technical representation of the
quotiented terms. However, we will remark that the most convenient induction
principle for the quotiented syntax would arrive from a presentation as a
\emph{quotient inductive type}; moreover, our inductive definition falls under
a schema for \emph{finitary} quotient inductive types which is already known to
be interpretable in the setoid model of type
theory~\citep{dybjer-moeneclaey:2018}.
\subsection{The classifying category of a \texorpdfstring{$\lambda$}{lambda}-signature}
We can see that the language of substitutions above has the structure of
category; this category is in fact called $\ClCat{\Sigma}$, the
\emph{classifying category} or \emph{Lawvere category} of the
$\lambda$-signature $\Sigma$. The classifying category is also just
called the (pure) $\lambda$-\emph{theory} generated by the signature.
Concretely, it has contexts $\Gamma$ as objects, and equivalence classes of
substitutions $\Seq{\Sigma}{\Gamma}{\delta}{\Delta}$ as morphisms.
\begin{proposition}
The classifying category $\ClCat{\Sigma}$ is cartesian closed.
\end{proposition}
\subsection{The category of renamings}\label{sec:ren-cat}
Every $\lambda$-signature gives rises to another category, namely the
\emph{category of renamings} $\Ren{\Sigma}$. Abstractly, this can be characterized as the
free strictly associative cartesian category generated by $\Ty{\Sorts[\Sigma]}$;
concretely, its objects are contexts $\Gamma$, and its morphisms
$\psi:\Hom[\Ren{\Sigma}]{\Gamma}{\Delta}$ are vectors of projections (indices) from $\Gamma$
into the types in $\Delta$.
An explicit presentation of $\Ren{\Sigma}$ appears
in~\cite{fiore:2002,fiore:2005} as the opposite of the comma category
$\lfloor-\rfloor\downarrow\mathsf{Const}(\Ty{\Sorts[\Sigma]})$, where
$\lfloor-\rfloor:\mathbb{F}\to\SET$ takes a finite cardinal to a set. Fiore
writes $\mathbb{F}\downarrow\Ty{\Sorts[\Sigma]}$ for this comma construction, and
$\mathbb{F}[\Ty{\Sorts[\Sigma]}]$ for its opposite. Another possible presentation of
$\Ren{\sigma}$ is as the subcategory of $\ClCat{\Sigma}$ which has the same
objects, but whose morphisms are all of the form $\SbIdn{\Gamma}$ or
$\SbExt{\SbExt{\SbIdn{\Gamma}}{t_0}}{\cdots{}t_n}$ for some $n> 0$, with $t_i$ of the
form $\Var[\SbProj^k]$, writing $\SbProj^k$ for the $k$-fold composition of
$\SbProj$ with itself.
\section{Normalization and the Yoneda embedding}\label{sec:yoneda}
Working in a constructive metatheory, we can see that the intensional content
of a certain natural isomorphism hides within it a normalization function for
the lambda calculus over $\Sigma$, as observed
in~\cite{cubric-dybjer-scott:1998}. Let $\Psh{\ClCat{\Sigma}}$ denote the
category of presheaves over the classifying category of $\Sigma$.
The Yoneda embedding is a cartesian closed functor
$\Yo:\ClCat{\Sigma}\to\Psh{\ClCat{\Sigma}}$, defined as $\Yo{\Delta} =
\Hom[\ClCat{\Sigma}]{-}{\Delta}$. Within the presheaf topos, it is easiest to think of
the representable objects $\Yo{\Delta}$ as the ``type of substitutions into
$\Delta$''.
There is another way to define the Yoneda embedding, which we will see is
naturally isomorphic to what is written above. In this version, we define a
functor $\FancyYo:\ClCat{\Sigma}\to\Psh{\ClCat{\Sigma}}$ by recursion on the
objects of $\ClCat{\Sigma}$. For an atomic type $\tau\in{}\Sorts[\Sigma]$, $\FancyYo\tau
= \Yo\tau$; but the remainder of the cases are defined using the cartesian
closed structure of the presheaf topos instead of the cartesian closed
structure of the classifying category:
\begin{align*}
\FancyYo(\sigma\times\tau) &= \FancyYo{\sigma}\times\FancyYo{\tau}
\\
\FancyYo(\sigma\to\tau) &= {\FancyYo{\tau}}^{\FancyYo{\sigma}}
\end{align*}
Now, because the Yoneda embedding is cartesian closed, it is easy to see that
we have a natural isomorphism $\Yo\cong\FancyYo$. However, observe that the
elements in the fibers of $\FancyYo$ are not $\lambda$-terms, but a kind of
eta-long Boehm-tree representation of $\lambda$-terms.
That is, whereas the action of $\Yo$ on a syntactic morphism/term
$\Seq{\Sigma}{\Gamma}{t}{\sigma\times\tau}$ is to simply embed $t$ into the
appropriate presheaf fiber, the action of $\FancyYo$ on the same term must take
$t$ to an element of $\FancyYo{\sigma}\times\FancyYo{\tau}$, that is, an actual
\emph{pair}. Considering the case where $t$ is actually a variable, we can see
that the action of these two embeddings is intensionally quite different.
The other side of the natural isomorphism is witnessed by a ``readback''
operation, which reads one of these expanded Boehm trees into a syntactic term
(which can be seen to be $\beta$-normal and $\eta$-long). The normalization
operation obtained by composing these operations can be seen to be an instance
of \emph{normalization by evaluation}.
The problem with this kind of result, however, is that the categories have
quotiented too much for us to be able to say in mathematical (rather than
merely intuitive) language that we have obtained a normalization function. In
particular, the normalization operation that we describe above is actually
equal as a function to the identity. This is because the classifying category
is already quotiented by definitional equivalence.
As summarized in~\citet{streicher:1998}, there are two ways out of this
situation. One is to use a higher-dimensional structure, such as partial
equivalence relations or setoids, in order to structure the ambient category
theory; then, in the spirit of Bishop's constructive mathematics, we can
observe the intension of the normalization operation at the same time as seeing
that it is the identity in its extension. This approach was carried out
in~\cite{cubric-dybjer-scott:1998} using P-category theory, a variant of
E-category theory in which setoids are replaced by PERs.
Another more direct way is obtained from the \emph{gluing construction} in
category theory, where we will choose a different semantic domain which allows
us to see the difference between the two ways of interpreting syntax into the
presheaf category. This was carried out in detail
in~\citet{altenkirch-hofmann-streicher:1995}, but in a manner that is
unfortunately different enough from the classical construction that it is
unclear how it relates. In~\citet{fiore:2002}, normalization by evaluation for
typed lambda calculus is related explicitly to gluing; what we present in these
notes can be seen as an explicit instantiation of Fiore's frameork.
\section{Normalization by gluing}\label{sec:nbg}
To resolve the problem described above in Section~\ref{sec:yoneda}, we will
work with a more refined base category, namely the category of renamings
$\Ren{\Sigma}$ defined in Section~\ref{sec:ren-cat}. First observe that there
is an inclusion of categories $i : \Ren{\Sigma}\to\ClCat{\Sigma}$, since every
context renaming can be represented as a substitution, a sequence of extensions
by variables.
We have a reindexing functor $i^*:\Psh{\ClCat{\Sigma}}\to\Psh{\Ren{\Sigma}}$ by
precomposition. Composing with the Yoneda embedding, we can define a new
functor $\TM:\ClCat{\Sigma}\to\Psh{\Ren{\Sigma}}$:
\[
\begin{tikzcd}[sep=huge,cramped]
\ClCat{\Sigma}
\arrow[d,hook,swap,"\Yo"]
\arrow[dr,dashed,"\TM"]
\\
\Psh{\ClCat{\Sigma}}
\arrow[r,swap,"i^*"]
&
\Psh{\Ren{\Sigma}}
\end{tikzcd}
\]
\paragraph{Relative hom functor}
As described in~\citet{fiore:2002}, the functor $\TM$ is called the ``relative
hom functor'' of $i$, taking $\Delta:\ClCat{\Sigma}$ to
$\Hom[\ClCat{\Sigma}]{i(-)}{\Delta}$. In~\citet{fiore:2002}, this functor is
written $\langle{i}\rangle$, whereas we write $\TM$ in order to suggest the
intuition that it defines a presheaf of open terms.
We have constructed $\TM$ from the perspective of ``adjusting'' the Yoneda
embedding from $\ClCat{\Sigma}$, but~\citet{fiore:2002} explains another
characterization of the same functor from the perspective of the Yoneda
embedding from $\Ren{\Sigma}$. In particular, $\TM$ is the left Kan extension
of $\Yo:\Ren{\Sigma}\to\Ren{\Sigma}$ along $i$:
\[
\begin{tikzcd}[cramped]
\Ren{\Sigma}
\arrow[rr,hook,"\Yo"]
\arrow[dr, swap,"i"]
&
\arrow[d,phantom,"\Downarrow"]
&
\Psh{\Ren{\Sigma}}
\\
&
\ClCat{\Sigma}
\arrow[ur,dashed,swap,"\TM"]
\end{tikzcd}
\]
\subsection{Presheaves of neutrals and normals}
In $\Psh{\Ren{\Sigma}}$, we can construct presheaves of neutral terms and
normal terms for each type; note that such presheaves cannot be defined in
$\Psh{\ClCat{\Sigma}}$, because they crucially cannot be closed under arbitrary
substitutions (whereas they happen to be closed under renamings). The fibers of
these presheaves will have the property that the equality relation for their
elements is \emph{discrete}.
To be concrete, let us begin by defining some restricted typing judgments for
neutrals and normals.
\begin{mathparpagebreakable}
\inferrule[variable]{}{
\SeqNE{\Sigma}{\Gamma}{\Var[\SbProj^k]}{\Gamma_{\vert{\Gamma}\vert-k-1}}
}
\and
\inferrule[operation]{
\Ops[\Sigma](\vartheta) \equiv (\Delta,\tau)
\\
\SeqNF{\Sigma}{\Gamma}{\delta}{\Delta}
}{
\SeqNE{\Sigma}{\Gamma}{\vartheta[\delta]}{\tau}
}
\and
\inferrule[app]{
\SeqNE{\Sigma}{\Gamma}{t}{\sigma\to\tau}
\\
\SeqNF{\Sigma}{\Gamma}{s}{\sigma}
}{
\SeqNE{\Sigma}{\Gamma}{t(s)}{\tau}
}
\and
\inferrule[proj1]{
\SeqNE{\Sigma}{\Gamma}{t}{\sigma\times\tau}
}{
\SeqNE{\Sigma}{\Gamma}{t.1}{\sigma}
}
\and
\inferrule[proj2]{
\SeqNE{\Sigma}{\Gamma}{t}{\sigma\times\tau}
}{
\SeqNE{\Sigma}{\Gamma}{t.2}{\tau}
}
\\
\inferrule[shift]{
\SeqNE{\Sigma}{\Gamma}{t}{\tau}
\\
\tau\in{}\Sorts[\Sigma]
}{
\SeqNF{\Sigma}{\Gamma}{t}{\tau}
}
\and
\inferrule[abstraction]{
\SeqNF{\Sigma}{\Gamma,\sigma}{t}{\tau}
}{
\SeqNF{\Sigma}{\Gamma}{
\ObjLam{\sigma}{t}
}{\sigma\to\tau}
}
\and
\inferrule[pair]{
\SeqNF{\Sigma}{\Gamma}{s}{\sigma}
\\
\SeqNF{\Sigma}{\Gamma}{t}{\tau}
}{
\SeqNF{\Sigma}{\Gamma}{(s,t)}{\sigma\times\tau}
}
\\
\inferrule[sb/nf/proj]{}{
\SeqNF{\Sigma}{\Gamma}{\SbProj^{\vert\Gamma\vert}}{[\,]}
}
\and
\inferrule[sb/nf/ext]{
\SeqNF{\Sigma}{\Gamma}{\delta}{\Delta}
\\
\SeqNF{\Sigma}{\Gamma}{t}{\tau}
}{
\SeqNF{\Sigma}{\Gamma}{\SbExt{\delta}{t}}{\Delta,\tau}
}
\\
\inferrule[sb/ne/proj]{}{
\SeqNE{\Sigma}{\Gamma}{\SbProj^{\vert\Gamma\vert}}{[\,]}
}
\and
\inferrule[sb/ne/ext]{
\SeqNE{\Sigma}{\Gamma}{\delta}{\Delta}
\\
\SeqNE{\Sigma}{\Gamma}{t}{\tau}
}{
\SeqNE{\Sigma}{\Gamma}{\SbExt{\delta}{t}}{\Delta,\tau}
}
\end{mathparpagebreakable}
\paragraph{Admissible substitutions}
We have restricted the language of normal substitutions to consist in vectors
of terms, constructed using the \textsc{sb/proj} and \textsc{sb/ext} rules. The
identity substitution is \emph{admissible} as a neutral substitution, but is
not one of the generators. We define
$\SeqNE{\Sigma}{\Gamma}{\SbIdn{\Gamma}}{\Gamma}$ by recursion on $\Gamma$ as follows:
\begin{align*}
\FmtNE{\SbIdn{[\,]}} &= \FmtNE{\SbProj^0}
\\
\FmtNE{\SbIdn{\Gamma,\tau}} &= \FmtNE{\SbExt{\SbIdn{\Gamma}}{\Var[\SbProj^0]}}
\end{align*}
\paragraph{\texorpdfstring{$\eta$}{Eta}-long normal forms}
Observe that we have ensured an $\eta$-long normal form by restricting the
\textsc{shift} rule above to apply only at atomic types. It is easy to see that
these judgments are closed under context renamings, i.e.\ support a
$\Ren{\Sigma}$-action.
Therefore, we can use these judgments as the raw material from which to build
the presheaves of neutrals and normals for each type $\tau\in\Ty{\Sorts[\Sigma]}$ as
follows:
\begin{align*}
\NE{\tau} &: \Psh{\Ren{\Sigma}}
\\
\NE{\tau}(\Gamma) &\equiv \SetCompr{\FmtNE{t}}{\SeqNE{\Sigma}{\Gamma}{t}{\tau}}
\\[6pt]
\NF{\tau} &: \Psh{\Ren{\Sigma}}
\\
\NF{\tau}(\Gamma) &\equiv \SetCompr{\FmtNF{t}}{\SeqNF{\Sigma}{\Gamma}{t}{\tau}}
\end{align*}
\subsection{Syntax with binding, internally}
So far we have developed three presheaves of syntax in $\Psh{\Ren{\Sigma}}$:
the presheaf of typed terms $\TM(\tau)$, the presheaf of neutrals $\NE{\tau}$
and the presheaf of normals $\NF{\tau}$. Using the internal language of the
functor category, we can justify a simpler ``higher-order'' notation for
working with elements of these presheaves
internally~\citep{hofmann:1999,fiore-plotkin-turi:1999,staton:2007,harper-honsell-plotkin:1993}.
First observe that exponentiation of a presheaf
$\mathcal{F}:\Psh{\Ren{\Sigma}}$ by a representable has a simpler
characterization using the Yoneda lemma (in fact, this works for any base
category that has finite products):
\begin{align*}
\mathcal{F}^{\Yo{\Delta}}(\Gamma)
&\cong \Hom[\Psh{\Ren{\Sigma}}]{\Yo\Gamma\times\Yo\Delta}{\mathcal{F}}
\\
&\cong \Hom[\Psh{\Ren{\Sigma}}]{\Yo(\Gamma\times\Delta)}{\mathcal{F}}
\\
&\cong \mathcal{F}(\Gamma\times\Delta)
\end{align*}
Writing $\VAR{\tau}$ for the representable presheaf
$\Yo{\tau}:\Psh{\Ren{\Sigma}}$ of variables, we can equivalently use
a higher-order notation for terms from inside the topos, with constructors like
the following:
\begin{align*}
\FmtNE{\Var} &: \VAR{\tau}\to\NE{\tau}
\\
\FmtNF{\lambda^{\Mute\sigma}} &: (\VAR{\sigma}\to\NF{\tau})\to\NF{\sigma\to\tau}
\\
&\ldots
\end{align*}
This is justified by the fact that all the generators of $\TM$, $\NE{}$ and
$\NF{}$ commute with the presheaf renaming action. When working internally, we
will implicitly use these notations as a simplifying measure.
We will also employ the internal substitution constructors $\FmtNF{\SbEmp} :
\mathbf{1}\to\NF{[\,]}$ and $\FmtNE{\SbEmp} : \mathbf{1}\to\NE{[\,]}$ defined as
follows:
\begin{align*}
\FmtNF{\SbEmp[\Gamma]}(\star) &= \FmtNF{\SbProj^{\vert\Gamma\vert}}
\\
\FmtNE{\SbEmp[\Gamma]}(\star) &= \FmtNE{\SbProj^{\vert\Gamma\vert}}
\end{align*}
\subsection{The gluing construction}
Next, we will construct the \emph{gluing category} which will serve as our
principal semantic domain for the model construction. Consider the comma
category $\GlCat{\Sigma} \equiv \Psh{\Ren{\Sigma}}\downarrow\TM$, which
``glues'' syntactic contexts together with their semantics in
presheaves.\footnote{Careful readers will note that this is a notation for the
actual instance of the comma construction,
$\mathbf{id}_{\Psh{\Ren{\Sigma}}}\downarrow\TM$.} Concretely, an object of
$\GlCat{\Sigma}$ is a tuple $(\mathcal{D}:\Psh{\Ren{\Sigma}},
\Delta:\ClCat{\Sigma}, \Quo{\Delta} : \mathcal{D}\to\TM(\Delta))$;
a morphism
$(\mathcal{G},\Gamma,\Quo{\Gamma})\to(\mathcal{D},\Delta,\Quo{\Delta})$ is a
commuting square of the following form, which we will suggestively write
$\Realize{\delta}{d}$:
\[
\begin{tikzcd}[sep=huge,cramped]
\mathcal{G}
\arrow[r,"\FmtVal{d}"]
\arrow[d,swap,"\Quo{\Gamma}"]
&
\mathcal{D}
\arrow[d,"\Quo{\Delta}"]
\\
\TM(\Gamma)
\arrow[r, swap, "\FmtTm{\delta}"]
&
\TM(\Delta)
\end{tikzcd}
\]
The gluing category $\GlCat{\Sigma}$ is the category of proof-relevant logical
predicates, and is known to be cartesian closed, and thence a model of simply
typed lambda calculus. To use this information to our advantage, we will need
to ``unearth'' its cartesian closed structure in explicit terms.
\paragraph{Presentation as a pullback}
Following~\citet{frey:2013}, we can give a more intuitive presentation of the
gluing construction as a pullback of the fundamental fibration along $\TM$:
\[
\begin{tikzcd}[sep=huge,cramped]
{\GlCat{\Sigma}}
\arrow[r, dashed]
\arrow[d, -{Triangle[open]}, dashed,swap,"\GlFib{\Sigma}"]
\arrow[dr, phantom, pos = 0, "\lrcorner"]
&
\ARR{\Psh{\Ren{\Sigma}}}
\arrow[d,-{Triangle[open]}, "\Cod"]
\\
\ClCat{\Sigma}
\arrow[r,swap,"\TM"]
&
\Psh{\Ren{\Sigma}}
\end{tikzcd}
\]
From the pullback above, we have the \emph{gluing fibration}
$\GlFib{\Sigma}:\GlCat{\Sigma}\to\ClCat{\sigma}$ which acts on objects
$(\mathcal{D},\Delta,\Quo{\Delta})$ by projecting $\Delta$, and on morphisms
$\Realize{\delta}{d}:\Hom[\GlCat{\Sigma}]{(\mathcal{G},\Gamma,q_\Gamma)}{(\mathcal{D},\Delta,\Quo{\Delta})}$ by projecting
$\delta$.\footnote{\citet{streicher:1998} calls this the ``codomain functor'',
but to avoid confusion with the codomain functor that it is a pullback of, we
use a different terminology.}
\subsection{Reification, reflection and logical predicates}
Observe that there are obvious natural embeddings
$\RbNF{\tau}:\NF{\tau}\hookrightarrow\TM(\tau)$ and
$\RbNE{\tau}:\NE{\tau}\hookrightarrow\TM(\tau)$ for each
$\tau\in\Ty{\Sorts[\Sigma]}$, called ``readback''.
In order to give an explicit character to the cartesian closed structure of
$\GlCat{\Sigma}$, we will define a proof-relevant family of logical predicates
$\Pred{\tau}:\Psh{\Ren{\Sigma}}$ by induction on $\tau\in\Ty{\Sorts[\Sigma]}$,
simultaneously exhibiting natural transformations
$\Reflect{\tau}:\NE{\tau}\to\Pred{\tau}$ (pronounced ``reflect'') and
$\Reify{\tau}:\Pred{\tau}\to\NF{\tau}$ (pronounced ``reify'') such that the
following triangle commutes:
\[
\begin{tikzcd}[sep=huge,cramped]
\NE{\tau}
\arrow[rr,"\Reify{\tau}\circ\Reflect{\tau}"]
\arrow[dr,hook,swap,"\RbNE{\tau}"]
&&
\NF{\tau}
\arrow[dl,hook,"\RbNF{\tau}"]
\\
&\TM(\tau)
\end{tikzcd}
\tag{reify-reflect yoga}
\]
\begin{remark}
An alternative to this approach is to
follow~\citet{altenkirch-hofmann-streicher:1995} and employ an ad hoc
``twisted gluing'' category, in which the data of the gluing objects contains
the reification and reflection maps. This has the benefit of leading to a
proof which is more self-contained, but the disadvantage is that it is not
clear how to connect this twisted gluing category to the classical
construction.
\end{remark}
\paragraph{Atomic types}
For an atomic type $\sigma\in{}\Sorts[\Sigma]$, we define $\Pred{\sigma}=\NF{\sigma}$,
$\Reflect{\tau}=1$, $\Reify{\tau}=1$; it is easy to see that the
reify-reflect yoga is upheld. Next, we come to the compound types.
\paragraph{Product types}
Fixing types $\sigma,\tau\in\Ty{\Sorts[\Sigma]}$, we define the logical predicate and the reflection and reification maps, using the internal language of $\Psh{\Ren{\Sigma}}$:
\begin{align*}
\Pred{\sigma\times\tau} &= \Pred{\sigma}\times\Pred{\tau}
\\
\Reflect{\sigma\times\tau}[t]
&=
\FmtVal{
\Parens{
\Reflect{\sigma}[t.1], \Reflect{\tau}[t.2]
}
}
\\
\Reify{\sigma\times\tau}[\FmtVal{v_0},\FmtVal{v_1}]
&=
\FmtNF{
\Parens{
\Reify{\sigma}[v_0],
\Reify{\tau}[v_1]
}
}
\end{align*}
To execute the reify-reflect yoga, working internally, we fix
$t:\NE{\sigma\times\tau}$; we need to observe that
$\RbNF{\sigma\times\tau}[\Reflect{\sigma\times\tau}[\Reify{\sigma\times\tau}[t]]]
= \RbNE{\sigma\times\tau}[t]$.
\begin{align*}
\RbNF{\sigma\times\tau}[\Reify{\sigma\times\tau}[\Reflect{\sigma\times\tau}[t]]]
&=
\RbNF{\sigma\times\tau}[
\Reify{\sigma\times\tau}[
\Reflect{\sigma}[t.1], \Reflect{\tau}[t.2]
]
]
\\
&=
\RbNF{\sigma\times\tau}[
\Reify{\sigma}[\Reflect{\sigma}[t.1]],
\Reify{\tau}[\Reflect{\tau}[t.2]]
]
\\
&=
\FmtTm{
\Parens{
\RbNF{\sigma}[\Reify{\sigma}[\Reflect{\sigma}[t.1]]],
\RbNF{\tau}[\Reify{\tau}[\Reflect{\tau}[t.1]]]
}
}
\\
&=
\FmtTm{
\Parens{
\RbNE{\sigma}[t.1],
\RbNE{\tau}[t.2]
}
}
\tag{i.h., i.h.}
\\
&=
\FmtTm{
\Parens{
(\RbNE{\sigma\times\tau}[t]).1,
(\RbNE{\sigma\times\tau}[t]).2
}
}
\\
&=
\RbNE{\sigma\times\tau}[t]
\tag{\textsc{pair/eta}}
\end{align*}
Above, the steps that commute readback of (neutrals, normals) with the syntax
of the $\lambda$-theory follow from the fact that normals and neutrals actually
embed directly into the syntax unchanged.
\paragraph{Function types}
To interpret function types, we cannot simply use the exponential in
$\Psh{\Ren{\Sigma}}$, as this would take us outside the realm of
\emph{definable} functions. In a move apparently inspired by Kreisel's
\emph{modified realizability}, we include in the logical predicate both a
definable function and its meaning, taking the pullback
\[
\begin{tikzcd}[cramped]
{\Pred{\sigma\to\tau}}
\arrow[r,dashed]
\arrow[d,dashed]
\arrow[dr, phantom, pos = 0, "\lrcorner"]
&
{\Pred{\tau}}^{\Pred{\sigma}}
\arrow[d,"\FmtVal\phi"]
\\
\TM(\sigma\to\tau)
\arrow[r,swap,"\FmtVal\psi"]
& \Parens{\TM(\tau)}^{\Pred{\sigma}}
\end{tikzcd}
\]
where for clarity, we define arrows $\FmtVal{\phi},\FmtVal{\psi}$ in the internal language of $\Psh{\Ren{\Sigma}}$ as follows:
\begin{align*}
\FmtVal{F}:{\Pred{\tau}}^{\Pred{\sigma}}\vdash\FmtVal{\phi} & \equiv
\MetaLam{\FmtVal{v}}{
\RbNF{\tau}[\Reify{\tau}[F(v)]]
}
\\
\FmtTm{t}:\TM(\sigma\to\tau)\vdash\FmtVal{\psi}&\equiv
\MetaLam{\FmtVal{v}}{
\FmtTm{
t\Parens{
\RbNF{\sigma}[\Reify{\sigma}[v]]
}
}
}
\end{align*}
Abusing notation slightly, we will write an element of $\Pred{\sigma\to\tau}$
as $\Realize{t}{F}$ where $\FmtTm{t}:\TM(\sigma\to\tau)$ and
$\FmtVal{F}:{\Pred{\tau}}^{\Pred{\sigma}}$.
Next, we need to define reflection of neutrals and reification into normals:
\begin{align*}
\Reflect{\sigma\to\tau}[t]
&=
\Realize{
\RbNE{\sigma\to\tau}[t]
}{
\MetaLam{\FmtVal{v}}{
\Reflect{\tau}[t(\Reify{\sigma}[v])]
}
}
\\
\Reify{\sigma\to\tau}[\Realize{t}{F}]
&=
\FmtNF{
\ObjLam{\sigma}{
\MetaLam{\FmtNF{x}}{
\Reify{\tau}[
F\Parens{
\Reflect{\sigma}[\Var(x)]
}
]
}
}
}
\end{align*}
To prove the reify-reflect yoga, (working internally) fix
$t:\NE{\sigma\to\tau}$.
\begin{align*}
\RbNF{\sigma\to\tau}[\Reify{\sigma\to\tau}[\Reflect{\sigma\to\tau}[t]]]
&=
\RbNF{\sigma\to\tau}[
\Reify{\sigma\to\tau}[
\Realize{
\RbNE{\sigma\to\tau}[t]
}{
\MetaLam{v}{
\Reflect{\tau}[
t(\Reify{\sigma}[v])
]
}
}
]
]
\\
&=
\RbNF{\sigma\to\tau}[
\ObjLam{\sigma}{
\MetaLam{x}{
\Reify{\tau}[
\Reflect{\tau}[
t\Parens{
\Reify{\sigma}[\Reflect{\sigma}[\Var(x)]]
}
]
]
}
}
]
\\
&=
\FmtTm{
\ObjLam{\sigma}{
\MetaLam{x}{
\RbNF{\tau}[
\Reify{\tau}[
\Reflect{\tau}[
t\Parens{
\Reify{\sigma}[
\Reflect{\sigma}[\Var(x)]
]
}
]
]
]
}
}
}
\\
&=
\FmtTm{
\ObjLam{\sigma}{
\MetaLam{x}{
\RbNE{\tau}[
t\Parens{
\Reify{\sigma}[
\Reflect{\sigma}[\Var(x)]
]
}
]
}
}
}
\tag{i.h.}
\\
&=
\FmtTm{
\ObjLam{\sigma}{
\MetaLam{x}{
\Parens{\RbNE{\tau}[t]}\Parens{
\RbNF{\sigma}[
\Reify{\sigma}[
\Reflect{\sigma}[\Var(x)]
]
]
}
}
}
}
\\
&=
\FmtTm{
\ObjLam{\sigma}{
\MetaLam{x}{
\Parens{\RbNE{\tau}[t]}\Parens{\RbNE{\sigma}[\Var(x)]}
}
}
}
\\
&=
\FmtTm{
\ObjLam{\sigma}{
\MetaLam{x}{
\Parens{\RbNE{\tau}[t]}\Parens{\Var(x)}
}
}
}
\tag{i.h.}
\\
&=
\RbNE{\tau}[t]
\tag{\textsc{abs/eta}}
\end{align*}
\paragraph{Contexts}
The interpretation is now extended to contexts $\Gamma\in\Ty{\Sorts[\Sigma]}^\star$,
which are the ``types'' of substitutions; the interpretation is essentially the
same as the one for products.
\begin{align*}
\Pred{[\,]} &= \mathbf{1}
\\
\Reflect{[\,]}[\SbEmp] &= \FmtVal{\star}
\\
\Reify{[\,]}[\star] &= \FmtNF{\SbEmp}
\\[6pt]
\Pred{\Gamma,\tau} &= \Pred{\Gamma}\times\Pred{\tau}
\\
\Reflect{\Gamma,\tau}[\SbExt{\gamma}{t}] &= (\Reflect{\Gamma}[\gamma], \Reflect{\tau}[t])
\\
\Reify{\Gamma,\tau}[g,v] &= \SbExt{\Reify{\Gamma}[g]}{\Reify{\tau}[v]}
\end{align*}
The reify-reflect yoga follows in exactly the same way as it did for products.
\bigskip
\bigskip
Observe that for any $\tau\in\Ty{\Sorts[\Sigma]}$, the triple
$\Sem{\tau}\equiv \Parens{\Pred{\tau},\tau,\Quo{\tau}\equiv\RbNF{\tau}\circ\Reify{\tau}}$ is an object in
$\GlCat{\Sigma}$.
This brings us to an explicit characterization of the cartesian closed
structure of $\GlCat{\Sigma}$.
\begin{theorem}
For $\sigma,\tau\in\Ty{\Sorts[\Sigma]}$, $\Sem{\sigma\times{}\tau}$ is the
cartesian product $\Sem{\sigma}\times\Sem{\tau}$ in $\GlCat{\Sigma}$.
\end{theorem}
\begin{proof}
We will establish that $\Sem{\sigma\times\tau}$ is the cartesian product
$\Sem{\sigma}\times\Sem{\tau}$ by exhibiting its universal property. We
need to exhibit a span in $\GlCat{\Sigma}$ with the following property for any
$D,d_1,d_2$:
\[
\begin{tikzcd}[sep=large,cramped]
D
\arrow[dr,dashed,"{\exists!\bar{d}}"]
\arrow[ddr,swap,bend right=20,"d_1"]
\arrow[drr,bend left=20,"d_2"]
\\
&
\Sem{\sigma\times\tau}
\arrow[d,"\pi_1"]
\arrow[r,"\pi_2"]
&
\Sem{\tau}
\\
&
\Sem{\sigma}
\end{tikzcd}
\]
The projections $\pi_1,\pi_2$ are the following commuting squares:
\begin{mathpar}
\begin{tikzcd}[sep=large]
\Pred{\sigma}\times\Pred{\tau}
\arrow[r,"{\FmtVal{(p,q)}\vdash{}\FmtVal{p}}"]
\arrow[d,swap,"\Quo{\sigma\times\tau}"]
&
\Pred{\sigma}
\arrow[d,swap,"\Quo{\sigma}"]
\\
\TM(\sigma\times\tau)
\arrow[r,swap,"{\FmtTm{t}\vdash\FmtTm{t.1}}"]
&
\TM(\sigma)
\end{tikzcd}
\and
\begin{tikzcd}[sep=large]
\Pred{\sigma}\times\Pred{\tau}
\arrow[r,"{\FmtVal{(p,q)}\vdash\FmtVal{q}}"]
\arrow[d,swap,"\Quo{\sigma\times\tau}"]
&
\Pred{\tau}
\arrow[d,swap,"\Quo{\tau}"]
\\
\TM(\sigma\times\tau)
\arrow[r,swap,"{\FmtTm{t}\vdash{}\FmtTm{t.2}}"]
&
\TM(\tau)
\end{tikzcd}
\end{mathpar}
We show that the first square commutes (the second is identical); fixing
$\FmtVal{p}:\Pred{\sigma},\FmtVal{q}:\Pred{\tau}$, we calculate.
\begin{align*}
\FmtTm{
\Parens{
\RbNF{\sigma\times\tau}[\Reify{\sigma\times\tau}[p,q]]
}.1
}
&=
\FmtTm{
\Parens{
\RbNF{\sigma\times\tau}[
\Reify{\sigma}[p],
\Reify{\tau}[q]
]
}.1
}
\tag{def.}
\\
&=
\FmtTm{
\Parens{
\RbNF{\sigma}[\Reify{\sigma}[p]],
\RbNF{\tau}[\Reify{\tau}[q]]
}.1
}
\tag{def.}
\\
&=
\FmtTm{
\Parens{
\RbNF{\sigma}[\Reify{\sigma}[p]]
}
}
\tag{\textsc{fst/beta}}
\end{align*}
Next, we need to show that there is a unique mediating arrow
$\bar{d}:D\to\Sem{\sigma\times\tau}$ such that the two triangles commute.
Unfolding what we are given, we have $D\equiv(\mathcal{D},\Delta,\Quo{\Delta})$
and two commuting squares:
\begin{mathpar}
\begin{tikzcd}[sep=large,cramped]
\mathcal{D}
\arrow[r,"\FmtVal{d_{00}}"]
\arrow[d,swap,"\Quo{\Delta}"]
&
\Pred{\sigma}
\arrow[d,swap,"\Quo{\sigma}"]
\\
\TM(\Delta)
\arrow[r,swap,"\FmtTm{d_{01}}"]
&
\TM(\sigma)
\end{tikzcd}
\and
\begin{tikzcd}[sep=large,cramped]
\mathcal{D}
\arrow[r,"\FmtVal{d_{10}}"]
\arrow[d,swap,"\Quo{\Delta}"]
&
\Pred{\tau}
\arrow[d,swap,"\Quo{\tau}"]
\\
\TM(\Delta)
\arrow[r,swap,"\FmtTm{d_{11}}"]
&
\TM(\tau)
\end{tikzcd}
\end{mathpar}
We define the mediating map $\bar{d}$ as the following square:
\[
\begin{tikzcd}[sep=large,cramped]
\mathcal{D}
\arrow[r,"\FmtVal{(d_{00},d_{10})}"]
\arrow[d,swap,"\Quo{\Delta}"]
&
\Pred{\sigma\times\tau}
\arrow[d,swap,"\Quo{\sigma\times\tau}"]
\\
\TM(\Delta)
\arrow[r,swap,"\FmtTm{(d_{01},d_{11})}"]
&
\TM(\sigma\times\tau)
\end{tikzcd}
\]
To see that the square commutes, fix $\FmtVal{p}:\mathcal{D}$ and calculate.
\begin{align*}
\RbNF{\sigma\times\tau}[
\Reify{\sigma\times\tau}[
d_{00}(p),d_{10}(p)
]
]
&=
\RbNF{\sigma\times\tau}[
\Reify{\sigma}[d_{00}(p)],
\Reify{\tau}[d_{10}(p)]
]
\\
&=
\FmtTm{
\Parens{
\RbNF{\sigma}[
\Reify{\sigma}[d_{00}(p)]
],
\RbNF{\sigma}[
\Reify{\sigma}[d_{00}(p)]
]
}
}
\\
&=
\FmtTm{
\Parens{
d_{01}(\Quo{\Delta}[p]),
d_{11}(\Quo{\Delta}[p])
}
}
\end{align*}
It is easy to see that $\pi_1\circ\bar{d}=d_1$ and $\pi_2\circ\bar{d}=d_2$.
The uniqueness of $\bar{d}$ with this property follows from the fact that its
components are unique: $\FmtVal{(d_{00},d_{10})}$ is the unique mediating arrow given
by the universal property of the product
$\Pred{\sigma\times\tau}\equiv\Pred{\sigma}\times\Pred{\tau}$; moreover,
because $\TM$ preserves finite products, we can say the same of
$\FmtTm{(d_{01},d_{11})}$.\footnote{Recall that $\TM$ is defined as $i^*\circ\Yo$
with $i^*$ the reindexing functor induced by
$i:\Ren{\Sigma}\to\ClCat{\Sigma}$. Because $i^*$ has a left adjoint, given by
Kan extension, it preserves limits; because the Yoneda embedding also
preserves limits, $\TM$ preserves limits too.}
\end{proof}
\begin{exercise}
For $\sigma,\tau\in\Ty{\Sorts[\Sigma]}$, show that $\Sem{\sigma\to\tau}$ is the
exponential in $\Sem{\tau}^{\Sem{\sigma}}$ in
$\GlCat{\Sigma}$~\citep[Example~2.1.12]{johnstone:2002}.
\end{exercise}
\begin{corollary}\label{cor:gluing-model}
$\GlCat{\Sigma}$ is a model of the free $\lambda$-theory generated by
$\Sigma$, with interpretation functor $\Sem{-}:\ClCat{\Sigma}\to\GlCat{\Sigma}$.
\end{corollary}
\begin{theorem}\label{thm:eval-cod}
The composite functor $\GlFib{\Sigma}\circ\Sem{-}$ is the identity endofunctor on $\ClCat{\Sigma}$:
\[
\begin{tikzcd}[sep=huge]
\ClCat{\Sigma}
\arrow[r,"\Sem{-}"]
\arrow[swap,dr,"\mathbf{id}_{\ClCat{\Sigma}}"]
&
\GlCat{\Sigma}
\arrow[d,"\GlFib{\Sigma}"]
\\
&
\ClCat{\Sigma}
\end{tikzcd}
\]
\end{theorem}
\begin{proof}
This follows immediately from the fact that $\ClCat{\Sigma}$ is the
classifying category of the theory $\Sigma$, so it is the initial category
with the structure of $\Sigma$. Therefore, any $\Sigma$-homomorphism
$\ClCat{\Sigma}\to\ClCat{\Sigma}$ must be the identity, including the composite above.
\end{proof}
Now, working externally in the category $\SET$, we can explicitly construct the
normalization function,
$\NfFun{\Gamma}{\Delta}:\Hom[\ClCat{\Sigma}]{\Gamma}{\Delta}\to\NF{\Delta}(\Gamma)$
as the following composite:
\[
\begin{tikzcd}
\Hom[\ClCat{\Sigma}]{\Gamma}{\Delta}
\arrow[swap,r,"\Sem{-}"]
\arrow[rrrr,bend left=20,"\NfFun{\Gamma}{\Delta}"]
&
\Hom[\GlCat{\Sigma}]{\Sem{\Gamma}}{\Sem{\Delta}}
\arrow[swap,r, "\pi"]
&
\Hom[\Psh{\Ren{\Sigma}}]{\Pred{\Gamma}}{\Pred{\Delta}}
\arrow[swap,r, "\phi"]
&
\Pred{\Delta}(\Gamma)
\arrow[swap,r, "{\Reify[\Gamma]{\Delta}}"]
&
\NF{\Delta}(\Gamma)
\end{tikzcd}
\]
where
\begin{align*}
\phi(\FmtVal{v}) &= \FmtVal{v}_{\Mute\Gamma}(\Reflect[\Gamma]{\Gamma}[\SbIdn[\Gamma]{\Gamma}])
\end{align*}
\begin{theorem}[Completeness]
If $\SeqEq{\Sigma}{\Gamma}{t_0}{t_1}{\tau}$, then
$\SeqEq{\Sigma}{\Gamma}{\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t_0]]}{\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t_1]]}{\tau}$.
\end{theorem}
\begin{proof}
This is immediate from the fact that we have defined a function out of the
morphisms of $\ClCat{\Sigma}$, which are already quotiented under
definitional equivalence.
\end{proof}
\begin{theorem}[Normalization]\label{thm:normalization}
If $\Seq{\Sigma}{\Gamma}{t}{\tau}$, then $\SeqEq{\Sigma}{\Gamma}{\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t]]}{t}{\tau}$.
\end{theorem}
\begin{proof}
Suppose $\FmtVal{\Sem{t}} = \FmtVal{\Parens{\Realize{t_0}{v}}}$. Now calculate.
\begin{align*}
\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t]]
&=
\RbNF[\Gamma]{\tau}[
\Reify[\Gamma]{\tau}[
v_{\Mute\Gamma}\Parens{
\Reflect[\Gamma]{\Gamma}[\SbIdn[\Gamma]{\Gamma}]
}
]
]
\\
&=
\Parens{
\RbNF{\tau}\circ\Reify{\tau}\circ{}\FmtVal{v}
}_{\Mute\Gamma}(\Reflect[\Gamma]{\Gamma}[\SbIdn[\Gamma]{\Gamma}])
\\
&=
\Parens{
\FmtTm{t_0}\circ\RbNF{\Gamma}\circ\Reify{\Gamma}
}_{\Mute\Gamma}(\Reflect[\Gamma]{\Gamma}[\SbIdn[\Gamma]{\Gamma}])
\tag{comma condition}
\\
&=
\FmtTm{
t_0^{\Mute\Gamma}\Parens{
\RbNF[\Gamma]{\Gamma}[
\Reify[\Gamma]{\Gamma}(\Reflect[\Gamma]{\Gamma}(\SbIdn[\Gamma]{\Gamma}))
]
}
}
\\
&=
\FmtTm{
t_0^{\Mute\Gamma}\Parens{
\RbNE[\Gamma]{\Gamma}[\SbIdn[\Gamma]{\Gamma}]
}
}
\tag{reify-reflect yoga}
\\
&=
\FmtTm{
t_0^{\Mute\Gamma}\Parens{
\SbIdn[\Gamma]{\Gamma}
}
}
\end{align*}
$\SeqEq{\Sigma}{\Gamma}{t_0^{\Mute\Gamma}(\SbIdn[\Gamma]{\Gamma})}{t}{\tau}$. Writing
$\FmtTm{\lfloor{}t\rfloor}$ for the induced natural transformation
$\TM(\Gamma)\to\TM(\tau)$ such that
$\SeqEq{\Sigma}{\Gamma}{\lfloor{}t\rfloor_{\Mute\Gamma}(\SbIdn[\Gamma]{\Gamma})}{t}{\tau}$, it
suffices to show that $\FmtTm{t} = \FmtTm{t_0}$. Because $\GlFib{\Sigma}\Parens{\Sem{\FmtTm{t}}}=\FmtTm{t_0}$, by
Theorem~\ref{thm:eval-cod} we have $\FmtTm{t}=\FmtTm{t_0}$.
\end{proof}
\begin{corollary}[Soundness]
If
$\SeqEq{\Sigma}{\Gamma}{\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t_0]]}{\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t_1]]}{\tau}$,
then $\SeqEq{\Sigma}{\Gamma}{t_0}{t_1}{\tau}$.
\end{corollary}
\begin{proof}
To see that $\FmtTm{t_0}=\FmtTm{t_1}$, observe that by Theorem~\ref{thm:normalization} we have $\RbNF[\Gamma]{\tau}[\NfFun{\Gamma}{\tau}[t_i]] = \FmtTm{t_i}$, so by transitivity and assumption we have $\FmtTm{t_0}=\FmtTm{t_1}$.
\end{proof}
\section{Perspective}
\subsection{Global sections and the Freyd cover}
\label{sec:freyd-cover}
A more common use of the gluing technique lies in the construction of the Freyd
cover (also called the ``scone'', which is short for ``Sierpinski cone'') of a
topos in order to prove properties of \emph{closed proofs} in intuitionistic
higher-order logic, such as the disjunction and existence properties, which
correspond in $\lambda$-calculus to instances of the closed canonicity
result~\citep[p.\ 228]{lambek-scott:1986}.
As an example, we will prove both these properties for intuitionistic
higher-order logic over simple types and the natural numbers.\footnote{This
section is an expanded version of material which appears
in~\citet{shulman:blog:scones-logical-relations}, with some more details filled in.}
Writing $\mathcal{F}$ for the free topos generated by a natural numbers object
$\mathsf{N}$, observe that we have the global sections functor
$\Gamma\equiv\mathcal{F}(\mathbf{1},-)$ which takes every object to its global
elements.
We define the Freyd cover $\breve{\mathcal{F}}$ over $\mathcal{F}$ as the gluing
category obtained by pulling back the fundamental fibration along the global
sections functor:
\[
\begin{tikzcd}[sep=huge,cramped]
{\breve{\mathcal{F}}}
\arrow[r,dashed]
\arrow[d,-{Triangle[open]}, dashed,swap,"\pi_1"]
\arrow[dr, phantom, pos = 0, "\lrcorner"]
&
{\SET}^\to
\arrow[d,-{Triangle[open]}, "\Cod"]
\\
\mathcal{F}
\arrow[r,swap,"\Gamma"]
&
\SET
\end{tikzcd}
\]
Because the global sections functor preserves finite limits, the Freyd cover
$\breve{\mathcal{F}}$ is again a topos with $\pi_1$ a \emph{logical
functor}~\citep[Example~2.1.12]{johnstone:2002}; moreover, $\pi_1$ preserves
the natural numbers object~\citep[Corollary~7.7.2]{taylor:1999}. We also have a
functor $\pi_0:\breve{\mathcal{F}}\to\SET$, which merely preserves finite
limits.
The Freyd cover $\breve{\mathcal{F}}$ has a natural numbers object
$\breve{\mathsf{N}}$ given by $(\mathsf{N},\mathbb{N},n\mapsto\bar{n})$, where
$\bar{n}$ takes a set-theoretic natural number to the corresponding global section in
$\mathcal{F}$, and $\pi_1$ preserves the natural numbers object.
Because $\mathcal{F}$ is the initial topos with a natural numbers object, for
any other such topos $\mathcal{E}$ we have a \emph{unique} map $I_{\mathcal{E}}
: \mathcal{F}\to\mathcal{E}$.
\begin{lemma}\label{lem:gluing-retract}
The logical functor $\pi_1:\breve{\mathcal{F}}\to\mathcal{F}$ is a retract of $I_{\breve{\mathcal{F}}}$:
\[
\begin{tikzcd}[sep=large,cramped]
\mathcal{F}
\arrow[r,"I_{\breve{\mathcal{F}}}"]
\arrow[dr, swap,"1"]
&
\breve{\mathcal{F}}
\arrow[d,"\pi_1"]
\\
&
\mathcal{F}
\end{tikzcd}
\]
\end{lemma}
\begin{proof}
We have the ``additional'' identity morphism $1:\mathcal{F}\to\mathcal{F}$, so by initiality of $\mathcal{F}$, we must have $1=\pi_1\circ{}I_{\breve{\mathcal{F}}}$.
\end{proof}
\begin{theorem}[Natural number canonicity]
Any global section $n:\Gamma(\mathsf{N})$ in $\mathcal{F}$ is equal to some
numeral $\bar{k}$.
\end{theorem}
\begin{proof}
The functor $I_{\breve{\mathcal{F}}}$ necessarily preserves $\mathsf{N}$.
Therefore, the global section $n$ lifts in $\breve{\mathcal{F}}$ to a square
in $\SET$ as follows:
\[
\begin{tikzcd}[sep=large,cramped]
\mathbf{1}
\arrow[r]
\arrow[d]
&
\mathbb{N}
\arrow[d]
\\
\Gamma(\mathbf{1})\cong\mathbf{1}
\arrow[r]
&
\Gamma(\mathsf{N})
\end{tikzcd}
\]
The upstairs morphism gives us a numeral $k$; because the diagram commutes
and using Lemma~\ref{lem:gluing-retract}, we have $n=\bar{k}$.
\end{proof}
\begin{theorem}[Existence property]\label{thm:existence-property}
Suppose that $\mathcal{F}\models\exists x:X.\phi(x)$; then there is a global
element $\alpha:\mathbf{1}\to{}X$ in $\mathcal{F}$ such that
$\mathcal{F}\models\phi(\alpha)$.
\end{theorem}
\begin{proof}
We will use the Kripke-Joyal semantics of the
topos~\citep{maclane-moerdijk:1992}; unwinding our assumption
$\mathbf{1}\Vdash\exists x:X.\phi(x)$, we have that there exists an
epimorphism $p:V\twoheadrightarrow\mathbf{1}$ and a morphism $\beta:V\to{}X$
such that $V\Vdash\phi(\beta)$.\footnote{Please note that the symbol $\Vdash$
here denotes the forcing relation, rather than the gluing relation.}
The logical functor
$I_{\breve{\mathcal{F}}}:\mathcal{F}\to\breve{\mathcal{F}}$ lifts $p$ to an epimorphism $I_{\breve{\mathcal{F}}}(p) :
I_{\breve{\mathcal{F}}}(V)\twoheadrightarrow{}I_{\breve{\mathcal{F}}}(\mathbf{1})$ in $\mathcal{F}$. Since $I_{\breve{\mathcal{F}}}$ preserves
the terminal object, this is actually to say $I_{\breve{\mathcal{F}}}(p) :
I_{\breve{\mathcal{F}}}(V)\twoheadrightarrow{}\mathbf{1}$ in $\breve{\mathcal{F}}$. $I_{\breve{\mathcal{F}}}(p)$ must be
a square in $\SET$ of the following kind:
\[
\begin{tikzcd}[sep=large,cramped]
\pi_0(I_{\breve{\mathcal{F}}}(V))
\arrow[r, two heads]
\arrow[d]
&
\mathbf{1}
\arrow[d]
\\
\Gamma(V)
\arrow[r]
&
\Gamma(\mathbf{1})
\end{tikzcd}
\]
Because the upstairs morphism is a surjection, we know that $\pi_0(I_{\breve{\mathcal{F}}}(V))$ is non-empty;
therefore, because we have a map $\pi_0(I_{\breve{\mathcal{F}}}(V))\to\Gamma(V)$, we can see that
$\Gamma(V)$ is non-empty, i.e.\ we have a global section of $\rho :
\mathbf{1}\to{}V$ in $\mathcal{F}$.
By precomposition and Kripke-Joyal monotonicity, then, we have a global
section $\beta\circ\rho:\mathbf{1}\to{}X$ such that
$\mathbf{1}\Vdash\phi(\beta\circ\rho)$.
\end{proof}
\begin{theorem}[Disjunction property]
Suppose that $\mathcal{F}\models\phi(\alpha)\lor\psi(\alpha)$ for some
$\alpha:\mathbf{1}\to{}X$; then either $\mathcal{F}\models\phi(\alpha)$ or
$\mathcal{F}\models\psi(\alpha)$.
\end{theorem}
\begin{proof}
We will use essentially the same technique as in our proof of
Theorem~\ref{thm:existence-property}. Unwinding the Kripke-Joyal semantics of
the topos, we have morphisms $p:V\to\mathbf{1}$ and $q:W\to\mathbf{1}$ such
that $[p,q]:V+W\to\mathbf{1}$ is an epimorphism and moreover
$V\Vdash\phi(\alpha\circ{}p)$ and $W\Vdash\psi(\alpha\circ{}q)$. As above,
$[p,q]$ lifts to an epimorphism in $\breve{\mathcal{F}}$ as follows:
\[
\begin{tikzcd}[sep=large,cramped]
\pi_0(I_{\breve{\mathcal{F}}}(V+W))=\pi_0(I_{\breve{\mathcal{F}}}(V))+\pi_0(I_{\breve{\mathcal{F}}}(W))
\arrow[r, two heads]
\arrow[d]
&
\mathbf{1}
\arrow[d]
\\
\Gamma(V+W)=\Gamma(V)+\Gamma(W)
\arrow[r]
&
\mathbf{1}
\end{tikzcd}
\]
Note that the global sections functor for $\mathcal{F}$ preserves finite
colimits, and $\pi_0$ preserves all
colimits~\citep[Proposition~7.7.1(l)]{taylor:1999}. Because the upstairs
morphism is a surjection, we know that $\pi_0(I_{\breve{\mathcal{F}}}(V))+\pi_0(I_{\breve{\mathcal{F}}}(W))$ is
non-empty, whence we must have either a global section $r\in\Gamma(V)$ or a
global section $s\in\Gamma(W)$.
Supposing we have a global section $r:\mathbf{1}\to V$ in $\mathcal{F}$, then
by Kripke monotonicity, we have
$\mathbf{1}\Vdash\phi(\alpha\circ{}p\circ{}r)$. On the other hand, if we have
a global section $s:\mathbf{1}\to{}W$, then we would have
$\mathbf{1}\Vdash\psi(\alpha\circ{}q\circ{}s)$.
\end{proof}
\subsection{Connection with the method of computability}
As we have alluded to in the previous section, the gluing category always
functions as the ``category of suitable logical predicates'', with the meaning
of ``suitable'' negotiated by choice of gluing functor. Most instances of the
logical relations/predicates technique can be phrased as an instance of the
more general gluing construction.
\subsubsection{Proof (ir)relevance}
The native notion of logical ``predicate'' which is induced by the gluing
construction is a proof-relevant one, whereas in the method of computability,
one generally studies predicates in the classical, proof-irrelevant sense. This
restriction is easily accounted for by making a slight adjustment to the
categories involved.
Writing $\mathbb{C}$ for the classifying category of our theory, if we take a
category $\mathcal{E}$ to be our semantic domain, we can form categories of
proof-relevant logical predicates and proof irrelevant logical predicates
respectively along a functor $F:\mathbb{C}\to\mathcal{E}$ as
follows:\footnote{For intuition, consider the specific example where
$\mathcal{E}$ is $\SET$ and $\mathcal{F}$ is the global sections functor, as in
Section~\ref{sec:freyd-cover}.}
\begin{mathpar}
\begin{tikzcd}[sep=large,cramped]
{\mathsf{Gl}}
\arrow[r]
\arrow[d, -{Triangle[open]}]
\arrow[dr, phantom, pos = 0, "\lrcorner"]
&
\mathcal{E}^{\to}
\arrow[d,-{Triangle[open]}, "\Cod"]
\\
\mathbb{C}
\arrow[r, swap, "F"]
&
\mathcal{E}
\end{tikzcd}
\and
\begin{tikzcd}[sep=small,cramped]
{\mathsf{Gl}_{\mathit{irr}}}
\arrow[rr]
\arrow[dd]
\arrow[drr, phantom, pos = 0, "\lrcorner"]
&&
\mathcal{E}^{\to}_{\mathit{mono}}
\arrow[d]
\\
&&
\mathcal{E}^{\to}
\arrow[d,-{Triangle[open]},"\Cod"]
\\
\mathbb{C}
\arrow[rr, swap, "F"]
&&
\mathcal{E}
\end{tikzcd}
\end{mathpar}
When $F$ is the global sections functor (and thence $\mathsf{Gl}$ is the Freyd
cover or the scone of $\mathbb{C}$), $\mathsf{Gl}_{\mathit{irr}}$ is often
called the ``subscone'' of $\mathbb{C}$.
\subsubsection{Relations vs predicates}
What we have seen so far corresponds to the technique of unary logical
relations, but the abstraction scales easily to the case of binary (and
$n$-ary) logical relations by replacing $\mathbb{C}$ with
$\mathbb{C}\times\mathbb{D}$, as described in~\citet{mitchell-scedrov:1993}. To
see the connection with binary logical relations, it will be instructive to
work out explicitly the case for exponentials in the subscone of
$\mathbb{C}\times\mathbb{C}$, which we will write
$\widetilde{\mathbb{C}\times\mathbb{C}}$.
First, observe that the exponential in the product of two cartesian closed
categories is calculated pointwise; so for $(A_0,A_1), (B_0,B_1) :
\mathbb{C}\times\mathbb{C}$, we have ${(B_0,B_1)}^{(A_0,A_1)} = ({B_0}^{A_0},
{B_1}^{A_1})$.
An object in $\widetilde{\mathbb{C}\times\mathbb{C}}$ is a monomorphism
$R\rightarrowtail\Gamma(A,B)$ where $\Gamma$ is the global sections functor for
$\mathbb{C}\times\mathbb{C}$. Because the global sections functor preserves
finite limits, this is to say that we have a monomorphism
$R\rightarrowtail\Gamma(A)\times\Gamma(B)$, in other words \emph{a relation on
the closed terms of type $A$ and $B$ in the language $\mathbb{C}$}.
We wish to inspect for ourselves the exponential object in
$\widetilde{\mathbb{C}\times\mathbb{C}}$. As we saw earlier on, to form the
exponential in the gluing category we first take the following pullback:
\[
\begin{tikzcd}[sep=large]
{E}
\arrow[dr, phantom, pos = 0, "\lrcorner"]
\arrow[r]
\arrow[d, tail]
&
{S^R}
\arrow[d, tail]
\\
{\Gamma\Parens{{B_0}^{A_0}}\times\Gamma\Parens{{B_1}^{A_1}}}
\arrow[r]
&
\Parens{\Gamma(B_0,B_1)}^R
\end{tikzcd}
\]
Then, we define the exponential
${\Parens{S\rightarrowtail\Gamma(B_0)\times\Gamma(B_1)}}^{\Parens{R\rightarrowtail\Gamma(A_0)\times\Gamma(A_1)}}$
to be the monomorphism on the left. Now, unfolding definitions, a global element of this
exponential is simply a pair of closed terms $\cdot\vdash{}F_0:A_0\to{}B_0$ and
$\cdot\vdash{}F_1:A_1\to{}B_1$ together with a function $H : S^R$ which is
tracked by $(F_0,F_1)$; unwinding further, this means only that for all
$\cdot\vdash{}a_0:A_0$ and $\cdot\vdash{}a_1:A_1$, if $(a_0,a_1)\in{}R$, then
$(F_0(a_0), F_1(a_1))\in{}S$.
\subsubsection{Kripke/Beth/Grothendieck logical relations}
A common generalization of the method of computability is to use a logical
relation which is indexed in some partial order (or even a category), subject
to a functoriality condition. In the literature, these are called \emph{Kripke
logical relations}, and indeed, the construction that we used to prove
normalization of free $\lambda$-theories in Section~\ref{sec:nbg} is the
proof-relevant unary Kripke instance of the gluing abstraction, where the
worlds are contexts of variables linked by renamings.
Many other variations of indexed logical relations appear in the wild, and
nearly all of these are already accounted for within the abstraction. For
instance, by imposing Grothendieck topology on the base poset or category and
requiring a \emph{local character} condition in addition to monotonicity, one
can develop something which might be called \emph{Beth/Grothendieck logical
relations} (see~\citet{coquand-mannaa:2016},
\citet{altenkirch-dybjer-hofmann-scott:2001} and \citet{fiore-simpson:1999} for
examples).
\begin{remark}[Terminology]
In the literature~\citep{jung-tiuryn:1993,fiore-simpson:1999,fiore:2002}, the
proof irrelevant version of this construction appears under the somewhat
confusing name ``Kripke Relations of Varying Arity''---confusing because it is
not immediately clear what it has to do with the arity of a relation.
In the early literature (such as~\citet{jung-tiuryn:1993}), there was
some resistance to explaining what these were in a more conceptual way, but
as described in~\citet{fiore-simpson:1999}, these have a simple
characterization as \emph{internal relations} of a certain kind within a
presheaf topos which corresponds exactly to a proof irrelevant version of the
construction we describe in these notes.
\end{remark}
\begin{example}[Independence of Markov's Principle]
In~\citet{coquand-mannaa:2016}, the method of computability was used to
establish the independence of Markov's Principle from Martin-L\"of Type
Theory using a forcing extension over Cantor space $\mathcal{C}$. We will
briefly describe how the construction in that paper fits into the framework
of gluing.
Letting $\mathbb{C}$ be the classifying category of the forcing extension of
type theory, we have a fibration $\pi_{\mathcal{C}}:\mathbb{C}\to\mathcal{C}$
which projects the forcing condition (a representation of compact open in
Cantor space). Writing $\Sh{\mathcal{C}}$ for the topos of sheaves over
Cantor space, we evidently have a functor $\TM:\mathbb{C}\times\mathbb{C}\to\Sh{\mathcal{C}}$
which takes a pair of contexts $(\Delta_0,\Delta_1)$ to the following sheaf:
\[
(p : \mathcal{C}) \mapsto
\SetCompr{
(\delta_0,\delta_1)
}{
p \leq \pi_{\mathcal{C}}(\Delta_0)
\land
p \leq \pi_{\mathcal{C}}(\Delta_1)
\land
{\cdot\mathrel{\vdash_p}\delta_0:\Parens{\Delta_0}_{\vert{}p}}
\land
{\cdot\mathrel{\vdash_p}\delta_1:\Parens{\Delta_1}_{\vert{}p}}
}
\]
(The above is a sheaf, because the topology on $\mathcal{C}$ is subcanonical,
and because the calculus contains a rule for local character.)
Now, consider the gluing category obtained from the following pullback:
\[
\begin{tikzcd}[sep=large,cramped]
{\mathsf{Gl}}
\arrow[r]
\arrow[dd]
\arrow[ddr, phantom, pos = 0, "\lrcorner"]
&
\Sh{\mathcal{C}}^{\to}_{\mathit{mono}}
\arrow[d]
\\
&
\Sh{\mathcal{C}}^{\to}
\arrow[d,-{Triangle[open]}, "\Cod"]
\\
\mathbb{C}\times\mathbb{C}
\arrow[r, swap, "F"]
&
\Sh{\mathcal{C}}
\end{tikzcd}
\]
Viewed externally, the objects of $\mathsf{Gl}$ are $\mathcal{C}$-indexed
binary relations on closed terms in $\mathbb{C}$ which enjoy both
monotonicity and local character. By examining the cartesian closed structure
of $\mathsf{Gl}$, it can be seen (as above) that the logical relations for
each connective match the na\"ive ones.
\end{example}
\section*{Acknowledgments}
We are thankful to Jonas Frey for explaining the Freyd cover construction for
the free topos; and to Marcelo Fiore, Michael Shulman and Darin Morrison for
helpful conversations about conceptual proofs of normalization. Thanks also to
Daniel Gratzer and Lars Birkedal for their comments on a draft of this note.
The authors gratefully acknowledge the support of the Air Force Office of
Scientific Research through MURI grant FA9550-15-1-0053 and the AFOSR project
`Homotopy Type Theory and Probabilistic Computation', 12595060. Any opinions,
findings and conclusions. Any opinions, findings and conclusions or
recommendations expressed in this material are those of the authors and do not
necessarily reflect the views of the AFOSR.%
\nocite{coquand-dybjer:1997, shulman:2015, crole:1993, dagand-scherer:2015, nlab:free-topos, nlab:freyd-cover}
\nocite{clairambault-dybjer:2015}
\bibliographystyle{plainnat}
|
\section{Introduction}
In the last decade, the rapid growth and adoption of online social networks, such as Facebook, Twitter,
Linkedin, has fundamentally changed the way people interact with each other.
There are many people who would rather spend more time on these social networking sites than
traditional media.
With this trend, a great deal of data has been generated from the increasing
number of online social networking users.
Therefore, it is important to study the structure of social networks, which can
provide meaningful insight to Sociology, Communications, Economics, Marketing or even Epidemiology.
One important type of structure of social networks is how the entities are divided into
different groups.
Basically, there is no formal definition of community, but it is believed
that entities are densely connected inside each community with less links
between different communities \cite{wasserman1994social}.
This community structure plays a significant role in
visualization \cite{kang2007visual}, dynamic community detection
\cite{tang2008community}, opinion mining \cite{wang2013influence}, and behavior
prediction \cite{tang2009scalable}.
Previous research work on community detection generally dealt with the single-view setting.
Views are independent data sources or datasets.
One classic example is the web-page classification \cite{blum1998combining}, in
which one view is the content of web-page, and the other is comprised of the hyperlinks
pointing to it.
In social networks, such as YouTube and Flickr, the interactions between users are
complex \cite{tang2009uncovering}.
Similarly, on Facebook, users like, comment and share content, and
interact with each other through these activities.
Specifically, in the same page, the activities on posts and those on comments can
form two views. From each view, we can generate features and construct a graph to
find community structures within a given page.
Apart from the complexity of interactions of users, the volume of data
derived from social media has increased exponentially.
At the end of 2017, Facebook had more than $2$ billion users
\cite{facebook2018} and up to $40$ million small businesses had public
pages \cite{page2018}.
We find that with an increase in the number of users, and corresponding
interactions, there is naturally a significant correlative increase in difficulty
of discovering community structures within Facebook pages.
In this paper, we propose to model Facebook page as a weighted graph that is
generated by two views (posts and comments).
Then we examine the community structure of CBS News and The New Times
Facebook pages in last week of 2012.
In addition, the community structure for common users across multiple pages is
studied.
Our findings show that combining different views can remarkably reduce the
number of isolates in a single-view and make the community structure more
cohesive in networks because both views can mutually benefit from each other.
The rest of the paper is organized as follows.
The next section introduces the related work.
Section $3$ describes the issue of single-view methods.
The method is presented in Section $4$.
Section $5$ gives a detailed empirical study and results.
Finally, discussion is made in Section $6$.
\section{Related Work}
Before community detection became a trending research topic, data clustering has been always
a basic problem in machine learning research.
Although clustering is more general in terms of the non-overlapping
characteristic of data points, it can be applied to the community detection problem.
There are basically two types of clustering algorithms \cite{berkhin2006survey}.
One is model-based or so-called generative approach and the other is
similarity-based or discriminative approach. Furthermore, in similarity-model
methods, spectral clustering \cite{shi2000normalized,ng2002spectral} has prevailed in
the last several years for its performance, efficiency and robustness.
On the other hand, multi-view clustering is an effective tool for complex social networks.
The earliest work of multi-view clustering was proposed by Blum and Mitchell in
their co-training algorithm \cite{blum1998combining}.
Their idea was based on the assumption that learning can progress with enough
labeled data in each view otherwise each view mutually benefits from the labeled
data to each other.
Steffen Bickel and Tobias Scheffer \cite{bickel2004multi} proposed to alternately cluster
each view and exchange information during the learning process.
Multi-view idea met spectral clustering in Virginia's work
\cite{de2005spectral}, in which the disagreement between two views was minimized.
It outperformed the result from each view in a sense that each view can leverage
the information from each other. Later, spectral clustering was extended to
clustering multiple graphs by Zhou et al \cite{zhou2007spectral}, where a
generalized Laplacian was built based on a random walk and its eigenvectors were computed.
Kumar et al. added co-training flavor to spectral clustering
\cite{kumar2011cotraining}.
First, original two sets of eigenvectors of graphs were computed.
Second, each eigenvector was used to modify the structure of the other graph.
Third, the clustering result was generated from the column-wise concatenation of the
two eigenvector matrices.
Also, Kumar et al. \cite{kumar2011coregularized} allowed the pair-wise co-regularizers
to be included in the objective function to decrease the disagreement between any pair of views.
Xiang et al. \cite{wang2013multi} were the first to bring the Pareto optimization into spectral
clustering to find optimal cuts for multiple graphs via multi-objective
functions.
No parameter was needed to set for Pareto frontier and it explored all possible good cuts.
Xia et al. proposed a novel Markov chain method for Robust Multi-view Spectral
Clustering (RMSC) to deal with the noise problem \cite{xia2014robust}.
They used the Lagrangian Multiplier scheme to optimize RMSC.
In addition to spectral clustering, many other clustering algorithms were applied to multi-view setting. Cai et al. extended K-means to multi-view clustering on a large scale amount of data \cite{cai2013multi}.
Linked Matrix Factorization \cite{tang2009clustering} was proposed for clustering multiple graphs.
And Liu et al. \cite{liu2013multi} formulated the problem as a joint non-negative matrix factorization process that forces each result of view to hold a consensus.
Zhang et al. \cite{zhang2015constrained} also used similar NMF method but with constrains to handle unmapped data.
Other related work includes Linked Matrix Factorization \cite{tang2009clustering}, tensor methods \cite{liu2013multiview}, fuzzy clustering \cite{jiang2015collaborative} and belief propagation \cite{wang2016multi}.
Moreover, heterogeneous networks represent multi-relational in social networks
and some interesting work \cite{greene2009multi, tang2009uncovering,
tang2012community} should be noticed recently.
However, little work has been carried out on Facebook public pages.
This paper focuses on modeling, testing and discovering the community
structure of Facebook newsgroup pages. To represent the users interactions
in public pages precisely, two graphs are built and merged in a way that
modularity is optimized.
Analysis is conducted on not only a single page but also against multi-page clusters.
\section{Issues of Single View Community Detection}
In the traditional research of community detection in the single-view setting, the
procedure is the following: first construct a graph based on the connection between
users and then apply some algorithm to partition the graph.
However, in real social networks, the interaction between users are complex.
For example, users can like or comment a post so it is hard to use only one
graph to represent different interactions between users.
For CBS News page, we extract $205$ sample users from our database and
construct two graphs based on the interactions of different content: posts and comments.
The definition is shown in the next section.
Then, we apply the multi-level algorithm \cite{blondel2008fast,csardi2006igraph,hagberg2008exploring} to find
the community structure in the two graphs.
The connected nodes with the same color belong to the same community.
It is clear that there are more isolates in the post graph in Figure \ref{co_like_communities_cbs} but the
comment graph in Figure \ref{like_comment_communities_cbs} shows a regular and balanced community structure.
Therefore, the issue is that single-view or one graph, can't best
represent the interactions between users.
In particular, in this example, the users that are considered as isolates in the
post graph actually have connections with others in the comment graph.
\begin{figure}[tbp]
\centering
\includegraphics[clip,width=0.9\linewidth]{cbs_co_like_communities.png}
\caption{The Community Structure of Post Graph for CBS}
\label{co_like_communities_cbs}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[clip,width=0.9\linewidth]{cbs_like_comment_communities.png}
\caption{The Community Structure of Comment Graph for CBS}
\label{like_comment_communities_cbs}
\end{figure}
\section{Method}
\subsection{Problem Formulation}
Before we formulate the problem, it is necessary to mention the terminology in
this paper.
We consider networks as graphs, where a node represents a Facebook user and an edge
represents interaction between a pair of nodes.
Community and cluster are interchangeable as well.
A multi-view dataset can be represented by \textbf{k} graphs that have the same
set of nodes but with a different set of edges.
Formally, given $m$ graphs $\mathcal{G}1=(\mathcal{V}1, \mathcal{E}1), \mathcal{G}2=(\mathcal{V}2,
\mathcal{E}2), ..., \mathcal{G}m=({\mathcal{V}m, \mathcal{E}m})$ and the number of communities
\textbf{k}, our goal is to find a vector \textbf{v}=($v_{1}$, $v_{2}$, ..., $v_{n}$) such that \textbf{v}
gives an optimal community structure for all graphs, where $v_{i}$ represents that node i belongs
to community $j$ and $1 \leq i \leq n$, $1 \leq j \leq k$.
In this paper, we focus on two views (activities on posts and comments) in
Facebook public pages.
Thus, $k = 2$.
\subsection{Graph Construction}
To some extent, community detection is a graph partitioning problem.
So, it is important to define the appropriate graph for our purpose.
In Facebook, users have three basic types of actions: comment, like and share.
Specifically, in newsgroup pages, post and comment are the basic blocks in which users interact with each other.
In data clustering, a matrix is used to represent and analyze a graph.
Here we use adjacency matrices to represent our social interaction graphs.
If a pair of users $i$ and $j$ concurrently like a post, then we put $1$ in the cell $(i,
j)$ of the matrix.
And we call this the post graph.
The adjacency matrix \ref{adjacency1} for post view/graph is defined as follows:
\begin{equation}\label{adjacency1}
A_{ij} = \left\{ \begin{array}{ll}
1 & \textrm{if \textit{i} and \textit{j} concurrently like the same post}\\
0 & \textrm{otherwise}\\
\end{array} \right.
\end{equation}
The other adjacency matrix \ref{adjacency2} for the comment view/graph is defined by likes on comment.
And we call this comment graph.
If user $i$ likes the comment of user $j$ or vice versa or they concurrently like a comment, we assign the
weight $B_{ij}$ to be 1, otherwise 0.
\begin{equation}\label{adjacency2}
B_{ij} = \left\{ \begin{array}{ll}
1 & \textrm{if \textit{i} likes \textit{j}'s comment or vice versa or}\\
& \textrm{\textit{i} and \textit{j} concurrently like a comment} \\
0 & \textrm{otherwise} \\
\end{array} \right.
\end{equation}
Then we define our weighted graph by combining the two graphs into one and
assigning each graph a weight based on importance factor.
Moreover, it can be easily extended to multiple views.
Its formal definition \ref{adjacency} is as follows:
\begin{equation}\label{adjacency}
W = \sum_{i=1}^{n}{\alpha_{i} X_{i}}
\end{equation}
where n is the number of views/graphs, $0 \leq \alpha \leq 1$ and $\sum_{i=1}^{n}{\alpha_{i}} = 1$.
And when $n=2$, it becomes the adjacency matrix for two views \ref{two_view}.
\begin{equation}\label{two_view}
W = \alpha X_{1} + (1 - \alpha) X_{2}
\end{equation}
It turns out when $\alpha = 0$ or $\alpha = 1$, it is reduced to single-view.
\subsection{Multi-View Community Detection via Weighted Graphs}
To learn the optimal parameter in equation \ref{two_view}, modularity
\cite{newman2006modularity} is introduced.
Modularity \ref{modularity} is a measurement that evaluates
how apposite community structure is for any given network.
It ranges from $-1$ to $1$ inclusively and the larger it is, the better the
community structure is.
From the definition, modularity essentially is the value that the real weight of
an edge minus the probability of generating it and sum them all.
\begin{equation}\label{modularity}
Q=\frac{1}{2m}\sum_{ij}{\biggl[W_{ij}-\frac{d_{i}d_{j}}{2m}\biggr]}\delta(c_{i},c_{j})
\end{equation}
\begin{equation}
\delta(c_{i},c_{j}) = \left\{ \begin{array}{ll}
1 & \textrm{if \textit{i} and \textit{j} are in the same community}\\
0 & \textrm{otherwise}\\
\end{array} \right.
\end{equation}
where m is the number of edges, W is the adjacency matrix, $d_{i}, d_{j}$ are
the degree of node i and j respectively.
Our algorithm borrows the idea of modularity maximization \cite{chen2014community}.
First, we generate a set of parameters and calculate the modularity for each
network's structure.
Then, we pick the largest modularity value and its corresponding community structure
as the result.
The details can be seen in Algorithm \ref{weighted_algorithm}.
\begin{algorithm}
\caption{Multi-View Community Detection via Weighted Graphs}
\label{weighted_algorithm}
\KwIn {Adjacency matrices $X_{i}$, and its parameter $\alpha_{i}$,
where $\sum_{i=1}^{n}{\alpha_{i}} = 1$, $0 < \alpha_{i} < 1$ and $1 \leq i \leq n$}
\KwOut {Indicator vector v}
Initialize an empty vector set $V$ \;
\ForEach {possible combination of $\alpha_{i}$}{
generate the unified similarity matrix $W_{j}$ by \ref{adjacency};
compute community structure $v_{j}$ using matrix $W_{j}$ \;
put $v_{j}$ into $V$ \;
}
pick the $v$ in $V$ with largest modularity value as the final indicator vector \;
\end{algorithm}
The time complexity of Algorithm \ref{weighted_algorithm} is based on two aspects.
The first is to construct the adjacency matrix, which takes $O(n^{2})$ time
theoretically, but this can be reduced to $O(m)$ because most social networks
are sparse $(m \ll n^{2})$, where $n$ and $m$ are the number of nodes and edges of the network.
In addition, the second is to run the core community detection algorithm
\cite{blondel2008fast} $k$ times that takes $O(km)$, where $k$ is the number of parameters.
In total, the time complexity of our algorithm is $O(m+k*m)$, which is $O(m)$ since $k \ll m$.
\section{Empirical Study}
\subsection{Data Collection}
Our Facebook dataset \cite{erlandsson2015crawling} is crawled from public pages by using Facebook Graph API.
In this paper, we only consider the public pages like CBS News, Fox News and New
York Times, etc.
One characteristic of these pages is that only the administrators can post
information.
Users can only comment, like and share posts or reply and like comments.
In our database, we have several tables such as comment, likedby and shares, etc.
Since comment and like are the most common activities in Facebook newsgroup
pages, we focus on the comment table and the likedby table.
Specifically, the schema of comment is composed of $id$, $post$ $id$, $page$
$id$, $fb$ $id$, $message$, $can$ $remove$ and $created$ $time$.
And the schema of likedby is composed of $page$ $id$, $post$ $id$, $common$
$id$ and $fb$ $id$.
Until June 2018, we have crawled millions of pages, hundreds of millions of posts,
billions of comments and likes by near $10$ billions application-scoped users.
Table \ref{crawled} shows the statistics in detail.
\begin{table}[thb]
\centering
\small
\caption{The Metadata of Facebook Pages Crawled}
\label{crawled}
\begin{tabular}{lll}
\hline
Table Name & Number of rows & Size in GB \\
\hline
page & 4,706,324 & 0.30 \\
shares & 160,098,978 & 8.54 \\
post & 343,000,084 & 246.66 \\
comment & 6,429,663,271 & 1,213.59 \\
likedby & 36,897,398,296 & 4,028.59 \\
reaction & 95,831,789,680 & 9,973.33 \\
\hline
\end{tabular}
\end{table}
We collected two datasets from CBS News and The New York Times in the last week of
2012.
The statistics of the two datasets are listed in Table \ref{statistics_datasets}.
\begin{table}[thb]
\centering
\small
\caption{The Statistics of Facebook Newsgroups}
\label{statistics_datasets}
\begin{tabular}{lll}
\hline
Category & CBS & NY Times \\
\hline
Users & 11,610 & 42,001 \\
Posts & 42 & 57 \\
Comments & 5,488 & 3,244 \\
Likes & 15,000 & 64,104 \\
\hline
\end{tabular}
\end{table}
According to previous work \cite{faloutsos1999power, clauset2009power}, most variables in internet or social networks display power-law/long tail distributions.
To verify this, we plot the users distribution with the number of activities in
Facebook public pages, where the x-axis is the number of likes of users on comments or posts and the
y-axis is the number of users.
Figure \ref{user_distribution}\subref{1a} and Figure \ref{user_distribution}\subref{2a} reveal the users distributions with likes on
posts and on comments respectively for CBS News.
And Figure \ref{user_distribution}\subref{3a} plots the distributions of users with comments.
In addition, Figure \ref{user_distribution}\subref{1b}, Figure \ref{user_distribution}\subref{2b} and Figure \ref{user_distribution}\subref{3b} are those for
The New York Times.
It can be seen that with the increase in number of likes, the
number of users drops significantly, approximately power distributions.
Most Facebook users in the two pages have less than $10$ likes and $10$ comments in a week.
\begin{figure}[t]
\centering
\subfloat[The User Distribution with Likes on Posts in CBS News Page]
{\includegraphics[width=0.45\linewidth]{cbs_user_post_distribution_line.png}\label{1a}}
\hfill
\subfloat[The User Distribution with Likes on Posts in The New York Times Page]
{\includegraphics[width=0.45\linewidth]{nyt_user_post_distribution_line.png}\label{1b}}
\hfill
\subfloat[The User Distribution with Likes on Comments in CBS News Page]
{\includegraphics[width=0.45\linewidth]{cbs_user_comment_distribution_line.png}\label{2a}}
\hfill
\subfloat[The User Distribution with Likes on Comments in The New York Times
Page]
{\includegraphics[width=0.45\linewidth]{nyt_user_comment_distribution_line.png}\label{2b}}
\hfill
\subfloat[The User Distribution with Comments in CBS News Page]
{\includegraphics[width=0.45\linewidth]{cbs_user_on_comment_distribution.png}\label{3a}}
\hfill
\subfloat[The User Distribution with Comments in The New York Times Page]
{\includegraphics[width=0.45\linewidth]{nyt_user_on_comment_distribution.png}\label{3b}}
\caption{The User Distribution in CBS and NY Times}
\label{user_distribution}
\end{figure}
\subsection{Parameter Tuning}
After defining the graphs, we study how the parameter affects the community
structure of pages.
Comment graphs show less density than post graphs, which implies that users are inclined to like posts rather than like comments.
Moreover, the number of clusters decrease tens of times in Table \ref{all_cbs}
and hundreds of times in Table \ref{all_nyt} for each page, which shows
our algorithm can effectively reduce isolates and uncover a more cohesive structure of networks.
View $1$ and View $2$ represent post graph and comment graph respectively.
Merged represents the weighted graph from the two views.
On the other hand, we observe that modularity is not a perfect measurement especially for networks
with a number of isolates because the more isolates are, the larger modularity is.
\begin{table}[thb]
\begin{center}
\small
\caption{The Statistics of CBS News Graph} \label{all_cbs}
\begin{tabular}{cccc}
\hline
Category & View 1 & View 2 & Merged \\
\hline
Users & 10,535 & 10,535 & 10,535 \\
Edges & 2,448,338 & 28,208 & 2,475,896 \\
Clusters & 3,338 & 6,821 & \bf{120} \\
Isolates & 3,321 & 6,647 & \bf{4} \\
Modularity & 0.8334 & \bf{0.9135} & 0.8350 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[hb]
\begin{center}
\small
\caption{The Statistics of The New York Times Graph} \label{all_nyt}
\begin{tabular}{cccc}
\hline
Category & View 1 & View 2 & Merged \\
\hline
Users & 41,252 & 41,252 & 41,252 \\
Edges & 106,115,374 & 213,746 & 106,296,974 \\
Clusters & 3,395 & 36,340 & \bf{31} \\
Isolates & 3,383 & 36,279 & \bf{0} \\
Modularity & 0.6054 & \bf{0.8057} & 0.6050 \\
\hline
\end{tabular}
\end{center}
\end{table}
In order to find the optimal parameter, we generate a set of parameters as the candidates, \{0.0, 0.2, 0.4, 0.6, 0.8, 1.0\}.
Then we run Algorithm \ref{weighted_algorithm} repetitively for each parameter and calculate the corresponding modularity.
We plot the relation between the parameters and their modularity in Figure
\ref{parameter_q} for both CBS News and The New York Times.
\begin{figure}[t]
\centering
\includegraphics[clip,width=0.9\linewidth]{parameter_q.png}
\caption{The Relation between Weight and Modularity for CBS News and The New York Times}
\label{parameter_q}
\end{figure}
Also we examine how the parameter affects the number of communities.
Table \ref{number_communities} shows the relation between the parameters and
their number of communities for the two pages.
The results demonstrate the parameter does not change the modularity value and
the number of communities.
In this study, the parameter value is to set to be $0.5$.
It is worth noting that the parameter improves the modularity of the post graph and it makes the both graphs denser.
Modularity is inclined to increase with more isolates for the same set of
users, which can explain why it reaches the maximum value when the parameter is $0$.
\begin{table}[thb]
\begin{center}
\small
\caption{The Relation between Parameter and Number of Communities in CBS
News and The New York Times} \label{number_communities}
\begin{tabular}{lll}
\hline
Parameter & CBS & NY Times \\
\hline
0 & 6,821 & 36,340\\
0.2 & 120 & 31 \\
0.4 & 120 & 31 \\
0.6 & 120 & 31 \\
0.8 & 120 & 31 \\
1 & 3,338 & 3,395 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Selection of Active Users}
Facebook itself is a huge dataset with billions of users so that it is
not easy to analyze the activities of all users.
Moreover, there are a number of users who have limited activities and don't
contribute to providing any structural information.
Therefore, for our good, we define the following popular content and active users.
For popular content (post and comment) \ref{popular_content} in public pages is
defined by the following:
\begin{Definition}\label{popular_content}
Popular Facebook Content is a post liked by at least 2 users or a comment liked by at least 1 user.
\end{Definition}
For active Facebook users, the definition \ref{active_user} is as follows:
\begin{Definition}\label{active_user}
An Active Facebook User must have the following conditions: at least likes 1
popular post; at least has 1 popular comment or at least likes 1 popular comment.
\end{Definition}
For CBS News and The New York Times, we selected popular content and then the active users in the same page.
The result \ref{statistics_active} is shown in the following:
\begin{table}[thb]
\begin{center}
\small
\caption{The Statistics of Active Users in Facebook Pages} \label{statistics_active}
\begin{tabular}{lll}
\hline
Category & CBS & NY Times \\
\hline
Users & 575 & 1,582 \\
Posts & 42 & 54 \\
Comments & 3,015 & 1,799 \\
Likes & 15,000 & 64,014 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Comparison of Different Methods}
We choose the multi-view clustering\cite{bickel2004multi} for the baseline.
To be fair, the weighted graph is taken as the base on which modularity is
calculated for all methods.
Multi-Level \cite{blondel2008fast} and LPA \cite{raghavan2007near} are selected
as our core community detection algorithms.
We tested these methods on the active users data described in Table \ref{statistics_active}.
Table \ref{cbs_comparison} and Table \ref{nyt_comparison} show the modularity
values calculated by a different method under the same graph and number of
communities.
We can see that our method based on the Multi-Level algorithm outperforms the other two in CBS
News and The New York Times pages.
On the other hand, like other clustering algorithms, MVC has to estimate the
optimal number of clusters so it repetitively runs at least K times, where K is
the maximum number of communities.
\begin{table}[thb]
\begin{center}
\caption{The Comparison of Methods for CBS} \label{cbs_comparison}
\begin{tabular}{ccc}
\hline
Method & Modularity & Clusters \\
\hline
MVC & 0.8196 & 14 \\
\hline
Weighted LPA & 0.8200 & 26 \\
\hline
Weighted Multi-Level & \bf{0.8373} & 20 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[thb]
\begin{center}
\caption{The Comparison of Methods for The New York Times} \label{nyt_comparison}
\begin{tabular}{ccc}
\hline
Method & Modularity & Clusters \\
\hline
MVC & 0.3899 & 7 \\
\hline
Weighted LPA & 0.1641 & 4 \\
\hline
Weighted Multi-Level & \bf{0.4241} & 10 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Community Detection in Multiple Pages}
In reality, Facebook users often participate in multiple pages.
For example, one user can like a post about the presidential election in the CBS News page
and write a comment on the post with a similar topic on the NBC page.
Therefore, we try to discover communities of users across multiple pages in this part.
The first dataset is ABC News and CBS News in the December of 2012.
The reason why we chose the two pages is that they share enough Facebook users.
The statistics of the two pages can be seen in Table \ref{abc_cbs_stats}:
\begin{table}[hb]
\begin{center}
\small
\caption{The Statistics of ABC and CBS in December 2012} \label{abc_cbs_stats}
\begin{tabular}{lll}
\hline
Category & ABC & CBS \\
\hline
Total users & 362,722 & 61,576 \\
Common users & 961 & 961 \\
Posts & 613 & 220 \\
Comments & 108,116 & 24,564 \\
Likes & 674,791 & 85,639 \\
\hline
\end{tabular}
\end{center}
\end{table}
Here, $961$ common users who are involved in the two pages and let the parameter
$\alpha$ be $0.5$.
Initially, we plotted the post graph in Figure \ref{co_like_post_two} and the comment graph in Figure \ref{comment_two}.
By applying Algorithm \ref{weighted_algorithm}, we get the communities structure
for the graphs in Table \ref{abc_cbs_community}.
\begin{table}[thb]
\begin{center}
\small
\caption{The Community Structure Information of ABC and CBS in December 2012} \label{abc_cbs_community}
\begin{tabular}{llll}
\hline
Category & View 1 & View 2 & Merged \\
\hline
Nodes & 961 & 961 & 961 \\
Edges & 17,188 & 169 & 17,343 \\
Clusters & 233 & 850 & 160 \\
Isolates & 217 & 818 & \bf{137} \\
Modularity & 0.6881 & \bf{0.8480} & 0.6912 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[tp]
\centering
\includegraphics[clip,width=0.9\linewidth]{co_like_post_graph.png}
\caption{The Community Structure of Post Graph for ABC and CBS}
\label{co_like_post_two}
\end{figure}
\begin{figure}[tp]
\centering
\includegraphics[clip,width=0.9\linewidth]{comment_graph.png}
\caption{The Community Structure of Comment Graph for ABC and CBS}
\label{comment_two}
\end{figure}
\begin{figure}[tp]
\centering
\includegraphics[clip,width=0.9\linewidth]{combined_graph.png}
\caption{The Community Structure of Weighted Graph in ABC and CBS}
\label{weighted_two}
\end{figure}
Both Figure \ref{co_like_post_two} and Figure \ref{comment_two} show sparse
structure. After we apply our multi-view community detection algorithm, it
\ref{weighted_two} erases most of the outliers in the comment graph.
The second dataset is ABC News, CBS News and NBC in the last half of $2012$.
In this dataset, we are interested in the users who have comment in ABC News and
likes in the rest two.
Consequently, we change the view $1$ into the graph in which any pair of users
concurrently like posts or comments and view $2$ into the graph in which any pair of users
like each other's comment.
The data statistics is shown in Table \ref{abc_cbs_nbc_data}.
And the community structure information is described in Table \ref{abc_cbs_nbc_community}.
\begin{table}[thb]
\begin{center}
\small
\caption{The Statistics of ABC, CBS and NBC of the Last Half of 2012} \label{abc_cbs_nbc_data}
\begin{tabular}{llll}
\hline
Category & ABC & CBS & NBC \\
\hline
Total users & 311,332 & 102,986 & 220,868 \\
Common users & 533 & 533 & 533 \\
Posts & 3,280 & 1,252 & 6,899 \\
Comments & 588,708 & 160,097 & 486,728 \\
Likes & 2,479,937 & 725,646 & 2,622,432 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[thb]
\begin{center}
\small
\caption{The Community Structure Information of ABC, CBS and NBC in the Last Half
2012} \label{abc_cbs_nbc_community}
\begin{tabular}{llll}
\hline
Category & View 1 & View 2 & Merged \\
\hline
Nodes & 533 & 533 & 533 \\
Edges & 129,674 & 612 & 129,708 \\
Clusters & 4 & 376 & \bf{6} \\
Isolates & 1 & 361 & 1 \\
Modularity & 0.0605 & \bf{0.5917} & 0.0604 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[tbp]
\centering
\includegraphics[clip,width=0.9\linewidth]{co_like_graph.png}
\caption{The Community Structure of Post Graph for ABC, CBS and NBC}
\label{co_like_three}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[clip,width=0.9\linewidth]{like_comment_graph.png}
\caption{ The Community Structure of Comment Graph for ABC, CBS and NBC}
\label{comment_three}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[clip,width=0.9\linewidth]{weighted_graph.png}
\caption{The Community Structure of Weighted Graph for ABC, CBS and NBC}
\label{weighted_three}
\end{figure}
Figure \ref{co_like_three} illustrates the users are located in three communities but Figure \ref{comment_three} shows the sparse structure.
Figure \ref{weighted_three} describes that most isolates are effectively removed in the post graph.
Interestingly, when there are more than two pages, users within a page
interact more than different pages so each community represents each page.
\section{Discussion}
Community detection in social networks as a research topic has attracted a significant amount of attention in the past few years.
Prior research provided innovative theories, algorithms, and applications. However little work has been done with regards to exploring community structure in Facebook public
pages.
In this work, we propose a weighted multi-view community detection method and
apply it to the Facebook newsgroup pages, CBS News, and The New York Times, etc.
We not only investigate the community structure in a single page but across two
or even three pages.
The results reveal three advantages of our method:
1) it can alleviate the isolates issue in the sparse network/view.
For example, in the CBS News page (last week of 2012), the isolates are decreased from $3,321$ and $6,647$ to $4$.
2) more cohesive community structure can be found during the process, take the New
York Times page (last week of 2012) as an example, the number of communities
becomes $31$ from $3,395$ and $36,340$.
3) it discovers latent communities across multiple pages.
In the ABC, CBS and NBC pages, the common users show strong cohesion in post
graph, but in our weighted graph, two more communities are discovered, which might
provide useful information for recommend systems.
However, our method still has two limitations.
First, it can not handle directed graphs.
For example, like actions of users are one-way, which means one edge should
point from the user who has the like to the user who posts the comment.
Under such situation, it is unknown whether the community structure could be
changed.
Second, overlapping communities can not be discovered, but it is interesting because it has
more reasonable assumption that one user can belong to multiple communities.
More importantly, such overlapping structure might help to recommend users
items more precisely in real recommendation systems.
\section{References}
\bibliographystyle{ieeetr}
|
\section{Introduction}
Identifying speakers and their names in movies, and videos in general, is a primary task for many video analysis problems, including automatic subtitle labeling \cite{Hu15}, content-based video indexing and retrieval \cite{Zhang09}, video summarization \cite{Tapaswi14}, and video storyline understanding \cite{Tapaswi14}. It is a very challenging task, as the visual appearance of the characters changes over the course of the movie due to several factors such as scale, clothing, illumination, and so forth \cite{Arandjelovic05,Everingham06}. The annotation of movie data with speakers' names can be helpful in a number of applications, such as movie question answering \cite{Tapaswi16}, automatic identification of character relationships \cite{Zhang09}, or automatic movie captioning \cite{Hu15}.
\begin{figure}[t]
\begin{center}
\scalebox{0.99}{
\includegraphics[width=\linewidth]{figures/introduction-ilovepdf-compressed.pdf}
}
\end{center}
\vspace{-0.2cm}
\caption{Overview of our approach for speaker naming.}
\vspace{-0.2cm}
\label{fig:long}
\end{figure}
Most previous studies relied primarily on visual information \cite{Arandjelovic05,Everingham06}, and aimed for the slightly different task of face track labeling; speakers who did not appear in the video frame were not assigned any names, which is common in movies and TV shows. Other available sources of information such as scripts were only used to extract cues about the speakers' names to associate the faces in the videos with their corresponding character name \cite{Everingham06,Tapaswi15,Baeuml13,Sivic09}; however since scripts are not always available, the applicability of these methods is somehow limited.
Other studies focused on the problem of speaker recognition without naming, using the speech modality as a single source of information. While some of these studies attempted to incorporate the visual modality, their goal was to cluster the speech segments rather than name the speakers \cite{Erzin05,Bost14,Kapsouras15,Bredin16,Hu15,Ren16}. None of these studies used textual information (e.g., dialogue), which prevented them from identifying speaker names.
In our work, we address the task of speaker naming, and propose a new multimodal model that leverages in an unified framework of the visual, speech, and textual modalities that are naturally available while watching a movie. We do not assume the availability of a movie script or a cast list, which makes our model fully unsupervised and easily applicable to unseen movies.
The paper makes two main contributions. First, we introduce a new unsupervised system for speaker naming for movies and TV shows that exclusively depends on videos and subtitles, and relies on a novel unified optimization framework that fuses visual, textual, and acoustic modalities for speaker naming. Second, we construct and make available a dataset consisting of 24 movies with 31,019 turns manually annotated with character names. Additionally, we also evaluate the role of speaker naming when embedded in an end-to-end memory network model, achieving state-of-the-art performance results on the subtitles task of the MovieQA 2017 Challenge.
\section{Related Work}
\label{sec:related}
The problem of speaker naming in movies has been explored by the computer vision and the speech communities. In the computer vision community, the speaker naming problem is usually considered as a face/person naming problem, in which names are assigned to their corresponding faces on the screen \cite{Everingham06,Cour10,Baeuml13,Haurilet16,Tapaswi15}. On the other hand, the speech community considered the problem as a speaker identification problem, which focuses on recognizing and clustering speakers rather than naming them \cite{Reynolds02,Campbell97}. In this work, we aim to solve the problem of speaker naming in movies, in which we label each segment of the subtitles with its corresponding speaker name whether the speaker's face appeared on in the video or not.
Previous work can be furthered categorized according to the type of supervision used to build the character recognition and speaker recognition models: supervised vs. weakly supervised models. In the movie and television domains, utilizing scripts in addition to subtitles to obtain timestamped speaker information was also studied in \cite{Everingham06,Tapaswi15,Baeuml13,Sivic09}. Moreover, they utilized this information to resolve the ambiguity introduced by co-occurring faces in the same frame. Features were extracted through the period of speaking (detected via lip motion on each face). Then they assigned the face based on candidate names from the time-stamped script. Thus, these studies used speaker recognition as an essential step to construct cast-specific face classifiers. \cite{Tapaswi12} extended the face identification problem to include person tracking. They utilized available face recognition results to learn clothing models for characters to identify person tracks without faces.
In \cite{Cour10,Haurilet16}, the authors proposed a weakly supervised model depending on subtitles and a character list. They extracted textual cues from the dialog: first, second, and third person references, such as ``I'm Jack'', ``Hey, Jack!'', and ``Jack left''. Using a character list from IMDB, they mapped these references onto true names using minimum edit distance, and then they ascribed the references to face tracks. Other work removed the dependency on a true character list by determining all names through coreference resolution. However, this work also depended on the availability of scripts \cite{Ramanathan14}. In our model, we removed the dependency on both the true cast list and the script, which makes it easier to apply our model to other movies and TV shows.
Recent work proposed a convolutional neural network (CNN) and Long Short-Term Memory (LSTM) based learning framework to automatically learn a function that combines both facial and acoustic features \cite{Hu15,Ren16}. Using these cues, they tried to learn matching face-audio pairs and non-matching face-audio pairs. They then trained a SVM classifier on the audio-video pairings to discriminate between the non-overlapping speakers. In order to train their models, they manually identified the leading characters in two TV shows, Friends and The Big Bang Theory (BBT), and collected their face tracks and corresponding audio segments using pre-annotated subtitles. Despite the very high performance reported in these studies, it is very hard to generalize their approach since it requires a lot of training data.
On the other hand, talking faces have been used to improve speaker recognition and diarization in TV shows \cite{Bredin16,Bost14,Li04}. In the case of \cite{Liu08}, they modeled the problem of speaker naming as facial recognition to identify speakers in news broadcasts. This work leveraged optical character recognition to read the broadcasters' names that were displayed on screen, requiring the faces to already be annotated.
\section{Datasets}
\label{sec:data}
Our dataset consists of a mix of TV show episodes and full movies. For the TV show, we use six full episodes of season one of the BBT. The number of named characters in the BBT episodes varies between 5 to 8 characters per episode, and the background noise level is low. Additionally, we also acquired a set of eighteen full movies from different genres, to evaluate how our model works under different conditions. In this latter dataset, the number of named characters ranges between 6 and 37, and it has varied levels of background noise.
We manually annotated this dataset with the character name of each subtitle segment. To facilitate the annotation process, we built an interface that parses the movies subtitles files, collects the cast list from IMDB for each movie, and then shows one subtitle segment at a time along with the cast list so that the annotator can choose the correct character. Using this tool, human annotators watched the movies and assigned a speaker name to each subtitle segment. If a character name was not mentioned in the dialogue, the annotators labeled it as ``unknown.'' To evaluate the quality of the annotations, five movies in our dataset were double annotated. The Cohen's Kappa inter-annotator agreement score for these five movies is 0.91, which shows a strong level of agreement.
To clean the data, we removed empty segments, as well as subtitle description parts written between brackets such as ``[groaning]'' and ``[sniffing]''. We also removed segments with two speakers at the same time. We intentionally avoided using any automatic means to split these segments, to preserve the high-quality of our gold standard.
Table \ref{table:datastats} shows the statistics of the collected data. Overall, the dataset consists of 24 videos with a total duration of 40.28 hours, a net dialogue duration of 21.99 hours, and a total of 31,019 turns spoken by 463 different speakers. Four of the movies in this dataset are used as a development set to develop supplementary systems and to fine tune our model's parameters; the remaining movies are used for evaluation.
\begin{table}[h]
\centering
\scalebox{0.8}{
\begin{tabular} { l c c c c c } \hline
& Min & Max & Mean & $\sigma$ \\ \hline
\# characters/video & 5 & 37 & 17.8 & 9.55 \\
\# Subtitle turns/video & 488 & 2212 & 1302.4 & 563.06 \\
\# words/turn & 1 & 28 & 8.02 & 4.157 \\
subtitles duration (sec) & 0.342 & 9.59 & 2.54 & 1.02 \\%& 2.38 \\
\hline
\end{tabular}
}
\vspace{-0.2cm}
\caption{Statistics on the annotated movie dataset.}
\vspace{-0.2cm}
\label{table:datastats}
\end{table}
\section{Data Processing and Representations}
\label{sec:modalities}
We process the movies by extracting several textual, acoustic, and visual features.
\subsection{Textual Features}
We use the following representations for the textual content of the subtitles:
\noindent \textbf{SkipThoughts} uses a Recurrent Neural Network to capture the underlying semantic and syntactic properties, and map them to a vector representation \cite{Kiros15}. We use their pretrained model to compute a 4,800 dimensional sentence representation for each line in the subtitles.\footnote{https://github.com/ryankiros/skip-thoughts}
\noindent \textbf{TF-IDF} is a traditional weighting scheme in information retrieval. We represent each subtitle as a vector of tf-idf weights, where the length of the vector (i.e., vocabulary size) and the idf scores are obtained from the movie including the subtitle.
\subsection{Acoustic Features}
For each movie in the dataset, we extract the audio from the center channel. The center channel is usually dedicated to the dialogue in movies, while the other audio channels carry the surrounding sounds from the environment and the musical background. Although doing this does not fully eliminate the noise in the audio signal, it still improves the speech-to-noise ratio of the signal. When a movie has stereo sound (left and right channels only), we down-mix both channels of the stereo stream into a mono channel.
In this work, we use the subtitles timestamps as an estimate of the boundaries that correspond to the uttered speech segments. Usually, each subtitle corresponds to a segment being said by a single speaker.
We use the subtitle timestamps for segmentation so that we can avoid automatic speaker diarization errors and focus on the speaker naming problem.
To represent the relevant acoustic information from each spoken segment, we use iVectors, which is the state-of-the-art unsupervised approach in speaker verification \cite{Dehak11}. While other deep learning-based speaker embeddings models also exist, we do not have access to enough supervised data to build such models. We train unsupervised iVectors for each movie in the dataset, using the iVector extractor used in \cite{Khorram16}. We extract iVectors of size 40 using a Gaussian Mixture Model-Universal Background Model (GMM-UBM) with 512 components. Each iVector corresponds to a speech segment uttered by a single speaker. We fine tune the size of the iVectors and the number of GMM-UBM components using the development dataset.
\subsection{Visual Features}
We detect faces in the movies every five frames using the recently proposed MTCNN \cite{zhang2016joint} model, which is pretrained for face detection and facial landmark alignment. Based on the results of face detection, we apply the forward and backward tracker with an implementation of the Dlib library \cite{King09,Danelljan14} to extract face tracks from each video clip. We represent a face track using its best face in terms of detection score, and use the activations of the fc7 layer of pretrained VGG-Face \cite{Parkhi15} network as visual features.
We calculate the distance between the upper lip center and the lower lip center based on the 68-point facial landmark detection implemented in the Dlib library \cite{King09,Kazemi14}. This distance is normalized by the height of face bounding boxes and concatenated across frames to represent the amount of mouth opening. A human usually speaks with lips moving with a certain frequency (3.75 Hz to 7.5 Hz used in this work) \cite{Tapaswi15}. We apply a band-pass filter to amplify the signal of true lip motion in these segments. The overall sum of lip motion is used as the score for the talking face.
\section{Unified Optimization Framework}
\label{sec:approach}
We tackle the problem of speaker naming as a transductive learning problem with constraints. In this approach, we want to use the sparse positive labels extracted from the dialogue and the underlying topological structure of the rest of the unlabeled data. We also incorporate multiple cues extracted from both textual and multimedia information. A unified learning framework is proposed to enable the joint optimization over the automatically labeled and unlabeled data, along with multiple semantic cues.
\subsection{Character Identification and Extraction}
\label{subsec:char}
In this work, we do not consider the set of character names as given because we want to build a model that can be generalized to unseen movies. This strict setting adds to the problem's complexity. To extract the list of characters from the subtitles, we use the Named Entity Recognizer (NER) in the Stanford CoreNLP toolkit \cite{Manning14}. The output is a long list of person names that are mentioned in the dialogue. This list is prone to errors including, but not limited to, nouns that are misclassified by the NER as person's name such as ``Dad" and ``Aye", names that are irrelevant to the movie such as ``Superman" or named animals, or uncaptured character names.
To clean the extracted names list of each movie, we cluster these names based on string minimum edit distance and their gender. From each cluster, we then pick a name to represent it based on its frequency in the dialogue. The result of this step consists of name clusters along with their distribution in the dialogue. The distribution of each cluster is the sum of all the counts of its members. To filter out irrelevant characters, we run a name reference classifier, which classifies each name into first, second or third person references. If a name was only mentioned as a third person throughout the whole movie, we discard it from the list of characters. We remove any name cluster that has a total count less than three, which takes care of the misclassified names' reference types.
\subsection{Grammatical Cues}
\label{subsec:ref}
We use the subtitles to extract the name mentions in the dialogue. These mentions allow us to obtain cues about the speaker name and the absence or the presence of the mentioned character in the surrounding subtitles. Thus, they affect the probability that the mentioned character is the speaker or not. We follow the same name reference categories used in \cite{Cour10,Haurilet16}. We classify a name mention into: first (e.g., ``I'm Sheldon''), second (e.g., ``Oh, hi, Penny'') or third person reference (e.g., ``So how did it go with Leslie?''). The first person reference represents a positive constraint that allows us to label the corresponding iVector of the speaker and his face if it exists during the segment duration. The second person reference represents a multi-instance constraint that suggests that the mentioned name is one of the characters that are present in the scene, which increases the probability of this character to be one of the speakers of the surrounding segments. On the other hand, the third person reference represents a negative constraint, as it suggests that the speaker does not exist in the scene, which lowers the character probability of the character being one of the speakers of the next or the previous subtitle segments.
To identify first, second and third person references, we train a linear support vector classifier. The first person, the second and third person classifier's training data are extracted and labeled from our development dataset, and fine tuned using 10-fold cross-validation. Table \ref{table:ref_class} shows the results of the classifier on the test data. The average number of first, second and third-person references in each movie are 14.63, 117.21, and 95.71, respectively.
\begin{table}[ht]
\centering
\scalebox{0.9}{
\begin{tabular}{l c c c}
\hline
& Precision & Recall & F1-Score \\ \hline
First Person & 0.625 & 0.448 & 0.522 \\
Second Person & 0.844 & 0.863 & 0.853 \\
Third Person & 0.806 & 0.806 & 0.806 \\
Average / Total & 0.819 & 0.822 & 0.820 \\ \hline
\end{tabular}
}
\vspace{-0.2cm}
\caption{Performance metrics of the reference classifier on the test data.}
\vspace{-0.2cm}
\label{table:ref_class}
\end{table}
\subsection{Unified Optimization Framework}
\label{subsec:semi}
Given a set of data points that consist of $l$ labeled\footnote{Note that in our setup, all the labeled instances are obtained automatically, as described above.} and $u$ unlabeled instances, we apply an optimization framework to infer the best prediction of speaker names. Suppose we have $l+u$ instances $X=\{x_1,x_2,...,x_l,x_{l+1},...,x_{l+u}\}$ and $K$ possible character names. We also get the dialogue-based positive labels $y_i$ for instances $x_i$, where $y_i$ is a $k$-dimension one-hot vector and $y_{i}^j=1$ if $x_i$ belongs to the class $j$, for every $1\leq i\leq l$ and $1\leq j \leq K$. To name each instance $x_i$, we want to predict another one-hot vector of naming scores $f(x_i)$ for each $x_i$, such that $\mbox{argmax}_{j} f^j(x_i) = z_i$ where $z_i$ is the ground truth number of class for instance $x_i$.
To combine the positive labels and unlabeled data, we define the objective function for predictions $f$ as follows:
\begin{equation}
\begin{aligned}
L_{initial}(f) &= \frac{1}{l}\sum_{i=1}^l ||f(x_i)-y_i||^2 \\
&+ \frac{1}{l+u} \sum_{i=1}^{l+u}\sum_{j=1}^{l+u} w_{ij} ||f(x_i)-f(x_j)|| ^2
\end{aligned}
\label{eqn:semi}
\end{equation}
Here $w_{ij}$ is the similarity between $x_i$ and $x_j$, which is calculated as the weighted sum of textual, acoustic and visual similarities. The inverse Euclidean distance is used as similarity function for each modality. The weights for different modalities are selected as hyperparameters and tuned on the development set. This objective leads to a convex loss function which is easier to optimize over feasible predictions.
Besides the positive labels obtained from first person name references, we also introduce other semantic constraints and cues to enhance the power of our proposed approach. We implement the following four types of constraints:
\mysubsection{Multiple Instance Constraint.} Although the second person references cannot directly provide positive constraints, they imply that the mentioned characters have high probabilities to be in this conversation. Following previous work \cite{Cour10}, we incorporate the second person references as multiple instances constraints into our optimization: if $x_i$ has a second person reference $j$, we encourage $j$ to be assigned to its neighbors, i.e., its adjacent subtitles with similar timestamps. For the implementation, we simply include multiple instances constraints as a variant of positive labels with decreasing weights $s$, where $s=1/(l-i)$ for each neighbor $x_l$.
\mysubsection{Negative Constraint.} For the third person references, the mentioned characters may not occur in the conversation and movies. So we treat them as negative constraints, which means they imply that the mentioned characters should not be assigned to corresponding instances. This constraint is formulated as follows:
\vspace{-0.2cm}
\begin{equation}
L_{neg}(f) = \sum_{(i,j)\in N} [f^j(x_i)]^2
\end{equation}
where $N$ is the set of negative constraints $x_i$ doesn't belong class $j$.
\mysubsection{Gender Constraint.} We train a voice-based gender classifier by using the subtitles segments from the four movies in our development dataset (5,543 segments of subtitles). We use the segments in which we know the speaker's name and manually obtain the ground truth gender label from IMDB. We extract the signal energy, 20 Mel-frequency cepstral coefficients (MFCCs) along with their first and second derivatives, in addition to time- and frequency-based absolute fundamental frequency (f0) statistics as features to represent each segment in the subtitles. The f0 statistics has been found to improve the automatic gender detection performance for short speech segments \cite{Levitan16}, which fits our case since the median duration of the dialogue turns in our dataset is 2.6 seconds.
The MFCC features are extracted using a step size of 16 msec over a 64 msec window using the method from \cite{Mathieu10}, while the f0 statistics are extracted using a step size of 25 msec over a 50 msec window as the default configuration in \cite{Eyben13}. We then use these features to train a logistic regression classifier using the Scikit-learn library \cite{scikit-learn}. The average accuracy of the gender classifier on a 10-fold cross-validation is 0.8867.
Given the results for the gender classification of audio segments and character names, we define the gender loss to penalize inconsistency between the predicted gender and character names:
\vspace{-0.1cm}
\begin{equation}
\begin{aligned}
L_{gender}(f) &= \sum_{(i,j)\in Q_1}P_{ga}(x_i) (1-P_{gn}(j))f^j(x_i)\\
&+\sum_{(i,j)\in Q_2}(1-P_{ga}(x_i))P_{gn}(j)f^j(x_i)
\end{aligned}
\end{equation}
where $P_{ga(x_i)}$ is the probability for instance $x_i$ to be a male, and $P_{gn(j)}$ is the probability for name $j$ to be a male, and $Q_1=\{(i,j)|P_{ga}(x_i)<0.5,P_{gn}(j)>0.5\}$, $Q_2=\{(i,j)|P_{ga}(x_i)>0.5,P_{gn}(j)<0.5\}$.
\mysubsection{Distribution Constraint.} We automatically analyze the dialogue and extract the number of mentions of each character in the subtitles using Stanford CoreNLP and string matching to capture names that are missed by the named entity recognizer. We then filter the resulting counts by removing third person mention references of each name as we assume that this character does not appear in the surrounding frames. We use the results to estimate the distribution of the speaking characters and their importance in the movies. The main goal of this step is to construct a prior probability distribution for the speakers in each movie.
To encourage our predictions to be consistent with the dialogue-based priors, we penalize the square error between the distributions of predictions and name mentions priors in the following equation:
\vspace{-0.5cm}
\begin{equation}
L_{dis}(f) = \sum_{j=1}^K (\sum(f^j(x_i)) - d_j)^2
\end{equation}
\noindent where $d_j$ is the ratio of name $j$ mentions in all subtitles.
\mysubsection{Final Framework.} Combining the loss in Eqn. \ref{eqn:semi} and multiple losses with different constraints, we obtain our unified optimization problem:
\vspace{-0.02cm}
\begin{equation}
\begin{aligned}
f^{\ast} &= \arg\min_{f} \lambda_1 L_{initial}(f)+ \lambda_2 L_{MI}(f) \\
& + \lambda_3 L_{neg}(f) + \lambda_4 L_{gender}(f) + \lambda_5 L_{dis}(f)
\end{aligned}
\label{eqn:unify}
\end{equation}
All of the $\lambda$s are hyper-parameters to be tuned on development set. We also include the constraint that predictions for different character names must sum to 1. We solve this constrained optimization problem with projected gradient descent (PGD). Our optimization problem in Eqn. \ref{eqn:unify} is guaranteed to be a convex optimization problem and therefore projected gradient descent is guaranteed to stop with global optima. PGD usually converges after 800 iterations.
\section{Evaluation}
\label{sec:eval}
We model our task as a classification problem, and use the unified optimization framework described earlier to assign a character name to each subtitle.
Since our dataset is highly unbalanced, with a few main characters usually dominating the entire dataset, we adopt the weighted F-score as our evaluation metric, instead of using an accuracy metric or a micro-average F-score. This allows us to take into account that most of the characters have only a few spoken subtitle segments, while at the same time placing emphasis on the main characters. This leads sometimes to an average weighted F-score that is not between the average precision and recall.
\begin{table}[htb]
\centering
\scalebox{0.75}{
\begin{tabular}{ l c c c } \hline
& Precision & Recall & F-score \\\hline
B1: MFMC & 0.0910 & 0.2749 & 0.1351 \\
B2: DRA & 0.2256 & 0.1819 & 0.1861 \\
B3: Gender-based DRA & 0.2876 & 0.2349 & 0.2317 \\ \hline
Our Model (Skip-thoughts)\mbox{*} & 0.3468 & 0.2869 & 0.2680 \\
Our Model (TF-IDF)\mbox{*} & 0.3579 & 0.2933 & 0.2805 \\
Our Model (iVectors) & 0.2151 & 0.2347 & 0.1786 \\
Our Model (Visual)\mbox{*} & 0.3348 & 0.2659 & 0.2555 \\
Our Model (Visual+iVectors)\mbox{*} & 0.3371 & 0.2720 & 0.2617 \\
Our Model (TF-IDF+iVectors)\mbox{*} & 0.3549 & 0.2835 & 0.2643 \\
Our Model (TF-IDF+Visual)\mbox{*} & 0.3385 & 0.2975 & 0.2821 \\ \hline
Our Model (all)\mbox{*} & \textbf{0.3720} & \textbf{0.3108} & \textbf{0.2920} \\ \hline
\end{tabular}
}
\vspace{-0.2cm}
\caption{Comparison between the average of macro-weighted average of precision, recall and f-score of the baselines and our model. * means statistically significant (t-test p-value $<$ 0.05) when compared to baseline B3.}
\vspace{-0.2cm}
\label{table:eval}
\end{table}
One aspect that is important to note is that characters are often referred to using different names. For example, in the movie ``The Devil's Advocate," the character Kevin Lomax is also referred to as Kevin or Kev. In more complicated situations, characters may even have multiple identities, such as the character Saul Bloom in the movie ``Ocean's Eleven,'' who pretends to be another character named Lyman Zerga. Since our goal is to assign names to speakers, and not necessarily solve this coreference problem, we consider the assignment of the subtitle segments to any of the speaker's aliases to be correct. Thus, during the evaluation, we map all the characters' aliases from our model's output to the names in the ground truth annotations. Our mapping does not include other referent nouns such as ``Dad,'' ``Buddy,'' etc.; if a segment gets assigned to any such terms, it is considered a misprediction.
We compare our model against three baselines:
\noindent \textbf{B1: Most-frequently mentioned character} consists of selecting the most frequently mentioned character in the dialogue as the speaker for all the subtitles. Even though it is a simple baseline, it achieves an accuracy of 27.1\%, since the leading characters tend to speak the most in the movies.
\noindent \textbf{B2: Distribution-driven random assignment} consists of randomly assigning character names according to a distribution that reflects their fraction of mentions in all the subtitles.
\noindent \textbf{B3: Gender-based distribution-driven random assignment} consists of selecting the speaker names based on the voice-based gender detection classifier. This baseline randomly selects the character name that matches the speaker's gender according to the distribution of mentions of the names in the matching gender category.
The results obtained with our proposed unified optimization framework and the three baselines are shown in Table \ref{table:eval}. We also report the performance of the optimization framework using different combinations of the three modalities. The model that uses all three modalities achieves the best results, and outperforms the strongest baseline (B3) by more than 6\% absolute in average weighted F-score. It also significantly outperforms the usage of the visual and acoustic features combined, which have been frequently used together in previous work, suggesting the importance of textual features in this setting.
\begin{figure}[!tb]
\begin{center}
\vspace{-0.09cm}
\begin{subfigure}[t]{1\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/BBT.png}
\vspace{-0.5cm}
\caption{The Big Bang Theory}
\end{subfigure}%
\vspace{-0.1cm}
\begin{subfigure}[t]{1\linewidth}
\centering
\includegraphics[width=\linewidth]{figures/tt0120338.png}
\vspace{-0.5cm}
\caption{Titanic}
\end{subfigure}%
\end{center}
\vspace{-0.2cm}
\caption{For each speech segment, we applied t-SNE \cite{van2014accelerating} on their corresponding iVectors. The points with the same color represent instances with the same character name.}
\vspace{-0.3cm}
\label{fig:ivec_cluster}
\end{figure}
The ineffectiveness of the iVectors might be a result of the background noise and music, which are difficult to remove from the speech signal. Figure \ref{fig:ivec_cluster} shows the t-Distributed Stochastic Neighbor Embedding (t-SNE) \cite{van2014accelerating}, which is a nonlinear dimensionality reduction technique that models points in such a way that similar vectors are modeled by nearby points and dissimilar objects are modeled by distant points, visualization of the iVectors over the whole BBT show and the movie ``Titanic.'' In the BBT there is almost no musical background or background noise, while, Titanic has musical background in addition to the background noise such as the screams of the drowning people. From the graph, the difference between the quality of the iVectors clusters on different noise-levels is clear.
Table \ref{table:eval_loss} shows the effect of adding components of our loss function to the initial loss $L_{init}$ function. The performance of the model using only $L_{init}$ without the other parts is very low due to the sparsity of first person references and errors that the person reference classifier introduces.
\begin{table}[htb]
\centering
\scalebox{0.8}{
\begin{tabular}{ l c c c } \hline
& Precision & Recall & F-score \\ \hline
$L_{initial}$ & 0.0631 & 0.1576 & 0.0775 \\
$L_{initial}$ + $L_{gender}$ & 0.1160 & 0.1845 & 0.1210 \\
$L_{initial}$ + $L_{negative}$ & 0.0825 & 0.0746 & 0.0361 \\
$L_{initial}$ + $L_{distribution}$ & 0.1050 & 0.1570 & 0.0608 \\
$L_{initial}$ + $L_{Multiple Instance}$& 0.3058 & 0.2941 & 0.2189 \\
\hline
\end{tabular}
}
\vspace{-0.2cm}
\caption{Analysis of the effect of adding each component of the loss function to the initial loss.}
\vspace{-0.2cm}
\label{table:eval_loss}
\end{table}
In order to analyze the effect of the errors that several of the modules (e.g., gender and name reference classifiers) propagate into the system, we also test our framework by replacing each one of the components with its ground truth information. As seen in Table \ref{tab:evalgt}, the results obtained in this setting show significant improvement with the replacement of each component in our framework, which suggests that additional work on these components will have positive implications on the overall system.
\begin{table}[htb]
\centering
\scalebox{0.8}{
\begin{tabular}{ l c c c } \hline
& Precision & Recall & F-score \\ \hline
Our Model & 0.3720 & 0.3108 & 0.2920 \\
Voice Gender (VG) & 0.4218 & 0.3449 & 0.3259 \\
VG + Name Gender (NG) & 0.4412 & 0.3790 & 0.3645 \\
VG + NG + Name Ref & 0.4403 & 0.3938 & 0.3748 \\
\hline
\end{tabular}
}
\vspace{-0.1cm}
\caption{Comparison between our model while replacing different components with their ground truth information.}
\vspace{-0.1cm}
\label{tab:evalgt}
\end{table}
\begin{comment}
\section{Additional Analyses}
To gain further insights into the system, we perform several analyses.
First, we look at the performance of the system when considering the top two characters returned by our system. That is, if the correct character is listed in this top two, the system receives full credit. The average weighted precision, recall, and F-score grow to 0.5069, 0.4230, and 0.4177, which indicates there are many cases where the system finds the correct character as the second option.
Second, in order to analyze the effect of the errors that several of the used modules (e.g., gender and name reference classifiers) propagate into the system, we test our framework by replacing each one of the components with its ground truth information. As seen in Table \ref{tab:evalgt}, the results obtained in this setting show significant improvement with the replacement of each component in our framework, which suggests that additional work on these components will have positive implications on the overall system.
\begin{table}[htb]
\centering
\scalebox{0.8}{
\begin{tabular}{ l c c c } \hline
& Precision & Recall & F-score \\ \hline
Our Model & 0.3720 & 0.3108 & 0.2920 \\
Voice Gender (VG) & 0.4218 & 0.3449 & 0.3259 \\
VG + Name Gender (NG) & 0.4412 & 0.3790 & 0.3645 \\
VG + NG + Name Ref & 0.4403 & 0.3938 & 0.3748 \\
\hline
\end{tabular}
}
\vspace{-0.2cm}
\caption{Comparison between our model while replacing different components with their ground truth information.}
\label{tab:evalgt}
\end{table}
Third, to isolate the role played by each of the constraints used in the optimization model, we perform an evaluation where we add to the initial loss function one constraint at a time. Table \ref{table:eval_loss} shows the results of these evaluations. As seen in this table, the biggest improvement in our framework is brought by the multiple instance constraint, followed by the gender constraint. That is expected because the first person reference instances are rare and the only constraint that provides weak supervision is the multiple instance constraint, which uses the second person references for the labels. Then the gender loss makes sure that the speaker is not assigned a name of an opposite gender.
Finally, we also analyze the effect of the number of characters on the performance of our framework. The number of characters in the movies in our dataset varies from 5 to 37 characters. Figure \ref{fig:char_no} shows the performance of our model on videos categorized according to the number of their named speakers. As expected, the performance is better when there is a limited number of characters, however, from the figure we can see that the performance does not decrease by much after a certain number of characters. This could be explained by the fact that even movies with many characters are dominated by a few leading characters, and a large number of characters that only utter a few subtitle segments is not significantly affecting the performance of the system.
\begin{figure}[t]
\begin{center}
\scalebox{0.8}{
\includegraphics[width=\linewidth]{figures/char_no.pdf}
}
\end{center}
\vspace{-0.3cm}
\caption{The average weighted precision, recall and f-score of our model across videos with different number of characters.}
\vspace{-0.2cm}
\label{fig:char_no}
\end{figure}
\end{comment}
\section{Speaker Naming for Movie Understanding}
Identifying speakers is a critical task for understanding the dialogue and storyline in movies. MovieQA is a challenging dataset for movie understanding.
The dataset consists of 14,944 multiple choice questions about 408 movies. Each question has five answers and only one of them is correct. The dataset is divided into three splits: train, validation, and test according to the movie titles. Importantly, there are no overlapping movies between the splits. Table \ref{table:MQAExamples} shows examples of the question and answers in the MovieQA dataset.
\begin{table*}[ht]
\centering
\scalebox{0.87}{
\begin{tabular}{l l p{4.8cm}p{4cm}}
\hline
Movie & Question & \multicolumn{2}{l}{Answers}\\\hline
\multirow{3}{1.2cm}{Fargo}&
\multirow{3}{5cm}{What did Mike's wife, as he says, die from?} &
A1: She was killed &\textbf{A2: Breast cancer} \\
& & A3: Leukemia & A4: Heart disease \\
& &
\multicolumn{2}{l}{A5: Complications due to child birth}
\\\hline
\multirow{3}{1.2cm}{Titanic}&
\multirow{3}{5cm}{What does Rose ask Jack to do in her room?}
&
A1: Sketch her in her best dress & A2: Sketch her nude \\
& & A3: Take a picture of her nude & \textbf{A4: Paint her nude} \\
& &
\multicolumn{2}{l}{A5: Take a picture of her in her best dress}
\\\hline
\end{tabular}
}
\vspace{-0.2cm}
\caption{
Example of questions and answers from the MQA benchmark. The answers in bold are the correct answers to their corresponding question.}
\vspace{-0.2cm}
\label{table:MQAExamples}
\end{table*}
The MovieQA 2017 Challenge\footnote{http://movieqa.cs.toronto.edu/workshops/iccv2017/} consists of six different tasks according to the source of information used to answer the questions. Given that for many of the movies in the dataset the videos are not completely available, we develop our initial system so that it only relies on the subtitles; we thus participate in the challenge subtitles task, which includes the dialogue (without the speaker information) as the only source of information to answer questions.
\begin{figure}[t]
\begin{center}
\scalebox{0.99}{
\includegraphics[width=\linewidth]{figures/arch.pdf}
}
\end{center}
\vspace{-0.2cm}
\caption{The diagram describing our Speaker-based Convolutional Memory Network (SC-MemN2N) model.}
\vspace{-0.2cm}
\label{fig:scmemn}
\end{figure}
To demonstrate the effectiveness of our speaker naming approach, we design a model based on an end-to-end memory network \cite{Sukhbaatar15}, namely Speaker-based Convolutional Memory Network (SC-MemN2N), which relies on the MovieQA dataset, and integrates the speaker naming approach as a component in the network. Specifically, we use our speaker naming framework to infer the name of the speaker for each segment of the subtitles, and prepend the predicted speaker name to each turn in the subtitles.\footnote{We strictly follow the challenge rules, and only use text to infer the speaker names.} To represent the movie subtitles, we represent each turn in the subtitles as the mean-pooling of a 300-dimension pretrained word2vec \cite{Mikolov13} representation of each word in the sentence. We similarly represent the input questions and their corresponding answers. Given a question, we use the SC-MemN2N memory to find an answer. For questions asking about specific characters, we keep the memory slots that have the characters in question as speakers or mentioned in, and mask out the rest of the memory slots. Figure \ref{fig:scmemn} shows the architecture of our model.
Table \ref{table:movieqa_results} includes the results of our system on the validation and test sets, along with the best systems introduced in previous work, showing that our SC-MemN2N achieves the best performance. Furthermore, to measure the effectiveness of adding the speaker names and masking, we test our model after removing the names from the network (C-MemN2N). As seen from the results, the gain of SC-MemN2N is statistically significant\footnote{Using a t-test p-value$<$0.05 with 1,000 folds each containing 20 samples.} compared to a version of the system that does not include the speaker names (C-MemN2N). Figure \ref{fig:scmemn_q} shows the performance of both C-MemN2N and SC-MemN2N models by question type. The results suggest that our speaker naming helps the model better distinguish between characters, and that prepending the speaker names to the subtitle segments improves the ability of the memory network to correctly identify the supporting facts from the story that answers a given question.
\begin{table}[!htb]
\centering
\scalebox{0.9}{
\begin{tabular}{ l c c} \hline
\multirow{2}{0em}{Method} & \multicolumn{2}{c}{Subtitles} \\
& val & test \\ \hline
SSCB-W2V \cite{Tapaswi16} & 24.8 & 23.7 \\
SSCB-TF-IDF \cite{Tapaswi16} & 27.6 & 26.5 \\
SSCB Fusion \cite{Tapaswi16} & 27.7 & - \\
MemN2N \cite{Tapaswi16} & 38.0 & 36.9 \\
Understanding visual regions & - & 37.4 \\
RWMN \cite{Na17} & 40.4 & 38.5 \\ \hline
C-MemN2N (w/o SN) & 40.6 & - \\
SC-MemN2N (Ours) & \textbf{42.7} & \textbf{39.4} \\
\hline
\end{tabular}
}
\vspace{-0.1cm}
\caption{Performance comparison for the subtitles task on the MovieQA 2017 Challenge on both validation and test sets. We compare our models with the best existing models (from the challenge leaderboard).
}
\vspace{-0.1cm}
\label{table:movieqa_results}
\end{table}
\begin{figure}[htb]
\begin{center}
\scalebox{0.99}{
\includegraphics[width=\linewidth, height=5cm]{figures/qbreak_down.pdf}
}
\end{center}
\vspace{-0.2cm}
\caption{Accuracy comparison according to question type.}
\vspace{-0.2cm}
\label{fig:scmemn_q}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
In this paper, we proposed a unified optimization framework for the task of speaker naming in movies. We addressed this task under a difficult setup, without a cast-list, without supervision from a script, and dealing with the complicated conditions of real movies. Our model includes textual, visual, and acoustic modalities, and incorporates several grammatical and acoustic constraints. Empirical experiments on a movie dataset demonstrated the effectiveness of our proposed method with respect to several competitive baselines. We also showed that an SC-MemN2N model that leverages our speaker naming model can achieve state-of-the-art results on the subtitles task of the MovieQA 2017 Challenge.
The dataset annotated with character names introduced in this paper is publicly available from \url{http://lit.eecs.umich.edu/downloads.html}.
\section*{Acknowledgments}
We would like to thank the anonymous reviewers for their valuable comments and suggestions. This work is supported by a Samsung research grant and by a DARPA grant HR001117S0026-AIDA-FP-045.
|
\section{Introduction}
Control of robotic manipulators on a fixed or wheeled base is well studied in the robot control community; see, e.g., \cite{borst2009rollin, englsberger2014overview, kim2015bringing, kim2017enhancing}. In order to further increase the workspace of the manipulation, a recently proposed solution is given by the combination of a robotic manipulator with a flying base (Fig. \ref{fig:configuration}a), e.g. a helicopter \cite{kondak2013closed, huber2013first, kim2018stabilizing}. However it has a drawback due to the large rotor blades which may cause safety issues when operating in a complex environment. To alleviate this drawback, a new configuration, which can decouple the helicopter and manipulator, is proposed. In this configuration, the robotic manipulator is mounted on a floating platform, which can be suspended on the helicopter by means of wires (Fig. \ref{fig:configuration}b). If the flying helicopter is controlled to be fixed, then the hanging platform is nothing but the oscillating pendulum.
\begin{figure}
\centering
\subfigure[Heli-manipulator system developed in DLR]{
\centering
\includegraphics[scale=0.70]{fig1a.png} } \\
\subfigure[]{
\centering
\includegraphics[scale=0.55]{fig1b.pdf} }
\caption{Schematic diagram of the system of interest.}
\label{fig:configuration}
\end{figure}
In the control point of view, the main challenge is the oscillations caused by the external force and the movement of the robotic manipulator. \cite{potter2015planar, chen2016swing, vyhlidal2017time} solved similar problems by controlling the base. In this paper, because the precise control of the base (helicopter) is not so trivial for our system, additional moving masses are installed to damp out the oscillations of the pendulum-like hanging platform. Furthermore, to reduce the complexity of the problem, so that we can focus only on the oscillation damping of the platform, propellers are considered to control yaw angle of the platform. Notice that the proposed configuration is an under-actuated mechanical system with a fewer control input than the degrees of freedom (DoF).
Control of under-actuated systems is still an active research field, as the stabilization is not always trivial \cite{shiriaev2014controlled, lee2015robust}. \cite{spong1997control} introduced the energy-based methods with saturation function to swing up the Acrobot, which is a two-link planar robot with one actuator at the elbow.
\cite{albu2012energy} have also applied the potential energy shaping control for the underactuated Euler-Lagrange systems by introducing a new feedback state variable, which are statically equivalent to the non-collocated state variables. However from the configuration (Fig. \ref{fig:configuration}b), it can be seen that when the system is stable at a fixed height, the potential energy reaches a constant value, no matter where are the moving masses at the platform. Therefore, the moving masses can have multiple equilibrium points, which is not appealing. Based on this observation, it can be concluded that (potential) energy shaping method is not suitable for this configuration.
\cite{spong1994partial} made use of the collocated/non-collocated partial feedback linearization (PFL) to control the output dynamics by exploiting the strong inertial coupling under a certain condition. However the remaining dynamic after PFL (internal dynamics) should be carefully taken into account to achieve asymptotic stability of the overall dynamics.
In this paper, we present an extension of PFL technique to asymptotically stabilize the output dynamics as well as internal dynamics. In this paper, within the PFL framework, the dynamics of the moving mass is the output dynamics, and that of the platform orientation is the internal dynamics which does not have collocated actuation. To overcome the absence of the actuation, the internal dynamics is accessed using the reference acceleration of the moving masses. By designing it properly, the internal dynamics as well as the output dynamics can be asymptotically stabilized at the same time. The proposed approach is verified through simulation studies.
The rest the paper is organized as follows. Section \ref{sec:moelding_problem} introduces the modeling of the system and states the problem definition. Moreover, a motivating example that shows limitation of the standard PFL approach will be shown. To overcome the limitation, the extension of PFL with stabilization of internal dynamics will be proposed in Section \ref{sec:control}. Simulation validation will be illustrated in Section \ref{sec:simulation}. Section \ref{sec:conclusion} concludes the paper.
\section{System Modeling and Problem Definition }
\label{sec:moelding_problem}
\subsection{System description and control goal}
A schematic diagram of the system of interest is shown in Fig. \ref{fig:configuration}b. A robotic manipulator is mounted on the platform which is suspended by means of the wire which can be connected to, for example, a flying helicopter or a crane. Because the pendulum-like platform may oscillate due to the external disturbances (e.g. wind gust) and the movement of the robotic manipulator, the moving masses are installed to damp out the oscillation. The main control goal is to achieve oscillation damping of the platform (i.e., achieving zero roll and pitch angles) using moving masses. Propellers are considered to control yaw angle independently (see Fig. \ref{fig:configuration}b), so that we can focus only on the oscillation damping problem.
The equation of motion can be expressed as
\begin{align}
\mathbf{M}(\mathbf{q})\ddot{\mathbf{q}}+\mathbf{C}(\mathbf{q},\dot{\mathbf{q}})\dot{\mathbf{q}}+\mathbf{g}(\mathbf{q})=\boldsymbol{\tau}
\label{eq:whole1}
\end{align}
where the vector {$\mathbf{q}=[\mathbf{q}_p^T \; \mathbf{q}_m^T \; \mathbf{q}_r^T]^T \in \mathbb{R}^{3+2+n}$} consists of platform orientation $\mathbf{q}_p \in \mathbb{R}^3$, moving mass coordinates $\mathbf{q}_m \in \mathbb{R}^2$, and the robotic arm coordinates $\mathbf{q}_r \in \mathbb{R}^n$. In this paper, $\mathbf{q}_p$ is composed of roll, pitch, and yaw angles $\alpha$, $\beta$, $\gamma$; i.e., $\mathbf{q}_p=[\alpha \; \beta \; \gamma]^T$
The torque $\boldsymbol{\tau}$ is given by
\begin{align}
\begin{matrix}
\boldsymbol{\tau}=\begin{pmatrix}
\mathbf{0}_{2 \times 1} \\ \mathbf{u}
\end{pmatrix}, & \text{with\:\:} \mathbf{u}=\begin{pmatrix}
\tau_{yaw} \\ \boldsymbol{\tau}_m \\ \boldsymbol{\tau}_r
\end{pmatrix}, \boldsymbol{\tau}_m \in \mathbb{R}^2
\text{\; and \;} \boldsymbol{\tau}_r \in \mathbb{R}^n
\end{matrix}.
\end{align}
Equivalently, $\boldsymbol{\tau}$ can be represented as
\begin{align}
\boldsymbol{\tau} = {\mathbf{B}} \mathbf{u},
\end{align}
where
\begin{align}
\label{eq:def_B}
{\mathbf{B}} = \left[
\begin{array}{c}
\mathbf{0}_{2 \times (3+n)} \\
\bm{I}_{3+n}
\end{array}
\right] \in \Re^{(6+n)\times(3+n)}.
\end{align}
$\tau_{yaw}$ is the torque generated by propellers to control yaw angle. $\boldsymbol{\tau}_m$ and $\boldsymbol{\tau}_r$ are forces/torques to control moving masses and robotic manipulator, respectively. Note that there is no actuation along the roll ($\alpha$) and pitch ($\beta$) angles of the platform, and this under-actuated nature makes it difficult to achieve the control goal which can be summarized as follows.
\begin{enumerate}
\item {Robotic manipulator variable $\mathbf{q}_r$ converges to the desired value $\mathbf{q}_r^{des}$.}
\item {Platform yaw angle $\gamma$ converges to the desired value $\gamma^{des}$.}
\item {Platform roll and pitch angles $\alpha$, $\beta$ converge to zero (oscillation damping and gravity compensation of the pendulum-like platform).}
\item {Moving mass variable $\mathbf{q}_m$ converges to a certain equilibrium point $\mathbf{q}_m^{*}$ at which the moving masses balance the static gravity torque when $\mathbf{q}_r=\mathbf{q}_r^{des}$. Hereinafter, $\mathbf{q}_m^*$ will be assumed to be known as it can be pre-calculated.}
\end{enumerate}
\subsection{Motivating example using standard PFL approach}
In this section, a standard PFL technique will be applied to achieve oscillation damping of the platform (i.e., $\alpha=\beta=0$). To begin with, let us define the following notations for later convenience.
\begin{itemize}
\item {$\bm{\phi}=[\alpha \; \beta]^T$: Roll and pitch angles.}
\item {$\mathbf{M}_{\phi\phi}$: First 2 by 2 block matrix of $\mathbf{M}$.}
\item {$\mathbf{M}_{\phi m}$: First 2 by (4:5) block matrix of $\mathbf{M}$. Maps $\ddot{\mathbf{q}}_m$ to $\ddot{\bm{\phi}}$.}
\item {$\mathbf{C}_{\phi \phi} \in \Re^{2 \times 2}$ and $\mathbf{C}_{\phi m} \in \Re^{2 \times 2}$ are defined similarly.}
\item {$\mathbf{g}_{\phi}$: First two components of $\mathbf{g}$.}
\end{itemize}
To apply PFL, define the output variable by
\begin{align}
\label{eq:y_PFL}
\mathbf{y} =
\left(
\begin{array}{c}
\gamma \\
\mathbf{q}_m \\
\mathbf{q}_r
\end{array}
\right)
=
{\mathbf{B}}^T \mathbf{q}
\end{align}
with the selection matrix ${\mathbf{B}}$ defined in (\ref{eq:def_B}). Then,
\begin{align}
\nonumber \ddot{\mathbf{y}} =& {\mathbf{B}}^{T} \ddot{\mathbf{q}} \\
=& {\mathbf{B}}^{T} ( \mathbf{M}^{-1} -\mathbf{C} \dot{\mathbf{q}} - \mathbf{g} + {\mathbf{B}} \mathbf{u} ).
\end{align}
Using
\begin{align}
\mathbf{u} = ({\mathbf{B}}^{T} \mathbf{M}^{-1} {\mathbf{B}})^{-1} \left( {\mathbf{B}}^T \mathbf{M}^{-1} (\mathbf{C} \dot{\mathbf{q}} + \mathbf{g}) + \ddot{\mathbf{y}}^{ref} \right),
\end{align}
where $\ddot{\mathbf{y}} = [\ddot{\gamma}^{ref} \; \ddot{\mathbf{q}}_m^{ref,T} \; \ddot{\mathbf{q}}_r^{ref,T}]^T$, we obtain
\begin{align}
\ddot{\mathbf{y}} = \ddot{\mathbf{y}}^{ref}.
\end{align}
Defining
\begin{align}
\label{eq:PFL_gamma_ref}
\ddot{\gamma}^{ref} = -D_\gamma \dot{\gamma} - K_\gamma (\gamma - \gamma^{des}),\\
\label{eq:PFL_qr_ref}
\ddot{\mathbf{q}}_r^{ref} = -\mathbf{D}_r \dot{\mathbf{q}}_r - \mathbf{K}_r (\mathbf{q}_r - \mathbf{q}_r^{des}),
\end{align}
the configuration of the robotic manipulator and yaw angle will converge to the desired values. $\ddot{\mathbf{q}}_m^{ref}$ will be designed shortly to damp out the oscillation of the platform.
At this point, we introduce a well-known result from the cascade control literature \cite{seibert1990global}.
\begin{thm}
\label{thm:cascade}
Consider a system
\begin{align}
\label{eq:cascade_1}
\dot{\mathbf{x}}_1 =& \bm{f}_1(\mathbf{x}_1), \\
\label{eq:cascade_2}
\dot{\mathbf{x}}_2 =& \bm{f}_2(\mathbf{x}_1,\mathbf{x}_2) .
\end{align}
If $\dot{\mathbf{x}}_1=\bm{f}_1(\mathbf{x}_1)$ is locally asymptotically stable to $\mathbf{x}_1=\mathbf{0}$ and $\dot{\mathbf{x}}_2=\bm{f}_2(\mathbf{0},\mathbf{x}_2)$ is locally asymptotically stable to $\mathbf{x}_2=\mathbf{0}$, then (\ref{eq:cascade_1})-(\ref{eq:cascade_2}) is locally asymptotically stable to $\mathbf{x}_1=\mathbf{0}$ and $\mathbf{x}_2=\mathbf{0}$
\end{thm}
By the cascade control theory (Theorem \ref{thm:cascade}), $\gamma \equiv \gamma^{d}$, $\mathbf{q}_r \equiv \mathbf{q}_r^d$, and $\ddot{\mathbf{q}}_m \equiv \ddot{\mathbf{q}}_m^{ref}$ can be used in the internal dynamics which represent the dynamics of $\alpha$ and $\beta$ (roll and pitch angles of the platform):
\begin{align}
\mathbf{M}_{\phi \phi} \ddot{\bm{\phi}}
+ \mathbf{M}_{\phi m} \ddot{\mathbf{q}}_m^{ref} +
\mathbf{C}_{\phi \phi} \dot{\bm{\phi}}
+
\mathbf{C}_{\phi m} \dot{\mathbf{q}}_m + \mathbf{g}_\phi = \mathbf{0}.
\label{eq:mot_example_internal_dyn}
\end{align}
From this equation, note that we can access the platform roll and pitch dynamics by means of $\ddot{\mathbf{q}}_m^{ref}$. Assuming that $\mathbf{M}_{\phi m}$ is invertible\footnote{As a matter of fact, this is true, but will not be discussed in detail since this fact will not be used in the main part (Sec. \ref{sec:control}) of the paper.}, let us define $\ddot{\mathbf{q}}_m^{ref}$ as follows to achieve balancing of the platform.
\begin{align}
\label{eq:ddot_q_m_ref_mot_ex}
\ddot{\mathbf{q}}_m^{ref} = \mathbf{M}_{\phi m}^{-1} ( \mathbf{D}_\phi \dot{\bm{\phi}} + \mathbf{K}_\phi \bm{\phi} - \mathbf{C}_{\phi m}\dot{\mathbf{q}}_m - \mathbf{g}_\phi).
\end{align}
Then, the exponential convergence of $\bm{\phi}$ is trivial, but the behavior of $\mathbf{q}_m$ becomes unclear. Applying Theorem \ref{thm:cascade} once again, the closed-loop dynamics of $\mathbf{q}_m$ is
\begin{align}
\mathbf{M}_{\phi m}\ddot{\mathbf{q}}_m + \mathbf{C}_{\phi m} \dot{\mathbf{q}}_m + \mathbf{g}_\phi(\mathbf{q}_p=\mathbf{q}_p^{des}, \mathbf{q}_m, \mathbf{q}_r=\mathbf{q}_r^{des}) = \mathbf{0}.
\label{eq:motivating_ex_mass_dyn}
\end{align}
Therefore, the moving masses are likely to experience pendulum-like behavior due to the gravity force.\footnote{To be precise, this argument is not correct as $\mathbf{M}_{\phi m}$ is not always positive definite. This fact may make the actual analysis complicated.}
\begin{figure}
\centering
\subfigure[]{
\centering
\includegraphics[scale=0.35]{fig2a.png} }
\subfigure[]{
\centering
\includegraphics[scale=0.35]{fig2b.png} }
\caption{Simulation environment. (a) Initial configuration of the system. (b) Final configuration to achieve; the platform is balanced (zero roll and pitch angles); the robotic manipulator is in desired position; the moving masses are properly located to balance the static gravity of manipulator.}
\label{fig:simul_env}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.47]{fig3-eps-converted-to.pdf}
\caption{Simulation result for motivating example. Although the platform was balanced, the moving masses fell into an oscillatory motion.}
\label{fig:simul_motivating_ex}
\end{figure}
The presented control approach was simulated. Fig. \ref{fig:simul_env} shows our simulation environment which is 3D extension of Fig. \ref{fig:configuration}b. Platform was hanged on a fixed base by means of wires, and a 7-DoF robotic manipulator was mounted on the platform. Two moving masses moves in $x-$, $y-$directions on the platform. The height from the platform to the base was 10 $\mathrm{m}$. The weight of each moving mass was 10 $\mathrm{kg}$, and that of platform was also 10 $\mathrm{kg}$. Total weight of the robotic manipulator was 15 $\mathrm{kg}$. Wires were massless and no friction was considered. The following scenario was simulated:
\begin{itemize}
\item $\gamma^{des}=0$.
\item $\mathbf{q}_r^{des}=[\pi/4 \; \pi/2 \; \mathbf{0}_{1 \times 5}]^T$ with zero initial condition.
\item $\mathbf{q}_m^*=[-0.430813 \; 0.434241]$ is a position to balance the static gravity torque when $\mathbf{q}_r=\mathbf{q}_r^{des}$.
\item See Fig. \ref{fig:simul_env}a (initial configuration) and Fig. \ref{fig:simul_env}b (desired configuration to achieve).
\end{itemize}
Fig. \ref{fig:simul_motivating_ex} shows the simulation results. As expected, the oscillation damping of the platform was achieved successfully. However, the moving masses converged to a oscillatory motion, which indicates the failure of the fourth requirement of the control goal. In the following section, to achieve the control goals, $\ddot{\mathbf{q}}_m^{ref}$ will be designed to stabilize not only the output dynamics, but also the internal dynamics.
\section{Oscillation Damping Control with Stabilization of Internal Dynamics}
\label{sec:control}
The limitation of previous approach is that the internal dynamics and output dynamics cannot be asymptotically stabilized at the same time, which is not appealing because the moving mass converges to a certain limit cycle rather than staying in the equilibrium point $\mathbf{q}_m^*$. Therefore, this section proposes an approach that can asymptotically stabilize the internal dynamics as well as the output dynamics.
More specifically, we seek $\ddot{\mathbf{q}}_m^{ref}$ that stabilizes
\begin{align}
\label{eq:closed_original_phi_dyn}
\mathbf{M}_{\phi \phi}\ddot{\bm{\phi}}+\mathbf{M}_{\phi m}\ddot{\mathbf{q}}_m^{ref}+\mathbf{C}_{\phi \phi}\dot{\mathbf{q}}_p+\mathbf{C}_{\phi m}\dot{\mathbf{q}}_m+\mathbf{g}_\phi= \mathbf{0},& \\
\label{eq:closed_original_q_m_dyn}
\ddot{\mathbf{q}}_m=\ddot{\mathbf{q}}_m^{ref}&.
\end{align}
However, as can be deduced from the previous example, it might be hard to find such $\ddot{\mathbf{q}}_m^{ref}$ mainly because of the gravity $\mathbf{g}_\phi$. Even if $\bm{\phi}$ converges to zero in (\ref{eq:closed_original_phi_dyn}), $\mathbf{g}_\phi$ is not zero because of nonzero $\mathbf{q}_m$. Unfortunately, nonzero gravity will somehow appear in (\ref{eq:closed_original_q_m_dyn}) (recall (\ref{eq:motivating_ex_mass_dyn})), and excites the moving masses. Converse is also true. Even if we design $\ddot{\mathbf{q}}_m^{ref}$ so that $\mathbf{q}_m$ converges to a proper equilibrium in (\ref{eq:closed_original_q_m_dyn}), the gravity $\mathbf{g}_\phi$ is not zero due to nonzero $\bm{\phi}$. Again, nonzero gravity will excite $\bm{\phi}$ in the dynamics (\ref{eq:closed_original_phi_dyn}).
This difficulty can be overcome by expressing the dynamics (\ref{eq:whole1}) using $x,y$ components of CoM of the overall system (represented in the global frame) $\mathbf{x}_{c}$ instead of $\bm{\phi}$. This approach is appealing because the gravity force acting on $\mathbf{x}_{c}$ is zero when $\mathbf{x}_{c}=\mathbf{0}$. To this end, consider the following coordinate transformation:
\begin{align}
\dot{\bar{\mathbf{q}}} = \bm{T} \dot{\mathbf{q}},
\end{align}
where
\begin{align}
\dot{\bar{\mathbf{q}}} =
\left(
\begin{array}{c}
\dot{\mathbf{x}}_{c} \\
\dot{\gamma} \\
\dot{\mathbf{q}}_m \\
\dot{\mathbf{q}}_r
\end{array}
\right)
\end{align}
and $\bm{T}$ is the properly defined transformation matrix. Using this transformation, (\ref{eq:whole1}) can be expressed as
\begin{align}
\bar{\mathbf{M}}\ddot{\bar{\mathbf{q}}}+\bar{\mathbf{C}}\dot{\bar{\mathbf{q}}}+\bar{\mathbf{g}}=\bm{T}^{-T} \boldsymbol{\tau} = \bm{T}^{-T} {\mathbf{B}} \mathbf{u}.
\label{eq:whole2}
\end{align}
Here, noting that
\begin{align}
\bm{T}^{-T}=
\left[
\begin{array}{cc}
* & \mathbf{0} \\
* & \bm{I}
\end{array}
\right],
\end{align}
the control input $\mathbf{u}$ does not affect the CoM dynamics.
Let us now apply PFL to the output defined in (\ref{eq:y_PFL}). From
\begin{align}
\nonumber \ddot{\mathbf{y}} =& {\mathbf{B}}^{T}\ddot{q} = {\mathbf{B}}^{T} \bm{T}^{-1} (\ddot{\bar{\mathbf{q}}}-\dot{\bm{T}}\dot{\mathbf{q}}) \\
=& {\mathbf{B}}^{T} \bm{T}^{-1} \left( \bar{\mathbf{M}}^{-1}( -\bar{\mathbf{C}} \dot{\bar{\mathbf{q}}} - \bar{\mathbf{g}} + \bm{T}^{-T}{\mathbf{B}} \mathbf{u}) - \dot{\bm{T}}\dot{\mathbf{q}} \right),
\end{align}
the control input $\mathbf{u}$ is defined by
\begin{align}
\nonumber \mathbf{u} =& ({\mathbf{B}}^{T} \bm{T}^{-1} \bar{\mathbf{M}}^{-1} \bm{T}^{-T} {\mathbf{B}})^{-1} \times \\
&\left( {\mathbf{B}}^T \bm{T}^{-1} \left( \bar{\mathbf{M}}^{-1} (\bar{\mathbf{C}} \dot{\bar{\mathbf{q}}} + \bar{\mathbf{g}}) + \dot{\bm{T}}\dot{\mathbf{q}} \right) + \ddot{\mathbf{y}}^{ref} \right).
\end{align}
Furthermore, using (\ref{eq:PFL_gamma_ref})-(\ref{eq:PFL_qr_ref}), the closed-loop dynamics after convergence of $\gamma$ and $\mathbf{q}_r$ is
\begin{align}
\label{eq:closed_new_com_dyn}
\bar{\mathbf{M}}_{cc}\ddot{\mathbf{x}}_{c}+\bar{\mathbf{M}}_{c m}\ddot{\mathbf{q}}_m^{ref}+\bar{\mathbf{C}}_{cc}\dot{\mathbf{x}}_{c}+\bar{\mathbf{C}}_{c m}\dot{\mathbf{q}}_m+\bar{\mathbf{g}}_{c}=& \mathbf{0} \\
\label{eq:closed_new_q_m_dyn}
\quad \ddot{\mathbf{q}}_m=&\ddot{\mathbf{q}}_m^{ref}
\end{align}
with $\gamma \equiv \gamma^{des}$ and $\mathbf{q}_r \equiv \mathbf{q}_r^{des}$. Similar to previous, $\bar{\mathbf{M}}_{cc}$ denotes the first 2 by 2 submatrix of $\bar{\mathbf{M}}$ (the subscript `c' stands for CoM). $\bar{\mathbf{M}}_{c m}$, $\bar{\mathbf{C}}_{cc}$, $\bar{\mathbf{C}}_{c m}$, and $\bar{\mathbf{g}}_{c}$ are defined in the same manner.
The following theorem states the main result.
\begin{thm}
Define
\begin{align}
\nonumber\ddot{\mathbf{q}}_m^{ref} = \mathbf{D}\Big(&\bar{\mathbf{M}}_{cm}^T (\mathbf{D}_c \dot{\mathbf{x}}_{c} + \mathbf{K}_c \mathbf{x}_c) \\
\label{eq:qm_ref_final}
& \;\; \;- \mathbf{D}_m \dot{\mathbf{q}}_m - \mathbf{K}_m (\mathbf{q}_m - \mathbf{q}_m^*) \Big).
\end{align}
If the control gains $\mathbf{D}_{(\cdot)}$ and $\mathbf{K}_{(\cdot)}$ are chosen sufficiently large with $\mathbf{D}_c, \mathbf{K}_c \gg \mathbf{D}_m, \mathbf{K}_m$, then the closed-loop dynamics is exponentially stable to $\mathbf{x}_{c}=\mathbf{0}$, $\mathbf{q}_m=\mathbf{q}_m^*$ which implies oscillation damping because $\bm{\phi}=\mathbf{0}$.
\end{thm}
{\bf{Sketch of proof.}} The proof consists of two parts: The first part shows boundedness, and the second part shows exponential stability based on the first part. Before going into the proof, we keep in mind that the state $\mathbf{q}_m=\mathbf{q}_m^*$ and $\mathbf{x}_{c}=\mathbf{0}$ with zero derivatives is the unique equilibrium point, because $\bar{\mathbf{g}}_\phi=\mathbf{0}$ implies $\mathbf{x}_c=\mathbf{0}$.
For the first part, apply singular perturbation approach by letting $\mathbf{D}=\frac{1}{\epsilon_1} \mathbf{I}$ and $\mathbf{K}_m=\frac{1}{\epsilon_1} \mathbf{I}$ with $\epsilon_1>0$. Then $\ddot{\mathbf{q}}_m^{ref}$ itself becomes a fast variable, and converges to $\ddot{\mathbf{q}}_m^{ref}= \bar{\mathbf{M}}_{cm}^T (\mathbf{D}_c \ddot{\mathbf{x}}_{c} + \mathbf{K}_c\dot{\mathbf{x}}_c)+\dot{\bar{\mathbf{M}}}_{cm}^T (\mathbf{D}_c \dot{\mathbf{x}}_{c} + \mathbf{K}_c \mathbf{x}_c) - \mathbf{K}_m \dot{\mathbf{q}}_m$ exponential fast in a new time scale $t/\epsilon_1$ when $\epsilon_1 \rightarrow 0 $. This observation implies $\bar{\mathbf{M}}_{cm}^T (\mathbf{D}_c \dot{\mathbf{x}}_{c} + \mathbf{K}_c \mathbf{x}_c)- \mathbf{D}_m \dot{\mathbf{q}}_m - \mathbf{K}_m (\mathbf{q}_m - \mathbf{q}_m^*) =\mathbf{0}$, and another fast variable $\mathbf{K}_m (\mathbf{q}_m - \mathbf{q}_m^*)$ converges to $\bar{\mathbf{M}}_{cm}^T (\mathbf{D}_c \dot{\mathbf{x}}_{c} + \mathbf{K}_c \mathbf{x}_c)$ in the new time scale. Then, the reduced system (\ref{eq:closed_new_com_dyn}) becomes
\begin{align}
\nonumber (\bar{\mathbf{M}}_{cc}+\bar{\mathbf{M}}_{cm}\bar{\mathbf{M}}_{cm}^T \mathbf{D}_c)\ddot{\mathbf{x}}_c + \bar{\mathbf{M}}_{cm}\bar{\mathbf{M}}_{cm}^T \mathbf{K}_c \dot{\mathbf{x}}_c + \bar{\mathbf{g}}_c + \bm{\delta} = \mathbf{0},
\end{align}
where the other terms are collected in $\bm{\delta}$ of which linearization is zero around the equilibrium point. Therefore, local asymptotic stability of $\mathbf{x}_c$, $\dot{\mathbf{x}}_c$ can be shown easily. As a result, by the Tikhonov Theorem (\cite{khalil1996noninear}), the resulting trajectory is bounded around the unique equilibrium point when $\epsilon_1$ is sufficiently small.
For the second part, to apply singular perturbation theory in a different way, let $\mathbf{D}=\mathbf{I}$, $\mathbf{D}_c=\frac{1}{\epsilon_2}\mathbf{I}$, $\mathbf{K}_c=\frac{1}{\epsilon_2}\mathbf{I}$ with $\epsilon_1 \gg \epsilon_2$ to avoid conflict with the first part. Then, (\ref{eq:closed_new_q_m_dyn}) becomes fast dynamics with the fast variable $\bm{z}=\frac{1}{\epsilon^2}\mathbf{x}_c$. Since $\mathbf{g}_c$ vanishes with $\mathbf{x}_c=0$, the fast variable converges to $\bm{z}=(\bar{\mathbf{M}}_{cm}\bar{\mathbf{M}}_{cm}^T)^{-1}\big(\bar{\mathbf{C}}_{cm}\dot{\mathbf{q}} +\bar{\mathbf{M}}_{cm} (\mathbf{D}_m \dot{\mathbf{q}}_m + \mathbf{K}_m (\mathbf{q}_m - \mathbf{q}_m^*)) \big)$ and $\dot{\bm{z}}=\mathbf{0}$. Therefore (\ref{eq:closed_new_com_dyn}) becomes $\bar{\mathbf{M}}_{cm}\ddot{\mathbf{q}}_m+\bar{\mathbf{C}}_{cm}\dot{\mathbf{q}}_m =\mathbf{0}$ (note that $\bar{\mathbf{C}}_{mc}\dot{\mathbf{x}}_c=\mathbf{0}$) which is slow dynamics. Therefore, $\mathbf{q}_m$ will either constantly drift away or stay in a certain equilibrium point. Since the boundedness is already guaranteed from the first part, $\mathbf{q}_m$ can only stay in the unique equilibrium point $\mathbf{q}_m=\mathbf{q}_m^*$, which can be interpreted as exponential stability. Hence, the original system (\ref{eq:closed_new_com_dyn})-(\ref{eq:closed_new_q_m_dyn}) is exponentially stable with sufficiently small $\epsilon_1$, $\epsilon_2$ satisfying $\epsilon_1 \gg \epsilon_2$. \QEDA
In the following remarks, to further motivate (\ref{eq:qm_ref_final}), we introduce a couple of variations which may seem plausible at a glance, but do not achieve control goals.
\begin{rem}
Consider
\begin{align}
\label{eq:qm_ref_only_com}
\ddot{\mathbf{q}}_m^{ref} = \bar{\mathbf{M}}_{cm}^T (\mathbf{D}_c \dot{\mathbf{x}}_{c} + \mathbf{K}_c \mathbf{x}_c),
\end{align}
which does not contain $\mathbf{q}_m$-related terms. Obviously, this choice will result in $\mathbf{x}_c \rightarrow \mathbf{0}$. However, $\mathbf{x}_c=\mathbf{0}$ does not always imply $\bm{\phi}=\mathbf{0}$ because $\mathbf{x}_c$ is the CoM represented in the global reference frame. This variation will be discussed in more detail with simulation results in next section.
\label{rem:ext_1}
\end{rem}
\begin{rem}
One may want to extend (\ref{eq:ddot_q_m_ref_mot_ex}) to have PD term of $\mathbf{q}_m$ similar to (\ref{eq:qm_ref_final}). However, correct extension is not trivial as naive extension lead to non-unique equilibrium point. For example, consider
\begin{align}
\nonumber \ddot{\mathbf{q}}_m^{ref} = & \mathbf{M}_{\phi m}^{-1} ( \mathbf{D}_\phi \dot{\bm{\phi}} + \mathbf{K}_\phi \bm{\phi} - \mathbf{C}_{\phi m}\dot{\mathbf{q}}_m - \mathbf{g}_\phi) \\
\label{eq:qddot_m_ref_variation}
& - \mathbf{D}_{m} \dot{\mathbf{q}}_m - \mathbf{K}_m (\mathbf{q}_m - \mathbf{q}_m^{*}).
\end{align}
Equilibrium point is given by solving
\begin{align}
\mathbf{M}_{\phi m}^{-1} ( \mathbf{D}_\phi \dot{\bm{\phi}} +\mathbf{K}_\phi \bm{\phi})+\mathbf{K}_m (\mathbf{q}_m - \mathbf{q}_m^{*}) =& \mathbf{0},\\
\mathbf{g}_\phi=&\mathbf{0}.
\end{align}
Unfortunately, there is no unique solution for this set of equations. This variation will be discussed in more detail with simulation results in next section.
\label{rem:ext_2}
\end{rem}
\section{Simulation}
\label{sec:simulation}
In this section, after showing the simulation results for the variations introduced in Remarks \ref{rem:ext_1} and \ref{rem:ext_2}, the results for the proposed approach is shown.
\begin{figure}
\centering
\includegraphics[scale=0.47]{fig4-eps-converted-to.pdf}
\caption{Simulation result with variation in Remark \ref{rem:ext_1}. Although $\mathbf{x}_c$ converged to $\mathbf{0}$, $\mathbf{q}_m$ and $\mathbf{q}_p$ diverged out.}
\label{fig:variation_com}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.47]{fig5-eps-converted-to.pdf}
\caption{Simulation result with variation in Remark \ref{rem:ext_2}. Although $\mathbf{q}_m$ is included in $\ddot{\mathbf{q}}_m^{ref}$ design (\ref{eq:qddot_m_ref_variation}), $\mathbf{q}_m$ and $\mathbf{q}_p$ did not converge to the desired equilibrium point.}
\label{fig:variation_from_orig}
\end{figure}
\subsection{Simulation result for the variation in Remark \ref{rem:ext_1}}
In Remark \ref{rem:ext_1}, only $\mathbf{x}_c$ is considered in $\ddot{\mathbf{q}}_m^{ref}$ design. By doing so, $\mathbf{x}_c$ converged to zero, as shown in the second row of Fig. \ref{fig:variation_com}. However, as pointed out in Remark \ref{rem:ext_1}, $\mathbf{x}_c=\mathbf{0}$ does not imply $\bm{\phi}=\mathbf{0}$ (and also $\mathbf{q}_m=\mathbf{q}_m^*$). As shown in the third and fourth row of Fig. \ref{fig:variation_com}, $\mathbf{q}_p$ and $\mathbf{q}_m$ can diverge out while maintaining $\mathbf{x}_c=\mathbf{0}$. Note that $\mathbf{q}_r=\mathbf{q}_r^{des}$ and $\gamma=\gamma^{des}$ were still achieved because feedback linearization with pole-placement (\ref{eq:PFL_gamma_ref})-(\ref{eq:PFL_qr_ref}) was applied to these variables.
\subsection{Simulation result for the variation in Remark \ref{rem:ext_2}}
In Remark \ref{rem:ext_2}, both $\mathbf{q}_\phi$ and $\mathbf{q}_m$ were considered in $\ddot{\mathbf{q}}_m^{ref}$ design. This design may seem plausible as it tries to stabilize $\mathbf{q}_\phi$ as well as $\mathbf{q}_m$ at the same time. However, as pointed out in Remark \ref{rem:ext_2}, equilibrium point is not unique. As shown in Fig. \ref{fig:variation_from_orig}, $\mathbf{q}_m$ and $\mathbf{q}_\phi$ oscillated around the desired point, but could not converge to it.
\subsection{Simulation with the proposed approach}
Fig. \ref{fig:main_res} shows simulation results with the proposed $\ddot{\mathbf{q}}_m^{ref}$ in (\ref{eq:qm_ref_final}). As expected, $\mathbf{x}_c=\mathbf{0}$ and $\mathbf{q}_m=\mathbf{q}_m^*$ could be achieved. Note that $\mathbf{x}_c=\mathbf{0}$ and $\mathbf{q}_m=\mathbf{q}_m^*$ imply $\bm{\phi}=\mathbf{0}$, as shown in the fourth row of Fig. \ref{fig:main_res}. As discussed previously, $\mathbf{q}_r$ and $\gamma$ converged and stayed in the desired equilibrium because feedback linearzation with pole-placement technique was applied to these variables.
\section{Conclusion}
\label{sec:conclusion}
In this paper, a pendulum-like hanging manipulation platform is presented to overcome limitation of the aerial manipulation system. The main challenge in the control point of view is the oscillation damping of the platform. Since there is no actuation for the platform orientation (i.e., under-actuated), it can be controlled only indirectly. To alleviate this problem, moving masses were installed on top of the hanging platform. This paper tries to damp out the platform oscillation by defining the reference acceleration of the moving masses properly. In particular, it is shown that the extension of PFL can be used to asymptotically stabilize not only output dynamics (robotic manipulator and yaw angle of the platform) but also the internal dynamics (platform oscillation and moving mass). The proposed approach was validated by simulation studies.
\begin{figure}
\label{fig:main_res}
\centering
\includegraphics[scale=0.47]{fig6-eps-converted-to.pdf}
\caption{Simulation result with proposed approach. Both $\mathbf{x}_c$ and $\mathbf{q}_m$ converged to the desired equilibrium point. Note that this result indicates successful oscillation damping of the platform as shown in the fourth row.}
\end{figure}
|
\section{Introduction}
Customized grippers have been broadly applied in industry to execute complex tasks such as assembly and packaging. Compared with general parallel grippers, customized grippers usually consist of curved fingers and have large contact surfaces to match with the geometries of the objects. Therefore, the customized grippers can generate more stable and robust grasps compared with those generated by general parallel grippers.
However, the grasp planning for customized grippers is challenging. On one hand, it has potentially large contact areas with workpieces, thus the traditional point contact model assumption~\cite{murray1994mathematical} and the quality metrics~\cite{ferrari1992planning, li1988task} built upon the point contact model are not applicable. As a result, the performance of the grasp planning algorithms based on those metrics~\cite{mahler2016dex, hang2014hierarchical} would be downgraded.
On the other hand, the grasp policies learned from end-to-end manner~\cite{levine2016learning, pinto2016supersizing} use general parallel grippers with millions of grasping data, and changing fingertips require re-collecting data and re-training the policies. With the tend of mass customization and requirement to change fingertips frequently, the end-to-end learning becomes less efficient.
There are a few works that consider the surface information during grasp planning. In~\cite{ciocarlie2007dexterous}, the optimal grasp is searched by enclosing the object with more contacts to match the object surface. However, it has a heavy computation load, thus is not suitable for online implementation. The grasp synthesis of human hands with shape matching is introduced in~\cite{li2007data}. However, this approach requires offline human demonstration and exhaustive search in the database.
In~\cite{klingbeil2011grasping,ten2018using}, the curve of the parallel gripper is matched to the surface of objects to speed up the searching process. However, it cannot be applied to the precise matching for customized grippers with complicated shapes.
In this paper, we propose a learning framework to plan grasps for customized grippers. To consider more complicated gripper shapes and retrieve reliable and secure grasps for different objects, it is desired to match the surfaces on the gripper to the object more precisely.
We use an iterative surface fitting (ISF) algorithm proposed in our previous work~\cite{fan2018grasp} to fit the contact surface of multiple fingers to the object surfaces while satisfying the structural constraints. A guided sampling is also introduced to avoid the local optima and encourage the exploration of the regions with high fitting scores.
ISF achieves efficient and precise grasp planning for a single object or slight clutter environment with small searching regions. However, the searching becomes less efficient in heavy clutter environment due to the excessively sampling candidates~\cite{fan2018grasp}.
To improve the efficiency of grasp planning in a heavy clutter environment, we propose a topological learning approach using regions with convolutional neural networks (R-CNN)~\cite{girshick2014rich} to learn the essential features that affect the successful execution of grasps.
The features include the spatial relationships between objects which are generally difficult to model. Thus the input to the CNN is the patches of the candidate regions. This learning-based planner is connected hierarchically with ISF in order to reduce the effect of object variations and plan reliable/precise grasps under data shortage.
Therefore, the learning framework includes an optimization-based planner, which searches for precise grasp pose based on the fine details of the selected region using ISF, and a learning-based explorer, which learns the desired regions to start ISF searching.
Compared with end-to-end learning methods~\cite{mahler2016dex, levine2016learning}, the proposed learning-based explorer ignores the details of grasp planning by detecting the desired regions within the image plane for potentially high fitting scores and better collision avoidance performance,
thus the dimension of the learning module is lower than end-to-end learning. The optimization-based planner searches for optimal grasps precisely in the chosen region based on the object-specific features, which are generally not shared across objects and difficult to learn by end-to-end manner.
Therefore, the proposed learning framework is able to improve the learning efficiency and performance at the same time.
The contributions of this paper are as follows. First, the proposed low-level optimization-based planner includes both the fingertip surfaces and the structural constraints of the gripper such as the jaw width and the allowable degree-of-freedoms (DOFs). It also achieves simultaneous surface fitting and gripper kinematic planning.
Second, by combining with the dedicated gripper design, the proposed surface fitting algorithm can deal with objects with complicated shapes as well as a clutter task with unsegmented point clouds.
Furthermore, the proposed grasp exploration strategy avoids getting stuck in the local optima by learning from the previous grasping experience. The grasp planning by ISF and R-CNN achieves an efficient planning and the time to search for a collision-free grasp is 1.5 secs for the objects in heavy clutter environment. The experimental videos are available at~\cite{website}.
The remainder of this paper is described as follows. The grasp planning problem for customized gripper is stated in Section~\ref{sec:problem_statement}, followed by the introduction of the proposed learning framework in Section~\ref{sec:learning_framework}. The experimental verification of the proposed framework on grasp planning with customized grippers are presented in Section~\ref{sec:experimental_results}.
Section~\ref{sec:conclusions} concludes the paper and proposes future works. The experimental videos are available at~\cite{website}.
\section{Problem Statement}
\label{sec:problem_statement}
The objective of the grasp planning is to search for the optimal gripper pose and finger configuration by maximizing a quality metric considering the constraints of the customized grippers. A grasp planning example is illustrated in Fig.~\ref{fig:gripper_object}. The customized gripper has parallel jaws with customized curved fingertip surfaces. Taking the gripper in Fig.~\ref{fig:gripper_object} as an example, the grasp planning problem is defined as
\begin{subequations}
\label{eq:general_form}
\begin{align}
\max_{R, t, \delta d, \mathcal{S}_j^f, \mathcal{S}_j^o} &\ Q(\mathcal{S}_1^f, \mathcal{S}_2^f,\mathcal{S}_1^o, \mathcal{S}_2^o) \label{eq1:cost}\\
s.t. \quad
& \mathcal{S}_j^f \subset \mathcal{T}(\partial \mathcal{F}_j;R,t,\delta d), \quad j = 1,2 \label{eq1:surface_finger}\\
& \mathcal{S}_j^o = NN_{\partial \mathcal{O}} (\mathcal{S}_j^f), \quad j = 1,2 \label{eq1:surface_object}\\
& (\mathcal{S}_1^f, \mathcal{S}_2^f) \in \mathcal{W}(d_0 + \delta d) \label{eq1:constraint1}\\
& d_0 + \delta d \in [d_{\text{min}}, d_{\text{max}}] \label{eq1:constraint2}
\end{align}
\end{subequations}
where $j \in \{1,2\}$ is the finger index, $R \in SO(3), t\in \mathbb{R}^3$ are the rotation and translation of the gripper jaw, $\delta d\in \mathbb{R}$ is the jaw displacement from the original width $d_0$, and $Q$ is the quality metric in terms of the finger contact surfaces $\mathcal{S}_j^f$ and the object contact surface $\mathcal{S}_j^o$. The finger contact surface $\mathcal{S}_j^f$ lies on the finger surface $\partial \mathcal{F}_j$ transformed by $\mathcal{T}$ with amount of $R,t$, as shown in~(\ref{eq1:surface_finger}). The object contact surface $\mathcal{S}_j^o$ is defined by searching the nearest neighbor of the $\mathcal{S}_j^f$ on the object surface $\partial \mathcal{O}$, as shown in~(\ref{eq1:surface_object}).
\begin{figure}[t]
\begin{center}
{\includegraphics[width =0.8\linewidth]{gripper_object.jpg}}
\caption{An grasp example by a customized gripper with curved fingertip surfaces. A natural quality metric is the surface fitting error between the gripper and the object, where the blue object in (a) has more steady grasp compared with the orange object in (b).}
\label{fig:gripper_object}
\end{center}
\end{figure}
Constraint~(\ref{eq1:constraint1}) indicates that the finger contact surfaces should be constrained in the space $\mathcal{W}$ parameterized by the jaw width, and (\ref{eq1:constraint2}) describes the finger displacement constraint.
Problem~(\ref{eq:general_form}) would be a standard grasp planning if the contact surface degenerates into a single contact point. In general, however, the point contact model may not be able to directly include gripper surface into the planning. Problem~(\ref{eq:general_form}) is challenging to solve by either learning or optimization. On one hand, the learning requires training on objects with large variety. On the other hand, the optimization with gradient or sampling can either be nontrivial to use contact surfaces as decision variables, or require to search the whole state space, which is not appropriate for real-time implementation.
A natural surface-related quality can be constructed by matching the surfaces of the gripper towards the object, as shown in Fig.~\ref{fig:gripper_object}. Intuitively, the grasp with small surface fitting error~(Fig.~\ref{fig:gripper_object}(a)) is more reliable and robust compared with the one with large surface fitting error~(Fig.~\ref{fig:gripper_object}(b)). With the surface fitting error as the quality metric, the problem can be addressed by point set registration algorithm, such as the rigid registration (e.g. ICP~\cite{besl1992method}) or non-rigid registration (e.g. CPD~\cite{myronenko2010point} or TPS~\cite{bookstein1989principal}). The non-rigid registration allows arbitrary deformation of the point set without considering the structural constraints (Constraint~\ref{eq1:constraint1}~\ref{eq1:constraint2}), while the rigid registration assumes rigid transformation of the point set without considering the allowable motion between different portions of the points.
Both the methods tend to be trapped into local optima during the searching.
\section{The Learning Framework}
\label{sec:learning_framework}
\subsection{Overview}
In this paper, we assume that the customized gripper has rigid fingertip surfaces and allows motion in certain DOFs. Therefore, the proposed searching algorithm is modified from the rigid registration by including the structural constraints of different fingers such as the width and DOFs. To avoid being trapped into local optima and achieve better collision avoidance performance, we propose a learning-based explorer to detect the desired regions to initialize the low-level search.
Figure~\ref{fig:learning_framework} shows the overall learning framework.
The proposed learning framework decouples the end-to-end learning into a low-level optimization-based planner and a high-level learning-based explorer. The optimization-based planner searches for optimal gripper pose and finger configuration with the iterative surface fitting~(ISF) we proposed in~\cite{fan2018grasp}. The learning-based explorer detects the desired region to explore from the previous grasping experiences with the regions with R-CNN. Compared with end-to-end learning methods, the optimization-based planner is more precise and reliable when grasping on various unknown objects, and the learning-based explorer is more efficient with much less training data and lower learning dimension.
\begin{figure}[t]
\begin{center}
{\includegraphics[width=3.4in]{learning_framework.png}}
\caption{Block diagram of the overall learning framework with RCNN.}
\label{fig:learning_framework}
\end{center}
\end{figure}
\subsection{Optimization-based Planner}
\label{sec:optimization_based_planner}
The optimization-based planner searches for the optimal gripper pose and finger configuration within some particular regions on the object surface. The block diagram of the optimization-based planner is shown in Fig.~\ref{fig:framework}. It contains an iterative surface fitting (ISF) and a guided sampling. ISF is to register multiple fingertip surfaces to the target workpiece considering the allowable motions of fingers, and the guided sampling is to guide the ISF to search in different portions within the region to reduce the effect of the local optima.
\begin{figure}[t]
\begin{center}
{\includegraphics[width =3.4in]{framework.png}}
\caption{Block diagram of the low-level optimization-based planner. }
\label{fig:framework}
\end{center}
\end{figure}
\subsubsection{Iterative Surface Fitting}
\begin{figure}[bt]
\begin{center}
{\includegraphics[width =1\linewidth]{gripper_isf.png}}
\caption{(Left) Customized gripper used in the paper, and (Right) illustration of the iterative surface fitting (ISF). (a) The correspondence matching, where the corresponding point $q_{i_j}$ on the object surface is found by searching with the nearest neighbor and removing outliers and duplicates. (b) The surface fitting, in which the gripper transformation $(R,t)$ and finger displacement $\delta d$ are optimized. The green gripper is updated by the optimization result, and the pink one is the converged result after several iterations of ISF.}
\label{fig:isf}
\end{center}
\end{figure}
The algorithm is illustrated by a customized gripper with two fingers and one DOF, where the two fingers move in opposite directions to adjust the jaw width, as shown in Fig.~\ref{fig:isf} (Left). ISF iteratively executes two modules: the correspondence matching and the surface fitting. The correspondence matching contains the nearest neighbor search and outlier/duplicate filtering, as shown in Fig.~\ref{fig:isf} (Right)(a).
The surface fitting searches over the gripper transformation $(R,t)$ and the jaw width $d = d_0 + \delta d$, as shown in Fig.~\ref{fig:isf} (Right)(b). More specifically, the surface fitting is to minimize the surface fitting error $E$:
\begin{subequations}
\label{eq3:overall}
\begin{align}
\min_{R, t, \delta d} &\ E(R,t,\delta d) \label{eq3:cost}\\
s.t. \quad
& \delta d + d_0 \in [d_\text{min}, d_\text{max}] \label{eq3:surface_finger}
\end{align}
\end{subequations}
Problem (\ref{eq3:overall}) is nonlinear due to the coupling between the $R$ and $\delta d$ and can be solved by an iterative palm-finger optimization (IPFO)~\cite{fan2018grasp}.
The ISF algorithm is summarized in Alg.~\ref{alg:isf}. Inspired by~\cite{jost2003multi}, ISF is optimized hierarchically by searching with a multi-resolution pyramid. The inputs to ISF include the initial gripper state $R_c, t_c, d_0$, the surfaces $\partial \mathcal{O}$ and $\partial \mathcal{F}$. $L, I_0$ and $\epsilon_0$ denote the level number, maximum iteration and error bound for convergence, respectively. The gripper surface $\partial F$ is first transformed to the specified initial state~(Line~\ref{isf:init}). In each level of pyramid, the $\partial \mathcal{F}$ is downsampled adaptively to $\mathcal{S}^f$ with different resolutions (Line~\ref{isf:down}). The while loop iteratively searches correspondence and solves for desired palm transformation and finger displacement. The while loop is terminated if IPFO gives a similar transformation in adjacent iterations.
\begin{algorithm}[tb]
\caption{Iterative Surface Fitting (ISF)}\label{alg:isf}
\begin{algorithmic}[1]
\State \textbf{Input:} Initial state $R_c,t_c, d_0$, $\partial \mathcal{O}$, $\partial \mathcal{F}$, $L, I_0, \epsilon_0$ \label{isf:input}
\State \textbf{Init:} $\partial \mathcal{F} = \mathcal{T}(\partial \mathcal{F}; R_c, t_c, d_0)$ \label{isf:init}
\For {$l = L-1, \cdots, 0$} \label{isf:paraymid}
\State $\mathcal{S}^f\leftarrow \texttt{downsample}(\partial \mathcal{F}, 2^l)$, $I_l = I_0/2^l$, $\epsilon_l= 2^l\epsilon_0$\label{isf:down}
\State $\mathcal{S}_{0}^f \leftarrow \mathcal{S}^f$, $e_{s} \leftarrow \infty$, $\eta \leftarrow 0$
\While {$\eta \notin [1 - \epsilon_l, 1 + \epsilon_l] $ and $it\texttt{++} \leq I_l$}
\State $e_{s,p} \leftarrow e_s$
\State $\mathcal{S}^o\leftarrow NN_{\partial O}(\mathcal{S}^f)$ \label{isf:nn}
\State $\{\mathcal{S}^f, \mathcal{S}^o\} \leftarrow \texttt{filter}(\mathcal{S}^f, \mathcal{S}^o)$ \label{isf:filter}
\State $\{R^*, t^*, \delta d^*, error\} \leftarrow \mathbf{IPFO}(\mathcal{S}^f, \mathcal{S}^o, d_0)$ \label{isf:IPFO}
\State $\mathcal{S}^f \leftarrow \mathcal{T}(\mathcal{S}^f; R^*, t^*, \delta d^*)$ \label{isf:updateS}
\State $\partial \mathcal{F} \leftarrow \mathcal{T}(\partial \mathcal{F}; R^*, t^*, \delta d^*)$ \label{isf:update}
\State $d_0 \leftarrow d_0 + \delta d^*$
\State $e_s \leftarrow \|\mathcal{S}^f - \mathcal{S}_{0}^f\|, \ \eta \leftarrow e_s/e_{s,p}$
\EndWhile
\EndFor \label{isf:paraymid2}
\State \Return $\{ error, \partial \mathcal{F}\}$
\end{algorithmic}
\end{algorithm}
\subsubsection{Baseline Initialization and Guided Sampling}
\label{sec:guided_sampling}
The initialization of ISF is important since ISF converges to local optima. Various initial points are desired for ISF to explore different portions of the region, so as to avoid getting trapped in bad local optima and achieve better collision-avoiding solutions.
In the baseline initialization, the region is partitioned into $K$ clusters by k-means clustering. The center of each cluster is regarded as a candidate initial position of the gripper for ISF. The ISF with this baseline initialization is called baseline-ISF in the remaining of the paper.
We build an empirical model to guide the sampling among $K$ candidates. Similar to the multi-armed bandit model, we record the fitting error, collision of ISF, and the reachability for each cluster and compute the average regret accordingly. The weight for ISF to be initialized in the $k$-th center is decreased if this center has larger regret value, and the cluster center with the minimum regret is chosen as the initial position of the gripper for the following ISF, while the initial orientation $R_c$ is randomly sampled.
Figure~\ref{fig:with_wo_collision} shows the baseline-ISF result on an Oscar model. The object and gripper surfaces are fed into the algorithm as shown by the blue and red dots in Fig.~\ref{fig:with_wo_collision}(a), after which the k-means clustering ran for centers of initialization, as shown by bold dots.
Multiple grasps were generated (red patches in~Fig.~\ref{fig:with_wo_collision}(a)) and passed through the collision check function. The planned collided grasp is shown with transparency, as shown in Fig.~\ref{fig:with_wo_collision}(b). The centers with small surface fitting error and no collision would have small regret values, thus will be sampled more frequently.
\begin{figure}[t]
\begin{center}
{\includegraphics[width =1\linewidth]{with_wo_collision_png.png}}
\caption{ (a) Illustration of the grasp planning on an Oscar model, where the blue and red dots represent the object and the gripper surfaces, and the bold dots represent the centers of k-means. (b) The visualization of planned grasps, where the collided grasp is represented by the transparent one.}
\label{fig:with_wo_collision}
\end{center}
\end{figure}
The optimization-based planner (baseline-ISF) with the k-means clustering and guided sampling is able to search the optimal collision-free grasps efficiently in simple environment with small regions. However, the searching would be less efficient and sub-optimal in clutter environment where an excessively large number of clusters are required to be searched. Moreover, while the guided sampling distinguishes different centers of clusters by taking into account of the fitting error, collision and reachability, this allocation only explores around the predefined $K$ positions without considering other portions of the region. Finally, the guided sampling can only exploit the experience of the current grasp searching in the current environment, since the previous grasping is conducted in different environment with different distribution of centers and regrets.
\subsection{Learning-based Explorer}
\label{sec:Learning-based Explorer}
The optimization-based planner solved by baseline-ISF is inefficient and sub-optimal in clutter environment. Meanwhile, we found that human tends to decouple the process of choosing the desired grasp region from that of searching specific grasp poses inside that region to increase the efficiency of the grasping. With this observation, we design a learning-based explorer to initialize the baseline-ISF search. The explorer selects the regions with potential low regret based on the previous grasp experiences. We use R-CNN to learn a classifier in order to detect the desired regions for initialization.
\subsubsection{R-CNN Pipeline}
The pipeline of R-CNN is shown in Fig.~\ref{fig:rcnn}. R-CNN is first introduced in~\cite{girshick2014rich} for object detection. R-CNN contains a region proposal block to provide possible choices of regions. The region proposal selects 2K regions with different sizes using the method such as selective search. The regions are resized and fed into a CNN for feature extraction. The CNN can be pre-trained by AlexNet~\cite{Krizhevsky2012ImageNet} or VGG-16/19~\cite{simonyan2014very}. The outputs of CNN are used to represent the features of the region proposals. SVMs are applied to classify the regions and bounding box regression is applied to further correct the positions of the bounding boxes.
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{rcnn_pipeline.png}}
\caption{Illustration of R-CNN pipeline. A region proposal block proposes regions and sends to CNN to extract features. SVMs and bounding box regression are applied to classify the proposed region and refine the bounding boxes, respectively. }
\label{fig:rcnn}
\end{figure}
\subsubsection{R-CNN Training}
We use a grasping pool with 25 objects of different categories. In industrial bin picking applications, this number is usually sufficiently large.
We randomly choose some of the objects and place them in the workspace. The scene is observed by two stereo cameras and the collected point cloud is rendered as depth images, as shown in Fig.~\ref{fig:depthImgRendering}.
R-CNN takes the rendered depth images as inputs, and produces regions of interest (ROI) to initialize the ISF searching. The ROI used for training is generated based on the optimal grasps found from baseline-ISF. The training process is illustrated by~Fig.~\ref{fig:training_framework}.
The R-CNN in this paper is pre-trained by AlexNet and is fine-tuned by the data we collected. In this stage, we use 250 data pairs with data augmentation, producing 2000 data pairs to fine tune the network. Some of the data pairs are illustrated in Fig.~\ref{fig:training_data}.
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{DepthImgRendering.png}}
\caption{The depth rendering. (a) Original scene. (b) Point cloud observed by two stereo cameras. (c) Rendered depth images. The closer of the points to the camera, the more white they are in the depth image. (d) The image is further rendered into jet colormap. }
\label{fig:depthImgRendering}
\end{figure}
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{training_framework.png}}
\caption{Illustration of the training framework. The framework contains the baseline-ISF and the R-CNN training. }
\label{fig:training_framework}
\end{figure}
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{training_data.png}}
\caption{The data used to fine-tune the R-CNN for good region detection.}
\label{fig:training_data}
\end{figure}
\subsubsection{R-CNN Testing}
At the test time, the trained R-CNN is applied to generate the ROI on the colormap. The desired regions in the point cloud are then computed based on the box coordinates of ROI in the image plane as well as their depth values.
The centers of these regions are regarded as the good initialization and ISF searches from these initial positions.
The proposed learning ISF is called RCNN-ISF and the framework is illustrated in Fig.~\ref{fig:test_framework}. The proposed learning framework with RCNN-ISF not only searches grasps locally within a small region, but also learns the large-scale grasp exploration using R-CNN. Therefore, it tends to provide better initialization more efficiently than the baseline-ISF.
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{test_framework.png}}
\caption{The learning framework with RCNN-ISF implementation. The R-CNN detects desired low regret regions and ISF searches grasps on the selected regions.}
\label{fig:test_framework}
\end{figure}
Compared with end-to-end learning~\cite{mahler2016dex,levine2016learning}, the proposed RCNN-ISF method has the following advantages. First, the learning dimension of the RCNN-ISF is generally lower than that of the end-to-end learning. More specifically, RCNN-ISF searches ROI in the two-dimensional image plane, while the end-to-end learning searches over higher dimension depending on the grasps and grippers. For example, a grasp planning for a eight-DOF hand with three fingers has 32 dimensions~\cite{fan2018real}.
Therefore, the end-to-end learning requires much more data than the proposed method.
Secondly, ISF searches for optimal grasps based on object-specific features that are not shared cross objects, instead of learning the behavior end-to-end from millions of data. Consequently, ISF tends to generate more precise and robust grasps.
Moreover, RCNN-ISF is able to produce versatile grasp as the experiment shows. On the contrary, the learned networks in~\cite{mahler2016dex, levine2016learning} produce planar grasps with simple parallel grippers.
\section{Experimental Results}
\label{sec:experimental_results}
This section presents the experimental results for baseline-ISF and RCNN-ISF to verify the effectiveness of the learning framework. The experimental videos are available at~\cite{website}. The host computer we used was a desktop with 32GB RAM and 4.0GHz CPU. All the computations were conducted by Matlab.
We used a SMC LEHF20K2-48-R36N3D parallel jaw gripper with the specialized curved fingers as shown in Fig.~\ref{fig:isf}~(Left). The desired contact surfaces are marked by red. The gripper was equipped on a FANUC LRMate 200iD/7L industrial manipulator. Two IDS Ensenso N35 stereo camera sets were used to capture the point cloud of the object. The point cloud was smoothed and the normal vectors were calculated. Despite the preprocessing, the point cloud produced by Ensenso cameras was not able to reflect the object precisely due to occlusion and noise.
\subsection{Parameter Lists}
The gripper width was constrained by $\left[d_\text{min},d_\text{max} \right] = \left[1,3\right]$ cm. The initial gripper width was set as $d_0$ = 2 cm. The pyramid level $L$, the maximum iteration $I_0$ and the tolerance $\epsilon_0$ were set as $4$, $200$ and $0.008$ respectively in Alg.~\ref{alg:isf}. The k-means center number $K = 6$.
\subsection{Baseline-ISF Experiment}
\label{sec:baseline-isf}
This section presents the experimental results of the optimization-based planner with baseline-ISF. The baseline-ISF algorithm considers the surface fitting around different initial positions specified by k-means. The guided sampling enables more effective exploration based on the grasp experience of the current environment.
Figure~\ref{fig:graspExp1} shows the grasp planning results on six different objects. The left and right sides of each subfigure show the grasp optimized by ISF and the result in lifting the object by 10 cm, respectively. Baseline-ISF was able to find a grasp that matched the fingertips to the object in spite of the complicated shapes of the surface. The algorithm was able to handle the missing points properly. For example, the gripper fingertip was able to locate to the bottom of the link (Fig.~\ref{fig:graspExp1}(b)), despite the points were missing since both cameras were looking from the top.
The geometry similarity between the gripper and the object is beneficial for the gripper to resist the effect of calibration uncertainty and external disturbance, and increase the stability and the payload, as it increases the potential contact surfaces between the gripper and the object.
In average, it takes less than 1.5 secs to find 10 collision-free optimal grasps.
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{GraspExp1_png.png}}
\caption{The results of the grasp planning experiment on six objects.}
\label{fig:graspExp1}
\end{figure}
\subsection{RCNN-ISF Experiment}
\label{res:rcnn_exp}
This section presents the experimental results of the proposed learning framework with RCNN-ISF. The experimental videos are available at~\cite{website}. The learning framework includes the low-level optimization-based planner with baseline-ISF, and the high-level learning-based explorer with R-CNN for grasp exploration in heavy clutter environment. The training process of R-CNN is shown in Fig.~\ref{fig:training_data}. Same experimental setup was used as Section~\ref{sec:baseline-isf}.
Figure~\ref{fig:graspExp2} illustrates the performance of the RCNN-ISF in light clutter environment. In this environment, RCNN-ISF achieved comparable performance with the baseline-ISF.
The initial configuration of the object set is shown in Fig.~\ref{fig:graspExp2}(a). Figure~\ref{fig:graspExp2}(b)-(f) show the consecutive grasps in the task. The left side of each subfigure is the depth image and the R-CNN output of the confidence map for the desired regions, after which the baseline-ISF was performed on the chosen regions, and the best grasp was executed by the FANUC industrial manipulator, as shown in the right side of each subfigure.
Even though the surface composed by the cluttered objects became more complicated than a single object, RCNN-ISF was able to successfully detect the desired regions for grasping and search on the selected regions for the optimal collision-free grasp.
\begin{figure}[t]
\centering
{\includegraphics[width =1\linewidth]{ClutterExp_RCNN.png}}
\caption{Grasp planning experiment in a clutter environment with the proposed learning framework. (a) The initial object clutter. (b)-(f) The consecutive grasps in the task.}
\label{fig:graspExp2}
\end{figure}
Figure~\ref{fig:GraspExp_RCNN} shows the grasp planning results on heavy clutter environments by RCNN-ISF. The first row shows different clutter environments. R-CNN took rendered depth images as inputs and generated the desired regions to initialize ISF searching, as shown in Fig.~\ref{fig:GraspExp_RCNN} (Middle). The optimization-based planner produced optimal grasp pose by minimizing the fitting error and checking the collision/feasibility. The optimal grasps were executed by the FANUC manipulator, as shown in Fig.~\ref{fig:GraspExp_RCNN} (Bottom).
\begin{figure}
\centering
\includegraphics[width = 1\linewidth]{GraspExp_RCNN_edit.png}
\caption{Grasp planning results on four different heavy clutter environments by RCNN-ISF. (Up) Clutter environments. (Middle) Selected regions and their confidence values by R-CNN to initialize ISF searching. (Bottom) Execution of the grasps planned by ISF. }
\label{fig:GraspExp_RCNN}
\end{figure}
The comparison between the baseline-ISF and RCNN-ISF is shown in Table~\ref{tab:comparison} for 4 heavy clutter environments in Fig.~\ref{fig:GraspExp_RCNN}.
The searching on heavy clutter environments are more difficult due to the collision with environments and the excessive search space for guided sampling. Consequently, the initialization becomes more important for efficient exploration. The baseline-ISF requires to initialize and execute ISF multiple times in order to explore broader regions. On the contrary, RCNN-ISF employs the previous experience for initialization and generates only a few low-regret regions to start ISF searching. Therefore, RCNN-ISF is able to perform grasp planning more efficiently.
In heavy clutter environment, the baseline-ISF spent 17.23 secs to find 1 optimal grasp, while the proposed RCNN-ISF spent 1.52 secs to find 7 optimal grasps.
\begin{table}[t]
\centering
\caption{Comparison of Baseline-ISF and RCNN-ISF}
\label{tab:comparison}
\begin{tabular}{r|cc}
Methods & Baseline-ISF & RCNN-ISF \\
\hline
Search Time (s) & 17.23 & 1.52 \\
\hline
Found Grasps\# & 1 & 7 \\
\hline
\end{tabular}
\end{table}
\section{Conclusions and Future Works}
\label{sec:conclusions}
This paper proposed a learning framework to plan robust grasps for customized grippers. The learning framework includes a low-level optimization-based planner and a high-level learning-based explorer.
The optimization-based planner used an iterative surface fitting (ISF) with guided sampling to search for optimal grasps by minimizing the surface fitting error. The performance of this low-level planner was locally effective and was sensitive to initialization. Therefore, the learning-based explorer was introduced with a region-based convolutional neural network (R-CNN) to search for desired low-regret regions to initialize ISF search. A series of experiments on robotic bin picking were performed to evaluate the proposed method. Experimental results show that the proposed learning framework with RCNN-ISF achieved a more efficient planning on heavy clutter environment, by significantly decreasing the average searching time from 17.23 secs to 1.52 secs.
There are several directions that we want to explore in the future. First, we would like to apply the algorithm to the customized grippers with more complicated shapes. In addition, we would like to employ the domain randomization~\cite{tobin2017domain} and train the learning-based explorer sufficiently and efficiently.
\addtolength{\textheight}{-1cm}
\bibliographystyle{IEEEtran}
|
\section{Introduction and Related Work}
\label{sec:intro and related work}
Human head detection in crowded scenes is an important problem in the field of computer vision because of its wide range of applications. One such use-case could be head detection based crowd counting where detection based approach tends to give more reliable results compared to previous density map based crowd counting techniques \cite{zhang2015cross, zhang2016single, sam2017switching}. This is because, in case of density maps, it is not always the correct location which contributes to final crowd count. This leads to unreliable results especially in case of false positives. Several approaches for head detection has been proposed previously which are used in different contexts \cite{mae2005head, zhang2013multi, merad2010fast, vu2015context, stewart2016end, marin2014detecting}. \cite{merad2010fast} proposed human body segmentation approach in order to localize heads in a scene. \cite{subburaman2012counting} trained a sophisticated classifier based on the cascade of boosted integral features. In order to perform head detection based analytics, Marin et. al. \cite{marin2014detecting} firstly localized upper part of the human body and then trained a DPM \cite{felzenszwalb2010object} in order to detect head in the localized upper body part. The purpose of this head detection strategy was to find out how two characters are interacting in a movie scene. Recent years have witnessed various deep learning based head detectors \cite{vu2015context, stewart2016end}. One of the best performing head detection model \cite{vu2015context} extends RCNN detector for head detection using two types of contextual cues. First cue leverages the person-scene relations through a global CNN model. Second cue models the pairwise relations among objects using a structured output. Both of these cues are combined in a joint RCNN framework for head detection. Both the approaches for head detection i.e. \cite{marin2014detecting} and \cite{vu2015context} are designed in the context of movies where the average head size is large whereas average head count is small. These models when deployed for head detection in crowded scenes might fail to generalize because of the complete opposite scenario for head detection i.e. small average head size and large average head count. This is because no special care is taken to incorporate large scale and density variations in the previous models. This makes the entire task of head detection in crowded scenes a challenging one to be solved. Our work closely relates to \cite{stewart2016end} where they propose an anchor-free detection framework for head detection in crowded scenes which uses RNN in order to exploit information from deep representations in order to make predictions.
In this work, we propose an architecture for head detection especially for detection in crowded scenes. Our model is based on general anchor based detection pipeline like Faster-RCNN \cite{ren2017faster}, where we have a set of pre-set anchors which are generated by regularly tiling a collection of boxes with different scales in an image. However unlike a two-stage pipeline of Faster-RCNN (Region Proposal Network + ROI Pooling and Classification), in this architecture, we have only a single fully-convolutional network that can perform both classification and bounding box prediction. In this paper, we make the following contributions: 1) A fully convolutional, end-to-end trainable head detection model with systematic anchor selection strategy based on effective receptive field of the network which helps to achieve good results in crowded scenes. 2) Our model has very low memory requirements and inference time which makes it suitable for edge deployments. 3) We demonstrate that our model achieves comparable results to many other architectures as well as outperforming some of the previous approaches on challenging head detection datasets.
\section{Method}
\label{sec: method}
\subsection{Model architecture}
Deep architectures like \cite{krizhevsky2012imagenet, szegedy2015going, simonyan2014very} are capable of learning ``generalized features", which are useful in a variety of tasks. We start off by leveraging one of such already existing pre-trained VGG16 \cite{simonyan2014very}, as our base model for our architecture. We remove the final layers succeeding the conv5 layer of VGG16, and use the remaining weights as a starting point for our new training. The architecture is shown in Fig. \ref{fig: overall pipeline}. This pre-trained base model is appended with $3$ new convolutional layers: (1) The first conv layer (conv6) slides over the entire feature map in order to make detections at all locations. (2) The second conv layer (conv\textsubscript{reg}) is the regression head responsible for predicting localization coordinates. (3) The third conv layer (conv\textsubscript{cls}) is the classification head which predicts a probability score of a head. The conv\textsubscript{reg} and conv\textsubscript{cls} layers are implemented using a $1\times1$ convolutional kernel. The outputs from the conv\textsubscript{reg} undergo bounding box transformation, where the scale and shift predictions are converted to spatial coordinates. Number of conv\textsubscript{reg} and conv\textsubscript{cls} kernels depends on number of anchors at each pixel location $(N)$.
\begin{figure*}[!t]
\centering
\includegraphics[scale=0.25]{g6904.png}
\caption{Architecture of our head detection model. A set of pre-defined anchors are generated and a fully convolutional network is used to estimate the coordinates from the anchors alongwith the probability score. There are two heads to the model, 1) conv\textsubscript{reg}, $40\times30\times(N\times4)$. 2) conv\textsubscript{cls}, $40\times30\times(N\times2)$. Each $1\times1\times(N\times4)$ output from the conv\textsubscript{reg} will indicate the scale and shift of $N$ anchors at that location in the feature map. Here, $N=2$ which is the number of anchors.}
\label{fig: overall pipeline}
\end{figure*}
\subsection{Design of Anchor Scales}
Anchors are obtained by placing a pre-defined scale of boxes at regular intervals in an image. It is with respect to these anchors, the scales and shifts are predicted in an object recognition system \cite{ren2017faster}. We select the anchor scales based on the ``effective receptive field" rather than the ``theoretical receptive field". As pointed out by \cite{luo2016understanding}, a unit in a CNN has two types of receptive field, one is a theoretical receptive field and the other is the effective receptive field. Only a few center pixels of a theoretical receptive field is responsible for triggering the neurons in the final output. This center region of the theoretical receptive field is the effective receptive field. In the conv5 layer of VGG16, the theoretical receptive field size is $228$ pixels. Taking note of this, we prefer anchors of scales $2$ and $4$ which leads to anchor sizes $32\times32$ and $64\times64$ respectively. Anchor size is computed as $layer\_stride \times aspect\_ratio \times anchor\_scale$ where the $layer\_stride$ of the conv5 layer is $16$. Anchor sizes are approximately scaled down by $3.5$ times the theoretical receptive field which will match the effective receptive field size. Designing anchor scales in this manner helps to achieve good detection results especially in crowded scenes where the head size is small. This can be concluded from our experiments.
\subsection{Training}
\subsubsection{Approach}
Using our designed set of anchor scales, a set of predefined anchors are generated for each image after the initial preprocessing stage. In the preprocessing, we resize the image to $640\times480$ resolution and then do mean subtraction and normalization of the image. The anchors are tiled at a stride of $16$ because of $4$ max-pooling layers prior to conv5 layer. We select two anchors/pixel of size $32\times32$ and $64\times64$. Because of $4$ max-pooling layers prior to conv5 layer the resolution of the feature map obtained at the conv5 layer is of size $40\times30$ (as input resolution of the image is $640\times480$). Thus, for each image we obtain $2400$ anchors $(40\times30\times2)$. For each anchor, the network will predict $4$ regression coordinates through conv\textsubscript{reg} and $2$ probability scores through conv\textsubscript{cls}. Regression coordinates are in the form of scales and shifts the anchors need to undergo in order to localize an actual head in the scene. We do not consider the anchors going out of the boundary of the image for training. In order to assign labels to the anchors for training, we follow these $2$ strategies: (1) The anchor that have IoU $\geq0.7$ with ground-truth is labeled positive. (2) The anchors with maximum IoU with the ground-truth is assigned a positive label. We follow both strategies since in some cases the first condition might fail to assign a positive label to any of the anchors. Negative labels are assigned to those anchors which have IoU $\leq0.3$. The anchors that do not satisfy both the conditions are don't care and are not included in the training. We process a single image at a time during training. And from the anchors with assigned labels, we pick a total of $32$ positive and negative anchors in a ratio of $1:1$.
\subsubsection{Loss Function}
The loss function used for the training of the model is a multi-task loss function, similar to that defined in training of an RPN by \cite{ren2017faster}. It is shown in Eq. \ref{eq:loss_function}.
\begin{equation}
L(\{p_{i}\}, \{t_{i}\}) = \dfrac{1}{N_{cls}}\sum_{i}L_{cls}(p_{i},p_{i}^{*}) + \dfrac{1}{N_{reg}}\sum_{i}p_{i}^{*}L_{reg}(t_{i},t_{i}^{*})
\label{eq:loss_function}
\end{equation}
where $i$ is the index of chosen anchor which will range through $32$ selected anchors. $p_{i}$ is the predicted probability that an anchor $i$ contains head and $p_{i}^{*}$ is the ground-truth label i.e. $1$ or $0$. $t_{i}$ is the parameterized coordinates of the predicted bounding box (i.e. scale and shift), and $t_{i}^{*}$ is the parameterized coordinates of the ground truth. This parameterized coordinates are defined in the same manner as by \cite{girshick2015fast}. $L_{cls}$ denotes the classification loss, whereas $L_{reg}$ denotes the regression loss. $L_{cls}$ is computed over all the anchors whereas $L_{reg}$ is computed over only positive anchors and this is taken care by $p_{i}^{*}$ in front of $L_{reg}$ in the equation. $L_{cls}$ is the cross entropy loss. Whereas $L_{reg}$ is the smooth $L_{1}$ loss as defined by \cite{girshick2015fast}. Both the loss terms are normalized by $N_{cls}$ and $N_{reg}$ which are the number of samples accounted for classification and regression respectively.
\subsubsection{Hyperparameters}
The base model is initialized with pre-trained VGG16, which is trained using ImageNet dataset \cite{russakovsky2015imagenet}. All the layers of the pre-trained model along with new layers are retrained. The new layers are initialized with random weights sampled from a standard normal distribution with $\mu = 0$ and $\sigma = 0.01$. Weight decay is set to $0.0005$. The entire model is fine-tuned using SGD. Learning rate for training is set to $0.001$, and the model is trained for $15$ epochs ($\sim 160k$ iterations). We decay the learning rate with a scale of $0.1$ after completing $8$ epochs. Whole training was performed in PyTorch framework \cite{paszke2017automatic}.
\begin{table}[]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Dataset} & \textbf{Scenario} & \textbf{Method} & \textbf{AP} \\ \hline
\multirow{5}{*}{Brainwash} & \multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Crowded \\ scenes\\ (7.89 avg. \\ head count)\end{tabular}} & Overfeat - AlexNet & 0.62 \\ \cline{3-4}
& & ReInspect, $L_{fix}$ & 0.60 \\ \cline{3-4}
& & ReInspect, $L_{firstk}$ & 0.63 \\ \cline{3-4}
& & ReInspect, $L_{hungarian}$ & 0.78 \\ \cline{3-4}
& & \textbf{Ours} & \textbf{0.70} \\ \hline
\multirow{5}{*}{Hollywood} & \multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Movie \\ scenes\\ (1.65 avg. \\ head count)\end{tabular}} & DPM Face & 0.37 \\ \cline{3-4}
& & R-CNN & 0.67 \\ \cline{3-4}
& & Context - Local & 0.71 \\ \cline{3-4}
& & \begin{tabular}[c]{@{}c@{}}Context - Local+\\ Global + Pairwise\end{tabular} & 0.72 \\ \cline{3-4}
& & \textbf{Ours} & \textbf{0.74} \\ \hline
\end{tabular}
\end{center}
\caption{Evaluation of our model across various datasets.}
\label{tab: average_precision}
\end{table}
\section{Experiments}
\label{sec: experiments}
In order to have a thorough evaluation of our model, we benchmark our results on two publicly available datasets mentioned in Table \ref{tab: average_precision}. We train the model using the train set of the respective dataset (Brainwash-10461, Hollywood-216694) and then evaluate the model on the test set. Both the datasets are captured in different scenarios. Brainwash dataset\cite{stewart2016end} is captured in crowded scenes where the average head count is large ($7.89$) whereas HollywoodHeads dataset is collected from movie scenes where an average head count is small ($1.65$). Following settings are kept constant throughout the experiment. Anchor scales are set to $2$ and $4$. Non Maximum suppression (NMS) is performed during the test phase and not in the training phase. NMS is a kind of post-processing technique to filter out highly overlapping detections. In this technique, any two detections are suppressed to one if the IoU among them is greater than a threshold. The threshold for NMS is set to $0.3$. Following evaluation metrics is used, i.e. head is considered to be correctly detected if the IoU of the predicted bounding box with the ground truth is $\geq0.5$ \cite{everingham2015pascal}. In order to quantify the performance of our model, we plot the precision-recall curve along with computing average precision (AP). We then compare the performance of other baselines with our technique using this precision-recall curve and average precision table (Table \ref{tab: average_precision}).
\subsubsection{Results on Brainwash Dataset}
We test our model using the test set of the Brainwash dataset which contains $484$ images. We consider same baselines for comparisons which were considered by \cite{stewart2016end} i.e. OverFeat with AlexNet architecture \cite{sermanet2013overfeat} and the current state-of-the-art model called ReInspect \cite{stewart2016end}. There are $3$ variants of ReInspect, which uses various types of losses i.e. $L_{fix}$, $L_{first}$, $L_{hungarian}$. Comparison with all of these variants is performed. Fig. \ref{fig: precision_recall}(a) shows the precision-recall curve on the test set. As can be seen from the figure, the precision-recall obtained by our technique is better than $3$ of the baselines i.e. Overfeat-AlexNet, ReInspect, $L_{fix}$ and ReInspect, $L_{firstk}$. ReInspect, $L_{hungarian}$ is the best performing model among all at EER (Equal Error Rate). Despite this, our model gives the best performance at a little lower recall i.e. best precision of $0.92$ at a recall of $0.65$ which can be observed from the precision-recall curve. AP attained by our model is $0.70$ which is better than $3$ baselines and comparable to best performing model (i.e. $0.78$). This can be seen in Table \ref{tab: average_precision}. Fig. \ref{fig: success cases} demonstrates some of the successful detection cases of our model under high occlusion and crowd density scenarios. Fig. \ref{fig: failure cases} shows some of the failure cases of our model.
\begin{figure}
\centering
\includegraphics[scale=.34]{precision_recall_comb.png}
\caption{Precision-Recall for Brainwash and HollywoodHeads dataset respectively.}
\label{fig: precision_recall}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=.5]{fig_result_1.png}
\caption{Successful detection (green arrows) under challenging cases like high occlusion and high crowd density.}
\label{fig: success cases}
\end{figure}
\subsubsection{Results on HollywoodHeads Dataset}
In order to experiment with the HollywoodHeads dataset we train a completely fresh model using the train set. After training, the model is then tested on the test set which contains $1297$ images. We compare our approach with techniques used for comparison in \cite{vu2015context} i.e. DPM based head detector \cite{mathias2014face}, RCNN based head detector \cite{girshick2014rich}, and two variants of same technique i.e. 1) Local context model (Context-Local) and 2) Combined Local$+$Global$+$Pairwise context model (Context-L$+$G$+$P) as proposed in \cite{vu2015context}. Precision-recall curve is shown in Fig. \ref{fig: precision_recall}(b). Our model gives best performance compared to all the baselines. We achieve an AP of $0.74$ which is $\sim2\%$ higher compared to the state-of-the-art as shown in Table \ref{tab: average_precision}.
\begin{figure}
\centering
\includegraphics[scale=.55]{fig_result_2.png}
\caption{Failure cases of our model. Gives false detections in scenarios where there is no or less heads to be detected.}
\label{fig: failure cases}
\end{figure}
\begin{table}[]
\centering
\begin{tabular}{|c|c|c|}
\hline
\textbf{Platform} & \textbf{Model} & \textbf{Speed} \\ \hline
\multirow{2}{*}{Quadro M1000M} & ReInspect & 1fps \\ \cline{2-3}
& FCHD & 5fps \\ \hline
\multirow{2}{*}{Jetson TX2} & ReInspect & NA \\ \cline{2-3}
& FCHD & 1.6fps \\ \hline
\end{tabular}
\caption{Run-time comparison at VGA resolution}
\label{tab: timings}
\end{table}
\subsubsection{Timings}
In this section, we compare the time taken by our model for inference on different hardware platforms with previous approaches. Timings of our model are benchmarked on two platforms. The first platform is NVidia Quadro M1000M GPU which has 512 CUDA cores. The second platform is NVidia Jetson TX2 which has $256$ CUDA cores which is the most suitable device for edge deployments. We record the inference time of all the images in the test set and then the final timing is computed as the average of inference time of all the images in the test set. The timings are shown in Table. \ref{tab: timings}. As can be seen from the table that our model runs at 5 FPS which is $5\times$ faster than ReInspect on the same platform. Moreover, our model runs at 1.6 FPS on Jetson TX2 embedded platform where the other model even fails to load because of memory requirements. This is a major advantage of our model as along with being accurate, it is also suitable for edge deployments.
\begin{table}[]
\begin{center}
\begin{tabular}{|c|c|}
\hline
\textbf{Anchor Size} & \textbf{AP} \\ \hline
$32\times32$, $64\times64$ & 0.70 \\ \hline
$64\times64$, $128\times128$ & 0.53 \\ \hline
$128\times128$, $256\times256$ & 0.45 \\ \hline
\end{tabular}
\end{center}
\caption{Effect of anchor size on AP}
\label{tab: anchor_scales}
\end{table}
\subsubsection{Ablation Experiments}
To make sure that the anchor size we choose for our experiments as well as for deployment are optimal, we do an ablation study where we train $3$ different models with varying anchor sizes i.e. (1) $32\times32$, $64\times64$ (default setting), (2) $64\times64$, $128\times128$, (3) $128\times128$, $256\times256$. These models are trained with the same hyperparameter settings. As it can be seen from Table \ref{tab: anchor_scales} that best AP is obtained when the anchor sizes are $32\times32$ and $64\times64$ which justifies our claim made in previous sections that, choosing anchor scales based on the effective receptive field gives better detection results.
\section{Conclusion}
\label{sec: conclusion}
In this work, we propose FCHD, an anchor based fully convolutional end-to-end trainable head detection model for crowded scenes. Our model achieves $0.70$ AP on challenging head detection dataset like Brainwash, which is comparable to previous techniques. Along with being accurate, it has very less run-time i.e. 5 FPS on Nvidia Quadro and 1.6 FPS on Jetson TX2. In the future, we plan to improve upon this work by getting better overall accuracy especially in cases where the head sizes are small. This could be achieved by combining the detections from various convolutional layers.
\bibliographystyle{IEEEbib}
|
\section{Introduction and statement of the results}
Oscillating multipliers on $\mathbb{R}^{n}$ are bounded functions of the
type
\begin{equation*}
m_{\alpha ,\beta }(\xi )=\left\Vert \xi \right\Vert ^{-\beta }e^{i\left\Vert
\xi \right\Vert ^{\alpha }}\theta \left( \xi \right) ,
\end{equation*
where $\alpha ,\beta >0$ and $\theta \in C_{0}^{\infty }\left( \mathbb{R
\right) $ which vanishes near zero, and equals to $1$ outside the ball
B\left( 0,2\right) $. Let $T_{\alpha ,\beta }$ be the operator which in the
Fourier transform variables is given by
\begin{equation*}
\widehat{(T_{\alpha ,\beta }f)}\left( \xi \right) =m_{\alpha ,\beta }(\xi
\hat{f}\left( \xi \right) \text{, \ }f\in C_{0}^{\infty }\left( \mathbb{R
^{n}\right) ,
\end{equation*
i.e. $T_{\alpha ,\beta }$ is a convolution operator with kernel the inverse
Fourier transform of $m_{\alpha ,\beta }$. The $L^{p}$-boundedness of
T_{\alpha ,\beta }$ on $\mathbb{R}^{n}$ is extensively studied. See for
example \cite{HIR,WAI,STE,FE,SCHO} for $\alpha \in \left( 0,1\right) $ and
\cite{PERAL} for $\alpha =1$.
The $L^{p}$-boundedness of oscillating multipliers has been studied also in
various geometric contexts as Riemannian manifolds, Lie groups and symmetric
spaces. See for example \cite{GIUME,ALEX,MAR,GEORG, LO} and the references
therein.
In the present work we deal with oscillating multipliers on rank one locally
symmetric spaces. To state our results, we need to introduce some notation
(for details see Section 2). Let $G$ be a semi-simple, non-compact,
connected Lie group with finite center and let $K$ be a maximal compact
subgroup of $G$. We consider the symmetric space of non-compact type $X=G/K
. Let $G=KAN$ be the Iwasawa decomposition of $G$. If $A\cong \mathbb{R},$
we say that $X$ has rank one. Recall that rank one symmetric spaces are the
real, complex, and quaternionic hyperbolic spaces, denoted $\mathbb{H}^{n}
\mathbb{R})$, $\mathbb{H}^{n}(\mathbb{C})$ and $\mathbb{H}^{n}(\mathbb{H})$,
$n\geq 2$, and the octonionic hyperbolic plane $\mathbb{H}^{2}(\mathbb{O})$.
Throughout this paper we shall assume that $\dim X=n\geq 2$ and rank$X=1$.
In \cite{GIUME}, Giulini and Meda consider the multiplier
\begin{equation*}
\ m_{\alpha,
\beta}(\lambda)=(\lambda^2+\rho^2)^{-\beta/2}e^{i(\lambda^2+\rho^2)^{
\alpha/2}}, \alpha>0, \operatorname{Re}{\beta} \geq 0, \lambda >0,
\end{equation*}
where $\rho $ is the half sum of positive roots, counted with their
multiplicities. As in the case of $\mathbb{R}^n$, we denote by $T_{\alpha,
\beta}$ the convolution operator with kernel $\kappa_{\alpha ,\beta }$, the
inverse spherical Fourier transform of $m_{\alpha ,\beta}$ in the sense of
distributions. Then,
\begin{equation}
T_{\alpha ,\beta }(f)(x)=\int_{G}\kappa _{\alpha ,\beta
}(xy^{-1})f(y)dy,\quad f\in C_{0}^{\infty }(X). \label{operatorX1}
\end{equation}
Note that
\begin{equation}
T_{\alpha ,\beta }=\Delta _{X}^{-\beta /2}e^{i\Delta _{X}^{\alpha /2}},\
\alpha >0,\ {\operatorname{Re}}{\beta }\geq 0, \label{om1}
\end{equation}
where $\Delta_{X}$ is the Laplace-Beltrami operator on $X$. In \cite{GIUME},
the $L^{p}$- boundedness of $T_{\alpha ,\beta }$ is investigated on rank one
symmetric spaces and the following theorem is proved:
\begin{theorem}[Giulini, Meda]
\label{GM}If $p\in (1,\infty )$, then
(i) If $\alpha <1$, then $T_{\alpha ,\beta }$ is bounded on $L^{p}(X)$,
provided that $\operatorname{Re}{\beta }>\alpha n\left\vert 1/p-1/2\right\vert $,
(ii) if $\alpha =1$, then $T_{\alpha ,\beta }$ is bounded on $L^{p}(X)$,
provided that $\operatorname{Re}{\beta }>\left( n-1\right) \left\vert
1/p-1/2\right\vert $,
(iii) if $\alpha >1$, then $T_{\alpha ,\beta }$ is bounded on $L^{p}(X)$ if
and only if $p=2$.
\end{theorem}
As it is noticed in \cite{GIUME}, the results above for the case $\alpha
\leq 1$ are less precise than in the Euclidean case, since it is not known
what happens at the critical indices $\operatorname{Re}{\beta }=\alpha n\left\vert
1/p-1/2\right\vert $ for $\alpha \in \left( 0,1\right) $ and $\operatorname{Re}
\beta }=\left( n-1\right) \left\vert 1/p-1/2\right\vert $ for $\alpha =1$.
Let us now present the case of locally symmetric spaces. Let $\Gamma $ be a
discrete and torsion free subgroup of $G$ and let us consider the locally
symmetric space $M=\Gamma \backslash X=\Gamma \backslash G/K.$ Then $M$,
equipped with the projection of the canonical Riemannian structure of $X$,
becomes a Riemannian manifold.
To define oscillating multipliers on $M$, we first observe that if $f\in
C_{0}^{\infty }(M)$, then, the function $T_{\alpha ,\beta }f$ defined by
\ref{operatorX1}), is right $K$-invariant and left $\Gamma $-invariant. So
T_{\alpha ,\beta }$ can be considered as an operator acting on functions on
M$, which we shall denote by $\widehat{T}_{\alpha ,\beta }$. Note that the
Laplace-Beltrami operator $\Delta _{M}$ is the projection of $\Delta _{X}$.
So from (\ref{om1}), it follows that
\begin{equation}
\widehat{T}_{\alpha ,\beta }f=\Delta _{M}^{-\beta /2}e^{i\Delta _{M}^{\alpha
/2}}f,\quad f\in C_{0}^{\infty }(M). \label{om2}
\end{equation
In this paper we deal with the $L^{p}$-boundedness of $\widehat{T}_{\alpha
,\beta }$ and we prove the analogue of Theorem \ref{GM}. We treat the cases
\alpha \in \left( 0,1\right) $, $\alpha =1$ and $\alpha >1$, separately.
\bigskip
\textbf{Case 1}\textit{. }$\alpha \in \left( 0,1\right) $. The main
ingredient for the proof of our results is the Kunze and Stein phenomenon on
locally symmetric spaces, proved in \cite{LOMAjga}, and which states that
there exist $\eta _{\Gamma }\in \mathfrak{a}^{\ast }$ and $s\left(
p\right)>0 $, $p\in \left( 1,\infty \right) $, such that for every $K$-bi-invariant function $\kappa $, the convolution operator $\ast \left\vert
\kappa \right\vert $ with kernel $\left\vert \kappa \right\vert $ satisfies
the estimate
\begin{equation}
\left\Vert \ast \left\vert \kappa \right\vert \right\Vert
_{L^{p}(M)\rightarrow L^{p}(M)}\leq \int_{G}\left\vert \kappa \left(
g\right) \right\vert \varphi _{-i\eta _{\Gamma }}\left( g\right) ^{s\left(
p\right) }dg, \label{ks00}
\end{equation}
where $\varphi _{\lambda }$ is the spherical function with index $\lambda $.
For more details see Section 2.
We say that $M=\Gamma \backslash X$ belongs in the class $(KS)$ if the Kunze
and Stein phenomenon is valid on it. The class $(KS)$ is described in detail
in \cite[Section 1]{LOMAjga}. We note that $M\in (KS)$ for all discrete
groups $\Gamma $, if $X=\mathbb{H}^{n}(\mathbb{H}),\mathbb{H}^{2}(\mathbb{O
) $, while if $X=\mathbb{H}^{n}(\mathbb{R}),\mathbb{H}^{n}(\mathbb{C})$,
then $M\in (KS)$, provided that $\Gamma $ is amenable \cite{FOMAMA, LOMAjga}.
We have the following theorem:
\begin{theorem}
\label{alphaless1}If $M$ belongs in the class $(KS)$, then for $\alpha \in
\left( 0,1\right) $, $\widehat{T}_{\alpha ,\beta }$ is bounded on $L^{p}(M)
, provided that $\operatorname{Re}{\beta >}\alpha n|1/p-1/2|.$
\end{theorem}
\textbf{Case 2}\textit{. }$\alpha =1$. This case is of particular interest
since by (\ref{om2}), $\widehat{T}_{1,\beta }=\Delta _{M}^{-\beta
/2}e^{i\Delta _{M}^{1/2}}$ and thus $\widehat{T}_{1,\beta }$ is related to
the wave operator.
The $L^{p}$-boundedness of the operator $T_{1,\beta }$ is investigated in
\cite{PERAL} for the case of $\mathbb{R}^{n}$, in \cite{LO} in a very
general geometric context including Riemannian manifolds of bounded geometry
and Lie groups of polynomial or exponential growth, and in \cite{GIUME} for
rank one symmetric spaces.
Denote by $\delta (\Gamma )$ the critical exponent of the group $\Gamma $:
\begin{equation}
\delta (\Gamma )=\inf \left\{ s>0:P_{s}(x,y)<+\infty \right\} ,
\label{criticalexp}
\end{equation
where
\begin{equation}
P_{s}(x,y)=\sum_{\gamma \in \Gamma }e^{-sd(x,\gamma y)} \label{poinc}
\end{equation
is the Poincar\'{e} series. Note that $\delta (\Gamma )\in \lbrack 0,2\rho ]
.
We say that $\Gamma $ is of convergence type according as the series
P_{\delta (\Gamma )}(x,y)$ converges and of divergence type otherwise.
We say that $\Gamma $ belongs in the class $(CT)$ if $\delta (\Gamma )=2\rho
$ and $\Gamma $ is of convergence type. For example, if $\Gamma \subset SO(n,1)$,
then $\Gamma \in (CT)$ if it is of the second kind, i.e. the limit set
$\Lambda (\Gamma )$ is not equal to the whole of
$\partial \mathbb{H}^{n}(\mathbb{R})$, \cite[Theorem 1.6.2]{NI}
In \cite{RO}, Roblin gives
a criterion of convergence for the case of rank one locally symmetric
spaces. Recall that a limit point $\xi$ belong in the conical limit set $\Lambda _{c}\left( \Gamma \right) $
if there is some geodesic ray $\beta_{t } $ tending to $\xi $ as $t\rightarrow \infty $
and a constant $c>0$ such that the orbit $\Gamma x_{0}$, $x_{0}\in X$, accumulates to $\xi $
within the closed $c$-neighborhood of $\beta_{t} $. In \cite[Theorem 1.7]{RO}
it is proved that $\Gamma$ is of convergence type iff
\mu _{y_{0}}\left( \Lambda _{c}\left( \Gamma \right) \right) =0$, where
$\mu _{y_{0}}$, $y_{0} \in X$ is a $\Gamma$-invariant Patterson-Sullivan density.
For more details for the class $(CT)$, see Section 2.
\begin{theorem}
\label{alphaeq1} If either $\delta (\Gamma )<2\rho $ or $\Gamma \in (CT)$,
then $\widehat{T}_{1,\beta }$ is bounded on $L^{p}(M)$, $p\in (1,\infty )$,
provided that $\operatorname{Re}{\beta >}(n-1)|1/p-1/2|$.
\end{theorem}
\textbf{Case 3}\textit{. }$\alpha >1$. As it is shown in Section 3,
\begin{equation*}
\widehat{T}_{\alpha ,\beta }=\frac{1}{\Gamma \left( \beta /\alpha \right)
\int_{0}^{\infty }\sigma ^{\left( \beta /\alpha \right) -1}e^{\left(
i-\sigma \right) \Delta _{M}^{\alpha /2}}d\sigma .
\end{equation*
But, by the spectral theorem
\begin{equation*}
\left\Vert e^{\left( i-\sigma \right) \Delta _{M}^{\alpha /2}}\right\Vert
_{L^{2}(M)\rightarrow L^{2}(M)}\leq e^{-\sigma \rho ^{\alpha }}.
\end{equation*
So,
\begin{equation*}
\left\Vert \widehat{T}_{\alpha ,\beta }\right\Vert _{L^{2}(M)\rightarrow
L^{2}(M)}\leq c,
\end{equation*
and this is the only we can say for the $L^{p}$-boundedness of $\widehat{T
_{\alpha ,\beta }$ if $\alpha >1$. On the contrary, in \cite{GIUME} Giulini
and Meda observe that the multiplier $m_{\alpha ,\beta }$ is not bounded on
any strip $S_{\varepsilon }=\left\{ |\operatorname{Im}\lambda |<\varepsilon \rho
\right\} $, $\varepsilon \leq 1$, and consequently by the necessary part of
the multiplier theorem of Clerc and Stein \cite{CLERC}, they conclude that
T_{\alpha ,\beta }$ is bounded on $L^{p}(X)$ \textit{iff} $p=2$.
The paper is organized as follows. In Section 2 we present the necessary
tools we need for our proofs, and in Section 3 we give the proofs of our
results.
\section{Preliminaries}
In this section we recall some basic facts about symmetric spaces and
locally symmetric spaces we will use for the proof of our results. For
details see \cite{AN, HEL, FOMAMA, LOMAjga}.
Let $G$ be a semisimple Lie group, connected, noncompact, with finite center
and let $K$ be a maximal compact subgroup of $G$ and consider the symmetric
space $X=G/K$. In the sequel we assume that $\dim X=n$. Denote by $\mathfrak
g}$ and $\mathfrak{k}$ the Lie algebras of $G$ and $K$. Let also $\mathfrak{
}$ be the subspace of $\mathfrak{g}$ which is orthogonal to $\mathfrak{k}$
with respect to the Killing form. The Killing form induces a $K$-invariant
scalar product on $\mathfrak{p}$ and hence a $G$-invariant metric on $G/K$.
Denote by $\Delta _{X}$ the Laplace-Beltrami operator on $X$, by $d(.,.)$
the Riemannian distance and by $dx$ the associated measure on $X$. Let
\Gamma $ be a discrete torsion free subgroup of $G$. Then, the locally
symmetric space $M=\Gamma \backslash X$, equipped with the projection of the
canonical Riemannian structure of $X$, becomes a Riemannian manifold. We
denote by $\Delta _{M}$ the Laplacian on $M$.
Fix $\mathfrak{a}$ a maximal abelian subspace of $\mathfrak{p}$ and denote
by $\mathfrak{a}^{\ast }$ the real dual of $\mathfrak{a}$. Let $A$ be the
analytic subgroup of $G$ with Lie algebra $\mathfrak{a}.$ Let $\mathfrak{a}
^{+}\subset \mathfrak{a}$ and let $\overline{\mathfrak{a}^{+}}$ be its
closure. Put $A^{+}=\exp \mathfrak{a}^{+}$. Its closure in $G$ is $\overline{
A^{+}}=\exp \overline{\mathfrak{a}^{+}}$. We have the Cartan decomposition
\begin{equation*}
\ G=K(\overline{A^{+}})K=K(\exp \overline{\mathfrak{a}^{+}})K.
\end{equation*}
We say that $a$ is a root vector if for every $H\in \mathfrak{a}$
\begin{equation*}
\ [H,X]=a(H)X,\;X\in \mathfrak{g}.
\end{equation*}
Denote by $\rho $ the half sum of positive roots, counted with their
multiplicities. Denote by $x_{0}=eK$ the origin of $X$. If $x,y\in X$, then
there are isometries $g,h\in G$ such that $x=gx_{0}$ and $y=hx_{0}$. Then,
\begin{equation}
d(x,y)=d(gx_{0},hx_{0})=d(x_{0},g^{-1}hx_{0}). \label{d(x,y)}
\end{equation}
But, by the Cartan decomposition
\begin{equation}
g^{-1}h=k\left( \exp H(g^{-1}h)\right) k^{\prime },\quad k,k^{\prime }\in
K,\;H(g^{-1}h)\in \overline{\mathfrak{a}_{+}}. \label{kak}
\end{equation}
In the rank one case, we have that
\begin{equation*}
\ A^{+}=\left\{ \exp H:H\in \overline{\mathfrak{a}_{+}}\right\} =\left\{
\exp tH_{0}:t>0\right\} ,
\end{equation*}
where $H_{0}\in \overline{\mathfrak{a}_{+}}$ with $\Vert H_{0}\Vert =1$.
From (\ref{d(x,y)}) and (\ref{kak}) it follows that
\begin{equation}
d(x,y)=\Vert H\Vert =t\Vert H_{0}\Vert =t. \label{tgamma}
\end{equation}
\subsection{The spherical Fourier transform and the Kunze and Stein
phenomenon}
Denote by $S(K\backslash G/K)$ the Schwartz space of $K$-bi-invariant
functions on $G$. The spherical Fourier transform $\mathcal{H}$ is defined
by
\begin{equation*}
\ \mathcal{H}f(\lambda )=\int_{G}f(x)\varphi _{\lambda }(x)\;dx,\quad
\lambda \in \mathfrak{a^{\ast }},\quad f\in S(K\backslash G/K),
\end{equation*}
where $\varphi _{\lambda }$ are the elementary spherical functions on $G$.
Let $S(\mathfrak{a^{\ast }})$ be the usual Schwartz space on $\mathfrak{\
a^{\ast }}$, and let us denote by $S(\mathfrak{a^{\ast }})^{W}$ the subspace
of $W$-invariants in $S(\mathfrak{a^{\ast }})$, where $W$ is the Weyl group
associated to the root system of $(\mathfrak{g},\mathfrak{a})$. Then, by a
celebrated theorem of Harish-Chandra, $\mathcal{H}$ is an isomorphism
between $S(K\backslash G/K)$ and $S(\mathfrak{a^{\ast }})$ and its inverse
is given by
\begin{equation*}
\ (\mathcal{H}^{-1}f)(x)=c\int_{\mathfrak{a\ast }}f(\lambda )\varphi
_{-\lambda }(x)\frac{d\lambda }{|\mathbf{c}(\lambda )|^{2}},\quad x\in
G,\quad f\in S(\mathfrak{a^{\ast }})^{W},
\end{equation*
where $\mathbf{c}(\lambda )$ is the Harish-Chandra function.
In \cite{LOMAjga} it is proved the following analogue \cite{CO} of Kunze and
Stein phenomenon, for convolution operators on a class of locally symmetric
spaces. Let $C_{\rho }$ be the convex body in $\mathfrak{a^{\ast }}$
generated by the vectors $\left\{ w\rho ;\;w\in W\right\} .$ Let also
\lambda _{0}$ be the bottom of the $L^{2}$-spectrum of $\Delta _{M}$. Then
there exists a vector $\eta _{\Gamma }\in C_{\rho }\cap S(0,(\rho
^{2}-\lambda _{0})^{1/2})$, where $S(0,r)$ is the Euclidean sphere of
\mathfrak{a^{\ast }}$, such that for all $p\in (1,\infty )$ and for every $K$-bi-invariant function $\kappa $, the convolution operator $\ast |\kappa |$
with kernel $|\kappa |$ satisfies the estimate
\begin{equation}
\Vert \ast |\kappa |\Vert _{L^{p}(M)\rightarrow L^{p}(M)}\leq
\int_{G}|\kappa (g)|\varphi _{-i\eta _{\Gamma }}(g)^{s(p)}dg, \label{Kunze}
\end{equation
where
\begin{equation*}
\ s(p)=2\min ((1/p),(1/p^{\prime })),
\end{equation*
and $p^{\prime }$ is the conjugate of $p$. For more details, see \cit
{FOMAMA, LOMAjga}.
\subsection{The class $(CT)$}
In \cite{RO}, Roblin investigates the question of convergence or divergence
type of a discrete group $\Gamma $ of isometries of $CAT(-1)$ spaces. Note
that $CAT(-1)$ spaces contain all rank one symmetric spaces. To state the
results of Roblin, we need to introduce some notation.
Denote by $\mu _{x}$, $x\in X$, a $\Gamma $-invariant Patterson-Sullivan
density: that is a family of finite and mutually absolutely continuous
measures on $\partial X$ satisfying the following conditions:
(i) for any $x,y\in X$,
\begin{equation*}
\frac{d\mu _{y}}{d\mu _{x}}\left( \xi \right) =e^{-\delta \left( \Gamma
\right) \beta _{\xi }\left( x,y\right) },\text{ }\xi \in \partial X.
\end{equation*
Here $\beta _{\xi }\left( x,y\right) $ is the Busemann function:
\begin{equation*}
\beta _{\xi }\left( x,y\right) =\lim_{t\rightarrow \infty }\left( d\left(
\xi _{t},x\right) -d\left( \xi _{t},y\right) \right) ,
\end{equation*
where $\xi _{t}$ is a geodesic ray tending to $\xi $ as $t\rightarrow \infty
$.
(ii) for any $\gamma \in \Gamma $ and $x\in X$, $\gamma ^{\ast }\mu _{x}=\mu
_{\gamma x}$.
Note that for any $x\in X$, $\mu _{x}$ is supported on the limit set
\Lambda \left( \Gamma \right) $. Note also that the celebrated Patterson-Sullivan
construction insures the existence of such conformal densities
in various geometric contexts as
Hadamard manifolds and $CAT\left( -1\right)$ spaces, \cite{COO,YUE}.
If $\mu _{x_{0}}$ is normalized to be a probability measure, then in \cite
Theorem 1.7]{RO}, Roblin proves that $\Gamma $ is of convergence type iff
\mu _{x_{0}}\left( \Lambda _{c}\left( \Gamma \right) \right) =0$.
As far as it concerns divergence type groups, the question is investigated
in \cite{SU,COR} for rank one symmetric spaces and in \cite{DAL,LEU} for
rank greater than $2$. More precisely in \cite[Proposition 2]{SU} Sullivan
treats the case of the real hyperbolic space and shows that $\Gamma\subset SO(n,1)$ is of
divergence type if it is geometrically finite and convex co-compact, i.e.
\Gamma \backslash C\left( \Lambda \left( \Gamma \right) \right) $ is
compact, where $C\left( \Lambda \left( \Gamma \right) \right) $ is the
convex hull of the limit set of $\Gamma $. In \cite[Proposition 3.7]{COR}
Corlette and Iozzi treat the case of all rank one symmetric spaces and
improve the result of \cite{SU} in proving that if $\Gamma $ is
geometrically finite subgroup of isometries of a rank one symmetric space,
then $\Gamma $ is of divergence type.
In \cite{LEU} Leuzinger investigates the case when $G$ possesses Kazhdan's
property $(T)$, i.e. when $G$ has no simple factors locally isomorphic to
SO(n,1)$ or $SU(n,1)$, \cite[Ch. 2]{HAVA}. For example, the rank one
symmetric spaces which possess property $(T)$ are $H^{n}\left( \mathbb{H
\right) =Sp(n,1)/Sp(n)$, and $H^{2}\left( \mathbb{O}\right)
=F_{4}^{-20}/Spin\left( 9\right) $. In \cite[Main Theorem]{LEU} it is proved
that if $G$ is as above, then the following are equivalent:
(i) $M$ is a lattice, i.e. Vol($M$)$<\infty $,
(ii) $\delta (\Gamma )=2\rho $,
(iii) $\Gamma $ is of divergence type.
It follows then that if $\Gamma\in Sp\left(n,1\right)$ (resp. $\Gamma\in F_{4}^{20})$
and $\Gamma \backslash \mathbb{H}^{n}(\mathbb{
})$ (resp. $\Gamma \backslash \mathbb{H}^{2}(\mathbb{O})$) is a lattice, then
$\Gamma$ is of divergence type.
So, only discrete subgroups $\Gamma $ of $SO(n,1)$ or $SU(n,1)$ with $\delta (\Gamma )=2\rho $
can possibly be in the class $\left(CT\right)$.
\section{Proof of the results}
\subsection{Proof of Theorem \protect\ref{alphaless1}}
We start by performing a decomposition of the kernel $\kappa _{\alpha ,\beta
}=\mathcal{H}^{-1}(m_{\alpha ,\beta }).$ We write
\begin{equation*}
\ \kappa _{\alpha ,\beta }=\kappa _{\alpha ,\beta }^{0}+\kappa _{\alpha
,\beta }^{\infty },
\end{equation*
with $\kappa _{\alpha ,\beta }^{0},\kappa _{\alpha ,\beta }^{\infty }$, $K
-bi-invariant functions that satisfy $supp(\kappa _{\alpha ,\beta
}^{0})\subset B(0,2)$ and $supp(k_{\alpha ,\beta }^{\infty })\subset
B(0,1)^{c}.$ Denote by $\widehat{T}_{\alpha ,\beta }^{0}$ and $\widehat{T
_{\alpha ,\beta }^{\infty }$ the corresponding convolution operators with
kernels $\kappa _{\alpha ,\beta }^{0}$ and $\kappa _{\alpha ,\beta }^{\infty
}$.
The operator $\widehat{T}_{\alpha ,\beta }^{0}$ is the \textquotedblleft
local\textquotedblright\ part of $\widehat{T}_{\alpha ,\beta }$ and has the
same behavior as its analogue in the Euclidean context.
\begin{proposition}
\label{localprop} \label{local part}For $\alpha \in \left( 0,1\right) $,
\widehat{T}_{\alpha ,\beta }^{0}$ is bounded on $L^{p}(M)$, provided that
\operatorname{Re}{\beta >}\alpha n|1/p-1/2|$.
\end{proposition}
\begin{proof}
The continuity of $\widehat{T}_{\alpha ,\beta }^{0}$ on $L^p(M)$ follows
from \cite[Proposition 13]{LOMAanna}. Indeed, observe first that $\widehat{T
_{\alpha ,\beta }^{0}$ can be defined as an operator on the group $G$, and
then, apply the local result of \cite{GIUME}, to conclude its boundedness on
$L^{p}(G)$, for all $p\in (1,\infty )$ such that $\operatorname{Re} {\beta >}\alpha
n|1/p-1/2|$. The continuity of $\widehat{T}_{\alpha ,\beta }^{0}$ on
L^{p}(M)$, follows by applying Herz's Theorem A, \cite{HE2}.
\end{proof}
Note that Proposition \ref{localprop} is valid for all discrete and torsion
free subgroups of $G$.
To finish the proof of Theorem \ref{alphaless1} it remains to prove the
L^{p}$ boundedness of $\widehat{T}_{\alpha ,\beta }^{\infty }$. Here we
shall need the assumption that the Kunze and Stein phenomenon is valid on $M
.
\begin{proposition}
\label{infinity}If $M$ belongs in the class $(KS)$, and $\alpha \in \left(
0,1\right) $, then $\widehat{T}_{\alpha ,\beta }^{\infty }$ is bounded on
L^{p}(M)$ for all $p\in \left( 1,\infty \right) $.
\end{proposition}
For the proof of the proposition above we shall make use of the Kunze and
Stein phenomenon. For that we need to introduce some notation.
For $p\in (1,\infty )$, set
\begin{equation}
v_{\Gamma }(p)=2\min \{(1/p),(1/p^{\prime })\}\frac{|\eta _{\Gamma }|}{\rho
+|(2/p)-1|, \label{v(p)}
\end{equation
where $p^{\prime }$ is the conjugate of $p$ and $\eta _{\Gamma }\in
\mathfrak{a^{\ast }}$ is the vector appearing in (\ref{Kunze}).
For $N\in \mathbb{N}$, $v\in \mathbb{R}$ and $\theta \in \left( 0,1\right) $, we say that the multiplier $m$ belongs in the class $\mathcal{M}
(v,N,\theta )$, if
\begin{itemize}
\item $m$ is analytic inside the strip $\mathcal{T}^v=\mathfrak{a^*}
+ivC_{\rho}$ and
\item for all $k\in \mathbb{N}$ with $k\leq N$, $\partial ^{k}m(\lambda )$
extends continuously to the whole of $\mathcal{T}^{v}$ with
\begin{equation}
|\partial ^{k}m(\lambda )|\leq c(1+|\lambda |^{2})^{-k\theta /2}:=<\lambda
>^{-k\theta }. \label{decayrate}
\end{equation}
\end{itemize}
We have the following:
\begin{proposition}
\label{propos1} Fix $p\in (1,\infty )$ and consider an even function $m\in
\mathcal{M}(v,N,\theta )$, with $v>v_{\Gamma }(p)$ and $N=\left[ \frac{n+1}{
2\theta }\right] +1$. If $\kappa $ is the inverse spherical transform of $m$
in the sense of distributions, and $\kappa ^{\infty }$ its part away from
the origin, then, the convolution operator $\widehat{T}_{\kappa }^{\infty }$
with kernel $\kappa ^{\infty }$, is bounded on $L^{p}(M).$
\end{proposition}
\begin{proof}
Since $M$ belongs in the class $(KS)$, then, according to Kunze and Stein
phenomenon, we have that
\begin{eqnarray}
||\widehat{T}_{\kappa }^{\infty }||_{L^{p}(M)\rightarrow L^{p}(M)} &\leq
&\int_{G}\left\vert \kappa ^{\infty }(g)\right\vert \varphi _{-i\eta
_{\Gamma }}(g)^{s(p)}dg \notag \label{SkLp} \\
&\leq &\int_{|g|\geq 1}\left\vert \kappa (g)\right\vert \varphi _{-i\eta
_{\Gamma }}(g)^{s(p)}dg:=I. \label{KS}
\end{eqnarray}
So, to finish the proof of Proposition \ref{propos1}, we have to show that
if $m$ satisfies (\ref{decayrate}), then the integral $I$ in (\ref{SkLp}) is
finite. To prove that, we will slightly modify the proof of the Theorem 1 in
\cite{LOMAjga} which is based on Proposition 5 of \cite{AN}. To estimate the
integral $I$, we shall estimate $\kappa ^{\infty }$ over concentric shells.
Set
\begin{equation*}
V_{r}=\{H\in \mathfrak{a}:|H|\leq r\},
\end{equation*}
and
\begin{equation*}
V_{r}^{+}=V_{r}\cap \overline{\mathfrak{a}_{+}}.
\end{equation*}
Note that in the rank one case, $V_{r}=\left[ -r,r\right] $, and $V_{r}^{+}=
\left[ 0,r\right] $.
Set also
\begin{equation*}
\ U_{r}=\{x=k_{1}(\exp H)k_{2}\in G:k_{1},k_{2}\in K,\;H\in
V_{r}^{+}\}=K(\exp V_{r}^{+})K.
\end{equation*}
Using the Cartan decomposition of $G$, the integral $I$ is written as
\begin{equation}
I=\sum_{j\geq 1}\int_{U_{j+1}\backslash U_{j}}|\kappa (g)|\varphi _{-i\eta
_{\Gamma }}(g)^{s(p)}dg:=\sum_{j\geq 1}I_{j}. \label{cocentric}
\end{equation}
Set $b=n-1$ and let $b^{\prime }$ be the smallest integer $\geq b/2$. Set
also $N=\left[ \frac{n+1}{2\theta }\right] +1$, $\theta \in (0,1)$. In \cite[
p.645]{LOMAjga}, using \cite[p.608]{AN}, it is proved that if $m\in \mathcal
\ M}\left( v,N,\theta \right) $, then for $v>v_{\Gamma }(p)$ and $j\geq 1$,
\begin{equation}
I_{j}\leq cj^{-N}\sum_{0\leq k\leq N}\left( \int_{\mathfrak{a\ast }
}(<\lambda >^{b^{\prime }-N+k}|\partial _{\lambda }^{k}m(\lambda +i\rho
)|)^{2}d\lambda \right) ^{1/2}. \label{Ij}
\end{equation}
Then, using the estimates of the derivatives of $m(\lambda )$
\begin{equation*}
|\partial ^{k}m(\lambda )|\leq c<\lambda >^{-k\theta },
\end{equation*}
we obtain that
\begin{align*}
I_{j}& \leq cj^{-N}\sum_{0\leq k\leq N}\left( \int_{\mathfrak{a\ast }
}(<\lambda >^{b^{\prime }-N+k}<\lambda >^{-k\theta })^{2}d\lambda \right)
^{1/2} \\
& \leq cj^{-N}\sum_{0\leq k\leq N}\left( \int_{\mathfrak{a\ast }}(<\lambda
>^{b^{\prime }-N+k(1-\theta )})^{2}d\lambda \right) ^{1/2} \\
& \leq cj^{-N}\left( \int_{\mathfrak{a\ast }}<\lambda >^{2(b^{\prime
}-\theta N)}d\lambda \right) ^{1/2} \\
& \leq cj^{-N}\left( \int_{0}^{\infty }(1+\lambda ^{2})^{b^{\prime }-\theta
\left( \lbrack \frac{n+1}{2\theta }]+1\right) }d\lambda \right) ^{1/2}.
\end{align*}
Using the fact that $b^{\prime }\leq n/2$ and that $[\frac{n+1}{2\theta }]=
\frac{n+1}{2\theta }-q$, $q\in \lbrack 0,1)$, it follows that
\begin{align}
I_{j}& \leq cj^{-N}\left( \int_{0}^{\infty }(1+\lambda ^{2})^{\frac{n}{2}
-\theta \left( \frac{n+1}{2\theta }-q\right) -\theta }d\lambda \right) ^{1/2}
\label{j^-N} \\
& \leq cj^{-N}\left( \int_{0}^{\infty }(1+\lambda ^{2})^{-\frac{1}{2}-\theta
(1-q)}d\lambda \right) ^{1/2} \notag \\
& \leq cj^{-N}, \notag
\end{align}
since $2\left( \frac{1}{2}+\theta (1-q)\right) >1$.
From (\ref{j^-N}) it follows that
\begin{equation*}
\ \int_{|g|\geq 1}\left\vert \kappa ^{\infty }(g)\right\vert \varphi
_{-i\eta _{\Gamma }}(g)^{s(p)}dg=\sum_{j\geq 1}I_{j}\leq c\sum_{j\geq
1}j^{-N}<\infty ,
\end{equation*}
since $N>1$.
\end{proof}
\bigskip
\begin{proof}[Proof of Proposition \protect\ref{infinity}]
Note first that $m_{\alpha ,\beta }\left( \lambda \right) $ has poles only
at $\lambda =\pm i\rho $. So, the function $\lambda \longrightarrow
m_{\alpha ,\beta }\left( \lambda \right) $ is analytic in the strip $S_{\rho
}=\left\{ z\in \mathbb{C}:\left\vert \operatorname{Im}z\right\vert <\rho \right\} $.
Secondly, for every $k\in \mathbb{N}$, it holds that
\begin{equation}
\left\vert \partial ^{k}m_{a,\beta }(\lambda )\right\vert \leq c(1+\lambda
)^{-Re\beta -k(1-a)}\leq c(1+\lambda ^{2})^{-k(1-\alpha )/2},\text{ }\lambda
\in S_{\rho }. \label{in2}
\end{equation}
Note also that for any $p\in \left( 1,\infty \right) $, from (\ref{v(p)}) it
follows that $v_{\Gamma }(p)<1$. So, for every $p\in \left( 1,\infty \right)
$, there exists $v^{\prime }\left( p\right) <1$, such that $v^{\prime
}\left( p\right) >v_{\Gamma }(p)$. It follows that $\mathfrak{a^{\ast }
+iv^{\prime }\left( p\right) C_{\rho }\subset S_{\rho }$, which combined
with (\ref{in2}), implies that for any $p\in \left( 1,\infty \right) $, and
N\in \mathbb{N}$, $m_{\alpha ,\beta }\in \mathcal{M}(v^{\prime }\left(
p\right) ,N,1-\alpha )$. Thus, Proposition \ref{propos1} applies and
Proposition \ref{infinity} follows.
\end{proof}
\subsection{Proof of Theorem \protect\ref{alphaeq1}}
To begin with, let us recall that in \cite{SCHO}, the operator $T_{\alpha
,\beta }$, $Re\beta \geq 0$, $\alpha >0$ is also expressed by the integral:
\begin{equation}
T_{\alpha ,\beta }(f)(x)=\frac{1}{\Gamma (\beta /\alpha )}\int_{0}^{\infty
}\sigma ^{\beta /\alpha -1}(f\ast q_{\sigma ,\alpha })(x)\;d\sigma ,\quad
f\in C_{0}^{\infty }(X), \label{Ta,b}
\end{equation
where $q_{\sigma ,\alpha }$, $\sigma >0$, is the inverse spherical Fourier
transform of the function $\lambda \rightarrow e^{(i-\sigma )(\lambda
^{2}+\rho ^{2})^{\alpha /2}}$. Note that $q_{\sigma ,\alpha }$ is $K$-bi-invariant as the inverse spherical Fourier transform of an even
function, \cite{AN}. Then, observe that the convolution operator
T_{q_{\sigma ,\alpha }}=\ast q_{\sigma ,\alpha }$ is equal to $e^{(i-\sigma
)\Delta _{X}^{\alpha /2}}$. So, by the spectral theorem, $T_{q_{\sigma
,\alpha }}$ is bounded on $L^{2}(X)$, with
\begin{equation}
\Vert T_{q_{\sigma }}\Vert _{L^{2}(X)\rightarrow L^{2}(X)}\leq \sup_{\lambda
>0}\left\vert e^{(i-\sigma )(\lambda ^{2}+\rho ^{2})^{\alpha /2}}\right\vert
\leq e^{-\sigma \rho ^{\alpha }}. \label{L(2(X))}
\end{equation}
For simplicity, for $\alpha =1$, set $q_{\sigma ,1}=q_{\sigma }$. Thus,
T_{q_{\sigma }}=\ast q_{\sigma }$ is equal to $e^{(i-\sigma )\Delta
_{X}^{1/2}}$. Let us now define the operator $\widehat{T}_{1,\beta }$ on the
quotient $M=\Gamma \backslash X$. For that, recall that $\widehat{T
_{q_{\sigma }}$, as it is the case of $\widehat{T}_{1,\beta }$, is initially
defined as a convolution operator
\begin{equation}
\widehat{T}_{q_{\sigma }}f(x)=\int_{G}q_{\sigma }(y^{-1}x)f(y)dy,\quad f\in
C_{0}^{\infty }(M). \label{33}
\end{equation}
Set $q_{\sigma }(x,y)=q_{\sigma }(y^{-1}x)$ and
\begin{equation}
\widehat{q_{\sigma }}(x,y)=\sum_{\gamma \in \Gamma }q_{\sigma }(x,\gamma
y)=\sum_{\gamma \in \Gamma }q_{\sigma }((\gamma y)^{-1}x). \label{conv}
\end{equation}
\begin{proposition}
\label{forallgroups}If either $\delta (\Gamma )<2\rho $ or $\Gamma \in (CT)
, then the series (\ref{conv}) is convergent and the operator $\widehat{T}
_{q_{\sigma }}$ on $M$ is given by
\begin{equation}
\widehat{T}_{q_{\sigma }}f(x)=\int_{M}\hat{q}_{\sigma }(x,y)f(y)dy,\quad
f\in C_{0}^{\infty }(M). \label{Tqhat}
\end{equation}
\end{proposition}
\begin{proof}
By the Cartan decomposition of $G$, we may write
\begin{equation*}
(\gamma y)^{-1}x=k_{\gamma }\exp (t_{\gamma }H_{0})k_{\gamma }^{\prime },
\end{equation*}
where $t_{\gamma }>0,\;k_{\gamma },k_{\gamma }^{\prime }\in K,$ and
\;H_{0}\in \mathfrak{a}_{+}$ with $\left\Vert H_{0}\right\Vert =1$. Then,
since $q_{\sigma }$ is $K$-bi-invariant, we get that $q_{\sigma }((\gamma
y)^{-1}x)=q_{\sigma }(\exp t_{\gamma }H_{0}).$
Recall that in \cite[p.103]{GIUME} it is proved that
\begin{equation}
\left\vert q_{\sigma }(\exp tH_{0})\right\vert \leq
\begin{cases}
c\sigma (t+1)^{-3/2}e^{-2\rho t}, & \sigma >1,t>0, \\
c(t+1)^{-3/2}e^{-2\rho t}, & 0<\sigma \leq 1,t>2
\end{cases}
\label{q1}
\end{equation
Then,
\begin{eqnarray*}
|\widehat{q}_{\sigma }(\exp tH_{0})| &\leq &\sum_{\gamma \in \Gamma
}|q_{\sigma }(\exp t_{\gamma }H_{0})| \\
&\leq &\sum_{\left\{ \gamma \in \Gamma :\;t_{\gamma }\leq 2\right\}
}q_{\sigma }(\exp t_{\gamma }H_{0})+\sum_{\left\{ \gamma \in \Gamma
:\;t_{\gamma }>2\right\} }q_{\sigma }(\exp t_{\gamma }H_{0}).
\end{eqnarray*
Note that the first sum above is finite, since $\Gamma $ is discrete. Thus,
we shall deal only with the second sum. Using (\ref{tgamma}) and the
estimates (\ref{q1}) of $q_{\sigma }$, we have
\begin{eqnarray*}
\sum_{\left\{ \gamma \in \Gamma :\;t_{\gamma }>2\right\} }q_{\sigma }(\exp
t_{\gamma }H_{0}) &\leq &\sum_{\left\{ \gamma \in \Gamma :\;t_{\gamma
}>2\right\} }c_{\sigma }(t_{\gamma }+1)^{-3/2}e^{-2\rho t_{\gamma }} \\
&\leq &c_{\sigma }\sum_{\left\{ \gamma \in \Gamma :d(x,\gamma y)>2\right\}
}e^{-2\rho d(x,\gamma y)} \\
&\leq &c_{\sigma }P_{2\rho }\left( x,y\right)
\end{eqnarray*
which is convergent, since by our assumption, either $\delta (\Gamma )<2\rho
$ or $\Gamma \in (CT)$.
To prove (\ref{Tqhat}), we follow \cite[Prop.4]{FOMAMA}. Since $q_{\sigma }$
and $f$ are right $K$-invariant, from (\ref{33}) we get that
\begin{equation*}
\widehat{T}_{q_{\sigma }}(f)(x)=\int_{G}q_{\sigma
}(xy^{-1})f(y)dy=\int_{X}q_{\sigma }(x,y)f(y)dy.
\end{equation*}
Further, since $f$ is left $\Gamma $-invariant, by Weyl's formula we obtain
that
\begin{align*}
\widehat{T}_{q_{\sigma }}(f)(x) &=\int_{X}q_{\sigma
}(x,y)f(y)dy=\int_{\Gamma \backslash X}\left( \sum_{\gamma \in \Gamma
}q_{\sigma }(x,\gamma y)f(\gamma y)\right) dy \\
&=\int_{M}\widehat{q}_{\sigma }(x,y)f(y)dy.
\end{align*}
\end{proof}
\begin{proposition}
\label{interpol}If either $\delta (\Gamma )<2\rho $ or $\Gamma \in (CT)$,
then for all $p\in (1,\infty )$, there are constants $c_{p},k_{p}>0$ such
that
\begin{equation}
\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{p}(M)\rightarrow L^{p}(M)}\leq
c_{p
\begin{cases}
e^{-k_{p}\sigma }, & \sigma \geq 1, \\
\sigma ^{(1-n)(1/2-1/p)}, & \sigma <1
\end{cases}
\label{pp}
\end{equation}
\end{proposition}
\begin{proof}
Recall that $\widehat{T}_{q_{\sigma }}=e^{(i-\sigma )\Delta_M ^{1/2}}$.
Thus, by the spectral theorem,
\begin{equation} \label{L2M}
\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{2}(M)\rightarrow L^{2}(M)} \leq
\sup_{\lambda >0}\left\vert e^{-(i-\sigma )(\lambda ^{2}+\rho
^{2})^{1/2}}\right\vert \leq e^{-\sigma \rho }.
\end{equation}
Next, note that if $f\in C_{0}^{\infty }(M)$, then $\Vert f\Vert _{L^{\infty
}(M)}=\Vert f\Vert _{L^{\infty }(X)}$, since $f$ is left $\Gamma $-invariant. Then, again by Weyl's formula it follows that
\begin{eqnarray}
|\widehat{T}_{q_{\sigma }}f(x)| &=&\left\vert \int_{M}\hat{q}_{\sigma
}(x,y)f(y)dy\right\vert =\left\vert \int_{\Gamma \backslash X}\sum_{\gamma
\in \Gamma }{q_{\sigma }}(x,\gamma y)f(y)dy\right\vert \notag \label{24} \\
&=&\left\vert \int_{X}q_{\sigma }(x,y)f(y)dy\right\vert \leq \Vert f\Vert
_{L^{\infty }(X)}\Vert q_{\sigma }\Vert _{L^{1}(X)} \label{++} \\
&=&\Vert f\Vert _{L^{\infty }(M)}\Vert q_{\sigma }\Vert _{L^{1}(X)}. \notag
\end{eqnarray
But, in \cite[p.101]{GIUME} it is proved that
\begin{equation}
\Vert {q_{\sigma }}\Vert _{L^{1}(X)}\leq
\begin{cases}
\sigma , & \sigma \geq 1, \\
\sigma ^{(1-n)/2}, & \sigma <1
\end{cases}
\label{q_sigmaest}
\end{equation
From (\ref{++}) and (\ref{q_sigmaest}), it follows that $\widehat{T
_{q_{\sigma }}$ is bounded on $L^{\infty }(M)$, with
\begin{equation}
\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{\infty }(M)\rightarrow L^{\infty
}(M)}\leq \Vert q_{\sigma }\Vert _{L^{1}(X)}. \label{Loo}
\end{equation}
By interpolation between (\ref{L2M}) and (\ref{Loo}) and duality, we deduce
the boundedness of $\widehat{T}_{q_{\sigma }}$ on $L^{p}(M)$, $p>1$.
Further, by the Riesz-Thorin interpolation theorem we have
\begin{equation*}
\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{p_{\theta }}(M)\rightarrow
L^{p_{\theta }}(M)}\leq c\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{{2
}(M)\rightarrow L^{{2}}(M)}^{1-\theta }\Vert \widehat{T}_{q_{\sigma }}\Vert
_{L^{{\infty }}(M)\rightarrow L^{{\infty }}(M)}^{\theta },
\end{equation*
with $\frac{1}{p_{\theta }}=\frac{1-\theta }{2}$. Choosing $\theta
=1-(2/p),\;p>2$, we have that
\begin{equation*}
\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{p}(M)\rightarrow L^{p}(M)}\leq
c\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{p}(M)\rightarrow
L^{p}(M)}^{2/p}\Vert \widehat{T}_{q_{\sigma }}\Vert _{L^{\infty
}(M)\rightarrow L^{\infty }(M)}^{1-(2/p)},
\end{equation*
and from (\ref{L2M}), (\ref{q_sigmaest}) and (\ref{Loo}), it is
straightforward to obtain (\ref{pp})\ for $p>2$. The claim for $p\in \left(
1,2\right) $ follows by duality.
\end{proof}
\begin{proof}[\textit{End of proof of Theorem \protect\ref{alphaeq1}}.]
Using (\ref{Ta,b}), we have that
\begin{eqnarray*}
\Vert \widehat{T}_{1,\beta }\Vert _{L^{p}(M)\rightarrow L^{p}(M)} &\leq
\frac{1}{\Gamma (\beta )}\int_{0}^{\infty }\sigma ^{\beta -1}\Vert \widehat{
}_{q_{\sigma }}\Vert _{L^{p}(M)\rightarrow L^{p}(M)}\;d\sigma \\
&\leq &\frac{1}{\Gamma (\beta )}\int_{0}^{1}\sigma ^{\beta -1}\Vert \widehat
T}_{q_{\sigma }}\Vert _{L^{p}(M)\rightarrow L^{p}(M)}\;d\sigma \\
&+&\frac{1}{\Gamma (\beta )}\int_{1}^{\infty }\sigma ^{\beta -1}\Vert
\widehat{T}_{q_{\sigma }}\Vert _{L^{p}(M)\rightarrow L^{p}(M)}\;d\sigma .
\end{eqnarray*
Applying the estimates (\ref{pp}), we get:
\begin{eqnarray*}
\Vert \widehat{T}_{1,\beta }\Vert _{L^{p}(M)\rightarrow L^{p}(M)} &\leq
\frac{1}{\Gamma (\beta )}\int_{0}^{1}\sigma ^{\beta -1}\sigma
^{(1-n)(1/2-1/p)}\;d\sigma \\
&+&\frac{1}{\Gamma (\beta )}\int_{1}^{\infty }\sigma ^{\beta
-1}e^{-k_{p}\sigma }\;d\sigma .
\end{eqnarray*
The second integral above is finite, while the first is convergent provided
that
\begin{equation*}
\ \operatorname{Re}\beta -1-(n-1)(1/2-1/p)>-1,\text{ or }\operatorname{Re}\beta
/(n-1)>1/2-1/p,
\end{equation*
and the claim follows for $p>2$. The case $p<2$ follows by duality.
\end{proof}
\textbf{Acknowledgement}. The author would like to thank Professor Michel
Marias for his generous and valuable help, and for bringing the problem to
our attention, as well as Professor Anestis Fotiadis for stimulating
discussions and support.
|
\section{Introduction}
Streamer discharges are fast filamentary transients that evolve due to self-enhanced electric fields at their tips. Streamers are the natural precursors of sparks, lightning, and sprites, and have found use in sterilization of polluted gases and breakup of molecules, plasma assisted combustion, and control of airflow in the boundary layer of airplane wings \cite{Boeuf2005, Moreau2007, Soloviev2009, 1167639, 18870, 4504897, 55956, Nair2004, Grymonpre2001, Ebert2006,Bogaerts2002}. The spatial scales for streamer discharges span several orders of magnitude. Firstly, a streamer is essentially a non-thermal plasma filament surrounded by a space charge layer with a thickness one to two orders of magnitude thinner than the filament diameter. Secondly, the length of the filament can be much longer than its thickness, leading to a numerical problem with widely different spatial scales. On the one hand, a fine numerical resolution is required for resolving the space charge layer. On the other, a large computational domain is required in order to facilitate the propagation of the streamer. Due to the nonlinearity of streamers, numerical solutions can usually not be obtained on coarse grids.
Streamers can be described by using either fluid or kinetic approaches, or a combination of them \cite{Li2010, Li2012}. In a fluid approach, electrons and ions are evolved according to their fluid moments (usually truncated to first order) by using tabulated or analytic transport data. Kinetic approaches approximate the phase space distribution function by evolving computational particles, using cross-sectional collision data as input parameters. The kinetic approach is computationally far more exhaustive, and includes far more physics. Fluid approximations are more common on larger scales, and this is the description that we use in this paper.
It is generally believed that spatially adaptive methods are ideally suited - and maybe even necessary - for large scale simulations of 3D streamers. The reason for this is that a streamer represents a dynamically evolving structure with possible stochastic behavior, restricting the use static grids to cases where the streamer path is a priori known, or small-scale cases where the entire domain can be resolved at the finest spatial resolution. Because of these difficulties, 3D simulations of streamers are rare \cite{Hallac2003, Pancheshnyi2008, Luque2008, Papageorgiou2011, Nijdam2016a, Teunissen2017, Teunissen2017, Plewa2018, Marskar2018}. In addition to multiple spatial scales, various characteristic time scales also exist for streamer discharges. For example, the avalanche-to-streamer time describes the time it takes for seed electrons to reach a critical size through impact ionization, whereas the dielectric relaxation time describes the local rate-of-change of the electric field due to motion of charge carriers. The relaxation time can be expressed in terms of the electric field $\mathbf{E}$ and the electric current density $\mathbf{J}$ as
\begin{equation}
\Delta t_{\mathbf{E}} = \frac{\epsilon_0\left|\mathbf{E}\right|}{\left|\mathbf{J}\right|},
\end{equation}
which follows from a first-order trunction of Ampere's law. This time scale is an important one for streamer simulations. Note that $\Delta t_\ensuremath{{\mathbf{E}}}$ is a \emph{physical} time scale that does not reference the grid resolution.
Computer codes have restrictions on which time step sizes can be used to evolve the fluid equations. For example, explicit methods for advection obey a Courant-Friedrichs-Lewy (CFL) time step restriction $\Delta t_{\ensuremath{\textrm{cfl}}} = \Delta x/|\mathbf{v}|$ where $\Delta x$ is the spatial resolution and $\mathbf{v}$ is the advective velocity. For parabolic equations, such as diffusion equations, the time step restriction on explicit methods is $\Delta t_{\ensuremath{D}} \sim \Delta x^2$, which quickly becomes even stricter than $\Delta t_\ensuremath{\textrm{cfl}}$ when diffusion is relevant. Furthermore, although not required for stability, one may derived a fourth time step constraint $\Delta t_{S} = \frac{S}{n}$, where $S$ is a source term and $n$ is a species density, which describes a characteristic time scale for growth due to ionization. For multiple components, the above time step sizes also minimized over species according to the worst offender, which is usually the electrons. On the other hand, kinetic models like Particle-In-Cell (PIC) choose their time steps differently, being restricted by the inverse plasma frequency and the inverse collision frequencies.
For streamer simulations, the above mentioned time scales are widely varying. Fully implicit methods are likely to be the most robust, but it is not yet clear how these should be implemented in the multiphysics environment of streamers, particularly if high-order shock-capturing methods on adaptive grids are involved for the convection part. For that reason, implicit-explicit (IMEX) methods are more attractive, allowing one to treat restrictive time scales with implicit methods and use explicit methods on others. Previous experience with large scale 3D simulations \cite{Marskar2018} show that such methods are often CFL bound on the time step, leading us to speculate that run-time performance can be improved through a better segregation between physical and numerical time scales. The rationale behind this idea is that it should be unnecessary to perform electric field updates when $\Delta t_{\ensuremath{\textrm{cfl}}} \ll \Delta t_{\ensuremath{{\mathbf{E}}}}$, since the correction to the $\ensuremath{{\mathbf{E}}}$-field would then be miniscule. Similar ideas have been adopted in particle codes \cite{Fierro2014, Teunissen2016}. Likewise, diffusive models approximations of the radiative transport equation reduce to Helmholtz equations, and it is neither not clear how often these need to be updated.
This paper quantifies the prospects of advancing fluid and elliptic equations using different time steps. Potentially, such techniques may lead to speedup (and flexibility) for fluid streamer simulation by elimination of extranous elliptic solves at very fine time scales, which are present for consistent schemes. The outline of this paper is as follows: In Sec.~\ref{sec:theory} we present a time stepping scheme that segregates the numerical and physical time scales. The numerical error of this scheme is then quantified in Sec.~\ref{sec:2D} by means of two-dimensional simulation experiments. Thus, we partially answer the question ''How often do we need to update the electric field?'', which is of interest to both the kinetic and fluid modeling parts of the plasma community. Finally, we present a high-performance computing example in Sec.~\ref{sec:hpc_example} that uses some of these techniques, and provide some concluding remarks in Sec.~\ref{sec:conclusions}.
\section{Theory}
\label{sec:theory}
We use a simplified fluid model for gas discharges, based on the following equations
\begin{eqnarray}
\label{eq:streamer_equations}
&\partial_tn_i = \nabla\cdot\left(D_i\nabla n_i - \mathbf{v}_in_i\right) + S_i, \\
&\nabla^2\phi = -\frac{\rho}{\epsilon_0}, \\
\label{eq:rte}
&\kappa_j\Psi_j - \nabla\cdot\left(\frac{1}{3\kappa_j}\nabla\Psi_j\right) = \frac{\eta_j}{c},
\end{eqnarray}
where $n_i$ is the density of species $i$, $D_i$ the diffusion coefficient, $\mathbf{v}_i$ the velocity, and $S_i$ the source term. The electric potential is given by $\phi$ and $\rho = \sum_{i}q_in_i$ is the space charge density. The symbols $\Psi_j$ denote the isotropic radiative density of a photon group $j$, $\kappa_j$ is the Beer's length and $\eta_j$ is an isotropic source term.
Equations \eref{eq:streamer_equations}-\eref{eq:rte} are solved with finite volumes over an adaptive mesh based on Cartesian grid patches, and solid boundaries are treated with an embedded boundary formalism. The spatial discretization of Eqs.~\eref{eq:streamer_equations} allows for arbitrarily cut cells (even ones that contain multiple cell fragments), and is as follows:
\begin{enumerate}
\item The convective term $\nabla\cdot\left(\mathbf{v}_in_i\right)$ is discretized with the unsplit Godunov's method. The state at a face center is given by the solution to a Riemann problem with slope-limited left and right states; the Riemann solution is the upwind state at the face. One-sided slopes are used if there are not enough cells available for the left or right slopes, which can occur if a cell face lies completely inside a material. On cut-cells, we require the flux on face centroids in addition to face centers. This is done by first computing fluxes at face centers, and then interpolating these to the respective face centroids. We additionally stabilize the convective derivative by computing a hybrid divergence with charge redistribution that allows us to use a standard CFL condition for cut cells with arbitrarily small volume fractions. Charge injection into the domain is a part of the advective discretization, and thus the injected charge is also redistributed in the neighborhood of the cut cells.
\item The diffusion advancement of $\partial_tn_i = \nabla\cdot\left(D_i \nabla n_i\right)$ is handled implicitly with the Twizell-Gumel-Arigu (TGA) scheme \cite{Twizell1996} (see equation~\eref{eq:tga}). This scheme is very stable in embedded boundary applications. The spatial discretization of the elliptic term $\nabla\cdot\left(D_i \nabla n_i\right)$ is done with a second-order accurate cell-centered discretization with Neumann boundary conditions everywhere. The resulting Helmholtz equations that arise from the TGA discretization are solved with a geometric multigrid method with red-black Gauss-Seidel smoothers and a biconjugate gradient stabilized method as the bottom solver.
\item The remaining elliptic equations for the Poisson and the radiative transfer equations are discretized with second order accurate cell-centered solvers. The spatial discretization is the same as for the diffusion equation above, with the exception of boundary conditions.
\end{enumerate}
The computer code that we use is compatible with adaptive mesh refinement and runs at high concurrencies (tested for up to 8k cores so far). A full discussion of the underlying software is not possible in the scope of this paper, but can be found elsewhere \cite{Marskar2018}. We will use this code to explore the numerical error obtained by using different-sized time steps for fluid and elliptic equations, thereby probing the importance of updating the Poisson and radiative transfer equations. We will then apply these findings to a large scale simulation example.
\subsection{Regular time stepping}
\label{sec:rk2_tga}
First, we present a consistent second order time stepper that is suitable for embedded boundary applications, which we will refer to as the consistent scheme. This integrator is based on a second order monotone-in-time Runge-Kutta method together with second order implicit diffusion and advances $t^k \rightarrow t^{k+1}$ as follows:
\begin{enumerate}
\item Compute $n^\ast = n^k + \Delta t\left[S^k - \nabla\cdot\left(\mathbf{v}^kn^k\right)\right]$.
\item Compute $\mathbf{E}^\ast$ by solving the Poisson equation with the space charge density $\rho^\ast=\rho(n^\ast)$.
\item Compute radiative transfer source terms $\eta^\ast = \eta\left(n^\ast, \mathbf{E}^\ast\right)$ and obtain $\Psi_\gamma^\ast$ by solving the RTE equations.
\item Advance $n^\dagger = \frac{1}{2}\left(n^k + n^\ast + \Delta t\left[S^\ast - \nabla\cdot\left(\mathbf{v}^\ast n^\ast\right)\right]\right)$ where $S^\ast = S\left(n^\ast, \Psi^\ast, \ensuremath{{\mathbf{E}}}^\ast \right)$ and $\mathbf{v}^\ast = \mathbf{v}\left(\ensuremath{{\mathbf{E}}}^\ast\right)$
\item Compute $\mathbf{E}^\dagger$ by solving the Poisson equation with $\rho^\dagger = \rho\left(n^\dagger\right)$.
\item Compute diffusion coefficients $D = D\left(\mathbf{E}^\dagger\right)$ and obtain $n^{k+1}$ with an implicit diffusion advance
\begin{equation}
\label{eq:tga}
\left(I-\mu_1L\right)\left(I - \mu_2L\right)n^{k+1} = \left(I+\mu_3L\right)n^\dagger,
\end{equation}
where $L$ is the diffusion operator. Expressions for $\mu_1$, $\mu_2$, and $\mu_3$ are found in \cite{Marskar2018}.
For non-diffusive species then $n^\dagger\rightarrow n^{k+1}$.
\item Obtain the final electric field $\mathbf{E}^{k+1}$ by solving the Poisson equation with $\rho^{k+1} = \rho\left(n^{k+1}\right)$.
\item Compute radiative transfer source terms $\eta^{k+1} = \eta\left(n^{k+1}, \mathbf{E}^{k+1}\right)$ and obtain $\Psi_\gamma^{k+1}$ by solving the RTE equations
\end{enumerate}
In the above, steps (i) through (iv) describe a consistent strongly stability preserving Runge-Kutta method of order two; steps (vi) and (vii) describe an implicit diffusion advance, and step (viii) describe the final radiative transfer update. For the physical model that we will consider, this scheme performs 11 elliptic updates per time step: Three Poisson updates, six radiative transfer equation solves, and two Helmholtz solves for the diffusion step. The computational bottlenecks are due to these elliptic updates.
\subsection{Probing elliptic equations - a modified time stepper}
The modified scheme segregates the evolution into different-sized time steps. The outline of this method is as follows: We assume the existence of two disparate time scales $\Delta t_f$, which is a ''fine'' numerical time scale, and $\Delta t_c$ which is a ''coarse'' numerical time scale. We assume $\Delta t_f < \Delta t_c$ and $\Delta t_c = m\Delta t_f$ where $m > 0$ is an integer such that the time scales synchronize. Typically, $\Delta t_f$ is is smaller than the CFL time $\Delta t_{\ensuremath{\textrm{cfl}}}$, whereas $\Delta t_c$ is a coarser numerical time scale, typically taken to be proportional to the dielectric relaxation time. The various time steps are summarized in Table~\ref{tab:timesteps}. We then evolve the equations of motions in the following way:
\begin{enumerate}
\item Evolve the advective-reactive part of the species equations from $t^k$ to $t^k + \Delta t_c$ by using $m$ non-diffusive steps with individual time step sizes $\Delta t_f < \Delta t_\ensuremath{\textrm{cfl}}$, corresponding to steps (i)-(iv) in \Sref{sec:rk2_tga}. The advective discretization over a time step $\Delta t_f$ then occurs as follows:
\begin{eqnarray}
\label{eq:advec_react}
n^\ast &= n^k + \Delta t_f\left[S^k - \nabla\cdot\left(\mathbf{v}^kn^k\right)\right], \\
\label{eq:s_modified}
S^\ast &= S\left(n^\ast, \Psi^k, \ensuremath{{\mathbf{E}}}^k \right), \\
\label{eq:v_modified}
\ensuremath{{\mathbf{v}}}^\ast &= \ensuremath{{\mathbf{v}}}\left(\ensuremath{{\mathbf{E}}}^k\right), \\
n^\dagger &= \frac{1}{2}\left(n^k + n^\ast + \Delta t_f\left[S^\ast - \nabla\cdot\left(\mathbf{v}^\ast n^\ast\right)\right]\right).
\end{eqnarray}
This process is repeated $m$ times such that the equations are advanced a time $\Delta t_c$.
\item Perform a diffusion advance $\left(I-\mu_1L\right)\left(I - \mu_2L\right)n^{k+1} = \left(I+\mu_3L\right)n^\dagger$. For non diffusive species, $n^\dagger \rightarrow n^{k+1}$.
\item Solve the Poisson equation with $\rho^{k+1} = \rho\left(n^{k+1}\right)$ and obtain the new electric field $\ensuremath{{\mathbf{E}}}^{k+1}$.
\item Compute radiative transfer source terms $\eta^{k+1} = \eta\left(n^{k+1}, \mathbf{E}^{k+1}\right)$ and obtain $\Psi_\gamma^{k+1}$ by solving the RTE equations
\end{enumerate}
We will examine three cases in total so that we can estimate which elliptic equations are important for numerical accuracy, and which ones are not:
\begin{description}
\item{Case 1:} Update all elliptic equations at the coarse time step $\Delta t_c$ as above.
\item{Case 2:} Also update $\ensuremath{{\mathbf{E}}}$ consistently. In this case, we additionally update the electric field at each Runge-Kutta stage of \eref{eq:advec_react}.
\item{Case 3:} Additionally update the RTE equations consistently. In this case, we additionally update the radiative transfer equations at each Runge-Kutta stage of \eref{eq:advec_react}.
\end{description}
Thus, the electric field and radiative transport equations are updated at the either the coarse or the fine numerical time step, allowing us to assess the importance of both types of elliptic updates. Note that, at worst, the above scheme is first order accurate in $\Delta t_c$ due the modified electric field coupling.
\begin{table}[h!t!b!]
\begin{tabular}{lll}
Symbol & Formula & Description \\
\hline
$\Delta t_c$ & - & Coarse numerical time step \\
$\Delta t_f$ & - & Fine numerical time step \\
$\Delta t_{\ensuremath{{\mathbf{E}}}}$ & $\epsilon_0\left|\ensuremath{{\mathbf{E}}}\right|/\mathbf{J}$ & Dielectric relaxation time \\
$\Delta t_{\ensuremath{\textrm{cfl}}}$ & $\Delta x/\left|\mathbf{v}\right|$ & CFL time step \\
$\Delta t_{S}$ & $S/n$ &Ionization time step \\
\end{tabular}
\caption{Time step size nomenclature. The time steps $\Delta t_{\ensuremath{\textrm{cfl}}}$ and $\Delta t_{S}$ are minimized according to the worst offender (usually electrons). }
\label{tab:timesteps}
\end{table}
The rationale for the above approach is removal of potentially redundant elliptic solves at every $\Delta t_f$, which would be present for a consistent scheme. Indeed, one should not need to update the Poisson equation if $\Delta t_f \ll \Delta t_{\ensuremath{{\mathbf{E}}}}$, but rather resolve its physical time scale $\Delta t_\ensuremath{{\mathbf{E}}}$ at reasonable accuracy. It is quite clear that an electric field update is redundant when space charge currents are negligible, which is e.g. the case in the avalanche stage of a streamer discharge. Elliptic solves are bottlenecks for most streamer simulation codes, and unnecessary updates can easily become performance killers. This is particularly true when direct solvers are involved, but also for iterative solvers when curved solid boundaries are present since this leads to deteriorated convergence rates for multigrid smoothing. One may possible extend these ideas to kinetic models, although the accuracy of this simplification must then be determined by a kinetic code, which is not a topic in this paper.
The use of longer time steps for the Poisson equation is not new. In the PIC code in \cite{Teunissen2016}, the authors update the electric field using a time step that may be as long as $\Delta t_{\ensuremath{{\mathbf{E}}}}$, i.e. $\Delta t_c \leq \Delta t_{\ensuremath{{\mathbf{E}}}}$. Likewise, the authors in \cite{Fierro2014} use a fixed time step for the Poisson update and a much finer time step for the particle push. The accuracy of this approximation was not quantified in either paper. Relatedly, the authors of the fluid code in \cite{Teunissen2017} and the authors of the PIC code in \cite{Fierro2018} use a much stricter condition, updating the Poisson equation consistently at each advective step or particle push. For the conditions in \cite{Fierro2018}, the time steps are on the order of $1\,\ensuremath{{\textrm{fs}}}$ whereas the dielectric relaxation time is estimated to be on the order of $100\,\ensuremath{{\textrm{fs}}}$. On the other hand, for most fluid simulations reported to date, the electric field updates are consistent, i.e. they are performed at each advective or diffusive step, regardless of the dielectric relaxation time.
\section{Two-dimensional simulation experiments}
\label{sec:2D}
In this section we estimate how often the $\ensuremath{{\mathbf{E}}}$-field should be updated by means of two-dimensional numerical experiments. We consider a Cartesian $(2\,\ensuremath{{\textrm{cm}}})^2$ domain with a rod-plane geometry. The rod is a cylinder with a hemispherical cap with a radius of $500\,\ensuremath{{\mu\textrm{m}}}$ protruding $1\,\ensuremath{{\textrm{cm}}}$ from the center of the top domain edge. This edge, and the needle, is live with a voltage of $40\,\ensuremath{{\textrm{kV}}}$ whereas the bottom edge is grounded. For simplicity, we use homogeneous Neumann boundary conditions on the left and right side edges.
For chemistry, we consider a three-species model for air at 1 bar that consists of electrons $n_\ensuremath{\textrm{e}}$, positive ions $n_+$, and negative ions $n_-$. The kinetic coefficients for this scheme are given in \cite{Morrow1997}. All three species are advected, but only electrons are diffusive. Note that for positive streamer discharges that do not touch the cathode, one might possibly simplify the system further by assuming immobile ions, although this is not possible for negative streamer due to the dynamics of the cathode sheath. For radiative transport, we consider the model by \cite{Segur2006, Bourdon2007}. The initial conditions are $n_\ensuremath{\textrm{e}} = n_+ = 10^{10}\,\ensuremath{{\textrm{m}}}^3$ and the equations are integrated for $2\,\ensuremath{{\textrm{ns}}}$, which is sufficiently long for capturing the three basic simulations phases: electron avalanche, streamer inception, and streamer propagation phase.
The numerical error of the modified scheme is evaluated by comparison with the consistent scheme at each step. Let $n^{k+1}$ denote the final state after advancing the species equation from $t_k$ to $t_k + \Delta t_c$ by using the modified scheme. Let $\widetilde{n}^{k+1}$ denote the final state after advancing the species equation from $t_k$ to $t_k+\Delta t_c$ by using the consistent scheme. The relative error obtained for $n^{k+1}$ at each grid point $\ensuremath{{\mathbf{i}}}$ is then
\begin{equation}
\label{eq:error}
\triangle n_{\ensuremath{{\mathbf{i}}}}^{k+1} = n_{\ensuremath{{\mathbf{i}}}}^{k+1} - \widetilde{n}_{\ensuremath{{\mathbf{i}}}}^{k+1}.
\end{equation}
We compute the $L_2$ error norm as
\begin{equation}
L_2\left(\triangle n^{k+1}\right) = \sqrt{\frac{\sum_{\ensuremath{{\mathbf{i}}}} \kappa_{\ensuremath{{\mathbf{i}}}}\left(\triangle n_{\ensuremath{{\mathbf{i}}}}^{k+1}\right)^2}{\sum_{\ensuremath{{\mathbf{i}}}} \kappa_{\ensuremath{{\mathbf{i}}}}\left(\widetilde{n}_{\ensuremath{{\mathbf{i}}}}^{k+1}\right)^2}},
\end{equation}
where $\kappa_{\ensuremath{{\mathbf{i}}}}$ is the volume fraction of a cut-cell (see \cite{Marskar2018} for details). We remark that $\triangle n^k$ is the step-wise error caused by inconsistent elliptic updates and not the accumulated error.
\begin{figure*}[h!t!b!]
\centering
\includegraphics[height=.8\textheight]{evolution_2d}
\caption{Two-dimensional streamer evolution with $\Delta t_c = \min\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_{S}\right)$ with radiative transport updates at $\Delta t_c$. The left hand side column shows the electric field magnitude $\left|\ensuremath{{\mathbf{E}}}\right|$ and the right hand side column shows the space charge density $\rho$. All numbers are in SI units and times are indicated in each frame. }
\label{fig:evolution_2d}
\end{figure*}
We use a fine spatial grid for all the simulations presented in this section. The grids use a base mesh of $(128)^2$ cells and includes four refinement levels; the refinement factor between each level is $4$ except for the last one where it is $2$. This yields an effective domain of $(16384)^2$, corresponding to a finest-level resolution of $1.2\,\ensuremath{{\mu\textrm{m}}}$.
\Fref{fig:evolution_2d} shows the time-evolution of a simulation that updates the elliptic equations at $\Delta t_c$ and restricts the time step according to
\begin{equation}
\Delta t_c = \min\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_{S}\right),
\end{equation}
where the fine step restriction is $\Delta t_f \leq 0.5\Delta t_\ensuremath{\textrm{cfl}}$. The equations are integrated for $2\,\ensuremath{{\textrm{ns}}}$ and evolve without obvious numerical errors such as spurious oscillations or numerical instabilities. This is true even though the number of advective integration steps per elliptic update ranged up to thirty, with an average of twenty advective steps per elliptic update throughout the simulation.
\begin{figure*}[h!t!b!]
\centering
\includegraphics{convergence_2d}
\caption{Error estimators for two-dimensional streamer simulation experiments. The various symbols show different time stepping criteria: If a symbol is empty (filled), the time step was restricted by $\Delta t_{S_e}$ ($\Delta t_{\ensuremath{{\mathbf{E}}}}$). The solid line shows the numerical error for a fully consistent scheme. a) Poisson and radiative transfer equations are updated at $\Delta t_c$. b) Radiative transport updates at $\Delta t_f$. c) Poisson updates at $\Delta t_f$.}
\label{fig:convergence_2d}
\end{figure*}
Next, to answer the question of when we need to update the electric field and radiative transport equations, we perform additional simulations by using smaller values of $\Delta t_c$ for each of the three cases listed in the preceding section. Specifically, we take $\Delta t_c = \alpha\min\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_{S_e}\right)$, where $\alpha = 0.1, 0.25, 0.5, 1.0$, and examine the numerical error for each case. The results of these comparisons are shown in \fref{fig:convergence_2d} for the electrons $n_e$. To aid in the interpretation of this error diagram, we mention that the error rise around $t=0.3\,\ensuremath{{\textrm{ns}}}$ coincides with the inception of the streamer; whereas the initial propagation phase takes place from $0.5\,\ensuremath{{\textrm{ns}}}$ and beyond. The various figures show the errors for the three different cases above. \Fref{fig:convergence_2d}a) shows the error norms when all elliptic equations are updated at $\Delta t_c$, corresponding to Case 1. All simulations are stable, although the numerical error can be significant in certain parts of the streamer evolution. For the propagation phase of the streamer, the error decreases to about $10^{-3}$, which is (in our opinion) a more acceptable error. For the avalanche phase, which occurs for $t<0.3\,\ensuremath{{\textrm{ns}}}$ in \fref{fig:convergence_2d}, the obtained errors are due to absence of consistently updated photoionization. However, the computed errors are smaller than $2\times10^{-4}$. \Fref{fig:convergence_2d}b) shows Case 2 where we consistently update the radiative transport equations. Differences between \fref{fig:convergence_2d}a) and \fref{fig:convergence_2d}b) are only seen in the avalanche phase $t<0.3\,\ensuremath{{\textrm{ns}}}$. In this evolution regime we have $\Delta t_{\ensuremath{\textrm{cfl}}} \ll \Delta t_{\ensuremath{{\mathbf{E}}}}$ so that electric field updates are redundant; solution errors occur due to consistent radiative transport updates, but even then the errors are quite small. For the streamer inception and propagation phases $t>0.3\,\ensuremath{{\textrm{ns}}}$, the error increases for both \fref{fig:convergence_2d}a) and b), but there is no notable quantitative difference between the computed errors. This error must therefore be due to inconsistent electric field updates. \Fref{fig:convergence_2d}c) correponds to Case 3 above, where radiative transport is updated inconsistently and the electric field is updated consistently. The initial errors for the avalanche phase $t<0.3\,\ensuremath{{\textrm{ns}}}$ correpond to the errors in \fref{fig:convergence_2d}a) and are due to inconsistent photoionization updates. For Case 3 the error is small throughout the entire evolution, showing that inconsistent photoionization updates leads to minor accuracy losses.
\begin{figure*}[h!t!b!]
\centering
\includegraphics{comparison}
\caption{Plots of the electric field magnitude and electron densities on the symmetry axis after integrating the equations for $2\,\ensuremath{{\textrm{ns}}}$ - corresponding to \fref{fig:convergence_2d}. The various rows show: a) Poisson and radiative transfer equations are updated at $\Delta t_c$. b) Radiative transport updates at $\Delta t_f$. c) Poisson updates at $\Delta t_f$. }
\label{fig:field_comparison}
\end{figure*}
To show the effect of the accumulated errors on the streamer for each of the nine simulations above, \fref{fig:field_comparison} shows the electric field and electron profiles on the symmetry axis at the end of the simulations. As before, \Fref{fig:field_comparison}a) shows Case 1, \fref{fig:field_comparison}b) shows Case 2, and \fref{fig:field_comparison}c) shows Case 3. We only find minor differences with respect to the initial velocities and peak field amplitude of the streamers. For the electron densities, \fref{fig:field_comparison} shows that omission of electric field updates may lead to significant variation of the predicted electron densities. For the worst resolved case in \fref{fig:field_comparison}a) which updates the electric field at the relaxation time, the predicted peak electron density is about $25\%$ higher than it should be. Likewise, \fref{fig:field_comparison}c) shows again that consistent updates of the radiative transport model do not lead to improved model accuracy.
\begin{table}[h!t!b!]
\centering
\begin{tabular}{c|c|c|c}
$\Delta t_c$ & Final $L_2(n_e)$ & \begin{tabular}{c} Poisson updates \\ (modified scheme) \end{tabular} & \begin{tabular}{c} Poisson updates \\ (original scheme) \end{tabular} \\
\hline
$\textrm{min}\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_S\right)$ & $2.43\times 10^{-1}$ & 133 & 8745 \\
$\frac{1}{2}\textrm{min}\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_S\right)$ & $1.10\times 10^{-1}$ & 237 & 8862 \\
$\frac{1}{4}\textrm{min}\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_S\right)$ & $4.09\times 10^{-2}$ & 450 & 9216 \\
$\frac{1}{10}\textrm{min}\left(\Delta t_{\ensuremath{{\mathbf{E}}}}, \Delta t_S\right)$ & $2.72\times 10^{-2}$ & 1095 & 9966
\end{tabular}
\caption{Total number of Poisson updates for the inconsistent (Case 1) and consistent schemes. }
\label{tab:elliptic_solves}
\end{table}
In summary, the preceding discussion shows that for the radiative transfer model considered here one does not need to compute the radiative transfer equations with a fine temporal resolution. Updates of the electric field should be evaluated based on a tradeoff between simulation time and accuracy need. For the model considered here, updating the electric field at $\Delta t_{\ensuremath{{\mathbf{E}}}}/10$ yields acceptable accuracy (a maximum step error of $10^{-3}$). We have also computed the final $L_2$ error norm for the electron density by computing error norms between the inconsistent (Case 1) solution and the consistent solution. These accumulated errors are reported in \tref{tab:elliptic_solves} where we also show the reduction of the number of elliptic solves for the inconsistent scheme. When we update the electric field at $\Delta t_{\ensuremath{{\mathbf{E}}}}/10$, the number of Poisson updates is reduced by approximately a factor nine, and the number of radiative transfer updates are reduced by a factor of eighteen. The price to pay for this is a $2.7\%$ increased error in the $L_2$ norm.
\section{A high performance computing example}
\label{sec:hpc_example}
As a large-scale computational example that uses some of these techniques, we next consider the inception and propagation of streamers in a needle-plane geometry. The domain is a $(2\,\ensuremath{{\textrm{cm}}})^3$ cube with a needle electrode with a $500\,\ensuremath{{\mu\textrm{m}}}$ radius protruding $1\,\ensuremath{{\textrm{cm}}}$ from the center of the top domain face (see \fref{fig:electron_initial}). This face, and the needle, is live with a voltage of $15\,\ensuremath{{\textrm{kV}}}$ whereas the bottom face is grounded. Homogeneous Neumann boundary conditions are used for the Poisson equation on the remaining domain faces. To demonstrate the use of spatially adaptive methods to streamers, we will consider initial conditions that provoke streamer branching and prevent prediction of their paths. Streamers are initiated by considering a stochastic preionization level with peak amplitudes up to $10^{14}\,\ensuremath{{\textrm{m}}}^{-3}$ where the initial density is prescribed by using a landscape function borrowed from computer graphics \cite{Perlin2002} which guarantees that the resulting noise is $C^1$ smooth (we use the original hash table in \cite{Perlin2002}). This function is then exponentiated in order to generate randomly placed plasma spots. The correlation length between these spots is roughly $500\,\ensuremath{{\mu\textrm{m}}}$. The initial electron distribution is shown in \fref{fig:electron_initial}.
\begin{figure*}[h!t!]
\centering
\includegraphics[width=0.6\textwidth]{electron_initial}
\caption{Initial preionization (in SI units) for the example simulation.}
\label{fig:electron_initial}
\end{figure*}
The simulation was run with a maximum time step $\Delta t_c = \textrm{Min}\left(\Delta t_\ensuremath{{\mathbf{E}}}, \Delta t_S\right)$ until $t=15\,\ensuremath{{\textrm{ns}}}$ with a time-to-solution of approximately \simtime\,hours. The electric field was updated at every advective evaluation whereas radiative transfer was updated at $\Delta t_c$ only (corresponding to Case 3 above). The simulation was run on 128 computing nodes interconnected with Infiniband. Each node contains dual-socket Intel Broadwell (E5-2683v4) chips with 32 cores per node in total. MPI ranks were mapped to cores with a one-to-one ratio for a total concurrency of 4096. For grids, we use a patch-based AMR grid generated by the Berger-Rigoutsos algorithm\cite{Berger1991} with a blocking factor of $16$ and a maximum patch size of $32$. This algorithm takes as input a number of ''tags'', which specify which cells should be refined, and the output of this algorithm is a properly nested hierarchy of Cartesian grids. The domain is discretized using this algorithm on a base mesh of $(256)^3$ cells with five levels of mesh refinement, which yields an effective domain of $(8192)^3$, corresponding to an effective resolution of $2.44\,\ensuremath{{\mu\textrm{m}}}$. This resolution is sufficient for numerical stability, but we remark that we have not performed grid convergence studies for this example. The refinement criteria that we use is based on resolving streamer heads and space charge layers \cite{Marskar2018}. The time step restriction on $\Delta t_f$ is done with $\Delta t_f \leq 0.5\Delta t_{\ensuremath{\textrm{cfl}}}$. The number of $\Delta t_f$ steps per coarse step $\Delta t_c$ was between $1$ and $20$, with an average of $2$ throughout the simulation. For an even finer spatial resolution, this ratio would be even higher, whereas this ratio would be 1:1 for a fully consistent scheme, which would increase the simulation time by approximately $40\%$.
\begin{figure*}[t]
\centering
\includegraphics[width=.45\textwidth]{rho_final}
\includegraphics[width=.45\textwidth]{mesh_final} \\[2em]
\includegraphics[width=.45\textwidth]{electrons_final}
\includegraphics[width=.45\textwidth]{source_final}
\caption{Final simulation state after $15\,\ensuremath{{\textrm{ns}}}$. a) Volume rendered space charge density. The data range has been adjusted to enhance visibility. b) Patch distribution at the finest level. Each brick represents a computational unit of minimum size $(16)^3$ and maximum size $(32)^3$. The ''inside'' of the brick agglomeration is empty (not shown). c) Volume rendered electron density. d) Volume rendered electron source term. All quantities are presented in SI units.}
\label{fig:solution}
\end{figure*}
\Fref{fig:solution} shows the final simulation state after $15\,\ensuremath{{\textrm{ns}}}$. The various subplots in the figure show snapshots of the electron density, space charge density, electron source term, and the mesh distribution on the finest level. Due to the perturbed initial conditions, we find that multiple streamers start from the electrode. During inception, we do not find uniform field screening over the needle tip (not shown) although this is something that we observe in simulations that do not use perturbed initial conditions. For this simulation, we observe five initial streamers, but two of these start from a position slightly higher up on the electrode and stop after a few nanoseconds. The other three filaments propagate into the gap. Their cross sections are generally not circular; for one of the branches the ratio between the major and minor radii is roughly $2$ so that the cross section is comparatively flat, which we believe is due to electrostatic repulsion from the other two nearby branches. Furthermore, this filament branches after approximately $10\,\ensuremath{{\textrm{ns}}}$, as seen in \fref{fig:solution}. Towards the end of the simulation, four individual streamer heads that propagate in different directions can clearly be observed, see e.g. \fref{fig:solution}d). The largest streamer propagates almost parallel with the rod, whereas the smallest streamer (see the branch in Fig.~\ref{fig:solution}) propagates almost perpendicular with it.
\section{Conclusion}
\label{sec:conclusions}
We have discussed recent advances in multiresolution computer models for streamer discharges. The use of longer time steps for the electric field and radiative transport leads to inherent numerical errors, which we quantify by means of two-dimensional simulations. We show that it is permissible to use a longer time step for the elliptic equations (e.g. Poisson) than for advection and chemistry, and believe that such techniques can be used to speed up simulation cases. The ideas can possibly also be extended to kinetic simulations. A high-performance computing example that uses some of these techniques was then presented.
\section*{Acknowledgements}
This work was financially supported by the Research Council of Norway through project 245422 and industrial partner ABB AS, Norway. The computations were performed on resources provided by UNINETT Sigma2 - the National Infrastructure for High Performance Computing and Data Storage in Norway.
|
\section{Introduction}
Cellular networks have enabled five billion people to connect to the Internet. This number is expected to rise further, with \ac{4G} and \ac{5G} standards meeting the high throughput and low latency requirements of mobile phones. In addition to this, it is estimated that billions of devices will also be directly connected to the Internet. The \ac{3GPP} has developed three new standards to address the needs of this \acf{IoT}: \ac{EC-GSM-IoT}, \ac{LTE} Cat-NB (\acs{NB-IoT}), and \ac{LTE} Cat-M (\acs{eMTC}) all provide better coverage and reduce both cost and power consumption for low-rate devices \cite{3gpp:rel13}.
The vast majority of \ac{cIoT} devices currently use \ac{GPRS}/\ac{EDGE} to connect to the Internet \cite{nokiaIoTConn}. For these, \ac{EC-GSM-IoT} provides an easy path to improved energy efficiency and a \SI{20}{\decibel} coverage improvement. \ac{EC-GSM-IoT} specifies \SI{164}{\decibel} of \ac{MCL} and an expected battery life of more than 10 years, matching the other \ac{cIoT} standards. Additionally, \ac{EDGE} support provides instantaneous global coverage and allows the throughput to scale up to \SI[per-mode=symbol,per-symbol = p]{355}{\kilo\bit\per\second}, comparable to half-duplex \ac{LTE} Cat-M1. Further, \ac{EC-GSM-IoT} is expected to have the lowest module cost of the \ac{3GPP} standards \cite{nokiaIoTConn}.
The problem of synchronizing to a legacy \ac{GSM} carrier has been extensively studied and several low-complexity solutions exist \cite{kroll2012low, jha2000acquisition}. For \ac{EC-GSM-IoT}, however, the sensitivity requirement is \SI{20}{\decibel} more stringent, and the signal level is now well below the thermal noise floor. The legacy algorithms are unable to cope with this, and new approaches are required to synchronize at \acp{SNR} as low as \SI{-8.5}{\decibel}.
In this paper, we present an algorithm, which is able to perform the \ac{EC-GSM-IoT} network synchronization down to \iac{SNR} of \SI{-11.5}{\decibel} in a low-band \ac{ST} channel. The support for frequency offsets of up to \SI{50}{\kilo\hertz} allows the use of a \SI{\pm 25}{\text{ppm}} crystal oscillator and, therefore, enables a low overall module cost. Compared to \cite{weber2017sawless}, the \ac{MF} boundary can be detected at \SI{-14.4}{\decibel} instead of \SI{-11}{\decibel} \ac{SNR} in \iac{ST} channel, and the average synchronization time at \SI{-8.5}{\decibel} \ac{SNR} has been reduced by \SI{56}{\percent} to \SI{1.3}{\second}. The proposed \ac{FOE} can accurately estimate the frequency offset at \SI{-8.5}{\deci\bel} \ac{SNR}.
The remainder of this paper is structured as follows: Section~\ref{sec:ec-gsm network sync} introduces the \ac{EC-GSM-IoT} \ac{MF}, the proposed synchronization procedure, the system model, and the synchronization requirements. The proposed algorithms are discussed in detail in Section~\ref{sec:algorithm}. Finally, the performance for synchronizing to a known carrier and the cell search are evaluated in Sections \ref{sec:performance} and \ref{sec:cell search}.
\section{\acs{EC-GSM-IoT} Network Synchronization}
\label{sec:ec-gsm network sync}
\subsection{\acl{MF} Structure}
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=\textwidth]{fig/mf_sync_flow_ts.pdf}
\caption[EC-GSM-IoT Multiframe]{Left: \acs*{EC-GSM-IoT} \acs*{MF}. In addition to the legacy \acsp*{FB} and \acsp*{SCH}, seven blind repetitions of the \acs*{EC-SCH} are present on \acs*{TS} 1 of the frames \numrange{0}{6}. \acp*{TS} \numrange{2}{7} are not shown, since they are not used for the synchronization. Right: Synchronization flow for the proposed algorithm.}
\label{fig:EC-GSM-IoT Multiframe}
\end{center}
\end{figure*}
Fig.~\ref{fig:EC-GSM-IoT Multiframe} shows the elements in the \ac{EC-GSM-IoT} \ac{MF}, which enable the network synchronization. This \SI{235}{\milli\second} long structure is repeated on the \ac{BCCH} carrier. It consists of \num{51} frames with eight \acp{TS} each. Five \acp{FB} are transmitted on \ac{TS} 0 of frames 0, 10, 20, 30, and 40. The legacy \ac{SCH} follows exactly one frame later and carries the current frame number and the \ac{BS} identity code. Devices may not be able to decode the \ac{SCH} below the noise floor, and can use the \ac{EC-SCH} with 28 blind repetitions over four \acp{MF} instead. The presence of the \ac{EC-SCH} indicates that the networks supports \ac{EC-GSM-IoT}.
\subsection{Synchronization Procedure}
The proposed algorithm performs the synchronization in four sequential steps as shown in Fig.~\ref{fig:EC-GSM-IoT Multiframe}. First, the \ac{MFD} finds a coarse estimate for the start time of the \ac{MF} and the frequency offset, using the \acp{FB}. After the frequency offset has been corrected for, the fine time offset is estimated using the \ac{EC-SCH} training sequences. At the same time, the decoding of the \ac{EC-SCH} is started. The fine \ac{FOE} can be started, as soon as the fine time offset is known. The device is ready to receive the \ac{EC-BCCH}, once the \ac{EC-SCH} has been successfully decoded, and may transmit, once the fine \ac{FOE} is completed.
\subsection{System Model}
We use the same noise figure as the \ac{3GPP} reference document: $\text{NF}=\SI{5}{\decibel}$ \cite{3gpp:45820}. The noise bandwidth is $\text{BW} = \SI{200}{\kilo\hertz}$ and the sampling frequency is $f_s = \SI{270.8}{\kilo\hertz}$. The receiver is powered on at a random time, and the initial frequency offset is uniformly distributed between $\pm f_{o, \max} = \pm R f_c$. $R$ is the error relative to the carrier frequency $f_c$. The simulations for the low bands were performed at $f_c = \SI{900}{\mega\hertz}$ and at $f_c = \SI{2}{\giga\hertz}$ for the high bands.
\acused{NF}
Since the oscillator is a significant factor in the system cost, it is desirable to use a low-cost crystal oscillator for \iac{cIoT} application. These typically have an initial accuracy of $R = \SI{25}{\text{ppm}}$. In the high bands, this corresponds to a maximum frequency offset of \SI{\pm 50}{\kilo\hertz}. This offset can be corrected by tuning the oscillator, or compensated for in the digital baseband.
\subsection{Synchronization Requirements}
A device is synchronized to the network, if it has successfully decoded the \ac{EC-SCH}, achieved timing synchronization, and corrected for the offset of the local oscillator. The requirements from the \ac{EC-GSM-IoT} standard are as follows \cite{3gpp:rel13}:
\begin{description}
\item[R1] The device must be able to synchronize to the \ac{BCCH} carrier within 2 seconds after power-on, which corresponds to successfully decoding the \ac{EC-SCH}.
\item[R2a] Before the first transmission, the timing offset with respect to the \ac{BS} must be below $\pm 1$ symbol periods.
\item[R2b] Before the first transmission, the relative frequency offset with respect to the \ac{BS} must be below \SI{\pm 0.1}{\text{ppm}}.
\end{description}
Requirement~R1 has to be met at the \ac{ISL} for reference performance of the \ac{EC-BCCH}, while Requirements R2a and R2b have to be met at the \ac{ISL} for reference performance of the \ac{EC-SCH}. These requirements differ for the low and high frequency bands, and are specified for three radio channels conditions: \ac{ST}, \ac{TU}1.2, and \ac{TU}50. The latter are fading with a mobile speed of \SI[per-mode=symbol]{1.2}{\kilo\meter\per\hour} and \SI[per-mode=symbol]{50}{\kilo\meter\per\hour}. The \ac{EC-SCH} and \ac{EC-BCCH} \ac{ISL} for the \ac{ST} channels correspond to \iac{SNR} of \SI{-8.5}{\decibel} and \SI{-7.5}{\decibel}. The \acp{SNR} for the low band \ac{TU}1.2 channels are \SI{-5.5}{\decibel} and \SI{-3.5}{\decibel} respectively. Since no maximum miss rate is specified for Requirement~R1, we will use the same as specified for the \ac{EC-SCH} \ac{BLER}: \SI{10}{\percent} misses.
The \SI{2}{\second} from Requirement~R1 equal approximately eight \acp{MF}. In the worst-case, it takes seven \acp{MF} to receive a complete set of 28 blind \ac{EC-SCH} repetitions. Therefore, the \ac{MFD} should usually only require a single \ac{MF}. If the \ac{MFD} takes more than four \acp{MF}, it is no longer possible to collect all 28 blind repetitions of the \ac{EC-SCH} within \SI{2}{\second}.
\section{Network Synchronization Algorithm}
\label{sec:algorithm}
\subsection{Fine Frequency Offset Estimation}
The fine \ac{FOE} is performed as the last step of the synchronization. This allows us to discuss it as a stand-alone problem, assuming perfect timing synchronization. The \ac{FOE} estimates the offset of the local oscillator compared to the \ac{BS} using the regularly transmitted \acp{FB}. Each \ac{FB} consists of a pure sine at $f_s /4 = \SI{67.7}{\kilo\hertz}$ for \num{148} symbol durations. The problem of estimating the frequency of a sine in \ac{AWGN} is a well-studied problem, and the \ac{CRLB} has been derived in \cite{rife1974single}. We have adapted it to the \ac{GSM} noise bandwidth:
\begin{equation} \label{eq:ml_crlb_b}
\text{Var}(\hat{\omega}) \geq \frac{6 \cdot f_s{}^3}{\text{SNR} \cdot \text{BW} \cdot N (N^2 -1)}. \\
\end{equation}
The \ac{ML} estimator for this problem selects the frequency with the maximum power spectral density $|A(f)|^2$:
\begin{equation} \label{eq:ml}
\hat{f} = \mathop{\arg\max}_f |A(f)|^2.
\end{equation}
It approaches the \ac{CRLB} for high \ac{SNR} and can be approximated using \iac{DFT} \cite{rife1974single}.
To fulfill Requirement~R2b at the minimum carrier frequency of \SI{869}{\mega\hertz}, the frequency has to be estimated with a maximum error of \SI{86}{\hertz}. At \iac{SNR} of \SI{-8.5}{\decibel}, the \ac{CRLB} for the \ac{RMS} error is \SI{182}{\hertz}, as is shown in Fig.~\ref{fig:FOE Combination Methods}. Clearly, this is not sufficient to achieve the required performance. The \ac{ML} estimator also suffers from a threshold effect, which renders it useless below \iac{SNR} of \SI{-6}{\decibel}.
The proposed algorithm combines the information from several \acp{FB} in order to improve the accuracy of the estimate. The phase of the \acp{FB} does not provide any information, since \ac{GSM} has no guaranteed phase coherency between frames. $N$ \acp{FB} are used by non-coherently accumulating the power of the individual \acp{FB} in the frequency domain:
\begin{equation} \label{eq:foe_comb_nc}
\hat{f} = \mathop{\arg\max}_f A'(f), \qquad A'(f) = \sum\limits_{i=0}^{N-1} |A_{i}(f)|^2.
\end{equation}
Our simulations show that this non-coherent combination achieves almost the same performance as a coherent combination with perfect knowledge of the phase.
The \ac{ML} estimator can be approximated using a two-step procedure to keep the implementation complexity low. In the first step a 256-point \ac{FFT} and the power for the frequency bins of interest are calculated. At this point, the discrete spectrum can be accumulated over multiple \acp{FB}. The maximum of these \ac{FFT} bins is then taken as the coarse location of the spectral peak. In a second step, a three-point interpolation is applied to produce a higher resolution estimate of the sine frequency. The algorithm from \cite{yang2011noniterative} calculates a correction term $f_a$ using only the power in the three \ac{DFT} bins, $Y_{-1}, Y_0,$ and $Y_1$, closest to the spectral peak:
\begin{equation} \label{eq:foe_rctsl}
f_{\alpha} = 2\pi \cdot \frac{|Y_1|^2 - |Y_{-1}|^2}{u(|Y_1|^2 + |Y_{-1}|^2) + v|Y_0|^2}.
\end{equation}
The original algorithm in \cite{yang2011noniterative} requires \iac{DFT} size of $2N$. We have modified the derivation of the constants $u$ and $v$ in order to allow for an arbitrary \ac{DFT} size, such that \iac{FFT} can be used:
\begin{align} \label{eq:foe_rctsl_l}
u & = \frac{sin^2 \left( \frac{\pi N}{2L} \right) \left[ \sin^2 \left( \frac{\pi N}{2 L} \right) \sin(\frac{\pi}{L})
- N \sin \left( \frac{\pi N}{L}\right) \sin^2 \left( \frac{\pi}{2 L} \right) \right]}
{\sin^2 \left( \frac{\pi}{2 L} \right) \left[ N^4 \sin^4 \left( \frac{\pi}{2L} \right) +
2 \sin^4 \left( \frac{\pi N}{2 L} \right) \right] }, \\
v &= u \cdot \frac{N^2 \sin^2 \left( \frac{\pi}{2 L} \right)}{\sin^2 \left( \frac{\pi N}{2 L} \right)}.
\end{align}
Fig.~\ref{fig:FOE Combination Methods} shows the simulated performance of the proposed \ac{FOE} algorithm in \iac{ST} channel, compared to the \ac{ML} estimator and the \ac{CRLB}. At the target \ac{SNR} of \SI{-8.5}{\decibel}, the \ac{RMS} frequency estimation error is \SI{36}{\hertz}, clearly meeting Requirement~R2b. Once the frequency of the sine has been estimated, it is also possible to estimate the amplitude of the sine from the same \ac{DFT} bins. This can be used to detect the presence of \iac{FB} and the \ac{MF} boundary.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/foe_comb.pdf}
\caption[FOE combination methods: RMS error]{\acs*{RMS} frequency offset improvement by considering 40 \acsp*{FB} in a static channel. The \acs*{CRLB} is shown twice, once for a single \acs*{FB} and once with a 40 times larger receive power. The ML estimator is approximated using a \SI{1}{\hertz} resolution \acs*{DFT}.}
\label{fig:FOE Combination Methods}
\end{center}
\end{figure}
\subsection{Multi-Frame Boundary Detection}
The \ac{MFD} is the first step in the synchronization process. Its main purpose is to estimate the start time of the \ac{MF}, and therefore the position of the \ac{EC-SCH}. One option is the use of \iac{ML} estimator, like the one proposed in \cite{kroll2017maximum} for \ac{NB-IoT}. However, the complexity of such an estimator scales with the maximum initial frequency offset and the number of correlated signals. For \ac{EC-GSM-IoT} the FCCH and the training sequences of the \ac{SCH} and \ac{EC-SCH} are known a-priori and can be used for the search. Based on \cite{kroll2017maximum}, we estimate the complexity for \ac{EC-GSM-IoT}: Three different signals need to be cross-correlated with, and the maximum frequency offset is three times larger than in \cite{kroll2017maximum}, resulting in a nine times larger real arithmetic complexity of \SI[per-mode=symbol]{2.5}{\giga\operation\per\second}. Storing the 9-bit correlation for \num{96} frequency and \num{63750} time offset candidates requires \SI{55}{\mega\bit} of memory. This is clearly not suited for a low-power and low-cost implementation and we therefore propose a reduced complexity \ac{MFD} algorithm.
The proposed \ac{MFD} algorithm processes the received data in sliding windows of 200 symbols, which are offset by 50 symbols. The \ac{FOE} procedure is re-used to look for the maximum spectral component in each window. The amplitude at this frequency, $C_{\text{FB}}[n]$, and the frequencies of the two highest components are stored. If the window contains \iac{FB}, there is a high chance that one of these two frequencies corresponds to the sine. Since \num{1275} windows, corresponding to one \ac{MF}, are considered, \num{3825} elements have to be stored in memory. Once an entire \ac{MF} worth of samples has been received, the algorithm accumulates the five \ac{FB} correlations for \iac{MF} start in each window:
\begin{equation} \label{eq:mf_corr}
C_{\text{MF}}[n] = \sum\limits_{i=0}^{4} C_{\text{FB}}[n + 250i \mod 1275].
\end{equation}
Fig.~\ref{fig:MFD: Correlation with the FBs} shows the two metrics in the noiseless case. Note that the normal bursts still result in a non-zero \ac{FB} correlation.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/mfd_corr_fb.pdf}
\caption[MFD: Correlation with the FBs]{MFD: Correlation with the \acsp*{FB} and resulting \acs*{MF} correlation in \iac*{ST} channel without noise. The \acs*{MF} starts at $\text{FN} = 0$.}
\label{fig:MFD: Correlation with the FBs}
\end{center}
\end{figure}
The stored frequencies for the maximum of $C_{MF}$ are then compared. If there is a set of frequencies with a spread below an empirical threshold of \SI{1}{\kilo\hertz}, the search is considered successful. Otherwise, additional \acp{MF} are searched and $C_{\text{MF}}$ is accumulated and the frequencies are updated. On a hit, the frequencies are also used as a coarse estimate of the frequency offset. At \iac{SNR} of \SI{-8.5}{\decibel} in \iac{ST} channel, the residual \ac{RMS} frequency offset is \SI{150}{\hertz}, which is sufficient to continue with decoding the \ac{EC-SCH}.
One of the main challenges for the \ac{MFD} is the structure of the \ac{GSM} \ac{MF}. The quasi-regular spacing of the \acp{FB} implies that the only difference between two \ac{MF} starts is the location of a single \ac{FB}. This results in a number of false side-peaks in the \ac{MF} correlation, as can be seen in Fig.~\ref{fig:MFD: Correlation with the FBs}. This is especially problematic for fast fading channels, like the TU50 test case. We propose the use of the \ac{EC-SCH} training sequences to find the actual start of the \ac{MF}, since they follow a different repetition scheme. To this end, the \ac{MFD} does not only return the peak of $C_{MF}$, but also the two positions 10 frames earlier and later as possible \ac{MF} start candidates. The correct candidate is later found using the \ac{EC-SCH} training sequences.
Fig.~\ref{fig:MFD: Spectrum method miss rate for the GSM channels} shows the performance of the Spectrum \ac{MFD} for the \ac{EC-GSM-IoT} channels in the low and high bands. In order to allow the \ac{EC-SCH} decoding sufficient time, the \ac{MFD} is terminated after searching 4 \acp{MF} and the most likely candidate is used.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/mfd_spec_miss_rate.pdf}
\caption[MFD: Spectrum method miss rate for the EC-GSM-IoT channels]{Spectrum \acs*{MFD} method miss rate for the \acs*{EC-GSM-IoT} channels. If none of the three candidates are within \num{\pm 80} symbols of the \acs*{MF} start, the attempt is considered \iac*{MFD} miss.}
\label{fig:MFD: Spectrum method miss rate for the GSM channels}
\end{center}
\end{figure}
The \ac{MFD} is the computationally most demanding part of the synchronization process. \num{5400} windows need to be processed every second, each requiring a 256-point \ac{FFT}. Including the calculation of the power and three-point interpolation, this results in an overall real arithmetic complexity of \SI[per-mode=symbol]{40}{\mega\operation\per\second}. This is suitable for a low-complexity hardware implementation and should result in a negligible power consumption, compared to the \ac{RF} receiver.
\subsection{Fine Time Offset Estimation and EC-SCH Decoding}
Once the \ac{MFD} is completed, the time offset is known to within \num{\pm 80} symbols. In a next step, the accuracy has to be improved to \num{\pm 1} symbol in order to decode the \ac{EC-SCH} and fulfill Requirement~R2a. This is done by cross-correlating the received \ac{EC-SCH} training sequences with the known symbols. The time offsets in the range of the residual time offset of \num{\pm 80} symbols are tried and the best match is selected. In a static channel at \iac{SNR} of \SI{-8.5}{\decibel}, the correct time offset is found in approximately \SI{60}{\percent} of all attempts. Blind repetitions of the training sequences are thus received and the correlation outputs $K_p[\eta]$ are combined non-coherently:
\begin{equation}
\hat{\eta} = \mathop{\arg\max}_{\eta} K'[\eta], \quad K'[\eta] = \sum\limits_p |K_p[\eta]|^2.
\end{equation}
Seven blind repetitions of the \ac{EC-SCH} are sufficient to determine the correct time offset at \iac{SNR} of \SI{-8.5}{\decibel} in \iac{ST} channel. The correct \ac{MF} start candidate is selected by choosing the one with the largest $K'[\hat{\eta}]$ after \num{28} blind repetitions have been received.
The last steps in the synchronization procedure are the fine frequency offset estimation and the decoding of the \ac{EC-SCH}, which can be performed simultaneously. The \ac{EC-SCH} decoding is started together with the fine time offset estimation and uses the most recent result of the latter. Seven repetitions of the \ac{EC-SCH} are transmitted at the start of every \ac{MF} and a total of 28 repetitions contain the same data. Repetitions in different \acp{MF} however, have different cyclic shifts, which depend on the frame number.
The actual decoding of the \ac{EC-SCH} is implemented, as described in \cite{weber2017sawless}, where the \acp{LLR} for the \ac{EC-SCH} repetitions are chase-combined for the four cyclic shift candidates. \ac{IQ}-combination can significantly improve the channel estimation, especially when short training sequences are used. However, the \ac{EC-SCH} uses longer training sequences and the phase between repetitions has to be estimated. Our simulations show that \ac{IQ}-combination does not improve the performance of the decoder due to the phase estimation error at \iac{SNR} of \SI{-8.5}{\decibel}.
One option to detect the cyclic shift is comparing the data received in two successive \acp{MF} using a cross-correlation with lag \num{1}. Unfortunately, the correlation output is dominated by noise in the target \ac{SNR} region, and this method fails. Alternatively, blind decoding attempts can be performed for all cyclic shifts, relying on the \ac{CRC} checksum to filter out the invalid results. But this method also has a flaw, because the decoding attempts on invalid data can result in false positives from the checksum. The problem is further complicated by the relatively short \ac{CRC} field length of 10 bits. If decoding is attempted after every received repetition, \ac{CRC} false positives occur up to \SI{19}{\percent} of the time. The proposed algorithm alleviates this problem by reducing the number of decoding attempts. In the low \ac{SNR} region, a decoding attempt is only performed after $7N, N \geq 4$ repetitions have been received. This reduces the maximum false positive rate to \SI{2.7}{\percent}. It is below \SI{1}{\percent} above \SI{-8.5}{\decibel} \ac{SNR}, which is acceptable and has to be handled after the synchronization.
\section{Performance Evaluation}
\label{sec:performance}
In order to evaluate the performance of the synchronization, we have performed simulations, where all the steps are performed in sequence. Fig.~\ref{fig:Synchronization miss rate} shows the resulting overall miss rate for all of the channels specified for \ac{EC-GSM-IoT}. \SI{90}{\percent} of the simulations succeed at \iac{SNR} of \SI{-11.5}{\decibel} in the low band \ac{ST} channel and at \iac{SNR} of \SI{-7.7}{\decibel} in the low band \ac{TU}1.2 channel. This exceeds the \ac{3GPP} goals by \SI{3}{\decibel} and \SI{2.2}{\decibel}, respectively. The \ac{CDF} for the synchronization time is shown in Fig.~\ref{fig:Synchronization time CDF}. It takes \SIrange{1.6}{1.8}{\second} to achieve \SI{90}{\percent} successful synchronizations, meeting Requirement~R1. The average synchronization time in \iac{ST} channel at \SI{-8.5}{\decibel} is \SI{1.3}{\second}, corresponding to a \SI{56}{\percent} reduction, compared to \cite{weber2017sawless}. The \ac{MFD} takes \SIrange{250}{425}{\milli\second} on average, the rest of the time is spent receiving the \ac{EC-SCH}. This is good from an energy standpoint, since the receiver can be turned off most of the time during the \ac{EC-SCH} reception, recalling the \ac{MF} structure in Fig.~\ref{fig:EC-GSM-IoT Multiframe}.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/sync_sim_bler_new.pdf}
\caption[Synchronization miss rate]{Synchronization miss rate for the different \ac*{EC-GSM-IoT} channels. A synchronization miss is either a timeout, or \iac*{CRC} hit with incorrect data. The initial oscillator error is uniformly distributed between \SI{\pm25}{\text{ppm}}.}
\label{fig:Synchronization miss rate}
\end{center}
\end{figure}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/sync_sim_time_cdf.pdf}
\caption[Synchronization time CDF]{Synchronization time \ac*{CDF} at the \ac*{ISL} for \ac*{EC-BCCH} reference performance, assuming \iac*{NF} of \SI{5}{\decibel}. Unsuccessful synchronizations are counted towards to total.}
\label{fig:Synchronization time CDF}
\end{center}
\end{figure}
Fig.~\ref{fig:Synchronization TO CDF} shows the fine time offset at the end of the synchronization process. It is below \num{\pm 1} symbol in more than \SI{99.5}{\percent} of the attempts for all the channels at the \ac{ISL} for \ac{EC-SCH} reference performance, which meets Requirement~R2a. As shown in Fig.~\ref{fig:Synchronization FO CDF}, the residual frequency offset after the fine \ac{FOE} over 40 \acp{FB} is below \SI{0.1}{\text{ppm}} in \SI{99}{\percent} of the test cases, which meets Requirement~R2b.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/sync_sim_toe_cdf.pdf}
\caption[Synchronization TO CDF]{Residual time offset \acs*{CDF} at the \acs*{ISL} for \acs*{EC-SCH} reference performance, assuming \iac*{NF} of \SI{5}{\decibel}.}
\label{fig:Synchronization TO CDF}
\end{center}
\end{figure}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=88mm]{fig/sync_sim_foe_cdf.pdf}
\caption[Synchronization FO CDF]{Residual frequency offset \ac*{CDF} at the \ac*{ISL} for \ac*{EC-SCH} reference performance, assuming \iac*{NF} of \SI{5}{\decibel}. Unsuccessful synchronizations are counted towards to total, but not shown.}
\label{fig:Synchronization FO CDF}
\end{center}
\end{figure}
\section{Cell Search}
\label{sec:cell search}
A cell search may have to be performed, whenever the device cannot reach \iac{BS} at a known frequency. Then, it has to scan all $298+673$ potential low and high band \ac{BCCH} carriers. Legacy \ac{GSM} devices look for a suitable carrier by performing receive power measurements, which are not feasible below the noise floor. The duration of the cell search becomes problematic in \ac{EC-GSM-IoT}, because the device must perform a partial synchronization to all potential carriers \cite{3gpp:gp160091}. Like other \ac{cIoT} standards, \ac{EC-GSM-IoT} trades synchronization time for an improvement in coverage. As in \cite{3gpp:gp160091}, we propose the use of the \ac{MFD} to detect the presence of \iac{GSM} carrier in extended coverage. Their algorithm requires 10 \acp{MF} to achieve a \SI{99}{\percent} detection rate in \iac{TU}\num{1.2} channel. Our proposed algorithm can achieve the same performance after searching only 4 \acp{MF}. However, the limiting case is the high band \ac{TU}50 channel with a maximum frequency offset of \SI{\pm 50}{\kilo\hertz}. In this case, 7 \acp{MF} are required to achieve a \SI{99}{\percent} detection rate. No false detections were simulated in \num{10000} test cases. With this \SI{30}{\percent} reduction compared to \cite{3gpp:gp160091}, the cell search takes \num{27} minutes and a maximum of \num{113} full cell searches are possible with a \SI{5}{\watt\hour} battery \footnote{Assuming a total receiver power consumption of \SI{100}{\milli\watt} \cite{3gpp:45820}. Note that the noise bandwidth definition in \cite{3gpp:gp160091} differs, and results in \iac{SNR} offset of \SI{1.3}{\decibel}.}.
\section{Conclusion}
We have presented the first complete set of algorithms, which is able to successfully synchronize at \iac{SNR} as low as \SI{-11.5}{\decibel}. It fulfills the requirements for the synchronization time and the residual time and frequency offset with a low computational complexity. Our simulations show that it exceeds the \ac{3GPP} target by \SI{3}{\decibel} in the best-case, resulting in \iac{MCL} of up to \SI{170.5}{\decibel}. In the worst-case the target is exceeded by \SI{2.2}{\decibel}, corresponding to \iac{MCL} of \SI{166.7}{\decibel}. The network synchronization supports a large frequency offset oscillator, enabling a low-cost \ac{EC-GSM-IoT} implementation.
\section*{Acknowledgment}
We would like to thank Innosuisse for their financial support through project 18987.1PFNM-NM.
\newpage
|
\section{Introduction}
Fix a base field $k$ of characteristic $0$ (or of a prime
characteristic $p$ which is larger than the dimension of the variety $X$ and
does not divide the size of the group $G$). All
varieties will be quasi-projective over $k$. We often take
$X^2=X\times X$ for $X$ a variety, or stack.
\par
In this paper we describe the multiplicative structure on the Hochschild cohomology $H\! H^\bullet([X/G])$ of the quotient stack of a smooth quasi-projective variety $X$ by a finite group $G$, which is assumed to act by automorphisms over $\operatorname{Spec} k$. We show first that the multiplication on Hochschild cohomology can be understood in terms of the geometry of $X$ as a $G$-variety, then proceed to discuss a number of applications to formality questions and orbifold cohomology. Such an object $[X/G]$ is also known as a global quotient orbifold.
\par
Our description of the cup product can be seen as a version of the Hochschild-Kostant-Rosenberg theorem for such
quotient stacks, taking into account the multiplicative structure. In
its simplest form, for smooth affine varieties $X$, this result states
that $H\! H^\bullet(X) \cong \wedge_{\mathscr{O}(X)}^\bullet T(X)$,
called polyvector fields on $X$~\cite{HKR} (as algebras), and it has
been generalized for example to the global setting in~\cite{yekutieli02,calaquevandenbergh10}. There are many applications
of understanding this together with its algebraic structure, such as
to deformation theory. In the case at hand of a global quotient orbifold $[X/G]$, the
Hochschild cohomology controls the deformation theory of the $G$-equivariant
geometry of $X$ (see Section~\ref{sect:remdefo} below).
\par
As a consequence of our description of the cup product we find that the multiplicative form of Kontsevich's formality theorem does not hold for smooth Deligne-Mumford stacks in general. We also observe a vector space identification between a variant of the Hochschild cohomology and the orbifold cohomology of $[X/G]$, in the particular instance in which $X$ is symplectic and $G$ acts by symplectic automorphisms. When $X$ is projective, in addition to being symplectic, we conjecture that there is furthermore an {\it algebra} isomorphism $H\! H^\bullet([X/G])\cong H^\bullet_\mrm{orb}([X/G])$. When the quotient variety $X/G$ admits a crepant resolution this conjectural algebra isomorphism reduces to Ruan's Cohomological Hyperk\"ahler Resolution Conjecture, and hence provides a resolution free variant of Ruan's conjecture.
We elaborate on our description of Hochschild cohomology, and subsequent applications, below. For the remainder of the introduction take $\mathscr{X}=[X/G]$, for $X$ a smooth quasi-projective variety and $G$ a finite group.
\subsection{The structure of Hochschild cohomology}
For us, the Hochschild cohomology of $\mathscr{X}$ will be the algebra of self-extensions
\[
H\! H^\bullet(\mathscr{X})=\mathrm{Ext}^\bullet_{\mathscr{X}^2}(\Delta_\ast\O_\mathscr{X},\Delta_\ast\O_\mathscr{X}).
\]
Motivated by methods from noncommutative algebra, we show that the cohomology can be given as an algebra of extensions between the pushforward $\Delta_\ast\O_X$ on $X^2$ and a certain smash product construction $\Delta_\ast\O_X\rtimes G$. Specifically, we explain below that the graded group of extensions $\mathrm{Ext}^\bullet_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X\rtimes G)$ is naturally a graded algebra with an action of $G$ by algebra automorphisms. We provide further an algebra identification $H\! H^\bullet(\mathscr{X})=\mathrm{Ext}^\bullet_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X\rtimes G)^G$ (see Section~\ref{sect:loc}).
\par
Our Hochschild cohomology with coefficients in the smash product has an obvious sheaf analog. For $\pi: X^2 \to X$ the first projection, we take
\[
\underline{H\! H}^\bullet(\msc{X})=\pi_* \mathscr{E}xt^\bullet_{X^2}(\Delta_* \O_X,\Delta_* \O_X\rtimes G),
\]
which we will refer to as the \emph{local equivariant Hochschild cohomology} of $\msc{X}$.
We will explain that the ``sheaf-valued'' cohomology $\underline{H\! H}^\bullet(\msc{X})$ is naturally a $G$-equivariant sheaf of graded algebras on $X$, i.e.\ a sheaf of algebras on $\msc{X}$ (see Remark~\ref{rem:intrinsic}), and that the usual local-to-global spectral sequence
\begin{equation}\label{eq:105}
E_2=
H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G\Rightarrow H\! H^\bullet(\mathscr{X})
\end{equation}
is multiplicative. By \cite{arinkinetal14}, this spectral sequence
degenerates at the $E_2$-page (see Section~\ref{ss:form-deg}). In this sense, the hypercohomology of $\underline{H\! H}^\bullet(\msc{X})$ provides an approximation of the Hochschild cohomology ring of the stack quotient.
\par
Our first main result can be summarized as follows. For every scheme $Y$, let $T^{poly}_{Y} := \bigwedge_{\O_Y} T_Y$ be the sheaf of polyvector fields on $Y$.
\begin{theorem}\label{thm:intro}
Let $G$ be a finite group acting on a smooth quasi-projective variety $X$, and consider the quotient orbifold $\msc{X}=[X/G]$. Then there is a canonical identification of equivariant sheaves
\begin{equation}\label{eq:166}
\underline{H\! H}^\bullet(\msc{X})=\bigoplus_{g\in G}T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g}).
\end{equation}
The induced algebra structure on the sum $\oplus_{g\in G} T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g})$ can be understood as follows:
\begin{enumerate}
\item[(i)] The identity component $T^{poly}_X$ is in the center of $\underline{H\! H}^\bullet(\msc{X})$.
\item[(ii)] As a $T^{poly}_X$-algebra, $\underline{H\! H}^\bullet(\msc{X})$ is generated by the determinants of the normal bundles $\oplus_{g\in G}\det(N_{X^g})$.
\item[(iii)] The sum of the determinants themselves form a $G$-graded, equivariant subalgebra in $\underline{H\! H}^\bullet(\msc{X})$.
\item[(iv)] Each multiplication
\[
m:\det(N_{X^g})\otimes_{\O_X} \det(N_{X^h})\to \det(N_{X^{gh}})
\]
is nonvanishing exactly on those components $Y$ of the intersection $X^g\cap X^h$ which are also components of $X^{gh}$, and along which $X^g$ and $X^h$ intersect transversely.
\end{enumerate}
Furthermore, there is a filtration on the global cohomology $H\! H^\bullet(\msc{X})$ under which we have an identification of algebras
\begin{equation}\label{eq:151}
\operatorname{gr} H\! H^\bullet(\mathscr{X})=H^\bullet\left(X,\oplus_{g\in G}\!\ T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g})\right)^G.
\end{equation}
\end{theorem}
In the unpublished manuscript~\cite{anno}, a version of the theorem above
appears in the affine setting (in terms of the obvious explicit
formula for cup product, rather than (i)--(iv); see Section~\ref{sect:altprod}). We note that the {\it vector space} identification~\eqref{eq:151} is due to Ginzburg-Kaledin in the affine case, and Arinkin--C\u{a}ld\u{a}raru--Hablicsek in general~\cite{arinkinetal14}.
\par
The subalgebra of (iii) is denoted $\underline{\mathrm{SA}}(\msc{X}):=\oplus_{g\in G}\det(N_{X^g})$ throughout. We see from (ii), or more precisely Theorem~\ref{thm:Tact} below, that this subalgebra provides all of the novel features of the Hochschild cohomology of the orbifold $\msc{X}$, as compared with that of a smooth variety. Furthermore, as the points of the fixed spaces $X^g$ account for the points of $\msc{X}$ which admit automorphisms, this algebra also represents a direct contribution of the stacky points of $\msc{X}$ to the Hochschild cohomology.
An explicit description of the multiplicative structure on the subalgebra $\underline{\mathrm{SA}}(\msc{X})$ is given in Theorem~\ref{thm:SA}, and the multiplicative structure on the entire cohomology $\underline{H\! H}^\bullet(\msc{X})$ is subsequently obtained from Theorem~\ref{thm:Tact}. The algebra identification~\eqref{eq:151} appears in Corollary~\ref{cor:arinkinetal_mult}.
\begin{remark}
Since $\underline{\mathrm{SA}}(\msc{X})$ is an equivariant sheaf of algebras on $X$, we may view it as a sheaf of algebras on the quotient $\msc{X}$. However, the reader should be aware that the algebra $\underline{\mathrm{SA}}(\msc{X})$ is not known to be an invariant of $\msc{X}$ at the moment. So a notation $\underline{\mathrm{SA}}(X\to \msc{X})$ may be more appropriate here.
\end{remark}
Of course, in the affine case the derived global sections vanish in positive degree. As a result, $\Gamma \underline{H\! H}^\bullet(\msc{X}) = H\! H^\bullet(\msc{X})$, and
we do not need a spectral sequence.
Hence Theorem~\ref{thm:intro}, in conjunction with the description of the algebra $\underline{\mathrm{SA}}(\msc{X})$ below, provides a complete description of the Hochschild cohomology in this case.
\begin{corollaryI}
Let $G$ be a finite group acting on a smooth affine variety $X$, and take $\mathscr{X}=[X/G]$. There is an algebra identification
\begin{equation}\label{eq:151-nogr}
H\! H^\bullet(\mathscr{X})=\Gamma\left(X,\oplus_{g\in G}\!\ T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g})\right)^G
\end{equation}
where the multiplication on the right hand side is as described above.
\end{corollaryI}
Note that there is no ``gr'' appearing in the corollary since the filtration is now trivial (it matches the cohomological grading).
The above corollary was established in work of Shepler and Witherspoon~\cite{SW} in the specific instance of a finite group acting linearly on affine space (cf.~\cite{anno}). We note that the description of cohomology given in the Corollary, as the global sections of our algebra of polyvector fields, is not exclusive to affine varieties. Indeed, we provide some projective examples in Section~\ref{sect:rem_deg} for which we have a similar equality~\eqref{eq:151-nogr}.
\begin{comment}
\subsection{The subalgebra $\underline{\mathrm{SA}}(\msc{X})$}
One can see from Theorem~\ref{thm:intro} that the subalgebra $\underline{\mathrm{SA}}(\msc{X}):=\oplus_{g\in G}\det(N_{X^g})$ essentially determines all of the novel behaviors of the cohomology of the orbifold $\msc{X}$, as compared with the cohomology of a smooth variety. We elaborate on this equivariant subalgebra here.
\par
Consider the maximal open subvariety $U=X-\cup_{g\in G} X^g$ of $X$ on which $G$ acts freely. We obtain an open embedding $U/G\to [X/G]$ which is an isomorphism when $U=X$. Hence, when all the fixed varieties vanish the quotient stack is a variety. Otherwise, each $k$-point in any of the fixed loci contributes a $k$-point of $\msc{X}=[X/G]$ with automorphisms, and we see that $\msc{X}$ is not a variety. In this way, the fixed loci $\{X^g:g\in G\}$ account for failure of $\msc{X}$ to be a variety.
\par
According to this information, the equivariant subsheaf
$\underline{\mathrm{SA}}^+(\msc{X}) = \bigoplus_{g \in G \setminus \{e\}} \det(N_{X^g})$ can be seen
as the contribution of the stacky points of $\msc{X}$ to the cohomology
$\underline{H\! H}^\bullet(\msc{X})$. The multiplication on
$\underline{\mathrm{SA}}(\msc{X})=\bigoplus_{g\in G}\det(N_{X^g})$ interprets the geometry of
the fixed loci as follows: Suppose for simplicity that all the fixed
loci $X^g$ are connected (and hence irreducible). Then we can inquire
as to when the following equality holds:
\begin{equation}\label{eq:163}
\mathrm{codim}(X^g)+\mathrm{codim}(X^h)=\mathrm{codim}(X^{gh}).
\end{equation}
Geometrically, the equality~\eqref{eq:163} is equivalent to the statement that $X^g$ and $X^h$ intersect transversely with $X^{gh}=X^g\cap X^h$.
\par
We show in Theorem~\ref{thm:SA} that the multiplication $m_{\underline{\mathrm{SA}}}$ on $\underline{\mathrm{SA}}(\msc{X})$ restricted to each pair of summands $\det(N_{X^g})$ and $\det(N_{X^h})$ has image in $\det(N_{X^{gh}})$, and that the restriction
\[
m_{\underline{\mathrm{SA}}}:\det(N_{X^g})\otimes_{\O_X}\det(N_{X^h})\to \det(N_{X^{gh}})\\
\hspace{5mm}\mathrm{is}\ \left\{\begin{array}{ll}
\mathrm{an\ }\cong &\mathrm{when~\eqref{eq:163}\ holds}\\
0 &\mathrm{else}.\end{array}\right.
\]
Furthermore, when~\eqref{eq:163} holds the isomorphism $m_{\underline{\mathrm{SA}}}$ is
induced by a corresponding decomposition of the normal bundle
$N_{X^{gh}}=N_{X^g}|_{X^{gh}}\oplus N_{X^h}|_{X^{gh}}$. These statements all
generalize to the case where $X^g$ need not be connected, by
considering the connected components of $X^g$ independently.
\end{comment}
\subsection{Formality, symplectic quotients, and orbifold cohomology}
We say that the Hochschild cohomology of a Deligne--Mumford stack $\msc{X}$ is formal if the sheaf-extensions $\msc{E}xt_\msc{X}^\bullet(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})$ and the derived endomorphisms $\mrm{R}\msc{H}om_\msc{X}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})$ are connected by a sequence of $A_\infty$ quasi-isomorphisms (see Section~\ref{sect:formality}). In characteristic zero, there is a highly nontrivial isomorphism between $T^{poly}_X$ and the derived endomorphisms of $\Delta_\ast\O_X$, for smooth $X$, using the composition of the Hochschild--Kostant--Rosenberg isomorphism with exterior multiplication by the square-root of the Todd class of $T_X$ (claimed on the level of global sections by Kontsevich~\cite[Claim, Section 8.4]{kontsevich03}, see also~\cite[Claim 5.1]{Cal-mp2}, and proved also by Calaque-Van den
Bergh~\cite[Corollary 1.5]{calaquevandenbergh10}; the local statement is a consequence of \cite[Theorem 1.1]{calaquevandenbergh10II}). We show that the analog of this result does not hold for projective Calabi--Yau quotient orbifolds with singular coarse space.
\begin{theoremI}[\ref{thm:informal}]
Suppose $X$ is projective and that $G$ is a finite subgroup in $\mrm{Aut}(X)$. Suppose additionally that the action of $G$ is not free, and that the quotient $\msc{X}=[X/G]$ is Calabi--Yau. Then the Hochschild cohomology of the orbifold $\msc{X}$ is not formal.
\end{theoremI}
One can construct such Calabi--Yau quotients via symplectic varieties equipped with symplectic group actions. Symplectic orbifolds feature prominently in the literature, particularly in studies of crepant resolutions and quantization.
\begin{corollaryI}[\ref{cor:symp_informal}]
Let $X$ be a projective symplectic variety and $G\subset \mrm{Sp}(X)$ be a finite subgroup for which the action is not free. Then the Hochschild cohomology of the corresponding orbifold quotient $\msc{X}$ is not formal.
\end{corollaryI}
Kummer varieties, i.e.\ quotients of abelian surfaces by the $\mbb{Z}/2\mbb{Z}$-inversion action, provide explicit examples of $\msc{X}$ as in the corollary, and hence explicit examples of smooth Deligne--Mumford stacks with non-formal Hochschild cohomology.
\par
We continue our analysis of symplectic quotients in Section~\ref{sect:symplectic}, where we provide a simplification of Theorem~\ref{thm:intro} in the symplectic setting over $\mbb{C}$. In this case one can produce simultaneous trivializions of the determinants of the normal bundles $\det(N_{X^g})$ which provide the following algebra identification.
\begin{theoremI}[\ref{thm:HHsymp}]
Suppose $X$ is a symplectic variety over $\mbb{C}$, and that $G$ acts on $X$ by symplectic automorphisms. Then there is an algebra isomorphism
\begin{equation}\label{eq:267}
\underline{H\! H}^\bullet(\msc{X})\cong \bigoplus_{g\in G} \Sigma^{-\mathrm{codim}(X^g)}T^{poly}_{X^g},
\end{equation}
where the (shifted) polyvector fields multiply in the expected manner.
\end{theoremI}
By the ``expected manner" we mean, vaguely, the multiplication one would expect given the vanishing result of Theorem~\ref{thm:intro} (iv) and the standard multiplication of polyvector fields. A precise description of this multiplication is given in the discussion following Theorem~\ref{thm:HHsymp}.
\par
We note one {\it cannot} achieve the above isomorphism by simply employing the trivializations $\det(N_{X^g})\cong \O_{X^g}$ provided by the symplectic form on $X$. One has to instead rescale these trivializations by the square roots of the determinants of the elements $(1-g)$, which we interpret as invertible operators on their respective normal bundles $N_{X^g}$. In Appendix~\ref{sect:pav}, with Pavel Etingof, we show in the particular case of a finite group acting linearly on a symplectic vector space that the rescaled trivializations produce the desired isomorphism~\eqref{eq:267}. For arbitrary $\msc{X}$ we reduce to the linear case to obtain the isomorphism~\eqref{eq:267}. We note that the introduction of these scaling factors corrects a persistent error in the literature (see Section~\ref{sect:remark}).
For a symplectic quotient $\msc{X}$, we also propose a number of relations between Hochschild cohomology and orbifold cohomology $H^\bullet_\mrm{orb}(\msc{X})$ for $\msc{X}$. In particular, we provide a linear identification between the Hochschild cohomology and the orbifold cohomology of a projective symplectic quotient orbifold at Corollary~\ref{cor:501}. We conjecture that there is furthermore an algebra identification between these two cohomologies.
\begin{conjectureI}[\ref{conj:HH_H_orb}]
For any projective, symplectic, quotient orbifold $\msc{X}$ over $\mbb{C}$ there is an algebra isomorphism $H\! H^\bullet(\msc{X})\cong H^\bullet_\mrm{orb}(\msc{X})$.
\end{conjectureI}
If the (generally singular) quotient variety $X/G$ admits a crepant resolution then our conjecture reduces to Ruan's Cohomological
Hyperk\"{a}hler Resolution Conjecture. (See Section~\ref{sect:orb_prod}.) As Ruan's conjecture relies on the existence of a crepant resolution, Conjecture~\ref{conj:HH_H_orb} represents a resolution free variant of the Cohomological
Hyperk\"{a}hler Resolution Conjecture.
\begin{remark}
Note that projective complex symplectic varieties (which pertain to the last two sections) are, from the differential geometric perspective, compact hyperk\"ahler manifolds.
This is due to Yau's solution to the Calabi conjecture~\cite{yau78}. The
converse also holds under our quasi-projectivity assumption.
\end{remark}
\subsection{Generalities for the Hochschild cohomology of algebraic stacks}
\label{sect:remdefo}
In principle, for a general algebraic stack $\msc{X}$, the Hochschild cohomology $H\! H^\bullet(\msc{X})$ defined via self-extensions of the pushforward $\Delta_\ast\O_\msc{X}$ need not be identified with the Hochschild
cohomology $H\! H^\bullet_\mrm{ab}(\mrm{Qcoh}(\msc{X}))$ of the category of quasi-coherent sheaves on $\mathscr{X}$, as defined
in~\cite{lowenvandenbergh05}. However, such an identification is desirable as the cohomology $H\! H^\bullet_\mrm{ab}(\mrm{Qcoh}(\msc{X}))$ has the correct theoretical interpretation, namely as the cohomology of the dg Lie algebra controlling the deformation theory of the category of sheaves on $\msc{X}$, while the self-extension algebra $H\! H^\bullet(\msc{X})$ is more accessible to computation.
\par
In an appendix by Pieter Belmans, Appendix~\ref{sect:cat_coho}, it is shown that for many reasonable stacks the algebra of self-extensions $H\! H^\bullet(\msc{X})$ does in fact agree with the Hochschild cohomology $H\! H^\bullet_\mrm{ab}(\mrm{Qcoh}(\msc{X}))$ of the abelian category $\mrm{Qcoh}(\msc{X})$, as desired.
\begin{propositionI}[\ref{prop:belmans}]
Let $\msc{X}$ be a perfect algebraic stack over
$\operatorname{Spec}k$. Then there exists an isomorphism of graded
algebras $H\! H^\bullet(\msc{X})\cong
H\! H^\bullet_\mrm{ab}(\mrm{Qcoh}(\msc{X}))$. In particular, we have
such an isomorphism for any global quotient orbifold $\msc{X}$.
\end{propositionI}
The above result implies that there is a natural graded Lie structure
on the self-extension algebra $H\! H^\bullet(\msc{X})$. Although we do
not pursue the topic here, one can attempt to construct a sufficiently
natural Lie algebra structure on the local Hochschild cohomology
$\underline{H\! H}^\bullet(\msc{X})$ and ask whether Kontsevich formality for the
{\it Lie} structure on the Hochschild cohomology of $\msc{X}$ is
similarly obstructed, as in Theorem~\ref{thm:informal}. However, we
expect that one has to at least push this sheaf
$\underline{H\! H}^\bullet(\msc{X})$ forward to the coarse space $X/G$ in order
to construct such an appropriate Lie structure. (Here, $X/G$ is
the categorical quotient scheme; in the case that $X$ is affine,
$X/G = \operatorname{Spec} \O(X)^G$.)
\begin{remark}
If instead of considering the Hochschild cohomology of
the abelian category, one works in the setting of derived algebraic
geometry and takes the Hochschild cohomology of the DG category
$\operatorname{QC}(\mathscr{X})$ of quasi-coherent sheaves on $\mathscr{X}$, then the
fact that one recovers self-extensions of $\Delta_\ast\O_\mathscr{X}$ is a direct consequence of
\cite[Theorem 1.2.(2)]{BFN-dag}, which identifies endofunctors of
$\operatorname{QC}(\mathscr{X})$ with quasi-coherent sheaves on $\mathscr{X}^2$, sending
the identity to $\Delta_\ast \O_\mathscr{X}$. For more details, see the last
two paragraphs of the proof of Proposition \ref{prop:belmans}.
\end{remark}
\subsection{References to related literature}
\begin{comment}
As stated previously, the sheaf identification~\eqref{eq:166} is due to Anno~\cite{anno}, and the vector space identification~\eqref{eq:151} appeared in work of Ginzburg--Kaledin and Arinkin--C\u{a}ld\u{a}raru--Hablicksek~\cite{ginzburgkaledin04,arinkinetal14}. For the case of a finite group acting linearly on affine space, the cup product on Hochschild cohomology was described explicitly in work of Shepler and Witherspoon~\cite{SW}, and an alternate formula for the cup product was given by Anno in the general affine setting~\cite{anno}. Our algebra $\underline{\mathrm{SA}}(\msc{X})$ is a generalization of the ``volume subalgebra" considered in~\cite{SW}. We note that the material presented here relies on~\cite{arinkinetal14} only to establish degeneration of the spectral sequence~\eqref{eq:105}.
\end{comment}
In the case of a finite group $G$ acting linearly on a vector space $V$ (considered here as an affine space) the Gerstenhaber structure on Hochschild cohomology $H\! H^\bullet([V/G])$ is understood. A concise algebraic description of the Gerstenhaber bracket can be found in work of the first author and Witherspoon~\cite{negronwitherspoon17} (see also~\cite{SW2,halbouttang10}). For more general orbifolds, even affine quotient orbifolds, an analogous description of the bracket has yet to appear.
\par
A general analysis of Hochschild homology and cohomology for orbifolds, from the perspective of the derived category, can be found in a sequence of papers by C\u{a}ld\u{a}raru and Willerton~\cite{caldararuI,Cal-mp2,caldararuwillerton10}.
\par
A $C^\infty$-analog of Theorem~\ref{thm:intro} appears in work of Pflaum, Posthuma, Tang, and Tseng~\cite{pflaumetal11}. In~\cite{pflaumetal11} the authors consider arbitrary ($C^\infty$-)orbifolds. Such objects may not be given as global quotients in general.
\subsection{Organization of the paper}
In Section~\ref{sect:rem_deg} we discuss some projective examples. Section~\ref{sect:bg} is dedicated to background material. In Section~\ref{sect:ss} we introduce the smash product $\Delta_\ast\O_X\rtimes G$ and establish the necessary relationships between the local Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$ and the cohomology of the stack quotient $\mathscr{X}$. In Section~\ref{sect:SPprelim} we establish some generic relations between normal bundles and tangent bundles for the fixed spaces, and in Sections~\ref{sect:A} and~\ref{sect:product} we use these relations to give a geometric description the Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$ as a sheaf of algebras on $\msc{X}$.
\par
We discuss formality and Calabi--Yau orbifolds in Section~\ref{sect:formality}. In Section~\ref{sect:symplectic} we describe the Hochschild cohomology of the quotient orbifold of a symplectic variety by symplectic automorphisms. Finally, in Section~\ref{sect:H_orb} we establish some linear identifications between Hochschild cohomology and orbifold cohomology, and provide a number of conjectures regarding the relationship between the cup product on Hochschild cohomology and the orbifold product on orbifold cohomology.
In Appendix~\ref{sect:cat_coho}, by Pieter Belmans, the Hochschild cohomology of the category $\mrm{Qcoh}(\msc{X})$ is identified with the derived endomorphisms of the pushforward $\Delta_\ast\O_\msc{X}$. In Appendix~\ref{sect:lin_alg} we cover some basic facts regarding linear representations of finite groups. In Appendix~\ref{sect:pav}, with Pavel Etingof, we show that a certain group cocycle which appears in the Hochschild cohomology of a symplectic quotient orbifolds is canonically bounded. This corrects a reccurring error in the literature.
\subsection{Acknowledgements}
Thanks to the participants and administrators of the Hochschild Cohomology in Algebra, Geometry, and Topology Workshop held at the Oberwolfach Research Institute for Mathematics in February 2016, where this project began in earnest. Thanks especially to Dmitry Kaledin and Damien Calaque for many helpful conversations, and to Pieter Belmans for helpful comments on a draft and pointing out references. We would like to thank Andrei C\u{a}ld\u{a}raru for many correspondences and useful clarifications, and Wendy Lowen and Michel Van den Bergh for explaining some aspects of categorical Hochschild cohomology. Thanks also to Heather Macbeth for offering many useful insights regarding symplectic geometry, to Allen Knutson for writing programs relevant to Remark \ref{r:other-hom}, to Haiping Yang for comments on this remark, and to Lie Fu for helpful comments on the Cohomological Hyperk\"{a}hler Resolution Conjecture.
\par
Thanks to the Hausdorff Institute for Mathematics in Bonn, Germany, where this project was continued during the Fall 2017 Junior Trimester Program on Symplectic Geometry and Representation Theory. The second author would also like to thank the Max Planck Institute for Mathematics for its hospitality on two research visits.
\section{Remarks on degeneration}
\label{sect:rem_deg}
Before we begin our main endeavor of computing Hochschild cohomology, let us make a few specific remarks on degeneration of the spectral sequence~\eqref{eq:105} and subsequent identification of the associated graded algebra of Hochschild cohomology.
\subsection{Examples of degeneration}
As mentioned in the introduction, for affine $X$, derived global sections vanish and we have $H\! H^\bullet(\msc{X})=\Gamma(X,\oplus_{g\in G}\!\ T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g}))^G$. Of course, such vanishing holds in many non-affine examples:
\begin{corollary}\label{cor-van}
Suppose that $G$ is a finite group acting on a smooth quasi-projective variety $X$ such that $H^i(X^g,
\wedge^j T_X \otimes \det(N_{X^g})
) = 0$ for $i > 0$ and all $j \geq 0$ and $g \in G$. Then there is an algebra isomorphism \eqref{eq:151-nogr}.
For instance, these conditions hold when $X=\mathbb{P}^n$.
\end{corollary}
The first assertion of this corollary is immediate; we sketch the
proof of the second assertion. We can assume that $G$ acts faithfully
on $X=\mathbb{P}^n$, so that $G < \operatorname{PGL}_{n+1}$. Every
element $g \in G$ has finite order, so is represented by a semisimple
element of $\operatorname{GL}_{n+1}$. Then $X^g$ is a disjoint union
of the projectivizations of the eigenspaces of $g$ in $k^{n+1}$.
Therefore $X^g$ is smooth and each connected component is
$\mathbb{P}^m \subseteq \mathbb{P}^n$. There is an isomorphism
$T_{\mathbb{P}^m} \oplus \mathscr{O}_{\mathbb{P}^m}(1)^{n-m} \to
T_{\mathbb{P}^n}|_{\mathbb{P}^m}$,
where the first component maps via the inclusion, and the second maps
by restricting the Euler surjection
$\msc{O}_{\mathbb{P}^n}(1)^{n+1} \twoheadrightarrow
T_{\mathbb{P}^n}$
to the last $n-m$ components and then restricting from $\mathbb{P}^n$
to $\mathbb{P}^m$. Thus $\det(N_{X^g}) = \mathscr{O}_{\mathbb{P}^m}(n-m)$. We therefore
have to show that all higher cohomology of $\wedge^j T_{\mathbb{P}^m}(n-m)$ vanishes.
This is well-known (see, e.g., \cite{Bot-hvb}).
We give an elementary proof using the Euler exact sequence
$0 \to \mathscr{O}_{\mathbb{P}^m} \to
\mathscr{O}_{\mathbb{P}^m}(1)^{m+1} \to T_{\mathbb{P}^m} \to 0$:
Replace $T_{\mathbb{P}^m}$ by the complex
$\mathscr{O}_{\mathbb{P}^m} \to \mathscr{O}_{\mathbb{P}^m}(1)^{m+1}$,
placing in degrees $-1$ and $0$, and take its $j$-th exterior
power. We get a complex in nonpositive degrees quasi-isomorphic to
$\wedge^{j} T_{\mathbb{P}^m}$. All terms of this complex are sums of
line bundles $\mathscr{O}_{\mathbb{P}^m}(\ell)$ for $\ell \geq 0$.
The same is true after twisting by
$\mathscr{O}_{\mathbb{P}^m}(n-m)$ with $n-m \geq 0$. Therefore taking derived global sections of the
complex results in a complex of vector spaces concentrated in
nonpositive degrees. So
$H^i(\mathbb{P}^m, \wedge^{j} T_{\mathbb{P}^m}(n-m))=0$ for $i \geq 1$.
\begin{remark}\label{r:other-hom}
It seems likely that the conditions of the corollary hold for some other homogeneous spaces. Let $G$ be a connected reductive group, $P<G$ a parabolic subgroup, and $X=G/P$. Then any finite subgroup $\Gamma < G$ acts by automorphisms on $G/P$.
Now let $g \in G$. In the case $g = 1 \in \Gamma$, so $X^g=X$, then
the relevant cohomology $H^i(X, \wedge^j T_X)$ is described in
\cite{Bot-hvb} (although making it explicit runs into the difficulty
that $\wedge^j T_X$ is not, in general, induced by a semisimple
representation of $P$; see also \cite[Exercise 4, Chapter 4]{Wey-cvbs}
for the vanishing for $i > 0$ in the Grassmannian case). Next consider an element
$g \in \Gamma$. Since it has finite order, it is semisimple, so its
centralizer $L:=Z_G(g)$ is reductive and contains a maximal torus $T$
of $G$. Replacing $P$ by a conjugate parabolic we may assume that
$T < P$. We have
$X^g = \{hP \in X \mid ghP=hP\}$.
The connected components of this are all isomorphic to $L^\circ/(L^\circ \cap P)$, where $L^\circ$ is the connected component of the identity of $L$. This is itself a homogeneous space of the same form, since $L^\circ \cap P$ is a parabolic subgroup of $L^\circ$.
Thus, one can again try to apply \cite{Bot-hvb} to each of these spaces.
We note, however, that the higher cohomology appearing in Corollary
\ref{cor-van} does not vanish for all homogeneous spaces (possibly not
even for most). Indeed, already for $g=1$ and $X$ the complete flag
variety ($G=\operatorname{GL}_n$ and $P=B$ a Borel subgroup), the
higher cohomology fails to vanish in general. For example, A.~Knutson
provided Macaulay2 code which demonstrates that for $n=6$, the Euler
characteristic of an isotypic component of $H^*(\wedge^4 T_{G/B})$
under the action of G is negative, and hence there exists a
nonvanishing cohomology group in odd degree. We have been informed
that P.~Belmans extended this using SageMath to flag varieties for
several other groups: in type $D_5$,
$H^{\text{odd}}(\wedge^k T_{G/B})$ does not vanish for
$k \in \{4,5,6\}$, and in type $F_4$,
$H^{\text{odd}}(\wedge^k T_{G/B})$ does not vanish for
$k \in \{3,4,5,6\}$. Moreover, ongoing work of P.~Belmans and
M.~Smirnov is investigating the case where $P$ is a maximal parabolic.
\end{remark}
\subsection{$\grH\! H^\bullet(\msc{X})$ versus $H\! H^\bullet(\msc{X})$ in general}
For general $\msc{X}$ there does not exist an algebra isomorphism $H\! H^\bullet(\msc{X})\cong \grH\! H^\bullet(\msc{X})$ (as in the above projective examples). Indeed, in our proof of the non-formality of Hochschild cohomology for projective Calabi--Yau quotient orbifolds (Theorem~\ref{thm:informal}) we observe an explicit obstruction to the existence of such an algebra isomorphism. So, for example, if we consider a projective symplectic quotient $\msc{X}$ with singular coarse space $X/G$ then we have explicitly $H\! H^\bullet(\msc{X})\ncong \grH\! H^\bullet(\msc{X})$. We refer the reader to Section~\ref{sect:formality}, and in particular the proof of Theorem~\ref{thm:informal}, for more details.
\par
We note that in the case of a projective symplectic quotient $\msc{X}$ we expect that the Hochschild cohomology $H\! H^\bullet(\msc{X})$ is strongly related to the orbifold cohomology of $\msc{X}$. The associated graded ring $\grH\! H^\bullet(\msc{X})$ in this case is expected to be the associated graded ring with respect to a natural ``codimension" filtration on orbifold cohomology. We discuss this topic in depth in Section~\ref{sect:H_orb}.
\section{Preliminaries}
\label{sect:bg}
Here we fix notations and record some basic
and well-known facts about fixed loci and resolutions which we will use throughout.
\subsection{Notations}
\label{sect:global_notes}
A {\it variety} is a quasi-projective scheme over $\operatorname{Spec} k$. For any variety $X$ we let $\O(X)$
denote the algebra of global functions. For
any variety map $\varphi:X\to Y$ we let
$\bar{\varphi}:\O_Y\to \varphi_\ast\O_X$ denote the underlying map of
sheaves of rings. By a vector bundle we will always mean a locally
free sheaf of {\it finite rank}.
\par
For a subvariety $i:Z\subset X$ and a quasi-coherent sheaf $M$ on $X$ we always let
$M|_Z$ denote the pullback $i^\ast M$ on $Z$, as opposed to $i^{-1}M$. At a geometric point $p:\operatorname{Spec}\bar{k}\to X$ we let $\O_{X,p}$ denote the corresponding local ring of the base change $X_{\bar{k}}$.
\par
Suppose $X$ is a variety with the action of a finite group $G$, which
we always assume to be an action by automorphisms of $X$ over
$\operatorname{Spec} k$. By applying the contravariant functor $\Gamma(X,?)$, the
action map $G\to \mathrm{Aut}_{\mathrm{sch}}(X)$ produces a map
$G^{op}\to \mathrm{Aut}_{\mathrm{alg}}(\O(X))$. Whence we get an action of
the opposite group $G^{op}$ on the global sections of $\O_X$, and an
action of $G$ by precomposing with the inversion operation on $G$.
Similarly, for any $G$-stable open $U$ we will get an action of $G$ on
$\Gamma(U,\O_X)$. For a group $G$ and $G$-representation $V$ in
general we often take ${^gv}=g\cdot v$, for $g\in G$ and $v\in V$.
Throughout $X$ will be a smooth (quasi-projective) variety equipped with
the action of a finite group $G$. We assume that the characteristic of
the base field $k$ does not divide $G$ and is greater than the
dimension of $X$.
We will often use the fact that such an $X$ admits a covering by
affines $\{U_i\}_i$ such that each $U_i$ is preserved under the action
of $G$. See~\cite[Appendix A]{mustata} or~\cite{grothendieck57}.
\par
To ease notation we omit the bullet from Ext and Hom notation, e.g.\ $\mathscr{E}xt_Y(M,N):=\mathscr{E}xt^\bullet_Y(M,N)$.
\subsection{Fixed locus under group actions}
We recall that a given variety $Y$ is smooth (over $k$) if and only if for any open affine subset $U\subset Y$ and any closed embedding $U\to \mathbb{A}^n_k$ with ideal of definition $\mathscr{I}$, the sequence
\[
0\to \mathscr{I}/\mathscr{I}^2\to \Omega_{\mbb{A}^n}|_U\to \Omega_U\to 0
\]
is split exact. A closed subvariety $Z\to Y$ in a smooth variety $Y$ is smooth if and only if the analogous sequence
\[
0\to \mathscr{I}_Z/\mathscr{I}_Z^2\to \Omega_Y|_Z\to \Omega_Z\to 0
\]
is exact with $\Omega_Z$ locally free.
\par
For any $g\in G$ we take $X^g$ to be the subvariety in $X$ fixed by $g$. This is the preimage of the graph of $g$ along $\Delta$, $X^g=\Delta^{-1}(\mathrm{graph}(g))$. Alternatively, it is the fiber product of $X$ with itself over $X^2$ along the two maps $\Delta:X\to X^2$ and $(1\times g)\Delta:X\to X^2$. For an arbitrary subset $S\subset G$ we let $X^S$ denote the intersection $\cap_{g\in S} X^g$ in $X$.
\begin{lemma}\label{lem:gbasics}
\begin{enumerate}
\item[(i)] Suppose $U$ is any open in $X$ with $g(U)=U$ for some $g\in G$. Then $U^g=X^g\cap U$.
\item[(ii)] Any point $p$ fixed by a subgroup $G'\subset G$ has an affine neighborhood $U$ with $g(U)=U$ for all $g\in G'$.
\item[(iii)] When $U$ is affine we have $\O(U^g)=\O(U)/(f-{^gf})_{f\in \O(U)}$.
\item[(iv)] The points of $X^g$ are all the points $p:\operatorname{Spec} k\to X$ with $g(p)=p$.
\end{enumerate}
\end{lemma}
\begin{proof}
(i) One simply shows that $X^g\cap U$, along with its inclusion into $U$, has the universal property of the fiber product. (ii) For an arbitrary affine neighborhood $W$ of $p$ we can take $U=\cap_{g\in G'}g(W)$. (iii) First note that $U^g$ will be affine, so we need only check that the map $\O(U)\to \O(U)/(f-{^gf})_{f\in \O(U)}$ has the appropriate universal property among maps to rings.
\par
For any algebra map $\phi:\O(U)\to B$ such that the two morphisms
\begin{equation}\label{eq:115}
\O(U)\otimes \O(U)\overset{1\otimes 1}{\underset{1\otimes \bar{g}}\rightrightarrows} \O(U)\otimes \O(U)\overset{\mathrm{mult}}\to \O(U)\overset{\phi}\to B
\end{equation}
agree we will have $\phi(a(f-\bar{g}(f)))=0$ for all $a,f\in \O(U)$. Replace $f$ with $-{^gf}=-\bar{g}^{-1}(f)$ to see that $\phi$ factors uniquely through $\O(U)/(f-{^gf})$. Conversely, for any map $\phi:\O(U)\to B$ factoring through the quotient we will have that the two maps~\eqref{eq:115} agree. So $\O(U)/(f-{^gf})$ has the appropriate universal property, $U^g=\operatorname{Spec}(\O(U)/(f-{^gf}))$, and $\O(U^g)=\O(U)/(f-{^gf})$. Claim (iv) follows by definition
(alternatively, one can reduce
to the affine case by (ii) and (i) and apply (iii)).
\end{proof}
\begin{lemma}\label{lem:smoothint}
For any subset $S\subset G$ the fixed locus $X^S$ is smooth. In particular, at any geometric point $p$ of the fixed space $X^S$ there is an isomorphism between the completed symmetric algebra $\hat{\mrm{Sym}}((T_p^\ast)_{S})$ generated by the $\langle S\rangle$-coinvariants of the tangent space at $p$ and the complete local ring $\hat{\O}_{X^S,p}$.
\end{lemma}
\begin{proof}
By changing base we may assume $k$ is algebraically closed. Note that $X^S=X^H$, where $H$ is the subgroup of $G$ generated by $S$. So we may also assume $S=G$. For smoothness, it suffices to show now that the complete local rings $\hat{\O}_{X^{G},p}$ at closed points $p$ on the fixed locus are power series rings~\cite{matsumura}.
\par
We may assume $X$ is affine. For any such point $p$, with corresponding maximal ideal $\mathfrak{m}_p\subset \O(X)$, we have the exact sequence $0\to \mathfrak{m}_p^2\to \mathfrak{m}_p\to \mathfrak{m}_p/\mathfrak{m}_p^2\to 0$ which admits a splitting over $G$, as $k[G]$ is semisimple. Any such splitting $T^\ast_p=\mathfrak{m}_p/\mathfrak{m}_p^2\to \mathfrak{m}_p$ gives compatible $G$-module algebra isomorphisms $\mathrm{Sym}(T^\ast_p)/(T^\ast_p)^n\to \O_{X,p}/\mathfrak{m}_p^n$ for each $n>0$, where $G$ acts linearly on the domain.
\par
Take $R_n=\mathrm{Sym}(T^\ast_p)/(T^\ast_p)^n$ and let $I=(f-{^gf})_{f\in \O(X),g\in G}$ to be the ideal of definition for $X^G$. Let $\bar{\mathfrak{m}}_p$ be the maximal ideal for the point $p$ in $X^G$. Then we have
\[
\frac{\mathrm{Sym}((T^\ast_p)_G)}{((T^\ast_p)_G)^n}=\frac{R_n}{(v-{^gv})_{v\in T^\ast_p,g}}=\frac{R_n}{(f-{^gf})_{f\in R,g}}\overset{\cong}\to \frac{\O(X)}{(I+\mathfrak{m}_p^n)}=\frac{\O(X^G)}{\bar{\mathfrak{m}}_p^n}
\]
and
\[
\hat{\mathrm{Sym}}((T^\ast_p)_G)\cong \varprojlim_n\O(X^G)/\bar{\mathfrak{m}}_p^n=\hat{\O}_{X^G,p}.
\]
Thus $X^G$ is smooth, and the complete local rings are as described.
\end{proof}
\subsection{Resolution property for stacks and the diagonal}
For a stack $\msc{Y}$ we say $\msc{Y}$ has the resolution property if any coherent sheaf $M$ on $\msc{Y}$ admits a surjection $E^0\to M$ from a vector bundle on $\msc{Y}$. The following lemma is well-known. One can see for example the comments of the introduction to~\cite{thomason87}.
\begin{lemma}\label{lem:resolprop}
Let $Y$ be a quasi-projective variety equipped with the action of a finite group $H$. Then the quotient stack $[Y/H]$ has the resolution property.
\end{lemma}
\begin{proof}
As in the proof of~\cite[Proposition IV.1.5]{knutson}, one can show that $Y$ admits a locally closed $H$-equivariant embedding $Y\to \mathbb{P}^n$ where $H$ acts linearly on projective space. Linearity of the $H$-action implies that the twist $\O(1)$ is an equivariant ample line bundle on $\mathbb{P}^n$, and the pullback $\O_Y(1)$ is therefore an equivariant ample line bundle on $Y$. Then by a slight variant of the usual argument (see e.g.~\cite[Corollary 5.18]{hartshorne}) one concludes that any equivariant coherent sheaf $M$ on $Y$ admits a equivariant surjection $E^0\to M$ from an equivariant vector bundle $E^0$ on $Y$.
\end{proof}
The resolution property allows us to derive the sheaf hom functor $\mathscr{H}om_{[Y/H]}$ via resolutions by vector bundles. Furthermore, when $Y$ is smooth, as will be the case for us, these resolutions can be taken to be bounded and the derived functor $\mrm{R}\mathscr{H}om_{[Y/H]}$ takes values in the derived category of coherent sheaves on $[Y/H]$. We will be applying this in the case $Y=X\times X$ and $H=G\times G$, so that $[Y/H]\cong \msc{X}\times\msc{X}$. We also include the following well-known result for the convenience of the reader.
\begin{lemma}\label{lem:aff_diag}
Let $\msc{Y}=[Y/H]$ be the quotient stack of a quasi-projective variety by a finite group action. Then the diagonal $\Delta:\msc{Y}\to \msc{Y}\times\msc{Y}$ is finite, affine, and separated.
\end{lemma}
\begin{proof}
By~\cite[\href{https://stacks.math.columbia.edu/tag/04XD}{tag 04XD}]{stacks-project} it suffices to show that the second projection $\msc{Y}\times_{\msc{Y}\times\msc{Y}}(Y\times Y)\to Y\times Y$ is finite, affine, and separated. But one calculates directly that the pullback $\msc{Y}\times_{\msc{Y}\times\msc{Y}}(Y\times Y)$ is isomorphic to the product $Y\times H$, and that the projection $Y\times H\to Y\times Y$ in this case is the coproduct of the graphs $(1\times h)\Delta:Y\to Y\times Y$. This map is indeed finite, affine, and separated.
\end{proof}
\section{The spectral sequence relating $\underline{H\! H}^\bullet(\msc{X})$ to $H\! H^\bullet(\msc{X})$} \label{sect:ss}
In this section we introduce the smash product sheaf $\Delta_\ast\O_X\rtimes G$ on $X^2$ and an equivariant dg algebra $\msc{E}(\msc{X})$ (defined up to quasi-isomorphism) for which the cohomology $H^\bullet(\msc{E}(\msc{X}))$ is the sheaf valued Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$. We apply results of~\cite{arinkinetal14} to find that $\msc{E}(\msc{X})$ is formal, as a sheaf, and subsequently deduce degeneration of the local-to-global spectral sequence~\eqref{eq:105}.
\subsection{The local algebras $\msc{E}(\msc{X})$ and $\underline{H\! H}^\bullet(\msc{X})$ on $X$}
\label{sect:loc}
Using descent, the local equivariant Hochschild cohomology and its relations to the Hochschild cohomology of $\msc{X}$ are explained as follows. Observe first that sheaves $\mathscr{F}$ on $\mathscr{X}$ are the same as $G$-equivariant sheaves $p^* \mathscr{F}$ on $X$ via the \'etale covering map $p:X \to \mathscr{X}$. We have the following cartesian diagram:
\begin{equation}
\xymatrix{
X \times G \ar[d] \ar[r]^{\widetilde \Delta} & X \times X \ar[d] \\
\mathscr{X} \ar[r]^{\Delta} & \mathscr{X} \times \mathscr{X},
}
\end{equation}
where $\widetilde \Delta(x,g) = (x,gx)$, i.e., $\Delta_g := \widetilde \Delta|_{X \times \{g\}}$ is the graph of $g: X \to X$.
Therefore, applying base change, we have
\begin{equation}
(p^* \times p^*) \mathrm{R}\mathscr{E}nd_{\mathscr{X}^2}(\Delta_*\O_{\mathscr{X}}) =
\mathrm{R}\mathscr{E}nd_{X^2}(\widetilde \Delta_* (\O_X \otimes \O_G)).
\end{equation}
We can identify $\widetilde \Delta_*(\O_X \otimes \O_G)$ with
the sheaf $\Delta_* \O_X \rtimes G$ (with $\O_{X^2}$ action given by the
usual $\O_X$-bimodule structure).
We can therefore study the Hochschild cohomology of $\mathscr{X}$ via the
$G \times G$-equivariant algebra $\mathrm{R}\mathscr{E}nd_{X^2}(\Delta_* \O_X \rtimes G)$
in the derived category of quasi-coherent sheaves on $X^2$. In particular, $H\! H^\bullet(\mathscr{X}) = \mrm{R}\Gamma(\mathscr{X}^2, \mathrm{R}\mathscr{E}nd_{\mathscr{X}^2}(\O_{\mathscr{X}})) = \mrm{R}\Gamma(X^2, \mathrm{R}\mathscr{E}nd_{X^2}(\Delta_* \O_X \rtimes G))^{G \times G}
= \mathrm{Ext}_{X^2}(\Delta_* \O_X \rtimes G,\Delta_* \O_X \rtimes G)^{G \times G}$. Note that we can
rewrite this as $\mathrm{Ext}_{X^2}(\Delta_* \O_X, \Delta_* \O_X \rtimes G)^G$, where the $G$ acts via the diagonal action.
On the derived level, we also have
$\mathrm{R}\mathscr{E}nd_{X^2}(\Delta_* \O_X \rtimes G) \cong \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \O_X, \Delta_* \O_X \rtimes G) \rtimes G$ in the derived category of quasi-coherent sheaves on $X^2$. Pushing forward under
the projection $\widetilde \pi: X^2 \to X \times X/G$
and taking invariants under $\{1\} \times G$ yields the $G$-equivariant algebra \begin{equation}
\msc{E}(\msc{X}) := \widetilde{\pi}_* \mathrm{R}\mathscr{E}nd_{X^2}(\Delta_* \O_X \rtimes G)^G \cong
\widetilde{\pi}_* \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \O_X, \Delta_* \O_X \rtimes G)
\end{equation}
in the derived category of quasi-coherent sheaves on $X \times X/G$. Note that
the map $X \to X/G$ is finite, and hence the pushforward $\widetilde{\pi}_*$ is underived.
Observe that the cohomology sheaves $H^\bullet (\msc{E}(\msc{X}))$ are scheme-theoretically supported on the image of the diagonal map $\overline{\Delta}: X \to X \times X/G$. Moreover their restriction to the diagonal is precisely $\underline{H\! H}^\bullet(\msc{X})$. We obtain the multiplicative spectral sequence \eqref{eq:105}.
Explicitly, the multiplication on $\msc{E}(\msc{X})$ can be
described as the sum over $g,h \in G$ of the bilinear maps
\begin{multline}\label{e:exp-mult}
\widetilde{\pi}_* \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \mathscr{O}_X, (\Delta_g)_*\mathscr{O}_X ) \times \widetilde{\pi}_* \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \mathscr{O}_X, (\Delta_h)_*\mathscr{O}_X ) \\ \to \widetilde{\pi}_* \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \mathscr{O}_X, \Delta_*(\Delta_{gh})_* \mathscr{O}_X )
\end{multline}
given by the identification
$\widetilde{\pi}_* \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \mathscr{O}_X, \Delta_*\mathscr{O}_X
\rtimes g) \cong \widetilde{\pi}_* \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \mathscr{O}_X \rtimes h,
\Delta_*\mathscr{O}_X \rtimes gh)$
from the diagonal action of $g$ followed by composition. Taking
cohomology sheaves and restricting to the diagonal yields the same statements for
$\underline{H\! H}^\bullet(\msc{X})$.
\begin{remark}
\label{rem:intrinsic}
We remark that the dg algebra $\msc{E}(\msc{X})$ is intrinsic to $\msc{X}$, as is the cohomology $\underline{H\! H}^\bullet(\msc{X})$. As noted above, sheaves on $\msc{X}\times X/G$ are equivalent to equivariant sheaves on $X\times X/G$, under the $G$-action on the first factor. Hence $\msc{E}(\msc{X})$ can be viewed as a sheaf of algebras on $\msc{X}\times X/G$. Indeed, $\msc{E}(\msc{X})$ is the pushforward of the derived endomorphism ring $\mrm{R}\mathscr{E}nd_{\msc{X}^2}(\Delta_\ast\O_\msc{X})$ along the finite map $\msc{X}^2\to \msc{X}\times X/G$. (Pushing forward from $\msc{X}$ to $X/G$ in the second factor automatically takes $\{1\}\times G$-invariants.) Since $X/G$ is the coarse space of $\msc{X}$, this construction of $\msc{E}(\msc{X})$ does not rely on the \'etale covering from $X$. Similarly, $\underline{H\! H}^\bullet(\msc{X})$ can be given as the pushforward of $\mathscr{E}xt_{\msc{X}^2}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})$ to $\msc{X}\times X/G$, or as the pushforward to $\msc{X}$ via the first projection.
\end{remark}
\subsection{Formality and degeneration}\label{ss:form-deg}
By \cite[Theorem 4.4]{arinkinetal14} (see also Theorem 4.9), the
object in the derived category of quasicoherent complexes on $X^2$,
\begin{equation}\label{eq:dsi}
(\Delta_* \O_X \rtimes G) \otimes^L_{\O_{X^2}} (\Delta_* \O_X \rtimes G)
\end{equation}
is formal. This object is $G \times G$-equivariant for the action of $G \times G$ on $X^2$. The second factor acts freely in the sense that, for $R^{(2)}_g: X \times X \to X \times X$ the action map on this factor, $R^{(2)}_g(x_1,x_2) = (x_1, g \cdot x_2)$, we have
\begin{equation}
(\Delta_* \O_X \rtimes G) \otimes^L_{\O_{X^2}} (\Delta_* \O_X \rtimes G)
\cong \bigoplus_{g \in G} (R_g)_* (\Delta_* \O_X \otimes^L_{\O_{X^2}} (\Delta_* \O_X \rtimes G)).
\end{equation}
Thus, it suffices to show that the following object is formal:
\begin{equation}
\Delta_* \O_X \otimes^L_{\O_{X^2}} (\Delta_* \O_X \rtimes G).
\end{equation}
We do this using coherent duality. For every scheme $Y$, let $\operatorname{pr}_Y: \to \operatorname{Spec} k$ be the
projection to a point. Let $\mathbb{D}(\mathcal{F}) :=
\mathrm{R}\mathscr{H}om_Y(\mathcal{F}, \operatorname{pr}_Y^! k)$. If $Y$ is smooth, then the (shifted) dualizing
complex $\operatorname{pr}_Y^! k$ equals $\omega_Y[\dim Y]$,
where $\omega_Y$ is the canonical line bundle. For $f: Y \to Z$ a proper
map, Grothendieck--Serre duality provides
a quasi-isomorphism $Rf_* \mathrm{R}\mathscr{H}om_Y(\mathcal{F}, f^! \mathcal{G}) \cong \mathrm{R}\mathscr{H}om_Z (Rf_* \mathcal{F}, \mathcal{G})$. We then have
\[
Rf_* \mathbb{D} \mathcal{F} := Rf_* \mathrm{R}\mathscr{H}om(\mathcal{F}, f^! \operatorname{pr}^! k)
\cong \mathrm{R}\mathscr{H}om(Rf_* \mathcal{F}, \operatorname{pr}^! k) =: \mathbb{D}(Rf_* \mathcal{F}).
\]
Thus $Rf_*$ preserves the class of self-dual objects under $\mathbb{D}$.
Now let us specialize to the case $Y=X, Z=X^2$.
Let $\Delta_g: X \to X \times X$,
the graph of $g: X \to X$. Then $\Delta_* \O_X \rtimes G = \bigoplus_{g \in G}
(\Delta_g)_* \O_X$, with each term self-dual under $\mathbb{D}$ (note that $R\Delta_* = \Delta_*$ since $\Delta$ is an embedding). Therefore,
\begin{equation}\label{e:selfdual-doxg}
\mathbb{D}(\Delta_* \O_X \rtimes G) \cong \Delta_* \O_X \rtimes G.
\end{equation}
Using the preceding identity in the second isomorphism, we obtain:
\begin{equation}
(\Delta_* \O_X \otimes^L_{\O_{X^2}} \mathbb{D}(\Delta_* \O_X \rtimes G))
\cong \mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \O_X, \mathbb{D} (\Delta_* \O_X \rtimes G)) \cong
\mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \O_X, \Delta_* \O_X \rtimes G),
\end{equation}
with the first isomorphism following from adjunction,
$\mathrm{R}\mathscr{H}om(\mathscr{E} \otimes \mathscr{F}, \mathscr{G}) \cong
\mathrm{R}\mathscr{H}om(\mathscr{E}, \mathrm{R}\mathscr{H}om(\mathscr{F}, \mathscr{G}))$.
Next, the cohomology sheaves of \eqref{eq:dsi}, which is just the sum
of derived intersections of graphs of elements $g \in G$, consists of
vector bundles on the fixed loci $X^g$, which are smooth closed
subvarieties of $X$. This is well-known (see, e.g., \cite[Proposition
A.5]{caldararuetal03}) and is also part of the statement of
\cite[Theorem 4.4]{arinkinetal14}. Then as above, the
dual of each cohomology sheaf is also a sum of vector bundles, the duals
to the ones before. Therefore
formality of \eqref{eq:dsi} implies formality of its dual,
$\mathrm{R}\mathscr{H}om_{X^2}(\Delta_* \O_X, \Delta_* \O_X \rtimes G)$.
Since the cohomology sheaves of $\msc{E}(\msc{X})$ are canonically isomorphic to $\overline{\Delta}_* \underline{H\! H}^\bullet(\msc{X})$ (for $\overline{\Delta}: X \to X \times X /G$ the diagonal map), we conclude that the spectral sequence \eqref{eq:105} is quasi-isomorphic to the
one computing the hypercohomology of $\overline{\Delta}_* \underline{H\! H}^\bullet(\msc{X})$. The quasi-isomorphism induces an isomorphism of bigraded vector spaces on the $E_2$-page, and for the hypercohomology of $\overline{\Delta}_* \underline{H\! H}^\bullet(\msc{X})$, we have $E_2=E_\infty$. Thus the same must be true for the spectral sequence computing the hypercohomology of $\msc{E}(\msc{X})$.
\section{Necessary relations between normal bundles and tangent sheaves}
\label{sect:SPprelim}
We cover some relations between the normal bundles for the fixed spaces $X^g$ and the restrictions of the tangent sheaf $T_X|_{X^g}$. These relations are required for our description of the cup product on the cohomology $\underline{H\! H}^\bullet(\msc{X})$. In particular, these relations ensure that our geometric description of the cup product is well-defined.
\subsection{Decompositions of the (co)tangent sheaf}
\label{sect:decomp}
For each $g\in G$ we consider the restriction $T_X|_{X^g}$. Note that $g$ now acts by sheaf automorphisms on $T_X|_{X^g}$. Since $X^g$ is a smooth embedded subvariety in $X$, we have the exact sequence
\[
0\to T_{X^g}\to T_X|_{X^g}\to N_{X^g}\to 0.
\]
We claim that this sequence identifies $T_{X^g}$ with the $g$-invariants in $T_X|_{X^g}$, i.e. the kernel of the endomorphism $(1-g):T_X|_{X^g}\to T_X|_{X^g}$. Hence $N_{X^g}$ is canonically identified with the image $(1-g)T_X|_{X^g}$. This fact can readily be checked complete locally, as we do below. It follows that we have canonical decompositions
\[
T_X|_{X^g}\cong T_{X^g}\oplus N_{X^g}\ \ \mrm{and}\ \ T^{poly}_X|_{X^g}\cong T^{poly}_{X^g}\otimes_{\O_{X^g}}(\wedge_{\O_{X^g}}^\bullet N_{X^g}).
\]
In particular, we have a canonical sheaf projection
\[
p_g:T^{poly}_X|_{X^g}\to T^{poly}_{X^g}\otimes_{\O_{X^g}}\det(N_{X^g}).
\]
(Cf.~\cite{anno} and~\cite[Sect.~1.4]{negronwitherspoon17}.) There is a similar decomposition for the cotangent sheaf $\Omega_X|_{X^g}\cong \Omega_{X^g}\oplus N^\vee_{X^g}$.
\par
Since we would like to make similar complete local reductions throughout the remainder of this document, let us provide an argument for the above claim in detail. The reader will not be harmed by skipping this on a first reading
and proceeding immediately to Section~\ref{sect:codimension}.
\par
We have the exact sequence $0\to N^\vee_{X^g}\to \Omega_X|_{X^g}\to \Omega_{X^g}\to 0$, which is locally split because $\Omega_{X^g}$ is locally free. Since $g$ has finite order, $\Omega_X|_{X^g}$ splits into eigensheaves $\Omega_1\oplus \Omega_1^\perp$, where $\Omega_1$ is the trivial eigensheaf for the action of $g$ and $\Omega_1^\perp$ is the sum of all non-trivial eigensheaves. Since the projection $\Omega_X|_{X^g}\to \Omega_{X^g}$ is $g$-linear, we have that $\Omega_1^\perp$ is in the kernel. Hence we have the proposed splitting $\Omega_X|_{X^g}\cong \Omega_{X^g}\oplus N^\vee_{X^g}$ if and only if $\Omega_{X^g}$ and $\Omega_1$ have the same rank. This property can be checked locally, or complete locally.
\par
At a geometric point $p$ of $X^g$ we take a $g$-linear section $T_p^\ast\to m_p$, where $m_p\subset \O_{X,p}$ is the maximal ideal, and let $V$ denote the image of this section. Note that $V$ is a subrepresentation in $\O_{X,p}$ for the action of the semisimple operator $g$. We are free to take a basis $\{x_1,\dots, x_n\}$ of $g$-eigenvectors for $V$.
\par
If $k$ is already algebraically closed, the elements $d_i=d_{x_i}$ provide a basis of $g$-eigenvectors for $\Omega_X|_{X^g}$ in an affine neighborhood around $p$. In general we may work complete locally to obtain a $g$-linear isomorphism
\begin{equation}\label{eq:cotang}
\hat{\O}_{X^g,p}\otimes V\overset{\cong}\longrightarrow \Omega_{\hat{X}_p}|_{\hat{X^g}_p},\ \ f\otimes x_i\mapsto fd_i.
\end{equation}
Dually, we have a $g$-linear isomorphism
\begin{equation}
\hat{\O}_{X^g,p}\otimes V^\ast\overset{\cong}\longrightarrow T_{\hat{X}_p}|_{\hat{X^g}_p},\ \ f\otimes x_i^\ast\mapsto fd_i^\vee
\end{equation}
where $d_i^\vee:\Omega_{\hat{X}_p}\to \hat{\O}_{X,p}$ is the dual function to $d_i$.
\par
From the above isomorphisms we see that the invariants $\Omega_1$ and $T_1$ are both of rank $\dim(V^g)$ around $p$. From Lemma~\ref{lem:smoothint} we see that $\Omega_{X^g}$ is of rank $\dim(V^g)=\dim(X^g)|_p$ around $p$ as well. Therefore
\[
\mrm{rank}(\Omega_{X^g})=\mrm{rank}(\Omega_1),\ \ \mrm{rank}(\Omega_1^\perp)=\mrm{rank}(N_{X^g}^\vee),\ \ \mrm{rank}(T_{X^g})=\mrm{rank}(T_1),\ \ \mrm{rank}(T_1^\perp)=\mrm{rank}(N_{X^g}).
\]
This implies that the inclusions $T_{X^g}\to T_X|_{X^g}$ is an isomorphism onto the invariants and, since $(1-g)T_X|_{X^g}=T_1^\perp$, the projection $T_X|_{X^g}\to N_{X^g}$ provides an isomorphism from the perpendicular
\begin{equation}\label{eq:793}
(1-g)T_X|_{X^g}\cong N_{X^g}.
\end{equation}
A similar result holds for the cotangent sheaf, and hence we have the proposed splittings.
\subsection{Codimensions and components of the fixed spaces}
\label{sect:codimension}
Recall that the codimension of a smooth subvariety $Y\subset X$ is a section of the sheaf of integers on $Y$, i.e. a continuous function $\mathrm{codim}(Y):Y\to \mathbb{Z}$. For sections of the sheaf of integers over a topological space we write $s_1<s_2$, $s_1\leq s_2$, etc. if the comparison holds pointwise. For any subspace $Z\subset Y$ we can restrict to get a section $\mathrm{codim}(Y)|_Z:Z\to Y\to \mathbb{Z}$ of the sheaf of integers over $Z$. Note that when $Z$ is connected $\mathrm{codim}(Y)|_Z$ is a constant.
\begin{lemma}\label{lem:codimineqs}
For any $g,h\in G$, and component $Y$ of $X^g\cap X^h$, we have
\[
\mathrm{codim}(X^{gh})|_Y\leq \mathrm{codim}(X^g\cap X^h)|_Y\leq \mathrm{codim}(X^g)|_Y+\mathrm{codim}(X^h)|_Y.
\]
If, additionally, $\mathrm{codim}(X^{gh})|_Y= \mathrm{codim}(X^g)|_Y+\mathrm{codim}(X^h)|_Y$ then $Y$ is also a component of $X^{gh}$.
\end{lemma}
\begin{proof}
The first inequality follows simply from the inclusion $X^g\cap X^h\subset X^{gh}$. The second inequality follows from the fact that we have the surjective map of vector bundles
\[
N_{X^g}^\vee|_{(X^g\cap X^h)}\oplus N_{X^h}^\vee|_{(X^g\cap X^h)}\to N_{X^g\cap X^h}^\vee.
\]
\par
Let $Z$ be a component of $X^{gh}$ containing $Y$. If we have an equality $\mathrm{codim}(X^{gh})|_Y= \mathrm{codim}(X^g\cap X^h)|_Y$ then, $Y$ is a codimension $0$ smooth embedded subvariety in $Z$. Since any component of smooth variety is irreducible this implies $Y=Z$. That is to say $Y$ is also a component of the intersection. Since we have the implication
\[
\begin{array}{l}
\mathrm{codim}(X^{gh})|_Y= \mathrm{codim}(X^g)|_Y+\mathrm{codim}(X^h)|_Y\\
\Rightarrow \mathrm{codim}(X^{gh})|_Y= \mathrm{codim}(X^g\cap X^h)|_Y
\end{array}
\]
the first inequality also implies that $Y$ is a component of $X^{gh}$.
\end{proof}
Below we describe the vanishing of the cup product on cohomology in terms of the inequality
\[
\mathrm{codim}(X^{gh})|_Y\leq \mathrm{codim}(X^g)|_Y+\mathrm{codim}(X^h)|_Y.
\]
\subsection{Relations between the normal bundles}
\label{sect:(co)normal}
Consider $g,h\in G$ and fix a component $Y$ of the intersection $X^g\cap X^h$. We then have the canonical map
\begin{equation}\label{eq:kappa}
\kappa:=\kappa(g,h,Y):N_{X^g}|_Y\oplus N_{X^h}|_Y\to N_{X^{gh}}|_Y
\end{equation}
whose restriction to each $N_{X^\tau}|_Y$, for $\tau=g,h$, is the inclusion $N_{X^\tau}|_Y\to T_X|_Y$ determined by the splitting of Section~\ref{sect:decomp} composed with the standard projection $T_{X}|_Y\to N_{X^{gh}}|_Y$.
\begin{lemma}\label{lem:esslem}
Suppose $Y$ is a component of the intersection $X^g\cap X^h$. The following are equivalent:
\begin{enumerate}
\item[(i)] $Y$ is a shared component with $X^{gh}$ and the intersection of $X^g$ and $X^h$ is transverse along $Y$.
\item[(ii)] $\mathrm{codim}(X^{gh})|_Y=\mathrm{codim}(X^g)|_Y+\mathrm{codim}(X^h)|_Y$.
\item[(iii)] The map $\kappa:N_{X^g}|_Y\oplus N_{X^h}|_Y\to N_{X^{gh}}$ of~\eqref{eq:kappa} is an isomorphism.
\item[(iv)] $N_{X^g}|_Y$ and $N_{X^h}|_Y$ have trivial intersection in $T_X|_Y$.
\end{enumerate}
\end{lemma}
This is a geometric analog of~\cite[Lemma 2.1]{SW}. We prove~\cite[Lemma 2.1]{SW} in finite characteristic in the appendix, Lemma~\ref{lem:dualrels}.
\begin{proof}
(i)$\Leftrightarrow$(ii): If either (i) or (ii) holds $Y$ is a shared component with $X^{gh}$. We consider the map of vector bundles $T_{X^g}|_Y\oplus T_{X^h}|_Y\to T_X|_Y$. The kernel of this map is $T_{(X^g\cap X^h)}|_Y=T_{X^{gh}}|_Y$, and one can check locally that the cokernel is a vector bundle on $Y$. By counting ranks, we see that transversality holds if and only if we have an equality
\[
\mathrm{rank}(T_{X^{gh}})|_Y= \mathrm{rank}(T_{X^g})|_Y+\mathrm{rank}(T_{X^h})|_Y-\mathrm{rank}(T_X)|_Y.
\]
This, in turn, occurs if and only if
\[
\begin{array}{rl}
\mathrm{codim}(X^{gh})|_Y&=\mathrm{rank}(T_X)|_Y-\mathrm{rank}(T_{X^{gh}})|_Y\\
&=\mathrm{rank}(T_X)|_Y-(\mathrm{rank}(T_{X^g})|_Y+\mathrm{rank}(T_{X^h})-\mathrm{rank}(T_X)|_Y)\\
&=\mathrm{rank}(T_X)|_Y-\mathrm{rank}(T_{X^g})|_Y+\mathrm{rank}(T_X)|_Y-\mathrm{rank}(T_{X^h})|_Y\\
&=\mathrm{codim}(X^g)|_Y+\mathrm{codim}(X^h)|_Y.
\end{array}
\]
Whence we have the desired equivalence between (i) and (ii).
\par
(ii)$\Leftrightarrow$(iv): For (iv), the vanishing of the intersection in $T_X|_Y$ is equivalent to the vanishing of the kernel $K$ in the sequence
\[
0\to K\to N_{X^g}|_Y\oplus N_{X^h}|_Y\to T_X|_Y.
\]
This property can be check complete locally at geometric points $p$ of $Y$, and hence can be reduced to the condition Lemma~\ref{lem:dualrels} (iii). The equivalent condition of Lemma~\ref{lem:dualrels} (ii) then assures as that at each such $p$ the proposed relation of codimensions holds. So we see that (iv) implies (ii), and we similarly find the converse implication.
\par
(iii)$\Leftrightarrow$(iv): The fact that (iii) implies (iv) is obvious, as the map $\kappa$ factors through $T_X|_Y$. Suppose now that (iv) holds, i.e.\ that the images of the normal bundles $N_{X^g}|_Y$ and $N_{X^h}|_Y$ intersect trivially in the tangent sheaf. Note that the image of $N_{X^g}|_Y\oplus N_{X^h}|_Y$ in $T_X|_Y$ lies in $(1-gh)T_X|_Y$ if and only if the the composition $N_{X^g}|_Y\oplus N_{X^h}|_Y\to T_X|_Y\to T_{X^{gh}}|_Y$ vanishes. Vanishing of this composition can be checked complete locally, and hence the inclusion $N_{X^g}|_Y+ N_{X^h}\subset (1-gh)T_X|_Y$ can be checked complete locally. Vanishing of the intersection is therefore reduced to the condition of Lemma~\ref{lem:dualrels} (iii), and the equivalent condition of Lemma~\ref{lem:dualrels} (ii) implies that $N_{X^g}|_Y+ N_{X^h}|_Y=(1-gh)T_X|_Y$. Since these normal bundles intersect trivially in $T_X|_Y$ the sum $N_{X^g}|_Y+ N_{X^h}$ remains direct. Finally, since $(1-gh)T_X|_Y$ is sent isomorphically to $N_{X^{gh}}|_Y$ under the projection from $T_X|_Y$, we see that $\kappa$ is an isomorphism.
\end{proof}
\section{The Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$ as an equivariant sheaf: Following~\cite{anno}}
\label{sect:A}
We recall an argument of Anno~\cite{anno}, which produces a canonical isomorphism of equivariant sheaves
\[
\underline{H\! H}^\bullet(\msc{X})\cong \oplus_{g\in G}T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g}).
\]
Note that the equivariant structure on the sum $\oplus_{g}T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})$ is induced by the isomorphisms $h:X^g\to X^{hgh^{-1}}$. To be precise, these isomorphisms fit into diagrams
\[
\xymatrix{
X\ar[r]^h & X\\
X^g\ar[u]^{\mrm{incl}}\ar[r]^h & X^{hgh^{-1}}\ar[u]_{\mrm{incl}},
}
\]
and we have induced isomorphisms $h_\ast T_{X^g}\cong T_{X^{hgh^{-1}}}$ and $h_\ast N_{X^g}\cong N_{X^{hgh^{-1}}}$ via conjugation by $h$, which produce the equivariant structure maps
\[
h_\ast(T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g}))\cong T^{poly}_{X^{hgh^{-1}}}\otimes_{\O_X}\det(N_{X^{hgh^{-1}}}).
\]
\subsection{The Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$ as an equivariant sheaf}
\label{sect:A1}
For any $g\in G$ we have the graph $\Delta_g:X\to X\times X$ of $g$ and extensions $\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,(\Delta_g)_\ast\O_X)$. We consider the projection $(\Delta_g)_\ast\O_X\to \Delta_\ast\O_{X^g}$, which is induced by the projection $\O_X\to \O_{X^g}$ over $X$ and the equality $(\Delta_g)_\ast\O_{X^g}=\Delta_\ast\O_{X^g}$. This provides a map on cohomology
\[
\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,(\Delta_g)_\ast\O_X)\to \mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_{X^g})=\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X)|_{X^g}=T^{poly}_X|_{X^g}.
\]
\par
The identification $\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_{X^g})=\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X)|_{X^g}$ employed above is induced by the affine local, natural, maps
\[
\begin{array}{rl}
\mathrm{Ext}_{\O(U^2)}(\O(U),\O(U))\otimes_{\O(U)}\O(U^g)\to &H^\bullet\left(\mrm{REnd}_{\O(U^2)}(\O(U))\otimes_{\O(U)}\O(U^g)\right)\\
&=H^\bullet\left(\mrm{RHom}_{\O(U^2)}(\O(U),\O(U^g)\right)\\
&=\mathrm{Ext}_{\O(U^2)}(\O(U),\O(U^g)).
\end{array}
\]
These local maps are isomorphisms since the cohomology $\mathrm{Ext}_{\O(U^2)}(\O(U),\O(U))$ is flat over $\O(U)$. Indeed, we have the classical HKR algebra isomorphism $\underline{H\! H}^\bullet(X)=\mathrm{Ext}_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X)=T^{poly}_X$ induced by the identification $\underline{H\! H}^1(X)=T_X$.
\par
We now have the composite map
\[
\mrm{A}_g:\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,(\Delta_g)_\ast\O_X)\to T^{poly}_X|_{X^g}\overset{p_g}\longrightarrow T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g}),
\]
where $p_g$ is as in Section~\ref{sect:decomp}. We recall that the determinant of the normal bundle $\det(N_{X^g})$ here lies in cohomological degree $\mrm{codim}(X^g)$. It is shown in~\cite{anno} that each $\mrm{A}_g$ is an isomorphism.
\begin{proposition}[{\cite[Proposition 5]{anno}}]\label{prop:anno}
For each $g\in G$, the map $\mrm{A}_g$ is an isomorphism of sheaves (supported) on $X^g$. Whence we have a canonical isomorphism
\[
\mrm{A}:\underline{H\! H}^\bullet(\mathscr{X})\overset{\cong}\to\bigoplus_{g\in G} T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g}).
\]
of $G$-equivariant sheaves (supported) on $X$.
\end{proposition}
We provide the proof for the reader's convenience.
\begin{proof}
As noted in~\cite{anno}, the fact that $\mrm{A}_g$ is an isomorphism can be checked complete locally at geometric points of $X^g=\mrm{Supp}\hspace{.5mm} \mathscr{E}xt_{X^2}(\Delta_\ast\O_X,(\Delta_g)_\ast\O_X)$.
\par
At any $g$-invariant geometric point $p$ of $X$, with corresponding maximal ideal $m_p\subset \O_{X,p}$, we choose a $g$-linear section $m_p/m_p^2\to m_p$ to deduce an isomorphism $\hat{\O}_{X,p}\cong \mrm{Sym}(m_p/m_p^2)$ of $g$-algebras. Hence by considering the complete localization $\hat{(\mrm{A}_g)}_p$ we reduce to the affine case, where $\mrm{A}_g$ is well-known to be an isomorphism (see~\cite{anno} or~\cite[Section 4.3]{negronwitherspoon17}).
\par
The fact that the total isomorphism $\mrm{A}$ is $G$-equivariant follows from the fact that the two constituent maps
\[
\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\oplus_g(\Delta_g)_\ast\O_X)\to \mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\oplus_g\Delta_\ast\O_{X^g})=\oplus_g T^{poly}_X|_{X^g}
\]
and
\[
\coprod_g p_g:\oplus_g T^{poly}_X|_{X^g}\to \oplus_g T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})
\]
are both $G$-equivariant (cf.~\cite[Lemma 4.3.2]{negronwitherspoon17}).
\end{proof}
As we proceed we will simply write an equality $\underline{H\! H}^\bullet(\mathscr{X})=\bigoplus_{g\in G} T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})$, by abuse of notation. Via degeneration of the local-to-global spectral sequence of Section~\ref{sect:ss} we recover the linear identification of Arinkin, C\u{a}ld\u{a}raru, and Hablicsek.
\begin{corollary}[\cite{arinkinetal14}]\label{cor:arinkinetal}
For $\msc{X}=[X/G]$ as above, there is an isomorphism of vector spaces
\[
\operatorname{gr} H\! H^\bullet(\mathscr{X})=H^\bullet\left(X,\oplus_{g\in G}\!\ T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g})\right)^G.
\]
\end{corollary}
\section{The cup product on $\underline{H\! H}^\bullet(\msc{X})$}
\label{sect:product}
We employ the identification $\mrm{A}:\underline{H\! H}^\bullet(\msc{X})\cong \oplus_g T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})$ of Proposition~\ref{prop:anno}, in conjunction with work of~\cite{SW}, in order to give a complete description of the cup product on local Hochschild cohomology.
\par
We prove below that the equivariant subsheaf $\underline{\mathrm{SA}}(\msc{X}):=\oplus_{g\in G}\det(N_{X^g})$ in $\underline{H\! H}^\bullet(\msc{X})$ is an equivariant subalgebra. As argued in the introduction, this subalgebra accounts for the points of $\msc{X}$ which admit automorphisms. It is also shown that the Hochschild cohomology of $\msc{X}$ is generated by the polyvector fields $T^{poly}_X\subset \underline{H\! H}^\bullet(\msc{X})$ and this distinguished subalgebra $\underline{\mathrm{SA}}(\msc{X})$.
\subsection{Technical points regarding complete localization and shrinking $G$}
We first recall some technical information which will be useful in our analysis. Let us consider a subgroup $G'$ in $G$, and the corresponding Hochschild cohomologies $\tilde{\pi}_\ast\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X\rtimes G')$ and $\tilde{\pi}_\ast\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X\rtimes G)$ as in Section~\ref{sect:SPprelim}. The split inclusion $i:\Delta_\ast\O_X\rtimes G'\to \Delta_\ast\O_X\rtimes G$ induces an inclusion
\[
i_\ast:\tilde{\pi}_\ast\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X\rtimes G')\to \tilde{\pi}_\ast\mathscr{E}xt_{X^2}(\Delta_\ast\O_X,\Delta_\ast\O_X\rtimes G),
\]
which we claim is an algebra map.
\par
In the ring theoretic setting, the fact that for any $G$-algebra $R$ the inclusion $H\! H^\bullet(R,R\rtimes G')\to H\! H^\bullet(R,R\rtimes G)$ is an algebra map is clear. Indeed, this is already clear at the level of the dg algebras of Hochschild cochains $C^\bullet(R,R\rtimes G')\to C^\bullet(R,R\rtimes G)$. Whence we see that $i_\ast$ is an algebra map by restricting to $G$-stable affines, as desired.
\par
We also recall that for coherent sheaves $M$ and $N$ on a smooth variety $Y$, and any geometric point $p$ of $Y$, the complete localization of the sheaf-extensions is canonically identified with the extensions of the complete localizations $\hat{M}_p$ and $\hat{N}_p$,
\[
\mathscr{E}xt_{Y}(M,N)^{\wedge}_p\underset{\mrm{canon}}\cong\mathrm{Ext}_{\hat{\O}_{Y,p}}(\hat{M}_p,\hat{N}_p).
\]
(See e.g.~\cite{hartshorne}.) To be clear, our completion at $p$ here is the completion of the pullback to $Y_{\bar{k}}$. This canonical isomorphism is one of algebras when $M=N$, and one of $\mathrm{Ext}_{\hat{\O}_{Y,p}}(\hat{N}_p,\hat{N}_p)-\mathrm{Ext}_{\hat{\O}_{Y,p}}(\hat{M}_p,\hat{M}_p)$-bimodules in general.
\par
Below, when we analyze the product on Hochschild cohomology complete locally at geometric points $p$ on the fixed spaces, we must first replace $G$ with the stabilizer $G_p$ of that point. This step is necessary so that $G$ actually acts on the completion $\hat{\O}_{X,p}$. The above argument assures us that there is no harm in making such local replacements.
\subsection{The action of $T^{poly}_X=\underline{H\! H}^\bullet(\msc{X})_e$ and generation}
For the moment, let us view the Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$ as a sheaf on $X\times X/G$ which is supported on the diagonal. We recall that the self-extension algebra $\mathscr{E}xt_{X^2}(\Delta_\ast\O_X\rtimes G,\Delta_\ast\O_X\rtimes G)$ is an $\O_{X^2}$-algebra which is supported on the union $\bigcup_{g\in G} \mrm{graph}(g)$, and hence the pushforward to $X\times X/G$ is supported on the diagonal $X\subset X\times X/G$. In particular, the algebra structure map
\[
\O_{X\times X/G} \to \underline{H\! H}^\bullet(\msc{X})
\]
factors through the projection $\O_{X\times X/G}\to \O_X$, where we abuse notation to identify $\O_X$ with its pushforward to $X\times X/G$. So we see that $\underline{H\! H}^\bullet(\msc{X})$ is naturally an $\O_X$-algebra. This is also clear from the description of the multiplication on $\underline{H\! H}^\bullet(\msc{X})$ given at~\eqref{e:exp-mult}. (The point here is that $\O_X$ is in the center of $\underline{H\! H}^\bullet(\msc{X})$.)
\par
By checking locally on $G$-stable affines, one readily verifies that the algebra
\[
\underline{H\! H}^\bullet(\mathscr{X})={\bigoplus}_{g\in G} T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})
\]
is graded by $G$ in the obvious way. That is to say, the multiplication on cohomology restricts to a map
\[
\underline{H\! H}^\bullet(\msc{X})_h\otimes_{\O_X}\underline{H\! H}^\bullet(\msc{X})_g\to \underline{H\! H}^\bullet(\msc{X})_{hg}
\]
for each $g,h\in G$, where $\underline{H\! H}^\bullet(\msc{X})_\tau$ is the summand $T^{poly}_{X^\tau}\otimes_{\O_X}\det(N_{X^\tau})$. In particular, $\underline{H\! H}^\bullet(\msc{X})_e$ is a subalgebra in the local Hochschild cohomology. The following lemma is easily seen by checking over $G$-stable affines.
\begin{lemma}
The identification $T^{poly}_X=\underline{H\! H}^\bullet(\msc{X})_e$, from Proposition~\ref{prop:anno}, is one of equivariant sheaves of algebras.
\end{lemma}
Let us consider the case where $h$, or $g$, is the identity. Here we have the inclusion $\mrm{det}(N_{X^g})\subset \underline{H\! H}^\bullet(\msc{X})_g$ and the identification $T^{poly}_X=\underline{H\! H}^\bullet(\msc{X})_e$, from which we restrict the multiplication to get a map
\[
T^{poly}_X\otimes_{\O_X}\det(N_{X^g})\to T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g}).
\]
Note that for any $g\in G$ we have an algebra projection
\begin{equation}\label{eq:1024}
T^{poly}_X\to T^{poly}_X|_{X^g}\to T^{poly}_{X^g}
\end{equation}
given by the counit of the pullback functor with the algebra map $T^{poly}_X|_{X^g}\to T^{poly}_{X^g}$ induced by the spitting $T_X|_{X^g}\cong T_{X^g}\oplus N_{X^g}$ of Section~\ref{sect:decomp}.
\begin{theorem}\label{thm:Tact}
The polyvector fields $T^{poly}_X$ are a (graded) central subalgebra in local Hochschild cohomology, so that $\underline{H\! H}^\bullet(\msc{X})$ is a $T^{poly}_X$-algebra. Furthermore, for each $g\in G$, the restriction of the multiplication
\[
\underline{H\! H}^\bullet(\msc{X})_e\otimes_{\O_X}\det(N_{X^g})=T^{poly}_X\otimes_{\O_X}\det(N_{X^g})\to T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})=\underline{H\! H}^\bullet(\msc{X})_g
\]
is the tensor product of the surjection~\eqref{eq:1024} with the identity on $\det(N_{X^g})$.
\end{theorem}
\begin{proof}
The fact that the multiplication factors through the reduction $T^{poly}_X\to T^{poly}_X|_{X^g}$, in the first coordinate, simply follows from the fact that the target sheaf is supported on $X^g$. Recall that the kernel of the projection $T^{poly}_X|_{X^g}\to T^{poly}_{X^g}$ is the ideal $\msc{I}$ generated by $N_{X^g}=(1-g)T_X|_{X^g}$. So we seek to show first that the multiplication map $T^{poly}_X|_{X^g}\otimes_{\O_X}\det(N_{X^g})\to T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})$ vanishes on $\msc{I}\otimes\det(N_{X^g})$, and that the restriction of the multiplication
\[
T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})\overset{\mrm{incl}}\longrightarrow T^{poly}_X|_{X^g}\otimes_{\O_X}\det(N_{X^g})\overset{\mrm{mult}}\longrightarrow T^{poly}_{X^g}\otimes_{\O_X}\det(N_{X^g})
\]
is the identity map. Both of these properties can be checked complete locally at points on $X^g$. Whence we may reduce to the affine case, as in the proof of Proposition~\ref{prop:anno}. Both the vanishing of the product on $\msc{I}\otimes_{\O_X}\det(N_{X^g})$ and the fact that the above composite is the identity are understood completely in the affine case~\cite{anno,SW}. Centrality of $T^{poly}_X$ in $\underline{H\! H}^\bullet(\msc{X})$ also follows by a complete local reduction to the linear setting.
\end{proof}
\begin{remark}
Theorem~\ref{thm:Tact} can also be deduced from the fact that the cohomology $\underline{H\! H}^\bullet(\msc{X})$ is a braided commutative algebra in the category of Yetter-Drinfeld modules for $G$. See~\cite{negron}.
\end{remark}
As an immediate corollary to Theorem~\ref{thm:Tact} one finds
\begin{corollary}
As a $T^{poly}_X$-algebra, the Hochschild cohomology $\underline{H\! H}^\bullet(\msc{X})$ is generated by the determinants of the normal bundles $\det(N_{X^g})$.
\end{corollary}
\subsection{The equivariant subalgebra $\underline{\mathrm{SA}}(\msc{X})$}
Recall that, since $X^g$ is sent isomorphically to $X^{hgh^{-1}}$ via the action of $h$ on $X$, the codimensions of $X^g$ and $X^{hgh^{-1}}$ agree. Rather, $\mrm{codim}(X^g)|_p=\mrm{codim}(X^{hgh^{-1}})|_{h(p)}$ at each point $p\in X^g$. Hence we see that the sum
\[
\underline{\mathrm{SA}}(\msc{X}):=\bigoplus_{g\in G} \det(N_{X^g})
\]
forms an equivariant subsheaf in $\underline{H\! H}^\bullet(\msc{X})$. Since each product $\det(N_{X^g})\otimes_{\O_X}\det(N_{X^h})$ is supported on the intersection $X^g\cap X^h$, one can describe the multiplication
\[
\det(N_{X^g})\otimes_{\O_X}\det(N_{X^h})\to \underline{H\! H}^\bullet(\msc{X})_{gh}
\]
by considering its restriction to the components of the intersection.
\begin{theorem}\label{thm:SA}
The equivariant subsheaf $\underline{\mathrm{SA}}(\msc{X})$ is a subalgebra in $\underline{H\! H}^\bullet(\msc{X})$. The multiplication
\[
m_{\underline{\mathrm{SA}}}:\det(N_{X^g})\otimes_{\O_X}\det(N_{X^h})\to \det(N_{X^{gh}})
\]
is non-vanishing only on those components $Y$ of the intersection $X^g\cap X^h$ which are shared components with $X^{gh}$, and along which $X^g$ and $X^h$ intersect transversely. On any such components $Y$, the multiplication $m_{\underline{SA}}$ is the isomorphism induced by the canonical splitting $\kappa:N_{X^g}|_Y\oplus N_{X^h}|_Y\cong N_{X^{gh}}|_Y$ of Lemma~\ref{lem:esslem} (iii), i.e.\ $m_{\underline{\mathrm{SA}}}=\det(\kappa)$.
\end{theorem}
\begin{proof}
Consider a geometric point $p$ in the intersection. By Lemma~\ref{lem:codimineqs} we have
\begin{equation}\label{eq:1085}
\mrm{codim}(X^{gh})|_p\leq \mrm{codim}(X^g)|_p+\mrm{codim}(X^h)|_p
\end{equation}
with equality holding if and only if the component $Y$ containing $p$ is shared with $X^{gh}$ and the intersection is transverse along $Y$. So we are claiming, first, that multiplication vanishes at each point $p$ at which~\eqref{eq:1085} is a strict inequality. At any such point $p$ we consider the complete localization $\hat{(m|_{\det(N_{X^g})\otimes_{\O_X}\det(N_{X^h})})}_p$ to reduce to the affine case, where the result is known to hold~\cite[Proposition~2.5]{SW}.
\par
We note that the equality $\mrm{codim}(X^{gh})|_p=\mrm{codim}(X^g)|_p+\mrm{codim}(X^h)|_p$ implies an equality of cohomological degrees
\[
\deg\big(\det(N_{X^{gh}})\big)|_p+\deg\big(\det(N_{X^{h}})\big)|_p=\deg\big(\det(N_{X^{gh}})\big)|_p
\]
in Hochschild cohomology. Since the cup product on $\underline{H\! H}^\bullet(\msc{X})$ is graded with respect to the cohomological degree, the vanishing result we have just established implies that the product on the Hochschild cohomology restricts to a product $m_{\underline{\mathrm{SA}}}$ on the equivariant subsheaf $\underline{\mathrm{SA}}(\msc{X})$. Whence we see that $\underline{\mathrm{SA}}(\msc{X})$ is an equivariant subalgebra in $\underline{H\! H}^\bullet(\msc{X})$.
\par
Now, at any point $p$ at which the inequality~\eqref{eq:1085} is an equality, we again reduce to the affine case via a localization $\hat{(m_{\underline{\mathrm{SA}}})}_p$ to see that the multiplication is in fact induced by the canonical decomposition $N_{X^g}|_Y\oplus N_{X^h}|_Y\cong N_{X^{gh}}|_Y$ of Lemma~\ref{lem:esslem} (see the proof of~\cite[Proposition~2.5]{SW}).
\end{proof}
\begin{remark}
Our algebra $\underline{\mathrm{SA}}(\msc{X})$ is a sheaf-theoretic variant of the ``volume subalgebra" $A_\mrm{vol}$ of~\cite{SW}.
\end{remark}
As remarked in the introduction, we can view $\underline{\mathrm{SA}}(\msc{X})$ as a sheaf on $\msc{X}$, or $\msc{X}\times X/G$. However, we are not certain that $\underline{\mathrm{SA}}(\msc{X})$ is an invariant of $\msc{X}$ as an orbifold. In particular, it is not immediately obvious how one should define this sheaf of algebras for orbifolds which are not global quotients.
\begin{question}
Does the algebra $\underline{\mathrm{SA}}(\msc{X})$, viewed as a sheaf of algebras on $\msc{X}$, have a natural interpretation as an orbifold invariant?
\end{question}
\subsection{Global descriptions}
One can readily combine Theorems~\ref{thm:Tact} and~\ref{thm:SA} to obtain a complete description of the algebra structure for $\underline{H\! H}^\bullet(\msc{X})$, as in Theorem~\ref{thm:intro} of the introduction. We observe what degeneration of the spectral sequence of Section~\ref{ss:form-deg} gives us, when paired with Theorems~\ref{thm:Tact} and~\ref{thm:SA}.
\begin{corollary}\label{cor:arinkinetal_mult}
The linear identification
\[
\operatorname{gr} H\! H^\bullet(\mathscr{X})=H^\bullet\left(X,\oplus_{g\in G}\!\ T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g})\right)^G.
\]
of Corollary~\ref{cor:arinkinetal} is an identification of graded algebras, where $\oplus_{g\in G}\!\ T^{poly}_{X^g}\otimes_{\O_X} \det(N_{X^g})$ has the equivariant algebra structure described in Theorems~\ref{thm:Tact} and~\ref{thm:SA}.
\end{corollary}
One can also see from Theorems~\ref{thm:Tact}/\ref{thm:SA} that the only obstruction to the Hochschild cohomology being commutative is the non-commutativity of $G$. Namely, there is an inequality of components $\underline{H\! H}^\bullet(\msc{X})_{gh}$ and $\underline{H\! H}^\bullet(\msc{X})_{hg}$ for non-abelian $G$. When $G$ is abelian, these two components agree and also $\det(N_{X^{gh}})=\det(N_{X^{hg}})$. Hence the generating subalgebra $\underline{\mathrm{SA}}(\msc{X})$ is commutative in this case, and it follows that entire cohomology $\underline{H\! H}^\bullet(\msc{X})$ is commutative as well. We record this observation.
\begin{corollary}
When $G$ is abelian, the cohomology $\underline{H\! H}^\bullet(\msc{X})$ is (graded) commutative.
\end{corollary}
\subsection{An alternate expression of the cup product}
\label{sect:altprod}
One can alternatively introduce a ``product" on local Hochschild cohomology via the composite
\begin{equation}\label{eq:altprod}
\xymatrix{
\underline{H\! H}^\bullet(\msc{X})_g\otimes_{\O_X}\underline{H\! H}^\bullet(\msc{X})_h\ar[r] & (T^{poly}_X|_{X^g})\otimes_{\O_X}(T^{poly}_X|_{X^h})=T^{poly}_X\otimes_{\O_X}T^{poly}_X|_{(X^g\cap X^h)}\ar[dl]|{\mrm{mult}}\\
T^{poly}_X|_{(X^g\cap X^h)}\ar[r]_(.35){P_{gh}} & T^{poly}_{X^{gh}}\otimes_{\O_X}\det(N_{X^{gh}})= \underline{H\! H}^\bullet(\msc{X})_{gh},
}
\end{equation}
where the first map is as in Section~\ref{sect:A1} and $P_{gh}$ is the projection $p_{gh}$ on those components $Y$ of $X^g\cap X^h$ which are shared components with $X^{gh}$, and vanishes elsewhere. One can in fact argue that the above sequence does provide a well-defined, associative, algebra structure which agrees with the cup product. Indeed, one can make a direct comparison between the two products via the equivalent conditions (iii) and (iv) of Lemma~\ref{lem:esslem}. The expression of the cup product via~\eqref{eq:altprod} provides a certain (re)reading of the last sentence of~\cite{anno}, and an alternate expression of~\cite[Theorem 3]{anno}.
\section{Formality and projective Calabi--Yau quotients}
\label{sect:formality}
We say the Hochschild cohomology of a Deligne--Mumford stack $\msc{X}$ is formal, as an algebra, if there is a sequence of quasi-isomorphism of $A_\infty$-algebras
\[
\msc{E}xt_{\msc{X}^2}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})\overset{\sim}\longrightarrow \mbb{M}_1\leftarrow \cdots\to\mbb{M}_r\overset{\sim}\longleftarrow \mrm{R}\msc{H}om_{\msc{X}^2}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X}),
\]
where the $\mbb{M}_i$ are $A_\infty$-algebras in the category of sheaves of vector spaces on (the small \'{e}tale site of) $\msc{X}$, with the tensor product over the base field.
\par
Formality for smooth varieties in characteristic zero was essentially claimed by Kontsevich~\cite[Claim 8.4]{kontsevich03}, and proved by Tamarkin for affine space~\cite{tamarkin98}. Tamarkin's proof was extended to the non-affine setting by Calaque and Van den Bergh~\cite{calaquevandenbergh10II}. We show here that this formality result does {\it not} extend to Deligne--Mumford stacks in general. Namely, we prove that certain Calabi--Yau orbifolds are not formal. Here symplectic varieties feature prominently, in examples and more in depth discussions, and we continue an analysis of symplectic quotient orbifolds in subsequent sections.
\begin{remark}
One can adopt a weaker version of formality, which proposes that a sheaf of dg algebras $\msc{R}$ on $\msc{X}$ is formal if there is an isomorphism of algebras $H^\bullet(\msc{R})\cong \msc{R}$ in the derived category of sheaves of vector spaces on $\msc{X}$ (cf.~\cite{calaquevandenbergh10}). Our method of proof shows that this weaker form of formality for Hochschild cohomology is also obstructed in general.
\end{remark}
\begin{remark}
The formality results of~\cite{kontsevich03,tamarkin98,calaquevandenbergh10,calaquevandenbergh10II} are much stronger than what is stated here.
\end{remark}
\begin{comment}
\subsection{Some results of~\cite{bridgelandkingreid01} and~\cite{caldararuI}}
In{bridgelandkingreid01} the authors consider a global quotient stack $\msc{X}$ equipped with a crepant resolution $Y\to X/G$ of the corresponding coarse space $X/G$. They prove the following result, among others.
\begin{theorem}[{[Corollary 1.3]{bridgelandkingreid01}}]
Suppose $X$ is a complex symplectic variety and $G$ acts by symplectic
automorphisms. Assume that $Y$ is a crepant resolution of $X/G$, and take $\msc{X}=[X/G]$. Then there is an equivalence $\Phi:\mrm{D}^b(Y)\to \mrm{D}^b(\msc{X})$ provided by an explicit integral transform.
\end{theorem}
By an integral transform, we mean a functor $\Phi$ We will also need some work of C\u{a}ld\u{a}raru{caldararuI} (see also{caldararuwillerton10,orlov03}).
\begin{theorem}[{[Theorem\ 8.1]{caldararuI}}]
Let $\msc{X}$ and $\msc{Y}$ be projective Deligne--Mumford
stack with Serre duality. Suppose $\msc{X}$ and $\msc{Y}$ admit a derived equivalence given by an integral transform. Then the Hochschild cohomologies admit a corresponding isomorphism $H\! H^\bullet(\msc{X})\cong H\! H^\bullet(\msc{Y})$ of graded rings.
\end{theorem}
\end{comment}
\subsection{Failure of formality for projective Calabi--Yau quotients}
Let us consider a Calabi--Yau triangulated category $\msc{D}$ over $k$ of dimension $d$~\cite{keller08}. So, $\msc{D}$ is a Hom-finite triangulated category for which the shift $\Sigma^d$ provides a Serre functor. Serre duality in this case can be expressed via a global, non-degenerate, degree $-d$, functorial pairing
\[
\mrm{tr}_{a,b}:\mathrm{Ext}_\msc{D}(a,b)\times \mathrm{Ext}_\msc{D}(b,a)\to k,
\]
for arbitrary objects $a,b$ in $\msc{D}$, and $\mathrm{Ext}_\msc{D}(a,b)=\oplus_{n\in \mbb{Z}}\mathrm{Hom}_\msc{D}(a,\Sigma^nb)$. In particular, each self-extension algebra $\mrm{Ext}_\msc{D}(a,a)$ is a graded Frobenius algebra.
\par
For $\msc{Z}$ a proper Calabi--Yau Deligne--Mumford stack, the derived category $\mrm{D}^b(\mrm{Coh}(\msc{Z}))$ is a Calabi--Yau category of dimension $d=\mrm{dim}(\msc{Z})$. We can construct examples of such projective Calabi--Yau $\msc{Z}$ as quotient stacks as follows: Consider a projective Calabi--Yau variety $Z$ equipped with an action of a finite group $\Pi$ such that $\omega_Z\cong \O_Z$ as a $\Pi$-equivariant sheaf. In this case the Grothendieck-Serre duality maps $\mathrm{Ext}^i_Z(M,N)\to \mathrm{Ext}^{d-i}_Z(N,M)^\ast$, for $\Pi$-equivariant $M$ and $N$, are $\Pi$-linear, by functoriality. Since taking $\Pi$-invariants gives equivariant maps $\mathrm{Ext}_Z^i(M,N)^\Pi=\mathrm{Ext}_{[Z/\Pi]}^i(M,N)$ we may take $\Pi$-invariants of Serre duality for $Z$ to obtain Serre duality for $\msc{Z}=[Z/\Pi]$. Whence the quotient $\msc{Z}=[Z/\Pi]$ is Calabi--Yau.
\par
We may consider specifically $Z$ a projective symplectic variety with $\Pi$ acting by symplectic automorphisms. In this case $\Pi$-invariance of the symplectic form implies that the isomorphism $\O_Z\cong \omega_Z$ provided by the corresponding symplectic volume form is an equivariant isomorphism.
\par
Since the product $\msc{Z}=\msc{X}\times \msc{X}$ of a Calabi--Yau Deligne--Mumford stack $\msc{X}$ with itself is again Calabi--Yau we find the following result.
\begin{lemma}
If $\msc{X}$ is a Deligne--Mumford stack which is proper and Calabi--Yau, then the Hochschild cohomology $H\! H^\bullet(\msc{X})$ is a graded Frobenius algebra.
\end{lemma}
We can now present the desired result.
\begin{theorem}\label{thm:informal}
Suppose $X$ is a projective variety and that $G$ is a finite subgroup in $\mrm{Aut}(X)$. Suppose additionally that the action of $G$ is not free, and that the quotient $\msc{X}=[X/G]$ is Calabi--Yau. Then the Hochschild cohomology of $\msc{X}$ is not formal.
\end{theorem}
\begin{proof}
In this case $H\! H^\bullet(\msc{X})$ is Frobenius, by the previous lemma. The fact that $G$ does not act freely implies that we have nonempty fixed spaces $X^g$ of positive codimension. Take $\msc{E}=\msc{E}xt_{\msc{X}^2}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})$ and $\msc{R}=\mrm{R}\msc{H}om_{\msc{X}^2}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})$.
\par
Let us suppose, by way of contradiction, that there is an isomorphism $\msc{E}\to \msc{R}$ of algebras in the derived category of sheaves of vector spaces on $\msc{X}$. Then we get an induced isomorphism on derived global sections, and subsequent algebra isomorphism on cohomology
\begin{equation}\label{eq:2026}
H^\bullet(\msc{X},\msc{E})=H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G\cong H\! H^\bullet(\msc{X}).
\end{equation}
(The first equality in the above equation follows from the material of Section~\ref{sect:ss}.) We claim, however, that $H^\bullet(X,\underline{H\! H}^\bullet(X))^G$ admits no graded Frobenius structure, and hence that no such isomorphism~\eqref{eq:2026} exists.
\par
Indeed, such a graded Frobenius structure would be given by a linear map from top degree
\[
f_\mrm{top}:\left(H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G\right)^{2\mrm{dim}(X)}\to k
\]
so that the Frobenius form is the product composed with this function, $\langle a,b\rangle=f_\mrm{top}(ab)$. But now, by our explicit description of the cohomology $\underline{H\! H}^\bullet(\msc{X})$ given in Section~\ref{sect:product}, we see that $H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G$ admits a natural algebra grading by codimensions, and the top degree occurs only in the codimension $0$ component
\[
\left(H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G\right)^{2\mrm{dim}(X)}=H^{\dim(X)}(X,\omega^\vee_X)^G.
\]
It follows that any such pairing on $H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G$ necessarily vanishes when restricted to any positive codimension component, and must therefore be degenerate, which is a contradiction. So $H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))^G$ admits no graded Frobenius structure, no such algebra isomorphism~\eqref{eq:2026} exists, and the Hochschild cohomology of $\msc{X}$ cannot be formal.
\end{proof}
Of course, there exist many $\msc{X}$ satisfying the hypotheses of Theorem~\ref{thm:informal}. We can consider symplectic quotients for example. More specifically, one can consider Kummer varieties, which are constructed as quotients of abelian surfaces by the $\mbb{Z}/2\mbb{Z}$ inversion action. There are $16$ fixed points in this instance, and the quotient scheme has $16$ corresponding isolated singularities.
\begin{corollary}\label{cor:symp_informal}
Let $X$ be a projective symplectic variety, and let $G\subset \mrm{Sp}(X)$ be a finite subgroup for which the corresponding action is not free. Then the Hochschild cohomology of the orbifold quotient $\msc{X}$ is not formal.
\end{corollary}
\begin{remark}
It seems likely that the conclusion of Theorem~\ref{thm:informal} holds for arbitrary proper, Calabi-Yau, Deligne-Mumford stacks $\msc{Z}$ with singular coarse space $\underline{Z}$.
\end{remark}
One observes that we have actually proved the following.
\begin{corollary}\label{cor:DM_informal}
There exist smooth Deligne--Mumford stacks for which the Hochschild cohomology is not formal.
\end{corollary}
\subsection{Alternate notions of formality and derived equivalence}
Let us say that the Hochschild cohomology for $\msc{X}$ is $\beta$-formal if the higher $A_\infty$-structure on the vector space valued cohomology $H\! H^\bullet(\msc{X})$ vanishes. This is equivalent to the statement that there exists a homotopy isomorphism, or $A_\infty$ quasi-isomorphism, $H\! H^\bullet(\msc{X})\overset{\sim}\to \mrm{RHom}_{\msc{X}^2}(\Delta_\ast\O_\msc{X},\Delta_\ast\O_\msc{X})$.
It is apparent from our above examples that the usual formality property for a stack $\msc{X}$ is not a derived invariant. Indeed, if we suppose that the coarse space $X/G$ of a projective symplectic global quotient $\msc{X}$ admits a crepant resolution $Y\to X/G$, then Bridgeland--King--Reid construct a derived equivalence $\mrm{D}^b(\msc{X})\cong \mrm{D}^b(Y)$ via an explicit integral transform. In this case $\msc{X}$ is as in Corollary~\ref{cor:symp_informal}, and hence the Hochschild cohomology is not formal. However, formality of Hochschild cohomology for smooth varieties tells us that the Hochschild cohomology of $Y$ is formal.
\par
On the other hand, $\beta$-formality is a derived invariant (see~\cite{caldararuI,orlov03,BFN-dag}). This $\beta$-formality is implied by usual formality for affine varieties, but the relationship between the two notions of formality is unclear in general. Indeed, by~\cite{calaquevandenbergh10II} the Hochschild cohomology of a smooth variety $Y$ is the $A_\infty$-algebra given as the cohomology of the derived global sections $\mrm{R}\Gamma^\bullet(Y,T^{poly}_Y)$, which is not apriori formal as a dg algebra in the category of vector spaces. So, for example, we do not know at this moment if an affine symplectic quotient $\msc{X}$ is $\beta$-formal even though any crepant resolution $Y$ of the coarse space is formal in the usual sense. This is because such a crepant resolution will be non-affine in general, and hence $\beta$-formality is not known for $Y$.
\section{Hochschild cohomology for Symplectic quotients}
\label{sect:symplectic}
We fix $k=\mbb{C}$ in this section. Here we consider a symplectic variety $X$ (over $\mbb{C}$) on which a finite group $G$ acts by symplectic automorphisms. We explain how the local Hochschild cohomology of the quotient $\msc{X}=[X/G]$ simplifies in this case. Our presentation involves a correction to an oft-repeated error in the literature (see Section~\ref{sect:remark} and Appendix~\ref{sect:pav}). In Section~\ref{sect:H_orb} we use our description of Hochschild cohomology to propose explicit relations between Hochschild cohomology and orbifold cohomology for symplectic quotients.
\subsection{Generalities for symplectic quotients}
Let $X$ be symplectic and $G$ be a finite group acting on $X$ via symplectic automorphisms. We let $\omega\in \Gamma(X,\Omega^2_X)$ be the symplectic form and $\pi=\omega^{-1}\in \Gamma(X,T^2_X)$ be the corresponding Poisson structure. We note that the fixed spaces for the $G$ action are symplectic subvarieties in $X$, and we denote the corresponding symplectic form on $X^g$ by $\omega_g$ for each $g\in G$. We fix $d_g=\frac{1}{2}\mrm{codim}(X^g)$ and $d=\frac{1}{2}\dim(X)$.
\par
Via the decomposition $\Omega_X|_{X^g}=\Omega_{X^g}\oplus \Omega_{X^g}^\perp$, where $\Omega_{X^g}^\perp$ is the sum of all nontrivial $g$-eigensheaves, we decompose the form as $\omega|_{X^g}=\omega_g+\omega_g^\perp$ with $\omega_g$ a section of $\Omega_{X^g}$ and $\omega_g^\perp$ a section of $\Omega_{X^g}^\perp\wedge \Omega_{X^g}^\perp$. Note that the mixed term $\omega_g^{mixed}$ over $\Gamma(X^g,\Omega_{X^g}\wedge \Omega_{X^g}^\perp)$ vanishes by $g$-invariance of the restriction $\omega|_{X^g}$. Hence the volume form is the product $(\omega|_{X^g})^d=\binom{d}{d_g}\omega_g^{d-d_g}(\omega_g^{\perp})^{d_g}$, and we have an isomorphism
\begin{equation}\label{eq:116}
\Omega_{X^g}^{2(d-d_g)}\otimes_{\O_X}\det(N_{X^g}^\vee)\to \Omega_X^{2d}|_{X^g},\ \ \omega_g^{(d-d_g)}\otimes (\omega_g^{\perp})^{d_g}\mapsto \binom{d}{d_g}^{-1}(\omega|_{X^g})^d
\end{equation}
given by the multiplication on $\Omega^{poly}_X|_{X^g}$. Since the respective volume forms trivialize $\Omega_{X^g}^{2(d-d_g)}$ and $\Omega_X^{2d}|_{X^g}$, the isomorphism~\eqref{eq:116} implies that $\det(N_{X^g}^\vee)$ is globally trivial and generated by the relative volume form $(\omega_g^{\perp})^{d_g}$. A similar analysis holds for the tangent sheaf, with $\det(N_{X^g})$ trivialized by the relative dual volume form $(\pi_g^\perp)^{d_g}$. We normalize and take
\[
\psi_g:=\frac{1}{d_g!}(\pi_g^\perp)^{d_g}.
\]
We have just established
\begin{lemma}
Let $X$ be a symplectic variety and $G$ be a finite group acting by symplectic automorphisms on $X$. Then the determinants of the normal bundles $\det(N_{X^g})$, for each $g\in G$, are globally trivial and generated by the global polyvector field $\psi_g$.
\end{lemma}
Recall that the subalgebra $\underline{\mathrm{SA}}(\msc{X})$ is generated by the determinants $\det(N_{X^g})$ as an $\O_X$-algebra, and each determinant is generated by the global polyvectors $\psi_g$. Hence $\underline{\mathrm{SA}}(\msc{X})$ is generated as an $\O_X$-algebra by the sections $\psi_g$, for $g$ not the identity. By the $G$-grading on $\underline{\mathrm{SA}}(\msc{X})$ we have
\begin{equation}\label{eq:126}
\psi_g\cdot \psi_h|_Y=a(g,h)\psi_{gh}|_Y,
\end{equation}
where the $Y$ range over shared components of $X^{gh}$ and $X^g\cap X^h$ and $a(g,h)$ is a section of the sheaf of units $(\prod_Y\O_Y)^\times$.
\par
By checking complete locally at closed points, to reduce to the affine case, one finds that the coefficients $a(g,h)$ are in the constant sheaf of units $\underline{\mbb{C}}$ on $\coprod_\mrm{shared\ cmpts} Y$. Whence we see that the multiplication on $\underline{\mathrm{SA}}(\msc{X})$, and the Hochschild cohomology in its entirely $\underline{H\! H}^\bullet(\msc{X})$, is determined by the coefficients $a(g,h)$. We will show that the function $a$ is, essentially, a boundary. (We use the word essentially here because $a$ is not actually a $2$-cocycle for $G$.)
\subsection{Hochschild cohomology for symplectic quotients}
\label{sect:scaling}
Consider the product $\mbb{C}^{\pi_0(X^g)}$ as the global sections of the constant sheaf $\underline{\mbb{C}}$ on $X^g$. We seek invertible scalars $\lambda_g\in \mbb{C}^{\pi_0(X^g)}$ for each $g\in G$ such that
\begin{equation}\label{eq:137}
(\lambda_g\lambda_h)^{-1}|_Y(\lambda_{gh}|_Y)=a(g,h)|_Y\ \ \text{on shared components $Y$ between $X^{gh}$ and }X^g\cap X^h,
\end{equation}
which satisfy the $G$-equivariance condition:
\begin{equation}\label{eq:137-eq}
\text{For every component
$Z \subseteq X^g$ and every $h \in G$, } \lambda_g|_Z = \lambda_{hgh^{-1}}|_{h(Z)}.
\end{equation}
In this case the scaled generators $\mu_g=\lambda_g\psi_g$ satisfy $\mu_g\mu_h|_Y=\mu_{gh}|_Y$, and under the $\O_X$-linear isomorphism
\[
\oplus_{g\in G}\O_{X^g}\overset{\sim}\longrightarrow \underline{\mathrm{SA}}(\msc{X}),\ \ 1_g\mapsto \lambda_g\psi_g,
\]
the induced algebra structure on the sum $\oplus_g \O_{X^g}$ is simply given by
\begin{equation}\label{eq:transp}
f_g\cdot f_h|_Y=\left\{\begin{array}{ll}
(f_g|_Y)(f_h|_Y) & \text{if $Y$ is a shared component with }X^g\cap X^h\\
0 & \text{otherwise},
\end{array}\right.
\end{equation}
where the $f_\sigma$ are sections of their respective structure sheaves and $Y$ is an arbitrary component of $X^{gh}$ in the above formula. We call the multiplication~\eqref{eq:transp} the {\it transparent multiplication} on the sum $\oplus_{g\in G}\O_{X^g}$.
\par
We note that the transparent multiplication on the sum $\oplus_g\O_{X^g}$ is related to the multiplication on the group ring $\O_X[G]$, although there is some nuance to consider here. For $X=V$ a symplectic vector space, for example, and $G\subset \mrm{Sp}(V)$, the fiber of the transparent algebra $\oplus_g\O_{X^g}$ at $0$ is the associated graded algebra $\operatorname{gr}_F\mbb{C}[G]$ of the group algebra under the codimension filtration $F_{-i}\mbb{C}[G]=\mbb{C}\{g\in G:\mrm{codim}(V^g)\geq i\}$ (see e.g.~\cite{etingofginzburg02}).
We will see below that the determinants of the operators $(1-g|_{N_{X^g}})$ on the normal bundles $N_{X^g}$, for each $g$, provide the desired bounding function as in~\eqref{eq:137}.
\begin{lemma}\label{lem:182}
For each $g\in G$, the determinant of the operator $(1-g|_{N_{X^g}})$ is a real, positive, global section of the constant sheaf of complex numbers over $X^g$.
\end{lemma}
\begin{proof}
Consider a component $Z$ of $X^g$. The space $T_X|_{Z}$ decomposes into eigensheaves $\oplus_{\chi}T_\chi$, where $\chi$ runs over the characters of the group $\langle g|_{T_X}\rangle$, and we have $N_{Z}=\oplus_{\chi\neq id}T_\chi$. Invariance of the non-degenerate form $\omega:T_X|_{Z}\wedge T_X|_{Z}\to \O_{Z}$ under $g$ implies that the pairing restricts to non-degenerate pairings $T_{\chi}\times T_{\chi^{-1}}\to \O_{Z}$. In particular, the eigensheaves come in pairs of the same rank, and if we take $\zeta_i=\chi_i(g)$ we have
\[
\det(1-g|_{N_{Z}})=2^l\prod_{i=1}^N(1-\zeta_i)(1-\bar{\zeta}_i)=2^l\prod_{i=1}^N|1-\zeta_i|^2,
\]
where $\{\chi_1,\dots,\chi_N\}$ is an enumeration of the characters for which $\zeta_i=\chi_i(g)$ is in the upper half plane in $\mbb{C}$, and $l$ is the rank of the eigensheaf with eigenvalue $-1$ for $g$.
\end{proof}
We provide a detailed proof of the following proposition in the case in which $X$ is a symplectic vector space in Appendix~\ref{sect:pav}, as there are some inaccuracies in the literature which need to be corrected. We discuss the proof for general $X$ in Section~\ref{sect:proof_disc}.
\begin{proposition}\label{prop:coboundary}
The global sections $\mrm{det}(1-g|_{N_{X^g}})^{1/2}\in \mbb{C}^{\pi_0(X^g)}$ collectively solve equation~\eqref{eq:137} and \eqref{eq:137-eq}, where $\mrm{det}(1-g|_{N_{X^g}})^{1/2}$ denotes the positive square root.
\end{proposition}
Note that this solution is indeed invariant under conjugation by $G$, as desired.
\begin{remark} \label{r:agh-nz}
If $G$ is abelian, then it is easy to check that the cocycle $a(g,h)=1$ for all $g$ and $h$, in spite of the fact that the determinants $\mrm{det}(1-g|_{N_{X^g}})$ are not equal to one. Thus in this case one could also use the trivial solution to \eqref{eq:137}. However, when $G$ is nonabelian, in general $a(g,h) \neq 1$. For example, if $G=S_n$ and $X=T^*\mbb{C}^n$,
it follows from Proposition \ref{prop:coboundary} (and one can verify directly) that $a((1 \cdots k)(k \cdots m)) = \frac{m}{k(m-k)}$.
\end{remark}
From Proposition~\ref{prop:coboundary} we obtain a simple description of the multiplication of $\underline{\mathrm{SA}}(\msc{X})$ for arbitrary symplectic quotients.
\begin{theorem}\label{thm:SAsymp}
Suppose $X$ is a symplectic variety and that $G$ acts on $X$ by symplectic automorphisms. Then there is a canonical isomorphism of $\O_X$-algebras $\underline{\mathrm{SA}}(\msc{X})\cong \oplus_{g\in G}\O_{X^g}$, where $\oplus_g\O_{X^g}$ is given the transparent multiplication~\eqref{eq:transp}.
\end{theorem}
This theorem extends to the ``sheaf-valued'' cohomology $\underline{H\! H}^\bullet(\msc{X})=\oplus_{g\in G} T^{poly}_{X^g}$ in the symplectic case as well, where $\msc{X}=[X/G]$, as usual.
\begin{theorem}\label{thm:HHsymp}
There is an isomorphism of equivariant sheaves of algebras
\[
\oplus_{g\in G}T^{poly}_{X^g}\overset{\sim}\longrightarrow\underline{H\! H}^\bullet(\msc{X}),\ \ 1_g\mapsto \det(1-g|_{N_{X^g}})^{1/2}\psi_g
\]
where the sum $\oplus_g T^{poly}_{X^g}$ is given the transparent multiplication.
\end{theorem}
Note that to include the grading we need to shift the polyvector fields as $\oplus_g T^{\bullet-\mrm{codim}(X^g)}_{X^g}$. We ignore this degree shift for the sake of notation.
\par
Let us state explicitly what we mean by the transparent multiplication here. Let $Y$ be a shared component for $X^{gh}$ and $X^g\cap X^h$ along which the fixed spaces intersect transversely. For $\tau=g,h,gh$, we have the projections
\[
p_\tau:T^{poly}_X|_Y\to T^{poly}_{X^\tau}|_Y
\]
with kernel generated by the normal bundle $N_{X^\tau}$. The containments $N_{X^g}|_Y,\ N_{X^h}|_Y\subset N_{X^{gh}}|_Y$ provided by (the proof of) Lemma~\ref{lem:esslem} (iii) imply that the projection $p_{gh}$ factors through $T^{poly}_{X^g}|_Y$ and $T^{poly}_{X^h}|_Y$, and we have induced projections $T^{poly}_{X^g}|_Y\to T^{poly}_{X^{gh}}|_Y$ and $T^{poly}_{X^h}|_Y\to T^{poly}_{X^{gh}}|_Y$. We let $p_{gh}$ denote these induced projections, by an abuse of notation. Whence we can define an obvious product by restricting to components $Y$ of $X^{gh}$,
\begin{equation}\label{eq:Ttransp}
(\xi_g\cdot\xi_h)|_Y=\left\{\begin{array}{ll}
p_{gh}(\xi_g|_Y)p_{gh}(\xi_h|_Y) &\text{if $Y$ is a shared component}\\
0 & \text{otherwise}.
\end{array}\right.
\end{equation}
This is what we have referred to as the transparent multiplication in Theorem~\ref{thm:HHsymp}.
\subsection{The proof of Proposition~\ref{prop:coboundary}}
\label{sect:proof_disc}
\begin{proof}[Proof of Proposition~\ref{prop:coboundary}]
The identity \eqref{eq:137-eq} is clear from the definition $\lambda_g=\mrm{det}(1-g|_{N_{X^g}})$.
To prove \eqref{eq:137},
it suffices to work complete locally at closed points on shared
components $Y$ of $X^{gh}$ and $X^g\cap X^h$. At such a point $p$ on
a shared component $Y$ we take $V=T_X|_{\hat{Y}_p}$. We have
$V^g=N_{X^g}|_{\hat{Y}_p}$,
$(V^g)^\perp=(1-g)V=T_{X^g}|_{\hat{Y}_p}$, and similar equalities for
$h$ and $gh$. We consider these objects as modules over the complete
local $\mbb{C}$-algebra $K=\hat{\O}_{Y,p}$.
\par
One replaces $\mbb{C}$ with $K$ and argues just as in the appendix to arrive at Proposition~\ref{prop:coboundary}. Namely, the proofs of Lemma~\ref{lem:pfaffian} and Proposition~\ref{p1} are equally valid over $K$. This gives
\[
a(g,h)=\left(\frac{\det(1-gh|_{N_{X^{gh}}})}{\det(1-g|_{N_{X^g}})\det(1-h|_{N_{X^h}})}\right)^{1/2}.
\]
We have seen at Lemma~\ref{lem:182} that all of the determinants in the above formula are positive real, and hence $a(g,h)$ is either the positive or negative square root on each component $Y$. One can check the sign of $a(g,h)$ on the fiber $V\otimes_K\mbb{C}$, which reduces us to an analysis of a symplectic vector space equipped with a finite symplectic group action. So we apply Proposition~\ref{p2} directly to find that $a(g,h)$ is in fact the positive square root.
\end{proof}
\subsection{A remark on the literature}
\label{sect:remark}
It is stated in multiple references that the coefficients $a(g,h)$ are identically $1$, in the specific case of a vector space $X=V$ and $G$ a finite subgroup in $\mrm{Sp}(V)$ (e.g.~\cite[proof of Theorem 1.8]{etingofginzburg02},~\cite[\S 3.10]{alvarez02},~\cite[Theorem 4]{anno}). However, this is not the case, as explained in Remark \ref{r:agh-nz}. The best that one could hope for is that the coefficients $a(g,h)$ form a $G$-equivariant ``coboundary", in the sense of~\eqref{eq:137},\eqref{eq:137-eq}, which is what is shown in Proposition~\ref{prop:coboundary}.
\section{Symplectic quotients and orbifold cohomology}
\label{sect:H_orb}
As in Section~\ref{sect:symplectic}, we suppose $X$ is a symplectic variety over $\mbb{C}$ equipped with an action of a finite group $G$ by symplectic automorphisms. We present relations between Hochschild cohomology and orbifold cohomology, motivated by findings of~\cite{ginzburgkaledin04,dolgushevetingof05,etingofoblomkov06}. Indeed, many of the results here can be seen as globalized variants of results from~\cite{ginzburgkaledin04}, and some materials intersect non-trivially with work of Ruan~\cite{ruan00,ruan06}.
\subsection{A dg structure on polyvector fields}
Let $F_\omega:T_X\to\Omega_X$ be the isomorphism provided by the symplectic structure, and let $\{-,-\}$ denote the Poisson bracket on polyvector fields.
\par
We have
\[
F_\omega\{\pi,f\}=2F_\omega\left(\pi(-,d_f)\right)=2\pi(F_\omega^\vee-,d_f)=2\langle -,d_f\rangle:T_X\to \msc{O}_X.
\]
This final element is simply $2 d_f\in \Omega_X$. Furthermore, one can checking complete locally to find that the sheaf isomorphism $F^{poly}_\omega$ provides an isomorphism of complexes
\begin{equation}\label{eq:226}
F^{poly}_\omega:(T_X^{poly},\frac{1}{2}\{\pi,-\})\to (\Omega_X^{poly},d_\mrm{dR}).
\end{equation}
\begin{comment}
More specifically, the elements $\xi_i:=\{\pi,x_i\}$ provide a basis for $T_X$ complete locally, in coordinates $\hat{\O}_{X,p}\cong \mbb{C}[\![x_i]\!]$. This is simply because the $F_\omega\xi_i=2d_{x_i}\in \Omega_{\hat{X}_p}$ provide a basis for the K\"ahler differentials. Now the fact that
\[
\frac{1}{2}\{\pi,\xi_i\}=\frac{1}{2}\{\pi,\{\pi,x_i\}\}=\frac{1}{4}\{\{\pi,\pi\},x_i\}=0
\]
for each $i$ implies that for arbitrary polyvectors
\[
\begin{array}{rl}
F_\omega^{poly}\left(\frac{1}{2}\{\pi,\sum_Ia_I\xi_{i_1}\dots\xi_{i_l}\}\right)&=\sum_I\frac{1}{2}F_\omega^{poly}\{\pi,a_I\} F^{poly}_\omega(\xi_{i_1}\dots\xi_{i_l})\\
&=2^l\sum_I d_\mrm{dR}(a_I)d_{x_{i_1}}\dots d_{x_{i_l}}\\
&=2^l\sum_I d_\mrm{dR}(a_Id_{x_{i_1}}\dots d_{x_{i_l}})=d_\mrm{dR}\left(F_\omega^{poly}(\sum_I a_I \xi_{i_1}\dots \xi_{i_l})\right),
\end{array}
\]
as claimed. Of course, we can repeat this at each of the fixed spaces to obtain isomorphisms $(T_{X^g}^{poly},\frac{1}{2}\{\pi_g,-\})\cong (\Omega_{X^g}^{poly},d_\mrm{dR})$.
\end{comment}
We define the operation
\[
\{\pi,-\}:\oplus_{g\in G}T^{poly}_{X^g}\to \oplus_{g\in G} T^{poly}_{X^g}
\]
formally as the sum $\oplus_{g\in G}\{\pi_g,-\}$.
\par
We consider $\oplus_{g\in G}T^{poly}_{X^g}$ as a sheaf of algebras on $X$ under the transparent multiplication~\eqref{eq:Ttransp}. By a nontrivial, but straightforward, computation one sees that the operator $\frac{1}{2}\{\pi,-\}$ is a degree one square zero derivation on the sheaf of algebras $\oplus_g T^{poly}_{X^g}$, with the transparent multiplication. Rather, $(\oplus_g T^{poly}_{X^g},\frac{1}{2}\{\pi,-\})$ is a dg algebra.
\begin{comment}
\begin{lemma}\label{lem:404}
The projection $p_g:T^{poly}_{X}|_{X^g}\to T^{poly}_{X^g}$ satisfies $p_g\{\pi,\xi\}|_{X^g}=\{\pi_g,p_g(\xi)\}$.
\end{lemma}
\begin{proof}
The result only relies on $G$-invariance of $\pi$. We may assume $G=\langle g\rangle$. It suffices to prove the equality complete locally at closed points in $X^{gh}$, and hence may assume further that $G$ is a finite group acting linearly on affine space. In this case $\Gamma(X,T^{poly}_X)$ is generated by $g$-eigenvectors $x_i$ and $\frac{\partial}{\partial x_i}$, where $\O(X)=\mbb{C}[x_1,\dots, x_N]$ with the $x_i$ eigenvectors, and the kernel of $p_g$ is generated by $I_g=(1-g)\Gamma(X,T^{poly}_X)$.
\par
We may write
\[
\pi=\sum_{i<j}a_{ij}\frac{\partial}{\partial x_i}\frac{\partial}{\partial x_j}
\]
and take $\pi_g\in \Gamma(X,T^{poly}_X)$ to be $\pi_g=\sum_{s<t}a_{st}\frac{\partial}{\partial x_s}\frac{\partial}{\partial x_t}$, where the sum is over the invariant $x_s$, $x_t$. This is an abuse of notation, as formally this element is a chosen lift of $\pi_g$. We note that in the above expression of $\pi$ the coefficients $a_{ij}$ must be $g$-eigenvectors, with the same eigenvalue as $x_ix_j$, by $g$-invariance of $\pi$.
\par
Via the Gerstenhaber identity and the fact that $p_g$ is an algebra map we need only show that
\[
\{\pi,x_i\}=\{\pi_g,x_i\}\ \mrm{mod} I_g\ \ \mrm{and}\ \ \{\pi,\frac{\partial}{\partial x_i}\}=\{\pi_g,\frac{\partial}{\partial x_i}\}\ \mrm{mod} I_g.
\]
for invariant $x_i$, and that
\[
\{\pi,x_j\}=0\ \mrm{mod} I_g\ \ \mrm{and}\ \ \{\pi,\frac{\partial}{\partial x_j}\}=0\ \mrm{mod} I_g
\]
for non-invariant $x_j$. The latter two equalities follow by $g$-invariance of the Schouten bracket, which implies
\[
(1-g)\cdot \{\pi,\eta\}=\{\pi,\eta\}\setminus \{\pi,g\cdot \eta\}=(1-\zeta)\{\pi,\eta\}
\]
for $\eta$ a $g$-eigenvector of eigenvalue $\zeta$. The first two equalities follow by a direct analysis using the explicit expressions of $\pi$ and $\pi_g$ given above.
\end{proof}
We employ this relation in the proof of the following lemma.
\begin{lemma}\label{lem:284}
The operator $\frac{1}{2}\{\pi,-\}$ is a degree one square zero derivation on the sheaf of algebra $\oplus_g T^{poly}_{X^g}$, with the transparent multiplication. Rather, $(\oplus_g T^{poly}_{X^g},\frac{1}{2}\{\pi,-\})$ is a dg algebra.
\end{lemma}
\begin{proof}
Take $p=p_{gh}$. It suffices to prove the equality
\[
\{\pi_{gh}|_Y,p(\xi_g|_Y)p(\xi_h|_Y)\}=p\{\pi_g,\xi_g\}|_Y\cdot p(\xi_h|_Y)\ \pm\ p(\xi_g|_Y)\cdot p\{\pi_h,\xi_h\}|_Y
\]
for $\xi_g$ and $\xi_h$ sections of $T^{poly}_{X^g}$ and $T^{poly}_{X^h}$ respectively, and $Y$ a shared component for $X^{gh}$ and $X^g\cap X^h$. Since $T^{poly}_{X^{gh}}$ is a dg algebra under $\{\pi_{gh},-\}$, we have
\[
\{\pi_{gh}|_Y,p(\xi_g|_Y)p(\xi_h|_Y)\}=\{\pi_{gh}|_Y,p(\xi_g|_Y)\}\cdot p(\xi_h|_Y)\ \pm\ p(\xi_g|_Y)\cdot p\{\pi_{gh}|_Y,(\xi_h|_Y)\}
\]
\par
We claim that $p\{\pi_g,\xi_g\}|_Y=\{\pi_{gh}|_Y,p(\xi_g|_Y)\}$. Indeed, we can choose an arbitrary lift $\tilde{\xi}_g$ of $\xi_g$ to a section of $T^{poly}_X|_Y$ (locally) and apply Lemma~\ref{lem:404} to find
\[
p\{\pi_g,\xi_g\}|_Y=p\{\pi,\tilde{\xi}_g\}|_Y=\{\pi_{gh}|_Y,p(\xi_g|_Y)\}.
\]
One similarly verifies $p\{\pi_h,\xi_h\}|_Y=p\{\pi_{gh}|_Y,(\xi_h|_Y)\}$. Whence we have the desired equality.
\end{proof}
Via the algebra isomorphism $\underline{H\! H}^\bullet(\msc{X})\cong \oplus_g T^{poly}_{X^g}$ of Theorem~\ref{thm:HHsymp} we transfer the operator $\frac{1}{2}\{\pi,-\}$ to a degree $1$, square $0$, derivation on the local Hochschild cohomology. Whence we have a dg algebra structure on the Hochschild cohomology induced by the symplectic structure $\pi$.
\begin{remark}
The cohomology $\underline{H\! H}^\bullet(\msc{X})$ is not expected to admit a Gerstenhaber structure. For the most na\"ive indication that no such structure exists, we note that the algebra $\underline{H\! H}^\bullet(\msc{X})$ is not (graded) commutative. However, the identity component $T^{poly}_X\subset \underline{H\! H}^\bullet(\msc{X})$ is naturally a Gerstenhaber algebra, and this Gerstenhaber structure should extend to a Gerstenhaber bimodule structure on the entire cohomology $\underline{H\! H}^\bullet(\msc{X})$. Under this conjectural Gerstenhaber module structure, we expect that the operation $\{\pi,-\}\in \mrm{End}_\mrm{shf}(\underline{H\! H}^\bullet(\msc{X}))$
\end{remark}
\begin{remark}
One could argue that the degree shift which we are ignoring in the isomorphism $\underline{H\! H}^\bullet(\msc{X})=\oplus_g T^{poly}_{X^g}$ should affect the action $\{\pi,-\}$ of $\pi$, since $\{\pi,-\}$ is essentially a graded commutator. However, since each of the $X^g$ are symplectic subvarieties in $X$, the degree shifts $\mrm{codim}(X^g)$ are all even and hence vanish when we consider the cohomology under the $\mbb{Z}/2\mbb{Z}$-parity, which is all that matters here.
\end{remark}
\end{comment}
\subsection{Linear identifications with orbifold cohomology}
We have the isomorphism of sheaves of complexes
\begin{equation}\label{eq:463}
(\underline{H\! H}^\bullet(\msc{X}),\frac{1}{2}\{\pi,-\})\cong \bigoplus_{g\in G}(\Omega_{X^g}^\bullet,d_\mrm{dR})
\end{equation}
given by $F^{poly}_\omega$, up to shifting by the codimensions. This induces a non-trivial dg algebra structure on the sum of the de Rham complexes. We discuss this dg structure in Section~\ref{sect:orb_prod}, but let us first record a number of observations and linear identifications. We define the dg algebra
\[
\underline{H\! H}^\bullet_\pi(\msc{X}):=(\underline{H\! H}^\bullet(\msc{X}),\frac{1}{2}\{\pi,-\})
\]
(cf.~\cite[Sect.\ 6]{ginzburgkaledin04}).
\begin{lemma}[{cf.~\cite{ginzburgkaledin04}}]\label{lem:HochFG}
There is an isomorphism of graded $G$-representations
\[
\mbb{H}^\bullet(X,\underline{H\! H}^\bullet_\pi(\msc{X}))\overset{\cong}\to \bigoplus_{g\in G}H^{\bullet+\mrm{codim}(X^g)}(X^g_\mrm{an},\mbb{C}),
\]
where $\mbb{H}^\bullet$ denotes the hypercohomology and $H^\bullet(Z_\mrm{an},\mbb{C})$ denotes the singular cohomology of the analytic space associated to a variety $Z$ over $\mbb{C}$.
\end{lemma}
\begin{proof}
This follows from the isomorphism~\eqref{eq:463}, Grothendieck's isomorphism between the algebraic de Rham cohomology and analytic de Rham cohomology~\cite{grothendieck66}, and the general isomorphism $H^\bullet_\mrm{dR}(Z_\mrm{an})\cong H^\bullet(Z_\mrm{an},\mbb{C})$. Each isomorphism here is sufficiently natural to account for the action of $G$.
\end{proof}
We take invariants to arrive at an identification with the orbifold cohomology.
\begin{theorem}\label{thm:HHorb}
There is a graded isomorphism between the hypercohomology of $\underline{H\! H}^\bullet_\pi(\msc{X})$, considered as a sheaf on $\msc{X}$, and the orbifold cohomology of $\msc{X}$, $\mbb{H}^\bullet(\msc{X},\underline{H\! H}^\bullet_\pi(\msc{X}))\cong H^\bullet_\mrm{orb}(\msc{X},\mbb{C})$.
\end{theorem}
\begin{proof}
In the symplectic setting the ``age shift" $a(g)$ of~\cite[Definition 1.5]{fantechigottsche03} is half the codimension of $X^g$. Hence the orbifold cohomology is given as the invariants
\[
H^\bullet_\mrm{orb}(\msc{X},\mbb{C})=\left(\oplus_g H^{\bullet+\mrm{codim}(X^g)}(X^g_\mrm{an},\mbb{C})\right)^G
\]
\cite[Definition 1.7]{fantechigottsche03}. Now the desired isomorphism is immediate from the equality $\mbb{H}^\bullet(\msc{X},\underline{H\! H}^\bullet_\pi(\msc{X}))=\mbb{H}^\bullet(X,\underline{H\! H}^\bullet_\pi(\msc{X}))^G$ and Lemma~\ref{lem:HochFG}.
\end{proof}
In the particular case of projective symplectic variety $X$, we have a direct comparison between Hochschild cohomology and orbifold cohomology.
\begin{corollary}\label{cor:501}
If $X$ is projective over $\mbb{C}$, then there is an isomorphism of graded vector spaces $H\! H^\bullet(\msc{X})\cong H^\bullet_\mrm{orb}(\msc{X})$.
\end{corollary}
\begin{proof}
By the Hodge decomposition we have
\[
H^\bullet_\mrm{orb}(\msc{X})=\big(\bigoplus_g H^\bullet(X^g_\mrm{an},\mbb{C})\big)^G=\big(\bigoplus_g H^\bullet(X^g,\Omega^\bullet_{X^g})\big)^G.
\]
Whence the identification follows from the isomorphisms $T_{X^g}\cong \Omega_{X^g}$ given by the symplectic form(s) and the vector space identification $H\! H^\bullet(\msc{X})=\big(\oplus_g H^\bullet(X^g,T^{poly}_{X^g})\big)^G$ of Corollary~\ref{cor:arinkinetal} (or rather~\cite{arinkinetal14}).
\end{proof}
\begin{remark}\label{rem:quant}
\begin{enumerate}
\item The isomorphism of Theorem~\ref{thm:HHorb} was obtained by Ginzburg and Kaladin in the case of a quotient $\msc{X}=[X/G]$ with $X$ affine~\cite{ginzburgkaledin04}.
\item A variant of Theorem~\ref{thm:HHorb} should hold, without the appearance of the differential $\frac{1}{2}\{\pi,-\}$, when one replaces $(\msc{X},\O_\msc{X})$ with a quantization of the symplectic structure, i.e.\ an equivariant quantization of $(X,\O_X)$. Specifically, if one takes $\msc{A}/\mbb{C}[\![\hbar]\!]$ a quantization of the symplectic orbifold $\msc{X}$, then one expects $H\! H^\bullet(\msc{A}[\hbar^{-1}])\cong H_\mrm{orb}^\bullet(\msc{X})\otimes\mbb{C}(\!(\hbar)\!)$. (Here we view $\msc{A}$ as a sheaf of rings on the \'etale site of $\msc{X}$, so that the Hochschild cohomology should implicitly involve a smash product construction for $\msc{A}$.) This relation was proved by Dolgushev and Etingof in the affine case. See also~\cite{ginzburgkaledin04,etingofoblomkov06}.
\item The isomorphism of Corollary~\ref{cor:501} can already be obtained from the formality result of~\cite{arinkinetal14}.
\end{enumerate}
\end{remark}
\subsection{Conjectural, stronger, relations with Fantechi-G\"ottsche cohomology and orbifold cohomology}
\label{sect:orb_prod}
We consider here the work of Fantechi and G\"ottsche~\cite{fantechigottsche03}, in the special case of a symplectic variety $X$ and a finite group $G$ acting via symplectic automorphisms. In~\cite{fantechigottsche03}, the authors define a noncommutative algebra structure on the sum
\[
H^\bullet_\mrm{FG}(X,G)=\bigoplus_{g\in G}H^{\bullet-\mrm{codim}(X^g)}(X_\mrm{an}^g,\mbb{C}),
\]
such that the natural action of $G$ is by algebra automorphisms. The orbifold cohomology is then recovered as the invariants
\[
H^\bullet_\mrm{orb}(\msc{X})=H^\bullet_\mrm{FG}(X,G)^G.
\]
\par
Of course, this resembles our situation for Hochschild cohomology, where we have produced an algebra structure on the sum $H^\bullet(X,\underline{H\! H}^\bullet(\msc{X}))$ such that the $G$-invariants approximately recovers the Hochschild cohomology algebra $H\! H^\bullet(\msc{X})$. However, for compact $X_\mrm{an}$ both the Fantechi-G\"ottsche cohomology and orbifold cohomology admit canonical Frobenius structures~\cite{fantechigottsche03,chenruan04}. As is argued in the proof of Theorem~\ref{thm:informal}, this Frobenius structure obstructs the existence of an algebra identification between Hochschild cohomology and Fantechi-G\"ottsche, or orbifold, cohomology.
\par
Let us define the codimension filtration on the Fantechi-G\"ottsche cohomology in the obvious way:
\[
F_i\ H^\bullet_\mrm{FG}(X,G)=\bigoplus_{\mrm{codim}(X^\sigma)\leq i}H^{\bullet-\mrm{codim}(X^\sigma)}(X_\mrm{an}^\sigma,\mbb{C}).
\]
This filtration restricts to a filtration on the invariant subalgebra $H^\bullet_\mrm{orb}(\msc{X})$. In the following conjecture, we denote the associated graded algebra for Hochschild cohomology appearing in the local-to-global spectral sequence simply by $\grH\! H^\bullet(\msc{X})$, as in the introduction.
\begin{conjecture}\label{conj:575}
Let $X$ be symplectic and $\msc{X}$ be the quotient stack of $X$ by a finite symplectic group action.
\begin{enumerate}
\item[(i)] The vector space isomorphism of Theorem~\ref{lem:HochFG} induces algebra isomorphisms
\[
\mbb{H}^\bullet(X,\underline{H\! H}^\bullet_\pi(\msc{X}))\cong \operatorname{gr}_FH^\bullet_\mrm{FG}(X,G)\ \ \mrm{and}\ \ \mbb{H}^\bullet(\msc{X},\underline{H\! H}^\bullet_\pi(\msc{X}))\cong \operatorname{gr}_FH^\bullet_\mrm{orb}(\msc{X}).
\]
\item[(ii)] When $X$ is projective, there is an algebra isomorphism $\grH\! H^\bullet(\msc{X})\cong \operatorname{gr}_FH^\bullet_\mrm{orb}(\msc{X})$.
\end{enumerate}
\end{conjecture}
When the quotient $X$ is projective and $X/G$ has isolated singularities the identification of $\operatorname{gr}_FH^\bullet_\mrm{orb}(\msc{X})$ with $\grH\! H^\bullet(\msc{X})$ follows from the fact that the Hodge decomposition respects the product on cohomology~\cite[Corollary 6.15]{voisin02}. When $X$ has fixed spaces of positive dimension an identification between the two cohomologies would already require some consideration.
\par
We note that Ginzburg and Kaledin make a related conjecture involving quantizations, as in Remark~\ref{rem:quant}. They propose that when one employs a quantization one should be able to avoid this associated graded procedure~\cite[Conjecture 1.6/1.3]{ginzburgkaledin04}. We repeat their conjecture here.
\begin{conjecture}[Ginzburg-Kaledin]
For an appropriate quantization $\msc{A}$ of a symplectic quotient orbifold $\msc{X}$, one has an algebra isomorphism $H\! H^\bullet(\msc{A}[\hbar^{-1}])\cong H^\bullet_\mrm{orb}(\msc{X})\otimes\mbb{C}(\!(\hbar)\!)$.
\end{conjecture}
Of course, point (ii) of Conjecture~\ref{conj:575} suggests that the Hochschild cohomology may be identified with the orbifold cohomology, before taking associated graded rings in both cases. The following is strongly related to a conjecture of Ruan, as explained below.
\begin{conjecture}[cf.~\cite{ruan00,ruan06}]\label{conj:HH_H_orb}
For $\msc{X}$ a projective symplectic quotient orbifold, there is an algebra identification between the Hochschild cohomology $H\! H^\bullet(\msc{X})$ and the orbifold cohomology $H^\bullet_\mrm{orb}(\msc{X})$.
\end{conjecture}
When $X/G$ admits a crepant resolution $Y\to X/G$, Conjecture~\ref{conj:HH_H_orb} reduces to Ruan's {\it Cohomological Hyperk\"{a}hler Resolution Conjecture}~\cite{ruan00,ruan06}, in the specific instance of a projective global quotient orbifold. The argument is as follows: In the presence of such a resolution $Y$, we have that $Y$ is projective and symplectic. There is then an equivalence of categories $\mrm{D}^b(Y)\cong \mrm{D}^b(\msc{X})$ by a result of Bridgeland-King-Reid~\cite[Corollary 1.3]{bridgelandkingreid01}, and subsequent algebra identification of Hochschild cohomologies $H\! H^\bullet(\msc{X})=H\! H^\bullet(Y)$~\cite{caldararuI} (see also~\cite{caldararuwillerton10,BFN-dag}). Now by the Hodge decomposition, formality for Hochschild cohomology~\cite{calaquevandenbergh10}, and the isomorphism $\Omega_Y\cong T_Y$ given by the symplectic form, we obtain further algebra identifications $H\! H^\bullet(Y)=H^\bullet_\mrm{dR}(Y)=H^\bullet(Y_\mrm{an},\mbb{C})$. This gives, in total, $H\! H^\bullet(\msc{X})=H^\bullet(Y_\mrm{an},\mbb{C})$. Ruan's conjecture proposes that we have $H^\bullet_\mrm{orb}(\msc{X})=H^\bullet(Y_\mrm{an},\mbb{C})$ as algebras, and we obtained the claimed reduction.
\par
We note that, in general, there will not exist a crepant resolution of
$X/G$ (a local obstruction to its existence is discussed in
\cite{Bel-sing,BS-sra2}, among other places).
For select verifications of Ruan's conjecture(s) one can see~\cite{lehnsorger03,uribe05,futian17,futianvial}, for example.
\begin{remark}
The statement of Conjecture~\ref{conj:HH_H_orb} makes perfect sense for an arbitrary projective symplectic orbifold $\msc{X}$, which needn't be a global quotient in general. It seems reasonable to propose the conjecture in this greater generality, although the statement should certainly undergo some additional scrutiny here. Note that our reduction of Conjecture~\ref{conj:HH_H_orb} to Ruan's conjecture does not immediately apply to more general $\msc{X}$, as the results of Bridgeland-King-Reid are only given for global quotients.
\end{remark}
\begin{remark}
It is proposed in~\cite[Proposition 6.2]{ginzburgkaledin04} that Conjecture~\ref{conj:575} (i) holds for affine $X$, in which case there are no derived global sections and the orbifold cohomology is already graded by codimension. So we have, in the affine case, a direct algebra isomorphism $H\! H_\pi^\bullet(\msc{X})\cong H^\bullet_\mrm{orb}(\msc{X})$.
\end{remark}
|
\section{Introduction}
Developing pure neural models to automatically
solve arithmetic expression calculation (AEC) is an interest and challenging task.
Recent research includes Neural GPUs~\citep{kaiser2015neural,freivalds2017improving}, Grid LSTM \citep{Kalchbrenner2015GridLS}, Neural Turing Machines \citep{graves2014neural}, and Neural Random-Access Machines~\citep{Kurach2016NeuralRM}. Most of these models just can deal with the addition calculation. Although Neural GPU has an ability to learn multi-digit binary multiplication, it does not work well in decimal multiplication \citep{kaiser2015neural}. The difficulty of multi-digit decimal multiplication lies in the fact that multiplication involves a complicated structure of arithmetic operations, which is hard for neural networks to learn. Considering how electronic circuit or human beings do multiplication, multi-digit multiplication can be decomposed into several subgoals, as shown in Figure \ref{fig:example}. High-level arithmetic tasks like multiplication iteratively use low-level operations like the addition to complete high-level tasks.
\begin{figure}[H]
\centering
\includegraphics[width=0.56\linewidth]{figs/sample.png}
\caption{An example of reusability and hierarchy in multiplication. $M+M$ denotes multi-digit addition, while $S\times S$ means single-digit multiplication.}\label{fig:example}
\end{figure}
The incapability of current models in solving arithmetic expression is because they fail to use two key properties of arithmetic operation: \textbf{reusability} and \textbf{hierarchy}. The arithmetic operation can be decomposed into a series of sub-operations, which form a hierarchical structure. Most of the sub-operations are reusable. When dealing with a complex arithmetic operation, we do not need to train a model from scratch. For the example in Figure \ref{fig:example}, the multi-digit multiplication involves several reusable sub-operations, such as $S\times S$ and $M + M$.
To leverage reusability and hierarchy in the arithmetic operation, we formulate this task as a Hierarchical Reinforcement Learning (HRL) problem~\citep{Sutton1999BetweenMA,Dietterich2000HierarchicalRL}, in which the task policy can be decomposed into several sub-task policies. Each sub-task policy is implemented by a \textit{skill module}, which can be used recursively. The skill module can be divided into two groups: basic skill module performing elementary single-digit operations, and interactive skill model performing complex operations by selectively invoking other skill modules.
There are two differences to the standard HRL. (1) One is that each invoked skill module can be executed with only its input, regardless of external environment state. Therefore, we propose \textbf{Interactive Skill Modules (ISM)} that can selectively interact with other skill modules by sending a partial expression and receiving answers returned. (2) Another is that the task hierarchy is multi-level, which is difficult to be learned from scratch. Therefore, we propose \textbf{Curriculum Teacher and Continual-learning Student (CTCS)} framework to overcome this problem. The skill modules are trained in a particular order, from easy to difficult tasks. The finally skill module would be a deep hierarchical structure.
The experiments show that our model has a strong capability to calculate arithmetic expressions.
The main contributions of the paper are:
\begin{itemize}
\item We propose a pure neural model to solve the (decimal) expression calculation problem, involving the $\{+, -, \times, \div, (, )\}$ operations. Both the input and output of our model is character sequence. To the best of our knowledge, this study is the first work to solve this challenging problem.
\item We regard arithmetic learning as a Multi-level Hierarchical Reinforcement Learning (MHRL) problem, and factorize a complex arithmetic operation into several simpler operations. The main component is the interactive skill module. A high-level interactive skill module can invoke the low-level skill modules by sending and receiving messages.
\item We introduce Curriculum Teacher and Continual-learning Student (CTCS), an automatic teacher-student framework that enables the model to be easier learned for the complex tasks.
\end{itemize}
\section{Related Work}
\paragraph{Arithmetic Learning}
In recent years, several models have attempted to learn arithmetic in deep learning. Grid LSTM \cite{Kalchbrenner2015GridLS} expands LSTM in multiple dimensions and can learn multi-digit addition. \citet{Zaremba2016LearningSA} use reinforcement learning to learn single-digit multiplication and multi-digit addition. Neural GPU \cite{kaiser2015neural} is noticeably promising in arithmetic learning and can learn binary multiplication. \citet{Price2016ExtensionsAL} and \citet{freivalds2017improving} improve Neural GPU to do multi-digit multiplication with curriculum learning. Nevertheless, there is no successful attempt to learn division or expression calculation.
\paragraph{Hierarchical Reinforcement Learning}
The first popular hierarchical reinforcement learning model may date back to the options framework \cite{Sutton1999BetweenMA}. The options framework considers the problem to have a two-level hierarchy. Recent work combines neural networks with this two-level hierarchy and has made promising results in challenging environments with sparse rewards, like Minecraft \cite{tessler2017deep} and ATARI games \cite{Baranes2013ActiveLO}. In contrast to the two-level hierarchy, the skill modules in our framework can selectively use other skill modules, which finally form a deep multi-level hierarchical structure.
\paragraph{Curriculum Learning}
Work by \citet{Bengio2009CurriculumL} brings general interests to curriculum learning. Recently, it has been widely used in many tasks, like learning to play first-person shooter games \cite{Wu2017TrainingAF}, and helping robots learn object manipulation \cite{Baranes2013ActiveLO}. It is noteworthy that the teacher-student curriculum learning framework proposed by \citet{matiisen2017teacher} can automatically sample tasks according to student's performance. However, it is limited to sampling data and can not help the student adapt to task switching with parameter adjustment.
\paragraph{Continual Lifelong Learning}
As proposed in \cite{tessler2017deep}, a continual lifelong learning model needs the ability to choose relevant prior knowledge for solving new tasks, which is named \textbf{selective transfer}. The main issue of continual learning models is that they are prone to \textbf{catastrophic forgetting} \cite{Mcclelland1995WhyTA,Parisi2018ContinualLL}, which means the model forgets previous knowledge when learning new tasks. To achieve continual lifelong learning, Progressive Neural Networks (PNN) \cite{Rusu2016ProgressiveNN} allocate a new module with access to prior knowledge to learn a new task. With this approach, prior knowledge can be used, and former modules are not influenced. Our model extends PNN with the ability to use helpful modules selectively.
\section{Model}
\paragraph{Task Definition}
We first formalize the task of arithmetic expression calculation (AEC) as follows. Given a character sequence, consisting of decimal digits $[0-9]$ and arithmetic operators of $\{+, -, \times, \div, (, )\}$, the goal is to output a sequence of digit characters representing the result, for example:
\begin{center}
\centering
\includegraphics[width=0.78\linewidth]{figs/introduction.png}
\end{center}
\subsection{Multi-level Hierarchical Reinforcement Learning}
As analyzed before, the arithmetic calculation can be decomposed into several sub-tasks, including single-digit multiplication, multi-digit addition and more.
Assuming we already have several modules for the simple arithmetic calculations, the key challenge is how to organize them to solve a more complex arithmetic calculation.
In this paper, we propose a multi-level hierarchical reinforcement learning (MHRL) framework to perform this task.
\paragraph{Hierarchical Reinforcement Learning (HRL)}
In HRL, the policy $\pi$ of an agent can be decomposed into several sub-policies from the set $\Pi=\{\pi_1,\pi_2,\cdots,\pi_K\}$. At time $t$, the policy $\pi: \mathcal{S}\rightarrow \Pi$ is a mapping from state $s^t\in \mathcal{S}$ to a probability distribution over sub-policies. Assuming the $k$-th sub-policy is chosen, the action $a_t$ is determined by $\pi_k(a|s^t)$.
The arithmetic calculation is a multi-level hierarchical reinforcement learning, in which the sub-policy could be further decomposed into sub-sub-policies. Suppose that each (sub-)policy is implemented by a skill module. There are two different kinds of modules: \textit{basic skill modules} (BSM) and \textit{interactive skill modules} (ISM). All modules use character sequences as inputs and produce character sequences as outputs.
\begin{figure}[h]\centering
\subfloat[Basic Skill Module (BSM)]{\label{fig:basic}
\includegraphics[width=0.325\textwidth]{figs/basic_module.png}
}
\subfloat[Interactive Skill Module (ISM)]{\label{fig:sampleISM} \includegraphics[width=0.5\textwidth]{figs/multiplication.png}
}
\caption{Demonstration of skill modules. (a) the basic skill module (BSM) is a neural network with Bi-RNN, which takes the character sequence as input and outputs the character sequence of the result. (b) The interactive skill module (ISM) interacts with other skill models to do the calculation. This illustrated module $\mathcal{M}_{M \times S}$ can perform multi-digit $\times$ single-digit calculation, using two BSMs $\mathcal{M}_{S + S}$ (single-digit addition), $\mathcal{M}_{S \times S}$ (single-digit multiplication) and another ISM $\mathcal{M}_{M + M}$ for multi-digit addition.}\label{fig:sample}
\end{figure}
\paragraph{Basic Skill Modules (BSM)}
The basic skill modules perform fundamental arithmetic operations like single-digit's addition or multiplication. The structure of basic skill modules is illustrated in Figure \ref{fig:basic}. Given a sequence $c_1, c_2, \cdots, c_l$ containing decimal and arithmetic characters of length $l$. We firstly map the sequence with character embeddings to $e_{c_1}, e_{c_2}, \cdots, e_{c_l}$. Then the inputs are fed into a bi-directional RNN (Bi-RNN). Outputs are generated by choosing characters with the max probability after $\mathrm{softmax}$ functions. Basic skill modules are trained in a supervised approach.
Each BSM provides a deterministic policy $a_{1:l}=\pi(c_{1:l})$, where $a_{1:l}$ is the calculated result in form of a digit character sequence.
\paragraph{Interactive Skill Modules (ISM)}
The interactive skill modules perform the arithmetic operations by invoking other skill modules. An example of interactive skill modules is shown in Figure \ref{fig:sampleISM}.
The policy of ISM is to select other skill modules to complete the partial arithmetic calculation.
Different from the standard HRL, each skill module performs a local arithmetic operation, and need not observe the global environment state. Therefore, when a skill module $\mathcal{M}_i$ chooses another skill module $\mathcal{M}_j$ as sub-policy, module $\mathcal{M}_i$ just sends character sequences to module $\mathcal{M}_j$ and receives character sequences as answers.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth,height=17em]{figs/memory_interactions.png}
\caption{Detailed structure of Interactive Skill Modules (ISM). In each time stamp, a bi-RNN encodes memory and output $o_1, o_2, \cdots, o_l$. The first and last outputs are concatenated as memory representations and are fed into the central RNN. The outputs of central RNN are used to select modules, generate read pointers and write pointers. The read pointers are used to read the sub-expression from memory. The selected module deal with the sub-expression and returns an answer. Then the answer is written to memory at positions indicated by write pointers. The dashed orange lines represent processes that are related to the policy of ISM.}\label{fig:memory}
\end{figure}
It is hard to train skill modules from scratch, so we use curriculum learning, which will be described in Section \ref{sec:curri}, to train skill modules in the order of increasing difficulty. Suppose that we already have $i-1$ well-trained skill modules $\Omega_\mathcal{M}=\{\mathcal{M}_1,\mathcal{M}_2,\cdots,\mathcal{M}_{i-1}\}$, the $i$-th ISM $\mathcal{M}_i$ is described as follows.
\subsubsection{Structure of Interactive Skill Module}
The detailed structure of ISM is shown in Figure \ref{fig:memory}.
First, each ISM is equipped with a memory $\xi$ to hold temporary information. Memory is composed of character slots with length $l$. When module $\mathcal{M}_i$ receives an expression $c_{1:l}$ , $\mathcal{M}_i$ first stores $c_{1:l}$ into memory.
The policy of ISM can be decomposed into three sub-policies: (1) selecting skill module, (2) reading memory, and (3) writing memory.
At time $t$, the memory contains characters $c_1^t, c_2^t, \cdots, c_l^t$, we first use a Bi-RNN to encode the state of memory.
\begin{align}
(o_1^t, o_2^t, \cdots, o_l^t) &= \mathrm{Bi\textrm{-}RNN}(e_{c_1^t}, e_{c_2^t}, \cdots, e_{c_l^t}),
\end{align}
where $e_{c_k^t}$ is the embedding of character $c_k^t$ for $1\leq k \leq l$.
The state $s^t$ of the environment is modeled by a forward RNN,
\begin{align}
h^t = \mathrm{RNNCell}[o_1^t\oplus o_l^t, h^{t-1}], \qquad
s^t = \mathrm{FNN}(h^t),\label{eq:memory1}
\end{align}
where $\mathrm{FNN}(\cdot)$ is one-layer forward neural network.
Given the state $s^t$, the agent chooses three actions according the three following sub-policies,
\begin{align}
\mathcal{M}^t &= \mathrm{ModuleSelect}([\mathcal{M}_1, \mathcal{M}_2, \cdots, \mathcal{M}_{i-1}], s^t),\\
R^t &= \mathrm{ReadPointer}([c_1, c_2, \cdots, c_l], s^t), \\
W^t &= \mathrm{WritePointer}([c_1, c_2, \cdots, c_l], s^t),
\label{eq:memory2}
\end{align}
where $\mathcal{M}^t, R^t, W^t$ denote the chosen module, the read pointers and write pointers at time $t$. $\mathrm{ModuleSelect}$, $\mathrm{Read}$, and $\mathrm{Write}$ are pointer functions described in Pointer Networks \cite{Vinyals2015PointerN}. Practically, there are two pairs of read pointers and one pair of write pointers specifying start and end positions of reading and writing. Additionally, Positional Embedding \cite{Vaswani2017AttentionIA} is combined with character embedding to provide the model with relative positional information.
Then the read pointer $R^t$ reads a sub-expression $\hat{e}_{1:p}^t$ from the memory and sends $\hat{e}_{1:p}^t$ to the selected module $\mathcal{M}^t$.
\begin{align}
\hat{e}_{1:p}^t &= \mathrm{Read}([c_1, c_2, \cdots, c_l], R^t),\\
\hat{c}_{1:q}^t &= \hat{M}^t(\hat{e}_{1:p}^t),
\label{eq:memory}
\end{align}
where $\hat{c}_{1:q}^t$ is output of module $\mathcal{M}^t$, which is further written into memory.
\begin{align}
(c_1^{t+1}, c_2^{t+1}, \cdots, c_l^{t+1}) &= \mathrm{Write}(\hat{c}_{1:q}^t, W^t).
\end{align}
\subsection{Optimization}
When the ISM generates the whole actions trajectory $\tau = (\mathcal{M}^1, R^1, W^1,\mathcal{M}^2, R^2, W^2,$ $ \cdots, \mathcal{M}^T, R^T, W^T)$, where $T$ number of the select skill modules, it can output an answer.
Finally, the ISM gets reward 1 when it gives the correct answer. If not, the reward is negative, based on character-level similarity to the solution.
Among reinforcement learning methods, Proximal Policy Optimization (PPO) \cite{schulman2017proximal} is an online policy gradient approach that achieves state-of-the-art on many benchmark tasks. Therefore, we implement PPO to train ISMs. We sample from policies $\pi_{\theta}$ where $\theta$ denotes model parameters. With every state $s$ and sampled action $a=\{\mathcal{M}, R, W\}$, we compute gradients to maximize the following objective function:
\begin{align}
O(\theta) & = O_{PPO}(\theta, \hat{A}) + \alpha * H(\pi_{\theta}(a|s)),
\end{align}
where $\hat{A}$ is the advantage function representing the discounted reward, $H(\cdot)$ is entropy regularizer to discourage premature convergence to suboptimal policies \cite{mnih2016asynchronous}, and $\alpha$ is the coefficient to balance the Exploration-Exploitation, which will be mentioned in CTCS framework (see Section \ref{sec:curri}).
\subsection{Curriculum Teacher and Continual-learning Student (CTCS)}\label{sec:curri}
We propose Curriculum Teacher and Continual-learning Student (CTCS) framework to help the model acquire knowledge efficiently.
The CTCS framework is illustrated in Figure \ref{fig:ctcs}. Given a set of $N$ tasks $\{t_1, t_2, \cdots, t_{N}\}$ that are ordered by increasing difficulty. Each task contains $M$ data samples: $t_i = \{x_1^{(i)}, x_2^{(i)}, \cdots, x_{M}^{(i)}\}$). The curriculum teacher gives tasks in the order of $1, 2, \cdots, N$, switching to the next task only when the student performs perfectly in the current task. In learning every task, the curriculum teacher uses difficulty sampling strategy to sample from $M$ data samples.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\linewidth]{figs/overview.png}
\caption{Curriculum Teacher and Continual-learning Student (CTCS) framework.}\label{fig:ctcs}
\end{figure}
\textbf{Difficulty Sampling} encourages learning difficult samples. Unlike most problems, arithmetic learning needs precise calculation, which requires complete mastery of training samples. However, the model tends to gain good performance, but not perfect scores. Inspired by Deliberate Practice, a common learning method for human beings, we use difficulty sampling to help the student achieve complete mastery.
To formalize, a difficulty score $d_j^{(i)}$ is the total number of incorrect attempts of sample $j$. Then the probability of each sample $p_{j}^{(i)}$ is determined by a parameterized $Softmax$ function:
\begin{equation}
p_{j}^{(i)} = {\frac {\exp(d_{j}^{(i)}/\tau)}{\sum _{k=1}^{M}\exp(d_{k}^{(i)}/ \tau )}}.
\label{eq:normalize}
\end{equation}
\textbf{Parameter Adjustment} encourages or discourages the exploration of the student. In reinforcement learning, adding entropy controlled by a coefficient $\alpha$ to loss is a commonly used technique \cite{mnih2016asynchronous} to discourage premature convergence to suboptimal policies. However, to what extent should we encourage the student to explore is a long-standing issue of Exploration-Exploitation Dilemma \cite{kaelbling1996reinforcement}. Intuitively, exploration should be encouraged when the student has difficulty doing some samples. Therefore, we employ the teacher to help the student change exploration strategy in keeping with its performance. To be specific, the entropy coefficient $\alpha$ is:
\begin{equation}
\alpha = \min(\beta , \gamma * \max_{j}{d_j^{(i)}})
,\label{eq:entropy}
\end{equation}
where $d_j^{(i)}$ is the difficulty score described in difficulty sampling, $\beta = 0.5$ and $\gamma = 0.01$. As shown in Section \ref{sec:discussion}, difficulty sampling and parameter adjustment methods are critical in achieving the perfect performance.
\section{Experiments}
\paragraph{Arithmetic Expression Calculation}
To train our model to calculate arithmetic expressions with curriculum learning, we define several sub-tasks, from basic tasks like the single-digit addition to compositional tasks like multi-digit division. Then we train our model with tasks in the order of increasing difficulty. The full curriculum list is shown in the appendix. The code is available here\footnote{The code is at (Anonymous).}.
The arithmetic expression data is generated through a random process. An expression of length 10 contains approximately 3 arithmetic operators of $\{+, -, \times, \div, (, )\}$ in average.
We compare our model with two baseline models:
\begin{itemize}
\item Seq2seq LSTM: A sequence to sequence model \cite{sutskever2014sequence} with LSTM \cite{Hochreiter1997LongSM} as encoder and decoder.
\item Neural GPU: An arithmetic algorithm learning model proposed by \citet{kaiser2015neural}. We use their open source implementation posted on Github.
\end{itemize}
To make an objective comparison, we also apply the same curriculum learning method to baseline models. The results are shown in Figure \ref{fig:train} and Figure \ref{fig:test}.
\begin{figure}[th]
\centering
\includegraphics[width=0.9\linewidth]{figs/train.png}
\caption{Accuracy in training set of size 1000. Training set is changing from easy tasks to difficult tasks as curriculum learning is applied. Every sudden drop indicates a task switching.}\label{fig:train}
\end{figure}
\begin{figure}[th]
\centering
\includegraphics[width=0.9\linewidth]{figs/validation.png}
\caption{Accuracy in test set of size 1000. Test set contains arithmetic expressions of length 10. As the figure shows, the accuracy of both Neural GPU and LSTM is constantly nearly zero.}\label{fig:test}
\end{figure}
\begin{table}[th]\small
\centering
\subfloat[][Addition.]{
\begin{tabular}{llll}
\toprule
Length & 5 & 10 & 20 \\
\midrule
Ours & \textbf{100\%} & \textbf{100\%} & 17\% \\
Neural GPU & \textbf{100\%} & 98\% & \textbf{84\%} \\
LSTM & 61\% & 33\% & 16\% \\
\bottomrule
\end{tabular}}
\qquad
\subfloat[][Substraction.]{
\begin{tabular}{llll}
\toprule
Length & 5 & 10 & 20 \\
\midrule
Ours & \textbf{100\%} & \textbf{100\%} & 19\% \\
Neural GPU & \textbf{100\%} & 72\% & \textbf{43\%} \\
LSTM & 95\% & 48\% & 20\% \\
\bottomrule
\end{tabular}}
\qquad
\subfloat[][Multiplication.]{
\begin{tabular}{llll}
\toprule
Length & 5 & 10 & 20 \\
\midrule
Ours & \textbf{100\%} & \textbf{100\%} & 0\% \\
Neural GPU & 30\% & 3\% & 0\% \\
LSTM & 12\% & 3\% & 0\% \\
\bottomrule
\end{tabular}}
\qquad
\subfloat[][Division.]{
\begin{tabular}{llll}
\toprule
Length & 5 & 10 & 20 \\
\midrule
Ours & \textbf{100\%} & 27\% & 15\% \\
Neural GPU & 30\% & 29\% & 21\% \\
LSTM & 28\% & 23\% & 19\% \\
\bottomrule
\end{tabular}}
\qquad
\subfloat[][Expression calculation.]{
\begin{tabular}{llll}
\toprule
Length & 5 & 10 & 20 \\
\midrule
Ours & \textbf{100\%} & \textbf{100\%} & \textbf{78\%} \\
Neural GPU & 48\% & 2\% & 0\% \\
LSTM & 5\% & 0\% & 0\% \\
\bottomrule
\end{tabular}}
\vspace{1em}
\caption{Results of addition, subtraction, multiplication, division, and expression calculation at different length of arithmetic expression in the test set. Task expressions contain one specific arithmetic operator except for the expression calculation task.}
\label{tab:per}
\end{table}
As the result shows, both the baseline models are striving to remember training samples, achieving relatively high accuracy in the training set, but nearly zero accuracy in the test set.
The LSTM model shows a powerful capability of remembering training samples. Every time the task switches, the performance suddenly drops down to zero and then increases to a high level. Although the Neural GPU seems to have better generalization ability, it still performs poorly in the test set.
In contrast, our model achieves almost 100 percent correctness in the experiment, which shows the effectiveness and generalization ability of our model.
\paragraph{Sub-task Performance}
We evaluate our model with different sub-tasks to see the performance of various arithmetic operations. The results are shown in Table \ref{tab:per}. It's noteworthy that the answer of the division is relatively small, so the models can guess the answer, resulting in nearly 20\% correctness in the division. As the result shows, our model achieves 100\% mastery much more than baseline models, especially in expression calculation task.
\paragraph{Hyperparameters}
The gradient-based optimization is performed using the Adam update rule \cite{kingma2014adam}. Every RNN in our model is GRU \cite{chung2014empirical} with hidden size 100. $\tau$ used in Equation \ref{eq:normalize} is 10. The consecutively sample number $N_c$ described in difficulty sampling is 64. In PPO, the reward discount parameter is 0.99, and the clipping parameter $\epsilon$ is 0.2.
\section{Discussion}\label{sec:discussion}
\subsection{Ablation Study}
\paragraph{Curriculum Learning and Continual Learning}
To test if our model can make use of prior knowledge when meeting a new task, we challenge our model with learning a new arithmetic operation: modular. We compare our proposed model with a baseline model that learns from scratch. The results are shown in Figure \ref{fig:curriculum}. Without curriculum learning and continual learning, the model fails to give any correct solutions. It shows the necessity of curriculum learning and continual learning.
\begin{figure}[h]
\centering
\includegraphics[width=0.78\linewidth]{figs/ablation_curriculum.png}
\caption{Results for ablation of curriculum learning and continual learning. The test set contains arithmetic expressions of length 10 with the modular operator. }\label{fig:curriculum}
\end{figure}
\paragraph{Difficulty Sampling and Parameter Adjustment}
In Curriculum Teacher Continual-learning Student (CTCS) framework, we present difficulty sampling and parameter adjustment to help the model produce the perfect performance. The effectiveness of them is illustrated in Figure \ref{fig:sampling}. Without difficulty sampling and parameter adjustment, the model shows convergence in suboptimal strategy. It shows that difficulty sampling and parameter adjustment are important in helping the model to achieve perfect mastery.
\begin{figure}[H]
\centering
\includegraphics[width=0.78\linewidth]{figs/ablation_difficulty_sampling.png}
\caption{Results for ablation of Difficulty Sampling and Parameter Adjustment. The test set contains arithmetic expressions of length 10.}\label{fig:sampling}
\end{figure}
\section{Conclusion}
In this paper, we propose a pure neural model to solve the arithmetic expression calculation problem. Specifically, we use the Multi-level Hierarchical Reinforcement Learning (MHRL) framework to factorize a complex arithmetic operation into several simpler operations.
We also present Curriculum Teacher Continual-learning Student (CTCS) framework where the teacher adopts difficulty sampling and parameter adjustment strategies to supervise the student. All these above contribute to solving the arithmetic expression calculation problem. Experiments show that our model significantly outperforms previous methods for arithmetic expression calculation.
\bibliographystyle{unsrtnat}
|
\section{Introduction}
\label{sec1}
The late-time cosmic acceleration was first discovered
in 1998 from the observations of distant supernovae
type Ia (SN Ia) \cite{SN1,SN2}.
The source for this phenomenon, which was dubbed
dark energy \cite{Huterer}, occupies about
70~\% of today's energy density of the Universe.
The existence of dark energy has been independently
confirmed from the observation data of
Cosmic Microwave Background (CMB) \cite{WMAP,WMAP9,Planck,Planck2015,Planck2018}
and baryon acoustic oscillations (BAO) \cite{BAO}.
Despite the tremendous observational and theoretical
progress over the past two
decades \cite{dreview1,dreview2,dreview2a,dreview3,dreview4,dreview5,Tsuji10,dreview6,dreview7,deRham12,dreview8,Joyce16,dreview9,dreview10,dreview11,dreview12}, the origin of dark energy
has not been identified yet.
A key quantity describing the property of dark energy is
the equation of state (EOS)
$w_{\rm DE}=P_{\rm DE}/\rho_{\rm DE}$,
where $\rho_{\rm DE}$ and $P_{\rm DE}$ are the dark energy density and pressure respectively.
If we use the EOS parametrization of the form
$w_{\rm DE}(a)=w_0+(1-a)w_a$ \cite{CP,Linder},
where $w_0, w_a$ are constants and $a$ is a scale
factor (with today's value $a=1$), the joint data analysis of
Planck 2018 combined with the SN Ia and BAO data placed the bounds
$w_0=-0.961 \pm 0.077$ and $w_a=-0.28^{+0.31}_{-0.27}$
at 68\,\% confidence level (CL) \cite{Planck2018}.
This shows the overall consistency with the EOS
$w_{\rm DE}=-1$, but the time variation of $w_{\rm DE}$
around $-1$ is also allowed from the data.
We also note that the parametrization $w_{\rm DE}(a)=w_0+(1-a)w_a$ is not necessarily
versatile to accommodate models with
a fast varying dark energy EOS \cite{Bruce1,Cora03,Bruce2,LH05} or models
in which $w_{\rm DE}$ has an extremum \cite{Savvas12,Lasenby}.
In the theoretical side,
the simplest candidate for dark energy is the
cosmological constant $\Lambda$ characterized
by the EOS $w_{\rm DE}=-1$.
If the cosmological constant originates from the
vacuum energy in particle physics, however,
the theoretically predicted value is
enormously larger than the observed dark energy
scale \cite{Weinberg,MartinLa,Tony}.
There is a local theory of the vacuum energy sequestering
in which quantum vacuum energy is cancelled by an auxiliary
four-form field \cite{Kaloper2}
(see also Ref.~\cite{Kaloper1}
for a nonlocal version).
In this formulation, what is left after the vacuum energy sequestering is
a radiatively stable residual cosmological
constant $\Delta \Lambda$.
The value of $\Delta \Lambda$ is not uniquely
fixed by the underlying theory, but it should be measured to match with observations.
To explain today's cosmic acceleration, it is fixed as $\Delta \Lambda \simeq
10^{-12}$ eV$^4$.
If the cosmological constant problem is solved in such
a way that the residual vacuum energy completely vanishes or it is much smaller than today's energy density of the Universe,
we need to find an alternative mechanism for explaining the
origin of dark energy.
There are dynamical models of dark energy
in which $w_{\rm DE}$ changes in time.
The representative example
is a minimally coupled scalar field $\phi$
with a potential energy
$V(\phi)$, which is dubbed
quintessence \cite{quin1,quin2,Ratra,quin3,quin4,quin4b,quin5,Joyce,quin6,quin7}.
The quintessence EOS varies
in the region $w_{\rm DE}>-1$.
While the ``freezing'' quintessence models \cite{Caldwell} in
which the deviation of $w_{\rm DE}$ from $-1$ has been large by today are in strong tension with observations,
the ``thawing'' models in which $w_{\rm DE}$ starts to
deviate from $-1$ at late times have been consistent
with the data for today's EOS
$w_{\rm DE}(a=1) \lesssim -0.7$ \cite{CDT}.
We also have other minimally coupled scalar-field theories
dubbed k-essence in which the Lagrangian $G_2$ is
a general function of $\phi$ and its derivative
$X=-\partial^{\mu} \phi \partial_{\mu}
\phi/2$ \cite{kinf,kes1,kes2,kes3}.
One of the examples of k-essence is the dilatonic
ghost condensate model
$G_2=-X+ce^{\lambda \phi}X^2$ \cite{gcon1,gcon2}
($c$ and $\lambda$ are constants), in which case
the typical evolution of $w_{\rm DE}$ is similar to
that in thawing quintessence.
It is also possible to construct unified models of dark energy
and dark matter in terms of a purely kinetic
Lagrangian $G_2(X)$ \cite{unifiedkes,Ber1}.
There are also theories in which the scalar field is nonminimally
coupled to the Ricci scalar $R$ in the form
$G_4(\phi) R$ \cite{Fujiibook}.
One of the representative examples is Brans-Dicke (BD)
theory \cite{Brans} with a scalar potential $V(\phi)$.
This theory is given by the Lagrangian
$L=\phi R/2+\omega_{\rm BD}X/\phi-V(\phi)$, where
$\omega_{\rm BD}$ is a constant called the BD parameter.
The $f(R)$ gravity whose equations of motion are derived under
the variation of metric tensor
$g_{\mu \nu}$ \cite{Bergmann,Ruz,Staro}
corresponds to the special case of BD theory with
$\omega_{\rm BD}=0$ \cite{Ohanlon,Chiba03}.
The application of $f(R)$ gravity to the late-time
cosmic acceleration has been extensively performed
in the literature \cite{fRearly1,fRearly1b,fRearly2,fRearly3,fR1,fR2,fR3,fR4,fR5}.
In this case, it is possible to
realize the dark energy EOS
smaller than $-1$ without having ghosts \cite{fR1,fR4,Moto}.
Moreover, the growth rate of matter perturbations is
larger than that in General Relativity (GR) \cite{fR2,fR4,Tsuji07}.
Hence it is possible to observationally distinguish
$f(R)$ gravity models from quintessence and k-essence.
The construction of dark energy models in BD theories with
the scalar potential is also possible
\cite{Brax04,BD1,BD2},
in which case observational signatures are different
depending on the BD parameter.
In these models, the propagation of fifth forces can be
suppressed in local regions of the
Universe \cite{Bunn,Navarro,fR1,Capo07,Tamaki,Brax08}
under the chameleon mechanism \cite{chame1,chame2}.
There are other modified gravity theories dubbed
Galileons \cite{Nicolis,Galileons,Galileons1} containing scalar
derivative self-interactions and nonminimal couplings to gravity.
In the limit of Minkowski space, the equations of motion
following from the Lagrangian of covariant Galileons
are invariant under the Galilean shift
$\partial_{\mu} \phi \to \partial_{\mu} \phi+b_{\mu}$
(see Ref.~\cite{brokengal} for the weakly broken version of Galilean shift symmetry).
The cubic Galileon of the form $X \square \phi$ arises
in the Dvali-Gabadadze-Porrati braneworld model
due to the mixture between longitudinal and transverse
gravitons \cite{DGP} and also in the Dirac-Born-Infeld decoupling theory
with bulk Lovelock invariants \cite{DTolley}.
This derivative self-interaction can suppress the propagation
of fifth forces in local regions of the Universe \cite{Cede,Luty,Ratta,Babichev,Babichev1,Burrage,Tana,Brax,DKT12,KKY12,deRham,Hira,Koyama13,Andrews,Babi13,Kase13}
under the Vainshtein mechanism \cite{Vain},
while modifying the gravitational interaction
at cosmological distances \cite{Kase10}.
For covariant Galileons including quartic and quintic
Lagrangians, there exist self-accelerating de Sitter attractors responsible
for the late-time cosmic acceleration \cite{Sami,DT10,DT10b}.
The self-accelerating solution of full covariant Galileons preceded by
a tracker solution with $w_{\rm DE}=-2$ is in tension with observational
data of SN Ia, CMB, and BAO \cite{NDT10,AppleLin,Neveu,Barreira1,Barreira2,Renk,Peirone2}.
However, this is not necessarily the case for covariant Galileons with a linear
potential \cite{KTD15} or the model in which an additional term $X^2$ is present to
the Galileon Lagrangian \cite{Kase18}.
The aforementioned theories belong to a subclass of more
general scalar-tensor theories--dubbed Horndeski
theories \cite{Horndeski}.
In 2011, Deffayet {\it et al.} \cite{Def11} derived the action
of most general scalar-tensor theories with second-order
equations of motion after the generalizations of
covariant Galileons.
Kobayashi {\it et al.} \cite{KYY} showed that the corresponding
action is equivalent to that derived by Horndeski
in 1974 \cite{Horndeski} (see also Ref.~\cite{Char11}).
The application of Horndeski theories to the late-time cosmic
acceleration was extensively performed
in the literature \cite{DKT11,DT12,Char12,Zhao,Cope12,Amen12,Zuma13,Sil13,Raveri14,Hu13,OTT13,HKi14,Bellini,Zuma17,Ken,Peirone,Ken}.
Since different modified gravity theories predict different background
expansion and cosmic growth histories, it is possible to distinguish between
dark energy models in Horndeski theories from
the observations of
SN Ia, CMB, BAO, large-scale structures, and weak lensing.
The recent detection of gravitational waves (GWs) by
GW170817 \cite{GW170817} from a binary neutron star
merger together with the gamma-ray burst GRB
170817A \cite{Goldstein} constrained the propagation
speed $c_t$ of GWs, as \cite{Abbott}
\begin{equation}
-3 \times 10^{-15} \le c_t-1 \le
7 \times 10^{-16}\,,
\label{ctbound}
\end{equation}
for the redshift $z \le 0.009$.
If we strictly demand that $c_t=1$ and do not allow any tuning
among functions in Horndeski theories, the Lagrangian
needs to be of the form
$L=G_2(\phi,X)+G_{3}(\phi,X)\square\phi+G_4(\phi) R$
\cite{Lon15,GWcon1,GWcon2,GWcon3,GWcon4,GWcon5,GWcon6}\footnote{
Even in the presence of quartic-order derivative and quintic-order couplings there are some specific models
consistent with the bound (\ref{ctbound}), see, e.g.,
Ref.~\cite{Amendola18} for interacting dark energy and dark matter models in which $c_t$ approaches 1 by today.}.
This includes the theories such as quintessence, k-essence,
$f(R)$ gravity, BD theories, and cubic Galileons, but the theories
with quartic-order derivative and quintic-order couplings
do not belong to this class\footnote{
This assumes that the effective field theory of dark energy
with derivative interactions is valid up to the energy scale
corresponding to the
gravitational-wave frequency observed by LIGO/Virgo
($f \sim 100$~Hz). If the effective field theory is modified
above a strong coupling scale (which is typically of the
same order as the LIGO/Virgo frequency) in such a way that
$c_t$ approaches 1, we may evade the
observational bound (\ref{ctbound}) \cite{deRham18}.
}.
Now, we are entering the era in which
dark energy models in Horndeski theories
can be tightly constrained from observations.
In this article, we review the application of Horndeski theories
to cosmology and discuss observational signatures of surviving
dark energy models. After reviewing Horndeski theories in
Sec.~\ref{Hornsec}, we devote Sec.~\ref{backsec} for deriving
the background equations of motion
and the second-order action of tensor perturbations
on the flat Friedmann-Lema\^{i}tre-Robertson-Walker
(FLRW) spacetime in the presence of matter.
In Sec.~\ref{scalarsec}, we obtain the second-order action of
scalar perturbations and their equations of motion
in full Horndeski theories without fixing any gauge conditions.
We note that this gauge-ready formulation was carried out in a subclass
of Horndeski theories \cite{Hwang} and
in scalar-vector-tensor theories \cite{HKT18}.
The formalism developed in Ref.~\cite{HKT18} encompasses both
Horndeski gravity and generalized Proca
theories \cite{Heisenberg14,Tasinato,Allys,Jimenez} as specific cases.
In Sec.~\ref{stasec}, we derive conditions for the absence of ghost and
Laplacian instabilities of scalar perturbations in the small-scale limit
by choosing three different gauges and show that these conditions are
independent of the gauge choices.
In Sec.~\ref{obsersec}, we also apply the results in Sec.~\ref{scalarsec}
to the computation of observables relevant to the linear growth of matter
density perturbations and the evolution of gravitational potentials in the gauge-invariant way.
We also classify surviving dark energy models into four classes
depending on their observational signatures.
In Secs.~\ref{quinsec}, \ref{fRBDsec}, \ref{classCsec}, and \ref{mostsec},
we review the behavior of dark energy observables in each class of theories, i.e.,
(A) quintessence and k-essence, (B) $f(R)$ gravity and BD theories,
(C) kinetic braidings, and (D) nonminimally coupled scalar with
cubic derivative interactions. This is convenient to distinguish between
surviving dark energy models in current and future observational data.
We conclude in Sec.~\ref{concludesec}.
Throughout the review, we adopt the metric signature
$(-,+,+,+)$. We also use the natural unit in which the
speed of light $c$, the reduced Planck constant $\hbar$,
and the Boltzmann constant $k_B$ are equivalent to 1.
The reduced Planck mass $M_{\rm pl}$ is related to the
Newton gravitational constant $G$, as
$M_{\rm pl}=1/\sqrt{8 \pi G}$.
\section{Horndeski theories}
\label{Hornsec}
The theories containing a scalar field $\phi$ coupled to gravity
are generally called scalar-tensor theories \cite{Fujiibook}.
This reflects the fact that the gravity sector has two tensor polarized degrees of freedom.
Horndeski theories \cite{Horndeski} are the most general
scalar-tensor theories with second-order equations of motion.
Due to the second-order property, there is no Ostrogradski
instability \cite{Ostro,Woodard} associated with
the Hamiltonian unbounded from below.
Horndeski theories are given by the action \cite{KYY}
\begin{equation}
{\cal S}_{\rm H}=\int d^4 x \sqrt{-g}\,L\,,
\label{action}
\end{equation}
where $g$ is a determinant of the metric tensor $g_{\mu \nu}$, and
\begin{eqnarray}
\hspace{-0.8cm}
L &=& G_2(\phi,X)+G_{3}(\phi,X)\square\phi
+G_{4}(\phi,X)\, R +G_{4,X}(\phi,X)\left[ (\square \phi)^{2}
-(\nabla_{\mu}\nabla_{\nu} \phi)
(\nabla^{\mu}\nabla^{\nu} \phi) \right] \nonumber \\
\hspace{-0.8cm}
& &
+G_{5}(\phi,X)G_{\mu \nu} \nabla^{\mu}\nabla^{\nu} \phi
-\frac{1}{6}G_{5,X}(\phi,X)
\left[ (\square \phi )^{3}-3(\square \phi)\,
(\nabla_{\mu}\nabla_{\nu} \phi)
(\nabla^{\mu}\nabla^{\nu} \phi)
+2(\nabla^{\mu}\nabla_{\alpha} \phi)
(\nabla^{\alpha}\nabla_{\beta} \phi)
(\nabla^{\beta}\nabla_{\mu} \phi) \right].
\label{LH}
\end{eqnarray}
Here, the symbol $\nabla_{\mu}$ stands for the covariant derivative operator with $\square \equiv \nabla^{\mu}\nabla_{\mu}$,
$R$ is the Ricci scalar, $G_{\mu \nu}$ is the Einstein tensor, and
\begin{equation}
X \equiv -\frac12 \nabla^{\mu}\phi \nabla_{\mu}\phi\,.
\label{defX}
\end{equation}
The functions $G_{2,3,4,5}$ depend on $\phi$ and $X$, with
$G_{i,\phi} \equiv \partial G_i/\partial \phi$ and
$G_{i,X} \equiv \partial G_i/\partial X$.
Originally, Horndeski derived the Lagrangian of
scalar-tensor theories with second-order equations
of motion in a form different from Eq.~(\ref{LH}) \cite{Horndeski},
but their equivalence was explicitly shown
in Ref.~\cite{KYY}.
Depending on the papers \cite{KYY,DT12,Bellini,Gleyzes14},
different signs and notations were used for
the quantities $G_3(\phi,X)$ and $X$, so we summarize them
in Appendix~\ref{notation} to avoid confusion.
Below, we list the theories within the framework of
the action (\ref{action}).
\begin{itemize}
\item (1) Quintessence and k-essence
\vspace{0.1cm}
K-essence \cite{kinf,kes1,kes2,kes3} is given by the functions
\begin{equation}
G_2=G_2(\phi,X)\,,\qquad G_3=0\,,\qquad
G_4=\frac{M_{\rm pl}^2}{2}\,,\qquad G_5=0\,.
\end{equation}
Quintessence \cite{quin1,quin2,Ratra,quin3,quin4,quin4b,quin5,Joyce,quin6,quin7}
corresponds to the particular choice:
\begin{equation}
G_2=X-V(\phi)\,,
\end{equation}
where $V(\phi)$ is the potential of $\phi$.
\item (2) BD theory
\vspace{0.1cm}
In BD theory \cite{Brans} with the scalar potential
$V(\phi)$, we have
\begin{equation}
G_2=\frac{M_{\rm pl}\omega_{\rm BD}}{\phi}X-V(\phi)\,,
\qquad G_3=0\,,\qquad
G_4=\frac{1}{2}M_{\rm pl}\phi\,,\qquad G_5=0\,.
\label{BDaction}
\end{equation}
In the limit that $\omega_{\rm BD} \to \infty$,
we recover GR with a quintessence scalar field.
We note that there are more general nonminimally coupled
theories given by the couplings
$G_2=\omega (\phi)X-V(\phi)$, $G_3=0$, $G_4=F(\phi)$,
$G_5=0$ \cite{Gas92,Damour1,Damour2,Amen99,Uzan,Chiba99,Bartolo99,Perrotta,Boi00,Gille}.
Since the basic structure of such theories is similar
to that in BD theories, we do not discuss them in this review.
\item (3) $f(R)$ gravity
\vspace{0.1cm}
The action of $f(R)$ gravity \cite{Bergmann,Ruz,Staro}
is given by
\begin{equation}
{\cal S}_{\rm H}=\int d^4x \sqrt{-g}\,\frac{M_{\rm pl}^2}{2}f(R)\,,
\label{fRaction}
\end{equation}
where $f(R)$ is an arbitrary function of $R$.
The metric $f(R)$ gravity, which corresponds to the variation
of (\ref{fRaction}) with respect to $g_{\mu \nu}$,
can be accommodated by the Lagrangian (\ref{LH})
for the choice
\begin{equation}
G_2=-\frac{M_{\rm pl}^2}{2} (RF-f),\qquad
G_3=0\,,\qquad G_4=\frac{M_{\rm pl}^2}{2}F\,,\qquad G_5=0\,,
\label{fRcase}
\end{equation}
where $F(R) \equiv \partial f/\partial R$.
In this case, the scalar degree of freedom
$\phi=M_{\rm pl}F(R)$ arises from the gravity sector.
Comparing Eq.~(\ref{BDaction}) with Eq.~(\ref{fRcase}),
it follows that metric $f(R)$ gravity is equivalent to BD theory with
$\omega_{\rm BD}=0$ and the scalar potential
$V=(M_{\rm pl}^2/2) (RF-f)$ \cite{Ohanlon,Chiba03}.
\item (4) Covariant Galileons
\vspace{0.1cm}
In original Galileons \cite{Nicolis}, the field
equations of motion are invariant under the shift
$\partial_{\mu} \phi \to \partial_{\mu} \phi+b_{\mu}$
in Minkowski spacetime \cite{Nicolis}.
In curved spacetime, the Lagrangian of covariant
Galileons \cite{Galileons} is constructed
to keep the equations of motion up to second order, while recovering
the Galilean shift symmetry in the Minkowski limit.
Covariant Galileons are characterized by the functions
\begin{equation}
G_2=\beta_1 X-m^3 \phi\,,\qquad G_3=\beta_3 X\,,\qquad
G_4=\frac{M_{\rm pl}^2}{2}+\beta_4 X^2\,,\qquad
G_5=\beta_5 X^2\,,
\end{equation}
where $\beta_{1,3,4,5}$ and $m$ are constants.
In absence of the linear potential $V(\phi)=m^3 \phi$,
there exists a self-accelerating de Sitter solution
satisfying $X={\rm constant}$ \cite{DT10,DT10b,Ginf1,Ginf2}.
\item (5) Derivative couplings
\vspace{0.1cm}
There is also a derivative coupling theory in which
the scalar field couples to the Einstein tensor of the form
$G_{\mu \nu}\nabla^{\mu}\phi \nabla^{\nu} \phi$ \cite{Amen93,Germani}.
This corresponds to the choice
\begin{equation}
G_2=X-V(\phi)\,,\qquad G_3=0\,,\qquad
G_4=\frac{M_{\rm pl}^2}{2}\,,\qquad
G_5=c \phi\,,
\end{equation}
where $V(\phi)$ is a scalar potential and
$c$ is a constant.
Integrating the term $c\phi\,G_{\mu \nu} \nabla^{\mu} \nabla^{\nu}\phi$ by parts, it is equivalent to
$-c\,G_{\mu \nu}\nabla^{\mu} \phi \nabla^{\nu} \phi$
up to a boundary term.
\item (6) Gauss-Bonnet couplings
\vspace{0.1cm}
One can consider a coupling of the form
$\xi(\phi){\cal G}$ \cite{Lovelock}, where
$\xi(\phi)$ is a function of $\phi$ and ${\cal G}$ is
the Gauss-Bonnet curvature invariant defined by
\begin{equation}
{\cal G}=R^2-4R_{\alpha \beta}R^{\alpha \beta}
+R_{\alpha \beta \gamma \delta}R^{\alpha \beta \gamma \delta}\,.
\end{equation}
The theories given by the action \cite{NOS,Mota,Mota2,TS06}
\begin{equation}
{\cal S}_{\rm H}=\int d^4 x \sqrt{-g}
\left[ \frac{M_{\rm pl}^2}{2}R+X-V(\phi)+\xi(\phi){\cal G} \right]
\end{equation}
can be accommodated in the framework of Horndeski theories
for the choice \cite{KYY}
\begin{eqnarray}
&&
G_2=X-V(\phi)+8\xi^{(4)} (\phi) X^2 \left( 3-\ln X \right)\,,\qquad
G_3=-4\xi^{(3)} (\phi) X \left( 7-3\ln X \right)\,,\nonumber \\
& &
G_4=\frac{M_{\rm pl}^2}{2}+
4\xi^{(2)}(\phi)X \left( 2-\ln X \right)\,,\qquad
G_5=-4\xi^{(1)} (\phi) \ln X\,,
\end{eqnarray}
where $\xi^{(n)}(\phi) \equiv \partial^n \xi (\phi)/\partial \phi^n$.
\item (7) $f({\cal G})$ gravity
\vspace{0.1cm}
There is also modified gravitational theories
given by the action \cite{NO05,DT08}
\begin{equation}
{\cal S}_{\rm H}=\int d^4 x \sqrt{-g}
\left[ \frac{M_{\rm pl}^2}{2}R+f({\cal G}) \right]\,,
\label{SfG}
\end{equation}
where $f$ is a function of ${\cal G}$.
Since the Lagrangian $f({\cal G})$ is equivalent to
$f(\phi)+f_{,\phi} ({\cal G}-\phi)$, the action (\ref{SfG}) belongs
to a subclass of Horndeski theories with the couplings
\begin{eqnarray}
&&
G_2=f(\phi)-f_{,\phi}\phi+8f^{(5)} (\phi) X^2 \left( 3-\ln X \right)\,,\qquad
G_3=-4f^{(4)} (\phi) X \left( 7-3\ln X \right)\,,\nonumber \\
& &
G_4=\frac{M_{\rm pl}^2}{2}+
4 f^{(3)}(\phi)X \left( 2-\ln X \right)\,,\qquad
G_5=-4 f^{(2)} (\phi) \ln X\,.
\end{eqnarray}
\item (8) Kinetic braidings and its extensions
\vspace{0.1cm}
There are theories given by the Lagrangian
\begin{equation}
L=G_2(\phi,X)+G_3(\phi,X) \square \phi+G_4(\phi)R\,.
\label{Lagkin}
\end{equation}
As we will show later, the Lagrangian (\ref{Lagkin}) corresponds
to most general Horndeski theories with the tensor propagation
speed $c_t$ equivalent to 1.
The kinetic braiding scenario \cite{braiding1,braiding2}
corresponds to the minimally coupled case, i.e., $G_4=M_{\rm pl}^2/2$.
The cubic Galileon given with
$L=\beta_1 X-m^3 \phi+\beta_3 X \square \phi+(M_{\rm pl}^2/2)R$
belongs to a subclass of kinetic braidings.
The dark energy scenario given by
$L=\beta_1 X+\beta_2 X^2+\beta_3 X \square \phi+(M_{\rm pl}^2/2)R$ \cite{Kase18}
is also in the framework of kinetic braidings.
In presence of the nonminimal coupling $G_4(\phi)R$,
it is known that a self-accelerating solution characterized by
$\dot{\phi}/\phi={\rm constant}$ exists for the model
$L=\omega (\phi/M_{\rm pl})^{1-n}X+(\lambda/\mu^3)
(\phi/M_{\rm pl})^{-n} X \square \phi+(M_{\rm pl}^2/2)
(\phi/M_{\rm pl})^{3-n}R$ with
$2 \le n \le 3$ \cite{Silva,Koba1,Koba2,DTge}.
There is also a nonminimally coupled model
given by
$L=\beta_1 \left( 1-6Q^2 \right) e^{-2Q \phi/M_{\rm pl}}X
-m^3 \phi+\beta_3 X \square \phi+(M_{\rm pl}^2/2)
e^{-2Q \phi/M_{\rm pl}}R$ \cite{Babi11,KTD15}, which recovers the minimally
coupled cubic Galileon in the limit $Q \to 0$.
\end{itemize}
Thus, Horndeski theories can accommodate a wide variety
of scalar-tensor theories with second-order equations of motion.
Except for quintessence and k-essence, the scalar field
has derivative self-interactions and nonminimal/derivative
couplings to gravity.
In such cases, we need to confirm whether
there are neither ghost nor Laplacian instabilities.
In Secs.~\ref{tensec} and \ref{stasec}, we will address this issue
after deriving the second-order actions of tensor and scalar perturbations
on the flat FLRW background in full Horndeski theories.
The same analysis also allows one to identify the speed of
gravitational waves on the isotropic cosmological background.
Under the observational bound (\ref{ctbound}),
this result can be used to pin down viable Horndeski theories
relevant to the late-time cosmic acceleration.
In Sec.~\ref{classsec}, we classify surviving dark energy models into
four classes depending on their observational signatures.
To discuss the dynamics of cosmic acceleration preceded
by the matter-dominated epoch, we take into account a matter
perfect fluid minimally coupled to gravity.
The vector perturbations are nondynamical in Horndeski theories,
so we only need to consider the evolution of scalar and tensor perturbations.
The perfect fluid without the vector degree of freedom can be
described by the Schutz-Sorkin action \cite{Sorkin,DGS}:
\begin{equation}
{\cal S}_m=-\int d^{4}x \left[ \sqrt{-g}\,\rho_m(n)
+J^{\mu} \partial_{\mu}\ell \right]\,,
\label{Spf}
\end{equation}
where $\rho_m$ is a fluid density, $\ell$ is
a scalar quantity, and $J^{\mu}$ is a four vector associated with
the fluid number density $n$, as
\begin{equation}
n=\sqrt{\frac{J^{\mu}J^{\nu}g_{\mu \nu}}{g}}\,.
\label{ndef}
\end{equation}
Varying the action (\ref{Spf}) with respect to $J^{\mu}$,
we obtain
\begin{equation}
u_\mu\equiv\frac{J_\mu}{n\sqrt{-g}}
=\frac{\partial_{\mu}\ell}{\rho_{m,n}}\,,
\label{umudef}
\end{equation}
where $u_{\mu}$ corresponds to a normalized four velocity, and
$\rho_{m,n} \equiv \partial\rho_m/\partial n$.
In this review, we focus on the theories given by the sum
of (\ref{action}) and (\ref{Spf}), i.e.,
\begin{equation}
{\cal S}={\cal S}_{\rm H}+{\cal S}_m\,.
\label{actionfull}
\end{equation}
After deriving the background and perturbation equations of motion,
they can be applied to subclasses of Horndeski theories listed above.
\section{FLRW background and tensor perturbations}
\label{backsec}
\subsection{Background equations of motion}
\label{backsec2}
Let us consider the flat FLRW spacetime given by the line element
\begin{equation}
ds^2=-N^2(t)dt^2+a^2(t) \delta_{ij}dx^i dx^j\,,
\label{FLRW}
\end{equation}
where $N(t)$ is a lapse and $a(t)$ is a scale factor.
The lapse is introduced here for deriving the Friedmann equation,
but we finally set $N=1$ after the variation of ${\cal S}$.
The scalar field $\phi$ depends on the cosmic time $t$ alone on
the background (\ref{FLRW}).
For the matter sector, the temporal component of $J^{\mu}$
in Eq.~(\ref{ndef}) is given by
\begin{equation}
J^{0}=n_0a^3\,,
\label{Jback}
\end{equation}
where $n_0$ is the background value of $n$.
Then, the Schutz-Sorkin action (\ref{Spf}) reduces to
\begin{equation}
{\cal S}_m= -\int d^4 x\,a^3
\left( N \rho_m+n_0 \dot{\ell}\right)\,.
\end{equation}
Varying this matter action with respect to $\ell$,
it follows that
\begin{equation}
{\cal N}_0 \equiv n_0 a^3={\rm constant}\,,
\label{N0con}
\end{equation}
where ${\cal N}_0$ corresponds to the total
conserved fluid number.
Now, we compute the action (\ref{actionfull}) on the spacetime
metric (\ref{FLRW}) and vary it with respect to
$N$, $a$ and $\phi$. Setting $N=1$ in the end, we obtain the
background equations of motion:
\begin{eqnarray}
& &
6 G_4H^2 +G_2 -\dot{\phi}^2 G_{2,X}
+\dot{\phi}^2 \left( 3H \dot{\phi}G_{3,X} -G_{3,\phi} \right)
+6 H \dot{\phi} \left( G_{4,\phi}+\dot{\phi}^2 G_{4,X \phi}
-2H \dot{\phi}G_{4,X}-H \dot{\phi}^3 G_{4,X X} \right)
\nonumber \\
& &
+H^2 \dot{\phi}^2 \left( 9 G_{5,\phi}+3\dot{\phi}^2 G_{5,X \phi}
-5H \dot{\phi}G_{5,X}-H \dot{\phi}^3 G_{5,X X} \right)=\rho_m\,,
\label{back1}\\
& &
2q_t \dot{H}-D_6 \ddot{\phi}+D_7 \dot{\phi}
=-\rho_m-P_m\,,
\label{back2}\\
& &
3D_6 \dot{H}+2D_1 \ddot{\phi}+3D_7 H-D_5=0\,,
\label{back3}
\end{eqnarray}
where $H \equiv \dot{a}/a$ is the Hubble expansion rate,
and $P_m$ is the matter pressure defined by
$P_m \equiv -n_0 \dot{\ell}-\rho_m$.
Substituting $J_0=-n_0a^3$ into Eq.~(\ref{umudef}), the temporal
component $u_{0}$ of four velocity is equivalent to $-1$
and hence
$\dot{\ell}=-\rho_{m,n}$. Thus, the matter pressure can be
written as
\begin{equation}
P_m=n_0 \rho_{m,n}-\rho_m\,.
\label{Pm}
\end{equation}
{}From Eq.~(\ref{N0con}), the fluid number density
obeys the differential equation $\dot{n}_0+3Hn_0=0$.
On using the property $\dot{\rho}_{m}=\rho_{m,n}
\dot{n}_0$ and the relation (\ref{Pm}), the conservation
of total fluid number translates to
\begin{equation}
\dot{\rho}_m+3H \left( \rho_m+P_m \right)=0\,.
\label{coneq}
\end{equation}
We note that this continuity equation also follows from
Eqs.~(\ref{back1})-(\ref{back3}).
The quantity $q_t$ in Eq.~(\ref{back2}) is defined by
\begin{equation}
q_t=2G_4-2\dot{\phi}^2 G_{4,X}
+\dot{\phi}^2 G_{5,\phi}-H \dot{\phi}^3 G_{5,X}\,,
\label{qt}
\end{equation}
which is associated with the no-ghost condition of
tensor perturbations discussed later in Sec.~\ref{tensec}.
The definitions of coefficients $D_{1,5,6,7}$ appearing in
Eqs.~(\ref{back2}) and (\ref{back3}) are presented in Appendix~\ref{coeff}.
As we will show in Sec.~\ref{scalarsec}, they also appear as coefficients
in the second-order action of scalar perturbations.
We note that Eq.~(\ref{back2}) has been derived by varying
the action with respect to $a$ and then eliminating the term
$G_2$ by using Eq.~(\ref{back1}).
Solving Eqs.~(\ref{back2})-(\ref{back3}) for $\dot{H}$
and $\ddot{\phi}$, it follows that
\begin{eqnarray}
\dot{H}
&=& -\frac{1}{q_s}
\left[ 2D_1 D_7 \dot{\phi}-D_6 (D_5-3D_7 H)
+2D_1 (\rho_m+P_m) \right]\,,\label{dH} \\
\ddot{\phi}
&=& \frac{1}{q_s} \left[
3D_6 D_7 \dot{\phi}+2 (D_5-3D_7H)q_t
+3D_6 (\rho_m+P_m) \right]\,,
\label{ddphi}
\end{eqnarray}
where
\begin{equation}
q_s \equiv 4D_1 q_t+3D_6^2\,.
\label{calD}
\end{equation}
To avoid the divergences on the right hand sides of
Eqs.~(\ref{dH}) and (\ref{ddphi}), we require that
\begin{equation}
q_s \neq 0\,.
\end{equation}
As we will show later, the determinant $q_s$ is related to the no-ghost condition of scalar perturbations.
The no-ghost condition corresponds to $q_s>0$,
in which case $\dot{H}$ and $\ddot{\phi}$ can
remain finite.
In the presence of matter, the condition for the cosmic
acceleration is given by
\begin{equation}
w_{\rm eff} \equiv -1-\frac{2\dot{H}}{3H^2}<-\frac{1}{3}\,.
\label{weff}
\end{equation}
We can express Eqs.~(\ref{back1}) and (\ref{back2}) in the forms
\begin{eqnarray}
& &
3M_{\rm pl}^2 H^2=\rho_{\rm DE}+\rho_m\,,
\label{Frieq}\\
& &
2M_{\rm pl}^2 \dot{H}=-\rho_{\rm DE}-P_{\rm DE}
-\rho_m-P_m\,,
\label{Frieq2}
\end{eqnarray}
where the density $\rho_{\rm DE}$ and pressure $P_{\rm DE}$
of the ``dark'' component are given, respectively, by
\begin{eqnarray}
\rho_{\rm DE}
&=& 3H^2 \left( M_{\rm pl}^2 -2G_4 \right)
-G_2+\dot{\phi}^2 G_{2,X}
-\dot{\phi}^2 \left( 3H \dot{\phi}G_{3,X} -G_{3,\phi} \right)
-6 H \dot{\phi} \left( G_{4,\phi}+\dot{\phi}^2 G_{4,X \phi}
\right. \nonumber \\
& &
\left.
-2H \dot{\phi}G_{4,X}-H \dot{\phi}^3 G_{4,X X} \right)
-H^2 \dot{\phi}^2 \left( 9 G_{5,\phi}+3\dot{\phi}^2 G_{5,X \phi}
-5H \dot{\phi}G_{5,X}-H \dot{\phi}^3 G_{5,X X} \right)\,,\\
P_{\rm DE}
&=& 2 \left( q_t-M_{\rm pl}^2 \right) \dot{H}-D_6 \ddot{\phi}
+D_7 \dot{\phi}-\rho_{\rm DE}\,.
\end{eqnarray}
We define the dark energy EOS, as
\begin{equation}
w_{\rm DE} \equiv \frac{P_{\rm DE}}{\rho_{\rm DE}}
=-1+\frac{2 ( q_t-M_{\rm pl}^2 ) \dot{H}-D_6 \ddot{\phi}
+D_7 \dot{\phi}}{\rho_{\rm DE}}\,,
\label{wdedef}
\end{equation}
which is different from the effective EOS (\ref{weff})
due to the presence of additional matter (dark matter, baryons, radiation).
The necessary condition for the late-time cosmic acceleration is
given by $w_{\rm DE}<-1/3$, but we caution that
this is not a sufficient condition.
In quintessence and k-essence we have $q_t=M_{\rm pl}^2$,
so the time variation of $\phi$ leads to the deviation
of $w_{\rm DE}$ from $-1$.
In other theories listed in Sec.~\ref{Hornsec}, the quantity $q_t$
generally differs from $M_{\rm pl}^2$, so the term
$2 (q_t-M_{\rm pl}^2 ) \dot{H}$ in Eq.~(\ref{wdedef})
also contributes to the additional deviation
of $w_{\rm DE}$ from $-1$.
Since the evolution of $w_{\rm DE}$ is different depending on
dark energy models, it is possible to distinguish between
them from the observations of SN Ia, CMB, and BAO.
\subsection{Tensor perturbations}
\label{tensec}
The speed of gravitational waves on the flat FLRW background
can be derived by expanding
the action (\ref{actionfull}) up to second order in tensor
perturbations $h_{ij}$. As a by-product, we can identify
conditions for the absence of ghost and Laplacian instabilities
in the tensor sector.
The perturbed line element, which contains
traceless and divergence-free tensor perturbations obeying
the conditions $h _{ii}=0$ and $\partial _{i}h_{ij}=0$,
is given by
\begin{equation}
ds^2=-dt^2+a^2(t)(\delta_{ij}+h_{ij})dx^i dx^j\,.
\end{equation}
Without loss of generality, we can choose nonvanishing components
of $h_{ij}$ in the forms
\begin{equation}
h_{11}=h_1(t,z)\,,\qquad
h_{22}=-h_1(t,z)\,,\qquad
h_{12}=h_{21}=h_2(t,z)\,,
\end{equation}
where $h_1$ and $h_2$ characterize the
two polarization states.
Expanding the Horndeski action (\ref{action}) up to second order
in perturbations and integrating it by parts, the quadratic action
$({\cal S}_{\rm H}^{(2)})_t$
contains the time derivative $\dot{h}_i^2$, the spatial derivative
$(\partial h_i)^2$, and the mass term $h_i^2$ (where $i=1,2$).
The second-order action of tensor perturbations arising from the
matter action (\ref{Spf}) can be written in the form
\begin{equation}
({\cal S}_m^{(2)})_t=-\int dtd^3x \left[
(\sqrt{-g})^{(2)} \rho_m
+\sqrt{-\bar{g}} \rho_{m,n}\delta n \right]\,,
\label{Sm2}
\end{equation}
where $(\sqrt{-g})^{(2)}=-a^3(h_1^2+h_2^2)/2$
and $\delta n=n_0 (h_1^2+h_2^2)/2$
with $\sqrt{-\bar{g}}=a^3$.
Then, Eq.~(\ref{Sm2}) reduces to
\begin{equation}
({\cal S}_m^{(2)})_t=-\int dt d^3 x \sum_{i=1}^{2}
\frac{1}{2}a^3 P_m h_i^2\,,
\end{equation}
where $P_m$ is the matter pressure given by Eq.~(\ref{Pm}).
Using the background Eqs.~(\ref{back1}) and (\ref{back2}),
the terms proportional to $h_i^2$ identically vanish from
the total second-order action
${\cal S}_t^{(2)}=({\cal S}_{\rm H}^{(2)})_t+({\cal S}_{m}^{(2)})_t$.
Finally, we can express ${\cal S}_t^{(2)}$ in a compact form:
\begin{equation}
{\cal S}_t^{(2)}=\int dt d^3x \sum_{i=1}^{2}
\frac{a^3}{4}q_t \left[ \dot{h}_i^2-\frac{c_t^2}{a^2}
(\partial h_i)^2 \right]\,,
\label{actionSt}
\end{equation}
where $q_t$ was already introduced in Eq.~(\ref{qt}),
and $c_t^2$ is the tensor propagation speed squared given by
\begin{equation}
c_t^2=\frac{1}{q_t}
\left( 2G_4-\dot{\phi}^2G_{5,\phi} -\dot{\phi}^2 \ddot{\phi} G_{5,X} \right)\,.
\label{ct}
\end{equation}
To avoid the ghost and Laplacian instabilities,
we require that
\begin{eqnarray}
& &
q_t>0\,,\label{Qtcon}\\
& &
c_t^2>0\,.
\label{ctcon}
\end{eqnarray}
Varying the action (\ref{actionSt}) with respect to $h_i$,
we obtain the tensor perturbation equation of motion
in Fourier space, as
\begin{equation}
\ddot{h}_i+\left(3H +\frac{\dot{q}_t}{q_t} \right) \dot{h}_i
+c_t^2 \frac{k^2}{a^2} h_i=0\,,
\label{heq}
\end{equation}
where $k$ is a coming wavenumber.
The time variation of $q_t$ and the deviation of $c_t^2$
from 1
lead to the modified evolution of $h_i$ compared to that in GR.
The observational bound (\ref{ctbound}) of $c_t$ places tight
constraints on surviving dark energy models.
To realize the exact value $c_t^2=1$, we require the following condition
\begin{equation}
2G_{4,X}-2G_{5,\phi}+\left( H \dot{\phi} -\ddot{\phi}
\right) G_{5,X}=0\,.
\end{equation}
If we do not allow any tuning among functions,
the dependence of $G_4$ on $X$ and $G_5$ on $\phi, X$
is forbidden. Then, the Horndeski Lagrangian is restricted
to be of the
form \cite{Lon15,GWcon1,GWcon2,GWcon3,GWcon4,GWcon5,GWcon6}
\begin{equation}
L=G_2(\phi,X)+G_3(\phi,X) \square \phi
+G_4(\phi)R\,.
\label{lagcon}
\end{equation}
Among the theories listed in Sec.~\ref{Hornsec}, the theories
(5), (6), (7) lead to $c_t^2$ different from 1.
In particular, as long as the derivative and Gauss-Bonnet couplings
contribute to the late-time cosmological dynamics, the deviation of
$c_t^2$ from 1 is large and hence such couplings are excluded from
the bound (\ref{ctbound}) as a source for dark
energy \cite{Gong}\footnote{Besides this problem,
the matter perturbation
in $f({\cal G})$ gravity is subject to strong instabilities
attributed to the negative sound speed squared \cite{FMT10},
so $f({\cal G})$ cosmological models are not
cosmologically viable unless the deviation from the
$\Lambda$CDM model is very tiny.}.
Quintessence and k-essence, BD theory, and $f(R)$ gravity give rise to
the exact value $c_t^2=1$, so they automatically satisfy the bound
(\ref{ctbound}). The quartic and quintic Galileon couplings
$G_4=\beta_4 X^2$ and $G_5=\beta_5 X^2$ lead to the deviation of
$c_t^2$ from 1, but the Galileon Lagrangian up to
the cubic interaction $G_3=\beta_3 X$ is allowed
(which includes the model with the additional term
$G_2=\beta_2 X^2$ to the cubic Galileon Lagrangian \cite{Kase18}).
The kinetic braidings \cite{braiding1,braiding2} and its extensions \cite{Silva,Koba1,Koba2,DTge,Babi11,KTD15} are also consistent
with the bound (\ref{ctbound}).
Even if $c_t^2$ is constrained to be close to 1, there is another modification
to Eq.~(\ref{heq}) arising from the time variation of $q_t$.
This leads to the modified luminosity distance of GWs
relative to that of electromagnetic
signals \cite{Maggiore1,Amen17,Maggiore2}.
Since $q_t=2G_4(\phi)$ for the Lagrangian (\ref{lagcon}),
it is possible to distinguish between nonminimally and
minimally coupled theories from the luminosity
distance measurements of GWs
together with the electromagnetic counterpart.
After the accumulation of GW events in future, the
luminosity distance can be a key observable to probe
the existence of nonminimal couplings.
For the Lagrangian (\ref{lagcon}), we have
$D_6=-\dot{\phi}^2 G_{3,X}-2G_{4,\phi}$ and
$D_7=\dot{\phi}(G_{2,X}+2G_{3,\phi}+2G_{4,\phi \phi})
-H(2G_{4,\phi}+3\dot{\phi}^2 G_{3,X})$, so
the dark energy EOS (\ref{wdedef}) reduces to
\begin{equation}
w_{\rm DE}
=-1+\frac{2(2G_4-M_{\rm pl}^2) \dot{H}
+(\dot{\phi}^2 G_{3,X}+2G_{4,\phi}) \ddot{\phi}
+[\dot{\phi}(G_{2,X}+2G_{3,\phi}+2G_{4,\phi \phi})-H(2G_{4,\phi}+3\dot{\phi}^2 G_{3,X})] \dot{\phi}}
{3H^2 ( M_{\rm pl}^2 -2G_4 )
-G_2+\dot{\phi}^2 G_{2,X}
-\dot{\phi}^2 ( 3H \dot{\phi}G_{3,X} -G_{3,\phi})
-6 H \dot{\phi}G_{4,\phi}}\,.
\label{wde}
\end{equation}
In Secs.~\ref{quinsec}-\ref{mostsec}, we will study
the evolution of
$w_{\rm DE}$ for concrete models
of the late-time cosmic acceleration in the framework of Horndeski theories.
\section{Gauge-ready formulation of scalar perturbations}
\label{scalarsec}
In this section, we derive the second-order action of scalar
perturbations in full Horndeski theories without fixing
gauge conditions. The resulting linear perturbation equations
of motion are written in a gauge-ready form \cite{Hwang,HKT18},
so that one can choose convenient gauges depending on the problems at hand.
For generality, we do not restrict the Horndeski Lagrangian to
the form (\ref{lagcon}).
We begin with the linearly perturbed line-element
given by \cite{Bardeen,Kodama,MFB92,BTW05,Malik}
\begin{equation}
ds^2=-(1+2\alpha)dt^2+2 \partial_i \chi dt dx^i
+a^2(t) \left[ (1+2\zeta) \delta_{ij}
+2\partial_i \partial_j E \right]dx^i dx^j\,,
\label{permet}
\end{equation}
where $\alpha, \chi, \zeta, E$ are scalar metric perturbations, and
the symbol $\partial_i$ stands for the partial derivative
$\partial/\partial x^i$.
We do not take into account vector perturbations,
as they are nondynamical in scalar-tensor theories.
The scalar field is decomposed into the form
\begin{equation}
\phi=\bar{\phi}(t)+\delta \phi (t, x^i)\,,
\end{equation}
where $\bar{\phi}(t)$ and $\delta \phi(t, x^i)$ are
the background and perturbed values, respectively.
In the following, we omit the bar from the background quantities.
\subsection{Second-order matter action}
We first expand the Schutz-Sorkin action (\ref{Spf}) up to
second order in perturbations.
We decompose the temporal and spatial components of $J^{\mu}$
into the background and perturbed parts, as \cite{GPcosmo,GPGeff}
\begin{equation}
J^{0} = \mathcal{N}_{0}+\delta J\,,\qquad
J^{i} =\frac{1}{a^2(t)}\,\delta^{ik}\partial_{k}\delta j\,,
\label{elldef}
\end{equation}
where $\mathcal{N}_{0}=n_0a^3$ is the conserved background
fluid number defined by Eq.~(\ref{N0con}),
and $\delta J, \delta j$ are scalar perturbations.
The spatial component of four velocity can be expressed as
$u_i=-\partial_i v$, where $v$ is the velocity potential.
On using Eq.~(\ref{umudef}), the scalar quantity $\ell$
contains the perturbation $-\rho_{m,n}v$. Then,
$\ell$ can be decomposed as
\begin{equation}
\ell=-\int^{t} \rho_{m,n}
(\tilde{t})d\tilde{t}
-\rho_{m,n}v\,,
\label{ells}
\end{equation}
where the first contribution to the right hand side corresponds
the background quantity.
Defining the matter density perturbation
\begin{equation}
\delta \rho_m \equiv \frac{\rho_{m,n}}{a^3} \left[\delta J
-\mathcal{N}_{0}(3\zeta+\partial^2E)\right]\,,
\label{drhom}
\end{equation}
where $\partial^2E \equiv \partial_i \partial_i E$
(the same latin subscripts are summed over),
the perturbation of fluid number density $n$, up to
second order, is expressed as
\begin{equation}
\delta n= \frac{\delta \rho_m}{\rho_{m,n}}
-\frac{({\cal N}_0 \partial \chi+\partial \delta j)^2}{2{\cal N}_0a^5}
-\frac{(3\zeta+\partial^2E)\delta \rho_m}{\rho_{m,n}}
-\frac{{\cal N}_0(\zeta+\partial^2E)(3\zeta-\partial^2E)}{2a^3}\,.
\end{equation}
Since $\delta n$ reduces to $\delta \rho_m/\rho_{m,n}$ at linear order,
this confirms the consistency of the definition (\ref{drhom}).
Now, we are ready for expanding the Schutz-Sorkin action (\ref{Spf})
up to quadratic order in scalar perturbations.
This manipulation gives the second-order matter
action \cite{HKT18}:
\begin{eqnarray}
({\cal S}_m^{(2)})_s
&=&\int dt d^3 x\, a^3
\Bigg[
\frac{\rhon}{2a^8n_0}(\pa \delj)^2+\frac{\rhon}{a^5}(\pa\chi+\pa v) (\pa\delj)
+\left(\dot{v}-3Hc_m^2v-\alpha\right)\delrho
-\frac{c_m^2}{2n_0\rhon}\delrho^2+\frac{\rho_m}{2}\alpha^2
\notag\\
&&
+\frac{n_0 \rhon-\rho_m}{2}\left\{\frac{(\pa\chi)^2}{a^2}
+(\zeta+\pa^2E)(3\zeta-\pa^2E)\right\}
+(3\zeta+\pa^2E)\left\{ n_0\rhon(\dot{v}-3Hc_m^2 v)-\rho_m\alpha\right\}
\Bigg]\,,
\label{SMS}
\end{eqnarray}
where $c_m^2$ is the matter sound speed squared given by
\begin{equation}
c_m^2=\frac{P_{m,n}}{\rho_{m,n}}
=\frac{n_0 \rho_{m,nn}}{\rho_{m,n}}\,.
\end{equation}
Variation of the action (\ref{SMS}) with respect to $\delj$
leads to
\begin{equation}
\partial \delta j=-a^3 n_0
\left( \partial v+\partial \chi \right)\,.
\label{deljre}
\end{equation}
Substituting the relation (\ref{deljre}) into
Eq.~(\ref{SMS}), we obtain
\begin{eqnarray}
\hspace{-0.3cm}
({\cal S}_m^{(2)})_s
&=&\int dt d^3x\,a^3
\Bigg[ \left( \dot{v}-3H c_m^2v-\alpha
\right) \delta \rho_m-\frac{c_m^2}{2(\rho_m+P_m)}\delrho^2
-\frac{\rho_m+P_m}{2a^2}
\left\{ (\partial v)^2+2\partial v \partial \chi \right\}
-\frac{\rho_m}{2a^2} (\partial \chi)^2
+\frac{\rho_m}{2}\alpha^2
\notag\\
\hspace{-0.3cm}
&&
+\frac{P_m}{2}(\zeta+\pa^2E)(3\zeta-\pa^2E)
+(3\zeta+\pa^2E)\left\{(\rho_m+P_m)(\dot{v}-3Hc_m^2 v)
-\rho_m\alpha\right\} \Bigg]\,,
\label{SMS2}
\end{eqnarray}
where we used Eq.~(\ref{Pm}).
\subsection{Full second-order scalar
action and linear perturbation equations of motion}
\label{actsec}
We also expand the Horndeski action (\ref{action}) up to second order
in scalar perturbations and take the sum with Eq.~(\ref{SMS2}).
Using the background Eq.~(\ref{back1}), the term
$\rho_m\alpha^2/2$ in Eq.~(\ref{SMS2}) is cancelled
by a part of contributions proportional to $\alpha^2$ arising from
the Horndeski action.
After the integration by parts, we can write the full second-order action
of ${\cal S}$ in the gauge-ready form
\begin{equation}
{\cal S}_s^{(2)}=\int dt d^3x
\left( L_{\rm flat}+L_{\zeta}+L_{E}\right)\,,
\label{Ss}
\end{equation}
where
\begin{eqnarray}
\hspace{-1.0cm}
L_{\rm flat}
&=& a^3\left[
D_1\dot{\dphi}^2+D_2\frac{(\partial\dphi)^2}{a^2}+D_3\dphi^2
+\left(D_4\dot{\dphi}+D_5\dphi+D_6\frac{\partial^2\dphi}{a^2}\right) \alpha
-\left(D_6\dot{\dphi}-D_7\dphi\right)\frac{\partial^2\chi}{a^2}
\right.\notag\\
\hspace{-1.0cm}
&&\left.~~~
+\left(\dot{\phi} D_6-2Hq_t\right)\alpha\frac{\partial^2\chi}{a^2}
+\left(\dot{\phi}^2 D_1+3H \dot{\phi} D_6-3H^2 q_t\right)\alpha^2
\right. \nonumber \\
\hspace{-1.0cm}
& &
\left.~~~+\left( \rho_m+P_m \right)v\frac{\partial^2 \chi}
{a^2}-v\dot{\delta \rho}_m-3H (1+c_m^2) v\delta \rho_m
-\frac{1}{2} (\rho_m+P_m) \frac{(\partial v)^2}{a^2}
-\frac{c_m^2}{2 (\rho_m+P_m)} \delta \rho_m^2
-\alpha \delta \rho_m \right]\,,
\label{LF}\\
\hspace{-1.0cm}
L_{\zeta}
&=& a^3\left[
\left\{
3D_6\dot{\dphi}-3D_7\dphi-3\left(\dot{\phi} D_6-2Hq_t\right)\alpha-3(\rho_m+P_m)v
+2q_t\frac{\pa^2\chi}{a^2}\right\}\dot{\zeta}-3q_t\dot{\zeta}^2
\right.\notag\\
\hspace{-1.0cm}
&&\left.~~~
-\left(B_1\dphi+2q_t\alpha\right)\frac{\pa^2\zeta}{a^2}
+q_t c_t^2\frac{(\pa\zeta)^2}{a^2}
\right]\,,
\label{Lze}\\
\hspace{-1.0cm}
L_{E}
&=& a^3\Bigg[
2q_t\ddot{\zeta}+2B_2\dot{\zeta}
-D_6\ddot{\dphi}-B_3\dot{\dphi}+B_4\dphi
+\frac{1}{a^3}\frac{d}{dt}\left\{a^3 \left(\dot{\phi} D_6-2Hq_t
\right) \alpha\right\}
+(\rho_m+P_m)(\dot{v}-3Hc_m^2v)
\Bigg]\,\pa^2 E,
\label{LE}
\end{eqnarray}
where the explicit forms of coefficients $D_{1,...,7}$ are
presented in Appendix~\ref{coeff}.
The quantities $q_t$ and $c_t^2$ are given by
Eqs.~(\ref{qt}) and (\ref{ct}), respectively.
The coefficients $B_{1,2,3,4}$ in Eqs.~(\ref{Lze}) and (\ref{LE})
can be expressed by using other quantities, as
\begin{equation}
B_1=\frac{2}{\tp}\left[\dot{q_t}+(1-c_t^2)Hq_t\right]\,,
\qquad
B_2=\dot{q_t}+3Hq_t\,,
\qquad
B_3=\dot{D_6}+3HD_6-D_7\,,
\qquad
B_4=\dot{D_7}+3HD_7\,.
\label{Bi}
\end{equation}
The Lagrangian $L_{\rm flat}$ is present for the flat
gauge $\zeta=0=E$ \cite{KTdark18},
while the other two Lagrangians
$L_{\zeta}$ and $L_{E}$ arise in the presence of metric perturbations
$\zeta$ and $E$.
We note that, in scalar-vector-tensor theories, the generalized version
of the action (\ref{Ss}) was derived in Ref.~\cite{HKT18}.
Since there are no kinetic terms for the variables $\alpha,\chi,v,E$
in Eqs.~(\ref{LF})-(\ref{LE}), they correspond to nondynamical perturbations.
Varying the second-order action (\ref{Ss}) with respect to $\alpha,\chi,v,E$,
their constraint equations in Fourier space are given, respectively, by
\begin{eqnarray}
\E_{\alpha} &\equiv&
D_4\dot{\dphi}-3\left(\dot{\phi} D_6-2Hq_t\right)\dot{\zeta}
+D_5\dphi+2\left(\dot{\phi}^2 D_1+3H \dot{\phi} D_6-3H^2 q_t\right)\alpha
\notag\\
&&
+\frac{k^2}{a^2}\left[ 2q_t\zeta-\left(\dot{\phi} D_6-2Hq_t\right)\left(\chi-a^2\dot{E}\right)
-D_6\dphi\right]-\delta \rho_m=0\,,
\label{eqalpha}\\
\E_{\chi} &\equiv&
D_6\dot{\dphi}-2q_t\dot{\zeta}-D_7\dphi-\left(\dot{\phi} D_6-2Hq_t\right)\alpha
-\left( \rho_m+P_m \right)v=0\,,
\label{eqchi}\\
\E_{v} &\equiv&
\dot{\delta \rho}_m+3H \left( 1+c_m^2 \right)
\delta \rho_m
+3(\rho_m+P_m)\dot{\zeta}
+\frac{k^2}{a^2} \left( \rho_m+P_m \right)
\left( v+\chi-a^2\dot{E} \right)=0\,,
\label{eqdrho} \\
\E_{E} &\equiv&
2q_t\ddot{\zeta}+2B_2\dot{\zeta}
-D_6\ddot{\dphi}-B_3\dot{\dphi}+B_4\dphi
+\frac{1}{a^3}\frac{d}{dt}\left\{a^3(\dot{\phi} D_6-2Hq_t)\alpha\right\}
+(\rho_m+P_m)(\dot{v}-3Hc_m^2v)=0\,,
\label{eqE}
\end{eqnarray}
where $k$ is a comoving wavenumber.
Variations of the action (\ref{Ss}) with respect to the remaining
variables $\dphi,\delrho,\zeta$ lead to
\begin{eqnarray}
\E_{\delta \phi} &\equiv&
\dot{\cal Z}+3H {\cal Z}+3D_7\dot{\zeta}-2D_3 \delta \phi-D_5 \alpha
-\frac{k^2}{a^2} \left( 2D_2 \delta \phi -D_6 \alpha
-D_7 \chi+B_1\zeta-a^2B_4E
\right)=0\,,\label{calZeq}\\
\E_{\delta \rho_m} &\equiv&
\dot{v}-3Hc_m^2 v-\frac{c_m^2}{\rho_m+P_m} \delta \rho_m
-\alpha=0\,,
\label{veq}\\
\E_{\zeta} &\equiv&
\dot{\cal W} +3H{\cal W}+(\rho_m+P_m)(\dot{v}-3Hc_m^2v)
+\frac{k^2}{3a^2}\left(2q_t\alpha+2q_tc_t^2\zeta
+B_1\dphi\right)=0\,,
\label{calWeq}
\end{eqnarray}
where
\begin{eqnarray}
{\cal Z} &\equiv& 2D_1 \dot{\delta \phi}+3D_6\dot{\zeta}
+D_4 \alpha+\frac{k^2}{a^2}\left[D_6\chi
-a^2(D_6\dot{E}+D_7E)
\right]\,, \label{Zdef}\\
{\cal W} &\equiv& 2q_t\dot{\zeta}-D_6\dot{\dphi}+D_7\dphi
+\left(\dot{\phi} D_6-2Hq_t\right)\alpha
+\frac{2k^2}{3a^2} q_t (\chi-a^2 \dot{E})\,.
\label{Wdef}
\end{eqnarray}
By combining Eq.~(\ref{eqE}) with Eq.~(\ref{calWeq}), we can eliminate
the time derivatives $\ddot{\zeta}$ and $\ddot{\dphi}$.
This manipulation leads to
\begin{equation}
q_t \left( \alpha+\dot{\chi}
+c_t^2 \zeta+H\chi-a^2 \ddot{E}-3a^2 H \dot{E}
\right)+ \dot{q}_t \left( \chi-a^2 \dot{E} \right)
+\frac{B_1}{2} \dphi=0\,.
\label{eqE2}
\end{equation}
The perturbation equations of motion (\ref{eqalpha})-(\ref{calWeq})
and (\ref{eqE2}) are written in the gauge-ready form, such that they can
be used for arbitrary gauge choices.
The consistency of the above perturbation equations can be
confirmed in terms of Bianchi identities.
For this purpose, we resort to the following properties:
\begin{eqnarray}
&&
2\tp^2 D_{2}=-2H\left[\dot{q_t}+H\left(1-c_t^2\right)q_t\right]
-\tp\left(\dot{D_6}+HD_6+D_7\right)\,,\label{conD2}\\
&&
2\tp D_{3}=\frac{1}{a^3}\frac{d}{dt}(a^3D_5)-\frac{3H}{a^3}\frac{d}{dt}(a^3D_7)\,,\\
&&
D_4=-2\tp D_1-3HD_6\,.\label{conD4}
\end{eqnarray}
Using also the background Eqs.~(\ref{back2}), (\ref{back3}) and (\ref{coneq}),
we find that the perturbation equations (\ref{eqalpha})-(\ref{calWeq})
satisfy the two particular relations:
\begin{eqnarray}
&&
\frac{1}{a^3}\frac{d}{dt}\left(a^3\E_{\alpha}\right)
-3H\E_{\zeta}-\tp\,\E_{\dphi}
-\frac{k^2}{a^2}\E_{\chi}
+3H(\rho_m+P_m) \E_{\delta \rho_m}+\E_v=0\,,
\label{Bianchi1}\\
&&
\E_{E}-\frac{1}{a^3}\frac{d}{dt}\left(a^3\E_{\chi}\right)
=0\,,
\label{Bianchi2}
\end{eqnarray}
which correspond to the temporal and spatial components
of Bianchi identities, respectively.
Thus, we have confirmed the consistency of
Eqs.~(\ref{eqalpha})-(\ref{calWeq}) with Bianchi identities.
\subsection{Gauge issues}
\label{gaugesec}
Since there are residual gauge degrees of freedom, we can
fix the gauges in the perturbation equations derived
in Sec.~\ref{actsec}.
We discuss the issues of gauge transformations,
gauge-invariant
variables, and gauge choices.
Let us consider a scalar infinitesimal gauge transformation from one coordinate
$x^{\mu}=(t, x^i)$ to another coordinate $\tilde{x}^{\mu}
=(\tilde{t}, \tilde{x}^i)$, as
\begin{equation}
\tilde{t}=t+\xi^{0}\,,\qquad
\tilde{x}^{i}=x^{i}+\delta^{ij} \partial_{j} \xi\,,
\label{gaugetra}
\end{equation}
where $\xi^{0}$ and $\xi$ are scalar quantities.
Then, the metric perturbations are subject to
the transformations:
\begin{equation}
\tilde{\alpha}=\alpha-\dot{\xi}^{0}\,,\qquad
\tilde{\chi}=\chi+\xi^{0}-a^2 \dot{\xi}\,,\qquad
\tilde{\zeta}=\zeta-H \xi^{0}\,,\qquad
\tilde{E}=E-\xi\,,
\label{gaugetra1}
\end{equation}
while the perturbations associated with the scalar field and matter
transform as
\begin{equation}
\widetilde{\delta \phi}=\delta \phi-\dot{\phi}\,\xi^{0}\,,
\qquad
\widetilde{\delta \rho_m}=\delta \rho_m-\dot{\rho}_m\,\xi^{0}\,,
\qquad
\tilde{v}=v-\xi^{0}\,.
\label{gaugetra2}
\end{equation}
One can construct a family of gauge-invariant variables whose
forms are unchanged under the transformation (\ref{gaugetra}).
{}From Eq.~(\ref{gaugetra1}), the Bardeen gravitational
potentials defined by \cite{Bardeen}
\begin{equation}
\Psi=\alpha+\frac{d}{dt}
\left( \chi - a^2 \dot{E} \right)\,,\qquad
\Phi=\zeta+H \left( \chi - a^2 \dot{E} \right)
\label{PsiPhi}
\end{equation}
are gauge-invariant.
On using the properties (\ref{gaugetra2}) as well,
there are the following gauge-invariant combinations:
\begin{eqnarray}
& &
\delta \phi_{\rm f}=\delta \phi
-\frac{\dot{\phi}}{H}\zeta\,,
\qquad
\delta \phi_{\rm N}=\delta \phi+\dot{\phi}\left(\chi-a^2 \dot{E}\right)\,,
\label{delphif}\nonumber \\
& &
\delta \rho_{\rm f}=\delta \rho_m -\frac{\dot{\rho}_m}{H}\zeta\,,\qquad
\delta \rho_{\rm u}=\delta \rho_m -\frac{\dot{\rho}_m}
{\dot{\phi}}\delta \phi\,,\qquad
\delta \rho_{\rm N}=\delta \rho_m+\dot{\rho}_m \left(\chi-a^2 \dot{E}\right)\,,
\label{rhof}\nonumber \\
& &
\delta_m=\frac{\delta \rho_m}{\rho_m}
+3H \left( 1+\frac{P_m}{\rho_m} \right)v\,,
\qquad
{\cal R}=\zeta-\frac{H}{\dot{\phi}} \delta \phi\,,
\qquad
{\cal B}=Hv-\zeta\,.
\label{delmB}
\end{eqnarray}
The Mukhanov-Sasaki variable $\delta \phi_{\rm f}$ \cite{Mukhanov,Sasaki}
is related to the gauge-invariant curvature perturbation
${\cal R}$ \cite{Lukash,Lyth}, as
\begin{equation}
{\cal R}=-\frac{H}{\dot{\phi}}\delta \phi_{\rm f}\,.
\label{R}
\end{equation}
The residual gauge degrees of freedom can be removed by
fixing $\xi^0$ and $\xi$ in the transformation (\ref{gaugetra2}).
There are several gauge conditions commonly
used in the literature:
\begin{eqnarray}
& &
\delta \phi=0\,,\quad ~\,E=0\,,\qquad\, \mathrm{(Unitary~gauge)}\,,
\label{Ugauge}\\
& &
\zeta=0\,,\qquad E=0\,,\qquad\, \mathrm{(Flat~gauge)}\,,
\label{Fgauge}\\
& &
\chi=0\,,\qquad E=0\,,\qquad
\mathrm{(Newtonian~gauge)}\,.
\label{Ngauge}
\end{eqnarray}
In the so-called synchronous gauge characterized by $\alpha=0$ and $\chi=0$,
the transformation scalar $\xi^0$ is not unambiguously fixed and hence
this gauge is not chosen in the following.
In Sec.~\ref{stasec}, we derive no-ghost conditions and the speed of scalar
perturbations in full Horndeski theories
by taking the three gauge conditions
(\ref{Ugauge})-(\ref{Ngauge}).
Independent of the gauge choices, we are dealing
with the same physics.
\subsection{Perturbation equations expressed in terms of
gauge-invariant variables}
\label{gauginsec}
We rewrite the perturbation equations of motion by using the gauge-invariant variables
$\Psi, \Phi, \dphiN, {\cal B}$ without fixing gauge conditions.
The perturbations $\alpha, \zeta, \dphi$ can be
expressed in terms of $\Psi, \Phi, \dphiN$ as well
as $E$, $\chi$ and their time derivatives.
On using the background Eqs.~(\ref{back2}), (\ref{back3}), (\ref{coneq}) with Eqs.~(\ref{conD2})-(\ref{conD4}),
the terms containing $E$, $\chi$ and their time derivatives
identically vanish.
First of all, the perturbation Eqs.~(\ref{calZeq}) and (\ref{calWeq}),
which correspond to the equations of motion for $\dphi$ and $\Phi$
respectively, can be expressed as
\begin{eqnarray}
&&
2D_1\ddot{\dphiN}+2(\dot{D_1}+3HD_1)\dot{\dphiN}
+\left(M_{\phi}^2-2D_2\frac{k^2}{a^2}\right)\dphiN
+3D_6\ddot{\Phi}+3(\dot{D_6}+3HD_6+D_7)\dot{\Phi}
-(2\tp D_1+3HD_6)\dot{\Psi}\notag\\
&&
-B_1\frac{k^2}{a^2}\Phi
-\left[2\tp(\dot{D_1}+3HD_1)+2D_5
+3H(\dot{D_6}+3HD_6-D_7)
-D_6\frac{k^2}{a^2}\right]\Psi=0\,,
\label{dphiNeq}\\
&&
-2q_t\ddot{\Phi}-\left(2\dot{q_t}+6Hq_t+\frac{\rho_m+P_m}{H}\right)\dot{\Phi}
+(2Hq_t-\tp D_6)\dot{\Psi}+\left[(\rho_m+P_m)\left(\frac{\dot{H}}{H^2}
+3c_m^2\right)-\frac{2q_tc_t^2}{3}\frac{k^2}{a^2}\right]\Phi
\notag\\
&&
-\left[\tp(\dot{D_6}+3HD_6+D_7)-2H(\dot{q_t}+3Hq_t)
+\rho_m+P_m+\frac{2q_t}{3}\frac{k^2}{a^2}\right]\Psi
-(\rho_m+P_m)\left[\frac{\dot{\cal B}}{H}
-\left(\frac{\dot{H}}{H^2}+3c_m^2\right){\cal B}\right]
\nonumber \\
& &
+D_6\ddot{\dphiN}+(\dot{D_6}+3HD_6-D_7)\dot{\dphiN}
-\left( \dot{D_7}+3HD_7+\frac{B_1}{3}
\frac{k^2}{a^2}\right) \dphiN=0\,,
\label{ddPhi}
\end{eqnarray}
where
\begin{equation}
M_{\phi}^2 \equiv -2D_3\,.
\end{equation}
The quantity $M_{\phi}^2$ corresponds to the mass
squared of scalar-field perturbation $\delta \phi$.
In quintessence with the Lagrangian $G_2=X-V(\phi)$,
the term $G_{2,\phi \phi}/2$ in $D_3$ gives rise to
the contribution $V_{,\phi \phi}\dphi_{\rm N}$
to Eq.~(\ref{dphiNeq}).
For the models in which $M_{\phi}^2$ exceeds
the order of $H^2$ in the past (like $f(R)$ models of the late-time
cosmic acceleration \cite{fR1,fR2,fR3,fR4,fR5}),
the condition $M_{\phi}^2>0$ is required for
avoiding the tachyonic instability\footnote{The model
$f(R)=R-\beta/R^n$ ($n>0$) proposed in
Refs.~\cite{fRearly1,fRearly1b,fRearly2,fRearly3} leads to
$M_{\phi}^2<0$, so it is ruled out by the tachyonic instability
of scalar perturbations.} \cite{fR1,fR3,fRreview0,fRreview}.
Equations (\ref{dphiNeq}) and (\ref{ddPhi}) can be
closed by solving them for $\ddot{\dphiN}$ and $\ddot{\Phi}$.
The remaining perturbation equations of motion can be
compactly expressed by using the following
dimensionless variables:
\begin{equation}
\aM \equiv \frac{\dot{q}_t}{Hq_t}\,,
\qquad
\aB\equiv-\frac{\tp D_6}{2Hq_t}\,,
\qquad
\aK\equiv\frac{2\tp^2D_1}{H^2q_t}\,,
\label{nodim}
\end{equation}
and
\begin{equation}
\dphiv_{\rm N} \equiv\frac{H}{\tp}\dphi_{\rm N}\,,
\qquad
h\equiv\frac{\dot{H}}{H^2}\,,
\qquad
\tilde{\Omega}_m \equiv \frac{\rho_m}{3H^2q_t}\,,
\qquad
w_m\equiv\frac{P_m}{\rho_m}\,,
\qquad
\K\equiv\frac{k}{aH}\,,
\qquad
{\cal N} \equiv \ln a\,.
\label{nodim2}
\end{equation}
The quantity $\aB$ introduced by Bellini and
Sawicki \cite{Bellini}, which is denoted as
$\aB^{\rm (BS)}$, is related to our $\aB$ according to
\begin{equation}
\aB=-\frac{1}{2}\aB^{\rm (BS)}\,.
\label{abre}
\end{equation}
Our notations of $\alpha_{\rm M}$, $\alpha_{\rm B}$, and
$\alpha_{\rm K}$ match with those used in Ref.~\cite{Gleyzes14}
in the context of effective field theory of
dark energy \cite{EFT1,EFT2,EFT3,EFT4,Piazza,EFT5,Gergely,Gao14,Kase15}.
{}From Eq.~(\ref{conD4}), the coefficient $D_4$ is written as
$D_4=H^2 q_t (6 \alpha_{\rm B}-\alpha_{\rm K})/\dot{\phi}$.
We also solve Eqs.~(\ref{back2}) and (\ref{back3}) for $D_5$ and $D_7$
to express them in terms of $\alpha_{\rm B}$, $\alpha_{\rm K}$,
and other dimensionless variables given in Eq.~(\ref{nodim2}).
When we take the time derivative of
$\delta_m$ in Eq.~(\ref{delmB}), we use Eq.~(\ref{coneq}) and
the relation $c_m^2=\dot{P}_m/\dot{\rho}_m$.
Then, Eqs.~(\ref{eqalpha})-(\ref{eqdrho}),
(\ref{veq}) and (\ref{eqE2}) reduce, respectively, to
\begin{eqnarray}
&&
(6\aB-\aK)\vphiN'+6(1+\aB)\Phi'
+\left[h(\aK-12\aB-6)-9(1+w_m)\tilde{\Omega}_m+2\aB\K^2\right]\vphiN\notag\\
&&
+\left[9(1+w_m)\tilde{\Omega}_m+2\K^2\right]\Phi+(\aK-12\aB-6)\Psi
-3\tilde{\Omega}_m\delta_m+9(1+w_m)\tilde{\Omega}_m{\cal B}=0\,,
\label{eqalpha2}\\
&&
\aB\vphiN'+\Phi'
-\left[h(1+\aB)+\frac32(1+w_m)\tilde{\Omega}_m\right]\vphiN
-(1+\aB)\Psi+\frac32(1+w_m)\tilde{\Omega}_m(\Phi+{\cal B})=0\,,
\label{eqchi2}\\
&&
\delta_m'-3(1+w_m){\cal B}'
+(1+w_m)\K^2(\Phi+{\cal B})
+3(c_m^2-w_m)\delta_m=0\,,
\label{eqdrho2}\\
&&
\Phi'+{\cal B}'-h(\Phi+{\cal B})-\Psi
-\frac{c_m^2}{1+w_m}\delta_m=0\,,
\label{veq2}\\
&&
\Psi+c_t^2 \Phi+\left( 1-c_t^2+\alpha_{\rm M}
\right) \delta \varphi_{\rm N}=0\,,
\label{aniso}
\end{eqnarray}
where a prime represents a derivative
with respect to ${\cal N}$.
Since ${\cal R}=\Phi-\delta \varphi_{\rm N}$, Eq.~(\ref{aniso}) can be also expressed as
\begin{equation}
\Psi+\left( 1+\alpha_{\rm M} \right) \Phi
+\left( c_t^2-1-\alpha_{\rm M} \right){\cal R}
=0\,.
\label{anire}
\end{equation}
The relation (\ref{anire}) shows that the time-dependence of $q_t$
and the deviation of $c_t^2$ from 1 lead to the difference
between $-\Psi$ and $\Phi$.
Under the bound (\ref{ctbound}), the gravitational slip
($-\Psi \neq \Phi$) mostly arises from the time variation
of $q_t$.
We solve Eqs.~(\ref{eqalpha2})-(\ref{veq2}) for
$\Phi', \vphiN', \delta_m', {\cal B}'$
by using Eq.~(\ref{aniso}) to eliminate the perturbation $\Psi$.
Then, we obtain
\begin{eqnarray}
\hspace{-.5cm}&&
\Phi'=
-(c_t^2-b_2+4\aB b_1\K^2)\Phi-(1-c_t^2+\aM-h+b_2+4\aB^2b_1\K^2)\vphiN
+6\aB b_1\tilde{\Omega}_m\delta_m+b_2{\cal B}\,,
\label{Phieq}\\
\hspace{-.5cm}&&
\vphiN'=
-(c_t^2+b_3-4b_1\K^2)\Phi-(1-c_t^2+\aM-h-b_3-4\aB b_1\K^2)\vphiN
-6b_1\tilde{\Omega}_m\delta_m-b_3{\cal B}\,,\label{vphieq}\\
\hspace{-.5cm}&&
\frac{\delta_m'}{3(1+w_m)}=
\notag\\
\hspace{-.5cm}&&
\left[h-b_2+\left(4\aB b_1-\frac13\right)\K^2\right]\Phi
-(h-b_2-4\aB^2b_1\K^2)\vphiN
+\left(\frac{w_m}{1+w_m}-6\aB b_1\tilde{\Omega}_m\right)\delta_m
+\left(h-b_2-\frac13\K^2\right){\cal B}\,,\label{dmeq} \\
\hspace{-.5cm}&&
{\cal B}'=
(h-b_2+4\aB b_1\K^2)\Phi-(h-b_2-4\aB^2b_1\K^2)\vphiN
+\left(\frac{c_m^2}{1+w_m}-6\aB b_1\tilde{\Omega}_m\right)\delta_m+(h-b_2){\cal B}\,,
\label{Beq}
\end{eqnarray}
where $b_1, b_2, b_3$ are dimensionless quantities
defined by
\begin{equation}
b_1 \equiv \frac{1}{2(\alpha_{\rm K}+6 \alpha_{\rm B}^2)}\,,
\qquad
b_2 \equiv \frac32(1+w_m)(12\aB^2b_1-1)\tilde{\Omega}_m\,,
\qquad
b_3 \equiv 18(1+w_m)\aB b_1\tilde{\Omega}_m\,.
\end{equation}
The evolution of $\Phi$, $\vphiN$, $\delta_m$, ${\cal B}$, and
$\Psi$ is known by integrating Eqs.~(\ref{Phieq})-(\ref{Beq})
with (\ref{aniso}) together with the background
Eqs.~(\ref{back1}), (\ref{coneq}), (\ref{dH}), (\ref{ddphi}).
In GR without the scalar field $\phi$, the gravitational
potentials and the matter density contrast evolve as
$-\Psi \simeq \Phi={\rm constant}$ and
$\delta_m' \simeq \delta_m \propto a$ in the
deep matter era. In modified gravity theories, the viable
models of late-time cosmic acceleration are usually
constructed to recover the behavior close to GR in the
asymptotic past. In such cases, the initial conditions of perturbations
can be chosen to satisfy $\Phi' \simeq 0$, $\vphiN' \simeq 0$,
and $\delta_m' \simeq \delta_m$ in the early matter era.
Hence $\Phi, \vphiN, {\cal B}$ and $\Psi$ can be expressed
in terms of $\delta_m$ by using Eqs.~(\ref{Phieq})-(\ref{dmeq})
and (\ref{aniso}).
The choice of these initial conditions amounts to neglecting
the oscillating mode of $\delta \phi_{\rm N}$
induced by a heavy mass term $M_{\phi}$ larger than $H$.
The large-field mass arises for $f(R)$ dark energy models
in the asymptotic past, so in such cases, there is a
fine-tuning problem of initial conditions for avoiding the
dominance of the oscillating mode over the mode induced
by $\delta_m$ \cite{fR2,fR4}.
This problem does not arise for dark energy models
in which $M_{\phi}$ does not exceed the order
$H$ in the past (like k-essence and Galileons).
\section{Stability conditions in the small-scale limit}
\label{stasec}
By using the scalar perturbation equations of motion obtained in Sec.~\ref{scalarsec},
we derive conditions for the absence of
ghost and Laplacian instabilities in the small-scale limit.
We choose the three different gauges (\ref{Ugauge})-(\ref{Ngauge})
and show that these stability conditions are independent of the choice of gauges.
\subsection{Unitary gauge}
We begin with the unitary gauge (\ref{Ugauge}), under which the dynamical perturbations
are given by ${\cal R}=\zeta$ and $\delta \rho_{\rm u}=\delta \rho_m$.
We solve Eqs.~(\ref{eqalpha}), (\ref{eqchi}), (\ref{eqdrho}) for $\alpha$, $\chi$, $v$, and substitute them
into Eq.~(\ref{Ss}).
After the integration by parts, the quadratic action in Fourier space is
expressed in the form
\begin{equation}
{\cal S}_s^{(2)}=\int dt d^3x\,a^{3}\left(
\dot{\vec{\mathcal{X}}}^{t}{\bm K}\dot{\vec{\mathcal{X}}}
-\frac{k^2}{a^2}\vec{\mathcal{X}}^{t}{\bm G}\vec{\mathcal{X}}
-\vec{\mathcal{X}}^{t}{\bm M}\vec{\mathcal{X}}
-\vec{\mathcal{X}}^{t}{\bm B}\dot{\vec{\mathcal{X}}}
\right)\,,
\label{Ss2}
\end{equation}
where ${\bm K}$, ${\bm G}$, ${\bm M}$, ${\bm B}$
are $2 \times 2$ matrices, and
\begin{equation}
\vec{\mathcal{X}}^{t}=\left({\cal R}, \delta \rho_{\rm u}/k \right) \,.
\end{equation}
The leading-order contributions to
${\bm M}$ and ${\bm B}$ are of order $k^0$.
In the small-scale limit ($k \to \infty$),
the non-vanishing components of
matrices ${\bm K}$ and ${\bm G}$ are give by
\begin{eqnarray}
&&
K_{11}^{(\rm u)}=\frac{\tp^2 q_t q_s}{(2Hq_t-\tp D_6)^2}\,,
\qquad
K_{22}^{(\rm u)}=\frac{a^2}{2(\rho_m+P_m)}\,,
\notag\\
&&
G_{11}^{(\rm u)}=-q_tc_t^2-\frac{\rho_m+P_m}{2Hq_t-\tp D_6}{\cal F}_1
+\frac{1}{a}\frac{d}{dt}\left(a{\cal F}_1\right)\,,
\qquad
G_{22}^{(\rm u)}=\frac{a^2c_m^2}{2(\rho_m+P_m)}\,,
\label{KGu}
\end{eqnarray}
where $q_s$ is defined by Eq.~(\ref{calD}), and
\begin{equation}
{\cal F}_1=\frac{2q_t^2}{2Hq_t-\tp D_6}\,.
\end{equation}
Since there are no off-diagonal components in
${\bm K}$ and ${\bm G}$, the matter perturbation
$\delta \rho_{\rm u}$ is decoupled from the other
field ${\cal R}$. For the matter sector, the conditions
for the absence of ghost and Laplacian instabilities
correspond to $\rho_m+P_m>0$ and $c_m^2>0$.
For the perturbation ${\cal R}$, the ghost does not arise for
\begin{equation}
q_s^{(\rm u)} \equiv K_{11}^{(\rm u)}
=\frac{\tp^2 q_t q_s}{(2Hq_t-\tp D_6)^2}>0\,.
\label{qsue}
\end{equation}
Since the absence of tensor ghosts requires that $q_t>0$,
the condition (\ref{qsue}) translates to
\begin{equation}
q_s>0\,,
\label{noghost}
\end{equation}
under which the denominators of Eqs.~(\ref{dH}) and
(\ref{ddphi}) do not cross 0.
Taking the small-scale limit in Eq.~(\ref{Ss2}), we obtain
the dispersion relation of the form
\begin{equation}
{\rm det} \left( c_{\rm s}^2 {\bm K}-{\bm G} \right)=0\,,
\end{equation}
where $c_{\rm s}$ is the propagation speed of scalar perturbations.
One of the solutions is the matter propagation speed
squared $c_m^2$, while the other solution is
$c_s^2={G_{11}^{(\rm u)}}/{q_s^{(\rm u)}}$.
On using Eqs.~(\ref{back2}), (\ref{Bi}), and (\ref{conD2})
to solve for $\ddot{\phi}$, $\dot{q}_t$, and $\dot{D_6}$,
the latter solution can be simply expressed as
\begin{equation}
c_s^2=\frac{G_{11}^{(\rm u)}}{q_s^{(\rm u)}}
=-\frac{c_t^2D_6^2+2B_1D_6+4q_t D_2}{q_s}\,.
\label{cs}
\end{equation}
The small-scale Laplacian instability can be avoided for
\begin{equation}
c_s^2>0\,.
\end{equation}
In sections after \ref{quinsec}, we will compute $c_s^2$
for concrete dark energy models in
the framework of Horndeski theories.
\subsection{Flat gauge}
In the flat gauge (\ref{Fgauge}), the dynamical
scalar perturbations are given by the matrix
$\vec{\mathcal{X}}^{t}=
\left(\dphi_{\rm f}, \delta \rho_{\rm f}/k \right)$.
Solving Eqs.~(\ref{eqalpha}), (\ref{eqchi}), (\ref{eqdrho}) for $\alpha$, $\chi$, $v$
and substituting them into Eq.~(\ref{Ss}), the second-order action reduces to
the same form as Eq.~(\ref{Ss2}) after the integration by parts.
In the small-scale limit, the matrix components
$K_{22}^{(\rm f)}$ and $G_{22}^{(\rm f)}$
are identical to those in the unitary gauge.
The other nonvanishing matrix components
are given by
\begin{eqnarray}
&&
q_s^{(\rm f)} \equiv K_{11}^{(\rm f)}=\frac{H^2 q_t q_s}
{(2Hq_t-\tp D_6)^2}\,,
\qquad
G_{11}^{(\rm f)}=-D_2+\frac{D_6D_7-(\rho_m+P_m){\cal F}_2}{2Hq_t-\tp D_6}
+\frac{1}{a}\frac{d}{dt}\left(a{\cal F}_2\right)\,,
\label{qsf}
\end{eqnarray}
where
\begin{equation}
{\cal F}_2=\frac{D_6^2}{2(2Hq_t-\tp D_6)}\,.
\end{equation}
Note that we eliminated the term $D_4$ by using the relation (\ref{conD4}).
As long as there is no tensor ghost, the condition for avoiding
the scalar ghost again corresponds to $q_s>0$.
On using Eqs.~(\ref{back2}) and (\ref{conD2}),
it follows that
\begin{equation}
\frac{q_s^{(\rm f)}}{q_s^{(\rm u)}}=
\frac{G_{11}^{(\rm f)}}{G_{11}^{(\rm u)}}=
\frac{H^2}{\tp^2}\,.
\end{equation}
Hence the scalar propagation speed squared
$c_s^2=G_{11}^{(\rm f)}/q_s^{(\rm f)}$
in the flat gauge is equivalent to that in the unitary gauge.
\subsection{Newtonian gauge}
\label{Newtonsec}
Let us finally consider the Newtonian gauge (\ref{Ngauge}), under
which $\Phi=\zeta$, $\delta \phi_{\rm N}=\delta \phi$, and
$\delta \rho_{\rm N}=\delta \rho_m$.
We first solve Eqs.~(\ref{eqalpha}) and (\ref{eqdrho}) for nondynamical perturbations
$\alpha$ and $v$, and substitute them into Eq.~(\ref{Ss}).
Then, the terms proportional to $\dot{\Phi}^2$,
$\dot{\dphi}_{\rm N}^2$, and $\dot{\Phi}\dot{\dphi}_{\rm N}$ appear in the
second-order action. Apparently, this looks as if the two fields
$\Phi$ and $\dphi_{\rm N}$ were dynamical, but the second-order action
in the small-scale limit is factorized as
\begin{equation}
{\cal S}_s^{(2)}=\int dt d^3x\,a^{3}\left[
\frac{3q_s}{4\{ 3H^2 q_t-\dot{\phi} (D_1 \dot{\phi}+3H D_6)\}}
\left( \dot{\phi} \dot{\Phi}-H \dot{\delta \phi}_{\rm N} \right)^2
+\frac{a^2}{2(\rho_m+P_m)} \frac{ \dot{\delta \rho}_{\rm N}^2}{k^2}
+\cdots
\right]\,,
\label{SsN}
\end{equation}
where the abbreviation corresponds to terms without containing the product
of first-order time derivatives of perturbed quantities.
If we consider the combination
\begin{equation}
{\cal R}=\Phi-\frac{H}{\dot{\phi}}\delta \phi_{\rm N}\,,\qquad {\rm or} \qquad
\delta \phi_{\rm f}=\delta \phi_{\rm N}-\frac{\dot{\phi}}{H}\Phi\,,
\end{equation}
then the action (\ref{SsN}) can be expressed in terms of $\dot{{\cal R}}^2$ or
$\dot{\delta \phi}_{\rm f}^2$ in addition to $\dot{\delta \rho}_{\rm N}^2$,
without any other products of first-order time derivatives of perturbations.
In other words, ${\cal R}$ or $\dphi_{\rm f}$ corresponds to
the dynamical perturbation besides the matter perturbation $\delta \rho_{\rm N}$.
If we choose ${\cal R}$ as a dynamical perturbation, the terms in the abbreviation
of Eq.~(\ref{SsN}) contain nondynamical variables like $\dot{\delta \phi}_{\rm N}$ and
$\delta \phi_{\rm N}$, which can be eliminated by using Eqs.~(\ref{eqchi}) and (\ref{eqE2}).
After the integration by parts, the second-order action (\ref{SsN}) reduces to
the same form as Eq.~(\ref{Ss2}) with dynamical perturbations
$\vec{\mathcal{X}}^{t}=
\left({\cal R}, \delta \rho_{\rm N}/k \right)$.
In the small-scale limit, the components of matrices ${\bm K}$ and ${\bm G}$
are identical to those in the unitary gauge given in Eq.~(\ref{KGu}).
If the combination $\delta \phi_{\rm f}$ is chosen as a dynamical perturbation besides
$\delta \rho_{\rm N}$,
the nondynamical variables like $\dot{\Phi}$ and $\Phi$ can be eliminated from
the action (\ref{SsN}) by using Eqs.~(\ref{eqchi}) and (\ref{eqE2}).
In the small-scale limit, the resulting second-order action is expressed
in the form (\ref{Ss2}) with the same matrix components of
${\bm K}$ and ${\bm G}$ as those in the flat gauge.
Thus, the conditions for the absence of ghost and Laplacian instabilities in the Newtonian gauge
are equivalent to those in the unitary and flat gauges.
\subsection{Summary}
We have shown that the quantities
$q_s^{(\rm u)}$ and $q_s^{(\rm f)}$
contain the common term $q_s$.
The positivity of this term, i.e.,
\begin{equation}
q_s=4D_1 q_t+3D_6^2>0\,,
\label{qscon}
\end{equation}
is the condition for the absence of scalar ghosts
for any gauge choices.
The small-scale Laplacian instability can be
avoided as long as the gauge-invariant sound speed
squared is positive, i.e.,
\begin{equation}
c_s^2
=\frac{G_{11}^{(\rm u)}}{q_s^{(\rm u)}}
=\frac{G_{11}^{(\rm f)}}{q_s^{(\rm f)}}
=-\frac{c_t^2D_6^2+2B_1D_6+4q_t D_2}{q_s}>0\,.
\label{csge}
\end{equation}
Depending on the problems at hand, we can choose most
convenient gauges.
We caution that the perturbations $\delta \phi_{\rm f}$
and $\delta \rho_{\rm f}$ contain the term $H$
in their denominators, so they are not well defined
at $H=0$. In this case, it apparently
looks possible to regulate the combination $\zeta/H$
by choosing the flat gauge ($\zeta=0$).
However, the problem arises for the gravitational
potential $\Phi$ and the curvature perturbation ${\cal R}$,
both of which exactly vanish at $H=0$ for the gauge
choice $\zeta=0$.
This suggests that, for the cosmological evolution crossing
$H=0$ (such as the bouncing cosmology),
it is not appropriate to choose the flat gauge \cite{HKT18}.
In the bouncing Universe, the quantity
$q_s^{(\rm f)}$ vanishes at the bounce ($H=0$). Apparently,
this looks as if a strong coupling problem arises, but
it simply comes from the inappropriate gauge choice
for this problem.
The real strong coupling problem arises when the
quantity $q_s$ in Eq.~(\ref{qscon}) approaches 0,
in which case the background Eqs.~(\ref{dH})
and (\ref{ddphi}) exhibit divergences.
In the unitary gauge, ${\cal R}$ and $\delta \rho_{\rm u}$ corresponds to dynamical perturbations, but
they are not well defined at $\dot{\phi}=0$.
For this gauge choice, the problems similar to those mentioned
above arise for the case in which $\dot{\phi}$ crosses 0.
If neither $H$ nor $\dot{\phi}$ vanishes during the cosmological evolution,
we can choose any gauge among
(\ref{Ugauge})-(\ref{Ngauge}).
\section{Growth of large-scale structures}
\label{obsersec}
The dark energy EOS $w_{\rm DE}$ introduced
in Sec.~\ref{backsec2} is a key quantity to distinguish
between different dark energy models at the background level.
In modified gravity theories, the gravitational coupling with
the matter sector is different from that in GR.
In this case, the growth rate of matter perturbations and
the evolution of gravitational potentials are subject to
modifications. Then, we can distinguish between different
models of the late-time cosmic acceleration from
the observations of large-scale structures, weak lensing,
and CMB.
In this section, we study observables associated
with the evolution of linear cosmological perturbations in full
Horndeski theories given by the action (\ref{actionfull}).
By using the linear perturbation equations of motion derived
in Sec.~\ref{scalarsec}, it is possible to estimate the effective gravitational coupling
with matter perturbations on scales
relevant to the growth of large-scale structures in the gauge-invariant way.
We also discuss the gravitational coupling around local objects
screened by nonlinear interactions.
\subsection{Observable quantities}
Since we are interested in the evolution of perturbations
after the end of the radiation-dominated epoch,
we consider nonrelativistic matter satisfying
\begin{equation}
P_m=0\,,\qquad c_m^2=0\,,
\end{equation}
for the action ${\cal S}_m$.
Taking the time derivative of Eq.~(\ref{eqdrho}) and
using Eq.~(\ref{veq}), the gauge-invariant matter
density contrast $\delta_m=\delta \rho_m/\rho_m+3Hv$
obeys
\begin{equation}
\ddot{\delta}_m+2H\dot{\delta}_m+\frac{k^2}{a^2}\Psi
=3 \left( \ddot{\cal B}+2H \dot{\cal B} \right)\,,
\label{delmeq}
\end{equation}
where $\Psi$ and ${\cal B}$ are gauge-invariant
perturbations defined in Eqs.~(\ref{PsiPhi}) and (\ref{delmB}).
We relate the gravitational potential $\Psi$ with
the density contrast $\delta_m$ through the modified
Poisson equation
\begin{equation}
\frac{k^2}{a^2}\Psi
=-4\pi G\mu \rho_m \delta_m\,,
\label{Poi}
\end{equation}
where
\begin{equation}
\mu=\frac{G_{\rm eff}}{G}\,.
\label{mudef}
\end{equation}
The quantity $\mu$ characterizes the ratio between the
effective gravitational coupling $G_{\rm eff}$ and
the Newton constant $G=1/(8\pi M_{\rm pl}^2)$.
In Sec.~\ref{quasisec}, we will derive an explicit form of
$\mu$ in full Horndeski theories by using a so-called
quasi-static approximation for perturbations deep
inside the sound horizon.
We also define the gravitational slip parameter $\eta$ and
the effective gravitational potential $\psi_{\rm eff}$
relevant to the light bending in weak lensing and CMB
observations, as \cite{weak1,weak2}
\begin{equation}
\eta=-\frac{\Phi}{\Psi}\,,\qquad
\psi_{\rm eff}=\Phi-\Psi\,.
\label{etadef}
\end{equation}
{}From Eqs.~(\ref{Poi}) and (\ref{etadef}), it follows that
\begin{equation}
\frac{k^2}{a^2}\psi_{\rm eff}=
8\pi G\,\Sigma\rho_m \delta_m\,,
\label{psieff}
\end{equation}
where
\begin{equation}
\Sigma=\frac{1+\eta}{2}\frac{G_{\rm eff}}{G}\,.
\label{Sigmadef}
\end{equation}
The dimensional ratios $\mu$ and $\Sigma$ are
two key quantities characterizing the linear growth of
$\delta_m$ as well as $\Psi, \psi_{\rm eff}$.
\subsection{Quasi-static approximation deep inside
the sound horizon}
\label{quasisec}
To confront dark energy models in the framework of Horndeski
theories with the observations of large-scale structures and
weak lensing, we would like to derive analytic expressions of
$\mu$ and $\Sigma$ for the perturbations
deep inside the sound horizon.
In doing so, we exploit the scalar perturbation equations of motion
expressed in terms of gauge-invariant variables.
As we will see below, we do not need to fix gauges for the derivation
of $\mu$ and $\Sigma$.
In Sec.~\ref{stasec}, we showed that the curvature perturbation
${\cal R}$ (or equivalently $\dphi_{\rm f}=-\tp {\cal R}/H$) and
the matter perturbation are the dynamical degrees of freedom.
Let us derive the closed-form equation of motion for ${\cal R}$
in the small-scale limit.
First of all, we replace $\Phi$ with ${\cal R}$ in the perturbation equations
(\ref{dphiNeq}), (\ref{ddPhi}) and (\ref{eqalpha2})-(\ref{aniso}) by using
the relation ${\cal R}=\Phi-\dphiv_{\rm N}$.
Secondly, we solve Eqs.~(\ref{eqalpha2})-(\ref{aniso}) and the
time derivative of Eq.~(\ref{aniso}) for $\dot{\cal B}$, ${\cal B}$,
$\dot{\delta \varphi_{\rm N}}$, $\delta \varphi_{\rm N}$, $\dot{\Psi}$ and $\Psi$
in order to express them in terms of $\dot{\cal R}$, ${\cal R}$,
$\dot{\delta_m}$ and $\delta_m$.
Finally, we substitute these solutions into the equation derived after
eliminating $\ddot{\dphi_{\rm N}}$ from Eqs.~(\ref{dphiNeq})
and (\ref{ddPhi}). Taking the $k\to\infty$ limit,
this equation reduces to
\begin{equation}
\ddot{\cal R}+\left(3H+\frac{\dot{q_s}^{(\rm u)}}{q_s^{(\rm u)}}\right) \dot{\cal R}
+c_s^2\frac{k^2}{a^2}{\cal R}\simeq
\frac{1}{2q_t q_s^{(\rm u)}}\left(q_t c_t^2+q_s^{(\rm u)} c_s^2-\frac{1+\aM}{1+\aB}q_t\right)
\rho_m\delta_m\,,
\label{ddotR}
\end{equation}
where we used Eqs.~(\ref{Bi}), (\ref{conD2}), (\ref{nodim}), (\ref{qsue}) and (\ref{cs})
to remove $D_1$, $D_2$, $D_6$, $D_7$ and $\dot{q_t}$.
The general solution to Eq.~(\ref{ddotR}) can be expressed in the form
\begin{equation}
{\cal R}={\cal R}_{\rm ho}+{\cal R}_{\rm ind}\,,
\end{equation}
where ${\cal R}_{\rm ho}$ is a homogenous solution obtained by setting
the right hand side of Eq.~(\ref{ddotR}) to be 0, and ${\cal R}_{\rm ind}$
is the special solution induced by the matter perturbation $\delta_m$.
The solution ${\cal R}_{\rm ho}$ contains an oscillating mode induced
by the second time derivative $\ddot{\cal R}$.
Provided that the variation of ${\cal R}$ over the Hubble time scale is not large such that
the conditions $|\ddot{{\cal R}}| \lesssim |H^2 {\cal R}|$ and $|\dot{{\cal R}}| \lesssim |H {\cal R}|$
are satisfied with
$|\dot{q_s}^{(\rm u)}/q_s^{(\rm u)}| \lesssim H$, the dominant contribution to
the left hand side of Eq.~(\ref{ddotR}) is the term $c_s^2k^2 {\cal R}/a^2$
for the modes deep inside the sound horizon ($c_s^2k^2/a^2 \gg H^2$).
This term balances with the matter perturbation $\delta_m$, in which case
the curvature perturbation ${\cal R}$ is dominated by the matter
induced mode ${\cal R}_{\rm ind}$.
In the so-called quasi-static approximation, the dominant contributions to
the perturbation equations are regarded as those containing the terms
$\delta_m$ and $k^2/a^2$, without any time derivatives
of metric perturbations \cite{Boi00,dreview4,Tsuji07,DKT12,Nesseris,Song}.
Provided that $c_s^2$ is not very close to 0, the quasi-static approximation
is sufficiently accurate for sub-horizon perturbations in dark energy models
where the mass $M_{\phi}$
associated with the field perturbation $\delta \phi_{\rm N}$ is at most of
order $H$. This is generally the case for a nearly massless scalar field
like quintessence, k-essence, and Galileons \cite{Kase10}.
One exception is $f(R)$ models of the late-time cosmic acceleration, in which case
the scalar mass $M_{\phi}$ becomes larger than $H$ as we go back to the past.
Then, the oscillating mode of field perturbations cannot be ignored
relative to the matter-induced mode in the asymptotic past \cite{fR2,fR4}.
In other words, in $f(R)$ gravity, we need a fine tuning for initial conditions
of perturbations such that $|{\cal R}_{\rm ind}| \gg |{\cal R}_{\rm ho}|$.
In the following, we apply the quasi-static approximation
to the scalar perturbation equations of motion.
First of all, we remind that there is the relation (\ref{aniso}) among two
gravitational potentials $\Psi$ and $\Phi$.
Applying the quasi-static approximation
to Eq.~(\ref{eqalpha2}), it follows that
\begin{equation}
2q_t \frac{k^2}{a^2} \left( \Phi + \aB \dphiv_{\rm N} \right)
\simeq \rho_m \delta_m\,.
\label{quasi2}
\end{equation}
In Eq.~(\ref{dphiNeq}), there exists the mass term
$M_{\phi}^2\dphi_{\rm N}$ of the scalar-field perturbation.
In viable models of the late-time cosmic acceleration
based on $f(R)$ gravity \cite{fR1,fR2,fR3,fR4,fR5}
and BD theories with the field potential \cite{BD1,BD2},
the mass of $\dphi_{\rm N}$ tends to be large
in the early cosmological epoch.
Taking into account its contribution
and using the quasi-static approximation
in Eq.~(\ref{dphiNeq}), we obtain
\begin{equation}
M_{\phi}^2 \delta \phi_{\rm N}
+\frac{k^2}{a^2} \left( D_6 \Psi-B_1 \Phi-2D_2
\delta \phi_{\rm N} \right) \simeq 0\,.
\label{quasi3d}
\end{equation}
{}From Eq.~(\ref{cs}), the coefficient $D_2$ can be expressed
as $D_2=-(q_s c_s^2+c_t^2 D_6^2+2B_1 D_6)/(4q_t)$.
We also note that $B_1$ defined in Eq.~(\ref{Bi})
is written as $B_1=2Hq_t(1-c_t^2+\alpha_{\rm M})/\dot{\phi}$.
Then, Eq.~(\ref{quasi3d}) can be expressed as
\begin{equation}
q_t\frac{k^2}{a^2}\left[\aB\Psi+(1-c_t^2+\aM)\Phi\right]
-\left[ q_t \frac{k^2}{a^2} \alpha_{\rm B}
\left\{ c_t^2 (2+\alpha_{\rm B})-2(1+\alpha_{\rm M})
\right\}+\frac{\dot{\phi}^2 (c_s^2 k^2 q_s
+2M_{\phi}^2a^2 q_t)}{4a^2H^2 q_t}
\right] \delta \varphi_{\rm N} \simeq 0\,,
\label{quasi3}
\end{equation}
where we used $D_6=-2Hq_t \alpha_{\rm B}/\dot{\phi}$.
Solving Eqs.~(\ref{aniso}), (\ref{quasi2}), and (\ref{quasi3})
for $\Psi, \Phi$, and $\dphiv_{\rm N}$, we obtain
\begin{equation}
\Psi=-\frac{1}{2\Delta_2}
\left(\Delta_1^2+\frac{c_t^2\Delta_2}{q_t}\right)
\frac{a^2}{k^2}\rho_m\delta_m\,,
\qquad
\Phi=\frac{1}{2\Delta_2}
\left(\aB\Delta_1+\frac{\Delta_2}{q_t}\right)
\frac{a^2}{k^2}\rho_m\delta_m\,,
\qquad
\dphiv_{\rm N}=-\frac{\Delta_1}{2\Delta_2}
\frac{a^2}{k^2}\rho_m\delta_m\,,
\end{equation}
where
\begin{equation}
\Delta_1 \equiv c_t^2(1+\aB)-1-\aM\,,
\qquad
\Delta_2 \equiv \frac{\tp^2q_sc_s^2}{4H^2q_t}
\left( 1+\frac{2a^2 M_{\phi}^2 q_t}{c_s^2 k^2 q_s}
\right)\,.
\label{Delta2}
\end{equation}
Then the quantities $\mu$, $\eta$, and $\Sigma$ defined
by Eqs.~(\ref{mudef}), (\ref{etadef}), and
(\ref{Sigmadef}) reduce to
\begin{eqnarray}
\mu
&=& \frac{c_t^2}{8\pi Gq_t}
\left(1+\frac{q_t \Delta_1^2}{c_t^2\Delta_2} \right)\,,
\label{Gefff}\\
\eta &=&
\frac{q_t \aB\Delta_1+\Delta_2}
{q_t \Delta_1^2+c_t^2\Delta_2}\,,
\label{etaf}\\
\Sigma &=&
\frac{1+c_t^2}{16\pi Gq_t}
\left[ 1+\frac{q_t (\aB+\Delta_1)\Delta_1}
{(1+c_t^2)\Delta_2} \right]\,.
\label{Sigmaf}
\end{eqnarray}
The effective gravitational coupling $\mu=G_{\rm eff}/G$ is composed of the following two contributions :
\begin{equation}
\mu_t=\frac{c_t^2}{8\pi G q_t}\,,\qquad
\xi_s=\frac{q_t \Delta_1^2}{c_t^2\Delta_2}\,.
\end{equation}
The term $\mu_t$ arises from the modification of gravity in the tensor sector. Under the no-ghost and stability conditions $q_t>0$ and $c_t^2>0$, we have $\mu_t>0$.
The term $\xi_s$ quantifies the interaction between
the scalar field $\phi$ and matter.
For the field mass squared $M_{\phi}^2>0$, the quantity
$\xi_s$ is positive under the no-ghost conditions
$q_t>0, q_s>0$ and the stability conditions $c_t^2>0, c_s^2>0$.
Thus, the scalar-matter interaction in full Horndeski theories
is attractive under theoretically consistent
conditions \cite{Tsuji15,Kase18}.
Since $\mu, \eta, \Sigma$ given in
Eqs.~(\ref{Gefff})-(\ref{Sigmaf}) have been derived without
fixing any gauge conditions, they are gauge-invariant
quantities for the modes deep inside the sound horizon.
The modification to the effective gravitational coupling
manifests itself in the ``massless'' regime in which the condition
$M_{\phi}^2 \lesssim (c_s^2 k^2/a^2)(q_s/q_t)$ holds.
For $M_{\phi}^2 \lesssim H^2$ with $q_s/q_t={\cal O}(1)$,
this condition is satisfied for perturbations inside
the sound horizon ($H^2 \lesssim c_s^2 k^2/a^2$).
Taking the massless limit $M_{\phi}^2 \to 0$ in $\Delta_2$,
the $k$ dependence disappears in the expressions of
$\mu, \eta, \Sigma$.
In Ref.~\cite{Kase18}, the present authors derived these quantities
by choosing the unitary gauge.
Since there is the relation
$q_s^{(\rm u)}/q_s=\tp^2/[4H^2q_t(1+\aB)^2]$ from Eq.~(\ref{qsue}),
the massless limits of Eqs.~(\ref{Gefff})
and (\ref{etaf}) yield
\begin{equation}
\mu=\frac{c_t^2}{8\pi G q_t} \left[ 1+
\frac{q_t \Delta_1^2}{q_s^{(\rm u)}
c_s^2 c_t^2 (1+\alpha_{\rm B})^2} \right]\,,\qquad
\eta=\frac{q_t \aB \Delta_1+q_s^{(\rm u)}
c_s^2 (1+\alpha_{\rm B})^2}
{q_t \Delta_1^2+q_s^{(\rm u)}
c_s^2 c_t^2 (1+\alpha_{\rm B})^2}\,,
\end{equation}
which coincide with Eqs.~(3.33) and (3.34) of
Ref.~\cite{Kase18}, respectively, after replacing $q_t$
with $2Q_t$.
The GW170817 event \cite{GW170817} placed the
tight bound (\ref{ctbound}) on the tensor propagation speed $c_t$.
Setting $c_t^2=1$ in Eqs.~(\ref{Gefff})-(\ref{Sigmaf}),
it follows that
%
\begin{eqnarray}
\mu
&=& \frac{1}{8\pi Gq_t}
\left[1+\frac{q_t (\aB-\aM)^2}{\Delta_2} \right]\,,
\label{Gefff2}\\
\eta &=&
1+\frac{q_t\aM(\aB-\aM)}{\Delta_2+q_t(\aB-\aM)^2}\,,
\label{eta2}\\
\Sigma &=& \frac{1}{8\pi Gq_t}
\left[ 1+\frac{q_t (2\aB-\aM)(\aB-\aM)}{2\Delta_2} \right]\,,
\label{Sigmaf2}
\end{eqnarray}
where we used the property $\Delta_1=\aB-\aM$.
The nonvanishing values of $\alpha_{\rm B}$ and
$\alpha_{\rm M}$ satisfying $\alpha_{\rm B}
\neq \alpha_{\rm M}$ lead to the enhancement
of the gravitational interaction with matter, such that
$\mu>1/(8\pi G q_t)$.
The gravitational slip ($\eta \neq 1$) arises for the theories
with $\alpha_{\rm M} \neq 0$ and $\alpha_{\rm B} \neq \alpha_{\rm M}$.
Under the no-ghost and stability conditions,
the quantity $\Sigma$ is larger
than $1/(8\pi Gq_t)$ for $(2\alpha_{\rm B}-\alpha_{\rm M})
(\alpha_{\rm B}-\alpha_{\rm M})>0$.
{}From Eq.~(\ref{veq2}), the perturbation ${\cal B}$ is at most
of order $\Psi$. Then the terms on the right hand side
of Eq.~(\ref{delmeq}) is at most of order $H^2 \Psi$, so
they can be ignored compared to those on its left hand side
for the perturbations deep inside the Hubble radius.
On using Eq.~(\ref{Poi}), the matter density contrast
approximately obeys
\begin{equation}
\ddot{\delta}_m+2H \dot{\delta}_m
-4\pi G \mu \rho_m \delta_m \simeq 0\,.
\label{delmeqf}
\end{equation}
For a given model the quantity $\mu$ is known
from Eq.~(\ref{Gefff2}), so we can integrate
Eq.~(\ref{delmeqf}) to solve for $\delta_m$.
The matter power spectrum and the growth rate of
matter perturbations can be constrained
from the measurements of galaxy
clusterings \cite{2df,Tegmark} and redshift-space
distortions \cite{Kaiser,RSD1,RSD2,RSD3,RSD4,Okumura},
so it is possible to distinguish between different dark energy models.
The evolution of two gravitational potentials is also
determined from Eqs.~(\ref{Poi}) and (\ref{psieff}).
This information can be used to place further constraints
on dark energy models from the observations of
CMB \cite{Planck2015,Planck2018}
and weak lensing \cite{lensing1,lensing2}.
\subsection{Screened gravitational coupling}
\label{screensec}
We have shown that Horndeski theories generally give
rise to modifications to the gravitational interaction
for scales relevant to the growth of large-scale structures
and weak lensing.
In local regions of the Universe, the fifth force induced by the scalar-matter interaction needs to be small
for the consistency with solar-system tests of gravity.
There are several mechanisms to suppress the propagation of fifth forces in regions of the high density:
(i) chameleon mechanism \cite{chame1}, and
(ii) Vainshtein mechanism \cite{Vain}.
The chameleon mechanism can be at work for a scalar
potential whose mass is different depending on the matter
densities in the surrounding environment.
If the effective scalar mass is sufficiently large in regions of the high density, the coupling between the field and matter can
be suppressed by having a thin shell inside a spherically symmetric body (see Ref.~\cite{chame2} for detail).
In $f(R)$ gravity or BD theories with
a scalar potential, it is possible to design functional forms of $f(R)$ or the potential $V(\phi)$ to realize the large mass
squared $M_{\phi}^2$ for increasing $R$, while realizing
the late-time cosmic acceleration by the potential of a light
scalar field \cite{fR1,fR2,fR3,fR4,BD1,BD2}.
Cosmologically, $M_{\phi}^2$ decreases in time,
so there is a transition from the ``massive'' regime
$M_{\phi}^2 \gg (c_s^2 k^2/a^2)(q_s/q_t)$ to
the ``massless'' regime
$M_{\phi}^2 \ll (c_s^2 k^2/a^2)(q_s/q_t)$.
In the massive limit $M_{\phi}^2 \to \infty$, we have
$\xi_s \to 0$ and $\mu \to c_t^2/(8\pi G q_t)$, by reflecting
the fact that the scalar degree of freedom does not propagate.
This is the regime in which the chameleon mechanism is
at work in the local region whose matter density $\rho_m$
is much higher than today's critical cosmological
density $\rho_c$.
The Vainshtein mechanism operates around local sources
in the presence of nonlinear scalar derivative interactions.
One of the representative examples is the cubic Galileon
given by the Lagrangian $X \square \phi$.
This nonlinear interaction leads to the decoupling of
the field $\phi$ from matter within a radius $r_V$ called
the Vainshtein radius (see Refs.~\cite{DKT12,KKY12,Kase13}
for detail).
For the Sun, the Vainshtein radius
can be of order $10^{20}$~cm, which is much larger
than the solar-system scale. On scales relevant to
the growth of large-scale structures ($\gtrsim 10^{24}$~cm),
the cubic Galileon modifies the effective gravitational
coupling $\mu$, while, in the solar system,
the scalar-matter interaction term
$\xi_s=q_t \Delta_1^2/(c_t^2 \Delta_2)$ in Eq.~(\ref{Gefff})
is much smaller than 1.
In the rest of this section, we focus the case in which
$c_t^2$ is close to 1, so that $\mu, \eta, \Sigma$ are
given by Eq.~(\ref{Gefff2})-(\ref{Sigmaf2}) on
scales relevant to the growth of large-scale structures.
If the screening of fifth forces occurs efficiently around
local sources, the screened gravitational coupling
$G_{\rm sc}$ is given by \cite{KWY}
\begin{equation}
G_{\rm sc}(t)=\frac{1}{8\pi q_t(t)}
=\frac{1}{16\pi} \left[ G_4-\tp^2G_{4,X}
+\frac12\tp^2G_{5,\phi}
-\frac12H \dot{\phi}^3G_{5,X} \right]^{-1}\,.
\label{Gsc}
\end{equation}
If we strictly demand that $c_t^2=1$, we have
$G_{\rm sc}(t)=1/[16\pi G_4(\phi)]$.
Since we live in a screened environment, today's value
of Eq.~(\ref{Gsc}), i.e.,
$G_{\rm sc}(t_0)=1/(8\pi q_t(t_0))$ should be close
to the Newton gravitational constant $G$, such that
\begin{equation}
q_t(t_0) \simeq \frac{1}{8 \pi G}\,.
\end{equation}
The quantity $\mu_t$ reduces to
\begin{equation}
\mu_t (t) \simeq \frac{q_t(t_0)}{q_t(t)}\,.
\end{equation}
If $q_t(t)<q_t(t_0)$ in the past, we have $\mu_t>1$
and hence $\mu>1$ under the conditions
$q_s>0, q_t>0, c_s^2>0$, and $M_{\phi}^2>0$.
In this case, the effective gravitational coupling
$G_{\rm eff}$ is larger than $G$ for scales relevant
to the linear growth of large-scale structures.
In the opposite case, $q_t(t)>q_t(t_0)$, $\mu_t$ is
smaller than 1. This is the necessary condition
for realizing $G_{\rm eff}<G$, but it is not sufficient
due to the existence of the positive term
$\xi_s=q_t (\aB-\aM)^2/\Delta_2$ in Eq.~(\ref{Gefff2}).
In other words, even if $q_t(t)>q_t(t_0)$ in the past,
the scalar-matter interaction $\xi_s$ can lead to
$G_{\rm eff}$ larger than $G$.
There are bounds on the variation of the gravitational coupling
constrained from Lunar Laser Ranging experiments \cite{William,Babi11}.
In the screened environment, the experimental bound
corresponds to $|\dot{G}_{\rm sc}/G_{\rm sc}|<0.02H_0$,
where $H_0$ is today's value of the Hubble parameter.
On using Eq.~(\ref{Gsc}), this bound translates to
\begin{equation}
|\alpha_{\rm M}(t_0)|<0.02\,.
\label{alpM}
\end{equation}
Assuming that the quantity $\alpha_{\rm M}$ is nearly constant
around today, we obtain
$q_t(t)=q_t(t_0) e^{(t-t_0)H_0 \alpha_{\rm M}(t_0)}$ and
hence $\mu_t \simeq e^{-(t-t_0)H_0 \alpha_{\rm M}(t_0)}$.
The difference of $\mu_t$ from 1 over the cosmological
time scale ($t-t_0 \sim 1/H_0$) is of order
$\alpha_{\rm M}(t_0)$, so the modification to $\mu=G_{\rm eff}/G$
arising from the tensor contribution $\mu_t$ is
suppressed under the bound (\ref{alpM}).
The scalar-matter contribution
$\xi_s=q_t (\aB-\aM)^2/\Delta_2$ is the main source
for modifying the gravitational interaction on scales
relevant to the growth of large-scale structures.
The two quantities $\alpha_{\rm B}$ and $\alpha_{\rm M}$
play important roles for the evolution of gravitational potentials
$\Psi$ and $\Phi$.
\subsection{Classification of surviving Horndeski
theories in terms of $\mu$ and $\Sigma$}
\label{classsec}
In the following, we focus on Horndeski theories given by
the Lagrangian (\ref{lagcon}), i.e., those satisfying
the condition $c_t^2=1$.
Then, the quantities $\alpha_{\rm B}$ and $\alpha_{\rm M}$
reduce, respectively, to
\begin{equation}
\alpha_{\rm B}=\frac{2\dot{\phi}G_{4,\phi}+\dot{\phi}^3G_{3,X}}
{4H G_4}\,,\qquad
\alpha_{\rm M}=\frac{\dot{\phi}G_{4,\phi}}{HG_4}\,,
\label{alBM}
\end{equation}
with $q_t=2G_4>0$.
Depending on the values of $\alpha_{\rm B}$ and $\alpha_{\rm M}$,
the surviving theories can be classified
into the following four classes.
\begin{itemize}
\item (A) $G_2=G_2(\phi,X)$, $G_3=0$, $G_4=M_{\rm pl}^2/2$.
This class accommodates both quintessence and k-essence. Since
$\alpha_{\rm B}=0=\alpha_{\rm M}$ in Eqs.~(\ref{Gefff2})-(\ref{Sigmaf2}),
it follows that
\begin{equation}
\mu=1\,,\qquad \eta=1\,,\qquad \Sigma=1\,.
\end{equation}
Hence $G_{\rm eff}$ is equivalent to $G$ without the gravitational slip.
\item (B) $G_2=G_2(\phi,X)$, $G_3=0$, $G_4=G_4(\phi)$.
This class includes metric $f(R)$ gravity and BD theories.
Since there is the specific relation
\begin{equation}
\alpha_{\rm B}=\frac{\alpha_{\rm M}}{2}\,,
\end{equation}
we have
\begin{equation}
\mu=\frac{1}{16\pi G\,G_4} \left( 1+\frac{G_4 \alpha_{\rm M}^2}
{2\Delta_2} \right)\,, \qquad
\eta=\frac{2\Delta_2-G_4\alpha_{\rm M}^2}
{2\Delta_2+G_4\alpha_{\rm M}^2}\,,\qquad
\Sigma=\frac{1}{16\pi G\,G_4}\,.
\end{equation}
The nonminimal coupling $G_4(\phi)$ enhances the gravitational
interaction with matter ($\xi_s=G_4 \alpha_{\rm M}^2/(2\Delta_2)>0$).
There is the difference between $\mu$ and $\Sigma$, so the Newtonian
gravitational potential $\Psi$ and the weak lensing potential $\psi_{\rm eff}$
evolve in different ways.
\item (C) $G_2=G_2(\phi, X)$, $G_3=G_3(\phi, X)$,
$G_4=M_{\rm pl}^2/2$.
The theories of this class are known as kinetic
braidings, which accommodate cubic Galileons
as a specific case. Since
\begin{equation}
\alpha_{\rm B}=\frac{\dot{\phi}^3 G_{3,X}}{2H M_{\rm pl}^2}\,,
\qquad \alpha_{\rm M}=0\,,
\end{equation}
we obtain
\begin{equation}
\mu=\Sigma=1+\frac{M_{\rm pl}^2 \alpha_{\rm B}^2}{\Delta_2}\,,
\qquad \eta=1\,.
\label{muC}
\end{equation}
Unlike the case (B) there is no gravitational slip.
The cubic derivative coupling $G_3(X)$ enhances the two gravitational
potentials $-\Psi$ and $\Phi$ in the same manner.
\item (D) $G_2=G_2(\phi, X)$, $G_3=G_3(\phi,X)$, $G_4=G_4(\phi)$.
This is the most general case including kinetic braidings and its extensions.
The relation between $\alpha_{\rm B}$ and $\alpha_{\rm M}$ is
\begin{equation}
\alpha_{\rm B}-\frac{\alpha_{\rm M}}{2}
=\frac{\dot{\phi}^3 G_{3,X}}{4H G_4}\,.
\end{equation}
The original kinetic braiding scenario \cite{braiding1} corresponds to
$G_4=M_{\rm pl}^2/2$, in which case
$\mu, \eta, \Sigma$ are of the same forms as those
given in Eq.~(\ref{muC}).
{}From Eqs.~(\ref{Gefff2}) and (\ref{Sigmaf2}), the difference between
$\mu$ and $\Sigma$ is
\begin{equation}
\mu-\Sigma=-\frac{\alpha_{\rm M}(\alpha_{\rm B}-\alpha_{\rm M})}
{16\pi G \Delta_2}\,.
\end{equation}
For the theories with $\alpha_{\rm M}=0$ (including the class (C)) or
$\alpha_{\rm B}=\alpha_{\rm M}$, $\mu$ is equivalent to $\Sigma$.
The case $\alpha_{\rm B}=\alpha_{\rm M}$ is special in that
Eqs.~(\ref{Gefff2})-(\ref{Sigmaf2}) reduce to
$\mu=\Sigma=1/(16\pi G\,G_4)$ and $\eta=1$.
\end{itemize}
In subsequent sections, we will discuss observational signatures
for concrete dark energy models which belong to the classes
(A), (B), (C), (D) in more detail.
\section{Class (A): Quintessence and k-essence}
\label{quinsec}
Quintessence and k-essence belong to the class (A)
given by the Lagrangian
\begin{equation}
L=G_2(\phi,X)+\frac{M_{\rm pl}^2}{2}R\,,
\end{equation}
which is within the framework of GR.
{}From Eq.~(\ref{wde}), the dark energy EOS
of k-essence yields
\begin{equation}
w_{\rm DE}=-\frac{G_2}{G_2-\dot{\phi}^2 G_{2,X}}\,.
\label{wdeke}
\end{equation}
The quantities $\alpha_{\rm K}$, $\alpha_{\rm B}$, and
$\alpha_{\rm M}$ are
\begin{equation}
\alpha_{\rm K}=\frac{\dot{\phi}^2}{H^2q_t}
\left( G_{2,X}+\dot{\phi}^2 G_{2,XX} \right)\,,
\qquad
\alpha_{\rm B}=0\,,\qquad
\alpha_{\rm M}=0\,,
\label{alM2}
\end{equation}
with $q_t=M_{\rm pl}^2>0$.
{}From Eqs.~(\ref{qscon}) and (\ref{csge}), we obtain
\begin{equation}
q_s=2M_{\rm pl}^2 \left( G_{2,X}+\dot{\phi}^2 G_{2,XX}
\right)\,,\qquad
c_s^2=\frac{G_{2,X}}{G_{2,X}+\dot{\phi}^2 G_{2,XX}}\,.
\label{qsku}
\end{equation}
The ghost and Laplacian instabilities of scalar perturbations
are absent for
\begin{equation}
G_{2,X}+\dot{\phi}^2\,G_{2,XX}>0\,, \quad
{\rm and} \quad
G_{2,X} > 0\,.
\label{kescon}
\end{equation}
The background Eqs.~(\ref{back1}), (\ref{dH}), and (\ref{ddphi})
reduce, respectively, to
\begin{eqnarray}
& &
3M_{\rm pl}^2 H^2=-G_2+\dot{\phi}^2 G_{2,X}+\rho_m\,,\label{bake1}\\
& &
2M_{\rm pl}^2 \dot{H}=-\dot{\phi}^2 G_{2,X}-\rho_m-P_m\,,\label{bake2}\\
& &
\ddot{\phi}=-\frac{3H\dot{\phi} G_{2,X}+\dot{\phi}^2 G_{2,X \phi}
-G_{2,\phi}}{G_{2,X}+\dot{\phi}^2 G_{2,XX}}\,.\label{bake3}
\end{eqnarray}
Provided that the field energy density
$\rho_{\rm DE}=-G_2+\dot{\phi}^2 G_{2,X}$ is
positive, the k-essence EOS (\ref{wdeke}) is in the range
\begin{equation}
w_{\rm DE}>-1\,,
\end{equation}
under the second condition of (\ref{kescon}).
The evolution of $w_{\rm DE}$ is different depending on
the potential of quintessence and the form of k-essence Lagrangian.
On using Eq.~(\ref{alM2}), the quantities (\ref{Gefff2})
and (\ref{Sigmaf2}) reduce, respectively, to
\begin{equation}
\mu=1\,,\qquad \Sigma=1\,,
\end{equation}
independent of the models of quintessence and k-essence.
Although this situation is degenerate, the evolution of
matter perturbations and gravitational potentials is affected
by the difference of the background cosmology \cite{TDA13}.
Moreover, as we will see
in the following, the scalar sound speed squared $c_s^2$
in k-essence differs from that in quintessence.
Hence there are still possibilities
for distinguishing between models of
quintessence and k-essence.
\subsection{Quintessence}
Quintessence is given by the function
\begin{equation}
G_2(\phi, X)=X-V(\phi)\,,
\end{equation}
under which Eq.~(\ref{wdeke}) reduces to
\begin{equation}
w_{\rm DE}=\frac{\dot{\phi}^2/2-V(\phi)}
{\dot{\phi}^2/2+V(\phi)}> -1\,.
\end{equation}
The cosmological constant $\Lambda$ corresponds to
the limit $\dot{\phi}^2/2 \to 0$ and $V(\phi) \to \Lambda$,
i.e., $w_{\rm DE} \to -1$.
For quintessence, the quantities given in Eq.~(\ref{qsku}) reduce to
$q_s=2M_{\rm pl}^2$ and $c_s^2=1$, so the conditions for the
absence of ghost and Laplacian instabilities are trivially satisfied.
To study the dark energy dynamics, we introduce the
following dimensionless quantities:
\begin{equation}
\Omega_{\rm DE} \equiv \frac{\rho_{\rm DE}}{3M_{\rm pl}^2 H^2}\,,\qquad
\Omega_{m} \equiv \frac{\rho_{m}}{3M_{\rm pl}^2 H^2}\,,\qquad
w_m \equiv \frac{P_m}{\rho_m}\,,
\end{equation}
where $\rho_{\rm DE}=\dot{\phi}^2/2+V(\phi)$.
For the matter sector, we consider nonrelativistic matter characterized
by the constant equation of state $w_m$ close to $+0$.
The Hamiltonian constraint (\ref{bake1}) gives the relation
$\Omega_{\rm DE}+\Omega_m=1$.
On using Eqs.~(\ref{bake2}) and (\ref{bake3}), we obtain the following
differential equations \cite{Sch,CDT}:
\begin{eqnarray}
w_{\rm DE}'
&=& (w_{\rm DE}-1) \sqrt{3(1+w_{\rm DE})}
\left[ \sqrt{3(1+w_{\rm DE})}
-\lambda \sqrt{\Omega_{\rm DE}}
\right]\,,
\label{quinw}\\
\Omega_{\rm DE}'
&=& 3(w_m-w_{\rm DE}) \Omega_{\rm DE}
\left( 1-\Omega_{\rm DE} \right)\,,
\label{quinOme} \\
\lambda'
&=& -\sqrt{3(1+w_{\rm DE}) \Omega_{\rm DE}}\,
(\Gamma-1) \lambda^2\,,
\label{quinlam}
\end{eqnarray}
where a prime represents a derivative with respect to
${\cal N}=\ln a$, and
\begin{equation}
\lambda \equiv -\frac{M_{\rm pl} V_{,\phi}}{V}\,, \qquad
\Gamma \equiv \frac{V V_{,\phi \phi}}{V_{,\phi}^2}\,.
\end{equation}
{}From Eq.~(\ref{quinw}), besides the trivial case
$w_{\rm DE} \simeq 1$, there are two important
situations in which
$w_{\rm DE}$ stays nearly constant\footnote{The case $w_{\rm DE} \simeq 1$ is irrelevant to the late-time cosmic acceleration, as Eq.~(\ref{quinOme})
shows that $\Omega_{\rm DE}$ decreases
for $0<\Omega_{\rm DE}<1$.}:
(1) $w_{\rm DE} \simeq -1$, and
(2) $\sqrt{3(1+w_{\rm DE})} \simeq \lambda \sqrt{\Omega_{\rm DE}}$.
In the following, we will discuss these two cases separately.
\subsubsection{Thawing quintessence}
\label{thawsec}
In case (1), $w_{\rm DE}$ is close to $-1$, but
the deviation of $w_{\rm DE}$ from $-1$ still occurs at late times.
{}From Eq.~(\ref{quinOme}), we find that
$\Omega_{\rm DE}$ increases according to
$\Omega_{\rm DE}' \simeq 3(1+w_m)\Omega_{\rm DE}
(1-\Omega_{\rm DE})>0$.
Eventually, the growth of $\Omega_{\rm DE}$ in Eq.~(\ref{quinw}) leads to
the variation of $w_{\rm DE}$, such that
$w_{\rm DE}' \simeq 2 \lambda \sqrt{3(1+w_{\rm DE}) \Omega_{\rm DE}}$.
This belongs to the class of {\it thawing quintessence} \cite{Caldwell} in which
the scalar field is nearly frozen by the Hubble friction in the early cosmological
epoch and it starts to evolve only recently.
The representative potential of this class is the one
arising from the pseudo-Nambu-Goldstone (PNGB)
boson \cite{Frieman}:
\begin{equation}
V(\phi)=\mu^4 \left[ 1+\cos \left(\frac{\phi}{f}
\right) \right]\,,
\label{pngbpo}
\end{equation}
where $\mu$ and $f$ are constants characterizing the
energy scale and the mass scale of spontaneous symmetry
breaking, respectively.
The axion can be the candidate for the ultra-light PNGB boson.
When a global $U(1)$ symmetry is spontaneously broken, the axion appears as an angular massless field $\phi$ with an expectation
value $\phi=f_s e^{i \phi/f_s}$ of a complex scalar
at a scale $f_s$ \cite{Quinn,Kim0,Shifman}.
If the field mass squared
$|m_{\phi}^2|=|V_{,\phi \phi}| \simeq \mu^4/f^2$ around
the potential maximum is smaller than $H^2$,
the field is stuck there with the dark energy EOS $w_{\rm DE} \simeq -1$.
After $H^2$ drops below $|m_{\phi}^2|$, $w_{\rm DE}$
starts to increase from $-1$.
For $|m_{\phi}^2|$ of order $H_0^2$
(i.e., $\mu \approx \sqrt{H_0 M_{\rm pl}} \approx 10^{-3}$~eV with the mass scale
$f \approx M_{\rm pl}$), the growth of
$w_{\rm DE}$ occurs at low redshifts.
The radiative correction, which is proportional to $\mu^4$,
does not give rise to explicit symmetry breaking terms.
Then, the small mass term $|m_{\phi}| \approx \mu^2/M_{\rm pl}
\approx 10^{-33}$~eV relevant to the late-time cosmic acceleration
can be protected against radiative corrections.
There are several interesting attempts for explaining the small
mass scale $\mu$ of order $10^{-3}$~eV
in the context of supersymmetric
theories \cite{Watari,Choi,Kim,Axiverse,Hall,Kim2}.
In case (a) of Fig.~\ref{fig1}, we plot one example for the evolution of
$w_{\rm DE}$ in terms of $z+1=1/a$, where $z$ is
the redshift (with today's value $z=0$).
In this case, $w_{\rm DE}$ starts to deviate from $-1$
around the redshift $z \lesssim 2$, with today's value
$w_{\rm DE}(z=0) \simeq -0.7$.
The likelihood analysis using CMB shift parameters measured
by WMAP7 \cite{WMAP7} combined with SN Ia and BAO data
placed the bound $w_{\rm DE}(z=0)<-0.695$ (95 \%\,CL) with
the quintessence prior $w_{\rm DE}>-1$ \cite{CDT}.
An updated data analysis based on the
Planck 2015+SN Ia+BAO data provided the bound
$w_{\rm DE}(z=0)<-0.473$ (95 \%\,CL) \cite{Ooba}.
This difference is mostly attributed to the fact that today's Hubble parameter
$H_0$ constrained from the Planck data \cite{Planck2015,Planck2018}
favors a lower value than that constrained from the WMAP7 data \cite{WMAP7}.
The bound of $H_0$ derived by the Planck team
has been in tension with direct measurements of
$H_0$ at low redshifts \cite{RiessH0}.
This is the main reason why the analysis of
Ref.~\cite{Ooba} provided a more conservative bound on
$w_{\rm DE}(z=0)$ than that obtained in Ref.~\cite{CDT}.
\begin{figure}[h]
\begin{center}
\includegraphics[height=3.4in,width=3.6in]{fig1.pdf}
\end{center}
\caption{\label{fig1}
Examples for the evolution of $w_{\rm DE}$ versus
$z+1$ in (a) thawing quintessence, (b) scaling freezing quintessence,
and (c) tracking freezing quintessence.
The present epoch is identified by the condition
$\Omega_{\rm DE}(z=0)=0.68$.}
\end{figure}
\subsubsection{Freezing quintessence}
In case (2), the field density parameter obeys the particular relation
\begin{equation}
\Omega_{\rm DE}=\frac{3(1+w_{\rm DE})}{\lambda^2}\,.
\label{Ometra}
\end{equation}
If $w_{\rm DE}=w_m$, then $\Omega_{\rm DE}=3(1+w_m)/\lambda^2=
{\rm constant}$ and hence $\lambda$ is constant.
This is known as a scaling solution \cite{CLW,Joyce}, along which the ratio
$\Omega_{\rm DE}/\Omega_{m}$ does not vary in time.
This can be realized by the exponential potential
$V(\phi)=V_0 e^{-\lambda \phi/M_{\rm pl}}$,
where $V_0$ is a constant. Since $\Gamma=1$ in this case,
Eq.~(\ref{quinlam}) is trivially satisfied.
The scaling solution has a nice feature in that the field
density is not much smaller than the background density
in the past, which can be compatibile with the energy scale
relevant to particle physics.
Hence it can alleviate the coincidence problem of
dark energy (see
Refs.~\cite{gcon2,Tsuji04,Quartin,Chiba14,Gomes,Amendola18,Fru18a,Fru18b}
for more general models
allowing for scaling solutions).
However, since $w_{\rm DE}=w_m \simeq 0$
for nonrelativistic matter, the scaling solution is not appropriate
to be used for the cosmic
acceleration. If the single exponential potential is
modified at late times to slow down the evolution
of $\phi$, it is possible to realize
a scaling matter-dominated epoch followed by the
accelerated expansion \cite{Nelson,Scap1,Scap2,Scap3,Scap4}.
We call this model {\it scaling freezing quintessence} \cite{CDT}.
The typical example of scaling freezing quintessence is
given by the potential \cite{Nelson}
\begin{equation}
V(\phi)=V_1 e^{-\lambda_1 \phi/M_{\rm pl}}
+V_2 e^{-\lambda_2 \phi/M_{\rm pl}}\,,
\label{scaling}
\end{equation}
where $\lambda_i$ and $V_i$ ($i=1, 2$) are constants.
We consider the case in which the slopes
$\lambda_1$ and $\lambda_2$ are in the ranges
$\lambda_1> {\cal O}(1)$ and $\lambda_2 <{\cal O}(1)$.
In the early cosmological epoch, the steep exponential
potential $V_1 e^{-\lambda_1 \phi/M_{\rm pl}}$
dominates over the other potential
$V_2 e^{-\lambda_2 \phi/M_{\rm pl}}$.
The solution enters the scaling radiation-dominated epoch
($w_{\rm DE}=1/3$ and $\Omega_{\rm DE}=4/\lambda_1^2$),
which is followed by the scaling matter era
($w_{\rm DE}=0$ and $\Omega_{\rm DE}=3/\lambda_1^2$).
{}From the big bang nucleosynthesis there is the bound
$\Omega_{\rm DE}<0.045$ in the radiation era \cite{Bean},
which translates to $\lambda_1>9.4$ \cite{Ohashi}.
The Universe finally enters the epoch of cosmic acceleration
after the second potential $V_2 e^{-\lambda_2 \phi/M_{\rm pl}}$
in Eq.~(\ref{scaling}) starts to contribute to the cosmological dynamics.
The fixed point relevant to the late-time cosmic acceleration
satisfies $\sqrt{3(1+w_{\rm DE})}=\lambda \sqrt{\Omega_{\rm DE}}$,
$\Omega_{\rm DE}=1$, $\Gamma=1$, with $\lambda=\lambda_2$
in Eqs.~(\ref{quinw})-(\ref{quinlam}), so that
\begin{equation}
w_{\rm DE}=-1+\frac{\lambda_2^2}{3}\,.
\label{wdefi}
\end{equation}
The necessary condition for the cosmic acceleration is given by
$\lambda_2^2<2$, under which the fixed point is a stable
attractor \cite{CLW,dreview4}.
In case (b) of Fig.~\ref{fig1}, we plot the evolution of
$w_{\rm DE}$ for $\lambda_1=50$ and $\lambda_2=0.1$ with
today's field density parameter $\Omega_{\rm DE}(z=0)=0.68$.
We observe that $w_{\rm DE}$ is close to 0 in the scaling matter era
and it approaches the asymptotic value (\ref{wdefi}).
In order to confront scaling freezing quintessence with observations,
it is convenient to quantify the evolution of $w_{\rm DE}$
in terms of the transition scale factor $a_t$ and the thickness of
transition to the freezing regime driven by the
potential $V_2 e^{-\lambda_2 \phi/M_{\rm pl}}$.
In Ref.~\cite{CDT}, it was shown that the change of $w_{\rm DE}$
from the scaling matter era to the accelerated attractor with
$\lambda_2^2 \ll 1$ can be well approximate by
the parametrization \cite{Linder05}:
\begin{equation}
w_{\rm DE}=-1+\frac{1}{1+(a/a_t)^{1/\tau}}\,,
\label{wdetran}
\end{equation}
with $\tau \simeq 0.33$.
The joint likelihood analysis based on the Planck 2015 data
combined with SN Ia and BAO data showed that the transition
scale factor $a_t$ is constrained to be $a_t<0.11$ (95\,\%\,CL) \cite{Ooba}
(which updated the bound $a_t<0.23$ (95\,\%\,CL) derived
in Ref.~\cite{CDT}).
This translates to the transition redshift $z_t>8.1$, so the scaling
matter era needs to end at quite early time.
This is mostly attributed to the fact that, since the sound
speed squared $c_s^2$ of quintessence is equivalent to 1,
the perturbation of scaling quintessence hardly contributes
to gravitational potentials. This slows down the growth of structures
and leads to a large early Integrated-Sachs-Wolfe (ISW)
effect \cite{Ooba}.
For decreasing $z_t$ in the range $z_t<8$, the CMB angular
power spectrum is subject to stronger modification.
There is the other class of freezing models dubbed
{\it tracking freezing quintessence}.
In this case, the slope $\lambda$ decreases in time with
$w_{\rm DE}$ nearly constant, so that $\Omega_{\rm DE}$
in Eq.~(\ref{Ometra}) grows in time.
{}From Eq.~(\ref{quinlam}), the condition for decreasing
$\lambda$ translates to
\begin{equation}
\Gamma>1\,.
\label{Gamcon}
\end{equation}
Since $\Omega_{\rm DE}$ increases, Eq.~(\ref{quinOme}) shows
that $w_{\rm DE}<w_m$ for this tracker solution.
This property is different from the scaling solution along which
$w_{\rm DE}=w_m$ and $\Omega_{\rm DE}={\rm constant}$.
Taking the ${\cal N}$ derivative of Eq.~(\ref{Ometra}),
it follows that $\Omega_{\rm DE}'/\Omega_{\rm DE}=-2\lambda'/\lambda$.
Substituting Eqs.~(\ref{quinOme}), (\ref{quinlam}),
and (\ref{Ometra}) into
this relation under the condition that $\Omega_{\rm DE} \ll 1$,
the dark energy EOS in the matter era
can be estimated as \cite{Paul98}
\begin{equation}
w_{\rm DE} \simeq w_{(0)} \equiv
-\frac{2(\Gamma-1)}{2\Gamma-1}\,,
\label{traeq}
\end{equation}
where we used $w_m \simeq 0$.
As long as $\Gamma$ evolves slowly, $w_{\rm DE}$ stays
nearly constant at high redshifts ($z \gg 1$).
The tracker solution can be realized by the inverse power-law
potential given by
\begin{equation}
V(\phi)=M^{4+p} \phi^{-p}\,,
\label{inpo}
\end{equation}
where $M$ and $p$ are positive constants.
This potential gives the value $\Gamma=1+1/p={\rm constant}>1$,
so it satisfies the tracking condition (\ref{Gamcon}).
The tracker EOS (\ref{traeq}) reduces to
\begin{equation}
w_{(0)}=-\frac{2}{p+2}\,.
\end{equation}
For $p$ closer to 0, $w_{(0)}$ approaches the cosmological
constant value $-1$.
As $\Omega_{\rm DE}$ grows to the order of 0.1,
$w_{\rm DE}$ starts to deviate from the value (\ref{traeq}), see
case (c) in Fig.~\ref{fig1}.
It is possible to estimate the deviation $\delta w_{\rm DE}$ from
$w_{(0)}$ by dealing with $\Omega_{\rm DE}$ as a perturbation to
the 0-th order solution (\ref{traeq}).
The leading-order solution to $\Omega_{\rm DE}$ can be derived by
substituting $w_{\rm DE} \simeq w_{(0)}$ into Eq.~(\ref{quinOme})
and integrating it with respect to $a$, such that
\begin{equation}
\Omega_{\rm DE}(a)=\Omega_{\rm DE}^{(0)}
\left[ \Omega_{\rm DE}^{(0)}+a^{3w_{(0)}}
\left( 1-\Omega_{\rm DE}^{(0)} \right)
\right]^{-1}\,,
\label{Omedetra}
\end{equation}
where $\Omega_{\rm DE}^{(0)}$ is today's value of $\Omega_{\rm DE}$.
Plugging Eq.~(\ref{Omedetra}) into Eq.~(\ref{quinw})
and integrating the linear perturbation equation
of $\delta w_{\rm DE}$ with
respect to $a$, it follows that \cite{Chiba10}
\begin{equation}
w_{\rm DE}(a)=w_{(0)}+\sum_{n=1}^{\infty}
\frac{(-1)^{n-1}w_{(0)}(1-w_{(0)}^2)}
{1-(n+1)w_{(0)}+2n(n+1)w_{(0)}^2}
\left( \frac{\Omega_{\rm DE}(a)}{1-\Omega_{\rm DE}(a)}
\right)^n\,.
\label{wtracker}
\end{equation}
As shown in Refs.~\cite{Chiba10,CDT}, the iterative formula (\ref{wtracker})
up to the order $n=3$ exhibits fairly good agreement with the numerically
integrated solution.
Under the quintessence prior $w_{(0)}>-1$,
the likelihood analysis using the iterative solution (\ref{wtracker})
put the bounds $w_{(0)}<-0.923$
and $0.675<\Omega_{\rm DE}^{(0)}<0.703$ (95 \%\,CL)
from the observational data
of Planck 2015, SN Ia, and CMB \cite{Ooba}.
For the potential (\ref{inpo}), this bound translates to $p<0.17$.
For example, the tracker solution (c) plotted in Fig.~\ref{fig1} corresponds
to $p=1$, which is strongly disfavored by the data.
Thus, the tracker solutions arising from the inverse power-law potential with
positive integer powers ($p \geq 1$) are observationally excluded.
\subsection{k-essence}
The Lagrangian of k-essence is a general function of $\phi$ and $X$.
For example, the tachyon field given by the Lagrangian
$G_2=-V(\phi) \sqrt{-{\rm det} \left(g_{\mu \nu}+
\nabla_{\mu} \phi \nabla_{\nu} \phi \right)}$, where
$V(\phi)$ is a tachyon potential, arises in open string theory
living on a non-BPS D3-brane \cite{tac1,tac2,tac3}.
There is also the so-called Dirac-Born-Infeld (DBI) scenario in which
the movement of a probe D3-brane along the radial direction in
the AdS$_5$ spacetime is described by
the action $P=-f(\phi)^{-1} \sqrt{1-2f(\phi)X}
+f(\phi)^{-1}-V(\phi)$ \cite{DBI1,DBI2}.
In the original setup of string theory, however, these two
scenarios do not account for the late-time cosmic acceleration.
Allowing for the freedom to modify the functions $V(\phi)$ and $f(\phi)$,
there are possibilities for applying such theories to dark
energy \cite{Paddy2002,Abramo,Lazkoz,CGST,Martin,Guo}.
In both tachyon and DBI theories, we require the existence of scalar potential $V(\phi)$
for driving the cosmic acceleration (as in quintessence).
There are k-essence scenarios in which nonlinear terms in $X$ play crucial roles
for the late-time cosmological dynamics \cite{kes1,kes2,kes3,gcon1,gcon2,unifiedkes}.
In the following, we review two typical k-essence theories of this type
and discuss their observational signatures.
\subsubsection{Ghost condensate}
If the scalar field has a negative kinetic energy $-X$,
this is typically a sign for the appearance of scalar ghosts.
However, the existence of an additional term $X^2$ to $-X$ can allow a possibility for avoiding the ghost \cite{gcon1}.
The so-called dilatonic ghost condensate \cite{gcon2}
given by the Lagrangian
\begin{equation}
G_2(\phi, X)=-X+e^{\lambda \phi/M_{\rm pl}}
\frac{X^2}{M^4}
\label{Lagghost}
\end{equation}
belongs to such a class (where $\lambda$ and $M$
are positive constants).
In the limit $\lambda \to 0$, there exists a de Sitter solution
at $X=M^4/2$. As we will see below, the exponential term
$e^{\lambda \phi/M_{\rm pl}}$ in Eq.~(\ref{Lagghost}) leads to
the deviation from the de Sitter solution.
To discuss the cosmological dynamics of dilatonic ghost condensate,
we introduce the dimensionless quantities:
\begin{equation}
x_1 \equiv \frac{\dot{\phi}}{\sqrt{6}H M_{\rm pl}}\,,\qquad
x_2 \equiv \frac{\dot{\phi}^2 e^{\lambda \phi/M_{\rm pl}}}{2M^4}\,.
\end{equation}
From Eq.~(\ref{qsku}), the quantities $q_s$
and $c_s^2$ are
\begin{equation}
q_s=2 \left( 6x_2-1 \right) M_{\rm pl}^2\,,\qquad
c_s^2=\frac{2x_2-1}{6x_2-1}\,,
\end{equation}
which are both positive for $x_2>1/2$.
The dark energy density parameter and its EOS
are given, respectively, by
\begin{equation}
\Omega_{\rm DE}=x_1^2 \left( -1+3x_2 \right)\,,\qquad
w_{\rm DE}=\frac{1-x_2}{1-3x_2}\,.
\end{equation}
The necessary condition for the cosmic acceleration
($-1<w_{\rm DE}<-1/3$) translates to $1/2<x_2<2/3$,
in which regime $0<c_s^2<1/9$ \cite{gcon2}.
On using the background Eqs.~(\ref{dH}) and (\ref{ddphi}),
we obtain the differential equations for
$w_{\rm DE}$ and $\Omega_{\rm DE}$, as
\begin{eqnarray}
w_{\rm DE}' &=&
\frac{(1-3w_{\rm DE})(1-w_{\rm DE})
[\sqrt{3(1-3w_{\rm DE})\Omega_{\rm DE}}\,\lambda-3(1+w_{\rm DE})]}{5-3w_{\rm DE}}\,,
\label{wdeke2}\\
\Omega_{\rm DE}' &=&
3(w_m-w_{\rm DE}) \Omega_{\rm DE}
\left( 1-\Omega_{\rm DE} \right)\,,\label{Omedeke}
\end{eqnarray}
where Eq.~(\ref{Omedeke}) is the same as Eq.~(\ref{quinOme})
derived for quintessence.
{}From Eq.~(\ref{wdeke2}), there are cases in which
$w_{\rm DE}$ stays nearly constant.
Among them, the fixed point corresponding
to $w_{\rm DE}<-1/3$ is characterized by
\begin{equation}
w_{\rm DE}=-1+\sqrt{\left( \frac{\lambda^2}{2}\Omega_{\rm DE} \right)^2
+\frac43 \lambda^2 \Omega_{\rm DE}}-\frac{\lambda^2}{2}\Omega_{\rm DE}\,.
\label{wdekes1}
\end{equation}
{}From Eq.~(\ref{Omedeke}), we have three fixed points satisfying $w_{\rm DE}=w_m$,
$\Omega_{\rm DE}=0$, and $\Omega_{\rm DE}=1$.
The first case corresponds to the scaling solution,
but this is not responsible for dark energy unless the form
of $G_2(\phi,X)$ is appropriately modified at late times.
The second fixed point ($\Omega_{\rm DE}=0$) is relevant
to the early matter era, during which $w_{\rm DE}$
is close to $-1$.
The third fixed point ($\Omega_{\rm DE}=1$)
is associated with
the late-time cosmic acceleration with
$w_{\rm DE}=-1+\sqrt{\lambda^4/4+4\lambda^2/3}-\lambda^2/2>-1$.
Then, the evolution of $w_{\rm DE}$ in the
diatonic ghost condensate is similar to that
in thawing quintessence discussed in Sec.~\ref{thawsec}.
Provided that $w_{\rm DE}$ does not significantly deviate from $-1$,
the integrated solution to Eq.~(\ref{Omedeke}) in the intermediate
regime $0<\Omega_{\rm DE}<1$ can be derived
by setting $w_{m} \simeq 0$ and $w_{\rm DE} \simeq -1$,
such that
\begin{equation}
\Omega_{\rm DE}(a)=a^3 \Omega_{\rm DE}^{(0)}
\left[ 1- \Omega_{\rm DE}^{(0)}+a^3 \Omega_{\rm DE}^{(0)}
\right]^{-1}\,.
\label{Omedekes1}
\end{equation}
Substituting Eq.~(\ref{Omedekes1}) into Eq.~(\ref{wdekes1}),
we obtain the approximate solution to $w_{\rm DE}$
as a function of $a$. If we adopt the observational bound
$w_{\rm DE}(a=1) \lesssim -0.7$ with $\Omega_{\rm DE}^{(0)}
=0.68$, the parameter $\lambda$ is constrained to be
$\lambda \lesssim 0.36$.
We note that, in the dilatonic ghost condensate, $c_s^2$ is
initially close to $+0$ and then it starts to deviate from $+0$
only recently. In this case, the k-essence scalar with $c_s^2 \simeq +0$
can work as a part of dark matter because of additional gravitational clusterings.
This property is different from that in quintessence where
$c_s^2$ is always equivalent to 1, so there is a possibility
for distinguishing between thawing quintessence and
diatonic ghost condensate at the level of perturbations.
\subsubsection{K-essence as unified dark energy and dark matter}
There is a unified k-essence model of
dark energy and dark matter given by \cite{unifiedkes}
\begin{equation}
G_2(X)=-b_0+b_2 \left( X-X_0 \right)^2\,,
\label{unilag}
\end{equation}
where $b_0, b_2, X_0$ are positive constants.
In this case, the k-essence pressure and density are
$P=-b_0+b_2 \left( X-X_0 \right)^2$ and
$\rho=b_0+b_2(X-X_0)(3X+X_0)$, respectively, with
the propagation speed squared:
\begin{equation}
c_s^2=\frac{X-X_0}{3X-X_0}\,.
\label{csuni}
\end{equation}
As long as $X$ stays around $X_0$, $c_s^2$ is close to 0.
Substituting $X=X_0 \left[ 1+ \epsilon(t) \right]$ with
$0<\epsilon \ll 1$ into Eq.~(\ref{bake3}), we obtain the
integrated solution $\epsilon(t)=\epsilon_1 (a_1/a)^3$,
where $\epsilon_1$ and $a_1$ are positive constants.
Then, the sound speed squared (\ref{csuni}) approximately
reduces to
\begin{equation}
c_s^2 \simeq \frac{\epsilon(t)}{2}
=\frac{\epsilon_1}{2} \left( \frac{a_1}{a} \right)^3\,,
\end{equation}
which always stays in the region $0<c_s^2 \ll 1$.
Around $X=X_0$, we have $P \simeq -b_0$ and
$\rho \simeq b_0+4b_2 X_0 (X-X_0)$, so the k-essence
EOS, $w=P/\rho$, reads
\begin{equation}
w \simeq -\left[ 1+\frac{4b_2}{b_0}X_0^2 \epsilon_1
\left( \frac{a_1}{a} \right)^3 \right]^{-1}\,.
\end{equation}
In the early matter era, the k-essence field behaves as dark matter
with $w \simeq 0$. Since $w$ approaches $-1$
at late times due to the dominance of the term $-b_0$
in Eq.~(\ref{unilag}), the same field behaves as dark energy.
Thus, the k-essence Lagrangian (\ref{unilag}) provides the
unified description of dark energy and dark matter
with $0<c_s^2 \ll 1$. We note that the purely k-essence
with the Lagrangian $G_2=G_2(X)$ is equivalent to a barotropic
perfect fluid on the cosmological background \cite{Hu05,Arroja,DMT10,KT14b}.
In Refs.~\cite{Ber1,Ber2}, the authors proposed several unified models
of dark matter and dark energy different from (\ref{unilag}).
In particular, there is a model in which $c_s^2$ starts to evolve from a
value close to 0 and approaches an asymptotic value $c_{\infty}^2$.
In such a case, the likelihood analysis based on the galaxy-ISW correlation
data showed that $c_{\infty}^2$ is constrained to be smaller than
$c_{\infty}^2 \lesssim 9 \times 10^{-3}$ \cite{Ber3}.
\section{Class (B): $f(R)$ gravity and Brans-Dicke theories}
\label{fRBDsec}
The theories of class (B) include BD theories and
$f(R)$ gravity as specific cases.
The Lagrangian of BD theories with a scalar
potential $V(\phi)$ is given by
\begin{equation}
L=\left(1-6Q^2 \right)e^{-2Q\phi/M_{\rm pl}}X
-V(\phi)+\frac{M_{\rm pl}^2}{2}
e^{-2Q\phi/M_{\rm pl}}R\,,
\label{lagbra}
\end{equation}
where $Q$ is a constant.
In terms of the redefined dimensionless field
\begin{equation}
\chi=e^{-2Q\phi/M_{\rm pl}}\,,
\end{equation}
the Lagrangian (\ref{lagbra}) is equivalent to
\begin{equation}
L=-\frac{M_{\rm pl}^2 \omega_{\rm BD}}{2\chi}
g^{\mu \nu} \nabla_{\mu}\chi \nabla_{\nu}\chi
-V(\phi(\chi))+\frac{M_{\rm pl}^2}{2} \chi R\,,
\label{lagbra2}
\end{equation}
where the BD parameter $\omega_{\rm BD}$
is related to the constant $Q$, as
\begin{equation}
3+2\omega_{\rm BD}=\frac{1}{2Q^2}\,.
\end{equation}
The original BD theory \cite{Brans} is written in the form
(\ref{lagbra2}) without the scalar potential ($V=0$).
The $f(R)$ gravity, which is given by the Lagrangian
$L=M_{\rm pl}^2 f(R)/2$, is equivalent to the scalar-tensor
theory with $L=M_{\rm pl}^2
[f(\varphi)+f_{,\varphi}(R-\varphi)]/2$,
where $\varphi$ is a scalar quantity.
Varying the latter Lagrangian with respect to $\varphi$,
we obtain $f_{,\varphi \varphi}(R-\varphi)=0$ and hence
$\varphi=R$ for $f_{,\varphi \varphi} \neq 0$.
By defining $\chi \equiv f_{,\varphi}=f_{,R}$, the Lagrangian
of $f(R)$ gravity reduces to
\begin{equation}
L=\frac{M_{\rm pl}^2}{2} \chi R-V(\chi)\,,\quad {\rm where}
\quad V(\chi)=\frac{M_{\rm pl}^2}{2} \left[ \chi \varphi (\chi)
-f(\varphi (\chi)) \right]\,.
\label{LfR}
\end{equation}
Comparing this Lagrangian with Eq.~(\ref{lagbra2}),
it follows that $f(R)$ gravity corresponds to BD theory
with $\omega_{\rm BD}=0$ \cite{Ohanlon,Chiba03}.
Under the conformal transformation
$\hat{g}_{\mu \nu}=\chi g_{\mu \nu}$, the Lagrangian
(\ref{LfR}) is transformed to that
in the Einstein frame with a canonical
scalar field \cite{Maeda89,Bunn,fRreview0,fRreview}
\begin{equation}
\phi=\sqrt{\frac{3}{2}}
M_{\rm pl} \ln f_{,R}\,,
\label{phifR}
\end{equation}
which corresponds
to the field $\phi$ appearing in the Jordan-frame Lagrangian
(\ref{lagbra}) of BD theory.
On using the correspondence $\chi=e^{-2Q \phi/M_{\rm pl}}=f_{,R}$,
the constant $Q$ in metric $f(R)$ gravity,
which plays the role of couplings between the field
$\phi$ and matter in the Einstein frame \cite{AmenPRL,BD1},
is given by
\begin{equation}
Q=-\frac{1}{\sqrt{6}}\,.
\end{equation}
In BD theory with an arbitrary BD parameter
$\omega_{\rm BD}$, the modification of gravity
is significant for $|Q|>{\cal O}(0.1)$.
In the limit that $Q \to 0$, i.e., $\omega_{\rm BD} \to \infty$,
the Lagrangian (\ref{lagbra}) recovers the minimally coupled quintessence in GR.
If the scalar potential $V(\phi)$ is absent, gravitational experiments in the solar system
constrain the BD parameter to be $\omega_{\rm BD}>40000$ \cite{Will},
which translates to $|Q|<2.5 \times 10^{-3}$. In the presence of $V(\phi)$, as in the case
of $f(R)$ gravity, it is possible to suppress the fifth force
induced by the scalar-matter interaction in the solar
system under the chameleon mechanism even for
$|Q|>{\cal O}(0.1)$ \cite{Bunn,Navarro,fR1,Capo07,Tamaki,Brax08}.
Indeed, the viable dark energy models in $f(R)$
gravity \cite{fR1,fR2,fR3,fR4,fR5}
and BD theory \cite{BD1,BD2} are constructed to have a large mass in regions of the high density, while the field mass is light on cosmological scales such that the effective potential of $\phi$ drives the late-time cosmic acceleration.
For the Lagrangian (\ref{lagbra}), the background Eqs.~(\ref{back1}) and (\ref{ddphi}) reduce,
respectively, to
\begin{eqnarray}
& &
3M_{\rm pl}^2 H^2=
\frac{1}{2}(1-6Q^2) \dot{\phi}^2
+6Q M_{\rm pl}H \dot{\phi}+\left( \rho_m+V
\right)e^{2Q\phi/M_{\rm pl}}\,,\label{BDeq1}\\
& &
\ddot{\phi}+3H \dot{\phi}+V_{,\phi}e^{2Q\phi/M_{\rm pl}}
+\frac{Q}{M_{\rm pl}} \left[ \left( \rho_m-3P_m+4V
\right)e^{2Q\phi/M_{\rm pl}}-2\dot{\phi}^2
\right]=0\,,\label{BDeq2}
\end{eqnarray}
where, for the derivation of Eq.~(\ref{BDeq2}), we used
Eq.~(\ref{BDeq1}) to eliminate $H^2$.
{}From Eq.~(\ref{BDeq2}), the coupling $Q$ can induce
a local minimum even for a runaway potential $V(\phi)$.
If the scalar field is nearly frozen around the local minimum
during the matter era characterized by
$\rho_m \gg \{P_m, V \}$, it follows that
\begin{equation}
M_{\rm pl}V_{,\phi}+Q \rho_m \simeq 0\,.
\label{mire1}
\end{equation}
If $Q>0$ (or $Q<0$), then the solution to Eq.~(\ref{mire1})
exists for $V_{,\phi}<0$ (or $V_{,\phi}>0$).
Since $\rho_m$ decreases in time, the field evolves slowly
along the instantaneous minima determined
by Eq.~(\ref{mire1}).
The scalar potential $V$ is responsible for the late-time cosmic
acceleration. After $V$ dominates over the matter density
$\rho_m$, there exists a de Sitter solution characterized by
\begin{equation}
M_{\rm pl}V_{,\phi}+4QV=0\,.
\label{mire2}
\end{equation}
{}From the above discussion, the field $\phi$ is almost
frozen apart from the transient period from the matter era
to the de Sitter solution. In other words, the deviation of
$w_{\rm DE}$ from $-1$ starts to occur around today and
$w_{\rm DE}$ finally approaches the asymptotic value $-1$.
We note that, with decreasing $\rho_m$, the effective mass
around the potential minimum induced by the coupling $Q$
gets smaller. In the early Universe, the field is nearly frozen
due to the heavy mass associated with the large matter density
$\rho_m$. In local regions of the Universe where the density
$\rho_m$ is much larger than today's critical density $\rho_c$,
the similar suppression of the propagation of fifth forces can
occur under the operation of
the chameleon mechanism \cite{chame1,chame2}.
We also compute quantities relevant to the evolution
of linear cosmological perturbations for the Lagrangian (\ref{lagbra}).
Since $q_t=M_{\rm pl}^2 e^{-2Q\phi/M_{\rm pl}}$,
the no-ghost condition of tensor perturbations
is automatically satisfied. The quantities
$\alpha_{\rm K}$, $\alpha_{\rm B}$, and
$\alpha_{\rm M}$ are given by
\begin{equation}
\alpha_{\rm K}=
\frac{(1 -6Q^2) \dot{\phi}^2}{M_{\rm pl}^2 H^2}\,,\qquad
\alpha_{\rm B}=\frac{\alpha_{\rm M}}{2}
=-\frac{Q\dot{\phi}}{M_{\rm pl}H}\,.
\label{alBB}
\end{equation}
{}From Eqs.~(\ref{qscon}) and (\ref{csge}), it follows that
\begin{equation}
q_s=2e^{-4Q\phi/M_{\rm pl}}M_{\rm pl}^2 \,,\qquad
c_s^2=1\,,
\end{equation}
so there are neither ghost nor Laplacian instabilities
of scalar perturbations in BD theories.
The dark energy EOS (\ref{wde})
reduces to
\begin{equation}
w_{\rm DE}=-1-\frac{4M_{\rm pl}^2 (1-e^{-2Q \phi/M_{\rm pl}}) \dot{H}
+4M_{\rm pl}Q e^{-2Q \phi/M_{\rm pl}} \ddot{\phi}
-2e^{-2Q \phi/M_{\rm pl}}\dot{\phi}[(1-2Q^2) \dot{\phi}+2M_{\rm pl}H Q]}
{(1-6Q^2)e^{-2Q \phi/M_{\rm pl}} \dot{\phi}^2+2V
+6M_{\rm pl}^2H^2(1-e^{-2Q \phi/M_{\rm pl}})
+12M_{\rm pl}H Qe^{-2Q \phi/M_{\rm pl}} \dot{\phi}}\,.
\label{wdeBD}
\end{equation}
Existence of the nonvanishing coupling $Q$ allows
the possibility for realizing
$w_{\rm DE}<-1$ \cite{fR1,fR4,Moto}, without
having ghost and Laplacian instabilities.
Applying the bound (\ref{alpM}) to $\alpha_{\rm M}$ in
Eq.~(\ref{alBB}), the field time derivative
today (denoted as $\dot{\phi}_0$) is constrained to be
\begin{equation}
\left| \frac{Q \dot{\phi}_0}{M_{\rm pl}H_0}
\right| <0.01\,,
\label{Qbound}
\end{equation}
which limits the large deviation of $w_{\rm DE}$ from
$-1$ at low redshifts.
{}From Eqs.~(\ref{Gefff2}) and (\ref{Sigmaf2}),
we obtain
\begin{equation}
\mu=e^{2Q\phi/M_{\rm pl}} \frac{1+2Q^2+{\cal F}_M (k)}
{1+{\cal F}_M(k)}\,,
\qquad
\Sigma=e^{2Q\phi/M_{\rm pl}}\,,
\label{SigmaBD}
\end{equation}
where
\begin{equation}
{\cal F}_M (k) \equiv \frac{a^2M_{\phi}^2
e^{2Q \phi/M_{\rm pl}}}{k^2}\,.
\label{calFM}
\end{equation}
As we will see below, the viable dark energy models
in $f(R)$ gravity and BD theory have been
constructed to have a growing mass $M_{\phi}$
in the asymptotic past with the scalar field nearly frozen around
$\phi \simeq 0$. Since $M_{\phi}^2$ becomes much
larger than $H^2$, the quantity ${\cal F}_M(k)$
can exceed the order 1 even for perturbations deep
inside the Hubble radius ($k^2 \gg a^2 H^2$).
Taking the limit ${\cal F}_M(k) \to \infty$
in Eq.~(\ref{SigmaBD}) in this massive regime,
it follows that $\mu \simeq \Sigma=
e^{2Q\phi/M_{\rm pl}} \simeq 1$.
Then, the evolution of linear perturbations is similar to that
in GR for the same background expansion history.
Since $M_{\phi}^2$ gradually decreases in time,
there is an instant at which ${\cal F}_M(k)$ crosses 1.
For larger $k$,
the entry to the regime ${\cal F}_M(k)<1$ occurs earlier.
Taking the limit ${\cal F}_M(k) \to 0$ in Eq.~(\ref{SigmaBD}), we obtain
\begin{equation}
\mu \simeq \left( 1+2Q^2 \right)\Sigma\,,\qquad
\Sigma=e^{2Q\phi/M_{\rm pl}}\,.
\label{muBD}
\end{equation}
This is the regime in which the modification of gravity
manifests itself on scales relevant to the linear growth of
large-scale structures.
In particular the coupling $Q$ leads to the enhancement
of $G_{\rm eff}$, so the growth rate of matter perturbations
gets larger than that in quintessence and k-essence.
This allows the possibility for distinguishing $f(R)$ gravity
and BD theories from dark energy models
in the framework of GR.
\subsection{$f(R)$ gravity}
\label{fRsec}
In $f(R)$ gravity, the examples of models relevant to
the late-time cosmic acceleration are
given by \cite{fR1,fR2}
\begin{eqnarray}
& & {\rm (i)}~f(R)=R-\lambda R_{0}\frac{(R/R_{0})^{2n}}{(R/R_{0})^{2n}+1}\,,\label{imodel}\\
& & {\rm (ii)}~f(R)=R-\lambda R_{0}\left[1-\left(1+\frac{R^{2}}{R_{0}^2}\right)^{-n}
\right]\,,\label{iimodel}
\end{eqnarray}
where $\lambda$, $R_0$, and $n$ are positive constants.
These models satisfy $f(R=0)=0$, so
the cosmological constant disappears in the limit
of flat spacetime.
In the high-curvature regime satisfying $R \gg R_0$,
they have the asymptotic behavior
\begin{equation}
f(R) \simeq R-\lambda R_0
\left[ 1-\left( \frac{R_0}{R} \right)^{2n}
\right]\,,
\label{fRasy}
\end{equation}
so that they approach the $\Lambda$CDM model
($f(R)=R-\lambda R_0$).
We note that the model $f(R)=R-\lambda R_0 \tanh (R/R_0)$ \cite{fR4}
also has the same asymptotic behavior.
The deviation from GR manifests itself after $R$
decreases to the order of $R_0$.
On using the asymptotic form (\ref{fRasy}), the scalar field $\phi$
defined by Eq.~(\ref{phifR}), which is called the scalaron \cite{Staro},
is expressed as
\begin{equation}
\phi \simeq \sqrt{\frac32}M_{\rm pl}\,\ln \left[1-2n \lambda
\left( \frac{R}{R_0} \right)^{-(2n+1)} \right]<0\,,
\label{phifRre}
\end{equation}
which approaches 0 as $R \to \infty$.
In the regime $R \gg R_0$,
the scalar potential (\ref{LfR}), i.e., $V=
(M_{\rm pl}^2/2)(f_{,R}R-f)$, reduces to
\begin{equation}
V(\phi) \simeq \frac{\lambda R_0M_{\rm pl}^2}{2} \left[ 1-\frac{2n+1}
{(2n\lambda)^{2n/(2n+1)}} \left( 1-e^{\sqrt{2/3}\phi/M_{\rm pl}}
\right)^{2n/(2n+1)} \right]\,,
\label{Vphiap}
\end{equation}
which approaches the constant $\lambda R_0M_{\rm pl}^2/2$
for $R \to \infty$.
Since the potential (\ref{Vphiap}) satisfies $V_{,\phi}>0$ for $\phi<0$,
the solutions to Eqs.~(\ref{mire1}) and (\ref{mire2}) exist for the
above $f(R)$ models.
The field mass squared $M_{\phi}^2=V_{,\phi \phi}$ increases
for larger $R$, with the divergence
$M_{\phi}^2 \to \infty$ for $R \to \infty$.
In terms of $f(R)$, the mass squared
can be expressed as
\begin{equation}
M_{\phi}^2=\frac{Rf_{,R}}{3m} \left( 1+m \right)\,,
\label{Mphi}
\end{equation}
where the quantity $m \equiv Rf_{,RR}/f_{,R}$
characterizes the deviation from the $\Lambda$CDM
model \cite{fRsta3}.
The increase of $M_{\phi}^2$ in the asymptotic past is
attributed to the decrease of
$f_{,RR}=2\lambda n (2n+1)R_0^{-1}(R/R_0)^{-2n-2}$
toward 0. Since $f_{,RR}$ is positive for
the models (\ref{imodel}) and (\ref{iimodel}),
the mass squared (\ref{Mphi}) is positive and hence
there is no tachyonic instability.
Unlike the $\Lambda$CDM model in which $f_{,RR}$ is exactly 0,
the above $f(R)$ models give rise to a large mass
in the asymptotic past due to the small deviation of
$f_{,RR}$ from $0$.
In the regime $m \ll 1$ the mass squared (\ref{Mphi})
can be estimated as $M_{\phi}^2 \simeq
Rf_{,R}/(3m) \simeq H^2/m \gg H^2$, where
we used the properties $R=6(2H^2+\dot{H}) \simeq 3H^2$
and $f_{, R} \simeq 1$ in the matter era.
Unless the initial conditions of $\phi$ are carefully
chosen to be close to
the instantaneous minimum given by Eq.~(\ref{mire1}),
the scalar field
oscillates due to the heavy mass in the early
cosmological epoch \cite{fR2,fR4}.
Then, the system can even access the curvature singularity
at $\phi=0$ \cite{Frolov}.
This is the general problem of late-time $f(R)$
cosmic acceleration models. For relativistic stars
in $f(R)$ gravity, the similar problem also arises as
a fine-tuning of boundary
conditions \cite{MaedafR,Tsuji09,Babi09,Upadhye}.
Provided that the initial conditions are fine-tuned such that
the scalar field is very close to the position determined
by Eq.~(\ref{mire1}), the field evolves slowly along the
instantaneous minima with decreasing $\rho_m$.
In the regime $R \gg R_0$, the field stays in the region around $\phi=0$
and hence the potential $V(\phi)$ is nearly constant:
$V(\phi) \simeq \lambda R_0 M_{\rm pl}^2/2$.
In this region, $w_{\rm DE}$ is close to $-1$.
After $R$ decreases to the order $R_0$, the field $\phi$ tends to
be away from 0, so the deviation of $w_{\rm DE}$ from $-1$
starts to occur by today \cite{Moto}.
Finally, the solution reaches the de Sitter
fixed point satisfying Eq.~(\ref{mire2}).
In $f(R)$ gravity, the condition (\ref{mire2}) is equivalent to \cite{fRsta3,fRreview}
\begin{equation}
Rf_{,R}=2f\,.
\label{fRcon1}
\end{equation}
The stability analysis around this fixed point shows that the
de Sitter point satisfying (\ref{fRcon1}) is stable
if the parameter $m=Rf_{,RR}/f_{,R}$ is in
the range \cite{fRsta1,fRsta2,fRsta3}
\begin{equation}
0<m \le 1\,.
\label{fRcon2}
\end{equation}
In model (i), the conditions (\ref{fRcon1}) and (\ref{fRcon2})
translate, respectively, to
\begin{eqnarray}
& &
\lambda x_1^{2n-1}(2+2x_1^{2n}-2n)=(1+x_1^{2n})^2\,,
\label{imodellam}\\
& &
2x_1^{4n}-2(n+2)(2n-1)x_1^{2n}+
(2n-1)(2n-2) \ge 0\,,
\label{imodelcon}
\end{eqnarray}
where $x_1 \equiv R_1/R_0$, and $R_1$ is the value of $R$
at the de Sitter point.
For given $n$, the condition (\ref{imodelcon})
gives a lower bound on the parameter $\lambda$.
If $n=1$, we have $x_1 \geq \sqrt{3}$ and
$\lambda \geq 8\sqrt{3}/9$.
The de Sitter fixed point corresponds to a stable spiral for
$0<m<16/25$ \cite{fRsta3}, in which case the dark energy
EOS approaches the asymptotic value $-1$ with oscillations
around $w_{\rm DE}=-1$.
The phantom EOS $w_{\rm DE}<-1$ can be
realized by today without having ghost
instabilities \cite{fR1,fR4,AT08,Moto}.
This property is different from that in quintessence and
k-essence where $w_{\rm DE}$ is always in the range
$w_{\rm DE}>-1$. We note that, under the bound (\ref{Qbound}),
the variation of $w_{\rm DE}$ given by Eq.~(\ref{wdeBD})
is limited around today, such that $|w_{\rm DE}+1|<{\cal O}(0.01)$.
At the background level, the $f(R)$ models (\ref{imodel}) and (\ref{iimodel}) are
consistent with the observational data \cite{fRob1,fRob2,fRob3}.
For $m \ll 1$, the mass squared (\ref{Mphi}) in the matter era
can be estimated as $M_{\phi}^2 \simeq H^2/m$, so the function (\ref{calFM}) is approximately given by
${\cal F}_M (k) \simeq a^2 H^2/(mk^2)$.
This means that the entry from the massive
regime (${\cal F}_M (k)>1$) to the massless regime
(${\cal F}_M (k)<1$) occurs at
\begin{equation}
m \simeq \left( \frac{aH}{k} \right)^2\,.
\label{mre}
\end{equation}
For the models (\ref{imodel}) and (\ref{iimodel}) the quantity
$m$ grows in time during the matter era, while
$(aH/k)^2$ decreases. Extrapolating the relation (\ref{mre})
up to the present epoch ($z=0$), the perturbations relevant to
the linear growth of large-scale structures ($k \lesssim 300 H_0$)
enter the massless regime by today
under the criterion
\begin{equation}
m_0 \gtrsim 10^{-5}\,,
\end{equation}
where $m_0$ is today's value of $m$.
Since $\mu$ and $\Sigma$ are close to 1 in the massive
regime (${\cal F}_M (k)>1$), the evolution of
perturbations is similar
to that in GR, i.e., $\delta_m \propto t^{2/3}$ and
$\psi_{\rm eff}=-2\Psi=2\Phi={\rm constant}$
during the matter dominance.
After the entry to the regime ${\cal F}_M (k)<1$,
the quantities $\mu$ and $\Sigma$ are
given by Eq.~(\ref{muBD}) with $Q=-1/\sqrt{6}$.
Solving Eq.~(\ref{delmeqf}) with Eqs.~(\ref{Poi}) and (\ref{psieff}),
the perturbations evolve as \cite{fR2,fR4}
\begin{equation}
\delta_m \propto t^{(\sqrt{33}-1)/6}\,,\qquad
\psi_{\rm eff}=-\frac{3}{2}\Psi=3\Phi \propto
t^{(\sqrt{33}-5)/6}\,,
\label{delmfR}
\end{equation}
during the matter dominance.
The growth rates of matter perturbations and
gravitational potentials are enhanced compared
to those in GR.
For $k \gtrsim 300 H_0$, the nonlinear effects on the evolution
of linear perturbations can be important, see Refs.~\cite{non1,non2,non3,non4,non5,non6,non7}
for $N$-body simulations in $f(R)$ gravity.
The modified growth of perturbations affects the observables
associated with the power spectra of galaxy clusterings, CMB, and
weak lensing.
In the left panel of Fig.~\ref{fig2}, we show the CMB angular
temperature power spectrum for three different $f(R)$ models
and the $\Lambda$CDM model.
The background expansion history
is fixed to be the same as that in the $\Lambda$CDM model,
but the evolution of perturbations is qualitatively similar to
that in the $f(R)$ models (\ref{imodel}) and (\ref{iimodel}).
We define the following quantity \cite{fR1}
\begin{equation}
B \equiv m \frac{H}{\dot{H}} \frac{\dot{R}}{R}\,,
\label{Bdef}
\end{equation}
whose today's value is denoted as $B_0$.
We note that $B$ is of the same order as $m$.
For increasing $B_0$, the CMB power spectrum exhibits stronger
amplification relative to that in the $\Lambda$CDM model.
This is attributed to the late-time ISW effect
induced by the enhanced gravitational potential $\psi_{\rm eff}$.
As we see in the middle panel of Fig.~\ref{fig2}, for increasing $B_0$,
the weak lensing power spectrum is also subject to stronger enhancement.
In the right panel, we can also confirm that the modified growth
of $\delta_m$ leads to the amplification of the matter power spectrum.
\begin{figure}[h]
\begin{center}
\includegraphics[height=2.5in,width=7.0in]{fig2.pdf}
\end{center}
\caption{\label{fig2}
The CMB angular temperature power spectrum (left), the lensing power spectrum
(middle), and the linear matter power spectrum (right) in $f(R)$ models of
the late-time cosmic acceleration with the $\Lambda$CDM background.
We plot theoretical predictions of three $f(R)$ models
($B_0=5, 1, 0.1$) besides that in the $\Lambda$CDM model.
Taken from Ref.~\cite{Battye}. }
\end{figure}
For the $f(R)$ model (\ref{imodel}), the likelihood analysis based on the data
of CMB, galaxy clusterings, weak gravitational lensing combined
with the SN Ia, BAO, and $H_0$
data showed that the parameter $B_0$ is constrained to be \cite{Lomb}
\begin{equation}
B_0<1.1 \times 10^{-3}\,,
\end{equation}
at 95 \% CL.
The similar bounds were also derived in Refs.~\cite{Sil13,Raveri14}
and \cite{Battye} by taking the effective field theory approach and
the designer approach to $f(R)$ gravity, respectively.
In the designer approach, the time-dependence of $w_{\rm DE}$ is
a priori assumed, under which the functional form of $f(R)$ is determined.
For designer $f(R)$ models with constant $w_{\rm DE}$, the joint data analysis
based on CMB, BAO, and lensing data placed the bounds $|w_{\rm DE}+1|<0.002$
and $B_0<0.006$ at 95 \% CL \cite{Battye}.
We also note that there are bounds on the value of $|f_{,R}-1|$ constrained
from the suppression of fifth forces in dense environments.
For the model (\ref{imodel}), the screening of fifth forces in the Milky Way
and dwarf galaxies put the bounds $|f_{,R0}-1| \lesssim 10^{-5}$ \cite{fR1} and
$|f_{,R0}-1| \lesssim 10^{-7}$ \cite{Jain12}, respectively, at 95 \% CL,
where $f_{,R0}$ is today's value of $f_{,R}$.
These bounds show that there have been no observational signatures for the $f(R)$
modification of gravity. The crucial point is that the gravitational coupling
with matter is large in the massless regime
${\cal F}_M(k)<1$, such that $\mu=(4/3)e^{-2\phi/(\sqrt{6}M_{\rm pl})}$.
To avoid the large enhancement of $\delta_m$, the perturbations
need to be in the massive regime ${\cal F}_M(k)>1$ until recently. In this case, however,
the scalar field hardly evolves by today, so the $f(R)$ models are not be distinguished from
the $\Lambda$CDM model both at the levels of background and perturbations.
\subsection{BD theory with a scalar potential}
The BD theory with a light scalar (satisfying $M_{\phi}^2 \lesssim H^2$) can
be consistent with local gravity constraints for the coupling
$|Q|<2.5 \times 10^{-3}$.
In this case, however, it is difficult to distinguish BD theory from quintessence.
If $|Q| \gtrsim 10^{-3}$, then the potential needs to be designed to have a large
mass in regions of the high density for suppressing the propagation of fifth forces.
The potential of this type is given by \cite{BD1}
\begin{equation}
V(\phi)=V_0 \left[ 1-C (1-e^{-2Q \phi/M_{\rm pl}})^p \right]\,,
\label{Vphige}
\end{equation}
where $V_0>0, 0<C<1, 0<p<1$ are constants.
This is the generalization of the potential (\ref{Vphiap}) in
$f(R)$ gravity to arbitrary couplings $Q$.
The background cosmological dynamics in BD theory with (\ref{Vphige})
is similar to that in $f(R)$ models explained in Sec.~\ref{fRsec}.
As long as $|\phi/M_{\rm pl}|\ll1$ in the matter era, the field slowly evolves along the instantaneous minimum
at $\phi=\phi_m$ determined by the condition (\ref{mire1}), i.e.,
\begin{equation}
\phi_m \simeq \frac{M_{\rm pl}}{2Q}
\left( \frac{2V_0p\,C}{\rho_m} \right)^{1/(1-p)}\,.
\label{phimexpre}
\end{equation}
In the limit $\rho_m \to \infty$, $\phi_m$ approaches 0 with the
divergent mass squared $M_{\phi}^2=V_{,\phi \phi}$. Hence
the fine-tuning problem of initial conditions mentioned in
Sec.~\ref{fRsec} also persists in this case.
During the early matter era, the variation of $\phi$ is suppressed due to
the heavy mass, so $w_{\rm DE}$ is close to $-1$.
The large deviation of $w_{\rm DE}$ from $-1$ starts to occur after
$\rho_m$ drops to the order of $V(\phi)$.
Finally, the solution approaches the de Sitter fixed point
given by Eq.~(\ref{mire2}).
Under the bound (\ref{Qbound}), the deviation of $w_{\rm DE}$ from
$-1$ around today should be also limited as in the case of $f(R)$ gravity.
Since $M_{\phi}^2$ is large in the early matter era, the perturbations
relevant to the linear growth of large-scale structures are initially
in the massive region ${\cal F}_M(k)>1$.
After the entry to the massless regime ${\cal F}_M(k)<1$,
the quantities $\mu$ and $\Sigma$ approach the values
given by Eq.~(\ref{muBD}).
Solving Eq.~(\ref{delmeqf}) with Eqs.~(\ref{Poi}) and (\ref{psieff})
in the massless regime, the evolution of perturbations
during the matter-dominated epoch yields
\begin{equation}
\delta_{m}\propto t^{(\sqrt{25+48Q^{2}}-1)/6}\,,\qquad
\psi_{\rm eff}=-\frac{2}{1+2Q^2}\Psi
=\frac{2}{1-2Q^2}\Phi \propto
t^{(\sqrt{25+48Q^{2}}-5)/6}\,.
\label{delmBD}
\end{equation}
For the coupling $|Q|>1/\sqrt{6}$, the growth rate of $\delta_m$ is
larger than that in $f(R)$ gravity.
This is also the case for the gravitational potentials
$\psi_{\rm eff}, \Psi, \Phi$.
Then, the observational constraints on BD theory with $|Q|>1/\sqrt{6}$ should be more stringent than those on
$f(R)$ models. As we explained in Sec.~\ref{fRsec} in $f(R)$ gravity,
it is hard to observationally distinguish BD theory with the potential
from the $\Lambda$CDM model.
If the chameleon mechanism works in local regions of the Universe,
there is the bound arising from the violation of equivalence principle
in the solar system \cite{Will,chame2}.
For the potential (\ref{Vphige}), this bound translates to \cite{BD1}
\begin{equation}
p>1-\frac{5}{13.8-{\rm log}_{10}|Q|}\,.
\end{equation}
For $|Q|=0.1$ and $|Q|=0.01$, we have $p>0.66$ and
$p>0.68$, respectively.
It may be of interest to study whether BD theory with
$10^{-3} \lesssim |Q| \lesssim 0.1$ can leave some
observational signatures different from those in the
$\Lambda$CDM model
(see Refs.~\cite{Fine1,Fine2,Fine3} for
related works).
If we transform the action (\ref{lagbra}) to that in the Einstein frame
under the conformal transformation, the scalar field $\phi$
has a direct coupling $Q$ with nonrelativistic matter
(dark matter and baryons) \cite{chame2,fRreview}.
Instead, we may start from the
Einstein frame action by assuming that the field $\phi$
is coupled to dark matter alone \cite{Amenco}.
Then, the solar-system constraints are evaded due
to the absence of interactions between baryons and the scalar field.
In this coupled quintessence scenario with the
inverse power-law potential
$V(\phi)=V_0\phi^{-p}$ ($p>0$),
the Planck CMB data placed the upper bound
$|Q|<0.066$ at 95 \% CL \cite{Planckdark}.
Combined with the BAO and weak lensing data, there is
a peak around $|Q|=0.04$ for the marginalized posterior
distribution of $Q$ \cite{Petto,Planckdark}.
The statistically strong observational evidence for supporting
the nonvanishing coupling $Q$ has been still lacking, but the future
high-precision observations may clarify whether the coupling
of order $|Q|={\cal O}(0.01)$ is really favored from the data.
\section{Class (C): Kinetic braidings}
\label{classCsec}
The Lagrangian of class (C), which is known as
kinetic braidings, is expressed in the form
\begin{equation}
L=G_2(\phi, X)+G_3(\phi, X) \square \phi
+\frac{M_{\rm pl}^2}{2}R\,.
\label{modelC}
\end{equation}
Since $q_t=M_{\rm pl}^2$, the no-ghost
condition of tensor perturbations is satisfied.
The parameter $\alpha_{\rm M}$ vanishes due
to the absence of nonminimal couplings $G_4(\phi)$.
Then, the above theories automatically evade the
bound (\ref{alpM}).
Unlike $f(R)$ gravity, we do not need to consider a heavy mass of $\phi$ in dense regions to suppress the propagation
of fifth forces, so we will focus on the case in which the field
mass squared $M_{\phi}^2=-G_{2,\phi \phi}$ does not
exceed the order of $H^2$.
Provided that the ratio $q_t/q_s$ is not much different from 1,
the term $2a^2 M_{\phi}^2q_t/(c_s^2 k^2 q_s)$ in
the expression of $\Delta_2$ of Eq.~(\ref{Delta2}) is
much smaller than 1 for the modes deep inside the
sound horizon.
Substituting $\Delta_2 \simeq \tp^2 q_s c_s^2/(4M_{\rm pl}^2H^2)$
into Eq.~(\ref{muC}), it follows that
\begin{equation}
\mu=\Sigma \simeq 1+\frac{4M_{\rm pl}^4H^2\aB^2}
{\tp^2q_sc_s^2}\,,
\label{muIII}
\end{equation}
which does not have the scale-dependence.
The evolution of linear perturbations is determined by
the product $4M_{\rm pl}^4H^2\aB^2/(\tp^2q_sc_s^2)$.
Due to the absence of nonminimal couplings in the
Lagrangian (\ref{modelC}), the typical
solution around a spherically symmetric and static body
corresponds to $\phi={\rm constant}$ \cite{DKT12,Kase13},
so that the propagation of fifth forces is suppressed in local regions of the Universe.
\subsection{Concrete dark energy models}
We present models of the late-time cosmic acceleration which
belong to the Lagrangian (\ref{modelC}).
The cubic covariant Galileon \cite{Galileons} corresponds to
\begin{equation}
{\rm Model~(C1)}:~~
L=\beta_1X-m^3 \phi+\beta_3X \square \phi
+\frac{M_{\rm pl}^2}{2}R\,,
\label{iii1}
\end{equation}
where $\beta_1, \beta_3, m$ are constants.
In the limit of Minkowski spacetime, the equations of motion following
from the Lagrangian (\ref{iii1}) respects the Galilean symmetry
$\partial_{\mu} \phi \to \partial_{\mu} \phi+b_{\mu}$.
In absence of the linear linear potential $V(\phi)=m^3 \phi$,
it is known that there exists a tracker solution for the full
Galileon Lagrangian containing quartic and quintic
couplings \cite{DT10,DT10b}.
The dark energy EOS along the tracker is
given by $w_{\rm DE}=-2$ during the matter era.
This is followed by the approach to the self-accelerating
de Sitter solution characterized by $X={\rm constant}$.
The tracker solution is disfavored from the joint data analysis
of SNIa, CMB, and BAO \cite{NDT10} because of the large
deviation of $w_{\rm DE}$ from $-1$ in the matter era.
Even in the case where the solutions approach the tracker
at late times, the covariant Galileon without the linear
potential is in strong tension with observational
data \cite{AppleLin,Neveu,Barreira1,Barreira2,Renk,Peirone2}
due to a very different structure formation pattern
compared to that in the
$\Lambda$CDM model \cite{Kase10}.
The likelihood analysis performed in
Refs.~\cite{Barreira1,Barreira2,Renk,Peirone2} assumed that the
Galileon potential is absent. In presence of the linear potential
$V(\phi)=m^3 \phi$, we do not have the self-accelerating solution
with $X={\rm constant}$, but the late-time cosmic acceleration
can be driven by $V(\phi)$ (see also Ref.~\cite{Ali12}
for other potentials).
After the field $\phi$ enters the region $V(\phi)<0$, the
Universe starts to enter the collapsing stage.
Existence of the linear potential can modify the dynamics of
background and perturbations, so it is worthy of studying
the compatibility of model (\ref{iii1}) with observations.
In Sec.~\ref{mostsec}, we will discuss the cosmological dynamics
of such a model by taking into account a more general nonminimal
coupling $G_4(\phi)$.
Instead of the linear potential $V(\phi)=m^3 \phi$, we can consider the quadratic kinetic term $X^2$.
This theory is described by the Lagrangian \cite{Kase18}:
\begin{equation}
{\rm Model~(C2)}:~~
L=\beta_1 X+\beta_2 X^2+\beta_3 X \square \phi
+\frac{M_{\rm pl}^2}{2}R\,,
\label{iii2}
\end{equation}
where $\beta_1,\beta_2,\beta_3$ are constants.
This can be regarded as the ghost condensate model
with the cubic Galileon term.
In absence of the term $\beta_2 X^2$, the solutions tend to
approach the tracker solution mentioned above, which is
observationally excluded.
Moreover, the dominance of the cubic Galileon term
$\beta_3 X \square \phi$ at low redshifts typically gives rise
to the galaxy-ISW anti-correlation, which is also
disfavored from the data \cite{Renk}.
For the model (\ref{iii2}), the new term $\beta_2 X^2$ prevents the
approach to the tracker solution, so that the deviation
of $w_{\rm DE}$ from $-1$ is smaller than that in the tracker case.
Moreover, the cubic Galileon term can be subdominant to
the contribution $G_2=\beta_1 X+\beta_2 X^2$, so it is expected that
the galaxy-ISW anti-correlation does not necessarily occur.
In Sec.~\ref{cghostsec}, we study the cosmological dynamics
and observational signatures for the model (\ref{iii2}).
\subsection{Cubic Galileon with ghost condensate}
\label{cghostsec}
As an example of class (C), we review the cosmology of
model (C2) given by the Lagrangian (\ref{iii2}).
In theories beyond Horndeski
gravity \cite{Zuma,GLPV,Langlois1,Langlois2,Crisostomi16},
it is possible to introduce the $X$ dependence in the
function $G_4$, while keeping the
value $c_t^2=1$ due to an additional term outside the
domain of Horndeski theories \cite{GWcon2,GWcon3}.
Since this review focuses on Horndeski theories, we will not
take into account such terms in the following.
For the matter sector, we take into account radiation
(density $\rho_r$ and pressure $P_r=\rho_r/3$) and
nonrelativistic matter (density $\rho_m$ and pressure $P_m=0$).
We also introduce the following dimensionless quantities:
\begin{equation}
\Omega_{\phi 1} \equiv
\frac{\beta_1 \dot{\phi}^2}{6M_{\rm pl}^2 H^2}\,,
\qquad
\Omega_{\phi 2} \equiv
\frac{\beta_2 \dot{\phi}^4}{4M_{\rm pl}^2 H^2}\,,
\qquad
\Omega_{\phi 3} \equiv
-\frac{\beta_3\dot{\phi}^3}{M_{\rm pl}^2 H}\,,
\qquad
\Omega_r \equiv \frac{\rho_r}{3 M_{\rm pl}^2 H^2}\,,
\end{equation}
which correspond to the density parameters arising from
$\beta_1 X$, $\beta_2 X^2$, $\beta_3 X \square \phi$,
and radiation, respectively.
The Hamiltonian constraint (\ref{Frieq}) gives the relation
\begin{equation}
\Omega_m \equiv \frac{\rho_m}{3M_{\rm pl}^2 H^2}
=1-\Omega_{\rm DE}-\Omega_r\,,\qquad
\Omega_{\rm DE} \equiv
\Omega_{\phi 1}+\Omega_{\phi 2}+\Omega_{\phi 3}\,.
\label{Omem}
\end{equation}
The density parameters obey the differential equations:
\begin{eqnarray}
\Omega_{\phi 1}' = 2\Omega_{\phi 1}
\left( \epsilon_{\phi}-h \right)\,,\qquad
\Omega_{\phi 2}' = 2\Omega_{\phi 2}
\left( 2\epsilon_{\phi}-h \right)\,,\qquad
\Omega_{\phi 3}' = \Omega_{\phi 3}
\left( 3\epsilon_{\phi}-h \right)\,,\qquad
\Omega_r' = -2\Omega_r \left( 2+h \right)\,,
\label{Omereq}
\end{eqnarray}
where, from Eqs.~(\ref{dH}) and (\ref{ddphi}),
the quantities $h \equiv \dot{H}/H^2$ and
$\epsilon_{\phi} \equiv
\ddot{\phi}/(H \dot{\phi})$ are given, respectively, by
\begin{eqnarray}
h &=&
-\frac{2(3\Omega_{\phi 1}+\Omega_{\phi 2}
+\Omega_r+3)(\Omega_{\phi 1}+2\Omega_{\phi 2})
+2\Omega_{\phi 3}(6\Omega_{\phi 1}+3\Omega_{\phi 2}
+\Omega_r+3)+3\Omega_{\phi 3}^2}
{4 \left( \Omega_{\phi 1}+2\Omega_{\phi 2}
+\Omega_{\phi 3} \right)+\Omega_{\phi 3}^2}\,,
\label{dh}\\
\epsilon_{\phi}
&=&-\frac{4(3\Omega_{\phi 1}+2\Omega_{\phi 2})
-\Omega_{\phi 3}(3\Omega_{\phi 1}+\Omega_{\phi 2}
+\Omega_r-3)}{4 \left( \Omega_{\phi 1}+2\Omega_{\phi 2}
+\Omega_{\phi 3} \right)+\Omega_{\phi 3}^2}\,.
\label{epphi}
\end{eqnarray}
The dark energy EOS defined by Eq.~(\ref{wdedef}) yields
\begin{equation}
w_{\rm DE}=\frac{3\Omega_{\phi 1}+\Omega_{\phi 2}
-\epsilon_{\phi}\Omega_{\phi 3}}
{3(\Omega_{\phi 1}+\Omega_{\phi 2}+\Omega_{\phi 3})}\,.
\label{wdecu}
\end{equation}
For the dynamical system given by Eq.~(\ref{Omereq}),
there exists a self-accelerating de Sitter solution
satisfying $\dot{\phi}={\rm constant}$ and $H={\rm constant}$,
in which case $\epsilon_{\phi}=0, h=0$, and $\Omega_r=0$.
From Eqs.~(\ref{dh}) and (\ref{epphi}), we obtain the following
two relations on the de Sitter solution:
\begin{equation}
\Omega_{\phi 1}=-2+\frac{1}{2}\Omega_{\phi 3}\,,\qquad
\Omega_{\phi 2}=3-\frac{3}{2}\Omega_{\phi 3}\,,
\label{Omephire}
\end{equation}
under which the dark energy EOS (\ref{wdecu})
reduces to $w_{\rm DE}=-1$.
As shown in Ref.~\cite{Kase18}, this de Sitter solution is
always a stable attractor.
In the limit $\Omega_{\phi 3} \to 0$, we have
$\Omega_{\phi 1}=-2$ and $\Omega_{\phi 2}=3$,
so that $\beta_1=-\beta_2 \dot{\phi}^2$.
Indeed, this is equivalent to the de Sitter solution
in ghost condensate corresponding to $G_{2,X}=0$
in Eq.~(\ref{wdeke}). As in the case of ghost condensate,
we consider the couplings $\beta_1<0$ and $\beta_2>0$
in the following discussion.
In the radiation and deep matter eras, we first consider the case
in which $\Omega_{\phi 3}$ dominates over
$|\Omega_{\phi 1}|$ and $\Omega_{\phi 2}$.
In these epochs, the field density parameters are smaller than
the order 1, so we obtain the approximate relations
$h \simeq -(\Omega_r+3)/2$ and
$\epsilon_{\phi} \simeq (\Omega_r-3)/4$
from Eqs.~(\ref{dh}) and (\ref{epphi}).
Substituting them into Eq.~(\ref{wdecu}), it follows that
\begin{equation}
w_{\rm DE} \simeq \frac{1}{4}-\frac{1}{12}\Omega_r\,,
\label{wdeG3}
\end{equation}
and hence $w_{\rm DE} \simeq 1/6$ in the radiation era and
$w_{\rm DE} \simeq 1/4$ in the matter era.
This behavior can be confirmed in the numerical simulation
of Fig.~\ref{fig3}.
Since $h \simeq -3/2$ and $\epsilon_{\phi} \simeq -3/4 $
during the matter dominance, we can integrate Eq.~(\ref{Omereq})
to give $|\Omega_{\phi 1}| \propto a^{3/2}, \Omega_{\phi 2}
\propto a^0$, and $\Omega_{\phi 3} \propto a^{-3/4}$.
Then, the density parameters $|\Omega_{\phi 1}|$ and
$\Omega_{\phi 2}$ eventually catch up with
$\Omega_{\phi 3}$. After this catch up, the evolution of
$w_{\rm DE}$ is no longer described by Eq.~(\ref{wdeG3}).
There is a transient period in which $w_{\rm DE}$ enters the
region $w_{\rm DE}<-1$.
The solutions finally approach the de Sitter fixed point
characterized by Eq.~(\ref{Omephire}).
\begin{figure}[h]
\begin{center}
\includegraphics[height=3.2in,width=3.4in]{fig3a.pdf}
\includegraphics[height=3.2in,width=3.4in]{fig3b.pdf}
\end{center}
\caption{\label{fig3}
(Left) Evolution of $w_{\rm DE}$ versus $z + 1$
for four different initial conditions:
(a) $\Omega_{\phi 1}=-1.8 \times 10^{-14}$,
$\Omega_{\phi 2}=1.0 \times 10^{-11}$,
(b) $\Omega_{\phi 1}=-1.0 \times 10^{-14}$,
$\Omega_{\phi 2}=3.0 \times 10^{-12}$,
(c) $\Omega_{\phi 1}=-4.0 \times 10^{-15}$,
$\Omega_{\phi 2}=5.0 \times 10^{-13}$, and
(d) $\Omega_{\phi 1}=-1.8 \times 10^{-15}$,
$\Omega_{\phi 2}=1.0 \times 10^{-13}$,
at the redshift $z=1.3 \times 10^5$.
The other initial conditions are chosen to be
$\Omega_{\phi 3}=1.0 \times 10^{-6}$ and
$\Omega_r=0.975$.
The present epoch ($z=0$) corresponds to
$\Omega_{\rm DE}=0.68$ and $\Omega_r \simeq 10^{-4}$.
(Right) Evolution of $\mu~(=\Sigma)$ corresponding to the cases
(a), (b), (c), (d) plotted in the left panel.
}
\end{figure}
For smaller initial conditions of $\Omega_{\phi 2}$, the minimum value
of $w_{\rm DE}$ reached during the transient period tends to
decrease toward $-2$, see the left panel of Fig.~\ref{fig3}.
This can be understood as follows. Taking the limit
$\Omega_{\phi 2} \to 0$ in Eq.~(\ref{Omereq}), the quantity
$y \equiv 2\Omega_{\phi 1}+\Omega_{\phi 3}$
obeys the differential equation
\begin{equation}
y'=y \frac{12\Omega_{\phi 1}^2+\Omega_{\phi 1} (4\Omega_r-12+
27\Omega_{\phi 3})+\Omega_{\phi 3} (5\Omega_r-3
+3\Omega_{\phi 3})}
{4\Omega_{\phi 1}+4\Omega_{\phi 3}+\Omega_{\phi 3}^2}\,,
\end{equation}
which admits the solution $y=0$, i.e., $\Omega_{\phi 3}=-2\Omega_{\phi 1}$. Substituting this relation and Eq.~(\ref{epphi})
into Eq.~(\ref{wdecu}), it follows that
\begin{equation}
w_{\rm DE}=-\frac{6+\Omega_r}{3(1-\Omega_{\phi 1})}\,.
\end{equation}
This is the tracker solution \cite{DT10,DT10b} along
which $w_{\rm DE}$ evolves as $-7/3$ (radiation era)
$\to -2$ (matter era) $\to -1$
(de Sitter era satisfying $\Omega_{\phi 1}=-1$ and
$\Omega_{\phi 3}=2$).
As we already mentioned, the tracker is in tension with
the observational data \cite{NDT10} due to the large deviation of
$w_{\rm DE}$ from $-1$.
Since $\Omega_{\phi 2} \neq 0$ in the present model,
the dominance of $\Omega_{\phi 2}$ over $\Omega_{\phi 3}$
prevents the approach to the tracker.
For larger initial values of $\Omega_{\phi 2}$, the deviation
of $w_{\rm DE}$ from $-1$ tends to be smaller.
Moreover, the approach to $w_{\rm DE}=-1$ occurs
at higher redshifts.
Thus, the existence of the term $\beta_2 X^2$ allows an
interesting possibility for realizing the phantom equation
of state ($w_{\rm DE}<-1$) which can be compatible
with observations.
The quantities $q_s$ and $c_s^2$ can be expressed,
respectively, as
\begin{equation}
q_s=\frac{3M_{\rm pl}^4H^2}{\tp^2}
\left( 4 \Omega_{\phi 1}+8\Omega_{\phi 2}
+4\Omega_{\phi 3}
+\Omega_{\phi 3}^2 \right)\,,\qquad
c_s^2=\frac{12\Omega_{\phi 1}+8\Omega_{\phi 2}
+4(\epsilon_{\phi}+2)\Omega_{\phi 3}-\Omega_{\phi 3}^2}
{3(4\Omega_{\phi 1}+8\Omega_{\phi 2}+4\Omega_{\phi 3}+
\Omega_{\phi 3}^2)}\,.
\label{stabC2}
\end{equation}
The conditions $q_s>0$ and $c_s^2>0$ hold for all
the cases shown in Fig.~\ref{fig3}, so
there are no ghost and Laplacian instabilities of scalar perturbations.
Substituting Eq.~(\ref{stabC2}) and $\aB=-\Omega_{\phi 3}/2$ into Eq.~(\ref{muIII}), we obtain
\begin{equation}
\mu=\Sigma \simeq 1+\frac{\Omega_{\phi 3}^2}
{12\Omega_{\phi 1}+8\Omega_{\phi 2}
+4(\epsilon_{\phi}+2)\Omega_{\phi 3}-\Omega_{\phi 3}^2}\,,
\label{muIII2}
\end{equation}
so that the deviations of $\mu$ and $\Sigma$ arise from
the cubic coupling. Using the relations (\ref{Omephire}) on the de Sitter
fixed point, it follows that
\begin{equation}
\mu=\Sigma \simeq 1+\frac{\Omega_{\phi 3}}{2-\Omega_{\phi 3}}\,.
\label{mudS}
\end{equation}
In the right panel of Fig.~\ref{fig3}, the evolution of $\mu$ is plotted
for four different cases corresponding to those in the left panel.
Solving the perturbation equations of motion numerically, we
compute $\mu$ and $\Sigma$ according to their definitions given
in Eqs.~(\ref{Poi}) and (\ref{psieff}). We confirmed that the analytic
estimation (\ref{muIII2}), which is derived under the quasi-static approximation,
is sufficiently accurate for the modes deep inside the sound horizon.
In the deep matter era, $\mu$ is close to 1 in all these cases,
but the difference arises at low redshifts depending on the values
of $\Omega_{\phi 3}$. For the cases (a), (b), (c), (d), the numerical
values of $\Omega_{\phi 3}$ today are given, respectively, by
$2.3 \times 10^{-2}$, $5.6 \times 10^{-2}$, $2.0 \times 10^{-1}$,
$5.1 \times 10^{-1}$, so the deviation of $\mu$ from 1 tends to be
more significant for larger $\Omega_{\phi 3}$.
This property is also consistent with the estimation (\ref{mudS})
on the de Sitter solution.
For the cases in which $\Omega_{\phi 3}$ gives the large contribution to the field density today
(say $\Omega_{\phi 3}(z=0)>{\cal O}(0.1)$),
the model may be tightly constrained from the galaxy-ISW correlation
data. Moreover, in such cases, the deviation of $w_{\rm DE}$
from $-1$ at low redshifts tends to be significant.
It will be of interest to place observational constraints on today's
value of $\Omega_{\phi 3}$ by performing the joint likelihood analysis
including the galaxy-ISW correlation data.
\section{Class (D): Nonminimally coupled scalar with cubic
derivative interactions}
\label{mostsec}
The class (D) is the most general Horndeski theories given by
the Lagrangian (\ref{lagcon}) in which $c_t^2$ is exactly equivalent to 1.
In this case, the quantities $\alpha_{\rm B}$ and
$\alpha_{\rm M}$ give rise to nonvanishing contributions
to the second terms in Eqs.~(\ref{Gefff2}) and (\ref{Sigmaf2}).
The difference between $\mu$ and
$\Sigma$ is given by
\begin{equation}
\mu-\Sigma=-\frac{2M_{\rm pl}^2H^2q_t\alpha_{\rm M}
(\alpha_{\rm B}-\alpha_{\rm M})}
{\tp^2 q_s c_s^2}\,,
\end{equation}
which means that the nonminimal coupling $G_4(\phi)R$
generally leads to $\mu \neq \Sigma$ due to the nonvanishing
$\alpha_{\rm M}$.
This property is different from that in the models of
classes (A) and (C).
In Sec.~\ref{fRBDsec}, we showed that it is difficult to observationally
distinguish $f(R)$ gravity and BD theory with a scalar potential
from the $\Lambda$CDM model.
In such class (B) theories, the field potential needs to be designed to have
a large mass in regions of the high density for the
chameleon mechanism to work.
There is yet another mechanism dubbed the Vainshtein
mechanism \cite{Vain} in which field nonlinear
self-interactions (like the cubic Galileon) can suppress
the propagation of fifth forces in local regions of the
Universe. In such cases, the background cosmological
dynamics and the evolution of perturbations are
different from those in class (B).
In class (D) the quantity $\Sigma$ can grow rapidly at
low redshifts, while this is not generally the case for
class (B) in which $\Sigma=1/(16\pi G G_4)$.
This allows the possibility for distinguishing the models
between classes (B) and (D) from the observations of
galaxy-ISW correlations further.
The class (D) is categorized as Horndeski theories containing
a nonminimally coupled light scalar field with cubic derivative
interactions.
\subsection{Concrete dark energy models}
One of the theories in class (D) is known as generalized
BD theories \cite{DTge} given by the Lagrangian
\begin{equation}
{\rm Model~(D1)}:~~L=\omega \left( \frac{\phi}{M_{\rm pl}}
\right)^{1-n}X+\frac{\lambda}{\mu^3} \left( \frac{\phi}{M_{\rm pl}}
\right)^{-n}X \square \phi+\frac{M_{\rm pl}^2}{2} \left( \frac{\phi}{M_{\rm pl}}
\right)^{3-n}R\,,
\label{LagGBD}
\end{equation}
where $\mu~(>0)$ is a constant having a dimension of mass,
and $n, \lambda, \omega$ are dimensionless constants.
The original BD theory \cite{Brans} without the cubic coupling corresponds to the power $n=2$.
In presence of the cubic derivative interaction,
there is a de Sitter solution characterized by
$\dot{\phi}/(H \phi)={\rm constant}$ even
without the scalar potential.
The case $n=2$ was first studied in Ref.~\cite{Silva}
(see also Refs.~\cite{Koba1,Koba2}) and then it was
generalized to the more general case with $n \neq 2$
in Ref.~\cite{DTge}.
The power $n$ is constrained to be
\begin{equation}
2 \le n \le 3\,,
\end{equation}
to satisfy theoretical consistent conditions throughout
the cosmic expansion history \cite{DTge}.
In model (D1), the cubic derivative coupling is the main
source for the late-time cosmic acceleration. For $n=2$, it was
shown in Ref.~\cite{Koba1} that the galaxy-ISW
anti-correlation tends to be stronger for larger $|\omega|$.
For $n \neq 2$, the galaxy-ISW correlation power spectrum
was not computed yet, but it is expected that the anti-correlation can still persist for the case in which the cubic
interaction provides the dominant contribution to the dark energy
density. It remains to be seen to put observational
constraints on model (D1) by using the galaxy-ISW
correlation data.
There is also the nonminimally coupled cubic Galileon model
given by the Lagrangian
\begin{equation}
{\rm Model~(D2)}:~~
L=\left( 1-6Q^2 \right) e^{-2Q (\phi-\phi_0)/M_{\rm pl}}X
-m^3 \phi+\beta_3 X \square \phi+\frac{M_{\rm pl}^2}{2}
e^{-2Q (\phi-\phi_0)/M_{\rm pl}}R\,,
\label{modelD2}
\end{equation}
where $Q, \phi_0, m, \beta_3$ are constants.
In the limit $Q \to 0$, Eq.~(\ref{modelD2}) recovers
the Lagrangian (\ref{iii1})
of model (C1), i.e., cubic Galileons with $\beta_1=1$.
In high-density regions, the cubic Galileon term can
suppress the fifth force
induced by the nonminimal coupling
$e^{-2Q (\phi-\phi_0)/M_{\rm pl}}R$.
The cosmological dynamics of model (D2) was studied
in Ref.~\cite{KTD15} in a more general context
beyond the Horndeski domain.
In this model, the main source for the late-time cosmic
acceleration is the linear potential $V(\phi)=m^3 \phi$
rather than the cubic Galileon term $\beta_3 X \square \phi$.
Then, unlike model (D1), the dominance of cubic interactions
over $V(\phi)$ does not typically occur at low
redshifts. In Sec.~\ref{cosnonsec}, we discuss the cosmological
dynamics and observational signatures of model (D2).
We note that it is also possible to take into account the
nonminimal coupling for model (C2) introduced
in Sec.~\ref{classCsec}.
It remains to be seen to explore whether such a model
is consistent with cosmological and local gravity constraints.
\subsection{Cosmology in nonminimally
coupled cubic Galileons}
\label{cosnonsec}
To study the cosmological dynamics in model (D2) given
by the Lagrangian (\ref{modelD2}), we introduce the following dimensionless quantities:
\begin{equation}
x_1 \equiv \frac{\dot{\phi}}{\sqrt{6}M_{\rm pl} H}\,,
\qquad
\Omega_{\phi_2} \equiv \frac{m^3 \phi}
{3M_{\rm pl}^2 H^2 F}\,,
\qquad \Omega_{\phi_3} \equiv
-\frac{\beta_3 \dot{\phi}^3}{M_{\rm pl}^2 H F}\,,
\qquad
\Omega_r \equiv \frac{\rho_r}{3M_{\rm pl}^2 H^2 F}\,,
\qquad
\lambda \equiv -\frac{M_{\rm pl}}{\phi}\,,
\label{moDdi}
\end{equation}
where $F \equiv e^{-2Q (\phi-\phi_0)/M_{\rm pl}}$.
The effective gravitational coupling in a screened
environment is given by $G_{\rm eff}=G/F$, so we consider
the case in which the field value $\phi$ today is
equivalent to $\phi_0$, i.e., $F(z=0)=1$.
Taking into account the radiation and nonrelativistic matter
as in the analysis of Sec.~\ref{cghostsec}, the Hamiltonian constraint (\ref{Frieq}) leads to
\begin{equation}
\Omega_m \equiv \frac{\rho_m}{3M_{\rm pl}^2 H^2 F}
=1-\left( 1-6Q^2 \right)x_1^2 -2\sqrt{6} Q x_1
-\Omega_{\phi 2}-\Omega_{\phi 3}-\Omega_r\,.
\label{Omem2}
\end{equation}
The quantities defined by Eq.~(\ref{moDdi}) satisfy
the differential equations:
\begin{eqnarray}
& &
x_1'=x_1 \left( \epsilon_{\phi}- h \right)\,,\qquad
\Omega_{\phi 2}'= \Omega_{\phi 2} \left[
\sqrt{6} \left( 2Q-\lambda \right)x_1-2h \right]\,,\qquad
\Omega_{\phi 3}'= \Omega_{\phi 3} \left(
2\sqrt{6} Q x_1+3\epsilon_{\phi}-h \right)\,,\nonumber \\
& &
\Omega_r'=2 \Omega_r \left( \sqrt{6} Q x_1-2-h \right)\,,
\qquad
\lambda'=\sqrt{6} \lambda^2 x_1\,,
\end{eqnarray}
where the quantities $h=\dot{H}/H^2$ and
$\epsilon_{\phi}=\ddot{\phi}/(H \dot{\phi})$ are
given, respectively, by
\begin{eqnarray}
h &=& -[ \Omega_{\phi 3}(6+2\Omega_r-6\Omega_{\phi 2}+3\Omega_{\phi 3})
+\sqrt{6}\Omega_{\phi 3}
(2Q-\lambda \Omega_{\phi 2})x_1
+2 \{ 3+\Omega_r-3\Omega_{\phi 2}+6\Omega_{\phi 3}
-6\lambda Q \Omega_{\phi 2} \nonumber \\
& &+6Q^2 (1-\Omega_r+3\Omega_{\phi 2}
-2\Omega_{\phi 3}) \}x_1^2+2\sqrt{6}Q (6Q^2-1)
(\Omega_{\phi 3}-2)x_1^3+6(12Q^4-8Q^2+1)x_1^4]
/{\cal D}\,,\label{hD}\\
\epsilon_{\phi} &=&
[ \sqrt{6} \Omega_{\phi 3} (\Omega_r-3\Omega_{\phi 2}
-3)+12 \{Q (\Omega_r-1-3\Omega_{\phi 2}
-2\Omega_{\phi 3})+\lambda \Omega_{\phi 2} \}
x_1 \nonumber \\
& &
+3\sqrt{6} \{ \Omega_{\phi 3}-4+2Q^2
(4+\Omega_{\phi 3}) \}x_1^2
+12Q (5-6Q^2)x_1^3]/(\sqrt{6}{\cal D})\,,
\label{epD}
\end{eqnarray}
with
\begin{equation}
{\cal D}=4x_1^2+4\Omega_{\phi 3}+4\sqrt{6} Qx_1
\Omega_{\phi 3}+\Omega_{\phi 3}^2\,.
\end{equation}
The dark energy EOS defined by Eq.~(\ref{wdedef})
can be expressed as\footnote{In Ref.~\cite{KTD15}, the dark energy
equation of state was defined in a different way
such that Eqs.~(\ref{Frieq}) and (\ref{Frieq2}) are expressed
to contain the terms $3M_{\rm pl}^2F_0H^2$ and $2M_{\rm pl}^2F_0 \dot{H}$
on the left hand sides of them, respectively, where $F_0$ is today's value of $F$.
The definition of $w_{\rm DE}$ in this review corresponds to the choice $F_0=1$,
which should be more suitable to constrain the model from observations.}
\begin{equation}
w_{\rm DE}=-\frac{3+2h-[3+2h+3(1+2Q^2)x_1^2-3\Omega_{\phi 2}
-\epsilon_{\phi}\Omega_{\phi 3}-2\sqrt{6}Q x_1(2+\epsilon_{\phi})]F}
{3-3[1-\Omega_{\phi 2}-\Omega_{\phi 3}+(6Q^2-1)x_1^2-2\sqrt{6}Qx_1]F}\,.
\label{wdeD}
\end{equation}
Let us consider the case in which the cubic Galileon density parameter
$\Omega_{\phi 3}$ (assumed to be positive) dominates over the other field density parameters
in the early cosmological epoch, except for $Qx_1$,
such that $1 \gg \Omega_{\phi 3} \gg \{ x_1^2, \Omega_{\phi 2} \}$.
Then, Eqs.~(\ref{hD}) and (\ref{epD}) approximately reduce to
\begin{eqnarray}
h \simeq -\frac{3}{2}-\frac{\Omega_r}{2}\,,\qquad
\epsilon_{\phi} \simeq -\frac{3}{4}+\frac{\Omega_r}{4}
-\frac{\sqrt{6}}{2}
\left( 1-\Omega_r \right) \frac{Qx_1}{\Omega_{\phi 3}}\,.
\label{hepes}
\end{eqnarray}
Depending on the initial ratio $Qx_1/\Omega_{\phi 3}$,
there are two qualitatively different cases for the evolution of
$x_1$, $\Omega_{\phi 2}$, and $\Omega_{\phi 3}$.
If $|Qx_1| \lesssim \Omega_{\phi 3}$, then we have
$h \simeq -2$ and $\epsilon_{\phi} \simeq -1/2$ during the
radiation dominance, so we obtain the integrated solutions
$x_1^2 \propto a^{3}$, $\Omega_{\phi 2} \propto a^4$,
and $\Omega_{\phi 3} \propto a^{1/2}$.
In this case, even if $\Omega_{\phi 3}$ initially
dominates over the other field densities,
$\Omega_{\phi 2}$ quickly catches up with $\Omega_{\phi 3}$.
After this catch up, both $x_1^2$ and $\Omega_{\phi 3}$ become
much smaller than $\Omega_{\phi 2}$, so the linear potential
$V(\phi)=m^3 \phi$ gives the dominant contribution to
the scalar-field dynamics.
The other case corresponds to the initial conditions satisfying
$|Qx_1| \gg \Omega_{\phi 3}$. Since there is the deviation of
$\Omega_r$ from 1 even during the radiation-dominated epoch,
it can happen that the last term in $\epsilon_{\phi}$ of Eq.~(\ref{hepes})
exceeds the order 1. For $Qx_1/\Omega_{\phi 3}<0$,
$\epsilon_{\phi}$ can be positive during the radiation era.
In the simulation of Fig.~\ref{fig4}, for example, the numerical
value of $\epsilon_{\phi}$ at the redshift $z \approx 10^6$
is around $0.5$, in which case the field density parameters evolve
as $x_1^2 \propto a^{5}$, $\Omega_{\phi 2} \propto a^4$,
and $\Omega_{\phi 3} \propto a^{3.5}$.
Then, even if $x_1^2<\{ \Omega_{\phi 2}, \Omega_{\phi 3} \}$
initially, the Universe enters the stage in which $x_1^2$
dominates over $\Omega_{\phi 2}$ and $ \Omega_{\phi 3}$,
see Fig.~\ref{fig4}.
In this case, after the radiation dominance, the solutions temporally
approach the so-called $\phi$-matter-dominated
epoch ($\phi$MDE) \cite{Amenco}
characterized by the fixed point
\begin{equation}
\left( x_1,\Omega_{\phi 2}, \Omega_{\phi 3}
\right) =\left( -\frac{\sqrt{6}Q}{3(1-2Q^2)},0,0 \right)\,,\quad
{\rm with} \quad w_{\rm eff}=\frac{4Q^2}{3(1-2Q^2)}\,,\quad
{\rm and} \quad \Omega_m=\frac{3-2Q^2}{3(1-2Q^2)^2}\,.
\label{phiMDE}
\end{equation}
For $Q>0$ satisfying the condition $Q^2<1/2$, the existence
of $\phi$MDE requires that $x_1<0$, i.e., $\dot{\phi}<0$.
In this case, the field rolls down along the linear potential
$V(\phi)=m^3 \phi$ with $m>0$.
The cosmic acceleration occurs in the region $\phi>0$,
so the existence of $\phi$MDE requires that $\lambda=-M_{\rm pl}/\phi<0$.
Alternatively, as in Ref.~\cite{KTD15}, we can consider the cases $Q<0, x_1>0, \lambda>0$.
In the left panel of Fig.~\ref{fig4}, we can confirm that $x_1^2$
temporally approaches a constant determined by Eq.~(\ref{phiMDE}).
As we observe in the right panel of Fig.~\ref{fig4}, the effective
EOS $w_{\rm eff}$ starts to evolve from the value close to $1/3$
and temporally approaches the $\phi$MDE value $4Q^2/[3(1-2Q^2)]$.
After the linear potential $V(\phi)=m^3 \phi$ dominates over
the other field densities, the Universe eventually enters the
stage of cosmic acceleration ($w_{\rm eff}<-1/3$).
In Fig.~\ref{fig4}, we also show the evolution of
$w_{\rm DE}$
given by Eq.~(\ref{wdeD}), whose value during the matter era
is slightly larger than 0. Compared to $w_{\rm eff}$,
the transition of $w_{\rm DE}$ to the value around $-1$
occurs earlier.
\begin{figure}[h]
\begin{center}
\includegraphics[height=3.2in,width=3.3in]{fig4a.pdf}
\includegraphics[height=3.2in,width=3.3in]{fig4b.pdf}
\end{center}
\caption{\label{fig4}
(Left) Evolution of $x_1^2, \Omega_{\phi 2},
\Omega_{\phi 3}$ versus $z + 1$
for $Q=0.02$ with the initial conditions
$x_1=-1.6 \times 10^{-13}$, $\Omega_{\phi 2}=7.0 \times
10^{-25}$, $\Omega_{\phi 3}=1.0 \times 10^{-18}$,
$\Omega_r=0.99968$, and $\lambda=-0.2$
at the redshift $z=1.01 \times 10^7$.
The present epoch is identified by
$\Omega_{m}=0.32$ and $\Omega_r \simeq 10^{-4}$.
(Right) Evolution of $w_{\rm eff}$ and $w_{\rm DE}$
for the same value of $Q$ and initial conditions as
those used in the left panel.}
\end{figure}
In comparison to model (C2) discussed in Sec.~\ref{cghostsec},
the contribution of $\Omega_{\phi 3}$ to the total field density
tends to be negligible at an earlier cosmological epoch.
This is attributed to the fact that $\Omega_{\phi 2}$ in model (D2)
evolves as $\Omega_{\phi 2} \propto a^3$ during the matter era,
whereas, in model (C2), the density parameter
$\Omega_{\phi 2}$ associated with the Lagrangian $\beta_2 X^2$
evolves much more slowly ($\Omega_{\phi 2} \propto a^0$).
In the latter case, the cubic Galileon density can dominate
over the other field densities at later cosmological epochs,
so the dark energy
EOS enters the region $w_{\rm DE}<-1$ by
temporally approaching the tracker ($w_{\rm DE}=-2$).
This behavior does not occur for model (D2), in which
$w_{\rm DE}$ typically stays in the region $w_{\rm DE}>-1$,
see the right panel of Fig.~\ref{fig4}.
Hence these two models can be observationally distinguished from
each other even at the background level.
In model (D2), $\Omega_{\phi 3}$ is usually much smaller than 1 today,
so it is expected that the galaxy-ISW correlation data do not put tight
constraints on the model parameters.
The quantities $\alpha_{\rm M}$ and $\alpha_{\rm B}$,
reduce, respectively, to
\begin{equation}
\alpha_{\rm M}=-2\sqrt{6} Q x_1\,,\qquad
\alpha_{\rm B}=-\sqrt{6} Q x_1-\frac{\Omega_{\phi 3}}{2}\,.
\end{equation}
After $\Omega_{\phi 3}$ becomes much smaller than $|Qx_1|$,
the simple relation $\alpha_{\rm M} \simeq 2\alpha_{\rm B}$ holds.
Indeed, this property can be confirmed in the left panel
of Fig.~\ref{fig5}. We recall that there is the bound (\ref{alpM})
on $\alpha_{\rm M}(z=0)$ arising from
Lunar Laser Ranging experiments.
In the numerical simulation on the left panel of Fig.~\ref{fig5},
which corresponds to $Q=0.02$, today's value of $\alpha_{\rm M}$
is $7.0 \times 10^{-3}$ and hence the bound (\ref{alpM}) is satisfied.
For the initial conditions similar to those used in Fig.~\ref{fig4},
the coupling needs to be in the range $|Q| \lesssim 0.04$
for the consistency with (\ref{alpM}).
\begin{figure}[h]
\begin{center}
\includegraphics[height=3.2in,width=3.3in]{fig5a.pdf}
\includegraphics[height=3.2in,width=3.4in]{fig5b.pdf}
\end{center}
\caption{\label{fig5}
Evolution of $\mu, 1-\eta, \mu-\Sigma, \alpha_{\rm M}, \alpha_{\rm B}$
(left) and $\delta_m'/\delta_m, -\Psi, \psi_{\rm eff}$ (right)
for $Q=0.02$ and the initial conditions same
as those used in Fig.~\ref{fig4}.
The perturbations correspond to the wavenumber $k=200H_0$
with $\sigma_8(z=0)=0.81$.
The gravitational potentials $-\Psi$ and $\psi_{\rm eff}$ are
normalized by their initial values.
In the right panel, we also show the evolution of $\delta_m'/\delta_m$
in the $\Lambda$CDM model as a dashed line.}
\end{figure}
The quantities associated with the stability of scalar perturbations are given by
\begin{equation}
q_s=\frac{M_{\rm pl}^2 F^2 (4x_1^2+4\Omega_{\phi 3}
+\Omega_{\phi 3}^2+4\sqrt{6}Q x_1 \Omega_{\phi 3})}
{2x_1^2}\,,\qquad
c_s^2=\frac{12x_1^2+4(2+\epsilon_{\phi})\Omega_{\phi 3}
-\Omega_{\phi 3}^2+4\sqrt{6}Q x_1 \Omega_{\phi 3}}
{3(4x_1^2+4\Omega_{\phi 3}+\Omega_{\phi 3}^2
+4\sqrt{6}Q x_1 \Omega_{\phi 3})}\,.
\label{qscsF}
\end{equation}
At the late cosmological epoch in which the contribution of
$\Omega_{\phi 3}$ is suppressed relative to the terms containing
$x_1$ in Eq.~(\ref{qscsF}), it follows that
$q_s \simeq 2M_{\rm pl}^2F^2$ and $c_s^2 \simeq 1$.
In the numerical simulations of Figs.~\ref{fig4} and \ref{fig5},
we confirmed that both $q_s$ and $c_s^2$ are
positive from the radiation era to today.
On using Eqs.~(\ref{Gefff2})-(\ref{Sigmaf2}) together
with the property $M_{\phi}^2=0$ and ignoring the contribution
$\Omega_{\phi 3}$ at late times, we obtain
\begin{equation}
\mu \simeq \frac{1}{F} \left( 1+2Q^2 \right)\,,\qquad
\eta \simeq \frac{1-2Q^2}{1+2Q^2}\,,\qquad
\Sigma \simeq \frac{1}{F}\,.
\end{equation}
For $Q>0$, the quantity $F=e^{-2Q(\phi-\phi_0)/M_{\rm pl}}$ is
smaller than 1 in the past ($\phi>\phi_0$).
Since $Q/\lambda<0$ for the existence of $\phi$MDE,
both $\mu$ and $\Sigma$ are larger than 1.
Since $Q^2=4 \times 10^{-4} \ll 1$ in the numerical simulation of Fig.~\ref{fig5},
the deviation of $\mu$ from 1 mostly arises from the term $1/F~(>1)$.
This property is different from that in $f(R)$ gravity, in which the large
coupling $Q^2=1/6$ gives rise to the strong gravitational interaction
with matter. Since $\mu-\Sigma \simeq 2Q^2/F$, the difference between
$\mu$ and $\Sigma$ is much smaller than 1, see Fig.~\ref{fig5}.
Let us estimate the evolution of $\delta_m$ during the $\phi$MDE.
Substituting $\mu \simeq (1+2Q^2)/F$ into Eq.~(\ref{delmeqf}), it
follows that
\begin{equation}
\frac{d^2 \delta_m}{d{\cal N}^2}+\frac{1}{2} \left( 1-3w_{\rm eff} \right)
\frac{d \delta_m}{d{\cal N}}-\frac{3}{2} \left(1 +2Q^2 \right)
\Omega_m \delta_m \simeq 0\,.
\label{delmap2}
\end{equation}
Using the values of $w_{\rm eff}$ and $\Omega_m$ given in Eq.~(\ref{phiMDE}),
the growing-mode solution to Eq.~(\ref{delmap2}) for $Q^2<1/2$ is
\begin{equation}
\delta_m \propto a^{(1+2Q^2)/(1-2Q^2)}\,.
\label{delmQ}
\end{equation}
In the limit $Q \to 0$, we recover the standard evolution $\delta_m \propto a$
during the matter era. Although the coupling $Q$ leads to the faster growth
of $\delta_m$, we recall that there exists the upper bound
$|Q|<{\cal O}(0.01)$ from (\ref{alpM}).
In the right panel of Fig.~\ref{fig5}, we plot the evolution of $\delta_m'/\delta_m$
for $Q=0.02$ together with that in the $\Lambda$CDM model.
{}From Eq.~(\ref{delmQ}), the value of $\delta_m'/\delta_m$ during the $\phi$MDE
is estimated as $\delta_m'/\delta_m \simeq 1.0016$, so it is hardly distinguishable
from that in the $\Lambda$CDM model.
This property can be confirmed in the numerical simulation of Fig.~\ref{fig5}
at high redshifts. Indeed, even at low redshifts,
the evolution of $\delta_m'/\delta_m$ is almost the same as
that in the $\Lambda$CDM model.
As we observe in the left panel of Fig.~\ref{fig5},
the deviation of $\mu$ from 1 is at most of
order $0.01$ and hence the growth of $\delta_m$
is hardly affected by the modification of gravity.
The gravitational potentials $-\Psi$ and $\psi_{\rm eff}$ also evolve in the similar
way to those in the $\Lambda$CDM model with a small gravitational slip, i.e.,
$1-\eta \simeq 4Q^2/(1+2Q^2) \ll 1$
for $|Q|<{\cal O}(0.01)$.
Although the model (D2) does not leave distinguished observational
signatures for the cosmic growth history, the background evolution of
$w_{\rm DE}$ can be clearly distinguished from
other (extended) Galileon models like (C2).
Finally, we should mention that, even in presence of the
quartic Galileon
term $G_4 \supset \beta_4 X^2$, there are cases in which the deviation of $c_t^2$ from 1
at low redshifts is so small that the bound (\ref{ctbound}) can be consistently satisfied \cite{KTD15}.
It remains to be seen whether the model with such a tiny value
of $|c_t^2-1|$ can survive in occurrence of further GW events
with electromagnetic counterparts.
\section{Conclusions}
\label{concludesec}
Two decades have passed after the first discovery of late-time cosmic acceleration,
but we did not identify the source for this phenomenon yet.
Although the cosmological constant $\Lambda$ is the simplest candidate
for dark energy, it is still a challenging problem to explain the very small value
of $\Lambda$ consistent with today's dark energy scale.
Moreover, in the $\Lambda$CDM model, the values of $H_0$ and
$\sigma_8 (z=0)$ constrained from the Planck CMB data have been in tension with
those measured in low-redshift observations.
If there are other possibilities to explain the cosmic acceleration,
it is worthwhile to construct theoretically consistent dark energy models
and explore their observational signatures to distinguish them from
the $\Lambda$CDM model.
In dynamical dark energy models, there are usually additional propagating
degrees of freedom to those appearing in standard model of particle physics.
A scalar field $\phi$, which is compatible with the isotropic and homogenous cosmological background, is one of the most natural candidates
for such a new degree of freedom.
If the scalar field has direct couplings to the gravity sector, Horndeski theories
are known as the most general scalar-tensor theories with second-order equations
of motion. As listed in Sec.~\ref{Hornsec}, Horndeski theories accommodate
a wide variety of dark energy models proposed in the literature.
In Sec.~\ref{backsec}, we obtained the background equations of motion in Horndeski
theories with a perfeclt fluid matter in the forms (\ref{back1}), (\ref{dH}), and (\ref{ddphi}),
which can be directly applied to concrete models of dark energy with a given Lagrangian.
We also computed the second-order action of tensor perturbations and derived the
speed of gravitational waves $c_t$ as Eq.~(\ref{ct}).
The gravitational wave event GW170817 together with the gamma-ray burst
GRB 170817A placed the tight observational bound on $c_t$.
If we strictly demand that $c_t^2=1$,
the Lagrangian of Horndeski theories is constrained to be
of the form (\ref{lagcon}). In extended versions of Horndeski gravity, such as
Gleyzes-Langlois-Piazza-Vernizzi (GLPV) theories \cite{GLPV} and
DHOST theories \cite{Langlois1,Langlois2,Crisostomi16},
there are models with $c_t^2=1$ even in presence of
quartic-order derivative
couplings \cite{GWcon1,GWcon2,GWcon3,GWcon4,GWcon5,GWcon6,Kase18}.
In this review, we did not explore the dark energy cosmology
beyond the domain of Horndeski theories.
In Sec.~\ref{scalarsec}, we derived the linear scalar perturbation equations of motion
in full Horndeski theories without fixing gauge conditions.
This gauge-ready formulation is versatile in that any convenient gauges can be
chosen depending on the problems at hand.
The perturbation equations expressed in terms of gauge-invariant variables are
directly applicable to the study of cosmic growth history (including initial
conditions of perturbations).
In Sec.~\ref{stasec}, we indentifid conditions for the absence of scalar ghost and
Laplacian instabilities in the small-scale limit by choosing three different gauges
and showed that they are given by Eqs.~(\ref{qscon}) and (\ref{csge})
independent of the gauge choices.
In Sec.~\ref{obsersec}, we introduced two quantities $\mu$ and $\Sigma$ associated
with the linear growth of nonrelativistic matter perturbations and the weak lensing
gravitational potential.
We computed them under the quasi-static approximation for the modes
deep inside the sound horizon, see Eqs.~(\ref{Gefff}) and (\ref{Sigmaf}).
Under the stability conditions of tensor and scalar perturbations,
the existence of scalar degree of freedom always enhances the gravitational
coupling with matter, such that $\mu>c_t^2/( 8\pi G q_t)$.
Taking the limit $c_t^2 \to 1$, we showed that $\mu$ and $\Sigma$ can be
expressed as Eqs.~(\ref{Gefff2}) and (\ref{Sigmaf2}) in terms of two parameters
$\alpha_{\rm M}$ and $\alpha_{\rm B}$ defined in Eq.~(\ref{nodim}).
We classified surviving theories of the late-time cosmic acceleration
into four classes according to these two parameters.
In Sec.~\ref{screensec}, we also showed that today's value of $\alpha_{\rm M}$
is constrained as $|\alpha_{\rm M}(t_0)|<0.02$ from the Lunar Laser Ranging
experiments measuring the variation of effective gravitational coupling
in screened environments. In particular, this information can be used to constrain
nonminimally coupled theories with the coupling $G_4(\phi)R$.
In Secs.~\ref{quinsec}-\ref{mostsec}, we reviewed observational signatures of four classes
of Horndeski theories with the exact value $c_t^2=1$ by paying particular attention to
the evolution of dark energy EOS and the cosmic growth
history. We summarize the main points in the following.
\begin{itemize}
\item Class (A): This is categorized as the theories with
$G_3=0$ and $G_4=M_{\rm pl}^2/2$.
In this case, we have $\alpha_{\rm B}=0=\alpha_{\rm M}$ and
hence $\mu=1=\Sigma$. Quintessence and k-essence belong to this class.
Depending on the evolution of $w_{\rm DE}$ (which is in the range $w_{\rm DE}>-1$
under stability conditions), we can classify quintessence into three subclasses:
(a) thawing, (b) scaling, and (c) tracking, see Fig.~\ref{fig1}.
The thawing quintessence is consistent with observations
for $w_{\rm DE}(z=0) \lesssim -0.7$.
The dark energy EOS in scaling quintessence realized by the potential
(\ref{scaling}) can be well approximated by Eq.~(\ref{wdetran})
for $\lambda_2^2 \ll 1$. The transition scale factor is constrained to be
$a_t<0.11$ (95 \%\,CL) to avoid the slow down of growth of structures
induced by the scaling scalar field with $c_s^2=1$.
In tracking quintessence with the inverse power-law potential (\ref{inpo}),
the power $p$ is constrained to be $p<0.17$ (95 \% CL) and hence positive
integers ($p \geq 1$) are excluded.
In k-essence, the ghost condensate model (\ref{Lagghost}) predicts the
evolution of $w_{\rm DE}$ consistent with observations for $\lambda \lesssim 0.36$.
The k-essence Lagrangian (\ref{unilag}) gives rise to the unified description of dark energy
and dark matter with the scalar sound speed $c_s^2$ much
smaller than 1.
\item Class (B): Compared to class (A), a nonminimal coupling of the form
$G_4(\phi)R$ is present in class (B).
In this case, there is the particular relation
$\alpha_{\rm B}=\alpha_{\rm M}/2 \neq 0$. The interaction between the scalar field and matter
enhances the gravitational coupling $\mu=G_{\rm eff}/G$, while this is not the case for
$\Sigma$. The $f(R)$ gravity and BD theory with the scalar potential belong to this class.
In $f(R)$ gravity, the models need to be constructed to suppress the propagation of fifth
forces in regions of the high density, in which case the deviation of $w_{\rm DE}$ from $-1$
occurs at low redshifts. After the perturbations enter the massless regime, the large coupling
$|Q|=1/\sqrt{6}$ leads to the strong amplification of matter density contrast and gravitational
potentials. For the consistency with CMB, BAO, and weak lensing data, the dark energy
EOS $w_{\rm DE}$ and the deviation parameter $B$ from the $\Lambda$CDM
model need to be in the ranges $|w_{\rm DE}+1|<0.002$ and $B(z=0)<0.006$ at 95 \% CL.
The similar tight constraints on the deviation from the $\Lambda$CDM
model also persist in BD theories with the potential (\ref{Vphige})
and the coupling
$|Q|>{\cal O}(0.1)$.
\item Class (C): In this class, the cubic coupling $G_3(\phi,X) \square \phi$
is added to the Lagrangian of class (A).
Since $\alpha_{\rm M}=0$, there is no gravitational
slip ($\eta=-\Phi/\Psi=1$) with $\mu=\Sigma>1$.
The bound (\ref{alpM}) from the Lunar Laser Ranging experiments is trivially satisfied.
In Sec.~\ref{classCsec}, we presented two models
of this class, including the cubic Galileon with a linear potential [model (C1)].
In model (C2) given by the Lagrangian (\ref{iii2}), the existence of an additional term $\beta_2 X^2$ to $\beta_1 X$
gives rise to a self-accelerating de Sitter attractor ($X={\rm constant}$).
If the dominance of this term over the cubic Galileon occurs
at a later cosmological epoch, the deviations of $w_{\rm DE}$ from $-1$
and $\mu$ from 1 tend to be more significant, see Fig.~\ref{fig2}.
The model (C2) allows an interesting possibility for realizing $w_{\rm DE}<-1$
without having ghosts, while modifying the large-scale structure growth at
low redshifts. We expect that the galaxy-ISW correlation data can provide
an upper bound on today's density parameter $\Omega_{\phi 3}$
of cubic Galileons.
\item Class (D): This class corresponds to the most general Lagrangian
in Horndeski theories with $c_t^2$ exactly equivalent to 1.
The difference from class (C) is that the nonminimal coupling
$G_4(\phi)R$ is present.
In model (D1) given by the Lagrangian (\ref{LagGBD}), the cubic
coupling $G_3(\phi,X)$ is the main source for the late-time cosmic acceleration,
but in such cases the model typically predicts
the galaxy-ISW anti-correlation which can be incompatible with
observational data. In model (D2), i.e., nonminimally coupled
cubic Galileons, there exists the $\phi$MDE followed by the
cosmic acceleration driven by the linear potential
$V(\phi)=m^3 \phi$. In this case, the cubic Galileon density
tends to be subdominant to the total field density at
early cosmological epochs, so the dark energy density
typically stays in the region $w_{\rm DE}>-1$.
{}From the bound (\ref{alpM}) the coupling $Q$ is constrained
to be $|Q|<{\cal O}(0.01)$, in which case the cosmic
growth pattern is not much different from that in GR.
However, it is possible to distinguish between model (D2) and
uncoupled model (C2) due to
different cosmic expansion histories and
galaxy-ISW correlations.
\end{itemize}
We have thus shown that the surviving dark energy models
after the GW170817 event predict different observational
signatures which can be probed in current and future
observational data.
The dawn of GW astronomies opened up new possibilities
for measuring today's Hubble parameter $H_0$ and
the luminosity distance of GWs \cite{GW170817,GWlu2}.
In the concordance cosmological model there has been a tension
of $H_0$ between CMB and low-redshift measurements,
so new GW constraints on $H_0$ may give us a hint
on whether we should go beyond the $\Lambda$CDM model
or not.
We also note that the GW luminosity distance is modified
by the nonminimal coupling $G_4(\phi)R$,
so the accumulation of GW events will provide important
information for the viability of models in classes (B) and (D).
The models in classes (A) and (C) can be distinguished further
by exploring whether the future observations favor
the region $w_{\rm DE}>-1$ or $w_{\rm DE}<-1$.
We hope that upcoming observational data will allow us
to distinguish between four classes of dark energy models
and that the origin of late-time cosmic acceleration is
eventually identified.
\section*{ACKNOWLEDGEMENTS}
We thank M.~Sami for the invitation to write this review.
We are grateful to Antonio De Felice and
Atsushi Naruko for useful discussions.
RK is supported by the Grant-in-Aid for Young Scientists B
of the JSPS No.\,17K14297.
ST is supported by the Grant-in-Aid for Scientific Research
Fund of the JSPS No.~16K05359 and
MEXT KAKENHI Grant-in-Aid for
Scientific Research on Innovative Areas ``Cosmic Acceleration'' (No.\,15H05890).
|
\section{Introduction}
The goal of Kinship Verification is to determine if two people are related and how (i.e., brothers, sisters, etc.).An automatic system that will verify the relation between two people can be beneficial in different areas. Such an automated system can help in finding the family of a known suspect; it can help determine the family of a lost child. In the field of biometrics, we can use such a concept in building a unified identity database per family and so on.\\
\begin{figure}[h]
\includegraphics[height=2.4in, width=3in]{openning.eps}
\caption{Illustration of the method proposed in the paper, using the face features from face recognition for the task of face verification. The weighted face features are self-learned for the task of kinship verification as will be explained in the paper}
\label{fig:main_idea}
\end{figure}
Although the potential for solving this problem and Although there is an increasing interest in the computer-science community for it, the progress made so far is limited and usually implemented in specific cases and hand-picked scenarios.
Recently, the most significant dataset for kinship verification was introduced \cite{robinson2016fiw} The dataset includes different types of kinship which serves as a resource for research, the dataset was released along with a challenge (Recognize Families In the Wild, FG2018) to build a classifier for different Kinship types automatically.\\
One of the promising methods for solving the Kinship problem is with using deep learning. Deep networks can learn a different representation of faces according to a specific task (i.e., face recognition, age estimation, etc.). When investigating the Kinship Verification task one can see that we do not have a theory for known features to be extracted from the faces for this task.
That is unlike the well studied Face Recognition task, where there are known features that were considered and proven to be accurate for building Face Recognition systems.
Furthermore, we can not point what makes us decide that two people are related, so it is impossible to code this knowledge or theory to an automated system.
The last justification was motivated us to use Deep Learning methods that can be thought by examples, and not by a theory to learn the features and classifier of Kinship-Verification problem.\\
Deep Learning methods can be implemented in different scenarios and setups.
In the training phase, we can distinguish between a relaxed scenario and a strict scenario, in the relaxed scenario we have information about each ID (i.e., tag ID for each photo).
The ID information can be used to build our dataset, (i.e., building more negative examples or even to fine tune our face recognition model to extract more accurate face features). In literature, it is also called the image-unrestricted scenario.
In the strict scenario, the dataset is composed of only having the examples of a Kin and non-kin photos. In literature, it is also called the image-restricted scenario.\\
We can also distinguish between relaxed and strict scenarios in the testing phase.
The relaxed scenario is where we have multiple photos per ID, and we want to jointly classify them as related to some other multiple other ID's photos.
The strict scenario is where we need to make a decision only on a pair of unique photos.
In our research, we challenged the strict in the notation as explained above. The illustration of our proposed method and concept can be seen in figure \ref{fig:main_idea}\\
We can summarize our main contribution as follows:
\renewcommand{\labelenumii}{\roman{enumi}}
\begin{enumerate}
\item We propose a novel method to make the network self-learn the needed features to complete the task of Face-Verification. Furthermore, we show that by learning those specific features we can reduce the number of parameters that the model uses by half with almost no reduction in performance.
\item We propose a local and global classifier to classify the different weighted features between two feature maps; we explain why this theory is needed in the task of Kinship Verification.
\item We train our model in the strict scenario of the Kinship-Verification problem and show that our trained model wins the RFIW2018 (FG2018) challenge.
\end{enumerate}
The rest of the paper is arranged as follows:
Section \ref{RELATED} will summarize the previous work on the area of kinship verification; Section \ref{ARCHITECTURE}, will describe in details the architecture we propose, focusing on the feature selection layer and the local-global classifier;
Section \ref{Dev} will include computation describing the proposed network, loss function, forward and backward calculations;
Section \ref{ImpDet} will include training information in details and technical information for training the parameterf our network.
Section \ref{Results}, we describe our results on RFIW's, and elaborate on showing the results for different setups of the proposed network;
Section \ref{CONCLUSION} will conclude and discuss future work.
\section{Related Work}\label{RELATED}
Kinship verification is a problem in the more general field of face verification and face recognition. The progress in face verification and face recognition can be divided to two main areas; the first area is the algorithm - creating more robust algorithms that describe the features of the face more distinctly. The second area is the data - getting access to a large scale of tagged data to develop more general and accurate face descriptors.\\
\textbf{Kinship datasets}: while describing the datasets that contribute to the progress of the research one can find: \cite{XIA}, there, kin-dataset was published (UB KinFace ver 2.0) the dataset consist with 600 images of parents and their children with varying ages. In\cite{Vieira2014}, a dataset of siblings - SibilingsDB was processed, the dataset includes pairs of better quality images corresponding to siblings. In \cite{FG2015Kinship}, a dataset of kinship images (KinFaceW-I, KinFaceW-II) with different family relation types (father-son, father-daughter, mother-son, mother-daughter) was introduced and was evaluated and tested as part of the FG-2015 challenge for kinship verification. In \cite{fang2010towards} the Cornell Database Group released a dataset of 150 images pairs of parents and children (i.e., F-S, F-D, M-S, M-D). In \cite{6909623} The Family101 dataset was introduced, It contains 101 families with 607 different individuals for the total of 14,816 images. Finally, as will be described on \ref{Results} the most extensive dataset of family members kinship images made public in \cite{robinson2016fiw}, and afterward has been expanded to three-generation families images in \cite{fiwpamiSI2018} this dataset was the ground for two kinship verification challenges, RFIW2017, RFIW2018(FG2018). The diversity of the latest can be seen in figure \ref{fig:three_gen}\\
\begin{figure}[h!]
\centering
\includegraphics[height=2.4in, width=3in]{three_gen_new.eps}
\caption{Example of a three generation family pictures from Families In the Wild dataset}
\label{fig:three_gen}
\end{figure}
\textbf{Algorithms}: in general, it is common to divide the face verification and recognition algorithms to \textbf{three} main approaches.
\begin{itemize}
\item \textbf{first}, engineered, hand selected features approach, where the features are engineered and tested from a database of faces and then picked to succeed in a specific task such as face verification.\\
Using this approach we can find \cite{fang2010towards} where different features such as LBP, HOG, and Gabor, were used to encode the faces in the dataset and then selected features was used to train KNN and Kernal SVM for kinship verification. Similarly, in \cite{Dandekar} used another classifier with the LBP as face descriptor. In \cite{Mathews} computed the HMM (Hidden Markov Model) with distance features between successive edges that derived from the DCT and Sobel operator. In \cite{FG2015Kinship,Qin} used SIFT face descriptor with sparse regularized regression in the problem of kinship verification to find the essential patches from the face images, those were then classified with SVM.
\item \textbf{second}, metric learning approach, where the idea is to select and learn features that can be similar for kin pairs and therefore have a higher score for similarity then non-kin pair.\\
As in \cite{KinFACEW}, Where a descriptor was learned to cast the two images to a space where a distance vector was minimized between a kin pair. In \cite{CMML} a Cross Model Metric Learning (CMML) approach was introduced, with the use of an asymmetric scheme, where each image is processed by a different deep net that is adjusted to the input, then the distance metric is being evaluated on the outputs of the two different deep nets. In\cite{Liris-7042} TSML (Triangular Similarity Metric Learning) was proposed with multiple dimensionalities reduced descriptors such as LBP and Fisher vectors. Finally, In \cite{Haibin} DMML(Discriminative Multimetric Learning) was proposed where different distance matrices are combined by learning the optimal weights per feature, each of those matrices is learned independently from different features.
\item \textbf{third}, Deep Networks approach, deep learning seems like a promising approach for describing and extracting the features needed for classification of kin pairs, due to the deep features it can extract as shown in \cite{DEEPID,imagenet,RESNET} and due to the fact that the features describing kin pairs are not explainable (opposed to the task of face recognition).\\
In \cite{Wang:2016} auto encoders were used to learn the representation of the difference between kin and non-kin pair, then classifying the mapping features for kin and non-kin pair. In \cite{6909623} the author introduced the idea to use deep neural encoders, to detect the most important facial features and use them as high-hierarchical features to classify between kin and non-kin pairs. Deep Learning approach was also developed by \cite{zhang12kinship}, there a softmax was used for classification and convolution layers for finding the features for representation of kin pairs.
\end{itemize}
\begin{figure}[H]
\includegraphics[height=1.6in, width=3in]{vgg16.eps}
\caption{Architecture of VGG-16}
\label{fig:vgg16}
\end{figure}
\section{Architecture Of SelfKin Network}\label{ARCHITECTURE}
In this section, we will start by describing our Architecture in general. Next, we will be focusing on each part of the Network and explain what is its rule in the solution of kinship verification.
\subsection{Architecture Overview}
The architecture of the SelfKin network composed of the idea to address a solution to two main challenges.
On the one hand, we would like that the representation (the features) of the images, fed to the network, will be as general as possible and not person-specific, so it will be possible to separate between kin and non-kin pairs in different poses, illuminations, etc.
On the other hand, we would require the representation to be flexible enough such in the face recognition task, so we will be able to learn an exact description of the faces in respect to some standard features of kin pairs.\\
As will be shown in the next subsections, the architecture of SelfKin network solves the need to compose features that were made for face recognition in a way that is more suitable for kinship verification.
As shown in \ref{fig:total_arc} our Architecture is constructed with four main blocks. Face recognition features, extracted per person, Self-weighted feature selection block, Local-Global classifier and Global averaging layer.
\subsection{Face Recognition Features}
During face recognition task, we represent each face as a set of features that can then be used to classify different identities.
It has been shown that those features are also invariant to different pose, illumination, gender, etc.
Those features are constructed based on all the face examples shown to the network.\\
Since those features are generalized across different faces and shown to be suitable to describe the face for a specific person, we will use them as descriptors for faces.
We will then need to compose those features in a way that will be possible to use them to classify kin and non-kin pairs.\\
For the face extract feature task, in the architecture that is proposed here, we used VGG-FACE \cite{Parkhi15}, VGG-FACE is a deep network with the architecture of VGG, composed of convolution, maxpool and fully connected layers, the architecture of this network can be seen in figure \ref{fig:vgg16}.
This network has been trained and evaluated for the face recognition task with a dataset of 2,622 celebrities containing a total of 982,803 images from the web, the dataset \cite{GBHuang}, is an open dataset for use.\\
To agree with the rules of the RFIW2018 it should be noticed that there are two different possible setups for the challenge, those setups influence the design of this network. The restrict setup - which means that the ID of the person is not known.
The unrestricted scenario where there is a label of ID for each photo.
In this work, we aim to solve the restrict mode of the challenge, so we did not fine-tune the VGG-FACE network with identities of the dataset.\\
The output of VGG-FACE is classification to one of the 2,622 identities, and the embedding layer size is a feature vector with 4,096 features.
For the task of kinship verification, we will extract the 4,096 features for each image in the questioned pair.
\subsection{Self-Weighted Features Layer}
The features extracted from the face images are features learned for the use of face recognition.
Those features are learned by tuning the network to recognize a specific person and ideally separate from one person to another.
This descriptor is not necessarily needed to classify kin and non-kin pairs and can be too informative and not generalize well for the kinship verification problem.\\
To overcome the un-generalization of the description for kinship verification we suggest inserting a self-adjusted weighting feature layer. This layer is learned during the task of kinship verification and aims to give high weight value to features that are needed for the task of classification and reduce the weights for the features that are not needed.\\
After training this layer we can threshold the weights and cancel features that do not contribute to the kinship verification task.
Not only it will help the network to focus on the right features it will also help to reduce the total number of weights the network has.
\begin{equation*}
\begin{split}
&{{X}_{\text{features}}} = {{x}_{1}},{{x}_{2}}...{{x}_{n}} \\
&{{X}_{\text{masked}}} = {{W}_{\text{mask}}}\cdot{{X}_{\text{features}}} = {{w}_{1}}{{x}_{1}},{{w}_{2}}{{x}_{2}}...{{w}_{n}}{{x}_{n}}
\end{split}
\end{equation*}
The weights of this layer $w_1\dots w_n$ are trained, end to end, with training the classifier of the network.
In order to preserve the symmetry of the problem, the weights of this network are shared between the pair of images.
When thresholding the weights, we choose a value of the threshold for all the weights in this layer and cancel the weights (e.g. assign the value of $0$) below this threshold.
The later leaves us with fewer features and fewer weights for the next layers.
\subsection{Local-Global Classifier}
Based on the weighted features gained from the face descriptor, we propose to develop a local and global classifier.
The local part of the classifier will convolve the matching features (e.g. by location) between the images, making a composed feature-map from the two separate feature-maps.
Mathematically, the local part of the classifier will take a nonlinear weighted sum of each two matching features to produce a new combined features map. An illustration of this part can be seen in figure \ref{fig:local_features}.
\begin{figure}[H]
\centering
\includegraphics[height=2.4in, width=3in]{local_features.eps}
\caption{Illustration of the local features convolution layer, each feature in the first image is weighted and summed with the corresponding weighted feature from the second picture. The resulting new feature is a nonlinear function of the above sum}
\label{fig:local_features}
\end{figure}
\begin{figure*}[t!]
\includegraphics[height=2.2in, width=6.2in]{full_arc_f.eps}
\caption{Network Architecture and Flow, the two images are input to the VGG-FACE, the output is face features that are then self weighted and classified}
\label{fig:total_arc}
\end{figure*}
As opposed to a "regular" convolution layer where the different weights of each filter are shared through the image, this proposed layer has local weights that are specific for each pair of features by there location. In this way, we can look at each of the new features as a composed nonlinear sum of the two original features. The weights of the nonlinear combination are adjusted with respect to the kinship verification task. To preserve the symmetry of the problem we are not adding a bias term to this layer.\\
By forcing the same one weight per original feature location, we are causing the network to effect the same way to all the same location features across the images during training.\\
As described above, the motivation to do this local convolution is to create one combined feature and train it in the task of kinship verification.
The local part of the classifier preserve the original size of the feature maps (e.g. if we had $4,096$ features per image, then the size of new feature map will also be $4,096$)\\
The global part of the classifier is a fully-connected layer that computes a nonlinear combination of the new fused feature map. This layer is aimed to make a combination of spatial features in the new feature map.
This part of the classifier decreases the size of the feature map to the size of the output of the fully connected layer, in our case, it will be $128$.
\subsection{Global Averaging Layer}
After applying the Local-Global classifier, the new feature map is a 128-features vector.
Following \cite{global_avg}, we use global averaging layer.
This layer helps to prevent overfitting the training data and minimizing the size of the feature-vector to classify it.
To correctly describe our classes (Kin, Non-Kin) we use two different sets of weights in the Local-Global classifier, one for each class. As a result, the outcome of the Local-Global classifier is two feature maps, one for each class.
To make a binary decision, we average the $128$ feature vector for each class and use softmax to classify it to a kin or non-kin pair.
\section{Derivative And Loss Function}\label{Dev}
In this section we compute the forward pass values and develop the derivative of the network, we will describe the loss function of the network.
\subsection{Forward Pass Computation}
The input to the network are two face images $224$ by $224$ pixels by $3$, we will assign them the notation $i\in (1,2)$, following the architecture is shown in figure \ref{fig:total_arc} the forward pass can be computed as:
\begin{equation*}
\begin{split}
&{{X}_{\text{features},i}} = {{x}_{1,i}}\ldots {{x}_{4096,i}}=\,VG{{G}_{16,\text{FACE}}}(\text{image}{_{i}})\\
&{{X}_{\text{masked},i}} = {{W}_{\text{mask}}}\centerdot {{X}_{\text{features},i}} = {{w}_{1}}{{x}_{1,i}}\ldots {{w}_{4096}}{{x}_{4096,i}}\\
&\phantom{\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,} i\in (1,2)\\
&{{X}_{\text{masked}}} =
\begin{bmatrix}
{{X}_{\text{masked},1}} \\
{{X}_{\text{masked},2}} \\
\end{bmatrix}\\
&X_{\text{local},j} = RELU({{W}_{\text{local},j}}{{X}_{\text{masked}}})\\
&\phantom{\,\,\,\,\,\,\,\,}= RELU({{W}_{\text{local},j,1}}{{X}_{\text{masked},1}}+{{W}_{\text{local},j,2}}{{X}_{\text{masked},2}})\\
&{X}_{\text{global},j} = RELU({{W}_{\text{global},j}}{{X}_{\text{local},j}})\\
&X_{\text{globavg},j} = \frac{1}{n}\sum\limits_{k}{{{X}_{\text{global},j,k}}}\\&\phantom{\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,} j\in (1,2)\\
&\phantom{\,\,}y=\text{softmax}({{X}_{\text{globavg},1}},{{X}_{\text{globavg},2}})
\end{split}
\end{equation*}
Where $X_\text{features}$ is given by the output of the face descriptors. $X_\text{masked}$ are the dot product of the weights $W_\text{mask}$ and the features.
It should be noted that, as explained above, there are two feature maps $X_local$ that are the convolution of the two original images feature maps. Those two new feature maps will eventually be computed for the probabilities of the two different classes for kin and non-kin.\\
The weights $W_\text{mask}$=$w_{\scriptscriptstyle{1}} \ldots w_{\scriptscriptstyle{4096}}$ are shared between both images and use as feature enhancement or suppression. the $W_\text{local}$ are weights that are shared between two image-specific features in order to create a convolution map from the two image's feature-maps.
\subsection{Loss Function}
The loss function of the network composed from the regularization part of the weights and the loss of the training examples computed by the softmax part.
Following \cite{FeatureSel} we are separating the regularization term for the mask layer and the regularization term for all the other weights in the network.
The total loss of the network is given by:
\begin{equation*}
\rm{Loss}_{{\rm{total}}} = \rm{Loss}{_{{\rm{reg,classifier}}}} + \rm{Loss}_{{{\rm{reg,mask}}}} + \rm{Loss}_{{{\rm{softmax}}}}
\end{equation*}
The regularization term for the weights in the local and the global part of the classifier is given by the $l^2$ norm, and can be expressed as:
\begin{equation*}
\begin{split}
\rm{Loss}_{\rm{reg,classifier}}& =\\
&-\lambda\left[{\sum\limits_{i=1}^{4096}{{w^2}_{local,i}}+\sum\limits_{k=1}^{4096\cdot128}{{w^2}_{global,k}}} \right]
\end{split}
\end{equation*}
Following \cite{FeatureSel}, the regularization term for the mask weights is computed by the $l^1$ norm and scaled by a factor proportional to the size of the layer:
\begin{equation*}
\rm{Loss}{_{\rm{reg,mask}}} = - \frac{\lambda }{d}\sum\limits_{t = 1}^{4096} {\left| {{w_t}} \right|}
\end{equation*}
Where $\lambda$ is the regularize value, and $d$ is the network size.
The loss function of the softmax, $\rm{Loss}_{\rm{softmax}}$, is given by the cross-entropy between the targets and computed outputs from the networks.
\subsection{Derivatives Computation}
For training the network we compute the derivatives of the different layers, the weights that need to be updated are:\\ ${{W}_{mask}},{{W}_{local,1}},{{W}_{local,2}},{{W}_{global,1}},{{W}_{global,2}}$. To simplify, we will omit the derivatives for the VGG network and the derivatives for the regularization terms.\\
Following the Architecture described above and using the notation as in figure \ref{fig:total_arc}, the backward-pass gradients of the network can be computed by:
\begin{equation*}
\frac{\partial{{L}_{\text{soft}}}}{\partial{{X}_{\text{globavg}{_{1}}}}} = {{{\hat{y}}}_{1}}-{{y}_{1}}
\end{equation*}
\begin{equation*}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{global}{_{1}}}}}=\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{globavg}{_{1}}}}}\frac{\partial {{X}_{\text{globavg}{_{1}}}}}{\partial {{X}_{\text{global}{_{1}}}}}=({{{\hat{y}}}_{1}}-{{y}_{1}})\left[ \frac{1}{n}..\frac{1}{n} \right]
\end{equation*}
\begin{equation*}
\begin{split}
&\frac{\partial {{L}_{\text{soft}}}}{\partial {{W}_{\text{global},1,l,m}}} = \frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{global}{_{1}}}}}\frac{\partial {{X}_{\text{global}{_{1}},(1\ldots t\ldots 128)}}}{\partial {{W}_{\text{global},1,l,m}}} \\ &\phantom{\,\,\,\,\,\,\,\,\,\,}=\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{global}{_{1}}}}}\left\{ \begin{matrix}
{{X}_{\text{local},1,{{m}_{if\,\left( {{W}_{\text{global},1,t}}{{X}_{\text{local},1}}>0 \right)}}}}\,\,\,l=t \\
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,l\ne t
\end{matrix} \right.
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1,l}}} = &\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{global}{_{1}}}}}\frac{\partial {{X}_{\text{global}{_{1}}(1\ldots t\ldots 128)}}}{\partial {{X}_{\text{local},1,l}}}\\ = &\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{global}{_{1}}}}}{{W}_{\text{global},1,t,{{l}_{if\,\left( {{W}_{\text{global},1,t}}{{X}_{\text{local},1}}>0 \right)}}}}
\end{split}
\end{equation*}
\begin{equation*}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{W}_{\text{local},1}}}=\left[ \frac{\partial {{L}_{\text{soft}}}}{\partial {{W}_{\text{local},1,i}}},\frac{\partial {{L}_{\text{soft}}}}{\partial {{W}_{\text{local},1,j}}} \right]
\end{equation*}
\begin{equation*}
\begin{split}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{W}_{\text{local},1,i,l}}} = &\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1}}}\frac{\partial {{X}_{\text{local},1,(1\ldots t\ldots 4096)}}}{\partial {{W}_{\text{local},1,i,l}}}\\ &\left\{ \begin{matrix}
{{X}_{\text{masked},i,{{l}_{if\,({{W}_{\text{local},1,t}}{{X}_{\text{masked}}}>0)}}}}\,\,\,\,\,l=t \\
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,l\ne t \\
\end{matrix} \right.\\
\frac{\partial {{L}_{\text{soft}}}}{\partial {{W}_{\text{local},1,j,l}}} = & \frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1}}}\frac{\partial {{X}_{\text{local},1,(1\ldots t\ldots 4096)}}}{\partial {{W}_{\text{local},1,j,l}}}\\ &\left\{ \begin{matrix}
{{X}_{\text{masked},j,{{l}_{if\,({{W}_{\text{local},1,t}}{{X}_{\text{masked}}}>0)}}}}\,\,\,\,\,l=t \\
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,l\ne t \\
\end{matrix} \right.
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
&\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{masked},l}}} = \left[ \begin{matrix}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{masked},i,l}}} \\
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{masked},j,l}}} \\
\end{matrix} \right] \\
= &\left[ \begin{matrix}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1}}}\frac{\partial {{X}_{\text{local},1,(1..t..4096)}}}{\partial {{X}_{\text{masked},i,l}}}+\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},2}}}\frac{\partial {{X}_{\text{local},2,(1..t..4096)}}}{\partial {{X}_{\text{masked},i,l}}} \\
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1}}}\frac{\partial {{X}_{\text{local},1,(1..t..4096)}}}{\partial {{X}_{\text{masked},j,l}}}+\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},2}}}\frac{\partial {{X}_{\text{local},2,(1..t..4096)}}}{\partial {{X}_{\text{masked},j,l}}} \\
\end{matrix} \right] \\
= &\left[ \begin{matrix}
\left\{ \begin{matrix}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1}}}{{W}_{\text{local},1,i,{{t}_{(')}}}}+\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},2}}}{{W}_{\text{local},2,i,{{t}_{('')}}}}l=t \\
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,l\ne t \\
\end{matrix} \right. \\
\left\{ \begin{matrix}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},1}}}{{W}_{\text{local},1,j,{{t}_{(')}}}}+\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{local},2}}}{{W}_{\text{local},2,j,{{t}_{('')}}}}l=t \\
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,l\ne t\\
\end{matrix} \right. \\
\end{matrix} \right] \\
&' =\rm{IF}\,({{W}_{\text{local},1,t}}{{X}_{\text{masked}}}>0) \\
&'' =\rm{IF}\,({{W}_{\text{local},2,t}}{{X}_{\text{masked}}}>0)
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
& \frac{\partial{{L}_{\text{soft}}}}{\partial{{W}_{\text{mask},l}}} =\frac{\partial{{L}_{\text{soft}}}}{\partial{{X}_{\text{masked},i}}}\frac{\partial{{X}_{\text{masked},i,(1\ldots t\ldots 4096)}}}{\partial{{W}_{\text{mask},l}}}\\
& \phantom{=\sum_{l=0}^{\infty}\,\,\,\,\,\,\,\,\,\,\,\Biggl(} +\frac{\partial{{L}_{\text{soft}}}}{\partial{{X}_{\text{masked},j}}}\frac{\partial{{X}_{\text{masked},j,(1\ldots t\ldots 4096)}}}{\partial{{W}_{\text{mask},l}}}\\
& = \Biggl\{ \begin{matrix}
\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{masked},i}}}{{X}_{\text{feature},i,t}}+\frac{\partial {{L}_{\text{soft}}}}{\partial {{X}_{\text{masked},j}}}{{X}_{\text{feature},j,t}}\,\,\,\,\,l=t\\
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,l\ne t
\end{matrix}
\end{split}
\end{equation*}
\section{Training Details}\label{ImpDet}
In this section, we will describe the different hyper-parameters that were chosen for training the network. we will continue with elaborate the different augmentation of the input data.
The pseudo-code for the training algorithm can be seen in algorithm \ref{alg:euclid}
\begin{algorithm}[H]
\caption{Training Algorithm}\label{alg:euclid}
\begin{flushleft}
\hspace*{\algorithmicindent} \textbf{Input: }\text{batch of image pairs}
\end{flushleft}
\begin{algorithmic}[1]
\Procedure{data augmentation}{}
\For{\text{im in images}}
\State $j\gets \text{rand(1,5)}$
\Switch{$j$}
\Case{$1$}
$\rm{im\_n}_i\gets \gamma(\rm{im},2)$
\EndCase
\Case{$2$}
$\rm{im\_n}_i\gets \gamma(\rm{im},\frac{1}{2})$
\EndCase
\Case{$3$}
$\rm{im\_n}_i\gets \rm{flip}(\rm{im})$
\EndCase
\Case{$4$}
$\rm{im\_n}_i\gets \rm{flip}(\rm{im}),\gamma(\rm{im},2)$
\EndCase
\Case{$5$}
$\rm{im\_n}_i\gets \rm{flip}(\rm{im}),\gamma(\rm{im},\frac{1}{2})$
\EndCase
\EndSwitch
\EndFor
\State \textbf{return} $\rm{im\_n}$
\EndProcedure
\Procedure{face descriptors}{}
\State $X_{\rm{features},(1,2)}\gets VGG_{\rm{FACE}}(\rm{image}_1,\rm{image}_2)$
\State \textbf{return} $X_{\rm{features},(1,2)}$
\EndProcedure
\Procedure{forward pass}{}
\State $X_{\rm{masked},(1,2)}\gets W_{\rm{mask}}\cdot X_{\rm{features},(1,2)}$
\State $X_{\rm{local},(1,2)}\gets\rm{conv} (X_{\rm{masked},(1,2)},W_{\rm{local}})$
\State $X_{\rm{global},(1,2)}\gets\rm{fc} (X_{\rm{local},(1,2)},W_{\rm{global}})$
\State $X_{\rm{average},(1,2)}\gets\frac{1}{n}\sum\limits_{k}{{{X}_{\rm{global},(1,2),k}}}$
\State $y\gets\rm{softmax}({{X}_{\rm{average},1}},{{X}_{\rm{average},2}})$
\State $\rm{L}_{\rm{soft}}\gets-{(y'\log(y) + (1 - y')\log(1 - y))}$
\State $\rm{L}_{\rm{class}}\gets-\lambda\left[l^2_{\rm{norm}}(W_{\rm{local}})+l^2_{\rm{norm}}(W_{\rm{global}})\right]$
\State $\rm{L}_{\rm{mask}}\gets-\frac{\lambda}{d}l^1_{\rm{norm}}(W_{\rm{mask}})$
\State $\rm{Loss}_{\rm{total}}\gets L_{\rm{soft}} + L_{\rm{class}} + L_{\rm{mask}}$
\State \textbf{return} $\rm{Loss}_{\rm{total}}$
\EndProcedure
\Procedure{backward pass}{}
\State \textbf{compute} $\frac{\partial {{L}_{\rm{total}}}}{\partial {{W}_{\rm{global}}}},\frac{\partial {{L}_{\rm{total}}}}{\partial {{W}_{\rm{local}}}},\frac{\partial {{L}_{\rm{total}}}}{\partial {{W}_{\rm{global}}}}$
\State ${W}_{mask,new}\gets{W}_{mask} + lr\frac{\partial {{L}_{\rm{total}}}}{\partial {{W}_{\rm{mask}}}}$
\State ${W}_{local,new}\gets{W}_{local} + lr\frac{\partial {{L}_{\rm{total}}}}{\partial {{W}_{\text{local}}}}$
\State ${W}_{global,new}\gets{W}_{global} + lr\frac{\partial {{L}_{\rm{total}}}}{\partial {{W}_{\rm{global}}}}$
\State \textbf{return} ${W}_{mask,new},{W}_{local,new},{W}_{global,new}$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\subsection{Network Parameters}
The face descriptor is implemented by the Keras version of VGG-FACE with depth $16$ as suggested in \cite{keras_vgg}.
VGG has an architecture of $16$ layers, composed of convolution layers, max pooling, and finally fully connected layers.
The output of this network is a classification for $1000$ classes. We used the output of the convolution layers (i.e., the $4096$ embeddings).
The VGG-FACE was pre-trained with one million celebrities dataset.\\
As described in \ref{Dev}, we proposed two regularization terms, one for the self-weighting layer, and the second for the local-global classifier.
For the first regularization term we used $l^1$ regularization with $\lambda=0.5$ and $d=4,096$.
For the second regularization term we used $l^2$ regularization with $\lambda=1e-5$\\
We Trained our network using Adam optimizer with a learning rate of $1e-4$ to $1e-5$ and $beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0$.
We used dropout layer with $p=0.8$ between the local part and the global part of the classifier.\\
For choosing those hyper-parameters we trained the different classifiers using the training set and validate it with the validation set, we stopped the training when the validation accuracy stopped raising. An example for the training results and the resulting mask layer can be seen in figure \ref{fig:loss_acc}. For the Final submission, we trained our model with the train and validation set.\\
We initialize the weights for the different layers of the network with random numbers. And initialize all the weights of the mask layer as $1$ since we want the network to self-adjust the weights for each of the features.
\begin{figure}[h]
\includegraphics[height=2.6in, width=3in]{loss_acc_weights.eps}
\caption{Example of the training results for sisters verification, upper left is the training accuracy, upper right is the training loss and in the bottom is the resulting masking layer}
\label{fig:loss_acc}
\end{figure}
\subsection{Data Augmentation}
Although the Data that was released for the RFIW2018 competition is the most massive kinship data available, still, it is proven that augmenting the face data can improve the performance of face recognition task. Further, when investigating the RFIW2018 data one can see the diversity in the face images, such as, pose, illumination, color, etc. This diversity can help to train a better and general network.
An example of this diversity can be seen in figure \ref{fig:diversity_dat}.\\
For each image we randomly select the augmentation method it will pass:
\renewcommand{\labelenumii}{\roman{enumi}}
\begin{enumerate}
\item randomly change the gamma factor of the image for different illumination, scale between $\gamma\in(0.5,1,2)$.
\item horizontal flipping the image.
\end{enumerate}
An augmentation for a single image can be any combination of one or more of the listed above.
\begin{figure}[h]
\includegraphics[height=2.6in, width=3.2in]{diversity_dat.eps}
\caption{Example for the diversity of the RFIW dataset in color, pose and illumination}
\label{fig:diversity_dat}
\end{figure}
\section{Experimental Result}
\label{Results}
In this section, we describe our result based on the challenge of RFIW2018.
We start by describing the dataset that was used. We then continue with describing our results on the different cases that ware trained. We conclude our work and suggesting new areas for research.
\begin{table}[h]
\caption{FIW Dataset Description and Details}
\label{tab:dataset}
\begin{tabular}{lccc}
\toprule
Type/Phase & Train & Validation & Test \\
\midrule
Brothers & 29,812 & 55,546 & 18,196 \\
Sisters & 19,778 & 35,024 & 4,796 \\
Siblings & 28,428 & 15,422 & 9,716 \\
Father-Daughter & 41,604 & 35,238 & 15,040 \\
Father-Son & 64,826 & 44,870 & 18,166 \\
Mother-Daughter & 39,110 & 29,012 & 14,394 \\
Mother-Son & 66,464 & 31,094 & 14,806 \\
Grfather-Grdaughter & 1,478 & 4,846 & 838 \\
Grfather-Grson & 1,388 & 1,926 & 1,588 \\
Grmother-Grdaughter & 1,580 & 3,768 & 952 \\
Grmother-Grson & 1,284 & 1,844 & 1,470 \\
\bottomrule
Total & 295,752 & 258,590 & 99,962 \\
\end{tabular}
\end{table}
\begin{table*}[t]
\centering
\caption{Results for different models trained with self-adjust weights layer and without}
\label{tab:Results}
\begin{adjustbox}{width=1\textwidth}
\begin{tabular}{l cccccccccccc}
\toprule
Model/Relation & BB & SS & SIBS & FD & FS & MD & MS & GFGD & GFGS & GMGD & GMGS & Average \\
\midrule
With Self-adjusted & 0.6914 & 0.7725 & 0.6864 & 0.6891 & 0.6889 & 0.7378 & 0.7085 & 0.6193 & 0.6366 & 0.6386 & 0.6319 & 0.6819 \\
With Self-adjusted t.h. at half & 0.6946 & 0.7733 & 0.6876 & 0.6890 & 0.6846 & 0.7389 & 0.7026 & 0.6229 & 0.6234 & 0.6344 & 0.6285 & 0.6800 \\
W/O Seld-adjusted & 0.7060 & 0.7723 & 0.6865 & 0.6232 & 0.6961 & 0.7348 & 0.7172 & 0.5763 & 0.6114 & 0.6523 & 0.6346 & 0.6737 \\\hline
\end{tabular}
\end{adjustbox}
\end{table*}
\begin{table*}[h]
\centering
\caption{Results from the challenge site}
\label{tab:SiteRes}
\begin{adjustbox}{width=1\textwidth}
\begin{tabular}{l cccccccccccc}
\toprule
Name/ Relation & MD & MS & SS & BB & SIBS & GMGD & GMGS & FS & GFGS & FD & GFGD & Average\\
\midrule
eranda-ours & 0.73794 & 0.70856 & 0.77251 & 0.69147 & 0.68649 & 0.63865 & 0.63265 & 0.68897 & 0.63665 & 0.68922 & 0.61933 & 0.68204 \\
Unknown & 0.66625 & 0.59840 & 0.72852 & 0.66256 & 0.62968 & 0.57563 & 0.61088 & 0.61527 & 0.56927 & 0.63304 & 0.57995 & 0.62449 \\
Unknown & 0.65888 & 0.60468 & 0.70204 & 0.68443 & 0.61620 & 0.56932 & 0.62789 & 0.61224 & 0.56360 & 0.63071 & 0.59427 & 0.62402 \\
Unknown & 0.59552 & 0.57814 & 0.64678 & 0.57221 & 0.57400 & 0.55357 & 0.57687 & 0.57420 & 0.55667 & 0.59355 & 0.55966 & 0.58011 \\
Unknown & 0.61553 & 0.60482 & 0.68015 & 0.62700 & 0.59016 & 0.54516 & 0.52176 & 0.58152 & 0.53274 & 0.58211 & 0.57875 & 0.58725 \\
Unknown & 0.56704 & 0.54086 & 0.60196 & 0.51319 & 0.55393 & 0.52416 & 0.50204 & 0.54123 & 0.52455 & 0.55957 & 0.55131 & 0.54362 \\\hline
\end{tabular}
\end{adjustbox}
\end{table*}
\subsection{Recognize Families in the Wild 2018}
The first track of Recognize Families In the Wild 2018 (RFIW2018) challenge deals with solving a classifier for kinship verification to $11$ types of relationships as shown in \ref{tab:dataset}.
The FIW (families in the wild) dataset contains images of $1,000$ different family-cells, for a total of $11,932$ images. The images are divided into $11$ family relations as mentioned above and arranged as positive and negative examples.
Each one of the family members has a variant amount of images from different ages. The total amount of pairs (positive and negative) is $654,304$. \\
The images also contain meta-data for gender. The images are characterized with a different background, face-color, size of the face in the images, pose and expressions.
The FIW dataset is the only dataset that also contains face images and relations for grandfather, grandmother/grandson, granddaughter.\\
The competition lasted for a defined period and was divided into different phases, in the first phase, training phase, the training data is released, then in the validation phase, the model trained is validated using a new validation data. Finally, in the test phase, the model is tested with unseen test data.
\subsection{Test Results and Discussion}
During training we optimized our architecture in the sense of choosing the different hyper-parameters for regularization terms as described in \ref{ImpDet}, choosing the size for the fully connected layers and selecting the dropout probability.
In this section, we will show the results of the optimization. We can distinguish the results from three experiments, training, and testing with the self-adjust layer, without it, and with threshold the self-adjusted layer by half.\\
\textbf{Without self-adjusted layer} - In this case, we trained our network with only updating the weights for the local-global classifier, and tested the test data with the networked trained.\\
\textbf{With self-adjusted layer} - In this case we trained our model, end to end, with the self-adjusted layer, allowing it to change the impact of each feature according to its contribution to the kinship-verification task. The weights of the local-global classifier also were modified during training.
When testing the network, we used the trained weights for the self-adjusted layer.\\
\textbf{With self-adjusted layer threshold at half} - In this case, we did the same training procedure as in with self-adjusted layer, while in the testing phase, we chose a threshold value that divides the features to half(median). we used this threshold to cancel all the features that were below the threshold and with this, also canceling all the weights that were connected to those original features.
The results of the above description are shown in table \ref{tab:Results}.\\
We show that there is no significant change (about $1$ percent on average) between the two models trained.
We saw that the potential of this self-adjusted layer is that by training this layer, we have the possibility to reduce the number of features by a factor of $2$ while not decreasing the performance of the network, and on average, the network is still better than a network that was not trained with this selecting layer.
It is noticeable that the average performance is reduced because of the poor performance of the GFGD, GFGS, GMGD, GMGS pairs which come with coherence to the lack of data for those pairs.
We conclude with presenting table \ref{tab:SiteRes} - the leader bored from RFIW challenge.
\section{CONCLUSIONS AND FUTURE WORKS}
\label{CONCLUSION}
In this work, we aimed to solve the Kinship Verification problem with the novel method of a deep learning network.
We presented an architecture that purposes to compete for the strict usage of kinship verification, without having the ID's of the people in the dataset. Moreover, not having the specific family classification for each photo, so we do not have the knowledge during training to select negative or positive examples that were not being given in the training set.\\
We presented a layer that self-adjust weights of face features for the task of kinship verification.
We show that we can use the self-adjusted weights to reduce the number of weights used in the network without the loss of performance.\\
We introduced the use of local and global classification, explaining why it is needed in combining two feature maps created for face recognition, we developed the derivatives for the network and this classifier as well.
Finally, we showed that our model wins the RFIW2018 (FG2018) challenge.\\
For future work, one can try using a more general face feature extractor with different architecture then VGG-16. one can also try to train the network with the relax scenario presented in the competition (use the ID's of people in the dataset for refining the face feature extractor or for creating more sophisticated examples from the datast)
it is also optional to use the self-adjusted weights layer, with a different purpose such as making the model invariant to age or color diversity in the dataset.
\FloatBarrier
\bibliographystyle{unsrt}
\bibliographystyle{ieeetr}
\nocite{*} |
\section{Introduction}
\label{sec:intro}
The solar system is composed of a variety of objects with different sizes: the Sun, the planets and their moons, other small bodies including dwarf planets, asteroids and comets, and \revise{20180710}{dust grains in interplanetary space}. The size of the small bodies in the solar system covers a wide dynamic range from 1,000\,km (\textit{e.g.}, Ceres) down to about 10\,nm ($\beta$-asteroids). The size distribution of the small bodies provides important information to understand the origin and the evolution of the solar system.
The amount of the interplanetary materials incoming the Earth is about $10^5\,\mathrm{kg\,day^{-1}}$. The particles in the mass range of $10^{-9}$--$10^{-2}\,$g are the major contributors of the incoming mass flux \citep{grun_collisional_1985}. Properties of such small particles are essential to understand what kind of materials fall into the Earth.
Particles smaller than about $10^{-6}\,$g \revise{20180710}{have been detected in place with dust counters on satellites and spacecrafts} \citep{gruen_ulysses_1992,grun_galileo_1992,gurnett_micron-sized_1997,szalay_student_2013}, while the number density of the large particles is so small that their size distribution is difficult to measure by in-situ observations.
The small particles incoming the Earth are observed as meteors, interacting with the Earth's atmosphere and \revise{20180710}{convert part of their kinetic energy into emission lines} \citep{baldwin_ablation_1971,ceplecha_meteor_1998,hawkes_quantitative_1975,jones_effects_1966,popova_meteoroid_2004}. Thus, the mass of the incoming particle can be derived from measuring the brightness of the meteor with an assumption of an energy conversion efficiency. The Earth works as an extremely wide-area \revise{20180710}{dust} detector, providing an indirect measurement of the size distribution of \revise{20180710}{interplanetary dust} around the Earth.
Meteors in the range between $-5$ and $+4\,\mathrm{mag}$. have been detected by large meteor survey networks, for example, Cameras for Allsky Meteor Surveillance (CAMS), and SonotaCo network \citep{jenniskens_meteor_2017}. CAMS obtained more than 135,000 meteor orbits in 2017. In this magnitude range, about 36\% of meteors belong to some meteor showers \citep{jenniskens_cams_2016-2}. The luminosity function of meteors in this magnitude range was intensively investigated by \citet{hawkins_influx_1958} based on a large survey with Super-Schmidt cameras. They showed that the meteor luminosity function was well-approximated by a power-law function. The meteors detected by these surveys, however, are mostly caused by particles of larger than ${\sim}1\,\mathrm{g}$. Observations of meteors fainter than $+5\,\mathrm{mag}$ are required to investigate the size distribution of \revise{20180710}{interplanetary dust}.
\citet{cook_flux_1980} obtained more than 2,000 sporadic meteors in four nights using a 10-m optical reflector and a photo multiplier. The photographic magnitudes of the detected meteors ranged between $+7$ and $+12\,\mathrm{mag}$. Comparing with literature, they concluded that the meteor luminosity function was well approximated by a power-law function between $-2.4$ to $+12.0\,\mathrm{mag}$. They observed meteors as light pulses since they used the photo multiplier instead of a camera. Thus, the information on the meteor trajectories was not obtained and the atmospheric correction to compensate a possibly unstable weather condition was not applied in real time. \citet{clifton_television_1973} and \citet{hawkes_television_1975} observed faint meteors down to about $8$ and $6\,\mathrm{mag}$, respectively, in television observations. Imaging observations of faint meteors using a wide-field video camera are favored to examine the meteor luminosity function in detail.
Radar observations detect meteors caused by particles in the mass range of about $10^{-6}$--$10^{-2}\,$g. The Canadian Meteor Orbit Radar has detected more than 3,000,000 meteors down to $10^{-4}$\,g in seven years \citep{brown_meteoroid_2008,brown_meteoroid_2010}. Large aperture radars can observe meteor head echoes. \revise{20180710}{Interplanetary dust grains} of about $10^{-6}\,g$ can be detected by these radars. \citet{kero_2009-2010_2012} obtained more than 100,000 meteors using the Shigaraki Middle and Upper atmosphere radar in Japan. Both the trajectories and the radar cross-sections of meteors were obtained simultaneously. This is a great advantage of radar head echo observations. On the other hand, the radar cross-sections are not easy to convert to the brightness of the meteors, or the mass of the meteoroids. Still, there is a need for optical observations of faint meteors.
The \revise{20180710}{interplanetary dust grains} in the mass range of $10^{-6}$--$10^{-3}\,\mathrm{g}$ correspond to about $+7$--$14\,\mathrm{mag}$ in optical observation. To obtain images of such faint meteors, we need a large photon-collecting mirror, a wide-field optics, and a high-sensitive video camera \citep{kresakova_activity_1955,poleski_1996-2001_2008}. \revise{20180710}{There are only a handful of studies on meteors detected with large telescopes. \citet{pawlowski_leonid_1998} detected 151 Leonid meteors down to about $13\,\mathrm{mag}$ with the Liquid Mirror Telescope in Cloudcroft Observatory, to constrain the mass distribution index of the Leonid shower. \citet{iye_suprimecam_2007} serendipitously detected sporadic meteors with \textit{Subaru}/SuprimeCam and constrained the diameter of the collisionally excited tubes caused by the meteors. \citet{bektesevic_linear_2017} developed a new method to discover faint meteors obtained in large optical surveys, such as the Sloan Digital Sky Survey. Currently, no facility conducts regular observations of faint meteors with a telescope larger than $1\,\mathrm{m}$.}
We carried out observations of faint meteors with a wide-field mosaic CMOS camera \tomoepm in 2016. This paper summarizes the results of the observations. This paper is organized as follows: details of observations are summarized in Seciton~\ref{sec:observation}; data reduction and observed meteors are presented in Section~\ref{sec:result}; the luminosity function of faint meteors is discussed in Section~\ref{sec:discussion}; we summarize the paper in Section~\ref{sec:conclusion}.
\section{Observations}
\label{sec:observation}
\subsection{\tomoepm}
\label{sec:tomoe-pm}
\tomoe is a wide-field camera being developed in Kiso observatory, the University of Tokyo, which will be the world largest video camera for astronomy. \tomoe, mounted on the 105-cm Kiso Schmidt Telescope\footnote{Kiso Observatory is located in $35\ifmmode^\circ\else$^\circ$\fi 47\ifmmode'\else$'$\fi 38.7\ifmmode''\else$''$\fi$\,N and $137\ifmmode^\circ\else$^\circ$\fi 37\ifmmode'\else$'$\fi 42.2\ifmmode''\else$''$\fi$\,E, and at the altitude of $1130\,\mathrm{m}$.} in Kiso observatory, will continuously monitor a \revise{20180710}{$20\,\mathrm{deg^2}$} area at 2\,Hz with 84 CMOS sensors developed by Canon Inc. The designed pixel scale is about \revise{20180710}{$1.19{\ifmmode''\else$''$\fi}{\times}1.19{\ifmmode''\else$''$\fi}$}. The limiting magnitude is estimated to be about $18.5\,\mathrm{mag}$. in the $V$-band in the 2\,Hz observation. The limiting magnitude for meteors will be about $13\,\mathrm{mag}$. in the $V$-band, on the assumption that meteors moves at \revise{20180710}{$10\ifmmode^\circ\else$^\circ$\fi \mathrm{s}^{-1}$, which corresponds to the angular velocity of a meteor at zenith, at 100\,km above an observer, at a velocity of $36\,\mathrm{km\,s^{-1}}$, and at an incident angle of $30\deg$}.
As a pilot project, we developed a prototype of \tomoe (\textit{hereafter}, \tomoepm), equipped with the 8 CMOS sensors \citep{sako_development_2016}. \tomoepm is mounted on the 105\,cm Schmidt telescope at Kiso Observatory, the University of Tokyo. The pixel scale is about \revise{20180710}{$1.19{\ifmmode''\else$''$\fi}{\times}1.19{\ifmmode''\else$''$\fi}$}. The detectors are aligned in a line along the direction of the right ascension with some gaps (\textit{see}, \citet{sako_development_2016}). The total area of the field-of-view is about \revise{20180710}{$1.98\,\mathrm{deg^2}$}. All the detectors are synchronously operated by the same control signal. The maximum frame rate is 2\,Hz. The overhead time due to readout is almost negligible thanks to rolling shutter. A NTP synchronized time is recorded as a time-stamp in a FITS header, although the time-stamp is not synchronized with a shutter timing. The time-stamp is as accurate as ${\sim}1\,\mathrm{s}$. The size of the imaging area is $2000{\times}1128$ pixels. The experimental observations were successfully completed, to confirm that \tomoepm achieved the designed sensitivity.
\subsection{Observations}
\label{sec:observations}
Meteor observations with \tomoepm were carried out on April 11 and 14, 2016 (UT). Details of the observations are summarized in Table~\ref{tab:obslog}. The sky was dark and clear on April 11, while part of the observations were \revise{20180710}{affected by} clouds on April 14.
Meteors are detected as a streak in the 2\,Hz observation, while artificial satellites and space debris are also detected as streaks. To eliminate these contaminations, the observations were conducted in UT 12--18, ${\pm}3\,$hours around midnight in JST, and the observed regions were set within the Earth's shadow. Since the shadow moves during a night, the observed regions were relocated in every two hours.
The captured movie data were compiled into FITS data cubes in every 3 minutes (\textit{hereafter}, an observation unit). Thus, each FITS data cube was composed of 360 frames. In total, 808 observation units (290,880 frames) were obtained on April 11, while 880 observation units (316,800 frames) were obtained on April 14.
\begin{table*}[tp]
\centering
\caption{Observation Logs}
\label{tab:obslog}
\begin{tabular}{lrr}
\hline\hline
Date
& 2016-04-11 & 2016-04-14 \\
Weather Condition
& Clear Sky & Partly Clouded \\
Lunar Age
& 3.6 & 6.6 \\
Elevation Range
& $13.7\ifmmode^\circ\else$^\circ$\fi$--$37.9\ifmmode^\circ\else$^\circ$\fi$ & $10.1\ifmmode^\circ\else$^\circ$\fi$--$37.1\ifmmode^\circ\else$^\circ$\fi$ \\
Typical Zero Magnitude
& $25.5\,\mathrm{mag}$ & $25.0\,\mathrm{mag}$ \\
Total Observing Time
& ${\sim}5.1\,$hours & ${\sim}5.5\,$hours \\
Number of Exposures
& 290,880 & 316,800 \\
Number of Detected Meteors
& 1,514 & 706 \\
Apparent Magnitudes\tablenotemark{\dagger}
& $4.5$--$12.5\,\mathrm{mag}$ & $5.5$--$11.0\,\mathrm{mag}$ \\
\hline
\end{tabular}
\tablenotetext{\dagger}{Video rate magnitudes defined in \citet{iye_suprimecam_2007}.}
\end{table*}
\section{Results}
\label{sec:result}
\subsection{Data Reduction}
\label{sec:reduction}
Streak-like signals were extracted from the movie data cube with the slant-and-compress algorithm developed by \citet{ohsawa_development_2016}. First, background emission was subtracted and stellar signals were masked. A frame was slanted by a certain angle and compressed into a one-dimensional array with the normalization such that the noise level should be uniform. A set of one-dimensional arrays were stacked into a two-dimensional array. The streak-like signals in the original image were converted as compact sources in the converted array. The signal-to-noise ratio of faint meteors in the original image was enhanced by about $\sqrt{n}$, where $n$ was the number of pixels along the streak, by integrating signals along with the streak. Refer to \citet{ohsawa_development_2016} for details of the detection algorithm.
A number of meteor candidates were detected by the algorithm. Then, genuine meteor events were counted by humans. The numbers of the real meteor events were 2002 and 926 on April 11 and 14, respectively. Some meteors were doubly counted due to the rolling shutter or passing through multiple detectors. Those duplicated events were merged. Finally, 1514 and 706 events were extracted as the individual meteor events detected on April 11 and 14, respectively. \revise{20180710}{Although no prominent meteor shower activity was predicted at that time, there could be some contributions from April Lyrids and $\eta$-Aquariids. The separation angles from the radiant points of these showers were calculated. The fractions of the meteors which might pass through within 3\ifmmode^\circ\else$^\circ$\fi of the radiants were at most 6\% and 3\% for April Lyrids and $\eta$-Aquariids, respectively. These fractions were consistent with the number of sporadic meteors randomly coming from the radiants. We concluded that the contribution from meteor showers was negligible and the detected meteors were attributed to sporadic meteors.}
The time variations in the number of the meteors on April 11 and 14 are illustrated in the top and bottom panels of Figure~\ref{fig:evrates}, respectively. The bar charts show the numbers of the meteors in the observation unit. The circles below the bar charts illustrates the time variations in the magnitude zero-point averaged over the eight detectors, indicating weather conditions in the observation unit. The small number of the detections on April 14 was simply attributed to the poor weather conditions.
\begin{figure*}[p]
\centering
\centering\includegraphics[width=.9\linewidth]{f001.pdf}
\centering\includegraphics[width=.9\linewidth]{f002.pdf}
\caption{Time variation in the number of meteors and the magnitude zero-point. The periods of cloudy weather are indicated by the green hatched regions. The gray cross-hatched regions mean no data due to pointings or troubles.}
\label{fig:evrates}
\end{figure*}
The brightness of the meteors was measured as follows. The projected profile of the meteor was obtained by averaging the signal along with the streak. The projected profile was broad and sometimes double-peaked since meteors were usually defocused \citep{bektesevic_linear_2018}. The line-intensity of the meteor, $\tilde{I}$ in units of $\mathrm{ADU\,pix^{-1}}$, was measured by fitting the projected profile with a combination of Gaussian profiles. Since the angular velocity of the meteors were not measured by \tomoepm observations, the line-intensity $\tilde{I}$ was not directly compared with the intensities of nearby stars. Thus, we adopted the video-rate magnitude defined in \citet{iye_suprimecam_2007}. Assuming that all the meteors moved at \revise{20180710}{an angular velocity of $10\ifmmode^\circ\else$^\circ$\fi s^{-1}$}, the travel distance of the meteors in $0.5\,$s was estimated to be $1.5{\times}10^3$ pixels, where $\theta$ was the pixel scale of \tomoepm in units of arcsecond. The total amount of the signals emitted by the meteor in $0.5\,$s was obtained by ${I} = 1.5{\times}10^3\tilde{I}\,\mathrm{ADU}$, which was transformed into the $V$-band magnitude by comparing with the intensities of nearby stars. The ranges of the estimated magnitudes are listed in Table~\ref{tab:obslog}. The meteors as faint as about $12.5\,\mathrm{mag}$ in the $V$-band were detected on April 11. Although the weather condition on April 14 was poorer than that on April 11, the meteors as faint as about $11.0\,\mathrm{mag}$ were successfully detected.
\subsection{System Efficiency}
\label{sec:syseff}
\begin{figure*}
\centering
\centering\includegraphics[width=.9\linewidth]{f003.pdf}
\caption{Definition of the meteor detectable volume. The blue dot is the location of a telescope, which is observing a sky at the elevation $\alpha$. The thin blue region indicates the meteor detectable layer. The blue cross-hatched region indicates the meteor detectable volume. The thick blue region is the test region, where meteors are randomly generated. A generated meteor with the incident angle $\theta$ is indicated by the red arrow.}
\label{fig:schematic-collect}
\end{figure*}
The meteor-collecting power of a observing system is usually evaluated in units of area. \citet{koschack_determination_1990} evaluated the effective meteor-collecting power from the area of the field-of-view projected onto a meteor level at an altitude of $100\,\mathrm{km}$ weighted by distance. In observation with naked eyes (visual observation), a field-of-view was as large as about $52.5\ifmmode^\circ\else$^\circ$\fi$ \citep{koschack_determination_1990}. The estimated effective meteor-collecting area (\textit{hereafter}, EMCA) in visual observation is $24,400\,\mathrm{km^2}$ when observing at zenith, whereas that area becomes $20,770\,\mathrm{km^2}$ when observing with an elevation angle of $50\ifmmode^\circ\else$^\circ$\fi$, in the case that a meteor population index\footnote{The number of meteors brighter than a magnitude $m$ in visible wavelengths should follow $N({<}m) \propto r^m$.} $r$ is 3.5. In Koschack's calculation, the region where meteors are observable is approximated by a thin spherical shell. Hereafter, Koschack's formalism is referred as to the thin shell approximation.
In meteor observations with telescopes, the field-of-view is usually smaller than 10\ifmmode^\circ\else$^\circ$\fi. In such cases, the thin shell approximation is not valid. \citet{kresakova_activity_1955} evaluated the EMCA by statistically estimating the averaged angular length of meteors. This method requires a large number of meteors. Since the efficiency in meteor detection with \tomoepm changes with elevation and \tomoepm has the multiple image sensors, their method is not applicable to our observations. Thus, we evaluated the EMCA of \tomoepm by a Monte Carlo simulation.
Figure~\ref{fig:schematic-collect} schematically describes the configuration of calculation. When \revise{20180710}{dust grains} penetrate into the upper atmosphere, they are heated by interaction and become bright in optical wavelengths at a certain altitude, observed as meteors. Then, when the \revise{20180710}{grain} penetrates into a lower atmospheric layer, they cease to be bright. Here, we define an upper and lower limit of an altitude where meteors are observable in optical wavelengths by $H_2$ and $H_1$, respectively. Thus, the region where meteors are observable in optical wavelengths is approximated by a spherical shell with a depth of $H_2{-}H_1$ (\textit{hereafter}, meteor shell). In Figure~\ref{fig:schematic-collect}, the meteor shell is indicated by the blue shaded region.
The region captured by a camera with a single rectangular field-of-view is defined by
\begin{equation}
\label{eq:mov}
\vec{x} \in \left\{~
c_0\vec{n}_0+c_1\vec{n}_1+c_2\vec{n}_2+c_3\vec{n}_3
~\middle|~
c_m > 0 ~\text{for}~ m=0,1,2,3
~\right\},
\end{equation}
where $\vec{n}_m (m=0,1,2,3)$ is a unit vector to point a corner of the field-of-view. The intersection of the meteor shell and the field-of-view is defined as a meteor-collecting volume (MCV). The MCV is approximated by a truncated pyramid as shown by the cross-hatched region in Figure~\ref{fig:schematic-collect}. This approximation does not affect results significantly as long as the dimension of the field-of-view is smaller than about $1\ifmmode^\circ\else$^\circ$\fi$. The distance to the MCV is represented by the distance to its geometric center. For a system with multiple image sensors such as \tomoepm, the union of the MCVs for each sensor is defined as the MCV of the system.
An incident meteor is defined by two properties: an incident point $\vec{p}$ and a pair of incident angles $(\omega,i)$. The incident point $\vec{p}$ means the location of \revise{20180710}{a dust grain} when they enter the meteor shell. The incident angles correspond to the velocity of \revise{20180710}{a grain} relative to the Earth. The angle $\omega$ is an azimuthal angle, while the angle $i$ is an incident angle measured from the normal vector of the meteor shell at $\vec{p}$. Incident meteors are illustrated by the orange arrows in Figure~\ref{fig:schematic-collect}. The evolution of the velocity by interaction with atmosphere is not taken into account. Thus, we assume that the direction of a meteor's velocity is constant.
Incident meteors are randomly generated on the test region, which is a sufficiently large part of the meteor shell, with a surface density of $\Sigma$ at an altitude of $H_2$. An effective meteor-collecting area (EMCA) at a nominal altitude of $100\,\mathrm{km}$ is defined by
\begin{equation}
\label{eq:mca}
A = \frac{N}{\Sigma}\left(\frac{100\,\mathrm{km}}{H_2}\right)^2
= N\frac{A_0}{N_0}\left(\frac{100\,\mathrm{km}}{H_2}\right)^2,
\end{equation}
where $N$ is the number of the meteors which enter the MCV, $A_0$ is the total area where the meteors are generated, $N_0$ is the total number of the generated meteor. We developed a code to calculate an EMCA for a given observing configuration. The shape, the elevation angle $\theta$, and the position angle $\phi$ of pointing are given. Incident meteors are generated until $N$ exceeds a threshold $\hat{N}$. Since the detections are given by a Poisson process, the statistical error of $A$ becomes negligible for sufficiently large $N_0$.
When the field-of-view is as large as ${\sim}50\ifmmode^\circ\else$^\circ$\fi$ in diameter, the depth of the meteor shell becomes negligible. Thus the EMCA calculated here becomes asymptotically identical to that in the thin shell approximation.
As the simplest case, we assumed that the number density of \revise{20180710}{interplanetary dust} around the Earth was uniform and the velocity of the \revise{20180710}{interplanetary dust grains} relative to the Earth was isotropic. Thus, the distribution of $\vec{p}$ was uniform on the top surface of the meteor shell, the distribution of $\omega$ distributed uniformly in $[0,\,2\pi)$, and the distribution of $i$ was set to be proportional to $\cos{i}$. In calculation, we adopted $H_1 = 80\,\mathrm{km}$ and $H_2 = 120\,\mathrm{km}$. The EMCA does not significantly depend on the choice of $H_1$ and $H_2$ (see, \ref{sec:depth_meteor_shell}). The threshold $\hat{N}$ was set to $100,000$ so that the statistical errors become negligible. Note that the effective meteor-collecting area calculated based on our assumption may contain a systematic error since the radiant distribution of sporadic meteor is not uniform \citep{jones_radiant_1994}, and may be incorrect especially for meteor showers.
\begin{figure*}
\centering
\centering\includegraphics[width=.9\linewidth]{f004.pdf}
\caption{Time variation in the meteor-collecting areas on April 11 (the red circles) and April 14 (the blue triangles).}
\label{fig:collecting-area}
\end{figure*}
The EMCAs of \tomoepm estimated with a Monte-Carlo simulation are shown in Figure~\ref{fig:collecting-area}. Random errors are as small as the symbols. Figure~\ref{fig:collecting-area} indicates that the meteor-collecting area of \tomoepm changed from about $1,000\,\mathrm{km}$ to $10,000\,\mathrm{km}$ with elevation. \revise{20180710}{Note that the EMCAs calculated here do not include the effect of the distance between the observer and meteors. When the elevation angle is low, a typical distance to meteors becomes large and faint meteors are not detected. Thus, the number of the detected meteors tends to decrease with decreasing elevation angle. Figure~\ref{fig:evrates} shows that the number of the detected meteors peaked around 15:00 UT on April 11. Detailed discussion is presented in \ref{sec:rmca}. The meteor travel distances and a possible bias in the meteor arrival directions are discussed in \ref{sec:travel_distances} and \ref{sec:position_angles}, respectively. }
\section{Discussion}
\label{sec:discussion}
\subsection{Absolute Magnitude Distribution}
\label{sec:absmags}
\begin{figure*}[tp]
\centering
\centering\includegraphics[width=.9\linewidth]{f005.pdf}
\caption{Distributions of the absolute magnitudes of the meteors detected by \tomoepm. The top and bottom panels respectively shows the distributions on April 11 and 14, 2016.}
\label{fig:hist}
\end{figure*}
The observed video-rate magnitudes $m_{V}$ were converted into absolute magnitudes $M_{V}$, which correspond to $V$-band magnitudes at the distance of $100\,\mathrm{km}$, by assuming that the distances to meteors were identical to those to the meteor detectable volume:
\begin{equation}
\label{eq:mag100km}
M_V = m_V
+ 5.0\log_{10}\frac{D(\alpha)}{100\,\mathrm{km}}.
\end{equation}
Figure~\ref{fig:hist} shows the distributions of the absolute magnitudes. Meteros of $M_V \lesssim 10$ and $9\,\mathrm{mag}$ were detected on April 11 and 14, respectively. The histograms increase exponentially with increasing magnitudes, and then start to decrease at magnitudes of 8th or 9th. The decrease is basically attributed to sensitivity limits. Detailed discussion is given in Section~\ref{sec:lf_analysis}.
\citet{jacchia_physical_1955} derived an equation to convert the magnitude of a meteor to the mass of a meteoroid, which were calibrated based on observations with Super-Schmidt cameras. A modified equation was given by \citet{jenniskens_meteor_2006}, which derives the mass of a meteoroid from the $V$-band magnitude of a meteors:
\begin{equation}
\label{eq:jenniskens}
\log_{10}{M_g} = 6.31 - 0.40M_V
- 3.92\log_{10}V_\infty - 0.41\log_{10}\sin(h_r),
\end{equation}
where $M_g$ is the mass of the meteoroid in units of grams, $M_V$ is the $V$-band magnitude of the meteor, $V_\infty$ is the incident velocity of the meteoroid in units of $\mathrm{km\,s^{-1}}$, and $h_r$ is the elevation of the radiant point of the meteor. By assuming that $V_\infty$ is $30\,\mathrm{km\,s^{-1}}$ and $h_r$ is $90\ifmmode^\circ\else$^\circ$\fi$, the mass of the meteors we detected ranges from about $8.3{\times}10^{-2}\,\mathrm{g}$ ($4\,\mathrm{mag}$) to about $3.3{\times}10^{-4}\,\mathrm{g}$ ($10\,\mathrm{mag}$).
\subsection{Luminosity Function Analysis}
\label{sec:lf_analysis}
Generally, a luminosity function of visible meteors is well approximated by an exponential function \citep{hawkins_influx_1958}:
\begin{equation}
\label{eq:lf}
\log_{10} N({<}M) = \log_{10}N_0 + M\log_{10}r,
\end{equation}
where $N({<}M)$ and $N_0$ are the event rates of meteors brighter than $M$-th and zero-th magnitude, respectively.
The present results are a composite of the observations in different conditions in terms of the meteor-collecting area and the limiting magnitudes. Thus, a naive fitting of the apparent magnitude distribution to derive luminosity functions may bring biased results. We introduce a statistical model to robustly estimate the slope parameter $r$ and the meteor rate $\log_{10}N_0$.
We assume that the luminosity function of meteors exactly follows Equation~(\ref{eq:lf}) and observational conditions are constant within the observation unit. The number of detectable meteors per observation unit is expected by
\begin{equation}
\label{eq:expnum}
\left\{
\begin{array}{ll}
\log_{10}\tilde{n}^i &= \log_{10}tN^i_0A_i + \mathcal{M}^i\log_{10}r \\
\log_{10}N^i_0 &= \log_{10}N_0 + \eta(T^i - \mathrm{15:00}_\mathrm{UTC})
\end{array},\right.
\end{equation}
where $t$ is the duration of an observation unit, $A_i$ is the meteor-collecting area of the $i$-th observation unit in units of $\mathrm{km^2}$, $\mathcal{M}^i$ is a limiting magnitude in the $i$-th observation unit, $\eta$ is a parameter to describe the diurnal variation \citep{hawkins_variation_1956,hawkins_influx_1958,fentzke_latitudinal_2009} in units of $\mathrm{dex\,hour^{-1}}$, and $T_i$ is the observation time of the $i$-th observation unit in hours in UTC. From Equation~(\ref{eq:lf}), expected magnitudes of detectable meteors in the $i$-th observation unit $\tilde{M}^i$ should follow
\begin{equation}
\label{eq:magexp}
\mathcal{M}^i - \tilde{M}^i \sim \mathrm{exponential}(\beta),
\end{equation}
where $\beta$ is a shape parameter of the exponential distribution, given by $\beta = \log{r}$.
\begin{figure*}[tp]
\centering
\centering\includegraphics[width=.9\linewidth]{f006.pdf}
\centering\includegraphics[width=.9\linewidth]{f007.pdf}
\centering\includegraphics[width=.9\linewidth]{f008.pdf}
\caption{Probability density funciton of the fitting parameters, $r$, $\log_{10}N_0$, and $\eta$, are shown in the top, middle, and bottom panels, respectively. The vertical lines with numbers indicate the posterior mean values.}
\label{fig:modelfit}
\end{figure*}
\begin{table*}
\centering
\caption{Statistical Inference with the Simple Model}
\label{tab:modelfit}
\begin{tabular}{cccccc}
\hline\hline
& Parameter & Mean & 2.5\% & 50\% & 97.5\% \\ \hline
2016-04-11
& $r$ &$\phm3.24$&$\phm3.01$&$\phm3.23$ &$\phm3.50$ \\
& $\log_{10}N_0$ & $-5.59$& $-5.91$& $-5.59$ & $-5.30$ \\
& $10^2\eta$ &$\phm1.16$& $-0.30$&$\phm1.16$ &$\phm2.62$ \\
\hline
2016-04-14
& $r$ &$\phm3.08$&$\phm2.74$&$\phm3.06$&$\phm3.50$ \\
& $\log_{10}N_0$ & $-5.34$& $-5.82$& $-5.33$& $-4.93$ \\
& $10^2\eta$ &$\phm0.50$& $-1.92$&$\phm0.50$&$\phm2.91$ \\
\hline
\end{tabular}
\end{table*}
The parameters are optimized to match $n^i \sim \tilde{n}^i$ and $M^i \sim \tilde{M}^i$. The optimization is carried out with \texttt{Stan} \citep{carpenter_stan:_2016}, which is a software for the Bayesian statistical inference with MCMC sampling \citep{carpenter_stan:_2016,homan_no-u-turn_2014}. The posterior probability distributions are shown in Figure~\ref{fig:modelfit}. The posterior mean values and the 95\% confidence intervals of the parameters are listed in Table~\ref{tab:modelfit}. No significant differences are detected between the results on April 11 and 14. The results suggest that the slope parameter is ${\sim}3.1\pm{0.4}$ and the meteor rate is about $-5.5{\pm}0.5$.
The data on April 11 marginally suggest a positive $\eta$ value, while the increase in the meteor rate is not confirmed on April 14. The present result, $\eta$ of ${\sim}1{\times}10^{-2}\,\mathrm{dex\,hours^{-1}}$, corresponds the increase by 30\% in 12 hours. The increase was much smaller than reported in previous studies \citep{murakami_annual_1955,kero_2009-2010_2012}. This could be in part attributable to our assumption in calculating the EMCA: the radiant distribution of meteors is uniform. Further observations are required to confirm the non-ditection of the diurnal variation with \tomoepm.
\begin{figure}[tp]
\centering
\centering\includegraphics[width=.9\linewidth]{f009.pdf}
\caption{Comparison of luminosity functions obtained in different observations. The present result is indicated by the blue region.}
\label{fig:luminosity_functions}
\end{figure}
We compare the present result with luminosity functions in literature in Figure~\ref{fig:luminosity_functions}. The purple solid line indicates the luminosity function derived by \citet{hawkins_influx_1958} based on surveys with Super-Schmidt cameras. \citet{hawkins_influx_1958} suggested that the $r=3.4{\pm}0.2$ and $\log_{10}N_0 = -5.2$ using meteors brighter than about $4\,\mathrm{mag}$. The green dashed line indicates the luminosity function provided by \citet{cook_flux_1980} using a 10-m reflector and phototubes. The slope parameter they derived was about $3.41$. The blue doted-dashed line and orange double-dotted-dashed line respectively show the luminosity functions in \citet{clifton_television_1973} and \citet{hawkes_inexpensive_1973}. They observed meteors using television systems. Since the brightness of meteors were not directly measured in \citet{clifton_television_1973}, we use the luminosity function re-calibrated by \citet{cook_flux_1980}. \citet{hawkes_television_1975} suggested that $r \sim 2.5$, while \citet{clifton_television_1973} suggested that $r \sim 3.4$. The red triple-dotted-dashed line shows the luminosity function of fireballs \citep{hawkins_relation_1959}, suggesting that $r \sim 2.5$. Here, we tentatively assume that the conversion factor from photographic to visual magnitude is ${+}1.0\,\mathrm{mag}$. The present result is shown by the blue region. The height of the region represents the uncertainties of the parameters.
The present observations constrain the luminosity function between about $+4$ to $+10\,\mathrm{mag}$, which is the deepest among the imaging observations. The slope parameter in the present result is roughly consistent with the other observations except for \citet{hawkins_relation_1959}. The meteor rate $N_0$ is by a factor of ${\sim}2.5$ lower than the value reported in \citet{hawkins_influx_1958}. This can be attributed to a seasonal variation in the sporadic meteor rate; The number of sporadic meteors in March and April is about a half of the annual average rate \citep{hawkins_variation_1956,kero_2009-2010_2012,murakami_annual_1955}. The meteor rate of the present result is by a factor of ${\sim}30$ lower than that of \citet{cook_flux_1980}. This difference could be in part attributable to the difference in the assumed EMCAs. \citet{cook_flux_1980} assumed that the EMCA was $3\,\mathrm{km^2}$, which was basically derived with the thin shell approximation and could be underestimated. Generally, the present result is consistent with the luminosity functions in literature. \citet{cook_flux_1980} suggested that the luminosity function of meteors was well approximated by a single slope power-law function from $-2.4$ to $+12\,\mathrm{mag}$. The present result is in line with the Cook's suggestion.
\subsection{System efficiency comparison between \tomoepm and \tomoe}
The operation of \tomoepm was completed in 2016. Observations with \tomoe started in February, 2018. Here, we compare the EMCAs between \tomoepm and \tomoe, to evaluate the efficiency in the meteor observations with \tomoe.
The EMCAs are calculated out for three systems. The shape of their \revise{20180710}{fields-of-view} are illustrated in Figure~\ref{fig:sys}. System 1 has a single field-of-view with dimensions of $39\ifmmode'\else$'$\fi.7{\times}22\ifmmode'\else$'$\fi.4$, corresponding to that of one CMOS sensor. System 2 emulates \tomoepm, a wide-field camera equipped with 8 CMOS image sensors mounted on 105\,cm Kiso Schmidt telescope \citep{sako_development_2016}. System 2 has eight image sensors, each of which has a field-of-view with dimensions of $39\ifmmode'\else$'$\fi.7 {\times} 22\ifmmode'\else$'$\fi.4$. System 3 is a counterpart of \tomoe. The field-of-view of System 3 is composed of 84 segments. The size of each segment is the same in Systems 1 and 2.
The MCVs with an elevation angle of $35\ifmmode^\circ\else$^\circ$\fi$, $60\ifmmode^\circ\else$^\circ$\fi$, and $85\ifmmode^\circ\else$^\circ$\fi$ are illustrated in Figure~\ref{fig:mcv_sys1}. The position angles of the field-of-view are fixed. Both the size of the MCV and the distance to the MCV increase with decreasing elevation angle.
\begin{figure*}
\centering\includegraphics[width=.9\linewidth]{f010.pdf}
\caption{\revise{20180710}{Fields-of-view} of the three systems. The violet cross-hatched region indicates the field-of-view of System 1. The green empty rectangles indicate the field-of-view of System 2. The blue filled rectangles indicate the field-of-view of System 3.}
\label{fig:sys}
\end{figure*}
\begin{figure*}
\centering\includegraphics[width=.9\linewidth]{f011.pdf}
\centering\includegraphics[width=.9\linewidth]{f012.pdf}
\centering\includegraphics[width=.9\linewidth]{f013.pdf}
\caption{Schematic view of meteor-collecting volumes (MCV). The top, middle, and bottom plots respectively show the MCVs for Systems 1, 2, and 3. The labels ``A'', ``B'', and ``C'' indicate the MCVs for elevation angles of $85\ifmmode^\circ\else$^\circ$\fi$, $60\ifmmode^\circ\else$^\circ$\fi$, and $35\ifmmode^\circ\else$^\circ$\fi$, respectively. The blue dots indicate the location of the observing system. The green solid curve in Panels B indicates the sea level. The gray dashed lines in Panels B respectively show the top and bottom surface of the meteor sphere.}
\label{fig:mcv_sys1}
\end{figure*}
The EMCAs for the three systems are calculated for elevation angles from $10\ifmmode^\circ\else$^\circ$\fi$ to $90\ifmmode^\circ\else$^\circ$\fi$ at $5\ifmmode^\circ\else$^\circ$\fi$ intervals. Figure~\ref{fig:emca} shows the dependence of the EMCA on the elevation angle $\alpha$. The EMCAs at zenith are about $43$, $281$, and $630\,\mathrm{km^2}$ for System 1, 2, and 3, respectively. At $\alpha = 10\ifmmode^\circ\else$^\circ$\fi$, the EMCAs increase to about $1.1{\times}10^3$, $7.8{\times}10^3$, and $25.8{\times}10^3\,\mathrm{km^2}$ for System 1, 2, and 3, respectively. The EMCAs for Systems 1, 2, and 3 similarly increase with decreasing elevation angle. The dependence of the EMCA on $\alpha$ is well approximated by $\sin^{-2}\alpha$. \revise{20180710}{Note that the EMCAs, which do not include the effect of the distance to meteors, do not reflect the expected number of meteors detected. Refer to \ref{sec:rmca} for detailed discussion.}
\begin{figure*}
\centering\includegraphics[width=.9\linewidth]{f014.pdf}
\caption{Effective meteor-collecting areas (EMCAs) in units of $\mathrm{km^2}$ against an elevation angle $\alpha$. The error bars are as small as the line width.}
\label{fig:emca}
\end{figure*}
The EMCAs for Systems 2 is about 7 times larger than that of System 1. This increase is almost proportional to the number of the image sensors. The efficiency of System 3 is about \revise{20180710}{two times} larger than that of System 2. Since the detectors of \tomoe are tiled in the image circle of the telescope, a large fraction of meteors will be detected first in peripheral detectors; the detectors around the center of the field-of-view have little contribution to the meteor-collecting efficiency.
\begin{table*}
\centering
\caption{EMCA for Square \revise{20180710}{Fields-of-View}}
\label{tab:emca_simple}
\begin{tabular}{p{9.0em}cccccc}
\hline\hline
& \multicolumn{1}{c}{$10\ifmmode'\else$'$\fi{\times}10\ifmmode'\else$'$\fi$}
& \multicolumn{1}{c}{$20\ifmmode'\else$'$\fi{\times}20\ifmmode'\else$'$\fi$}
& \multicolumn{1}{c}{$30\ifmmode'\else$'$\fi{\times}30\ifmmode'\else$'$\fi$}
& \multicolumn{1}{c}{$40\ifmmode'\else$'$\fi{\times}40\ifmmode'\else$'$\fi$}
& \multicolumn{1}{c}{$50\ifmmode'\else$'$\fi{\times}50\ifmmode'\else$'$\fi$}
& \multicolumn{1}{c}{$60\ifmmode'\else$'$\fi{\times}60\ifmmode'\else$'$\fi$}
\\\hline
Side Length (\ifmmode'\else$'$\fi)
& $10$
& $20$
& $30$
& $40$
& $50$
& $60$
\\
FOV area ($\mathrm{deg}^2$)
& $0.03$
& $0.11$
& $0.25$
& $0.44$
& $0.69$
& $1.00$
\\
EMCA ($\mathrm{km^2}$)
& $14.6$
& $27.2$
& $41.4$
& $55.1$
& $69.2$
& $83.2$
\\
Projected Area\tablenotemark{a} ($\mathrm{km^2}$)
& $0.09$
& $0.34$
& $0.76$
& $1.35$
& $2.12$
& $3.05$
\\
Event Rate\tablenotemark{b} ($\mathrm{hour^{-1}}$)
& $22$
& $45$
& $68$
& $90$
& $114$
& $136$
\\
\hline\hline
\end{tabular}
\tablenotetext{a}{The area of the field-of-view projected on the sphere at an altitude of $100\,\mathrm{km}$.}
\tablenotetext{b}{Estimated meteor event rates for meteors brighter than $10$ magnitude.}
\end{table*}
To investigate the dependence of the EMCA on the field-of-view, we calculate the EMCA for simple systems which have a single square field-of-view pointing at the zenith. The \revise{20180710}{side lengths} of the \revise{20180710}{fields-of-view} are $10\ifmmode'\else$'$\fi$, $20\ifmmode'\else$'$\fi$, $30\ifmmode'\else$'$\fi$, $40\ifmmode'\else$'$\fi$, $50\ifmmode'\else$'$\fi$, and $60\ifmmode'\else$'$\fi$. Calculated EMCAs are listed in Table~\ref{tab:emca_simple} as well as the areas of the \revise{20180710}{fields-of-view} projected on the surface at an altitude of $100\,\mathrm{km}$ above the sea level (\textit{hereafter}, referred as to projected \revise{20180710}{fields-of-view}).
The EMCA is larger than the area of the projected field-of-view. In the case of the $60\ifmmode'\else$'$\fi{\times}60\ifmmode'\else$'$\fi$ system, the EMCA is larger by a factor of about 30. This simply illustrates the thin shell approximation is not appropriate for systems with small \revise{20180710}{fields-of-view}.
The EMCAs are almost proportional to the \revise{20180710}{side lengths} of the \revise{20180710}{fields-of-view}, rather than the areas of the \revise{20180710}{fields-of-view}. \revise{20180710}{The meteor-collecting efficiency, in general, depends on the surface area of the MCV. For a narrow field-of-view, the surface area of the MCV is approximately proportional to the side length of the field-of-view. This simply explains the dependence of the EMCA on the side length; The EMCA of narrower field-of-view systems does not decrease as with the area of the field-of-view. This is generally consistent with the fact that the EMCA of System 3 is about two times larger than that of System 2; The summation of the most outer circumference sides of System 2 is about $11.3\ifmmode^\circ\else$^\circ$\fi$, while that of System 3 is about $19.5\ifmmode^\circ\else$^\circ$\fi$.}
The expected number of detected meteors per hour is estimated by $A N_0 r^{\mathcal{M}}$, where $\mathcal{M}$ is a limiting magnitude, $r$ is a meteor index, $A$ is an EMCA, and $N_0$ is a rate of meteors brighter than zeroth magnitude. Here, we adopt $r=3.4$ and $\log_{10}N_0 = -5.1$ \citep{hawkins_influx_1958}. Table~\ref{tab:emca_simple} lists the event rate calculated for $\mathcal{M} = 10\,\mathrm{mag}$ in optical wavelengths. A system with a $1\ifmmode^\circ\else$^\circ$\fi{\times}1\ifmmode^\circ\else$^\circ$\fi$ field-of-view and a limiting magnitude of ${\sim}10$ will detect more than 100 meteors per hour. This result encourages meteor survey observations with wide-field cameras mounted on large aperture telescopes (e.g. Subaru/HSC \citep{miyazaki_hyper_2012}).
\tomoe is expected to detect more than about 2,000 faint meteors a night. Since meteors will run across multiple detectors of \tomoe, faint meteors are more robustly detected with \tomoe than \tomoepm. \tomoe will be an ideal instrument to investigate variations or structures in the luminosity function of meteors brighter than about ${+}10\,\mathrm{mag}$.
\section{Conclusion}
\label{sec:conclusion}
\tomoe is a mosaic CMOS camera developed in Kiso Observatory, the University of Tokyo. \tomoe, equipped with 84 CMOS sensors, will continuously obtain images of $20\,\mathrm{deg^2}$ at $2\,\mathrm{Hz}$. In this sense, \tomoe will be the world largest video camera for astronomy. As a pilot project, we developed a prototype model of \tomoe, \tomoepm, which has 8 CMOS sensors. \tomoepm can monitor a sky of about $2\,\mathrm{deg^2}$ at $2\,\mathrm{Hz}$. We carried out imaging observations of faint meteors with \tomoepm on April 11 and 14, 2016. We set the observation field within the Earth's shadow to eliminate contamination from debris and satellites in the low Earth orbit. The total observing time was about 10 hours, and 2,220 meteor events were detected in total.
\revise{20180710}{Our observations provide a new measurement of the meteor luminosity function.} The video rate magnitudes of the meteors we detected are typically from $+4$ to $+12\,\mathrm{mag}$. The corresponding mass of the meteors ranges from $8.3{\times}10^{-2}$ to $3.3{\times}10^{-4}\,\mathrm{g}$. \revise{20180710}{The present results are consistent with a single power-law luminosity function.} A statistical model suggest that the slope parameter $r = 3.1{\pm}0.4$ and the meteor rate $\log_{10}N_0 = -5.5{\pm}0.5$. The diurnal variation in the sporadic meteor rate is marginally confirmed only on April 11. \revise{20180710}{The slope is roughly consistent with those of luminosity functions in literature. The meteor rate is lower than those in literature. The discrepancy is, in part, attributed to the seasonal variation and different assumptions on the EMCAs.}
The operation of \tomoepm was completed in 2016. Observations with \tomoe started in February, 2018. We compare the effective meteor-collecting areas between \tomoepm and \tomoe, suggesting that \tomoe is twice as efficient as \tomoepm in detecting meteors. More than 2,000 meteors will be detected every night with \tomoe. Kiso Observatory has \revise{20180710}{the capability to make it a leading place} in time-domain astronomy including observations of faint meteors.
\section*{Acknowledgments}
This research is supported in part by JSPS Grants-in-Aid for Scientific Research (KAKENHI): Grant Numbers JP16H02158, JP16H06341, JP18H01261, and JP18K13599. This research is in part supported by JSPS Grant-in-Aid for Scientific Research on Innovative Areas: Grant Number JP2905. This is also supported in part by PRESTO, Japan Science and Technology Agency (JST). This work is also supported by the Optical and Near-infrared Astronomy Inter-University Cooperation Program by the MEXT of Japan. This work is achieved using the grant of Joint Development Research by the Research Coordination Committee, National Astronomical Observatory of Japan (NAOJ). The fabrication of Tomo-e PM is conducted in collaboration with the Advanced Technology Center of NAOJ. This research is supported by Canon Inc. for developing the full-HD high-sensitivity CMOS sensors.
|
\section{Introduction}
One of the most basic questions in graph theory is, given graphs $H$ and $G,$ to decide whether $H$ is a subgraph of $G.$
If so, we also say that $H$ can be {\it embedded} into $G,$ and call $G$ the {\it host graph.}
In this paper we construct a class of bounded degree bipartite graphs that have small separator and large bandwidth, and prove that the graphs of this
class are spanning subgraphs of $n$-vertex graphs that have minimum degree just slightly larger than $n/2.$ We also show that using earlier methods these
graphs cannot be embedded in general into host graphs with such small minimum degree.
The structure of the paper is as follows. First we give an brief overview of the area, finishing with a list of our new results. Next we define an infinite class of bounded degree graphs having small separators and very large bandwidth. Finally, we embed such graphs using the Regularity lemma -- Blow-up lemma method.
\section {Overview of the area and our new results}
In this paper we consider only simple graphs. We use standard graph theory notation. In particular, if $F=(V, E)$ is a graph then
the degree of a vertex $v\in V(F)$ is denoted by $deg_F(v),$ or just $deg(v),$ if $F$ is understood from the context. The number of vertices of $F$
is denoted by $|F|,$ and we let $e(F)=|E(F)|.$ Given $A\subset V(F)$ we let $F[A]$ denote the subgraph of $F$ that is spanned by the vertices in $A.$ If $A, B\subset V(F)$ such that $A\cap B=\emptyset$ then $F[A, B]$ denotes the bipartite subgraph of $F$ which contains precisely the edges with one endpoint in $A$ and the other endpoint in $B.$
Given $S\subset V(F)$ then $N(S)$ denotes the set of those vertices that have at least one neighbor in $S.$ The maximum degree of $F$ is denoted by $\Delta(F),$ the minimum degree of $F$ is denoted by $\delta(F).$
Perhaps the most cited result in extremal graph theory is the celebrated theorem of Dirac, stating that if the minimum degree of a graph on $n\ge 3$
vertices is at least $n/2,$ then the graph contains a Hamilton cycle. This result was generalized in various ways. In a {\it Dirac-type} embedding problem one gives a
lower bound on the minimum degree of the host graph $G,$ and finds conditions for a graph $H$ on the same number of vertices which guarantees that $H\subset G.$
The famous Bollob\'as-Eldridge-Catlin conjecture~\cite{BE, C} is a Dirac-type question, in which we have a bound for the maximum degree of the graph to be embedded.
It asserts that if $G$ and $H$ are graphs on $n$ vertices, and $\delta(G)\ge (1-1/(\Delta(H)+1))n$ then $H\subset G.$
It is still open in general, only some special cases were solved (for large $n$), e.g., the cases $\Delta(H)=2$~\cite{AignerBrandt}, $\Delta(H)=3$~\cite{CsSSz}, and when
$\Delta(H)$ is bounded and $H$ is bipartite~\cite{BEparos}. Kaul, Kostochka and Yu proved an approximation result in~\cite{KKY}.
One may impose other kind of restrictions on $H$ and obtains still hard problems. For example, one may upper bound the so-called bandwidth of $H,$ this guarantees that $H$ is ``far from being an expander".
The bandwidth of a graph $H$ is denoted by $bw(H)$ and is defined to be the smallest positive integer $b$ such that there exists a
labelling of the vertices of $V(H)$ by the numbers $\{1,\ldots, n\}$ such that the labels of every pair of adjacent vertices
differ by at most $b.$
Note that a Hamilton path has bandwidth 1, a Hamilton cycle has bandwidth 2. Expander graphs have large, linear bandwidth, a star on $n$ vertices
has bandwidth $n/2,$ a complete graph has bandwidth $n-1.$
One of the most important open problems concerning the bandwidth was a conjecture by Bollob\'as and Koml\'os~\cite{Komlos}. This conjecture was proved by B\"ottcher, Schacht and Taraz~\cite{BST}\footnote{We remark, that the case of $k=2$ was first proved by Abbasi~\cite{Sarmad}, and is also a special case of a result by the author~\cite{wellsep}. Some of the ideas of the latter proof are used in the present paper as well.} using deep tools, in particular the Regularity lemma and the proof of the celebrated P\'osa-Seymour conjecture by Koml\'os, S\'ark\"ozy and Szemer\'edi~\cite{PosaSeymour}.
\begin{theorem}\label{BollKomlos
For every $D, k\in \mathbb{N}$ and $\varepsilon >0$ there exists $\beta>0$ such
that the following holds. Every $n$-vertex graph $G$ having minimum degree
$\delta(G)\ge (1 - 1/k+\varepsilon)n$ contains all $k$-chromatic $n$-vertex graphs of maximum degree
at most $D$ and bandwidth at most $\beta n$ as subgraphs.
\end{theorem}
B\"ottcher~\cite{Btezis} and B\"ottcher et al.~\cite{BPstb} went further and explored relations of bandwidth with other notions, like separability.
We say that an $n$-vertex graph $H$ is $\gamma$-separable if there exists a separator set $S\subset V(H)$ with $|S|\le \gamma n$ such that
every component of $H-S$ has at most $\gamma n$ vertices. B\"ottcher et al.~\cite{BPstb} observed that bandwidth and separability are closely related: they proved the Sublinear Equivalence Theorem. This states that, roughly speaking, in bounded degree graphs
sublinear bandwidth implies the existence of a sublinear sized separating set and vice versa.
It is easy to see that there are bounded degree graphs having linearly large bandwidth: it is well-known that a random $l$-regular graph with $l\ge 3$
has large bandwidth with positive probability. However, such random graphs do not have small separators.
One of our results shows that when the separating set has small (but not very small) linear size, the bandwidth can be very large even for bounded degree graphs.
\begin{theorem}\label{fo1}
Let $r\ge 35$ and $t\ge 1$ be integers and set $\gamma=\gamma(r)= 1/(8r2^r).$ Then one can construct an infinite class of graphs
$\mathcal{H}_{r, t}$ such that every element $H$ of $\mathcal{H}_{r, t}$ has a separator set of size $\le \gamma |H|,$ has bandwidth at least $0.3|H|/(2t+4),$ moreover, $\Delta(H)=O(1/\gamma).$
\end{theorem}
There is a recent interest in embedding graphs with sublinear bandwidth. Staden and Treglown~\cite{ST} embed graphs on $n$ vertices with sublinear bandwidth into locally
dense graphs on $n$ vertices having minimum degree at least $(1/2 + o(1))n.$ Condon, Kim, K\"uhn and Osthus~\cite{CKKO} find an approximate decomposition of a certain class of graphs into graphs that have sublinear bandwidth.
The result of Knox and Treglown~\cite{KT} is particularly interesting for us. They embedded bounded degree graphs with sublinear bandwidth into so called robust expanders.
Let $0< \nu \le \tau <1.$ Assume that $G$ is a graph of order $n$ and
$S\subset V(G).$ The $\nu$-robust neighborhood $RN_{\nu, G}(S)$ of $S$ is the set of vertices $v\in V(G)$ such that $|N(v)\cap S|\ge \nu n.$ We say that $G$ is a robust $(\nu, \tau)$-expander
if $|RN_{\nu, G}(S)|\ge |S|+\nu n$ for every $S\subset V(G)$ such that $\tau n\le |S|\le (1-\tau)n.$
We will also show that
elements of $\mathcal{H}_{r, t}$ (the graph class of Theorem~\ref{fo1}) cannot be embedded into arbitrary robust expanders.
However, if an $n$-vertex graph $G$ has minimum degree slightly larger than $n/2,$ then it contains the elements of $\mathcal{H}_{r, t}$ as spanning
subgraphs. We will prove the following.
\begin{theorem}\label{fo2}
Let $r\ge 35$ and $t\ge 1$ be integers and set $\gamma=\gamma(r)= 1/(8r2^r).$ Then there exists an $n_0=n_0(\gamma)$ such that the following holds.
Assume that $n\ge n_0$ and $G$ is an $n$-vertex graph having minimum degree $\delta(G)\ge (1/2+3\gamma^{1/3})n.$ If $H\in \mathcal{H}_{r,t}$ is a graph on
$n$ vertices, then $H\subset G.$
\end{theorem}
A standard example shows that one cannot significantly reduce the minimum degree of $G$ in the above theorem. Let $G$ be the union of two complete graphs on, say,
$n/2+\gamma n/100$ vertices that share $\gamma n/50$ vertices. Clearly, $\delta(G)=n/2+\gamma n/100-1.$ It is an easy exercise to prove\footnote{Of course, one first needs the definition of $\mathcal{H}_{r,t}$ for this proof, which is given in Section~\ref{harmadik} soon.} that if $H\in \mathcal{H}_{r,t}$ and $|H|=|G|=n,$ then $H\not\subset G.$ We leave this proof for the reader.
The proof of Theorem~\ref{fo2} will rely heavily on the proof method of~\cite{wellsep} and an important result of Fox and Sudakov~\cite{FS}. Let us remark that in~\cite{wellsep}
the size of the separator set was $o(n),$ and therefore, by the Sublinear Equivalence theorem, the bandwidth was also $o(n).$ This time the separator set is quite large
compared to previous results.
Finally, let us mention a tightness result for Theorem~\ref{BollKomlos} by Abbasi~\cite{Sarmad}. He proved that for infinitely many $n$ there exist graphs $G$ and $H$ on $n$ vertices such that $\delta(G)=(1/2 +\eta)n$ and $bw(H)\le 4\eta n,$ still $H\not\subset G.$ So in general if one wants to be able to embed every graph with bandwidth $\eta n,$ then the minimum degree bound for the host graph must
be larger than $n/2 + \eta n/4.$
\section{Construction of $\mathcal{H}_{r,t}$ and proof of Theorem~\ref{fo1}}\label{harmadik}
In order to exhibit the infinite family of graphs $\mathcal{H}_{r,t}$ we first need to construct certain kind of bipartite expander graphs.
We begin with defining a bipartite graph $F$ with vertex classes $V_1$ and $V_2$ such that $|V_1|=|V_2|=k$ and $F$ has relatively good expansion properties. Our construction of $F$ relies on the existence of so called Ramanujan graphs\footnote{We remark that here one may as well work with random regular bipartite graphs instead of explicit constructions.}: an $r$-regular (non-bipartite) graph $U$ is a Ramanujan graph if $\lambda \le 2\sqrt{r-1},$ where $\lambda$ is the second largest in absolute value of the eigenvalues of $U$ (since $U$ is $r$-regular, the largest eigenvalue is $r$). Lubotzky, Phillips and Sarnak~\cite{LPS}, and independently Margulis~\cite{Marg}, constructed for every $r=p+1$ where $p\equiv 1\mod 4$
infinite families of $r$-regular graphs with second largest eigenvalues at most $2\sqrt{r-1}.$
We need a fact about these graphs, a lower bound for the number of edges between subsets of $U.$
\begin{lemma}\label{elszam}
Let $U$ be a graph as above. Then for every two subsets $A, B \subset V(U)$ where $|A|=ak$ and $|B|=bk$ we have
$$|e(A, B)-abrk|\le 2\sqrt{r-1}\sqrt{ab}k.$$
\end{lemma}
The proof of Lemma~\ref{elszam} can be found for example in~\cite{AS}.
\begin{corollary}\label{kov}
Let $U$ be an $r$-regular Ramanujan graph on $k$ vertices with $r\ge 35.$ Let us assume that $A, B\subset V(U)$ with $|A|=|B|=k/3$ and $A\cap B=\emptyset.$ Then $e(A, B)\ge 1.$
\end{corollary}
\begin{proof}
It is easy to see that the expression of Lemma~\ref{elszam} gives a lower bound for $e(A, B)$ which is monotone increasing in $r.$ Hence it is sufficient to apply Lemma~\ref{elszam} with $r=35$ and $a=b=1/3.$
Straightforward computation gives what was desired.
\end{proof}
We are ready to discuss the details of the construction of $F.$ Given an $r$-regular Ramanujan graph $U$ with $r\ge 35$ the vertex classes of $F$ will be copies of $V(U)$: for every $x\in V(U)$ we have two copies of it, $x_1\in V_1$
and $x_2\in V_2.$ For every $xy\in E(U)$ we include the edges $x_1y_2$ and $x_2y_1$ in $E(F).$ Finally, for every $x\in V(U)$ we will also have the edge $x_1x_2$ in $E(F).$ Observe that $F$
is an $(r+1)$-regular bipartite graph.
The following claims are crucial for the construction of $\mathcal{H}_{r,t}.$
\begin{claim}\label{elszam2}
Let $A\subset V_1$ and $B\subset V_2$ be arbitrary such that $|A|=|B|=k/3.$ Then $e_F(A, B)\ge 1.$
\end{claim}
\begin{proof}
If there exists $x\in V(U)$ such that $x_1\in A$ and $x_2\in B$ then we are done since every $x_1x_2$ edge is present in $F.$ If there is no such $x\in V(U)$ then we can apply Corollary~\ref{kov}
and obtain what is desired.
\end{proof}
\begin{claim}\label{expanzio2}
For every $A\subset V_1$ we have $|N_F(A)|\ge |A|.$ Analogous statement holds for any subset $B\subset V_2.$
\end{claim}
\begin{proof}
The claim easily follows from the fact that we included a perfect matching in $F$ when for every $x\in V(F)$ we added the $x_1x_2$ edge to $E(F).$
\end{proof}
Observe that we can construct a bipartite graph $F$ with $|V(F)|=2k$ having the above properties whenever there exists a Ramanujan graph $U$ with $|V(U)|=k,$ for the latter we also assume that $r\ge 35.$
Thus, there exists an infinite sequence of $\{F_i\}_{i=1}^{\infty}$ graphs on increasing number of vertices,
say, $F_i$ has $2k_i$ vertices.
We are ready to define $\mathcal{H}_{r,t}.$ Each graph from this class
is $\gamma$-separable where $\gamma=\gamma(r)$ can be relatively small as we will see soon. Still, the bandwidth of each of them is very large. Hence, $\mathcal{H}_{r,t}$ demonstrates that in spite of sublinear equivalence of separability and bandwidth, there is no {\it linear} equivalence.
The construction of $\mathcal{H}_{r,t}$ is somewhat specific, we do it with foresight as our goal is not only to further explore the relation of separability and bandwidth but also to be able to embed the elements of $\mathcal{H}_{r,t}$ later.
\begin{definition} Let $n, m \in \mathbb{N}$ be sufficiently large, $r\ge 35,$ $t\ge 1$ be integers, and set $\gamma=\gamma(r)= 1/(8r2^r).$ Let $F_i$ be the $(r+1)$-regular
bipartite graph on $2k_i$ vertices given above such that
$k_i$ is the largest for which $\gamma n\ge 2k_i.$
The elements of $\mathcal{H}_{r,t}$ are constructed as follows. Given $n$ we let $H=(A,B;E)\in \mathcal{H}_{r,t}$ to be the following bipartite graph.
\begin{enumerate}
\item $V(H)=A\cup B$ and $|H| = |A\cup B| = n,$
\item let $S\subset V$ such that $S= S_A\dot{\cup} S_B$ and $|S_A| = |S_B| = k_i$,
\item $E(H[S_A])=E(H[S_B])=\emptyset$ and $H[S_A, S_B] = F_i,$
\item $D=\Delta(H)=O(r2^r),$
\item $H-S$ has exactly $|S|$ isomorphic components; each component is a tree $T$ which contains a path on $t$ vertices, and one of its endpoints, the {\em last vertex}, has $D-1$ leaves attached to it; we call the other endpoint of this path the {\em first vertex}.
\item every $x\in S$ has a unique neighbor $y$ in $V-S$ which is the first vertex of one of the tree components of $H-S,$ moreover, every first vertex has exactly one neighbor in $S.$
\end{enumerate}
\end{definition}
\begin{figure}[h!]
\centering
\begin{tikzpicture}[vertex/.style={draw,fill,circle,inner sep=0pt,minimum size=5pt}]
\foreach \x in {1,...,9}{
\foreach \t in {0,2}
{
\node[vertex] (s\x\t) at (\x*0.6,\t) {};
}
}
\foreach \x/\t in {4/8,1/4,9/7,3/8,9/5,3/5,2/3,4/7,1/1,4/9,5/2,9/8,4/7,2/9,6/9,7/6,1/6,1/5,2/6,3/1,8/8,6/4,8/5,3/5,8/9,5/3,7/2,4/2,6/3,5/5,2/1,8/7,3/7}
{
\draw (s\x0) -- (s\t2);
}
\draw[thick,rounded corners] ($(s10.north west)+(-0.3,0.15)$) rectangle ($(s90.south east)+(0.3,-0.15)$) node[below,yshift=-0.2cm,xshift=-2.7cm]{$S_B$};
\draw[thick,rounded corners] ($(s12.north west)+(-0.3,0.15)$) rectangle ($(s92.south east)+(0.3,-0.15)$) node[above,yshift=0.6cm,xshift=-2.7cm] {$S_A$};
\foreach \x/\t/\i in {0.8/3/0,0.6/3.5/1,0/4/2,-1/3.8/3,-1.5/3/4}
{
\node[vertex] (x\i) at (\x,\t) {};
}
\draw (s12) -- (x0) -- (x1) -- (x2) -- (x3) -- (x4);
\foreach \x in {1,...,6}
{
\node[vertex] (a\x) at (\x*0.4-3,2) {};
\draw (x4) -- (a\x);
}
\draw[thick,rounded corners] ($(a1.north west)+(-0.2,0.15)$) rectangle ($(a6.south east)+(0.2,-0.15)$) node[below,xshift=-1.25cm] {$\leq\Delta-1$};
\foreach \x/\t/\i in {5.5/-0.5/0,6/-1/1,6.6/-1.3/2,7/-1/3,7.5/-0.5/4}
{
\node[vertex] (y\i) at (\x,\t) {};
}
\draw (s90) -- (y0) -- (y1) -- (y2) -- (y3) -- (y4);
\foreach \x in {1,...,6}
{
\node[vertex] (b\x) at (\x*0.4+6.2,0) {};
\draw (y4) -- (b\x);
}
\draw[thick,rounded corners] ($(b1.north west)+(-0.2,0.15)$) rectangle ($(b6.south east)+(0.2,-0.15)$) node[above,xshift=-1.25cm,yshift=0.5cm] {$\leq\Delta-1$};
\end{tikzpicture}
\caption{The separator set $S$ and the way the components of $H-S$ are connected to $S$}\protect\label{fig1}
\end{figure}
Note that $S$ is a separator set of $H$ with $|S|=2k_i\approx \gamma n,$ every component of $H-S$ has less than $t+D$ vertices. From this one can easily obtain the bound $D< 3n/k_i\approx 6/\gamma.$ We remark that when $t=1$ then
the first and the last vertex in every path are the same.
The following lemma is crucial for bounding the bandwidth of $H\in \mathcal{H}_{r,t}.$
\begin{lemma} \label{shortpath}
Let $H$ be an element of $\mathcal{H}_{r,t}$ on $n$ vertices. Assume that $X,Y\subset V(H)$ with $|X|, |Y|\geq 0.35n$ and $X\cap Y=\emptyset.$ Then there exists an $x\in X$ and a $y\in Y$ such that the distance of $x$ and $y$
is at most $2t + 4.$
\end{lemma}
\begin{proof
Denote the vertices of $H-S$ closer to $S_A$ by $A^*$, and analogously, the vertices of $H-S$ closer to $S_B$ by $B^*.$ By the construction of $H$ we have $|A^*| = |B^*| \ge (1-\gamma)n/2$.
Note that $\gamma<0.01$ since $r\ge 35.$ Hence we have that $|X-S|\geq 0.34n$ and $|Y-S|\geq 0.34n$. Thus, either $|X\cap A^*| \geq |A^*|/3$ or $|X\cap B^*|\geq |B^*|/3$. Without loss of generality, suppose the former. This also implies that at least $1/3$ of the components of $A^*$ have vertices in $X.$ Denote the first vertices in these components by $X^*.$
We also let $Y^*_A$ denote the set of first vertices of those components in $A^*$ that have at least one vertex from $Y,$ and analogously, $Y^*_B$ denotes the set of first vertices of those components in $B^*$ that have at least one vertex from $Y.$
Let $X_A=N(X^*)\cap S,$ $Y_A=N(Y^*_A)\cap S$ and $Y_B= N(Y^*_B)\cap S.$ Using these notations we have that $|X_A|\ge k/3$ and either $|Y_A|\geq {k/3}$, or $|Y_B|\geq {k/3}.$
If $|Y_B|\geq {k/3}$, then by Claim~\ref{elszam2} there is an edge $z_1z_2$ between $X_A$ and $Y_B,$ and therefore we have a path $xv_s \ldots v_1 z_1z_2 u_1\ldots u_q y$ of length $\le 2t + 3$,
where $x\in X, y\in Y, v_i\in A^*$, $z_1\in X_A$, $z_2 \in Y_B$, $u_i\in B^*$.
If $|Y_B|<\frac{k}{3}$, then $|Y_A| \geq {k/3}.$ Let $Y'_B=N(Y_A)\cap S.$ Claim~\ref{expanzio2} implies that
$|Y'_B|\ge |Y_A| \ge k/3,$ so by Claim~\ref{elszam2}
$H$ has an edge $z_2z_1$ between $Y_B'$ and $X_A.$ Thus, we have a path $xv_s \ldots v_1 z_1z_2z_3\linebreak[1] u_1\ldots u_q y$ of length $\le 2t + 4$, where $x\in X, y\in Y, v_i \in A^*, u_i\in B^*, z_1\in X_A$, $z_3 \in Y_A$ and $z_2 \in Y_B'.$
\end{proof}
\begin{corollary}
Let $H$ be an element of $\mathcal{H}_{r,t}$ on $n$ vertices. Then the bandwidth of $H$ is at least $\frac{0.3n}{2t +4}$.
\label{corollary:large-bandwidth}
\end{corollary}
\begin{proof} Take an arbitrary ordering $\mathcal{P}$ of the vertices of $H.$ Let $X$ be the first $0.35n$ vertices, and
$Y$ be the last $0.35n$ vertices of $\mathcal{P}$. Using Lemma~\ref{shortpath} there is an $x\in X$ and an $y\in Y$ such that the distance of $x$ and $y$ is at most $2t + 4$. Their distance in $\mathcal{P}$ is at least $0.3n.$ Thus at least one of the edges of the shortest path between $x$ and $y$ must have ``length'' at least $\frac{0.3n}{2t + 4},$ from which the bound for the bandwidth follows immediately.
\end{proof}
With this we proved Theorem~\ref{fo1}.
Observe that choosing $t=1$ results in graphs having bandwidth at least $3n/60=n/20$ while being $\gamma$-separable.
It is easy to see that, using the above ideas, one can define a much wider class of graphs having very large bandwidth that
are also $\gamma$-separable. Our main goal, however, is not only to construct but also to be able to embed such graphs.
\medskip
Recall the notion of robust expanders.
As we mentioned in the introduction, Knox and Treglown~\cite{KT} embedded spanning subgraphs of sublinear bandwidth into robust expanders.
The following example shows that graphs of $\mathcal{H}_{r,t}$ not only have very large bandwidth,
these graphs are not necessarily subgraphs of robust expanders. Hence, in the theorem of Knox and Treglown one cannot replace sublinear bandwidth by
$\gamma$-separability, unless $\gamma$ is very small. In fact the proofs of~\cite{BPstb} and~\cite{KT} works only
in case $\gamma\le 1/\ell,$ where $\ell$ denotes the number of clusters in a sufficiently large graph after applying the Regularity lemma with some small parameter $\varepsilon>0.$
It is known (see~\cite{Gowers}) that $\ell$ is bounded from below by a tower function of $1/\varepsilon.$
Next we us construct a robust expander as follows. Let $G=(V,E)$ be a graph on $n$ vertices such that the vertex set of $G$ is
$V = A_1\dot{\cup}A_2\dot{\cup}\cdots\dot{\cup}A_{100}$, here $|A_i|=\frac{n}{100}$ for every $1 \leq i\le 100.$
The edges of $G$ are defined as follows: $G[A_i, A_{i+1}]$ is a complete bipartite graph for every $1 \le i\le 99,$ and
and $G[A_1], G[A_{100}]$ are complete graphs on $n/100$ vertices.
From the following remark one can easily see that $G$ is a $(\nu, \tau)$-robust expander, where $1/n<\nu$ and $\tau\ge 100\nu.$
Let $1<i<100$ and $Q\subset V(G)$ be an arbitrary set. Then every vertex in $A_{i-1}\cup A_{i+1}$ has $|Q\cap A_i|$ neighbors in $Q\cap A_i.$
Furthermore, every vertex in $A_1\cup A_2$ has at least $|Q\cap A_1|-1$ neighbors in $Q\cap A_1,$ unless $|Q\cap A_1|\le 1.$ Analogous
statement holds for $A_{100}.$
\begin{lemma}
Let $H$ be a graph from $\mathcal{H}_{r,t}$ on $n$ vertices and let $G$ be as above. Then $H\not\subset G$ if $t\leq 13.$
\end{lemma}
\begin{proof}
Let $X=A_1\cup \ldots \cup A_{35}$ and $Y=A_{66}\cup \ldots \cup A_{100}.$ If one embeds $H$ into $G,$ then between the
vertices that were mapped onto some vertex of $X$ and those that were mapped onto some vertex of $Y$ there is a path having length
at most $2t+4.$ In $G$ the shortest path between $X$ and $Y$ has length $31.$ Since $2t+4<31$ whenever $t\le 13,$ we proved what
was desired.
\end{proof}
\section{Proof of Theorem~\ref{fo2}}
The proof of Theorem~\ref{fo2} is very similar to the proof of the main result of~\cite{wellsep}.
First let us state a special case of the main theorem of~\cite{wellsep} for embedding bipartite graphs with small separators.
\begin{theorem}\label{wellsep2}\cite{wellsep}
For every $\epsilon>0$ and positive integer $D$ there exists an $n_0=n_0(\epsilon, D)$ such that the following holds. Assume that $H$ is a bipartite graph on $n\ge n_0$ vertices which has a separator set $S$ such that $|S|=o(n),$ and every component of $H-S$ has $o(n)$ vertices. Assume further
that $\Delta(H)\le D.$ Let $G$ be an $n$-vertex graph such that $\delta(G)\ge (1/2+\epsilon)n.$ Then $H\subset G.$
\end{theorem}
One can observe the similarities with Theorem~\ref{fo2}. The main difference is that in Theorem~\ref{fo2} the separator set can be very large compared to the separator set in Theorem~\ref{wellsep2}. This difference requires a new embedding tool.
\subsection{Main tools for the proof}
The Regularity lemma of Szemer\'edi~\cite{SzRL} and the Blow-up lemma~\cite{KSSz1} plays a very important role in the proof.
While we assume familiarity with these tools, below we give a brief summary of the necessary definitions and results.
The interested reader may consult with the survey paper by Koml\'os and Simonovits~\cite{KS} also for further details.
\begin{definition}
Given a graph $G$ and two disjoint subsets $X, Y \subset V(G),$ the \textit{density} between $X$ and $Y$ is
\begin{equation}
d(X,Y) = \frac{e(X,Y)}{|X||Y|}.
\end{equation}
\end{definition}
\begin{definition}
We call a pair $(A,B)$ of disjoint vertex sets in $G$ $\varepsilon$-regular, if for every $X\subset A$ and $Y\subset B$ we have $|d(X,Y)-d(A,B)|<\varepsilon$, whenever $|X|>\varepsilon|A|$ and $|Y|>\varepsilon |B|$.
\end{definition}
We will need the so called Degree Form of the celebrated Regularity lemma of Szemer\'edi:
\begin{lemma}
\label{reglemma_deg}
For every $\varepsilon>0$ there is an $M = M(\varepsilon)$ such that if $G = (V,E)$ is any graph and $d\in[0,1]$, then there is a partition of the vertex set $V$ into $\ell+1$ clusters $W_0,W_1,\ldots,W_\ell$ and there is a subgraph $G'$ of $G$ such that
\begin{itemize}
\item $\ell\leq M$,
\item $|W_0|\leq \varepsilon|V|$,
\item All clusters $W_i,\ i\geq 1$ are of the same size $m$
\item $deg_{G'}(v) > deg_G(v) - (d+\varepsilon)|V|$ for all $v\in V$,
\item $W_i$ is an independent set in $G'$ for all $i\geq 1$,
\item All pairs $(W_i,W_j)$ ($1\leq i<j\leq \ell$) are $\varepsilon$-regular, each with density either 0 or at least $d$ in $G'$.
\end{itemize}
\end{lemma}
We call $W_0$ the \textit{exceptional cluster}, $W_1, \ldots, W_{\ell}$ are the non-exceptional clusters.
\begin{definition}
Apply \autoref{reglemma_deg} to the graph $G=(V,E)$ with parameters $\varepsilon$ and $d.$ We construct the {\em reduced graph} $G_r$ as follows. Its vertices are the non-exceptional clusters, and two vertices are connected if the corresponding clusters form an $\varepsilon$-regular pair with density at least $d$.
\end{definition}
\begin{claim}\label{reduced-degree}
Let $G = (V, E)$ be a graph of order $n$ and $\delta(G) \geq cn$ for some $c > 0.$ Let $G_r$ be
the reduced graph of $G'$ after applying \autoref{reglemma_deg} with parameters $\varepsilon$ and $d$.
Then $\delta(G_r) \geq (c - \theta)\ell$, where $\theta = 2\varepsilon + d$.
\end{claim}
\begin{definition}
We say that the pair of disjoint vertex sets $(A,B)$ is $(\varepsilon,\delta)$-super-regular, if it is $\varepsilon$-regular, and $deg(a)>\delta|B|$ for all $a\in A$, and $deg(b)>\delta|A|$ for all $b\in B$.
\end{definition}
\begin{remark}\label{adas}
\label{regular-addition}
If $(A,B)$ is an $\varepsilon$-regular pair, and we add $\le 2\varepsilon|A|$ new vertices to $A$ resulting $A'$, then
the new pair $(A',B)$ is $\varepsilon'$-regular, where $\varepsilon' \le 2\sqrt{\varepsilon}.$
\end{remark}
\begin{remark}
\label{regularity-removal}
If $(A,B)$ is an $\varepsilon$-regular pair with density $d$, then for any $Y\subset B,~|Y|>\varepsilon|B|$ we have
\begin{equation}
|\{x\in A: |N(x)\cap Y|\leq (d-\varepsilon)|Y|\}|\leq \varepsilon|A|.
\end{equation}
\end{remark}
We need the following very important lemma:
\begin{theorem}[Blow-up Lemma, \cite{KSSz1}]
Given a graph $R$ of order $r$ and positive parameters
$\delta,\Delta$, there exists a positive $\varepsilon=\varepsilon(\delta,\Delta,r)$ such that the following holds: Let $n_1, n_2, \ldots, n_r$ be arbitrary positive integers and replace the vertices $v_1,v_2,\ldots,v_r$ of $R$ with pairwise disjoint sets $V_1, V_2, \ldots, V_r$ of sizes $n_1, n_2, \ldots, n_r$ (blowing up). We construct two graphs on the same vertex set $V = \cup V_i$. The first graph $F$ is obtained by replacing each edge $v_iv_j$ of $R$ with the complete bipartite graph between $V_i$ and $V_j$. A sparser graph $G$ is constructed by replacing each edge $v_iv_j$ arbitrarily with an $(\varepsilon, \delta)$-super-regular pair between $V_i$ and $V_j$. If a graph $H$ with $\Delta(H)\leq \Delta$ is embeddable into $F$, then it is also embeddable into $G$.
\end{theorem}
\begin{theorem}[Strengthening the Blow-up Lemma \cite{KSSz1}]\label{blow2}
Assume that $n_i \leq 2n_j$ for every $1\leq i, j\leq r$. Then we can strengthen the lemma: Given $c > 0$ there are positive numbers $\varepsilon=\varepsilon(\delta,\Delta,r,c)$ and $\alpha = \alpha(\delta,\Delta,r, c)$ such that the Blow-up Lemma remains true if for every $i$ there are certain vertices $x$ to be embedded into $V_i$ whose images are a priori restricted to certain sets $T_x\subset V_i$ provided that each $T_x$ within a $V_i$ is of size at least $c|V_i|$, and the number of such restrictions within a $V_i$ is not more than $\alpha|V_ i|$.
\end{theorem}
Another important tool for the proof is the following result by Fox and Sudakov~\cite{FS}.
\begin{theorem}\label{DRCh}
Let $H$ be a bipartite graph with $n$ vertices and maximum degree $\Delta \ge 1.$ Assume that $\rho \in (0, 1)$ is a real number. If
$G$ is a graph with $N \ge 8\Delta \rho^{-\Delta}n$ vertices and at least $\rho \binom{N}{2}$
edges, then $H \subset G.$
\end{theorem}
We are going to apply Theorem~\ref{DRCh} in the special case $\rho=1/2.$
Finally one more notation: we will write $a\ll b$ for two positive numbers $a$ and $b$ if $a$ is sufficiently smaller than $b.$
\subsection{Proof of Theorem~\ref{fo2}}
Below we present the embedding algorithm as an itemized list.
As we indicated above, certain parts of the proof of Theorem~\ref{fo2} is very similar to the proof of Theorem~\ref{wellsep2}
from~\cite{wellsep}. Hence, whenever it is possible we refer
to the claims, lemmas of~\cite{wellsep}.
\paragraph{Step 1} Assume that $H\in \mathcal{H}_{r,t}$ has $n$ vertices. Denote the separator set of $H$ by $S,$ here $|S|\le \gamma n.$ Observe that we can apply the deep result of Fox and Sudakov, Theorem~\ref{DRCh} above, for finding a copy of $H[S]$ in $G,$ since $\delta(G)>n/2.$ Let us call the uncovered part of $G$ by $\widetilde{G}$ after embedding $H[S].$
Note that $\delta(\widetilde{G})\ge (1/2+2\gamma^{1/3})n$ and $|\widetilde{G}|\ge (1-\gamma)n.$
\paragraph{Step 2} Next we apply the Degree form of the Regularity lemma for $\widetilde{G}$ with parameters $\varepsilon$ and
$d$ such that $\varepsilon \ll d=\sqrt{\gamma}.$ As a result, we have $\ell + 1$ clusters, $W_0, W_1, \ldots, W_\ell$. The exceptional cluster $W_0$ has at most $\varepsilon n$ vertices, while all other clusters have the same size $m$. We form the reduced graph $\widetilde{G}_r$, for which we have $\delta(\widetilde{G}_r) \geq (1/2+\gamma')\ell$, where $\gamma' = 2\gamma^{1/3} - \sqrt{\gamma} - \varepsilon > \gamma^{1/3}.$
\paragraph{Step 3} Since $\delta(\widetilde{G}_r)\ge \ell/2,$ we have a perfect (or almost perfect) matching $M$ in $\widetilde{G}_r$:
at most one cluster will remain uncovered by $M,$ and only if $\ell$ is odd. If there is such a cluster, we add it into $W_0$.
\paragraph{Step 4} Next we transform the edges of $M$ into super-regular pairs. Given a $\delta$ with $\varepsilon\ll \delta \ll d$ we have to remove at most $\varepsilon m$ vertices from a cluster to make a regular pair $(\varepsilon,\delta)$-super-regular. We discard the same number of vertices, $\varepsilon m,$ from every non-exceptional cluster, and place the discarded vertices into $W_0$.
Note that pairs become $\varepsilon'$-regular with $\varepsilon'<2\varepsilon$, and the sizes of the clusters are $m'=m-\varepsilon m$. For sake of simplicity we will still use $\varepsilon$ and $m$ in the rest of the paper.
\paragraph{Step 5} The next step is to distribute the vertices of the exceptional cluster $W_0$ among the non-exceptional clusters while maintaining the super-regularity of the edges of $M.$ We also require that the resulting clusters have about the same size (i.e. we need, that $\big||W_i|-|W_j|\big|\le 3\varepsilon m$ for all $1\leq i,j\leq \ell$). For this, we use the same procedure described in \cite{wellsep}, the reader may consult with that paper for the details, here we only sketch the argument. Let us denote the neighbor of a cluster $W_i$ in the matching $M$ by $N_M(W_i)$.
For the distribution of $W_0$ we define an auxiliary bipartite graph $J$ having vertex classes $W_0$ and $V(\widetilde{G}_r).$ Here $vW_i\in E(J)$ if $deg_G(v,N_M(W_i))\geq \delta m$.
Let $\gamma'' = 3(\gamma^{1/3} -2(\varepsilon +d))$. We need the following lemma, which is a special case of Lemma 10 in~\cite{wellsep}.
\begin{lemma}\label{auxiliary-lemma}
$deg_{J}(v) \geq (\frac{1}{2} + \gamma'')\ell$ for every $v\in W_0$.
\end{lemma}
Using \autoref{auxiliary-lemma} we can distribute the vertices $W_0$ among the clusters one-by-one, essentially greedily.
Assume that we have already found non-exceptional clusters for the first $s$ vertices of $W_0.$ Then for every $1\le i\le \ell$
we let $P_s(W_i)$ to be the number of $W_0$-vertices put into $W_i.$
When we look for a non-exceptional cluster for the $(s+1)$st vertex $v\in W_0,$ we take the smallest $P_s(W_i)$ value for every $W_i\in N_J(v)$ (we break ties arbitrarily). It is easy to see by the large minimum degree of $J$ that no non-exceptional
cluster will get more than $2|W_0|/\ell$ new vertices from $W_0.$ Since $|W_0|\le \varepsilon n,$ and $n/\ell \approx m,$ we get that
$\big||W_i|-|W_j|\big|<3\varepsilon m.$ Using Remark~\ref{adas} super-regularity of edges of $M$ is maintained, although instead of $\varepsilon$ we have at most $2\sqrt{2\sqrt{\varepsilon}}<3\sqrt[4]{\varepsilon}$ here.
\paragraph{Step 6} Next we assign the components of $H-S$ to the non-exceptional clusters. We do it using a random procedure as in ~\cite{wellsep} (randomness is not necessary here, but a simple choice), i.e., the components are assigned randomly to edges of $M.$
The algorithm is as follows. Consider a component $A$ of $H$, it is a tree, so it is bipartite.
Denote its vertex classes by $A_1$ and $A_2.$ Pick an edge $Q_1Q_2$ of $M$ randomly, uniformly. Let $\pi$ be a uniform random permutation on $\{1,2\}$, and assign the vertices of $A_i$ to $Q_{\pi(i)}$ for $i=1,2$. The following lemma holds (for a proof see~\cite{wellsep}).
\begin{lemma}
With positive probability the mapping algorithm assigns $n/\ell \pm \varepsilon m/\ell$ vertices of $H$ to every edge of $\widetilde{G}_r$.
\end{lemma}
For $x\in V(H)-S$ we let $C(x)$
denote the cluster to which $x$ is assigned to.
The algorithm immediately implies that whenever $x, y\in V(H)-S$ are adjacent, then $C(x)C(y)$ is an edge of $M.$
Recall that in Step 1 $H[S]$ was embedded using Theorem~\ref{DRCh}, before even applying the Regularity Lemma.
So the vertices of $S$ were not assigned to clusters but directly mapped onto vertices of $G.$ It is clear that adjacent vertices
of $S$ were mapped onto adjacent vertices of $G.$
However, for applying Theorem~\ref{blow2}, the Blow-up lemma with restrictions, we need that if $x\in S$ is mapped onto $v\in V(G),$ $y\in V(H)-S$ is assigned to
$C(y),$ then $v$ must have many neighbors in $C(y).$
In such a case we will assign $y$ to another cluster (again, we use a procedure from~\cite{wellsep} with a few minor modifications) as follows.
Let $L$ denote those clusters in which $v$ has at least $3\gamma^{2/3} m$ neighbors. Simple counting argument shows that
$|L|\ge \ell/2.$ Let $C(y)W_i$ denote the edge of
$M$ to which the component of $y$ was assigned. Then we locate a cluster $W_j\in L$ such that $W_j$ is adjacent to $W_i$
in $\widetilde{G}_r.$ Since $L$ has more than $\ell/2$ clusters, using the minimum degree of $\widetilde{G}_r$
we have at least $2\gamma^{1/3}\ell$ choices for $W_j$ in $L.$
Then we change the assignment of $y,$ we let $C(y)=W_j.$
This way $v$ will have many neighbors in $C(y)$ and $C(y)$ will be adjacent to the cluster of the neighbors of $y$ in its component.
Observe that if we locate the $W_j$ clusters as evenly as possible then we can achieve that at most
$|S|/(2\gamma^{1/3}\ell) \le \gamma n/(2\gamma^{1/3}\ell)\le \gamma^{2/3}m$
vertices are reassigned to a particular cluster, using that $|S|\le \gamma n$ and the number of choices for a new cluster is always at least $2\gamma^{1/3}\ell.$
Of course, changing the assignment of vertices of $H$ does not affect regularity or super-regularity between clusters of $\widetilde{G}_r.$
\paragraph{Step 7} In this step we will achieve that the number of vertices assigned to a cluster is the same as the size of that cluster. As before, we will reassign some of the vertices. While in the previous step we changed the assignment of first vertices of components, this time we will work with the leaves of the components.
Say, that $W_s$ has more vertices assigned to it than its size $|W_s|.$ Then there must be a cluster $W_i$ to which we assigned
less than $|W_i|$ vertices of $H.$ Let $W_j$ denote the neighbor of $W_s$ in the matching $M.$ If $W_jW_i$ is an edge
in $\widetilde{G}_r$ then we pick a vertex $x$ such that $C(x)=W_j$ and $deg_H(x)=D-1.$
This is possible since the number of last vertices (these have degree $D-1$) is at least $(1-\gamma)n/(t+D-1),$ and by Chernoff's inequality with probability $>1-1/n^2$ we assigned more than $(1-\gamma) m/(2 (t+D-1))$ to every cluster.
This implies that the number of leaves that are assigned to a particular cluster is at least
$(D-1)(1-\gamma) m/(2 (t+D-1)),$ which is much larger than $\gamma^{2/3}m.$
We reassign some of the {\it leaves} that are adjacent to $x,$ the necessary number will be assigned to $W_i.$
If $W_jW_i$ is not an edge, then by the minimum degree of $\widetilde{G}_r$ there must exist at least $\gamma^{1/3}\ell$ clusters $W_p$ and $W_q$ such that $W_qW_i$ and $W_jW_p$ are edges in
$\widetilde{G}_r,$ and $W_pW_q$ is an edge in $M.$ Then the above procedure is done in two steps: first we reassign
some vertices (always leaves) from $W_s$ to $W_p$ and then the same number of leaves from $W_p$ to $W_i.$
Note that the same computation works as above:
at most $\gamma^{2/3}m$ vertices are reassigned at every cluster.
Let us remark again that reassigning the vertices of $H$ during this step does not affect regularity and super-regularity of pairs in $\widetilde{G}_r,$ unlike
when vertices of $G$ were distributed in Step 5.
\paragraph{Step 8} Recall that the density of the regular pairs is at least $\sqrt{\gamma},$ and at most $\gamma^{2/3}m$ vertices are reassigned at every cluster. It is crucial here to look at the so called a priori restrictions for some of the vertices of $H$ before applying the Blow-up Lemma. These are the first vertices of components (since their neighbors belong to $S$ and are already mapped), and the leaves that were reassigned
during Step 7. As the number of these vertices is very small compared to $m$ in every cluster, we are able to apply the Blow-up lemma. This finishes the proof of Theorem~\ref{fo2}.
|
\section{Introduction}
\label{sec:intro}
In recent years, adversarial machine learning has received tremendous attention as it provides new means of improving machine leaning performance and studying model robustness in the adversarial setting, such as generative adversarial networks (GANs) \cite{goodfellow2014generative} and adversarial examples \cite{goodfellow2014explaining}. In image classification tasks, well-trained machine learning models such as deep convolutional neural networks have shown to be vulnerable to adversarial examples -- human-imperceptible perturbations to natural images can be easily crafted to mislead the decision of a target image classifier \cite{szegedy2013intriguing,biggio2013evasion,biggio2017wild}, leading to new challenges on model robustness. The adversarial perturbations are often evaluated by the $\ell_1$, $\ell_2$ and $\ell_\infty$ norms \cite{kurakin2016adversarial_ICLR,carlini2017towards,chen2017ead,weng2018evaluating,su2018robustness}. Beyond image classification, other machine learning tasks such as image captioning \cite{chen2018attacking} or sequence-to-sequence text learning \cite{cheng2018seq2sick} have also shown to be vulnerable to adversarial examples. Moreover, it has been made possible to generate adversarial examples in the so-called black-box setting by simply leveraging the input-output correspondence of a target model and performing zeroth-order optimization \cite{chen2017zoo,tu2018autozoom,liu2018zeroth}.
Motivated by the recent studies in adversarial machine learning, in this paper we shift our focus to the robustness of regression models to adversarial perturbations. To the best of our knowledge, regression is a fundamental task in machine learning but little has been explored in the setting of adversarial perturbations.
Specifically, we aim to investigate the robustness of the ordinary least-squared regression models regularized by the ordered weighted $\ell_1$ (OWL) norm \cite{bogdan2013statistical,zeng2014decreasing}.
The OWL family of regularizers is a widely adopted method for sparse regression with strongly correlated covariates. It is worth mentioning that the octagonal shrinkage and clustering algorithm for regression \cite{bondell2008simultaneous}, which is called as OSCAR, is in fact a special case of the OWL regularizer \cite{zeng2014solving}. OSCAR is known to be more effective in identifying feature groups (i.e., strongly correlated covariates) than other feature selection methods such as LASSO \cite{tibshirani1996regression}.
In this paper, we investigate the robustness of OSCAR to adversarial perturbations by formulating the process of finding the maximally disruptive noise of the measurement model as an optimization problem. Although the recent work in \cite{figueiredo2016ordered} has established a finite-sample error upper bound on the OWL-regularized regression models associated with a norm-bounded noise level, it still remains unclear whether an adversary can disrupt the identified feature groups by intentionally manipulating the measurement error within the same noise budget. In other words, our adversarial formulation is novel in the sense that it provides a worst-case robustness analysis of OSCAR by finding a disruptive measurement error (but still within a specified noise budget) in order to deviate the detected feature groups from the ground truths. More importantly, upon verifying the lack of robustness to adversarial perturbations, our method could be incorporated to devise resilient OWL-regularized regression models via adversarial learning techniques.
Perhaps surprisingly, the experimental results show that using our proposed approach, it is possible to generate small norm-bounded perturbations to the measurement model, in order to deviate the regression results of OSCAR from the ground truths.
Consequently, our results offer new insights on the robustness analysis of OWL-regularized regression methods in the adversarial setting.
\section{Background}
\label{sec_background}
For any real-valued vector $\mathbf{x}=[x_1,x_2,\ldots,x_p] \in \mathbb{R}^{p}$, let $|\mathbf{x}|$ denote the vector of its element-wise absolute value and let $|\mathbf{x}|^\downarrow$ denote the element-permuted vector of $|\mathbf{x}|$ such that $|x|^\downarrow_1 \geq |x|^\downarrow_2 \geq \ldots \geq |x|^\downarrow_p$, where $|x|^\downarrow_i$ is the $i$-th largest component of $\mathbf{x}$. Given a vector $\mathbf{w} \in \mathbb{R}^p_+$ such that $w_1 \geq w_2 \geq \ldots \geq 0$ and $w_1>0$, the OWL norm \cite{bogdan2013statistical,zeng2014decreasing} is defined as
\begin{align}
\label{eqn_OWL}
\Omega_{\mathbf{w}}(\mathbf{x})=\sum_{i=1}^p w_i |x|^\downarrow_i.
\end{align}
The OSCAR regularizer \cite{bondell2008simultaneous} is a special case of the OWL norm in \eqref{eqn_OWL} when $w_i=\lambda_1 + \lambda_2 (p-i)$, where $\lambda_1, \lambda_2 \geq 0$.
We consider the OWL-regularized linear regression problem taking the following form:
\begin{align}
\label{eqn_OWL_regression}
\textnormal{Minimize~}_{\mathbf{x} \in \mathbb{R}^p} \|\mathbf{y}-\mathbf{A} \mathbf{x} \|_2^2 + \lambda \Omega_{\mathbf{w}}(\mathbf{x}),
\end{align}
where $\mathbf{y} \in \mathbb{R}^n$ is the vector of $n$ noisy measurements, $\mathbf{A} \in \mathbb{R}^{n \times p}$ is the design matrix, and $\lambda \geq 0$ is the regularization parameter of the OWL norm.
The seminal work in \cite{figueiredo2016ordered} establishes a finite-sample error bound on the OWL-regularized regression method under the measurement model
\begin{align}
\label{eqn_measurement_model}
\mathbf{y}=\mathbf{A} \mathbf{x}^* + \boldsymbol{\nu},
\end{align}
where $\mathbf{x}^* \in \mathbb{R}^p$ is an $s$-sparse vector (i.e., the signal) and $\boldsymbol{\nu} \in \mathbb{R}^n$ is the measurement error (i.e., the noise). Let $\overline{\mathbf{x}}^*$ denote the vector with identical coefficients corresponding to identical columns in $\mathbf{A}$, and assume the entries in each column of $\mathbf{A}$ are i.i.d. $\mathcal{N}(0,1)$ (standard Gaussian random variables) but different columns could be strongly correlated.
Consider the $\ell_1$-norm bounded measurement error constraint $\frac{\| \boldsymbol{\nu}\|_1}{n} \leq \epsilon $, then the solution $\widehat{\mathbf{x}}$ to \eqref{eqn_OWL_regression} is guaranteed to satisfy the finite-sample error upper bound \cite{figueiredo2016ordered}:
\begin{align}
\label{eqn_upper_bound}
\mathbb{E} \|\widehat{\mathbf{x}} - \overline{\mathbf{x}}^* \|_2 \leq C_1 \left( C_2 \|\mathbf{x}^*\|_2 \sqrt{\frac{s \log p}{n}} + \epsilon \right),
\end{align}
where $\|\cdot\|_2$ denotes the $\ell_2$ norm, the expectation $\mathbb{E}$ is taken over the random design matrix $\mathbf{A}$, and $C_1$ and $C_2$ are some positive constants that we omitted for brevity (see Theorem 1.1 in \cite{figueiredo2016ordered}). Note that in this finite-sample analysis no distributional assumptions are imposed on the measurement error $\boldsymbol{\nu}$ other than its bounded $\ell_1$ norm.
Similar error bound can be obtained when the rows of $\mathbf{A}$ are i.i.d. Gaussian random vectors \cite{figueiredo2016ordered}.
In general, the solution $\widehat{\mathbf{x}}$ to \eqref{eqn_OWL_regression} can be efficiently obtained by leveraging the proximal operator $\textnormal{Prox}_{\Omega_{\mathbf{w}}}(\cdot)$ of OWL regularization. As illustrated in \cite{zeng2014ordered}, one can use the fast iterative shrinkage-thresholding algorithm (FISTA) \cite{beck2009fast} to obtain $\widehat{\mathbf{x}}$, which includes iterating the following optimization steps:
\begin{enumerate}
\item OWL proximal gradient descent -- \\ $\mathbf{x}^{(k)}=\textnormal{Prox}_{\lambda \Omega_{\mathbf{w}}} \left( \mathbf{u}^{(k)} - \mathbf{A}^T (\mathbf{A} \mathbf{u}^{(k)} - \mathbf{y}) / \alpha_k \right)$
\item Momentum -- $\mathbf{u}^{(k+1)}=\mathbf{x}^{(k)} + \beta_k \left( \mathbf{x}^{(k)} - \mathbf{x}^{(k-1)} \right) $
\item Update $\alpha_k$ and $\beta_k$ if not converged
\end{enumerate}
The index $k$ denotes the FISTA iteration, and $\alpha_k$ and $\beta_k$ denote the inverse of the step size and the momentum coefficient, respectively. The notation $\cdot^T$ denotes matrix transpose.
Specifically, when the OWL regularizer reduces to OSCAR, its approximate proximity operator (APO) has a closed-form expression given by \cite{zeng2014solving}
\begin{align}
\label{eqn_APO}
\textnormal{Prox}_{\textnormal{OSCAR-APO}}(\mathbf{z})=\textnormal{sign}(\mathbf{z}) \odot \max \{ |\mathbf{z}| - \widetilde{\mathbf{w}},0 \},
\end{align}
where $\textnormal{sign}(\cdot)$ is the vector of entry-wise sign function ($+1$ or $-1$), $\odot$ denotes entry-wise product, $\max\{\mathbf{z},0\}$ denotes entry-wise maximum value between $z_i$ and $0$, and $\widetilde{\mathbf{w}} = \mathbf{P}(\mathbf{z})^T \mathbf{w}$, where $\mathbf{P}(\mathbf{z})$ is a permutation matrix associated with a given vector $\mathbf{z} \in \mathbb{R}^p$ satisfying $\mathbf{P}(\mathbf{z}) |\mathbf{z}|= |\mathbf{z}|^\downarrow$, which can be obtained by taking $|\mathbf{z}|$ and sorting its entries in descending order. In addition, for OSCAR the vector $\mathbf{w}$ in OWL has the relation $w_i=\lambda_1 + \lambda_2 (p-i)$, where $\lambda_1, \lambda_2 \geq 0$.
\section{Main Results}
Although a finite-sample analysis for OWL-regularized regression has been established under the $\ell_1$-norm constrained measurement error $\epsilon$ in \cite{figueiredo2016ordered}, motivated by the recent advances in adversarial machine learning, we are interested in investigating its robustness to adversarially designed noises satisfying the same error budget $\epsilon$. In other words, given an $\ell_1$-norm bounded threat model $\frac{\|\boldsymbol{\nu}\|_1}{n} \leq \epsilon$ and a design matrix $\mathbf{A}$, we aim to find an optimal noise $\boldsymbol{\nu}^*$ that could maximally degrade the performance of OWL-regularized regression in terms of the detected feature groups. In this paper,
we particularly focus on the case of OSCAR with APO as its solver.
Under the same measurement model as in \eqref{eqn_measurement_model}, we formulate the problem of finding a norm-bounded noise $\boldsymbol{\nu}^*$ that could maximally deviate the OWL-regularized regression $\widehat{\mathbf{x}}$ from the ground-truth feature cluster membership vector $\overline{\mathbf{x}}^*$ by solving
\begin{align}
\label{eqn_Adv_OWL}
&\textnormal{Maximize}_{\boldsymbol{\nu} \in \mathbb{R}^n}~ \| \widehat{\mathbf{x}}(\boldsymbol{\nu}) - \overline{\mathbf{x}}^* \|_2^2 \\
&\textnormal{subject~to~} \|\boldsymbol{\nu}\|_1 / n \leq \epsilon, \\ &\widehat{\mathbf{x}}(\boldsymbol{\nu})=\arg \min_{\mathbf{x} \in \mathbb{R}^p} \| \mathbf{y} - \mathbf{A} \mathbf{x} \|_2^2 + \lambda \Omega_{\mathbf{w}}(\mathbf{x}).
\label{eqn_Adv_OWL_2}
\end{align}
Essentially, our adversarial formulation studies the robustness of OWL-regularized regression in the worst-case scenario by exploring the space of constrained measurement noise to maximize the feature group identification loss in \eqref{eqn_Adv_OWL}.
In our setting, we assume the adversary has access to the ground-truth vector $\mathbf{x}^*$ so that based on \eqref{eqn_measurement_model}, \eqref{eqn_Adv_OWL_2} can be written as
\begin{align}
\label{eqn_Adv_OWL_ground}
\widehat{\mathbf{x}}(\boldsymbol{\nu})=\arg \min_{\mathbf{x} \in \mathbb{R}^p} \| \mathbf{A} (\mathbf{x}^*-\mathbf{x}) + \boldsymbol{\nu} \|_2^2 + \lambda \Omega_{\mathbf{w}}(\mathbf{x}).
\end{align}
Next we specify how to solve the adversarial regression formulation in \eqref{eqn_Adv_OWL} to \eqref{eqn_Adv_OWL_2}
in the case of OSCAR with APO as its solver. Let $\mathbf{u}^*$ and $\alpha^*$ be the final iterates of the momentum and step size terms in FISTA as described in Section \ref{sec_background}. Given the measurement model \eqref{eqn_measurement_model} under OSCAR-APO, \eqref{eqn_Adv_OWL_ground} becomes
\begin{align}
\label{eqn_OSCAR_APO}
\widehat{\mathbf{x}}(\boldsymbol{\nu})&=\textnormal{Prox}_{\textnormal{OSCAR-APO}} \left( \mathbf{u}^* - \mathbf{A}^T (\mathbf{A} \mathbf{u}^*- \mathbf{y})/ \alpha^* \right) \nonumber \\
&=\textnormal{Prox}_{\textnormal{OSCAR-APO}} \left( \mathbf{u}^* - \mathbf{A}^T (\mathbf{A} \mathbf{u}^*- \mathbf{A} \mathbf{x}^* - \boldsymbol{\nu})/ \alpha^* \right),
\end{align}
where $\textnormal{Prox}_{\textnormal{OSCAR-APO}}(\cdot)$ is defined in \eqref{eqn_APO}.
With the method of Lagrange multipliers, we are interested in solving the following alternative optimization problem
\begin{align}
\label{eqn_Adv_OWL_Lagrange}
&\textnormal{Minimize}_{\boldsymbol{\nu} \in \mathbb{R}^n} -\| \widehat{\mathbf{x}}(\boldsymbol{\nu}) - \overline{\mathbf{x}}^* \|_2^2 + \frac{\gamma}{n} \|\boldsymbol{\nu}\|_1,
\end{align}
where $\gamma >0$ is a tunable regularization coefficient such that the solution $\boldsymbol{\nu}^*$ to \eqref{eqn_Adv_OWL_Lagrange} will satisfy the norm constraint $\|\boldsymbol{\nu}^*\|_1/n \leq \epsilon$.
We note that the formulation in \eqref{eqn_Adv_OWL_Lagrange} falls into the category of LASSO problem \cite{tibshirani1996regression} and can be efficiently solved by using optimization methods such as iterative shrinkage-thresholding algorithm (ISTA). Specifically, let $f(\boldsymbol{\nu})= -\| \widehat{\mathbf{x}}(\boldsymbol{\nu}) - \overline{\mathbf{x}}^* \|_2^2$. Then $\boldsymbol{\nu}^*$ (possibly a local optimum) can be obtained by iteratively solving
\begin{align}
\boldsymbol{\nu}^{(k+1)}=S_{\gamma/n} \left( \boldsymbol{\nu}^{(k)} - \eta_k \nabla f(\boldsymbol{\nu}^{(k)}) \right),
\end{align}
where $\boldsymbol{\nu}^{(k)}$ denotes the $k$-th iterate, $\eta_k$ is the step size, $\nabla f$ denotes the gradient of $f$\footnote{We use the subgradient of $f$ at points where $f$ is not differentiable.}, and $S_{\rho}(\mathbf{z}): \mathbb{R}^n \mapsto \mathbb{R}^n$ is an entry-wise function defined as
\begin{align}
S_{\rho}(z_i)=
\left\{
\begin{array}{ll}
z_i - \rho, & \text{~if~} z_i > \rho ; \\
0, & \text{~if~} |z_i| \leq \rho ; \\
-z_i + \rho, & \text{~if~} z_i < -\rho.
\end{array}
\right.
\end{align}
In what follows, we explicitly derive the gradient $\nabla f(\boldsymbol{\nu})=[\frac{\partial f}{\partial \nu_1},\frac{\partial f}{\partial \nu_2},\ldots,\frac{\partial f}{\partial \nu_n}]^T$ of $f(\boldsymbol{\nu})$ with respect to $\boldsymbol{\nu}$ for OSCAR-APO. For clarity, the index $i \in \{1,\ldots,n\}$ specifies the measurement instance, and the index $j \in \{1,\ldots,p\}$ specifies the covariate instance. To simplify the notation, let $\mathbf{b}(\boldsymbol{\nu})=\mathbf{u}^* - \mathbf{A}^T (\mathbf{A} \mathbf{u}^*- \mathbf{A} \mathbf{x}^* - \boldsymbol{\nu})/ \alpha^*$ such that $\widehat{\mathbf{x}}(\boldsymbol{\nu})=\textnormal{Prox}_{\textnormal{OSCAR-APO}}(\mathbf{b})=\textnormal{sign}(\mathbf{b}) \odot \max \{ |\mathbf{b}| - \widetilde{\mathbf{w}},0 \} $ based on \eqref{eqn_APO}, where $\widetilde{\mathbf{w}}=\mathbf{P}(\mathbf{b})^T \mathbf{w}$. Rewriting $f(\boldsymbol{\nu})=-\sum_{j=1}^p ([\widehat{\mathbf{x}}(\boldsymbol{\nu})]_j - \overline{x}^*_j)^2=-\sum_{j=1}^p [ \textnormal{sign}(b_j) \cdot \max\{|b_j| - \widetilde{w}_j,0\} - \overline{x}^*_j]^2$ and using chain rule, we can obtain
\begin{align}
\label{eqn_partial_der}
\frac{\partial f}{\partial \nu_i} &= \frac{\partial f}{\partial \mathbf{b}}^T \frac{\partial \mathbf{b}}{\partial \nu_i} \nonumber \\
& = \sum_{j=1}^p \frac{\partial f}{\partial b_j} \cdot \frac{\partial b_j}{\partial \nu_i} \nonumber \\
&= \sum_{j:|b_j| > \widetilde{w}_j} -2 \left( b_j - \textnormal{sign}(b_j) \widetilde{w}_j - \overline{x}^*_j \right) \cdot \frac{A_{ij}}{\alpha^*} \nonumber \\
&= \sum_{j=1}^p -2 \left( b_j - \textnormal{sign}(b_j) \widetilde{w}_j - \overline{x}^*_j \right) \cdot \frac{A_{ij}}{\alpha^*} \cdot I_{\{|b_i| > \widetilde{w}_j\}},
\end{align}
where $I_E$ is an indicator function such that $I_E=1$ if event $E$ is true; otherwise $I_E=0$.
\begin{algorithm}[t]
\caption{Adversarial perturbation for OSCAR-APO}
\label{algo_OSCAR_APO}
\begin{algorithmic}
\State \textbf{Input:} $\mathbf{A}$, $\mathbf{x}^*$, $\overline{\mathbf{x}}^*$, $\mathbf{w}$, $\mathbf{u}^*$, $\alpha^*$, $\epsilon$, $\{\eta_k\}$
\State \textbf{Output: $\boldsymbol{\nu}^*$}
\State \textbf{Initialization:} $k=0$, $\gamma=\gamma_0$, $\mathbf{g} \sim \mathcal{N}(0,\mathbf{I}_n)$ and
$\boldsymbol{\nu}^{(0)} = \epsilon \cdot \mathbf{g}/\|\mathbf{g}\|_1 $
\While{not converged}
\State 1. $\mathbf{b}=\mathbf{u}^* - \mathbf{A}^T \left( \mathbf{A} \mathbf{u}^*- \mathbf{A} \mathbf{x}^* - \boldsymbol{\nu}^{(k)} \right) / \alpha^*$
\State 2. Find the permutation $\mathbf{P}(\mathbf{b})$ s.t. $\mathbf{P}(\mathbf{b})|\mathbf{b}|=|\mathbf{b}|^\downarrow$
\State 3. $\widetilde{\mathbf{w}} = \mathbf{P}(\mathbf{b})^T \mathbf{w}$
\State 4. $\frac{\partial f}{\partial \nu_i}= \sum_{j:|b_j| > \widetilde{w}_j} -2 \left( b_j - \textnormal{sign}(b_j) \widetilde{w}_j - \overline{x}^*_j \right) \cdot \frac{A_{ij}}{\alpha^*}$
\State ~~~~~for all $i \in \{1,\ldots,n\}$
\State 5. $\boldsymbol{\nu}^{(k+1)}=S_{\gamma/n} \left( \boldsymbol{\nu}^{(k)} - \eta_k \nabla f(\boldsymbol{\nu}^{(k)}) \right)$
\State 6. $k \leftarrow k+1$
\EndWhile
\State $\boldsymbol{\nu}^* \leftarrow \boldsymbol{\nu}^{(k)}$
\If{ $\|\boldsymbol{\nu}^*\|_1 / n > \epsilon$}
\State Reinitialization with a larger $\gamma$ and redo the while loop
\EndIf
\end{algorithmic}
\end{algorithm}
\begin{figure*}
\centering
\includegraphics[scale=0.9]{figures/adversarial.pdf}
\vspace{-6mm}
\caption{Assessing effect of our proposed adversarial attack on OSCAR (rightmost column) by varying the attack strength $\epsilon$ with $\epsilon=0.05$ (top row), $\epsilon=0.1$ (second row), $\epsilon=0.2$ (third row) and $\epsilon=0.3$ (fourth row). In each column, the ground truth refers to $\mathbf{x}^*$, OSCAR refers to the regression results without noise, and OSCAR + Attack refers to the regression results against our designed adversarial noises. The feature groups can be adversarially misaligned even for small $\epsilon$.}
\label{fig:oscarattack}
\vspace{-2mm}
\end{figure*}
The detailed derivations are as follows. To obtain $\frac{\partial f}{\partial b_j}$, we divide the analysis into three cases based on the value of $b_j$:
\begin{itemize}
\item Case I -- If $|b_j| \leq \widetilde{w}_j $, then $[\widehat{\mathbf{x}}(\boldsymbol{\nu})]_j = 0$ and hence $\frac{\partial f}{\partial b_j}=0$.
\item Case II -- If $b_j > \widetilde{w}_j $, then $[\widehat{\mathbf{x}}(\boldsymbol{\nu})]_j=\textnormal{sign}(b_j) \cdot (b_j - \widetilde{w}_j)=b_j - \widetilde{w}_j$ since $\textnormal{sign}(b_j)=1$. Therefore, $\frac{\partial f}{\partial b_j}=-2 (b_j - \widetilde{w}_j - \overline{x}^*_j).$ We note that technically, $\widetilde{\mathbf{w}}$ is also a function of $\mathbf{b}$ and hence a function of $\boldsymbol{\nu}$. As a result, one needs further chain rule factorization $\frac{\partial f}{\partial b_j} = \frac{\partial f}{\partial~b_j - \widetilde{w}_j} \cdot \frac{\partial~b_j - \widetilde{w}_j}{\partial b_j}$. Here we implicitly use the fact that $\frac{\partial \widetilde{w}_j}{\partial b_j}=\frac{\partial [\mathbf{P}(\mathbf{b})^T \mathbf{w}]_j}{\partial b_j}=0$ since no matter how we permute the entries of $\mathbf{w}$, it is still a constant vector.
\item Case III -- If $-b_j> \widetilde{w}_j $, then $[\widehat{\mathbf{x}}(\boldsymbol{\nu})]_j= \textnormal{sign}(b_j) \cdot (-b_j - \widetilde{w}_j)=b_j + \widetilde{w}_j$. Similar to the analysis of Case II, we obtain
$\frac{\partial f}{\partial b_j}=-2 (b_j + \widetilde{w}_j - \overline{x}^*_j) $.
\end{itemize}
Summarizing these cases, we have a simplified expression $\frac{\partial f}{\partial b_j}=-2 (b_j - \textnormal{sign}(b_j) \widetilde{w}_j - \overline{x}^*_j) $ if $|b_j|>\widetilde{w}_j$ and $\frac{\partial f}{\partial b_j}=0$ otherwise. Next, to obtain $\frac{\partial b_j}{\partial \nu_i}$, based on the definition of $\mathbf{b}$ we have $b_j = \textnormal{constant} + \sum_{k=1}^n A^T_{jk} \nu_k/\alpha^*=\sum_{k=1}^n A_{kj} \nu_k/\alpha^*$. Therefore, $\frac{\partial b_j}{\partial \nu_i}=A_{ij}/\alpha^*$. Finally, combining the analysis of $\frac{\partial f}{\partial b_j}$ and
$\frac{\partial b_j}{\partial \nu_i}$, we obtain the results in \eqref{eqn_partial_der}.
The pipeline of crafting a norm-bounded adversarial perturbation $\boldsymbol{\nu}^*$ for OSCAR-APO is describe in Algorithm \ref{algo_OSCAR_APO}. We also note that beyond OSCAR-APO, it is possible to craft an adversarial noise $\boldsymbol{\nu}^*$ for generic OWL-regularized regression methods by treating \eqref{eqn_Adv_OWL_2} as a black-box function, which will be considered in our future work.
\section{Performance Evaluation}
In this experiment, we generated a synthetic dataset of $p=100$ features and $n=50$ instances with a pre-defined grouping structure among the features. The features (entries in $\mathbf{A}$) were generated using a standard Gaussian distribution and we modeled the response variable using \eqref{eqn_measurement_model}. Given a norm-bounded noise strength $\epsilon$, we call the adversarial perturbation found using our proposed approach (Algorithm \ref{algo_OSCAR_APO}) an ``attack'' on the considered regression method. For the attack parameters, we set $\gamma_0=0.1$ and $\eta_k=10^{-3}$.
In Figure~\ref{fig:oscarattack}, the x-axis represents the feature index and the y-axis represents the coefficient values. The left column represents the ground-truth $\mathbf{x}^*$ with two defined feature groups. The middle column shows the feature grouping obtained after running OSCAR algorithm in the noiseless setting. The right column shows how the grouping is adversely affected after our attack.
We varied the noise budget $\epsilon$ from 0.05 to 0.3 to assess the effect of the attack. One can observe that although some grouped features are retained up to a certain degree, the true effect of the attack can be seen on the features which are misaligned from their original feature groups, even for relatively small $\epsilon$.
\section{Conclusion and Future Work}
To study the robustness of OWL-regularized regression methods in the adversarial setting, this paper proposes a novel formulation for finding norm-bounded adversarial perturbations in the measurement model and illustrates the pipeline of adversarial noise generation in the case of OSCAR with APO as its solver. In the adversarial setting, the experimental results show that our proposed approach can effectively craft adversarial noises that severely degrade the regression performance in identifying ground-truth grouped features, even in the regime of small noise budgets. Our results indicate the potential risk of lacking robustness to adversarial noises in the tested regression method. One possible extension of our approach is to devise adversary-resilient regression methods. Our future work also includes developing a generic framework for generating adversarial noises for the entire family of OWL-regularized regression methods.
\vfill\pagebreak
\clearpage
|
\section{Introduction}
With the explosion of video data, automatic analysis of videos is increasingly becoming important. Examples of such videos include user videos, sports videos, TV videos or CCTV footages or for that matter videos coming from any other source. One of the popular requirements in this context is the ability to automatically summarize videos. Video summarization finds its uses in a wide variety of applications ranging from security and surveillance to compliance and quality monitoring to user applications aimed at saving time and storage. Broadly speaking, in terms of the kind of video summaries produced video summarization can be of two types - compositional video summarization, which aims at producing spatio-temporal synopsis \cite{pritch2007webcam,rav2006making,pritch2009clustered,pritch2008nonchronological} and extractive video summarization, which aims at selecting key frames (also called key frame extraction, static story boards or static video summarization, eg. \cite{de2011vsumm}) or shots (also called dynamic video summarization or dynamic video skimming, eg. \cite{GygliECCV14}). Past work has tried to address the problem of video summarization using unsupervised as well as supervised techniques. Early unsupervised techniques used attention models \cite{ma2005generic}, MMR \cite{li2010multi}, clustering \cite{de2011vsumm,mahmoud2013unsupervised} etc. and more recently auto encoder based techniques \cite{yang2015unsupervised} and LSTM based techniques \cite{mahasseni2017unsupervised} have been used. Use of supervised techniques began with forms of indirect supervision from external sources of information like related images \cite{khosla2013large}, similar videos \cite{chu2015video} or title of videos \cite{song2015tvsum}. Gong et al's work \cite{gong2014diverse} was the first work on video summarization which used direct high level supervision in form of user annotations. This was followed by Gygli et al's work \cite{gygli2015video}. Sharghi, Gong et al took this forward by incorporating a notion of user query in the produced video summaries thereby involving the user in the summary generation process \cite{sharghi2016query,sharghi2017query}. Theirs became the first work on query focussed video summarization. Then came several deep learning based video summarization techniques. For example, Zhang et al. \cite{zhang2016video} used LSTMs \cite{hochreiter1997long} to model long range dependencies among video key frames to produce summaries. Emphasis was given to sequential structures in videos and their modeling. Mahasseni et al.\cite{mahasseni2017unsupervised} used LSTM networks in an adversarial setting to generate summaries. Both these methods were domain agnostic. Some researchers viewed video summarization as a subset selection problem \cite{zhang2016summary,elhamifar2017online,sharghi2016query,sharghi2017query}. Another key approach was to combine the elements of both deep learning and subset selection in summary generation. Gong et al. achieved this using their seqDPP architecture \cite{gong2014diverse} where an LSTM network was coupled with a DPP (Determinantal Point Process) \cite{kulesza2012determinantal}.
In this work we address the challenge of domain specific dynamic video summarization. A good video summary should be a meaningful short abstract of a long video, it must contain \emph{important} events, it should exhibit some continuity and it should be free from redundancy. However, the notion of importance varies with domain. For example a "six" or a "wicket" could be considered important in cricket videos while "entry of birthday girl" and "cutting of cake" would be considered important in birthday videos. Further the characteristics of a good summary like representativeness, coverage or diversity also depends on the domain. For example, while for surveillance videos a good summary should contain outliers, for a user video coverage and representativeness become more important. We hereby propose a novel domain specific video summarization framework which automatically learns to produce summaries that possess desired characteristics suitable for that domain. Past work on video summarization has focused either on using supervised approaches for ranking the snippets/segments thereby producing a video summary (eg. \cite{potapov2014category,ma2005generic}) or on unsupervised approaches of generating the video summary as a subset of snippets/segments with desired characteristics of representativeness, diversity, coverage, etc. (eg. \cite{li2010multi,de2011vsumm}). Some work has also focused on learning the relative importance of uniformity, interestingness and representativeness for different domains (eg. \cite{gygli2015video,sahoo2017unified}). None of these works, however, have looked at the joint problem of generating domain specific summaries by automatically learning the concepts that are deemed important for a domain, together with having the desired summary characteristics like diversity, coverage etc. for that domain. Building upon the max margin structured learning framework in \cite{gygli2015video} we learn a mixture of modular and submodular terms. Modular terms help to capture shots more important to the domain under consideration and submodular terms help to capture characteristics of summary important to that domain. The different weights learnt for different components indicate the varying notion of importance from domain to domain.
Further, having many possible ground truth summaries has been posed as one of the challenges in video summarization~\cite{truong2007video}. Multiple ground truths are due to the difference in perspectives and the fact that different visual content can have the same semantic meaning. However, one reason is also a lack of any other information about the video. Dealing with domain specific videos, however, allows us to define a notion of importance ratings that are unique and unambiguous for that domain. Such ratings can be seen as a unified representation of all possible ground truth summaries of a video. We establish the importance of ratings in the training dataset in producing good summaries as against binary inclusion/exclusion information. Instead of getting ground truth user summaries from annotators we rather ask them to provide ratings for segments in the entire video. The framework learns to generate more accurate summaries when it is provided more supervision this way. We thus establish that the more efficient way of incorporating domain specific relevances into a summary is to provide a supervision in form of ratings as against multiple ground truths.
We also define a new evaluation measure which is more naturally suited for this task than other standard measures used in literature like F1 score. Our measure evaluates video summaries considering the ratings and not just binary inclusion/exclusion information. It also evaluates summaries not only with respect to what they \emph{should} contain but also with respect to what they \emph{should not} contain and the degree of diversity.
As a part of this work we will also release a gold standard dataset for furthering research in domain specific video summarization. To the best of our knowledge, ours is the first dataset with long videos across several domains with rating annotations.
\section{Related Work and Our Contributions}
\subsection{Domain specific video summarization}
One of the earliest works on domain specific video summarization was by Potapov et al.\cite{potapov2014category} in 2014. They were one of the first to realize the importance of building separate models for summarization for distinct categories of videos. They used an SVM\cite{hearst1998support} classifier conditioned on video category to produce summaries. The SVM learns to score the segments according to their importance to the domain. The segments having higher scores are then selected greedily and put in temporal order to create the final summary. Sun et al. \cite{sun2014ranking} analyzed edited videos of a particular domain as an indicator of highlights of that domain. After finding pairs of raw and corresponding edited videos, they obtain pair-wise ranking constraints to train their model. Zhang et al. \cite{zhang2016summary} use supervision in the form of human-created summaries to perform automatic keyframe-based video summarization. Their main idea is to nonparametrically transfer summary structures of a particular domain of videos from annotated videos to unseen test videos. By learning a joint model to understand what snippets and what characteristics are important for a domain, we use a more principled approach with a form of supervision which is more efficient.
\subsection{Submodular functions for video summarization}
Video summarization can be viewed as a subset selection problem subject to certain constraints. Given a set $V = \{1, 2, 3, \cdots, n\}$ of items which we also call the \emph{Ground Set}, define a utility function (set function) $f:2^V \rightarrow \mathbf{R}$, which measures how good a subset $X \subseteq V$ is. Let $c :2^V \rightarrow \mathbf{R}$ be a cost function, which describes the cost of the set (for example, the size of the subset). Often the cost $c$ is budget constrained (for example, a fixed set summary) and a natural formulation of this is the following problem:
\begin{align}
\max\{f(X) \mbox{ such that } c(X) \leq b\}
\end{align}
The goal is then to have a subset $X$ which maximizes $f$ while simultaneously minimizing the cost function $c$. It is easy to see that maximizing a generic set function becomes computationally infeasible as $V$ grows.
A special class of set functions, called submodular functions ~\cite{nemhauser1978analysis}, however, makes this optimization easy. A function $f : 2^V \rightarrow R$ is submodular if for every
$A \subseteq B \subseteq V$ and $e \in V$ and $e \notin B$ it holds that
\begin{align}
f(\{e\} \cup A) - f(A) \geq f(\{e\} \cup B) - f(B)
\end{align}
Likewise,a function $f : 2^V \rightarrow R$ is supermodular if for every
$A \subseteq B \subseteq V$ and $e \in V$ and $e \notin B$ it holds that
\begin{align}
f(\{e\} \cup A) - f(A) \leq f(\{e\} \cup B) - f(B)
\end{align}
Submodular functions exhibit a property that intuitively formalizes the idea of ``diminishing returns''. That is, adding some instance $x$ to the set $A$ provides more gain in terms of the target function than adding $x$ to a larger set $A'$, where $A \subseteq A'$. Informally, since $A'$ is a superset of $A$ and already contains more information, adding $x$ will not help as much. Using a greedy algorithm to optimize a submodular function (for selecting a subset) gives a lower-bound performance guarantee of around 63\% of optimal ~\cite{nemhauser1978analysis} to the above problem, and in practice these greedy solutions are often within 98\% of optimal ~\cite{krause2008optimizing}. This makes it advantageous to formulate (or approximate) the objective function for data selection as a submodular function.
This concept has been used in document summarization \cite{lin2012learning} and in image collection summarization \cite{tschiatschek2014learning}. More recently, Elhamifar et al. \cite{elhamifar2017online} used submodular optimization for online video summarization by performing incremental subset selection. One of the first attempts to summarize videos using a submodular mixture of objectives was by Gygli et al.~\cite{gygli2015video}. They, however, did not distinguish between various domains of videos and had a specially crafted video frame interestingness model which played a significant role in the summaries produced. Building upon the approach in~\cite{gygli2015video} we learn a mixture for a domain to produce summary specific to that domain. However, in our work, domain specific importance of snippets/shots is not predicted separately using another model. Rather weighted features are directly used in the mixture as modular terms. These modular components capture the shot level domain importance while the other submodular and supermodular components in the mixture correspond to different desired characteristics of the summary like diversity and coverage.
\subsection{Evaluation measures}
Different measures have been reported in literature for the purpose of accurately evaluating the quality of the video summary produced. VIPER \cite{doermann2000tools} addresses the problem by defining a specific ground truth format which makes it easy to evaluate a candidate summary. SUPERSEIV \cite{huang2004automatic} is an unsupervised technique to evaluate video summarization algorithms that perform frame ranking. VERT \cite{li2010vert} was inspired by BLEU in machine translation and ROUGE in text summarization. More recently approaches by Yeung et al. \cite{yeung2014videoset} and Plummer et al. \cite{plummer2017enhancing} also used text based evaluation methods. De Avila et al. \cite{de2011vsumm} propose a method of evaluation which considers several ground truth summaries. Others, like \cite{khosla2013large,GygliECCV14,elhamifar2017online} and \cite{xu2015gaze} more directly use precision and recall type measures. Kannappan et al. \cite{kannappan2016pertinent} propose an approach which is only for static video summaries. The evaluation measure proposed by Potapov et al. \cite{potapov2014category} is capable of evaluating a summary only against one ground truth. As annotations are done by several users producing several ground truth summaries, they evaluate those annotations against each other to form some kind of an upper bound on performance. Approaches like \cite{zhao2014quasi,conf/cvpr/SongVSJ15,gygli2015video} and \cite{zhang2016video} combine several ground truths into one before using them for evaluation. This comes at the cost of losing individual opinion. In search for a measure which would work directly on ratings (which is a potential generator of multiple ground truths) and having certain other desired characteristics (as enumerated in the corresponding section below) we developed our own evaluation measure.
\subsection{Datasets for video summarization}
Different researchers in the past have released different datasets for the purpose of video summarization. Examples include The Video Summarization (SumMe) dataset \cite{GygliECCV14}, MED Summaries dataset \cite{potapov2014category} and Title-based Video Summarization (TVSum) dataset\cite{conf/cvpr/SongVSJ15}. However, none of these existing datasets were found suitable for our work for following reasons. Firstly, we aim to summarize videos across a large number of domains like surveillance, sports, user etc. in a single framework. For that purpose we need a wide variety of videos summarized uniformly. The various datasets only provide certain subsets of types of videos and they use vastly different methods to annotate those videos. So, it was essential that we had a uniformly annotated, diverse set of videos from diverse domains. Secondly, we wanted to test our method on long videos, as the true benefit of a video summary in real world applications is seen only with respect to long videos. Thirdly, we wanted the annotations to not only capture what is important, but also what is \emph{not} important and what is repetitive. Identifying segments which are relatively long and contain repetitive information (for example, scene of spectators clapping for 5 minutes in a cricket video) and retaining only a fraction of them to be included in the summary, is essential to having a good quality summary.
\subsection{Our Contributions}
In the following, we summarize the main contributions of this paper.
\begin{itemize}
\item We address the problem of Domain specific video summarization, by jointly ranking the most important portions of the video for that domain (for example, a goal in Soccer), while simultaneously capturing diverse and representative shots. We do this by training a joint mixture model with features which capture domain importance along with diversity models.
\item We introduce a novel evaluation criteria which captures these aspects of a summary, and also introduce a large dataset for domain specific summarization. Our dataset comprises of several long videos for different domains (surveillance, personal videos, sports etc.) and to our knowledge is the first domain specific video summarization dataset with long videos.
\item We then empirically demonstrate various interesting insights. a) We first show that by jointly modeling diversity, relevance and importance, we can learn substantially superior summaries on all domains compared to just learning any one of these aspects. b) We next show that by learning on the same domain, we can obtain superior results than using learnt mixtures from other domains, thus proving the benefit for domain specific video summarization. c) We then look at the top components learnt for different domains and show how those individual components perform best for that domain if considered in isolation. Moreover, we observe how intuitively it makes sense that these components are important. For example, in surveillance, we see that diversity functions tend to have high ranking compared to other models (as finding outliers is more important there), while in personal videos (like birthday), we see that representation is important. We moreover also look at the highest ranking snippets based on these components and show how they capture the most important aspects of that domain (like a goal in soccer).
\end{itemize}
The major contribution of this work is that this is the first systematic study of domain specific video summarization on large videos and we provide several insights into the role of different summarization models for this problem.
\section{Methodology}
We begin by creating a training dataset comprising of videos from several categories annotated with ratings information. Our method works on ratings and hence better deals with issue of multiple ground truths. Building upon the approach in \cite{gygli2015video}, we create a mixture, but our mixture contains modular terms (to capture the domain specific importance of snippets) and submodular terms (for imparting certain desired characteristics to the summary). For each training video of a domain, the components of the mixture are instantiated and the weights of the complete mixture for that domain are learnt using max margin learning framework. After the training phase, for any given test video of that domain, the weighted mixture is then maximized to produce the desired summary video. Below we describe details of every step in the above methodology.
\subsection{Training Data}
In this work we focus on videos from five different domains - birthday, cricket, soccer, entry exit and office. The latter two are surveillance videos taken from CCTV cameras installed at various entry/exit locations and offices respectively. We have collected birthday, cricket and soccer videos from internet (existing published datasets / youtube). Due to privacy reasons and to be able to experiment with presence/absence of abnormal events in the surveillance videos, we have collected surveillance videos from our own setup of surveillance cameras. \arxiv{Table \ref{tab:table1} shows distribution of number and duration of videos across these categories.
\begin{table}[h!]
\begin{center}
\caption{Distribution of videos across categories}
\label{tab:table1}
\begin{tabular}{l|l|l}
\textbf{Category} & \textbf{Number of Videos} & \textbf{Duration in mins}\\
\hline
Cricket & 7 & 276\\
Birthday & 9 & 136\\
Soccer & 11 & 609\\
Entry Exit & 21 & 306\\
Office & 33 & 687\\
\end{tabular}
\end{center}
\end{table}}
\narxiv{We have 7 Cricket videos (of 276 mins), 9 of Birthday (136 minutes total length) and 21 of Entry Exit (306 minutes).}
Next, for each domain, we go over every video and first prepare a table of different kinds of scenes that occur across different videos in that category and using domain knowledge we assign ratings to those scenes. Negative ratings are assigned to segments which \emph{must not} be included in the summary. Since the ratings are relative (for example, a 2 rated scene is supposed to be more important than a 1 rated scene but less important than a 3 rated scene) it was necessary to gather information about all scenes before starting to rate the scenes in the specific videos. Also going through the extra step of creating such a scenes document for each category enabled us to come up with a consistent philosophy of ratings and consequently a very high inter annotator correlation. Using this scenes table as annotation guidelines for each category, the annotators then annotated the videos in each category. For each video, three different annotators were independently asked to identify the segments based on the scenes and label them. Segments that are long and contain repetitive content are explicitly marked repetitive in addition to their rating. For the purpose of annotating, we customized a tool called oTranscribe \cite{oTranscribe} to make the annotation task easy and to produce the desired annotation JSON. The oTranscribe interface was cleaned up and a lot more keyboard shortcuts were added in for ease of annotation. Shortcuts were added in to mark the beginning and ends of segments, to rate segments, to give a short description of the segment, to mark a segment repetitive and to skip to the end and beginning of the previous segment. Finally, hooks were added in to output the annotation as a JSON file. The annotations from different annotators were combined based on majority agreement. As a sanity check, we visually verified the annotations thus produced by looking at the annotated videos which were produced by overlaying the labels on top of the original videos.
\subsection{Learning framework}
The task of video summarization is posed as a discrete optimization problem for finding the best subset representing the summary. Given a video $V$ we split it into a snippets $v_i$ of fixed length. Now we have a set $Y_v$ of all snippets in the video. Our problem reduces to picking $y$ $\subset$ $Y_v$ such that $|y| \leq k$ that maximizes our objective.\looseness-1
\begin{align} \label{max}
y^* = \operatorname*{argmax}_{y\subseteq Y_v, |y| \leq k} o(x_v, y)
\end{align}
$y^*$ is the predicted summary, $x_v$ the feature representation of the video snippets and $o(x_v, y)$ is the weighted mixture of components each capturing some aspect of the domain. Different weights are learnt for different domains.\looseness-1
\begin{align}
o(x_v, y) = w^Tf(x_v, y)
\end{align}
where $f(x_v, y) = [f_1(x_v, y),..., f_n(x_v, y)]$ and $f_i(x_v, y)$ are the various modular, submodular and supermodular components. Given $N$ pairs of a video and a reference summary $(V, y_{gt})$, we learn the weight vector $w$ by optimizing the following large-margin \cite{taskar2005learning} formulation:\looseness-1
\begin{align} \label{min}
\min\limits_{w \geq 0} \frac{1}{N} \sum_{n=1}^{N} L_n(w) + \frac{\lambda_1}{2}||w_1||^2 + \frac{\lambda_2}{2}||w_2||^2
\end{align}
where $L_n(w)$ is the generalized hinge loss of training example $n$ and $w_1$ and $w_2$ are the weight vectors for the modular terms and the submodular terms respectively.\looseness-1
\begin{align} \label{loss-aug}
L_n(w) = \max\limits_{y \subseteq Y_v^n} (w^T f(x_v^n, y) + l_n(y)) - w^T f(x_v^n, y_{gt}^n)
\end{align}
This objective is chosen so that each human reference annotation scores higher than any other summary by some margin. For training example $n$, the margin we chose is denoted by $l_n(y)$. We use $1 - normalizedScore(y)$ as margin where normalized score is computed using min-max normalization of the score generated by our evaluation measure, given the ratings, as described below.
\subsection{Components of the mixture}
Our mixture contains several hand picked components. Every component serves to impart certain characteristics to the optimal subset (the predicted summary).
\textbf{Set Cover:} For a subset $X$ being scored, the set cover is defined as $f_{sc}(X) = \sum_{u \in U} min\{m_u(X), 1\}$ $u$ is a concept belonging to a set of all concepts $U$, $m_u(X) = \sum_{x \in X} w_{xu}$ and $w_{xu}$ is the weight of coverage of concept $u$ by element $x$. This component governs the coverage aspect of the candidate summary and is monotone submodular.\looseness-1
\textbf{Probabilistic Set Cover:} This variant of the set cover function is defined as $f_{psc}(X) = \sum_{u \in U} (1 - \prod_{x \in X} (1 - p_{xu}))$ where $p_{xu}$ is the probability with which concept $u$ is covered by element $x$. Similar to the set cover function, this function governs the coverage aspect of the candidate summary, viewed stochastically and is also monotone submodular.\looseness-1
\textbf{Facility Location:} The facility location function is defined as $f_{fl}(X) = \sum_{v \in V} \max_{x \in X} sim(v,x)$ where $v$ is an element from the ground set $V$ and $sim(v, x)$ measures the similarity between element v and element x. Facility Location governs the representativeness aspect of the candidate summaries and is monotone submodular.\looseness-1
\textbf{Saturated Coverage} Saturated Coverage is $f_{satc}(X) = \sum_{v \in V} min\{m_v(X), c\}$ where $m_v(X) = \sum_{x \in X} sim(v, x)$ measures the relevance of set $X$ to item $v \in V$ and $c$ is a saturation hyper parameter that controls the level of coverage for each item $v$ by the set $X$. Saturated Coverage is similar to Facility Location except for the fact that for every category, instead of taking a single representative, it allows for taking potentially multiple representatives. Saturated Coverage is also monotone submodular.\looseness-1
\textbf{Generalized Graph-Cut} Generalized Graph Cut is $f_{gc}(X) = \sum_{i \in V, j \in X} sim(i, j) - \lambda \sum_{i, j \in X} sim(i, j)$
Similar to above two functions, Generalized Graph Cut also models representation. When $\lambda$ becomes large, it also tries to model diversity in the subset. $\lambda$ governs the tradeoff between representation and diversity. For $\lambda < 0.5$ it is monotone submodular. For $\lambda > 0.5$ it is non-monotone submodular. \looseness-1
\textbf{Disparity-min:} Denoting the distance measure between snippet/shot $i$ and $j$ by $d_{ij}$, disparity-min is defined as a set function $f_{disp}(X) = \min_{i, j \in X, i \neq j} d_{ij}$. It is easy to see that maximizing this function involves obtaining a subset with maximal minimum pairwise distance, thereby ensuring a diverse subset of snippets or shots. In principle this is similar to determinantal point processes (DPP), but DPP becomes computationally expensive at inference time. This function, though not submodular, can be efficiently optimized via a greedy algorithm.
\textbf{Continuity:} We work on a set of 2 second snippets as ground set. A summary (subset) could thus may not look continuous enough to give good viewing experience. Thus we add this continuity term in the mixture which would give more score when nearby snippets are chosen - this would ensure a visually more coherent and appealing summary. Essentially it is modeled as a redundancy function (this function is super-modular) within a shot as follows: $f_{cont}(X) = \sum_{s \in S} \sum_{x, x' \in s \cap X} w_{x, x'}$ where $S$ is the set of shots as a result of a shot boundary detection algorithm and $w_{x,x'}$ is the similarity between two snippets which can be defined as how close they are to each other based on their index. That is, the features used here are the indices of the snippets.
\textbf{Modular components:}
We use weighted features of snippets (described in the next section) as modular terms in the mixture.\\
\subsection{Features used for instantiating the components}
Let the video $\mathcal{V}$ be a set of frames $f_i, i = \{0, 1, 2, \cdots, n\}$. Let us define the ground set $V = \{S_0, S_1, S_2, \cdots, S_k\}$ as a set of \emph{snippets} where each snippet is 2 seconds long. A snippet for a video with frame rate $r$ would thus contain $2r$ consecutive frames. Feature vectors are calculated for each snippet independently by averaging the feature vectors of the frames/images in that snippet. For \narxiv{Different components of the mixture (as above) are instantiated for each video using the features such as VGG~\cite{simonyan2014very}, GoogleNet~\cite{szegedy2015going}, YOLO entities and features from Pascal VOC and COCO~\cite{redmon2016you} and Color Histogram features. A comprehensive list of the features used and how they are computed can be found in the extended version of this paper in the supplementary material}. \arxiv{Different components of the mixture (as above) are instantiated for each video using the features mentioned below:
\begin{itemize}
\item \textbf{vgg\_features:} \textit{fc6} layer of VGG19 \cite{simonyan2014very} trained on ImageNet dataset\cite{deng2009imagenet} is used as the feature vector for each image in the snippet. The feature vector for a snippet is an average of all the image feature vectors in that snippet. The feature size is 4096.
\item \textbf{googlenet\_features:} \textit{pool5\_7x7\_s1} layer of GoogLeNet \cite{szegedy2015going} trained on MIT Places dataset \cite{zhou2014learning} is used as the feature vector for each image in the snippet. The feature vector for a snippet is 1024-d and is an average of all the image feature vectors in that snippet.
\item \textbf{vgg\_p\_concepts:} The output probability layer of VGG19 \cite{simonyan2014very} trained on ImageNet dataset \cite{deng2009imagenet} is used as the feature vector for each image in the snippet. The feature vector of a snippet is an average of all the image feature vectors in that snippet. The feature size is 1000.
\item \textbf{googlenet\_p\_concepts:} The output probability layer GoogLeNet \cite{szegedy2015going} trained on MIT Places dataset \cite{zhou2014learning} is used as the feature vector for each image in the snippet. The feature vector of a snippet is an average of all the image feature vectors in that snippet and is 365 dimensional.
\item \textbf{vgg\_concepts:} The output probability layer of VGG19 \cite{simonyan2014very} trained on ImageNet dataset \cite{deng2009imagenet} is used as the feature vector for each image in the snippet. To create the feature vector of a snippet, an average of all the image feature vectors in that snippet is taken. This vector is one-hot-encoded based on a 0.5 threshold. The feature size is 1000.
\item \textbf{goolenet\_concepts:} The output probability layer of GoogLeNet \cite{szegedy2015going} trained on MIT Places dataset \cite{zhou2014learning} is used as the feature vector for each image in the snippet. To create the feature vector of a snippet, an average of all the image feature vectors in that snippet is taken. This vector is one-hot-encoded based on a 0.5 threshold. The feature size is 365.
\item \textbf{yolo\_coco\_concepts:} A vector of size 80 (corresponding to 80 classes in COCO dataset) where each component represents the average number of objects of the respective COCO \cite{lin2014microsoft} class, found by the YOLO \cite{redmon2016you} network trained on COCO dataset, in all images in the snippet.
\item \textbf{yolo\_voc\_concepts:} A vector of size 20 (corresponding to 20 classes in PASCAL VOC dataset) where each component represents the average number of objects of the respective PASCAL VOC\cite{Everingham10} class, found by the YOLO \cite{redmon2016you} network trained on PASCAL VOC dataset, in all images in the snippet.
\item \textbf{yolo\_coco\_p\_concepts:} A vector of size 81 (corresponding to 80 classes in COCO dataset and 1 dummy class) where each component represents the average fraction of objects of the respective COCO \cite{lin2014microsoft} class relative to the total number of objects detected, found by the YOLO \cite{redmon2016you} network trained on COCO dataset, in all images in the snippet. The last component is 1 if no objects were detected.
\item \textbf{yolo\_voc\_p\_concepts:} A vector of size 21 (corresponding to 20 classes in PASCAL VOC dataset and 1 dummy class) where each component represents the average fraction of objects of the respective PASCAL VOC \cite{Everingham10} class relative to the total number of objects detected, found by the YOLO \cite{redmon2016you} network trained on PASCAL VOC dataset, in all images in the snippet. The last component is 1 if no objects were detected.
\item \textbf{color\_hist\_r\_features:} A vector of size 256 representing the average red color histogram of the images in the snippet.
\item \textbf{color\_hist\_g\_features:} A vector of size 256 representing the average green color histogram of the images in the snippet.
\item \textbf{color\_hist\_b\_features:} A vector of size 256 representing the average blue color histogram of the images in the snippet.
\item \textbf{color\_hist\_h\_features:} A vector of size 180 representing the average hue histogram of the images in the snippet.
\item \textbf{color\_hist\_s\_features:} A vector of size 256 representing the average saturation histogram of the images in the snippet.
\end{itemize}}
\subsection{Evaluation measure}
To serve the desired purpose and to be suitable to be used in our framework, we wanted an evaluation measure which would satisfy the following characteristics: 1) The reward for including an $r$ rated snippet must be greater than the reward for including an $r-1$ rated snippet, 2) A negative rated snippet must be penalized, 3) An $r$ rated segment, no matter how big, should not displace an $r+1$ rated segment from a budget constrained gold summary, 4) No number of $r$ rated segments should displace an $r+1$ rated segment from a budget constrained gold summary, 5) In the gold summary, segments marked non-repetitive should not be broken unless it is absolutely necessary (possibly the last one to fit within the boundary) and 6) No reward should be given for picking more than $\beta$ seconds of a segment marked repetitive. After a careful design involving interactions of different multiplicative and additive terms and their effects on the above characteristics, we came up with the following formulation which satisfies the above characteristics.
The score function for video V, $S_V: y \rightarrow R$, is defined as:
\begin{align*}
S_V(y) &= \sum\limits_{x_i \in X_P} |y \cap x_i| * (1 + \frac{|y \cap x_i|}{|x_i|}) * e ^ {\alpha * rating(x_i)} \\
&+ \sum\limits_{x_i \in X_R} \min(|y \cap x_i|, \beta) * (1 + \frac{\min(|y \cap x_i|, \beta)}{\min(|x_i|, \beta)}) \\
&* e ^ {\alpha * rating(x_i)} \\
&- \sum\limits_{x_i \in X_N} |y \cap x_i| * k
\end{align*}
where,
$X_P$ is the set of segments in V marked non-repetitive and rated positive, $X_R$ is the set of segments in V marked repetitive and rated positive, $X_R$ is the set of segments in V rated negative, $\alpha > 0$ is the reward scaling hyper-parameter, $\beta > 0$ is the repetitiveness cut-off factor and $k > 0$ is the penalty factor.
This function is neither submodular nor supermodular. However, this can be written as a sum of a submodular and a supermodular function \narxiv{ (details and proof in supplementary material)} and hence the bounds discussed below hold true when this appears as the margin in the discrete optimization of the loss augmented objective (Equation ~\ref{loss-aug}).
\arxiv{
\begin{lemma}
The evaluation function $S_V(y)$ can be written as a sum of a submodular $f(y)$ and a supermodular function $g(y)$ where,
\begin{align*}
f(y) &= \sum\limits_{x_i \in X_R} \min(|y \cap x_i|, \beta) * e ^ {\alpha * rating(x_i)} \\ &- \sum\limits_{x_i \in X_R} \frac{\max(|y \cap x_i|^2, \beta)}{\min(|x_i|, \beta)} \\ &- \sum\limits_{x_i \in X_N} |y \cap x_i| * k
\end{align*}
and
\begin{align*}
g(y) &= \sum\limits_{x_i \in X_P} |y \cap x_i| * (1 + \frac{|y \cap x_i|}{|x_i|}) * e ^ {\alpha * rating(x_i)} \\ &+ \sum\limits_{x_i \in X_R} \frac{|y \cap x_i|^2}{\min(|x_i|, \beta)}
\end{align*}
\end{lemma}
It is easy to see that the function $f$ and $g$ above are submodular and supermodular respectively, and moreover, expanding the terms we can see that by adding these two functions, we get back $S_V(y)$.
}
\subsection{Discrete Optimization}
Our framework entails two different discrete optimization problems - maximization of the weighted mixture in the loss augmented inference, Equation ~\ref{loss-aug}, and during inference to obtain the summary once the mixture is learnt (Equation ~\ref{max}). For efficient optimization with guaranteed bounds, it is important to understand certain characteristics of these components. Note that our mixture of set functions can be written as follows:
\begin{align*}
f(X) &=
\alpha f^{\mbox{msub}}(X) + \beta f^{\mbox{nmsub}}(X) + \gamma f^{\mbox{sup}}(X) \\
&+ \delta f^{\mbox{d}}(X)
\end{align*}
where $f^{\mbox{msub}}(X)$ is a monotone submodular function, $f^{\mbox{nmsub}}(X)$ is a non-monotone submodular function, $f^{\mbox{sup}}(X)$ is a monotone supermodular function, and $f^{\mbox{d}}(X)$ is a dispersion function (also called disparity-min) ($f^{\mbox{d}}(X) = \min_{i, j \in X} d_{ij}$), and $\alpha, \beta, \gamma, \delta \geq 0$. Moreover, we assume that each of the functions above are non-negative (without loss of generality). Note that in the above, we have grouped all monotone submodular, non-monotone submodular, supermodular functions together. The only function which is neither submodular nor supermodular is Disparity Min.
We would like to understand the theoretical guarantees for the following optimization problem:
\begin{equation} \label{optprob}
\max \{f(X) \ | \ X \subseteq V, |X| \leq k\}
\end{equation}
for various values of $\alpha, \beta, \gamma, \delta$.
\begin{theorem}
The following theoretical results hold for solving the optimization problem of Equation~\ref{optprob}:
\begin{enumerate}
\item We obtain an approximation factor of $1 - 1/e$ if $\alpha \geq 0$ and $\beta = \gamma = \delta = 0$.
\item We obtain an approximation factor of $1/2$ if $\delta > 0$ and $\alpha = \beta = \gamma = 0$.
\item We obtain an approximation factor of $1/e$ if $\alpha \geq 0, \beta > 0$ and $\gamma = \delta = 0$.
\item We obtain an approximation factor of $1/4$ if $\alpha > 0, \delta > 0$ and $\beta = \gamma = 0$
\item We obtain an approximation factor of $1/2e$ if $\alpha, \beta, \delta \geq 0$ and $\gamma = 0$.
\item We obtain an approximation factor of $\frac{(1 - e^{(1 - \kappa^l)\kappa_k})}{\kappa_k}$ where $k(X) = f^{\mbox{msub}}(X)$ and $l(X) = f^{\mbox{sup}}(X)$, if $\alpha, \gamma \geq 0$ and $\beta = \delta = 0$.
\item We obtain an approximation factor of $\frac{(1 - e^{(1 - \kappa^l)\kappa_k})}{2\kappa_k}$ where $k(X) = f^{\mbox{msub}}(X)$ and $l(X) = f^{\mbox{sup}}(X)$, if $\alpha, \gamma, \delta \geq 0$ and $\beta = 0$.
\item The optimization problem of Equation~\ref{optprob} is inapproximable unless P = NP, if $\beta, \gamma > 0$ and $\alpha, \delta \geq 0$.
\end{enumerate}
\end{theorem}
\arxiv{
\begin{proof}
The following paragraphs provide the proofs for each of the cases enumerated above:
\begin{enumerate}
\item The first result follows directly from~\cite{nemhauser1978analysis} since when $\beta = \gamma = \delta = 0$, $f$ is a monotone submodular function. Hence the greedy algorithm admits an approximation guarantee of $1 - 1/e$.
\item For the second result, notice that when $\alpha = \beta = \gamma = 0$, we have $f(X) = \delta f^{\mbox{d}}(X)$, the dispersion function. Following Lemma 1 from~\cite{dasgupta2013summarization}, the Dispersion function admits a $1/2$ approximation factor for the cardinality constrained optimization problem from Equation~\ref{optprob}.
\item The third result is a consequence of \cite{buchbinder2014submodular}, since we have a non-monotone submodular maximization problem subject to a cardinality constraint. Note that the Randomized Greedy algorithm achieves an approximation guarantee of $1/e$.
\item The fourth result follows the proof technique from Theorem 1 in~\cite{dasgupta2013summarization}. since when $\alpha, \delta > 0$, $f(X)$ is a sum of a monotone submodular function and a dispersion function. Denote $m(X)$ as the monotone submodular function, and $d(X)$ as the dispersion function. Note that $f(X) = \alpha m(X) + \delta d(X)$. For the algorithm, we first optimize $m(X)$ alone, say using a greedy algorithm and then we optimize $d(X)$ alone, again using the greedy algorithm. We then take the better amongst the solutions achieved. Let $S_m$ denote the solution obtained by optimizing the monotone submodular function, and $S_d$ denote the solution obtained by optimizing the Dispersion function. Denote $O_m$ and $O_d$ as the optimal solutions respectively. Note that $m(S_m) \geq (1 - 1/e) m(O_m) \geq 1/2 m(O_m)$ and $d(S_d) \geq 1/2 d(O_d)$ from the respective approximation guarantees of the individual algorithms. Denote $X$ as the better amongst $S_m, S_d$ in terms of the final objective $f$. Therefore, $f(X) \geq 1/2 (f(S_d) + f(S_m))$. Note that $f(S_m) \geq \alpha m(S_m)$ and $f(S_d) \geq \delta d(S_d)$ and therefore $f(X) \geq 1/2(\alpha m(S_m) + \delta d(S_d)) \geq 1/4(\alpha m(O_m) + \delta d(O_d)) \geq 1/4 f(O)$ where $O$ is the optimal solution (note that $m(O) \leq m(O_m)$ and $d(O) \leq d(O_d)$ by definition.
\item For the fifth result, we use a similar proof technique as above. Note that we can achieve an approximation factor of $1/e$ by maximizing a non-monotone submodular function subject to a cardinality constraint, and a factor of $1/2$ for maximizing the Dispersion function. Let $n(X)$ as the non-monotone submodular function, and $d(X)$ as the dispersion function. We know that $f(X) = \beta n(X) + \delta d(X)$ (w.l.o.g, take the monotone function term into the non-monotone function since the sum of a monotone and non-monotone function is in general non-monotone). For the algorithm, we first optimize $n(X)$ alone, say using a randomized greedy algorithm and then we optimize $d(X)$ alone, using the greedy algorithm. We then take the better amongst the solutions achieved. Let $S_n$ denote the solution obtained by optimizing the non-monotone submodular function, and $S_d$ denote the solution obtained by optimizing the Dispersion function. Denote $O_n$ and $O_d$ as the optimal solutions respectively. Note that $n(S_n) \geq 1/e n(O_n)$ and $d(S_d) \geq 1/2 d(O_d) \geq 1/e d(O_d)$ from the respective approximation guarantees of the individual algorithms. Denote $X$ as the better amongst $S_n, S_d$ in terms of the final objective $f$. Therefore, $f(X) \geq 1/2 (f(S_d) + f(S_n))$. Note that $f(S_n) \geq \beta n(S_n)$ and $f(S_d) \geq \delta d(S_d)$ and therefore $f(X) \geq 1/2(\beta n(S_n) + \delta d(S_d)) \geq 1/2e(\alpha n(O_n) + \delta d(O_d)) \geq 1/2e f(O)$ where $O$ is the optimal solution (note that $n(O) \leq n(O_n)$ and $d(O) \leq d(O_d)$ by definition.
\item The sixth result is a direct consequence of~\cite{bai2018greed}, since in this case $f(X)$ is a sum of monotone submodular and a monotone supermodular function. Note that the approximation bounds here depend on the curvature of the two functions.
\item The seventh result follows by combining the proof technique from the fourth and fifth result, with the bound above. In the interest of space, we skip the details of this.
\item Finally, when $\alpha, \beta, \gamma > 0$, we have the most general case of the sum of non monotone submodular and supermodular functions. This is then equivalent to a difference of submodular functions, which is known to be inapproximable, following Theorems 5.1 and 5.2 from~\cite{iyer2012algorithms}.
\end{enumerate}
\end{proof}
}
\narxiv{The proofs for each of the cases enumerated above are included in the extended version of the paper in supplementary material.}
\subsection{Generating Ground Truth Summaries}
The use of ratings allow us to generate multiple ground truth summaries for a video. The total number of possible summaries could be exponential in video duration (a variant of knapsack on duration of segments), so for our experiments we randomly generate up-to 500 ground truth summaries for each video for each budget percentage (5, 15 and 30). Starting from the highest rating, if all segments of that rating can be fit in the budget, they are included in the summary. If all segments of a rating cannot be included in the remaining budget, using a flavor of standard coin exchange problem, maximal combinations of segments are enumerated such that possibly only the last segment gets broken.
\arxiv{
Algorithm~\ref{alg:gtGen} demonstrates how we generate the ground truth summaries given a set of ratings and a budget.
\begin{algorithm}
\caption{Generate ground truth summaries of a video for a given $budget$}
\label{alg:gtGen}
\begin{algorithmic}
\Procedure{generateGTSummaries}{$segments$, $budget$}
\State $gtSet = \{\}$
\State $normalizedSegments = \{\}$
\State $remBudget = budget$
\For {$s$ in $segments$}
\If {$s.repetitive == true$}
\State $s' = s$ trimmed by $\beta$
\State $normalizedSegments.add(s')$
\Else
\State $normalizedSegments.add(s)$
\EndIf
\EndFor
\State $tempGt = \{\}$
\For {$r$ in $3,2,1,0$}
\State $currSegs =$ segments rated $r$ in $normalizedSegments$
\State $cumBudget =$ Total duration of $currSegs$
\If {$cumBudget < remBudget$}
\State $tempGt.addSnippets(currSegs)$
\State $remBudget -= currBudget$
\Else\If {$cumBudget == remBudget$}
\State $tempGt.addSnippets(currSegs)$
\State $gtSet.add(tempGt)$
\State \textbf{break}
\Else
\While {$i <= 500$}
\State $rndSegSet$ = $X \subset currSegs$ and $\sum\limits_{x \in X} x.duration = remBudget$
\State $gtSet.add(tempGt \cup toSnippets(rndSegs))$
\State $i++$
\EndWhile
\EndIf
\EndIf
\EndFor
\Return $gtSet$
\EndProcedure
\end{algorithmic}
\end{algorithm}
}
\section{Experiments and Results}
What follows is a description of various experiments performed and results observed using the above framework for domain specific video summarization. For surveillance videos (entry exit and office), since night videos are black and white we do not use the color histogram features based on hue and saturation. For videos in Soccer, Cricket and Birthday domains we additionally perform shot detection to identify distinct shots in the video and create a feature which keeps track of the snippets present in each shot. This is consumed by the continuity component in the mixture. However, we do not use this continuity component in the mixture for the surveillance videos where the notion of a shot is not well defined in those videos. Also unless explicitly stated, for training, during each epoch, a random ground truth summary, out of the many possible summaries, was chosen for each video so that over a large number of epochs, all ground truths get covered. We do a train test split of 70-30 with respect to the number of videos in each domain in the dataset. The hyper-parameters used for the evaluation measure while training were $\alpha = 1$, $\beta = 6$ and $k=2$. We arrive at best values for $\lambda_1$ and $\lambda_2$ by testing the models on the held-out validation set.
\paragraph{Sanity of ground truths and behavior of evaluation measure}
We perform the following sanity checks on the ground truth summaries produced and the behavior of our evaluation measure:
\begin{itemize}
\item scores of all ground truth summaries of a particular budget for a video should be same, asserting that the synthesis of ground truths as above is consistent with $S_V(y)$
\item scores of ground truth summaries should always be greater than randomly produced summaries - for all lengths, for all videos for all categories
\end{itemize}
In this experiment we compare the normalized scores of the ground truth summaries against the normalized scores of 1000 random summaries picked exclusively from segments rated highly positive. We do the standard min-max normalization. We plot the minimum, maximum and average scores for both the random summaries and the ground truth summaries. To ensure that the random summaries do not get very low scores (and hence favoring the ground truth summaries during comparison), the random summaries used in this experiment are not truly random. They do not include the negatively rated segments.
\arxiv{\begin{figure}
\centering
\begin{tabular}{@{}cccc@{}}
\includegraphics[width=.25\textwidth]{images/birthday_gt.png} &
\includegraphics[width=.25\textwidth]{images/cricket_gt.png} \\
\includegraphics[width=.25\textwidth]{images/entryexit_gt.png} &
\includegraphics[width=.25\textwidth]{images/office_gt.png} \\
\multicolumn{2}{c}{\includegraphics[width=.25\textwidth]{images/soccer_gt.png}}
\end{tabular}
\caption{Scores of ground truth summaries compared with random summaries}
\label{pic:rndvsGt}
\end{figure}}
\arxiv{The results as per figure \ref{pic:rndvsGt} show that our evaluation measure and our ground truth generation algorithm behave as expected. All ground truth summaries have the same normalized scores and hence the minimum, maximum and the average scores for all ground truth summaries coincide as one line in the plots. Also all ground truth summaries score higher than any random summary across all domains and all videos. }\narxiv{The results (included in the extended version) show that all ground truth summaries score higher than any random summary across all domains and all videos.}
We also verify the ground truth summaries visually by representing them as videos and visually assessing their quality.
\paragraph{Learning experiments}
For learning weights in the above formulation, we constrain the weights of all submodular components to be always positive. There is no such constraint for modular components in the mixture. We compare AdaGrad and stochastic gradient descent and find AdaGrad to work better for all experiments and hence all results reported use AdaGrad. The reported numbers are losses (i.e. $normalizedGTScore$ - $normalizedSummaryScore$). We compare the following results: a) \emph{All Modular: } Train with only modular terms in the mixture, b) \emph{All Submodular: } Train with only submodular terms in the mixture, c) \emph{Full: } Complete Mixture (all modular terms and all submodular terms).
We compare all these to random summaries, uniform summaries and the best individual component baselines instantiated with different features are used individually to produce summaries. Results are reported in Table \ref{tab:learningResults}. We observe that combining both the submodular and modular terms in the mixture (full) provides the best results, as compared to just using submodular and modular terms alone or as compared to any of the baselines. Moreover, the learnt mixtures for only modular and either submodular also outperform Random, Uniform and the average individual submodular functions. We see that the best individual submodular functions also perform better than random or uniform baselines, but not as well as the learnt mixtures, thus proving the benefit of learning for this problem. We also verify the goodness of a summary both quantitatively (using the scores from our evaluation measure) and qualitatively (by visualizing the summary produced). This establishes our hypothesis that joint training can significantly help.
\begin{table}[ht]
\begin{center}
\begin{tabular}{|l|l|r||}
\hline
\textbf{Domain} & \textbf{Method} & \textbf{ScoreLoss} \\
\hline
\multirow{6}{*}{Birthday} & All Modular & 0.7234 \\ \cline{2-3}
& All Submodular & 0.7307 \\ \cline{2-3}
& Full & \textbf{0.6625} \\ \cline{2-3}
& Random & 0.7378 \\ \cline{2-3}
& Uniform & 0.7569 \\ \cline{2-3}
& Submodular & 0.7432 \\ \cline{2-3}
\hline
\multirow{6}{*}{EntryExit} & All Modular & 0.5967 \\ \cline{2-3}
& All Submodular & 0.6306 \\ \cline{2-3}
& Full & \textbf{0.5884} \\ \cline{2-3}
& Random & 0.7706 \\ \cline{2-3}
& Uniform & 0.7785 \\ \cline{2-3}
& Submodular & 0.6306 \\ \cline{2-3}
\hline
\multirow{6}{*}{Cricket} & All Modular & 0.8140 \\ \cline{2-3}
& All Submodular & 0.8275 \\ \cline{2-3}
& Full & \textbf{0.7733} \\ \cline{2-3}
& Random & 0.8911 \\ \cline{2-3}
& Uniform & 0.8979 \\ \cline{2-3}
& Submodular & 0.8275 \\ \cline{2-3}
\hline
\multirow{6}{*}{Office} & All Modular & 0.3871 \\ \cline{2-3}
& All Submodular & 0.4783 \\ \cline{2-3}
& Full & \textbf{0.3696} \\ \cline{2-3}
& Random & 0.5743 \\ \cline{2-3}
& Uniform & 0.5399 \\ \cline{2-3}
& Submodular & 0.5590 \\ \cline{2-3}
\hline
\multirow{6}{*}{Soccer} & All Modular & 0.8849 \\ \cline{2-3}
& All Submodular & 0.7645 \\ \cline{2-3}
& Full & \textbf{0.6533} \\ \cline{2-3}
& Random & 0.9217 \\ \cline{2-3}
& Uniform & 0.8747 \\ \cline{2-3}
& Submodular & 0.9152 \\ \cline{2-3}
\hline
\end{tabular}
\end{center}
\caption{Learning experiments comparing all submodular, all modular, full mixture with random, uniform and submodular baselines for all domains}
\label{tab:learningResults}
\end{table}
\paragraph{Verification of learning domain specific characteristics}
To demonstrate that the learnt summaries are domain specific, we test the model learnt on one domain in producing summaries of another domain.
The results in Table \ref{tab:domSpec} shows that models learnt on one domain perform poorly on other domains, establishing that the model has indeed learnt domain specific characteristics.
\begin{table}[ht]
\begin{center}
\begin{tabular}{|l|l|r|}
\hline
\textbf{Model Trained On} & \textbf{Model Tested On} & \textbf{ScoreLoss} \\
\hline
\multirow{3}{*}{Birthday} & Birthday & \textbf{0.6625} \\ \cline{2-3}
& Soccer & 0.9753 \\ \cline{2-3}
& Cricket & 0.9177 \\ \cline{2-3}
\hline
\multirow{4}{*}{EntryExit} & EntryExit & \textbf{0.5884} \\ \cline{2-3}
& Soccer & 0.9900 \\ \cline{2-3}
& Cricket & 0.9710 \\ \cline{2-3}
& Birthday & 0.8009 \\ \cline{2-3}
\hline
\multirow{3}{*}{Cricket} & Cricket & \textbf{0.7733} \\ \cline{2-3}
& Soccer & 0.8284 \\ \cline{2-3}
& Birthday & 0.8103 \\ \cline{2-3}
\hline
\end{tabular}
\end{center}
\caption{Sample results for cross domain experiments}
\label{tab:domSpec}
\end{table}
Next, we look at the weights learnt for the different submodular components. Figure \ref{pic:compWeights} (top) shows magnitude of the learnt weights the different domains. We see that different domains prefer different submodular components and features to produce good summaries. For example, scene features (googlenet\_features) are important for Cricket, while object detection features (yolo\_voc\_p\_concepts and yolo\_coco\_concepts) are more important for Surveillance Videos. For Cricket, the scene of ground or pitch or crowd has a lot of bearing on the importance and for Surveillance, detection of entities assume more importance, given the static scene. Next, we look at the correlation between the components which achieve the best weight in the learnt mixture and the components which achieve the best score when run in isolation. We see in the bottom table of Figure~\ref{pic:compWeights} that there is a strong correlation between the two. In particular about 6 to 7 out of the top ten components are the same in both buckets. It is also informative to look at the components themselves. For Birthday and Cricket, we see that Saturated Coverage and Facility Location (i.e. the representative models) are the winners, while in Office (which are surveillance videos), we see a lot of Disparity Min (diversity) functions as winners. \narxiv{A more detailed view of weights learnt for different domains as well as tests demonstrating significance of our results are provided in the extended version of this paper in the supplementary material.}
\begin{figure}[!htbp]
\begin{center}
\includegraphics[scale=0.25]{images/weights.png}
\includegraphics[scale=0.2]{images/dsvs-top-comp.png}
\caption{Sample results for a few domains: (top) Mixture component weights and (bottom) the top components based on mixture weights and top components when evaluated in isolation.}
\label{pic:compWeights}
\end{center}
\end{figure}
\arxiv{The Figures \ref{pic:birthdayWeights}, \ref{pic:cricketWeights} and \ref{pic:entryExitWeights} show the different weights learnt for different domains.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[scale=0.5]{images/birthday_weights.png}
\caption{Birthday Weights}
\label{pic:birthdayWeights}
\end{center}
\end{figure}
\begin{figure}[!htbp]
\begin{center}
\includegraphics[scale=0.5]{images/cricket_weights.png}
\caption{Cricket Weights}
\label{pic:cricketWeights}
\end{center}
\end{figure}
\begin{figure}[!htbp]
\begin{center}
\includegraphics[scale=0.5]{images/entryexit_weights.png}
\caption{EntryExit Weights}
\label{pic:entryExitWeights}
\end{center}
\end{figure}
}
\begin{figure}[!htbp]
\begin{center}
\includegraphics[scale=0.18]{images/dsvs-vis.png}
\caption{Top frames based on the domain specific learnt mixtures for Office, Birthday and Cricket}
\label{pic:topFrames}
\end{center}
\end{figure}
Finally, we look at the top ranked frames based on the learnt mixture for each domain. We see that the the frames intuitively capture the most important aspects for that domain. For example, in office footages, the top frames are frames where people are either entering/leaving or meeting, in Birthday videos, we see a shot where people are taking a selfie and a shot where the birthday girl is posing are ranked the best. Similarly in Cricket, shots where the player is about to hit, and where there is a four being hit are selected. Hence we see that the joint training has learnt specific domain specific importance of events along with the right weights for diversity and representation.
\paragraph{Importance of ratings in generating multiple ground truths}
To show that ability to generate multiple ground truth from our ratings based annotation is beneficial, we compare the scores obtained by training with a single ground truth summary in every epoch against the scores obtained by training with a random ground truth summary in every epoch.
\begin{table}[ht]
\begin{center}
\scalebox{0.7}{
\begin{tabular}{|l|l|r|}
\hline
\multirow{2}{*}{Birthday} & Random GTs & \textbf{0.6625} \\ \cline{2-3}
& Same GT & 0.6818 \\
\hline
\multirow{2}{*} {EntryExit} & Random GTs & \textbf{0.5883} \\ \cline{2-3}
& Same GT & 0.6188 \\
\hline
\end{tabular}
}
\end{center}
\caption{Sample results for using multiple ground truths}
\label{tab:rndGtVsfixedGt}
\end{table}
The results on Table \ref{tab:rndGtVsfixedGt} suggests that the model indeed learns better when multiple ground truth summaries are used hence establishing the importance of the system of ratings which allow us to generate many ground truths.
\section{Conclusion}
Motivated by the fact that what makes a good summary differs across domains, we set out to develop a framework which would automatically learn what is considered important for a domain, both in terms of the kind of snippets to be selected and also in terms of the desired characteristics of the summary produced in terms of representativeness, coverage, diversity, {\em etc.} We also establish that ratings provide a more efficient way of supervision to impart domain knowledge necessary to create such summaries. Further, we propose a novel evaluation measure well suited for this task. In the absence of any existing dataset which would lend itself well to this particular problem, we created a gold standard dataset and will be making it public as a part of this work. Through several experiments we demonstrated the effectiveness of our solution in producing domain specific summaries which can be seen as a first significant breakthrough in this direction.
|
\subsection{Abstract}\label{abstract}}
In this study we present evidence that endowment effect can be elicited
merely by assigned ownership. Using Google Customer Survey, we
administered a survey were participants (n=495) were randomly split into
4 groups. Each group was assigned ownership of either legroom or their
ability to recline on an airline. Using this experiment setup we were
able to generate endowment effect, a 15-20x (at p\textless{}0.05)
increase between participant's willingness to pay (WTP) and their
willingness to accept (WTA).
\hypertarget{i.-introduction}{%
\subsection{I. Introduction}\label{i.-introduction}}
Many economic theories are developed under a rational agent model, where
consumers are expected to treat all information unbiasedly in their
decision making process. In the 1960s economists started to realize that
consumers did not always act in this anticipated rational manner.
Research lead to the development of behavioral economics, the study of
mental processes such as ``attention, language use, memory, perception,
problem solving, creativity, and thinking''{[}8{]}. One of the
hypothesis developed was the endowment effect, the hypothesis that
consumer ascribe more value to things merely because they own them. The
endowment effect captures the observation of the valuation paradigm,
where people will tend to pay more to retain something they own than to
obtain something they do not own --even when there is no cause for
attachment, or even if the item was only obtained minutes ago. Studies
of the effect typically focus on the difference in Willingness to Pay
(WTP) in comparison to the Willingness to Accept (WTA). In one famous
example participants were given a mug and then offered the chance to
sell or trade for an equally valued alternative, pens. The researchers
found that once participants had the mug their WTA was twice as high as
their WTP. In another similar study, researchers found that participants
selling price of NCAA final four tickets (WTA) was 14 times higher than
their WTP. With magnitudes ranging from two to 14 times in research, it
is clear that the endowment effect can cause people to assign
drastically different values due to simple possession. Loss aversion
(the disutility of giving up an object is greater that the utility
associated with acquiring it) and status quo bias (tendency of
individuals to remain in status quo than to leave it) {[}2{]}, were
initially shown to correlate with the endowment effect. Recent studies
have suggested that evolutionary, strategic and cognitive factors play a
part in eliciting endowment effect{[}4{]}.
Our study attempts to understand if an individual would respond
differently between a question where ownership is assigned and where it
is not. Our hypothesis is that assigning ownership will cause the
participant to price the object more highly. In this paper, we attempt
to answer this question by means of a questionnaire that sets up this
hypothetical ownership. This paper is organized as follows. Section II
describes our experiment design while in Section III we share data
analysis. Results and findings are discussed in section IV. Section V
concludes this paper.
\hypertarget{ii.-experimental-design}{%
\subsection{II. Experimental Design}\label{ii.-experimental-design}}
To test our hypothesis we set up a post-treatment measurement 2x2 test
design where we asked subjects to price seating features on a five hour
flight. Specifically, we asked them to either price legroom or the
recline feature. Half were asked their willingness to accept a payment
(WTA) to give up the feature and the other was asked the willingness to
pay (WTP) for the feature.
\hypertarget{data-gathering-and-subject-interaction}{%
\subsubsection{Data Gathering and Subject
Interaction}\label{data-gathering-and-subject-interaction}}
Our team decided to continue to use Google Consumer Surveys to gather
our data responses. Google surveys was chosen because of its relative
cost compared to other survey options, quick response time, and
additional information provided about the respondents like age, gender,
and location. Our team made a conscious decision to potentially trade
off quality of responses so that we could have more responses.
Our participants were mainly visitors of news websites who wanted to
read articles behind the website's paywall. Instead of paying, Google
allows customers to complete a surveys such as ours to read the article.
Each subject was asked only one of our four questions, and 133 responses
were gathered for each question. Participants could be reaching using
their computer or mobile device. We note that because most of our
respondents were attempting to get through paywalls they in a sense had
``self selected'' into our study. We also note that our sample
population may be similar in other ways due to this type of experiment,
specifically that they all are the type of people who desire to read
news articles.
\hypertarget{pilot-study}{%
\subsubsection{Pilot Study}\label{pilot-study}}
Before embarking on a full fledged study of endowment effect, we wanted
to ensure that different aspects of the study are working as expected.
To this effect we initiated a pilot study. Our object in running a pilot
study were:
\begin{itemize}
\tightlist
\item
Ensuring Google Customer Survey (GCS) randomization works,
\item
Survey takers understand our survey questions,
\item
We are able to deliver treatment as we intended to,
\item
Verify that responses are valid and meets expectation,
\item
Distill presence of Endowment effect and
\item
Get a baseline to calculate statistical power required for full
fledged study.
\end{itemize}
The pilot study ran on Google Customer Survey from 25th June - 27 June,
2017. The survey consisted of two questions.
\begin{longtable}[]{@{}ll@{}}
\toprule
\begin{minipage}[b]{0.22\columnwidth}\raggedright
Question\strut
\end{minipage} & \begin{minipage}[b]{0.72\columnwidth}\raggedright
Description\strut
\end{minipage}\tabularnewline
\midrule
\endhead
\begin{minipage}[t]{0.22\columnwidth}\raggedright
1\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
You are traveling on a 5 hour flight. What is the maximum amount you
would be willing to pay to recline your seat?\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
2\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
You are traveling on a 5 hour flight. What is the minimum amount you
would accept to not recline your seat?\strut
\end{minipage}\tabularnewline
\bottomrule
\end{longtable}
Each survey taker is randomly assigned to one of the two questions and
they are provided with a textbox to post their response. The textbox
doesn't run any validation on the responses and accepts free form text.
This was done so as to gauge the kind of responses we get. We set the
survey to end when GCS receives at least 50 responses per question. Both
the questions assigns the respondent ownership of the chair. Question 1
is testing for their willingness to pay while question 2 tests their
willingness to accept. Once the survey ends, GCS provides functionality
to export survey responses with additional attributes pertaining to
respondents. The data consists of following columns:
\begin{longtable}[]{@{}ll@{}}
\toprule
\begin{minipage}[b]{0.22\columnwidth}\raggedright
Data Column\strut
\end{minipage} & \begin{minipage}[b]{0.72\columnwidth}\raggedright
Description\strut
\end{minipage}\tabularnewline
\midrule
\endhead
\begin{minipage}[t]{0.22\columnwidth}\raggedright
User ID\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
A unique user ID for each survey respondent\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Time (UTC)\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The time in which the survey was completed\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Publisher Category\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The type of website the user was trying to view when filling out the
survey\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Gender\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Gender of survey respondent or Unknown\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Age\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The age bracket of the survey respondent or Unknown\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Geography\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The Country, region and State of the respondent\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Urban Density\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Population density of user's location (Urban, suburban, rural)\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Income\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Income bracket user falls into\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Parental Status\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Parental status of the user\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Question \#\# Response\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Raw response of the respondent\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Response Time \#1 (ms)\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Time taken to respond\strut
\end{minipage}\tabularnewline
\bottomrule
\end{longtable}
Once we exported the data, we loaded it in R and checked the validity of
the responses. Of the 101 responses, following are the non-numeric ones.
All but three (``Ly'', ``Cff'', ``5 hours'') of them can be easily
converted to numeric format. We will mark these three as NA so as not to
affect our calculations.
Odd responses
\begin{longtable}[]{@{}ll@{}}
\toprule
Odd & Responses\tabularnewline
\midrule
\endhead
Ly & \$2\tabularnewline
\$50 & \$10\tabularnewline
\$0 & 0\ldots{}I think it's rude to recline on any
flight\ldots{}ever\tabularnewline
\$100 & 50 dollars\tabularnewline
Cff & \$100\tabularnewline
5 hours & \$100\tabularnewline
\$150 & 300.00 us\tabularnewline
Zero & 1,000000000.00\tabularnewline
\bottomrule
\end{longtable}
Following this we plot parsed responses on a box plot to see wether the
distribution is as expected. We see that the median response for
question 1 is around \$10 while that for question 2 is around \$100. The
difference between responses to these two questions is the endowment
effect which we were able to elicit in our pilot study which is
promising.
\includegraphics{unnamed-chunk-5-1.pdf}
Next we check wether the responses are distributed uniformly among
different demographics. Below we show distribution of responses for each
question by Gender/Age/Income.
\includegraphics{unnamed-chunk-6-1.pdf}
\includegraphics{unnamed-chunk-7-1.pdf}
\includegraphics{unnamed-chunk-8-1.pdf}
Next, we ran a test for covariate balance. We create a subset of
responses by omitting NAs for the covariates. We create two regression
models, one with all the covariates and another with no covariates.
Using ANOVA we test whether any of these
\begin{verbatim}
## Analysis of Deviance Table
##
## Model 1: treatment ~ gender + region + age
## Model 2: treatment ~ 1
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 75 18.435
## 2 84 21.247 -9 -2.8117 0.2468
\end{verbatim}
The result is a non statistically significant p-value of 2468 which
suggests the randomization design was not violated. The data from the
pilot study shows that we were able to get right responses for our
questions and that GCS did a good job of randomizing respondents.
\hypertarget{actual-study}{%
\subsubsection{Actual Study}\label{actual-study}}
In our actual study we expanded our questions to four, gathering 133
responses per question. These questions were:
\begin{longtable}[]{@{}ll@{}}
\toprule
\begin{minipage}[b]{0.09\columnwidth}\raggedright
Question Number\strut
\end{minipage} & \begin{minipage}[b]{0.85\columnwidth}\raggedright
Question Text\strut
\end{minipage}\tabularnewline
\midrule
\endhead
\begin{minipage}[t]{0.09\columnwidth}\raggedright
q1\strut
\end{minipage} & \begin{minipage}[t]{0.85\columnwidth}\raggedright
On a 5 hour flight what is the maximum amount you would be willing to
pay to recline the seat?\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.09\columnwidth}\raggedright
q2\strut
\end{minipage} & \begin{minipage}[t]{0.85\columnwidth}\raggedright
On a 5 hour flight what is the minimum amount you would accept to not
recline your seat?\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.09\columnwidth}\raggedright
q3\strut
\end{minipage} & \begin{minipage}[t]{0.85\columnwidth}\raggedright
On a 5 hour flight what is the maximum amount you would be willing to
pay to stop the passenger in front of you from reclining into the space
in front of you?\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.09\columnwidth}\raggedright
q4\strut
\end{minipage} & \begin{minipage}[t]{0.85\columnwidth}\raggedright
On a 5 hour flight what is the minimum amount you would accept to allow
the passenger in front of you to recline into your space?\strut
\end{minipage}\tabularnewline
\bottomrule
\end{longtable}
\newpage
\hypertarget{iii.-data-analysis}{%
\subsection{III. Data Analysis}\label{iii.-data-analysis}}
Our raw data is provided to us by google surveys in the form of a xls
file. White our main concern is the answers provided by the survey
respondents, we are also provided additional information about our
sample population. We can run some statistical tests to determine if
these are the same for each group.
\begin{longtable}[]{@{}ll@{}}
\toprule
\begin{minipage}[b]{0.22\columnwidth}\raggedright
Data Column\strut
\end{minipage} & \begin{minipage}[b]{0.72\columnwidth}\raggedright
Description\strut
\end{minipage}\tabularnewline
\midrule
\endhead
\begin{minipage}[t]{0.22\columnwidth}\raggedright
User ID\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
A unique user ID for each survey respondent\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Time (UTC)\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The time in which the survey was completed\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Publisher Category\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The type of website the user was trying to view when filling out the
survey\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Gender\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Gender of survey respondent or Unknown\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Age\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The age bracket of the survey respondent or Unknown\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Geography\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
The Country, region, State and City of the respondent\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Question Raw Response\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Raw response of the respondent\strut
\end{minipage}\tabularnewline
\begin{minipage}[t]{0.22\columnwidth}\raggedright
Response Time \#1 (ms)\strut
\end{minipage} & \begin{minipage}[t]{0.72\columnwidth}\raggedright
Time taken to respond\strut
\end{minipage}\tabularnewline
\bottomrule
\end{longtable}
\hypertarget{response-filtering-and-noncompliance}{%
\subsubsection{Response filtering and
Noncompliance}\label{response-filtering-and-noncompliance}}
Our survey allowed the participant to type any text in the response box.
The text box, before the respondents start entering text stated ``Enter
your answer in US \$''. Because of this the responses were not all
easily converted into numerical values. Some manual interpretation was
required. Many of the responses could be converted, for instance `2
dollars' could be interpreted as `2'. Other responses such as those in
units of measurement (3 inches), units of time (3 hours) or nonsense
responses (wa) were all converted to NA values and ignored for our
analysis. These participants are considered `noncompliers' in our study.
Below are the responses that were manually converted to numeric values.
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{# Cases which can be turned into numerical values}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "none it should be free"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "1 000 000 000.69"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\FloatTok{1000000000.69}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "two"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{2}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "zero"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "nothing"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "none"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "195 500 812.50"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\FloatTok{195500812.50}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "100 000 000.836215"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\FloatTok{100000000.836215}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "non"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "10 000"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{10000}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "0.00 as long as the seat reclines it is their right to recline it"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "25 dollars"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{25}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "zedo"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\DecValTok{0}
\NormalTok{d[d}\OperatorTok{$}\NormalTok{response }\OperatorTok{==}\StringTok{ "1 000 000.00"}\NormalTok{]}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\FloatTok{1000000.00}
\NormalTok{d}\OperatorTok{$}\NormalTok{response <-}\StringTok{ }\KeywordTok{as.numeric}\NormalTok{(d}\OperatorTok{$}\NormalTok{response)}
\end{Highlighting}
\end{Shaded}
The following were responses that were converted to \texttt{NA}.
\begin{verbatim}
## [1] "i don't know" "i don't fly commercial"
## [3] "2 hours" "reject offer"
## [5] "250 free flight" "2 hours"
## [7] "2 hours" "2 min"
## [9] "do not follow question" "na"
## [11] "yes" "wa"
## [13] "2.5 hour" "not at all"
## [15] "yes" "3 hours"
## [17] "5 hours" "15 minutes"
## [19] "5 hours" "1 hour"
## [21] "not sure" "3 inches"
\end{verbatim}
We can plot a count by question to inspect noncompliance by question. We
note that visually it appears that the distribution of NA responses is
not even between questions. The `Willingness to Pay' questions (Question
\#1 and Question \#3) have fewer NA responses compared to our
`Willingness to Accept' questions (Question \#2 and Question \#4). This
may be attributed to the fact that the willingness to accept questions
were fundamentally harder to understand, which may have impacted our
randomization and must be accounted for in our final analysis.
\includegraphics{unnamed-chunk-13-1.pdf}
Additionally we conduct a \texttt{t.test} and \texttt{cohen.d} test to
compare the compliance rate of our treatment and control groups. The
results of the t-test show a statistically significant difference
between the compliance rate while the Cohen's D shows a small effect
size between the two groups. In this study we are mainly concerned with
the complier average causal effect (CACE). Simply looking at the
intention to treat (ITT) would give these nonsense responses some
weight, while the CACE give us the effect for those who actually
understood the question and answered the question in compliance with our
survey request of ``Enter your answer in US \$''.
\begin{verbatim}
##
## Welch Two Sample t-test
##
## data: d$compliance[d$treatment == 1] and d$compliance[d$treatment == 0]
## t = -3.0802, df = 382.93, p-value = 0.002218
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## -0.08608726 -0.01900442
## sample estimates:
## mean of x mean of y
## 0.9325843 0.9851301
\end{verbatim}
\begin{verbatim}
##
## Cohen's d
##
## d estimate: -0.2667107 (small)
## 95 percent confidence interval:
## inf sup
## -0.43716436 -0.09625695
\end{verbatim}
\hypertarget{check-for-randomization}{%
\subsubsection{Check for Randomization}\label{check-for-randomization}}
To check for randomization we will compare the responses between
questions. If randomization was done correctly there should be no
statistically significant difference between the response rates for each
question. First we will look at the results visually.
The distribution of gender per question shows a slightly less number of
known female respondents for all questions except question \#1.
\includegraphics{unnamed-chunk-17-1.pdf}
We can compare how long it took respondents to complete the question by
using a boxplot. Note the y-axis is on a logarithmic scale. Question 2
appears to be the only question with a slightly longer average response
time.
\includegraphics{unnamed-chunk-18-1.pdf}
Exploring the age distribution of respondents we see no obvious bias.
The number of respondents in the 18-24 age bracket appears to have the
largest difference between questions. With the small number of
respondents we cannot say this is statistically significant.
\includegraphics{unnamed-chunk-19-1.pdf}
\hypertarget{anova-test}{%
\subsubsection{ANOVA test}\label{anova-test}}
Next, we create a subset from our dataset and remove any observations
that do not include the categorical data we are leveraging to check
covariate balance.
We isolate the first question set's observations and check if any
categorical factors had an effect on whether treatment was assigned.
\begin{verbatim}
## Analysis of Deviance Table
##
## Model 1: treatment ~ gender + region + age
## Model 2: treatment ~ 1
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 199 278.43
## 2 209 290.44 -10 -12.006 0.2847
\end{verbatim}
The result is a non statistically significant p-value of 0.3341 which
suggests the randomization design was not violated.
We repeat this on the second test:
\begin{verbatim}
## Analysis of Deviance Table
##
## Model 1: treatment ~ gender + region + age
## Model 2: treatment ~ 1
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 190 268.68
## 2 200 278.04 -10 -9.3636 0.498
\end{verbatim}
Which also results in a non statistically significant p-value of 0.4674.
\hypertarget{transformations}{%
\subsubsection{Transformations}\label{transformations}}
Due to the continuous user input there are a handful of extreme
outliers, particularly in the WTA pool (e.g., willing to accept
\$10,000+).
\includegraphics{unnamed-chunk-23-1.pdf}
The responses of the raw data resulted in an average user WTP as
\$9126012 (\$5 median) and WTA \$794316 (\$75 median) for the first
question set (recline). WTP as \$15.27 (\$1 median) and WTA \$8259.81
(\$20 median) for the second question set (legroom).
At first, we considered removing these observations altogether since
they are disruptive to the response mean and likely an unconsidered
response. And perhaps they are. However, after discussing amongst
ourselves we decided we ultimately will treat these responses as valid
data despite our reservations. For the purposes of our analysis, we are
concluding these users are unwilling to negotiate an accept value which
is a legitimate response. Although this decision compromises the
usefulness of our data as is. In light of these considerations, our data
required transformation.
To get a better sense of the response data relative to itself we
explored the data using nonparametric statistics.
Resultantly, a bimodal distribution was unveiled. Collectively, the
largest density of responses were subjects who answered \$0 which
resulted in one of the two humps. The second is due to the subjects in
the treatment (WTA) consistently responding in a higher max value than
their control (WTP) counterparts. The first question set identified the
average WTP subject in the 0.381 percentile vs the 0.632 percentile for
WTA subjects. 0.427 percentile for WTP and 0.595 percentile for WTA in
the second question set
In addition to better understanding the distribution, both question sets
were highly statistically significant using the nonparametric Wilcoxon
Rank-sum test model. Despite the promising results, our analysis still
left us without an answer for the average a subject is willing to pay in
comparison to what they're willing to accept.
To better answer that question, we used a log transformation on the
data.
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{# Stacked plot}
\KeywordTok{grid.arrange}\NormalTok{(log_recline, log_legroom, }\DataTypeTok{ncol=}\DecValTok{1}\NormalTok{, }\DataTypeTok{nrow=}\DecValTok{2}\NormalTok{)}
\end{Highlighting}
\end{Shaded}
\includegraphics{unnamed-chunk-26-1.pdf}
\includegraphics{unnamed-chunk-27-1.pdf}
\hypertarget{iv.-results-discussion}{%
\subsection{IV. Results \& Discussion}\label{iv.-results-discussion}}
Given the non-parametric distribution of our data we first analyzed our
data using a Wilcoxon rank sum test.
\begin{verbatim}
##
## Wilcoxon rank sum test with continuity correction
##
## data: d$response[(d$recline == 1) & (d$treatment == 1)] and d$response[(d$recline == 1) & (d$treatment == 0)]
## W = 12232, p-value = 1.247e-11
## alternative hypothesis: true location shift is not equal to 0
\end{verbatim}
\begin{verbatim}
##
## Wilcoxon rank sum test with continuity correction
##
## data: d$response[(d$recline == 0) & (d$treatment == 1)] and d$response[(d$recline == 0) & (d$treatment == 0)]
## W = 10858, p-value = 5.68e-06
## alternative hypothesis: true location shift is not equal to 0
\end{verbatim}
Our results showed that for both our recline and legroom question sets
we reject our null hypothesis that there is no shift in rank. These
results gave us evidence to reject the null hypothesis that assigned
ownership does not have an impact on the value of an object, but failed
to show the magnitude of the effect.
We wanted to quantify the amount of difference observed between subjects
beyond simply stating their rank was different. Given the skewed
distribution of our observations and the large difference between the
mean and median in each respondent group we decided to analyze our
results using randomization inference on the median response.
\includegraphics{unnamed-chunk-30-1.pdf}
Subjects who answered our WTPxRecline question (q1) had a median
response of (\$5) and our WTAxRecline (q2) subjects had a median
response of (\$75). The difference in median of \$70 (15:1 WTA:WTP) is
statistically significant with 0 of 10000 draws showing a larger
difference (p value of 0).
Similar to our recline question set, subjects who answered our
WTPxLegroom question (q3) had a median response of \$1 and our
WTAxLegroom (q4) subjects had a median response of (\$20). The
difference in median of \$19 (20:1 WTA:WTP) is statistically significant
with 40 of 10000 draws showing a larger difference (p value of 0.004).
Given the statistically significant results of both the recline and
legroom question sets we can reject our null hypothesis that assigned
ownership does not have an impact on the value of an object. We observed
that the ratio of WTA:WTP ranged from 15:1 for reclining and 20:1 for
legroom.
Our 2x2 design allows us to set up hypothetical pairs of real world
`negotiations' between a recliner and the person behind them. We used
the same Randomization Inference methodology for the medians of q1 vs q4
(WTPxRecline vs WTAxLegroom) and q2 vs q3 (WTAxRecline vs WTPxLegroom).
As the endowment effect would predict, we observe that the WTA
individuals have a statistically significant higher median minimum
amount required than the median WTP. WTAxlegroom individuals wanted a
median minimum payment of 20 while the WTPxrecline individuals wanted to
spend a median maximum payment of 5. This difference is statistically
significant with a p-value of 0.0234 at an observed 4:1 WTA:WTP.
WTAxrecline individuals wanted a median minimum payment of 75 while the
WTPxlegroom individuals wanted to spend a median maximum payment of 1.
This difference is statistically significant with a p-value of 0 at an
observed 75:1 WTA:WTP. This has interesting real-world implications in
that the object, legroom/reclining in this study, has different
negotiated values depending on who is given the ownership - even though
rational-agent economic theory suggests individuals should value the
object identically!
\hypertarget{v.-conclusion}{%
\subsection{V. Conclusion}\label{v.-conclusion}}
Our results show that we can reject our null hypothesis that assigned
ownership does not have an impact on the value of an object. While this
is exciting, we need to be cautious about how far to extend the
implications of these results. There does not exist a standing framework
with airlines to allow passengers to negotiate seat privileges. We would
anticipate real world tests to have slightly different outcomes than our
hypothetical negotiation tests due to additional behavioral economic
interactions, such as price setting (from overheard negotiations). That
said, the intent for this study was to evaluate how assigned ownership
in a question causes an individual to price an object differently, not
how to help airline best set seat pricing.
The statistically significant results we observed have implications
outside of the airline industry. For example, voters are often asked
through ballot measures to vote on approval of public work projects or
government services. For many voters this decision includes determining
in their mind what the price of the good or service is and if the
proposal is an acceptable deal. Our research would indicate that the way
the question on the ballot is posed could have a material impact on the
outcome. Consistent with our results, we would anticipate voters to have
a higher assumed value to objects or services they perceive to ``own'',
such as a park or river. Just like the reclining seat in our study,
these objects or services on ballots rarely have a known market price.
Moreover just like our subjects, voters are simply given a question and
rarely involved in the negotiation to get to the ballot-stated value.
While our results lack a real world mechanism for seat negotiations they
have strong implications for similar real-world situations where
individuals are asked to determine the value of a good or service.
Our work gives strong evidence that assigned ownership changes how an
individual perceives the worth of an object. While we can't extend these
results too far, future research into how an individual comes to believe
ownership would be interesting. Additionally, given our positive result
it would be interesting to see how assigned ownership interacts with
political beliefs and if this effect extends into how individuals vote.
\hypertarget{references}{%
\subsection{References}\label{references}}
\begin{enumerate}
\def\arabic{enumi}.{\arabic{enumi}.}
\tightlist
\item
Kahneman, Daniel, Jack L. Knetsch, and Richard H. Thaler.
``Experimental tests of the endowment effect and the Coase theorem.''
Journal of political Economy 98.6 (1990): 1325-1348.
\item
Kahneman, Daniel, Jack L. Knetsch, and Richard H. Thaler. ``Anomalies:
The endowment effect, loss aversion, and status quo bias.'' The
journal of economic perspectives 5.1 (1991): 193-206.
\item
Ericson, Keith M. Marzilli, and Andreas Fuster. ``The endowment
effect.'' (2014).
\item
Morewedge, Carey K., and Colleen E. Giblin. ``Explanations of the
endowment effect: an integrative review.'' Trends in cognitive
sciences 19.6 (2015): 339-348.
\item
Horowitz, John K., and Kenneth E. McConnell. ``A review of WTA/WTP
studies.'' Journal of environmental economics and Management 44.3
(2002): 426-447.
\item
Bordalo, Pedro, Nicola Gennaioli, and Andrei Shleifer. ``Salience in
experimental tests of the endowment effect.'' The American Economic
Review 102.3 (2012): 47-52.
\item
Price, Curtis R., and Roman M. Sheremeta. ``Endowment origin,
demographic effects, and individual preferences in contests.'' Journal
of Economics \& Management Strategy 24.3 (2015): 597-619.
\item
``American Psychological Association (2013). Glossary of psychological
terms''. Apa.org.
\end{enumerate}
\end{document}
|
\section{INTRODUCTION}
\label{sec:intro}
The Square Kilometer Array (SKA) is a worldwide collaboration with radiotelecopes hosted in South Africa and Australia at a first stage, and in other eight African countries later on, namely, Botswana, Ghana, Kenya, Madagascar, Mauritius, Mozambique, Namibia and Zambia. This will be largest radiotelescope array in the world and it will end up being 50 times more sensitive and 100 times faster than current best radiotelescopes. Its life span is expected to be of at least 50 years, and although it uses the most recent antenna technology and signal processing and computing, it can be continuously upgraded as computing power increases. SKA will cover frequencies in the range from 70 MHz to 10 GHz.
The knowledge of the Universe will be dramatically increased with SKA, since it will provide insight on the evolution of galaxies and cosmology, on strong gravity through pulsars and black holes, on the origin and evolution of cosmic magnetism, on cosmic history of the Universe at dark ages and reionisation epochs, and on putative cradle of life\cite{ska}.
In what concerns Cosmology, radio observations from SKA are expected to constrain the equation of state parameter for dark energy, hence discriminating between several models of cosmic acceleration, to observe a dark matter power spectrum, to test modified theories of gravity using weak lensing observations, and to get and insight on the cosmic magnetism\cite{ska,bull,raccanelli}.
\section{COSMOLOGY}
\subsection{Dark Universe}
One of the science key goals of SKA concerns the understanding of the dark components of the Universe. Several missions have been designed to study the energy content of the Universe. The most recent, Planck mission\cite{planck}, showed that the Universe is filled with $4.9\%$ of ordinary matter, $26.8 \%$ of dark matter and $68.3\%$ of dark energy. Dark matter is responsible for the flattening of the galactic rotation curves, and gravitational lensing which cannot be explained only by regular matter. On its turn, dark energy is the smooth energy substratum behind the current acceleration of the Universe. These two dark components dominate the Universe. Nevertheless, aside from some observational signatures and dynamical effects, their inner nature still remains unknown.
SKA will be able to provide some answers about the components of the Universe through the 21cm radiation of the neutral hydrogen at high redshifts. The cosmic evolution of such well defined signal encodes information about the medium where it propagates, most particularly about the dark matter and dark energy effects. For instance, by measuring the BAO signature in the radio of the galaxy power spectrum, SKA will allow for precise measurements of the equation of state of dark energy, $w=p/\rho$, of the form:
\begin{equation}
w=w_0+w_a(1-a)~,
\label{eq:1}
\end{equation}
where $p$ and $\rho$ are the pressure and energy density of dark energy, $w_0$ and $w_a$ are some real parameters and $a$ is the scale factor. In Fig. \ref{fig:de} one can see that assuming a Gaussian distribution around the fiducial model $w=-1$, which corresponds to the cosmological constant as dark energy, the predicted SKA results combined with Planck \cite{planck} and Baryon Oscillation Spectroscopic Survey (BOSS) \cite{BOSS} priors constrain considerably the allowed region of parameters. The expected combined precision is compared with the one from Euclid mission \cite{euclid}.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{de.pdf}
\caption{Predicted SKA constraints at various stages on the equation of state parameter of dark energy (with BOSS and Planck priors) \cite{raccanelli}.}
\label{fig:de}
\end{figure}
Furthermore, by studying weak gravitational lensing the measurement of the dark matter power spectrum will be possible, and bounds on the mass and the number of neutrinos families can be obtained.
\subsection{Testing gravity to the limit}
As stated, testing the nature of dark matter or dark energy is a key issue in Cosmology. However, the effect of these putative entities can be manifestations of a yet unknown gravity theory beyond General Relativity \cite{whatif}. Thus, testing gravity at the strong limit provides a deeper understanding on the nature of this fundamental interaction. SKA, in particular, will be able to detect a vast number of pulsars, and binary systems of a pulsar orbiting a black hole, hence providing rich information about strong gravity. Thousands of millisecond pulsars might be detected, forming a “pulsar timing array” which can be a prime arena for detection of gravitational waves \cite{strong}. On the other hand, General Relativity and alternative theories of gravity will be tested through gravitational lensing observations. In general, there are three model dependent parameters which characterise the growth of structures in any modified gravity theory \cite{euclid}. The first is the modified gravitational constant, $G_N \tilde{\mu}(a,k)$, where $\tilde{\mu}(a,k)$ is a model dependent correction in the Poisson equation, which can be expressed in the Fourier space as:
\begin{equation}
-2k^2\Phi = 8\pi G_N a^2\rho D \tilde{\mu} (a,k) ~,
\end{equation}
where $k$ is the wavenumber and $D$ is the gauge-invariant density contrast. The second is the anisotropic stress or slip relation, $\tilde{\gamma}$, which relates the two gauge invariant Bardeen potentials, $\Phi$ and $\Psi$, which come from time and spacial scalar perturbations of the metric, respectively:
\begin{equation}
\Psi = \tilde{\mu}(a,k)\Phi~,
\end{equation}
and the third one which is the growth rate, $f(a,k)$, (or its index $\gamma$):
\begin{equation}
f(a,k)=\left(\frac{a^2 8\pi G_N \rho}{3H^2}\right)^{\gamma}~,
\end{equation}
where $H=\dot{a}/a$ is the Hubble expansion rate. Therefore, this may discriminate between models of gravity. In Fig. \ref{fig:gamma}, we can see the improvement of the measurement of the equation of state parameter for dark energy over the index of structure growth. In General Relativity, $\tilde{\mu}=\tilde{\gamma}=1$ and $\gamma\approx 0.545$. Thus, observational deviations from these values are evidences of modified gravity models, which can allow for predicting dependencies on the scale factor and on the wavenumber on the previous parameters. A particular example of such modified gravity theories is the one which admits an extension of the well known $f(R)$ theories with a non-minimal coupling between matter and curvature \cite{nmc}, whose cosmological perturbations give specific relationships for those parameters \cite{frazao}.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{gamma.pdf}
\caption{SKA predictions for the constraints on the growth of structures index (including BOSS and Planck priors) \cite{raccanelli}.}
\label{fig:gamma}
\end{figure}
Furthermore, it would also be interesting to investigate clusters rich in radiogalaxies in order to assess whether there is evidence for interacting dark matter-dark energy models \cite{ide1,ide2,ide3} or modified gravity theories \cite{lieq} through deviations from the virial theorem, which, at cosmic scales, is given by the so-called Layzer-Irvine equation.
\subsection{Cosmic magnetism}
Another relevant science objective of SKA is related to the cosmic magnetism, which is ubiquitous in the Universe, since interstellar gas, planets, stars and galaxies all exhibit magnetic fields. However, the shape and strength of such magnetic field in galaxies or even its origin is not known. The Universe itself can be magnetic. Therefore, radio observations can measure the Faraday rotation, the polarised synchrotron emission and the Zeeman effect, which gives a detailed information on such fields, in particular whether they are primordial\cite{turner,magnetic} or are generated later on via a dynamo mechanism. Another relevant issue is the impact of magnetic fields on the evolution of the Universe.
\section{CONCLUSIONS}
In this brief contribution we have presented the main contributions that SKA is expected to provide for Cosmology. As the world largest radiotelescope array, SKA will be able to access the parameters that characterise the dark components of the Universe with unprecedented detail: one expects to get a power spectrum for dark matter from weak lensing observations and to considerably constrain the equation of state of dark energy (cf. Eq. \ref{eq:1}). The acquired knowledge will allow for ruling out many models of dark matter and dark energy, including those where these two components interact with each other\cite{ide1,ide2,ide3} or come from a unique field or a fluid model\cite{kamenshchik,chaplygin, bilic}. Moreover, SKA data will provide essential information on putative alternative models of gravity beyond General Relativity.
As mentioned, SKA will allow to look at the strong regime of Einstein’s gravity and of any other gravitational models. On the other hand, by studying weak lensing distribution of matter and gravitational distortions, SKA will provide means to test the three parameters that characterise alternative theories of gravity: the growth of structure factor (or index), the modified gravitational constant and the anisotropic stress.
It is also expected that SKA will be an excellent tool to study cosmic magnetism and its characterisation over several redshifts, as well as the role of magnetic fields on the evolution of the Universe.
To conclude we can say that SKA will, most likely, revolutionise the way we perceive the Universe, as it will provide a huge amount of data on a wide range of astrophysical and cosmological issues and will allow for the understanding of many key features of the Cosmos.
\acknowledgments
C. G. acknowledges the support from Fundação para a Ciência e a Tecnologia (FCT) under the grant SFRH/BD/102820/2014.
|
\section{Introduction}
\label{sec:intro}
Buoyancy is the root cause of thermal convection in the atmosphere and fundamentally governs the vertical acceleration of clouds. A density anomaly's buoyancy is most often characterized by the Archimedean formula $B = - \rho'/\rho_og$. However, only an \emph{infinitesimal} buoyant parcel, whose motion has a negligible impact on the environment, actually accelerates at its familiar Archimedean value. A density anomaly of finite size disrupts its environment and sets the surrounding fluid in motion. This loss of momentum to the environment reduces the finite parcel's buoyant acceleration from its Archimedean buoyancy to its "effective buoyancy" $\beta$, as introduced by \cite{davies2003} and discussed by \cite{doswell_markowski_2004,xu_randall_2001,krueger_mclean_fu_1995}. The back-reaction exerted by the environment on the buoyant fluid is realized by a buoyancy perturbation pressure (to be fully defined in the next section).
The reduction from $B$ to $\beta$ is determined by the geometry of the density anomaly. As discussed in \cite{houze2014cloud}, for broad pancake-shaped density anomalies, buoyant motion is severely inhibited by the surrounding fluid, and $\beta$ can be several orders of magnitude less than $B$. On the other hand, narrow, pencil-shaped density anomalies impart negligible momentum to their environment, and $B$ serves as a reliable proxy for $\beta$.
Moving beyond this qualitative understanding, approximate formulae and scalings exist for the effective buoyancy in updrafts \citep{morrison_2016_1,morrison_2016_2,peters_2016}, of an isolated bubble \citep{pauluis_garner_2006}, and in hydrostatic versus non-hydrostatic models \citep{weisman_skamarock_klemp_1997,jeevanjee2017b}. Furthermore, exact analytical results for the effective buoyancy along the central axis of a buoyant cylinder were obtained by \cite{Jeevanjee2016}. While this progress is significant, we still lack a full 3D analytical solution for the effective buoyancy of a density anomaly even for idealized geometries, and thus have no quantitative picture for how the \emph{environment} responds to buoyant accelerations.
Another issue is that convection schemes often account for the buoyancy perturbation pressure gradient by simply multiplying $B$ by a fixed virtual (a.k.a. induced) mass coefficient. This coefficient is sometimes set equal to 2/3, which is the value for a solid sphere in potential flow (e.g., see \cite{romps2015sticky, romps2010undiluted,de2012parameterization,batchelor2000introduction}). In a similar fashion, \cite{turner_1964} approximated the perturbation pressure acting on an ellipsoidal thermal by appealing to the virtual mass coefficient of a solid ellipsoid. Not only does fixing a coefficient neglect the geometry-dependent nature of the buoyancy perturbation pressure, but there also does not seem to be any rigorous proof that the solid-body value is appropriate for a \emph{fluid} density anomaly, which may experience an internal circulation and thus deform as it accelerates.
In this paper, we aim to make progress on these questions by deriving the exact effective buoyancy for the idealized case of a uniformly buoyant ellipsoid. Our analytical results spring in part from the recognition of a novel mathematical equivalence between the equations of magnetostatics and effective buoyancy. We show that the magnetic field of a uniformly magnetized body of arbitrary shape is mathematically identical to the effective buoyancy of an identically-shaped fluid density anomaly.
This equivalence allows us to leverage the long-standing classical literature on magnetization to find the effective buoyancy of an ellipsoidal thermal. In addition, the correspondence provides theoretical justification for appealing to the solid-body virtual mass coefficients in the case of ellipsoids, yet also proves that the connection between solid-body motion and fluid anomaly acceleration breaks down for all other geometries.
The relevance of the ellipsoidal geometry examined in this work is supported by the observed shapes of convecting thermals. The actively convecting cores of Cumulus in Large Eddy Simulations \citep{sherwood_2013, romps2015sticky} and atmospheric observations \citep{damiani_2006, damiani_2007} are often ellipsoidal in shape. Laboratory experiments \citep[e.g.][]{scorer_1957,woodward_1959} of rising thermals also demonstrate this same ellipsoidal form.
Realistic turbulent thermals, of course, have non-uniform $B$ fields. How relevant are theoretical formulae for $\beta$ that assume uniform $B$? To answer this we present simulations of a turbulent thermal, and show that its \emph{average} $\beta$ can indeed be explained by our analytical formulae, when applied to the turbulent thermal's average $B$ field.
\section{Effective Buoyancy Preliminaries}
\label{sec:beta_prelim}
We begin with the Boussinesq momentum equation,
\begin{equation}
\frac{D \vec{v}}{D t}= - \frac{ \nabla p'}{\rho_o } + B \, \vec{e}_z + \nu \nabla^2 \vec{v} \, ,
\label{eqn:boussinesq}
\end{equation}
where $B = - \rho'/\rho_og$ is the Archimedean buoyancy and $\nu$ is the molecular viscosity \citep{markowski2011mesoscale}. We decompose the pressure according to $p'(x,y,z,t) = p(x,y,z,t) - \bar{p(z)}$, where $\bar{p(z)}$ is in hydrostatic balance with the constant reference density, $\rho_o$. Consider the divergence of the momentum equation,
\begin{equation}
\rho_o \frac{\partial }{\partial t} \left[ \nabla \cdot \vec{v} \right] = - \rho_o \nabla \cdot \left( \vec{v} \cdot \nabla \vec{v} \right) - \nabla^2 p' + \rho_o \pd{B}{z} + \rho_o \nu \nabla^2 \left[ \nabla \cdot \vec{v} \right] \,.
\end{equation}
Boussinesq mass continuity ($\nabla \cdot \vec{v} = 0$) dictates that the total convergence tendency is zero, yielding a Poisson's equation for $p'$,
\begin{equation}
\nabla^2 p' = - \rho_o \nabla \cdot \left( \vec{v} \cdot \nabla \vec{v} \right) + \rho_o\pd{B}{z} \, . \label{p'eqn}
\end{equation}
From this perspective, the perturbation pressure arises to enforce mass continuity by opposing any convergence tendencies generated by advective motion or buoyancy.
The distinct physical mechanisms generating the perturbation pressure motivates further decomposing $p'$ according to $ p' = p_b + p_d $, where $p_b$ is the buoyancy perturbation pressure driven statically by vertical gradients in the buoyancy field, and $p_d$ is the dynamic pressure generated by advective motion. Linearity of Poisson's equation then allows us to separate Eqn. \ref{p'eqn} into independent parts,
\begin{subequations}
\begin{gather}
\nabla^2 p_b =
\rho_o\pd{B}{z} \label{eqn:poisson_buoyancy} \\
\nabla^2 p_d =
- \rho_o \nabla \cdot \left( \vec{v} \cdot \nabla \vec{v} \right) \label{eqn:poisson} \, .
\end{gather}
\end{subequations}
Likewise, the Lagrangian acceleration admits a complimentary decomposition. As in \cite{Jeevanjee2015}, we define the effective buoyancy $\beta$ as the Lagrangian vertical acceleration that would result from instantaneously zeroing out the velocity field,
\begin{equation}
\beta = \frac{D w}{D t} \bigg|_{ \vec{v} \rightarrow 0} \,
= - \frac{1}{ \rho_o } \pd{p_b}{z} + B \, .
\label{eqn:beta_momentum}
\end{equation}
By accounting for $p_b$, the effective buoyancy is thus the true buoyant acceleration experienced by a finite-sized density anomaly.
We also define the dynamic acceleration as the Lagrangian acceleration that would result from instantaneously zeroing out the density anomaly field,
\begin{equation}
a_d = \frac{D w}{D t} \bigg|_{\rho' \rightarrow 0} = - \frac{1}{ \rho_o } \pd{ p_d}{z} + \nu \nabla^2 w \, .
\label{eqn:inert_accel}
\end{equation}
Together, the effective buoyancy and the inertial acceleration combine to give the Lagrangian vertical acceleration,
\begin{equation}
\frac{D w}{D t} = \beta + a_d.
\end{equation}
Note that these definitions fully specify the boundary conditions necessary to determine $p_b$ and $p_d$ as detailed in \cite{Jeevanjee2015}.
Our focus in this article is on the effective buoyancy, but this is not to suggest that the dynamic acceleration is of little importance in atmospheric convection. In a mature thermal, the dynamic acceleration is comparable in magnitude to the effective buoyancy and critical for establishing and maintaining the thermal's internal circulation. For the initial motion of a nascent thermal, however, the effective buoyancy dominates over the dynamic acceleration. In Section \ref{sec:motion_from_rest}, we further detail the transition between these two regimes.
\begin{figure*}
\centerline{\includegraphics[width=7in]{./test_arch_vs_effective.png}}
\caption{A uniform spherical thermal: a) vertical cross-sections of the Archimedean buoyancy profile, b) analytical solution for the effective buoyancy, and c) analytical solution for the Lagrangian acceleration with vector acceleration in arrows and acceleration magnitude in color.}\label{fig:sphere_plots}
\end{figure*}
\section{Effective Buoyancy of a Spherical Anomaly}
\label{sec:sphere_analytical_results}
We begin by seeking a solution for the effective buoyancy field of a spherical buoyancy anomaly of uniform Archimedean buoyancy $B_o$ and radius $R$. In spherical coordinates centered on the buoyant sphere with $\vec{r} = (r, \varphi, \theta)$, polar angle $\theta$, and azimuthal angle $\varphi$, the buoyancy field is described by
\begin{equation}
B(\mathbf{r} ) = B_o \mbox{H} (R -r ) \, ,
\end{equation}
where $\mbox{H}(R-r)$ is the Heaviside step function. Relying on Eqn. \ref{eqn:poisson_buoyancy}, we compute the buoyancy perturbation pressure as
\begin{equation}
- \nabla^2 p_b = \rho_o B_o \cos{\theta} \, \delta (R -r ) \, \label{eqn:pb_sphere_poisson}
\end{equation}
where $\delta (R -r ) $ is the Dirac delta function. Away from $r=R$, $p_b$ solves Laplace's equation, and thus admits a series solution in the spherical harmonics. The conditions that $p_b$ be finite at $r=0$ and vanish at infinity provide the piece-wise expansion,
\begin{equation}
p_b(r, \theta) = \left\{
\begin{array}{ll}
\displaystyle \sum_{\ell} a_{\ell} r^l P_{\ell}(\cos{\theta}) \, & r<R \\[5pt]
\displaystyle \sum_{\ell} \frac{ b_{\ell} } {r^{\ell+1} } P_{\ell}(\cos{\theta}) & r>R,
\end{array}
\right.
\end{equation}
where $P_{\ell}$ is the $\ell$-th member of the Legendre polynomials, which form a normalized basis for azimuthally-symmetric functions on the sphere. For future reference, we note that $P_1(\cos \theta ) = \cos \theta$ and $P_2(\cos \theta) = 1/2 \left(3\cos^2 \theta - 1\right)$. Dependence on $\varphi$ is precluded by azimuthal symmetry (whereas $\theta$ symmetry is broken by gravity).
To generate the boundary conditions, we multiply Eqn. (\ref{eqn:pb_sphere_poisson}) by $r^2$ and separate $r^2\nabla^2$ into a radial and angular part,
\begin{equation} \pd{}{r}\left(r^2 \pd{p_b}{r} \right)
+ \nabla^2_{\text{ang}} p_b = - \rho_o B_o \cos{\theta} \,r^2 \delta (R -r ) \,,
\label{eqn:boundary_cond_eqn}
\end{equation}
where we have defined
\begin{equation}
\nabla^2_{\text{ang}} \equiv
\frac{1}{\sin\theta} \frac{\partial}{\partial \theta}\left(\sin\theta \frac{\partial }{\partial \theta}\right)
+ \frac{1}{\sin^2\theta} \frac{\partial^2 }{\partial \varphi^2} \, .
\end{equation}
Suppose $p_b$ were discontinuous about $r=R$, then $\partial^2 p_b / \partial^2 r$ would generate a $\partial_r \left[ \delta (R - r) \right]$ on the L.H.S of Eqn. \ref{eqn:boundary_cond_eqn}. Since the singularities on the L.H.S and R.H.S. are now of different order (i.e., $\partial_r \left[ \delta (R - r) \right]$ is linearly independent from $\delta(R-r)$), we have a contradiction. Therefore, $p_b$ is continuous.
Integrating Eqn. \ref{eqn:boundary_cond_eqn} with respect to $r$ over $[R-\epsilon,R+\epsilon]$ then yields,
\begin{equation}
\lim_{\epsilon \rightarrow 0 } \, \left[ \,\pd{p_b}{ r } \bigg|_{R + \epsilon } - \, \pd{p_b}{ r }\bigg|_{R - \epsilon} \, \right] = - \rho_o B_o \cos{\theta} \,,
\label{eqn:pb_bc}
\end{equation}
where the angular term dropped out as $\epsilon \rightarrow 0$ since $p_b$ is continuous in $r$ and smooth in $\theta$.
Continuity of $p_b$ provides that $b_\ell = a_\ell R^{2\ell + 1}$, and we may write the expansion outside the spherical shell as
\begin{equation}
\sum_{\ell} \frac{ a_\ell R^{2 \ell + 1} } {r^{\ell+1} } P_{\ell}(\cos{\theta}) \,.
\end{equation}
Enforcing the boundary condition (\ref{eqn:pb_bc}) on the radial derivative gives
\begin{equation}
\lim_{\epsilon \rightarrow 0} \sum_{\ell=1}^{\infty} \left[ \frac{ - (\ell + 1) R^{2\ell + 1} }{ (R + \epsilon)^{\ell+2 } } - \ell (R-\epsilon)^{\ell-1} \right] a_\ell P_{\ell}(\cos{\theta}) = - \rho_o B_o \cos{\theta} .
\end{equation}
Recognizing $P_1(\cos{\theta})$ on the R.H.S and invoking the linear independence of the Legendre polynomials demands that any factors in front of the $l \neq 1$ Legendre polynomials are individually zero,
\begin{subequations}
\begin{gather}
\lim_{\epsilon \rightarrow 0} \left[ \frac{ - (\ell + 1) R^{2\ell + 1} }{ (R + \epsilon)^{\ell+2 } } - \ell (R-\epsilon)^{\ell-1} \right] a_\ell = 0 \\
\implies \quad (2\ell + 1) R^{\ell - 1}\ a_\ell = 0 \, .
\end{gather}
\end{subequations}
We find that $a_\ell = 0 $ for $\ell \neq 1$. For the $\ell = 1$ case, we have
\begin{subequations}
\begin{gather}
\lim_{\epsilon \rightarrow 0} \left[ \frac{ - 2 R^{3} }{ (R + \epsilon)^{3} } - 1 \right] a_1 = -\rho_o B_o \\
\implies \quad \ a_1 = \frac{\rho_o B_o}{3},
\end{gather}
\end{subequations}
yielding the solution for the buoyancy perturbation pressure:
\begin{equation}
p_b(r, \theta) = \left\{
\begin{array}{ll}
\displaystyle \frac{ \rho_o B_o }{3} r \cos{\theta} \, & r \leq R \\[9pt]
\displaystyle \frac{ \rho_o B_o R^3 }{3} \frac{\cos{\theta} }{r^2} & r \geq R \,. \label{eqn:pb_solution}
\end{array}
\right.
\end{equation}
From Eqn. \ref{eqn:beta_momentum}, the effective buoyancy field readily follows:
\begin{equation}
\beta (\mathbf{r}) = \left\{
\begin{array}{ll}
\displaystyle \frac{2 B_o}{3} \, & r < R \\[9pt]
\displaystyle \frac{B_oR^3}{3} \frac{\left( 3 \cos^2 \theta - 1 \right)}{r^3} & r > R \,.
\end{array}
\right. \label{eqn:effective_buoyancy_sphere}
\end{equation}
This analytical expression for $\beta(\mathbf{r})$ is a main result of this paper, and is plotted in the center panel of Figure \ref{fig:sphere_plots}. Note the presence of $P_2(\cos{\theta})$ in Eqn. (\ref{eqn:effective_buoyancy_sphere}), with characteristic lobes evident in Fig. \ref{fig:sphere_plots}. Furthermore, setting $\theta = \pi/2$ in Eqn. (\ref{eqn:effective_buoyancy_sphere}) yields a horizontal profile of $\beta$: the sphere itself accelerates vertically at $2/3B_o$, while the immediately adjacent environment subsides at $-\frac{1}{3}B$. This subsidence acceleration then decays fairly rapidly as $r^{-3}$.
One implication of these results is that the negative acceleration associated with near-field compensating subsidence is comparable in magnitude to the acceleration of the anomaly itself. This environmental acceleration declines rapidly with distance, yielding after finite time a descent of environmental air that is strongly enhanced near the anomaly. In a stably stratified environment (i.e., $d\theta/dz >0$), such a distribution of descent yields a corresponding distribution of positive buoyancy anomalies, which then generate their own convection and cause this chain to repeat itself. This is, of course, simply a description of how gravity waves propagate away from rising fluid, and we see that the horizontal inhomogeneity of the environmental acceleration evident in Fig. \ref{fig:sphere_plots} is critical to this propagation.
For completeness, we also note that the environmental fluid's horizontal acceleration can be computed as
\begin{equation}
\frac{D \vec{u}}{D t} = - \frac{1}{ \rho_o } \nabla_h p_b = \frac{ B_o R^3 z }{(x^2 + y^2 + z^2)^{5/2} } \big(x,y\big) \, .
\end{equation}
The environmental fluid above $z=0$ accelerates outward and is displaced away from the sphere while the fluid beneath $z=0$ accelerates inwards to fill the vacated space as depicted in Figure \ref{fig:sphere_plots}c.
\begin{figure*}
\centering \centerline{\includegraphics[width=7in]{./initial_regime_panel_updated3.png}}
\caption{a) log-log plot of the height of the anomaly versus time b) contour plot of the initial Lagrangian acceleration.}\label{fig:initial_regime_panel}
\end{figure*}
\section{Buoyant Motion From Rest}
\label{sec:motion_from_rest}
What is the relevance of the foregoing results? For a buoyancy anomaly accelerating from rest in a quiescent environment, the dynamic pressure is zero and the effective buoyancy gives the full Lagrangian acceleration. The motion will stay buoyancy-dominated until enough momentum has developed that the dynamic pressure force can no longer be neglected.
Since $L_{\text{th}}$ (the length of thermal) is the only length in the problem, it sets the length-scale of the buoyancy-dominated motion. The end of this regime occurs when $\nabla p_b \sim \nabla p_d$. Noting that $\nabla p_{\text{d}} \sim \rho W^2/L_{\text{th}} $ and $\nabla p_\text{b} \sim \rho B_o$ due to the Poisson's equations, we have the following scalings for the buoyancy-dominated regime's duration and final velocity:
\begin{align}
\tau & \sim \sqrt{\frac{L_{\text{th}}}{B_o}} \, , \\
W & \sim \sqrt{B_o L_{\text{th}}} \, . \label{eqn:L_scaling}
\end{align}
During this buoyancy-dominated regime, a spherical thermal has near uniform effective buoyancy and thus achieves solid body motion with $\Delta z \approx \frac{1}{2} \beta t^2$.
For a representative thermal with $L_{\text{th}} \sim .5$ km (\cite{romps2015sticky, hernandez2016numerical}), this parabolic motion would describe a thermal's rise through a typical boundary layer. With buoyancy in the characteristic range of $B=.01 - .1\ \mathrm{m/s^2}$, the scaling yields $W_{\text{b}} \sim 2-7\ \mathrm{m/s}$ in close agreement with velocities observed for thermals in numerical studies \citep{romps2015sticky, hernandez2016numerical}. The above scalings are also in consonance with the idealized model for a spherical thermal put forward by \cite{escudier_maxworthy_1973}, who under several assumptions (e.g., that the thermal's motion is detrainement-free, drag from dynamic pressure is negligible, and the entrainment of ambient fluid is proportional to surface area and velocity) obtained asymptotic solutions for the thermal's trajectory.
To validate this account of the buoyancy-dominated motion, we analyze the direct numerical simulations of a spherical thermal conducted in \cite{lecoanet_2018} using Dedalus,\footnote{For more information and links to the source code, see \url{dedalus-project.org}.} an open-source pseudo-spectral framework \citep{dedalus}. As detailed in Appendix A, we solve the non-dimensional Boussinesq equations for the turbulent motion of a heterogenous density anomaly rising through a neutrally-stratified environment.
As shown in Figure \ref{fig:initial_regime_panel}b, the spatial structure of the instantaneous Lagrangian acceleration from rest matches the analytical $\beta$ result of Eqn. \ref{eqn:effective_buoyancy_sphere} (displayed in Figure \ref{fig:sphere_plots}). Let $\left< \right>$ denote an average over the thermal's initial spherical volume. We find that $\left< \beta \right>/\left< B \right> = .65$, in close agreement with the theoretical prediction of 2/3.
To verify the length-scale of the buoyancy-dominated regime, we track the center of the density anomaly through time by computing
\begin{equation}
\vec{x}_{\text{AC}}(t) = \frac{ \displaystyle\int_{\Omega} \rho'(t) \, \vec{x} \, \, dV }{ \displaystyle\int_{\Omega} \rho'(t) \, \, dV } \, ,
\end{equation}
where $\Omega$ the full domain. As shown in Figure \ref{fig:initial_regime_panel}a, the buoyancy-dominated regime, characterized by $\Delta z \sim \beta t^2$ ascent, persists until the thermal rises a distance comparable to the thermal's own length, validating the analytical scaling of Eqn. \ref{eqn:L_scaling}.
\section{Effective Buoyancy of a Mature Thermal}
\label{sec:mature_thermal}
After the initial motion chronicled in the prior section, the thermal transitions to a turbulent regime distinguished by a \say{vortex ring} circulation. This structure, which loosely resembles the laminar Hill's vortex, consists of an ellipsoidal thermal that encases a rotating ring of more anomalously density; fluid ascends in the ring's interior and descends on the periphery \citep{scorer_1957, woodward_1959}.
Figure \ref{fig:ellipsoid_panel} demonstrates azimuthally-averaged $w$, $\rho'$, and $\beta$ fields for the ellipsoidal mature thermal. The $\beta$ distribution is found using the Poisson solver detailed in \cite{Jeevanjee2015}. We observe that $\beta$ also is non-uniform in the thermal's interior: the concentration of buoyant fluid in the ring imposes a secondary structure with a pattern akin to the superposition of two buoyant spheres. Accordingly, $\beta$ is enhanced over the ring and opposite-signed in the center. We find that the presence of this ring of relatively undiluted fluid depends on the level of turbulence in our simulations. More laminar simulations in \citet{lecoanet_2018} find a much stronger ring. Considering that thermals in the atmosphere are far more turbulent than our simulations, this suggest that the $\beta$ field is closer to uniform in real thermals.
Motivated by the findings of the prior section, we examine to what extent the mean $\beta$ of a mature thermal may be explained by considering a uniformly buoyant ellipsoid. With that goal in mind, we first derive the effective buoyancy of an ellipsoid via exploiting a mathematical equivalence between effective buoyancy and the equations of magnetostatics.
\subsection{Magnetostatics Correspondence}
\label{subsec:magnetostatics}
Understanding the connection between magnetostatics and effective buoyancy requires familiarity with the concept of magnetization, which we outline below. Readers are referred to \cite{griffiths} or \cite{jackson} for a thorough treatment of the topic.
Consider a small circular loop of wire through which a steady current flows. If the loop has vector area $\vec{a}$ and carries current $I$, taking $\vec{a} \rightarrow 0$ while fixing the \emph{dipole moment} $\vec{m} = \vec{a}I$ yields the \emph{dipole magnetic field}. This miniature loop circuit is a conceptual model for the magnetic dipole moments generated by atomic electrons. A physical magnet is composed of many such dipole moments aligned in a common direction and described by the magnetic dipole moment per unit volume, known as the \textit{magnetization}, $\vec{M} = \frac{d \vec{m}}{dV}$. Inside a magnet that is uniformly magnetized in the vertical, we would have $\vec{M} = M_o e_z$.
We will consider the case of a permanently magnetized object, where the interior dipoles are aligned even in the absence of external magnetic fields. The total magnetic field $\vec{B}$ due to the magnet's permanent dipoles is found by first accomplishing the mathematically simpler task of determining the auxiliary $\vec{H}$-field, defined as
\begin{equation}
\vec{H} = \frac{\vec{B}}{\mu_0} - \vec{M} \, .
\end{equation}
In the absence of any free currents (i.e., currents following through wires), the curl and divergence of the above definition give
\begin{align}
\nabla \times \vec{H} &= 0 \\
\nabla \cdot \vec{H} &= - \nabla \cdot \vec{M} \, .
\end{align}
Since $\vec{H}$ is curl-free, Helmholtz's theorem tells us that $\vec{H} = - \nabla \Phi_M$, where $\Phi_M$ is a scalar function known as the magnetic potential. We then have the Poisson's equation,
\begin{equation}
\nabla^2 \Phi_M = \nabla \cdot \vec{M} \, .\label{magnetic_poisson}
\end{equation}
We regard $\nabla \cdot \vec{M} $ as the \say{ bound magnetic-charge}, which is non-zero only on the surface of the magnetized object. If the magnet is magnetized only in the vertical, then
\begin{equation}
\nabla^2 \Phi_M = \pd{M}{z} \,.
\end{equation}
This equation is mathematically equivalent to Eqn. \ref{eqn:poisson_buoyancy}, the Poisson's equation for the buoyancy perturbation pressure. In addition, $\Phi_M$ and $p_b$ share identical boundary conditions: both must be finite at the origin, vanish at $\infty$, and continuous over the boundary of a magnetic/fluid body.
This mathematical equivalence spells out the following correspondence,
\begin{equation}
\begin{split}
\text{Fluid}\quad \quad & \text{Magnetostatics} \\
B\ \longleftrightarrow & \ M \\
\frac{ p_b }{\bar{\rho}} \ \longleftrightarrow &\ \Phi_M \\
- \frac{ \nabla p_b }{\bar{\rho}}\ \longleftrightarrow & \ \vec{H} \\
\vec{\beta} \ \longleftrightarrow & \ \frac{ \vec{B}}{\mu_0}
\end{split}
\label{magneto_corresp}
\end{equation}
The effective buoyancy of an arbitrarily-shaped fluid anomaly is mathematically equivalent to the magnetic field generated by a vertically magnetized body of the same shape.
The classical theory of magnetized bodies is replete with analytical results. Poisson is attributed with finding the magnetic field inside a uniformly magnetized sphere and Maxwell gives the corresponding result for an ellipsoid in Article 437 of \cite{maxwell}. A modern treatment of the magnetic field inside a magnetized ellipsoid can be found in \cite{stoner1945xcvii}, and \cite{tejedor1995external} derives formulas for the external magnetic field presented in the following section.
The magnetostatics correspondence also clarifies the relationship between effective buoyancy and induced mass.
We have shown that a spherical buoyancy anomaly \emph{accelerates uniformly}, moving as if it were a solid body. This explains why its induced mass is equal to that of a solid sphere, and why the spatial structure of $\beta$ in Eqn. (\ref{eqn:effective_buoyancy_sphere}) is identical to that of the potential flow of a solid sphere translating at a steady velocity $U\vec{e}_z$ in an infinite fluid \citep[simply let the buoyant sphere evolve for time $\Delta t$ and set $U= (2 B_o/3) \Delta t$]{batchelor2000introduction}. For an ellipsoid, the effective buoyancy result presented in the next section also agrees with the induced mass formulas computed in Articles 114 and 373 of \cite{lamb}. Does an arbitrarily-shaped density anomaly also have a uniform effective buoyancy,and thus exactly mimic its solid body counterparts?
Again, we turn to the magnetostatics literature. The effective buoyancy is uniform only if $\nabla p_b$ is uniform in the fluid anomaly's interior. In Article 437 of \cite{maxwell}. Maxwell proved that $\vec{H}$ -- the magnetostatics quantity analogous to $\nabla p_b$ -- is uniform in a magnetized body's interior only if the body's surface is quadratic. The only closed finite quadratic surface being the ellipsoid. Therefore, we conclude that only \emph{ellipsoidal} bodies have uniform $\beta$, and have induced mass coefficients equal to their solid body counterparts.
\begin{figure*}
\centering
\centerline{\includegraphics[width=6in]{./spheroid_panel.png}}
\caption{Evolution of the azimuthally-averaged $w$, $\rho'$, and $\beta$ fields. Black contours show the identified boundaries of the actively convecting ellipsoidal thermals.}\label{fig:ellipsoid_panel}
\end{figure*}
\subsection{Ellipsoidal Thermal}
\label{subsec:ellipsoidal}
\label{subsubsec:spheriod_analytical_results}
Consider an azimuthally-symmetric ellipsoidal density anomaly of uniform Archimedean buoyancy $B_o$ and specified by major axes $(W,W,H)$. Let $\alpha = W/H$ be the aspect ratio of the ellipsoid with $\alpha < 1$ indicating a prolate ellipsoid and $\alpha > 1$ an oblate ellipsoid. Adapting the corresponding result for the magnetized ellipsoid presented in Article 437 of \cite{maxwell} and fully derived in \cite{stoner1945xcvii}, the buoyancy perturbation pressure in the interior of the ellipsoid takes the form,
\begin{equation}
p_b = f(\alpha) \rho_o B_o z \, ,
\end{equation}
where we have defined
\begin{equation}
f(\alpha) =
\begin{cases}
\frac{\alpha^2}{1-\alpha^2} \left[ \frac{ \log \left( \frac{1+\sqrt{1-\alpha^2} }{\alpha} \right) }{\sqrt{1-\alpha^2}} - 1\right] & \alpha < 1 \\
\frac{1}{3}, & \alpha = 1 \\
\frac{\alpha^2}{1-\alpha^2} \left[1 - \frac{1}{\sqrt{\alpha^2 -1}} \sin^{-1} \left( \frac{\sqrt{\alpha^2 -1}}{\alpha} \right) \right] & \alpha > 1 \\
\end{cases}
\label{eqn:f_alpha}
\end{equation}
Note that the oblate formula ($\alpha >1$) can be obtained by taking the real part of the prolate formula ($\alpha < 1$) and applying the identity $\sin^{-1}{x} = -i \ln(\sqrt{1-x^2} + ix)$. For reference, $f(\alpha)$ corresponds to the demagnetization factor $D(\mu)$ of Eqn. 4.2 in \cite{stoner1945xcvii} with $\alpha = \mu$.
As in the sub-case of the sphere, $p_b$ is a linear function of $z$ inside the anomaly, and thus the whole ellipsoid uniformly accelerates in the vertical with an effective buoyancy of
\begin{equation}
\beta_{\text{ellipsoid}}(\alpha) = (1 - f(\alpha)) B_o \, , \label{eqn:beta_spheroid}
\end{equation}
as shown in Figure \ref{fig:beta_vs_B_comparison}. This formula is the quantitative foundation for the observation of \cite{houze2014cloud} that wide pancake-shaped thermals are severely inhibited by their environment with $\beta \approx 0$, while narrow pencil-shaped anomalies impart negligible momentum to their environment yielding $\beta \approx B$.
Importantly, it also enables us to move beyond the limiting cases and understand the regime relevant for atmospheric thermals. \cite{hernandez2016numerical} found evidence for slightly prolate thermal, while \cite{scorer_1957, woodward_1959, damiani_2006} observe moderately oblate thermals consistent with our simulations, which have $\alpha \approx 1.4$.
\begin{figure}
\centering \centerline{\includegraphics[width=3in]{./beta_vs_alpha_comparison2.pdf}}
\caption{Comparison between $\beta$ averaged over the ellipsoidal thermal as a function of the diagnosed aspect ratio and the analytical formula (Eqn. \ref{eqn:beta_spheroid}). Thermal identification was conducted every .1 $T$, yielding 5 snapshots of the initial regime and 70 snapshots of the mature thermal. Bars indicate the standard deviation of the data for each regime (error bars for the initial regime are smaller than marker). } \label{fig:beta_vs_B_comparison}
\end{figure}
\subsection{Comparison to Simulation}
\label{subsec:comparison}
Equipped with the idealized solution (Eqn. \ref{eqn:beta_spheroid}), we investigate its potential relevance to the simulated ellipsoidal thermals. Rigorous comparison to the simulation results requires introducing an objective method to identify the thermal's boundary. To achieve this, we implement a simple $\rho'$-based approach as described in Appendix B. Example identification results are shown in Figure \ref{fig:ellipsoid_panel}.
Thermal identification enables us to compute the volume average of $\beta$ over the thermal, again denoted as $\left< \beta \right> $. Figure \ref{fig:beta_vs_B_comparison} compares $\left< \beta \right>/\left< B \right>$ to the analytical result of Eqn. \ref{eqn:beta_spheroid}, evaluated with the thermal's aspect ratio. The analytical curve is found to accurately explain the effective buoyancy over the mature thermal's ascent. In particular, the time-mean $\left<\beta \right>$ of the turbulent thermal is within $.01 B_o$ of the analytical result --- a remarkable agreement. This finding allows us to go beyond scaling arguments and make a first analytical step towards rigorously understanding the forces acting on a turbulent thermal.
A lingering question, however, is: why do we find such close agreement despite the heterogeneous ring structure present in the $B$ field? Consider decomposing the thermal's Archimedean buoyancy field according to $B = \left<B\right> + B'$. By definition, we have that $\left<B'\right>=0$, and thus $B'$ contains opposite-signed regions that cancel when averaged over the thermal. Taking advantage of the linearity of Poisson's equation, $\beta$ admits the decomposition $\beta = \beta\left[\left<B\right>\right] + \beta\left[B'\right]$, where $\beta\left[\left<B\right>\right]$ is the effective buoyancy generated by a uniform ellipsoid of constant Archimedean buoyancy $\left<B\right>$. Likewise, $\beta[B']$ is the effective buoyancy resulting from the heterogeneous $B'$ field. See Figure \ref{fig:mean_vs_prime} for a plot of these quantities. As expected, $\beta\left[\left<B\right>\right]$ is constant in the ellipsoid and described by the analytical result of the prior section.
We observe that the $\beta\left[B'\right]$ field inherits the heterogeneous features of the $B'$ field and preserves the distribution of opposite-signed regions. As a consequence, if we average $\beta\left[B'\right]$ over the thermal, cancellation between the opposite-signed regions yields a value three orders of magnitude smaller than $\beta\left[\left<B\right>\right]$. Therefore, we find that the departures from $\left<B\right>$ make a negligible contribution to $\left< \beta \right>$, supplying the agreement observed in Figure \ref{fig:beta_vs_B_comparison}.
\begin{figure}
\centering \centerline{\includegraphics[width=3in]{./mean_vs_prime_new}}
\caption{a) vertical cross section at $t = 28.7$ of $\left< B \right>$, the volume average of the thermal's Archimedean buoyancy, b) $B' = B - \left< B \right>$, the departures from the mean ellipsoidal buoyancy, c) effective buoyancy generated by $\left< B \right>$ distribution, d) effective buoyancy generated by $B'$ distribution. See text for discussion. }\label{fig:mean_vs_prime}
\end{figure}
\section{Discussion}
\label{sec:discussion}
Our main findings are as follows:
\begin{itemize}
\item There exists an exact correspondence between magnetostatics and the fluid dynamics of effective buoyancy (Eqn. \ref{magneto_corresp})
\item This correspondence supplies analytical formula for the effective buoyancy of spherical and ellipsoidal anomalies---the only fluid geometries that experience a uniform buoyant acceleration, and thus have virtual mass coefficients identical to their solid-body counterparts.
\item The effective buoyancy of a heterogenous, turbulent thermal is captured by an analytical function of the thermal's aspect ratio (Fig. \ref{fig:beta_vs_B_comparison}).
\end{itemize}
The above result enables the computation of a turbulent thermal's buoyant acceleration from the coarse geometry of the density field, and thus can be applied to lower resolution observational or model data. Another application of these results is to cumulus parameterizations, and specifically those that parameterize the vertical velocity equation. As summarized by \cite{de2012parameterization}, a wide range of virtual mass coefficients are employed in the literature. However, the results presented here, and in particular Fig. \ref{fig:beta_vs_B_comparison}, suggest that the virtual mass coefficients much different from 2/3 are not justified by thermal-based reasoning.
A limitation of this study is that our simulations do not account for water vapor or a stratified environment. This poses a problem, however, only if a \say{wet} thermal in a stratified fluid deviates from the coarse ellipsoidal geometry necessary to apply our analysis. Fortunately, atmospheric observations (\cite{damiani_2006}) and \say{wet} stratified simulations (\cite{sherwood_2013, romps2015sticky, hernandez2016numerical}) find thermals that possess the necessary ellipsoidal shape.
Note that our analysis emphasized the buoyancy perturbation pressure, but the dynamic pressure force is equally important to the circulation of an ellipsoidal vortex ring. While there has been recent progress in visualizing and parameterizing the dynamic pressure force \citep{peters_2016}, a deeper quantitative understanding is still lacking. In particular, we would benefit from analytical solutions to \eqref{eqn:poisson} that reveal the dynamic pressure's spatial pattern and mean effect on a mature thermal. Future research should aim to address these questions.
\acknowledgments
NT thanks Kirk McDonald for discussion of the magnetostatics content, and for bringing Maxwell's proof to his attention. The authors thank three anonymous reviewers in addition to Howard Stone, Aaron Match, and David Romps for helpful discussions and feedback. NJ is supported by a Harry Hess fellowship from the Princeton Geosciences Department. DL is supported by a PCTS fellowship and a Lyman Spitzer Jr.~fellowship. Computations were conducted with support by the NASA High End Computing (HEC) Program through the NASA Advanced Supercomputing (NAS) Division at Ames Research Center on Pleiades with allocations GID s1647 and s1439.
|
\section{INTRODUCTION}
\label{sec:intro}
Distributed control has undergone an unprecedented growth during the past few years mainly due to the complexity in modeling and analysis of large scale systems.
In such scenarios, high-dimensional collective tasks conducted by members of a team are formed from local decisions of each member leading towards the global system-level final decision.
Accordingly, the main focus in distributed control design is finding the closest-to-optimal control mechanism for a large-scale system, making use of the structure in information exchange and decision-making.
Indeed, such an approach has found broad applications in such areas as robotic swarms \cite{bullo2009distributed}, structured robust learning \cite{sedghi2019robust}, and social networks \cite{alemzadeh2017controllability}.
Decentralized control of large-scale systems is not a new research area.
The roots of the field traces back to the socioeconomics literature of 1970's \cite{mcfadden1969controllability}; an early work in the control literature is~\cite{wang1973stabilization}.
The inspiration of these types of works is that the presupposition of centrality fails to hold due to the lack of either central intelligence or computational capability~\cite{sandell1978survey}.
This line of work was followed by the pioneering work \cite{corfmat1976decentralized}, where stability conditions for multi-channel linear systems were derived.
Fast forward a few decades, the stability of networks was studied in \cite{fax2004information}, where sufficient graph-theoretic conditions were provided for stability of formations comprised of identical vehicles.
%
Graph-based structured controller design was further examined in works such as \cite{borrelli2008distributed, massioni2009distributed}.
The topic was also studied from a spatially distributed control viewpoint~\cite{bamieh2002distributed} or a layered control design approach~\cite{alemzadeh2018influence}.
However, all of these studies are based on the knowledge of the underlying dynamics; as the system grows in scale, modeling becomes prohibitively difficult and uncertain due to complexities or potential perturbations in high dimensions.\footnote{For example, computational complexity of order $\mathcal{O}(n^3)$ for solving the Algebraic Riccati Equation (ARE) is not desirable for large-scale systems.}
This motivates a data-guided approach to evade the difficulties of model-based distributed control.
In this work, we focus on a model-free distributed control design using $Q$-learning \cite{watkins1992q}.
The approach has been used to find the optimal Linear Quadratic Regulator (LQR) feedback controller online for a single system \cite{bradtke1994incremental}.
In that sense, $Q$-learning can be thought of as an adaptive optimal control design method \cite{lewis2012reinforcement}.
More recent works on distributed adaptive systems can be found in \cite{gao2017data,nguyen2017selectively} to name a few.
Other related works are~\cite{kar2013cal} that considers network effects within $Q$-functions, and \cite{nguyen2017selectively} that introduces a decentralized $Q$-learning approach for a general framework but not necessarily on an underlying graph.
Moreover, there has emerged several recent works on data-driven control using finite sample opposed to the adaptive control design in the limit \cite{fattahisample,dean2017sample,alaeddini2018linear}.
Our contribution is mainly built upon the work of Bradtke~\cite{bradtke1994incremental}, in that we provide a distributed policy iteration to find a collective controller.
We assume that the distributed nature of the problem comes from the interconnection of identical dynamically decoupled systems that work together for a common system-level goal in a network.
In fact, we assume that the underlying interaction graph is only reflected in the performance index of the corresponding LQR problem.
We provide a graph theoretical framework under which each agent synthesizes an estimate of the optimal LQR controller.
%
The rest of the paper is organized as follows: In \S\ref{sec:math} we provide a quick overview of
mathematical tools that are used in the paper.
In \S\ref{sec:p-setup} we introduce the problem setup. In \S\ref{sec:DQL} the distributed setup and the main algorithm is provided along with its proof of convergence.
The section concludes with a discussion on the computational savings due to the adoption of the distributed algorithm.
An example is provided in \S\ref{sec:example} to validate our theoretical results.
Concluding remarks and future directions are discussed in \S\ref{sec:conclusion}.
\section{Mathematical Preliminaries}
\label{sec:math}
We denote by $\mathbb{R}$ the set of real numbers.
A column vector with $n$ elements is referred to as $v\in\mathbb{R}^n$, where $v_i$ represents the $i$th element in $v$.
The matrix $M\in \mathbb{R}^{p\times q}$ contains $p$ rows and $q$ columns with $[M]_{ij}$ denoting the element in the $i$th row and $j$th column of $M$.
The square matrix $N\in\mathbb{R}^{n\times n}$ is \emph{symmetric} if $N^{\top}=N$, where $N^\top$ denotes the \emph{transpose} of the matrix $N$.
%
The operator \emph{diag(.)} makes a square diagonal matrix out of the elements of its argument.
The $n\times n$ zero matrix is denoted by $\textbf{0}_n$ and $\textsc{I}_n = \text{diag}(1,1,\dots,1)$, is the \emph{identity matrix}.
We write $N\succ0$ ($\succeq0$) when $N$ is a positive-(semi)definite matrix, i.e., $x^{\top}Nx>0$ ($\geq0$) for all $x\neq0$.
To simplify the vector notation, we use semicolon (;) to concatenate column vectors, hence $[v^{\top}\ w^{\top}]^{\top} = [v;w]$.
We call the pair $(A,B)$ \emph{controllable}, if and only if the controllability matrix $\mathcal{C}=[B\quad AB\ \dots\ A^{n-1}B]$ has full-rank, where $n$ is the size of the system.
$A\otimes B\in\mathbb{R}^{p_1q_1\times p_2q_2}$ refers to the \emph{Kronecker product} of $A\in\mathbb{R}^{p_1\times q_1}$ and $B\in\mathbb{R}^{p_2\times q_2}$, and $\textsc{I}\otimes T$ gives a block diagonal square matrix, with $T$ on each diagonal block.
A graph is characterized by $\mathcal{G}=(\mathcal{V},\mathcal{E})$ where $\mathcal{V}$ is the set of nodes and $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$ denotes the set of edges.
An edge exists from node $i$ to $j$ if $(i,j)\in \mathcal{E}$;
this is also specified by writing $j\in\mathcal{N}_i$, where $\mathcal{N}_i$ is the set of neighbors of node $i$.
Finally, $\mathcal{G}$ can be represented by various matrices, in particular, by its \emph{graph Laplacian} denoted by $\mathcal{L}$.
\section{Problem Setup}
\label{sec:p-setup}
Herein, we provide the basic formulation and problem setup.
First, we introduce the distributed LQR control on a given graph.
This is mainly related to~\cite{borrelli2008distributed}, where the network contains identical dynamic agents,
yet decoupled from other agents' dynamics.
The only coupling between these agents is through a common network-level objective function.
Then we introduce the basic setup of $Q$-learning for linear dynamical systems and extend the formulation for the distributed setup.
As we shall see later, the distributed nature will be simplified into an additional interaction term in the output of a linear Recursive Least Squares (RLS) algorithm.
\subsection{Distributed LQR Problem}
\label{subsec:disLQR}
Assume that the system contains $N$ agents that form a graph $\mathcal{G}$ with each node of the graph indicating a linear time-invariant dynamical system corresponding to that agent as,
\begin{align*}
x_{t+1}^{(i)} = Ax_t^{(i)} + Bu_t^{(i)},
\qquad i = 1,2,\dots,N,
\end{align*}
where $x_t^{(i)}$ is the state of agent $i$ at time-step $t$, where $x_t^{(i)}\in\mathbb{R}^n$, $u_t^{(i)}\in\mathbb{R}^m$, $A\in\mathbb{R}^{n\times n}$, and $B\in\mathbb{R}^{n\times m}$.
The assumption that all agents have identical system matrices $A$ and $B$ is relevant in many applications such as formation flight, homogenous mobile robots, and power grids consisting of identical generators.
These dynamics can be integrated into a compact form as, $\tilde{\textbf{x}}_{t+1} = \tilde{\textbf{A}} \tilde{\textbf{x}}_{t} + \tilde{\textbf{B}} \tilde{\textbf{u}}_{t},$
where $\tilde{\textbf{x}}\in\mathbb{R}^{Nn}$ and $\tilde{\textbf{u}}\in\mathbb{R}^{Nm}$ are formed by concatenation of all states and inputs into one vector with $\tilde{\textbf{A}} = \textsc{I}_{N} \otimes A\in\mathbb{R}^{Nn\times Nn}$ and $\tilde{\textbf{B}} = \textsc{I}_{N} \otimes B\in\mathbb{R}^{Nn\times Nm}$.
The graph structure is reflected in the cost function of the associated LQR problem by the following definition,
\begin{equation}
\begin{aligned}
\textbf{J} &= \sum_{i=1}^{N} \big[ \underbrace{x_t^{(i)\top} Q_i x_t^{(i)} + u_t^{(i)\top} R_i u_t^{(i)} }_{\text{System $i$ Performance}} \big] \\
& \hspace{8mm}+\sum_{i=1}^{N} \sum_{j\neq i}^{N} \big[ \underbrace{\ \big( x_t^{(i)}-x_t^{(j)} \big)^{\top} Q_{ij} \big( x_t^{(i)} - x_t^{(j)} \big) }_{\text{Systems $i$ and $j$ Coupling}}\big],
\end{aligned}
\label{eq:cost}
\end{equation}
where the first term indicates the intra-systems cost while the second denotes the inter-system coupling.
We make the simplifying assumption,
\begin{align*}
Q_i = \bar{Q},\quad R_i = \bar{R},\quad Q_{ij} = \bar{Q}_{\{ j\in\mathcal{N}_i \}} = \begin{cases}
\textbf{0}_n \hspace{5mm} j\not\in\mathcal{N}_i \\
\bar{Q} \hspace{6mm} j\in\mathcal{N}_i
\end{cases}
\end{align*}
where $\bar{Q}\succeq 0$ and $\bar{R}\succ 0$.
The cost function can also be written in compact form as $\textbf{J} = \tilde{\textbf{x}}_t^{\top} \tilde{\textbf{Q}} \tilde{\textbf{x}}_t + \tilde{\textbf{u}}_t^{\top} \tilde{\textbf{R}} \tilde{\textbf{u}}_t$
where, $\tilde{\textbf{Q}} = (\mathcal{L}+\textsc{I}_n) \otimes \bar{Q} \succeq 0$ and $\tilde{\textbf{R}} = \textsc{I}_n \otimes \bar{R} \succ 0$.
Solution of the LQR problem in such systems is studied for a particular $\tilde{\textbf{Q}}$ resulting in a structured controller~\cite{borrelli2008distributed}.
Suboptimal solutions to the controller design consistent with the graph structure has also been proposed.
Nevertheless, in many real-world applications there is no \emph{a priori} knowledge of the system's model due to either complexities or model uncertainties~\cite{alemzadeh2018influence}.
We introduce a model-free approach while considering the optimality criteria for each subsystem.
We will show that for an interconnected system with identical dynamically decoupled agents as discussed above, $Q$-learning leads to each subsystem running their respective local LQR optimal controller independent of other agents in the network.
This phenomenon is shown to hold asymptotically after each agent collects enough data.
\vspace{2mm}
\begin{remark}
We note that the global cost in \cref{eq:cost} induces a structured way of steering the states of the agents to the origin through an auxiliary consensus term.
Our future work will consider further realizations of the global/local cost structure in the LQR setup--that might not be completely aligned with each other.
%
\end{remark}
\subsection{Centralized $Q$-Learning}
\label{subsec:disQfun}
To make the paper self-contained, we refer to some basics of $Q$-learning and its connections to LQR feedback control design.
$Q$-learning describes a methodology where an agent aims to optimize the value of a sum of reward functions from observing the results of its own actions.
This value is reformulated by the $Q$-function which is defined for a single agent as,
\begin{equation}
\begin{aligned}
Q(x_t,u_t) = R(x_t,u_t) + \gamma Q(x_{t+1},u_{t+1}),
\end{aligned}
\label{eq:Qdefine}
\end{equation}
where $Q(x_t,u_t)=x_t^{\top}Px_t$ is the \emph{state-action} $Q$-function, $P$ is the \emph{cost-to-go} matrix,\footnote{which is also the solution to the discrete-time ARE in LQR.} and $R(x_t,u_t)=x_t^{\top}\bar{Q} x_t + u_t^{\top}\bar{R} u_t$ is the \emph{one-step reward} with symmetric constant matrices $\bar{Q}\succeq 0$ and $\bar{R}\succ 0$.
\Cref{eq:Qdefine} is the simplified form of the well-known Bellman equation for the deterministic case of LQR.
Also, the control actions come from a set of optimal policies that assume the form of a feedback law $u_{t}=-Kx_{t}$ in the LQR framework.
Simplification of \eqref{eq:Qdefine} results in,
\begin{equation}
\begin{aligned}
Q(x_t,u_t) = z_t^{\top} H z_t,
\end{aligned}
\label{eq:Qfun}
\end{equation}
where $z_t = [x_t;u_t]$ and $H$ is a block matrix defined as,
\begin{align*}
H = \begin{bmatrix}
H_{11} & H_{12} \\
H_{21} & H_{22}
\end{bmatrix}
= \begin{bmatrix}
\bar{Q} + \gamma A^{\top}PA & \gamma A^{\top} P B \\
\gamma B^{\top} P A & \bar{R} + \gamma B^{\top} P B
\end{bmatrix}.
\end{align*}
Then the idea is to learn the parameters in $H$ through observations $z_t$ and update the estimate of the controller as,
\begin{align*}
K_{\text{new}} = -H_{22}^{-1} H_{21} = -\gamma (R + \gamma B^{\top} P B)^{-1} (B^{\top} P A),
\end{align*}
which can also be obtained by setting $\partial Q/\partial u_t = 0$.
The adaptive nature of the algorithm is originated from a linear RLS step to learn the parameters of $H$ in real-time.
Hence, we pursue \cite{bradtke1994incremental} to form a linear parameterization of \eqref{eq:Qfun} as,
\begin{align}
Q(x_t,u_t) = z_t^{\top} H z_t = \bar{z}^{\top}_t \theta_H,
\label{eq:quad}
\end{align}
where $\bar{z}_t,\theta_H\in\mathbb{R}^{(n+m)(n+m+1)/2}$ are quadratic basis of the elements in $z_t$ and vector of upper right triangle of symmetric $H$ in the correct order, respectively.
With these definitions,
\begin{equation}
\begin{aligned}
R(x_t,u_t) = r_t &= Q(x_t,u_t) - \gamma Q(x_{t+1},u_{t+1}) \\
&= z_t^{\top} H z_t - \gamma z_{t+1}^{\top} H z_{t+1} = \phi_t^{\top} \theta_H,
\end{aligned}
\label{eq:reward}
\end{equation}
where $\phi_t = \bar{z}_t - \gamma\bar{z}_{t+1}$.
Therefore, assuming that we know $R(x_t,u_t)$ and $\phi_t$, RLS can be employed to find an estimate of $\theta_H$.
According to \cite{goodwin2014adaptive}, this recursive algorithm converges in the limit if $\phi_t$ is persistently excited (PE), i.e.,
\begin{align}
\alpha\textsc{I} \leq \frac{1}{M} \sum_{i=1}^M \phi_{t-i}\phi_{t-i}^{\top} \leq \beta\textsc{I} \qquad \forall\ t,M\geq M_0,
\label{eq:PE}
\end{align}
for some positive parameters $M_0$, $\alpha$, and $\beta$.
Following the convergence of $\theta_H$, then $H$ is obtained using \eqref{eq:quad}.
\section{Distributed $Q$-learning}
\label{sec:DQL}
\subsection{Distributed $Q$-function}
\label{sec:DQfun}
We now switch to a multiagent setup, where several autonomous agents try to minimize their own discounted reward based on a global cost and single-agent control is not applicable since there exist multiple decision-makers.
In this section, we extend the $Q$-learning setup based on the distributed control framework defined in \cref{subsec:disLQR}.
To this end, we assume that each agent enjoys its own $Q$-function whose reward is a function of the state of the agent as well as the state of its neighbors.
For agent $i$ we define,
\begin{equation}
\begin{aligned}
Q^{(i)}(x_t^{(i)},u_t^{(i)}) &= R^{(i)}(x_t^{(i)},u_t^{(i)}) + \gamma Q^{(i)}(x_{t+1}^{(i)},u_{t+1}^{(i)}) \\
&= y_t^{(i)\top} \mathcal{Q}^{(i)} y_t^{(i)} + \gamma x_{t+1}^{(i)\top} P^{(i)} x_{t+1}^{(i)},
\end{aligned}
\label{eq:i}
\end{equation}
where $y_t^{(i)} = [x_t^{(i)} ;\ u_t^{(i)} ;\ x_t^{(j_1)} ;\ \dots\ ;\ x_t^{(j_{d_i})}]$, $d_i$ is the degree of agent $i$, $j_{d_k}\in\mathcal{N}_i$ for $k=1,\dots,i$, and $\mathcal{Q}^{(i)}$ is defined as, \footnotesize
\begin{equation}
\begin{aligned}
\mathcal{Q}^{(i)} = \begin{bmatrix}
(d_i+1)\bar{Q} & \textbf{0} & -\bar{Q} & \dots & -\bar{Q} \\
\textbf{0} & \bar{R} & \textbf{0} & \dots & \textbf{0} \\
-\bar{Q} & \textbf{0}& \bar{Q} & \dots & \textbf{0} \\
\vdots & \vdots & \vdots & \ddots& \vdots \\
-\bar{Q} & \textbf{0} & \textbf{0} & \dots& \bar{Q}
\end{bmatrix}\in\mathbb{R}^{(d_i+2)n\times (d_i+2)n}.
\end{aligned}
\label{eq:bigQ}
\end{equation} \normalsize
The structure of $\mathcal{Q}^{(i)}$ is resulting from \cref{eq:cost} and implies the new definition of reward function for multiple agents in the system.
Note that \cref{eq:i,eq:bigQ} make two implicit assumptions: ($i$) there is no control coupling amongst agents and, ($ii$) each agent has only access to the reward form the coupling between its own state and the states of neighbors.
This motivates the existence of zero blocks in \eqref{eq:bigQ}.
Similar to \eqref{eq:Qfun}, \cref{eq:i} can also be re-arranged into,
\begin{equation*}
\begin{aligned}
Q^{(i)}(x_i,u_i) = y_t^{(i)\top} H^{(i)} y_t^{(i)},
\end{aligned}
\end{equation*}
where, \scriptsize
\begin{equation*}
\hspace{-0.8mm}
\begin{aligned}
H^{(i)} = \begin{bmatrix}
(d_i+1)\bar{Q}+\gamma A^{\top}P^{(i)}A & \gamma A^{\top} P^{(i)} B & -\bar{Q} & \dots & -\bar{Q} \\
\gamma B^{\top} P^{(i)} A & \bar{R} + \gamma B^{\top} P^{(i)} B & \textbf{0} & \dots & \textbf{0} \\
-\bar{Q} & \textbf{0} & \bar{Q} & \dots & \textbf{0} \\
\vdots & \vdots & \vdots& \ddots & \vdots \\
-\bar{Q} & \textbf{0} & \textbf{0} & \dots & \bar{Q}
\end{bmatrix}
\end{aligned}
\end{equation*}
\normalsize
Since there is no control coupling, in order to update the controller for each agent at each iteration we set again,
\begin{align}
K_{\text{new}}^{(i)} = -H_{22}^{(i)-1} H_{21}^{(i)}.
\label{eq:update}
\end{align}
Finally, as in the centralized case, for each agent $i$ we define $\phi_t^{(i)} = \bar{z}_t^{(i)} - \gamma\bar{z}_{t+1}^{(i)}$.
\subsection{Main Results}
\label{sec:algorithm}
In this section, we introduce the distributed policy iteration algorithm.
The analysis in this part is mainly inspired by \cite{bradtke1994incremental}, however, there are fundamental differences as we only assume couplings through a global cost function; as such, the state transition or feedback of each agent only depends on their own history of states and actions.
Under these assumptions, we show that this way of coupling in the case of identical systems signifies the interdependency of the agents in the decision-making process.
\begin{algorithm}[H]
\label{alg:DRL}
\caption{The distributed Q-learning Algorithm}\label{euclid}
\begin{algorithmic}[1]
\State \textbf{Initialize}:
\State $\hspace{5mm}$ Random: $\hat{\theta}_0^{(1)}(0),\dots,\hat{\theta}_0^{(N)}(0)$
\State $\hspace{5mm}$ Stabilizable: $K_0^{(1)},\dots,K_0^{(N)}$
\State $\hspace{5mm}$ $t=0$, $k=1$
\State \textbf{while convergence:}
\State $\hspace{5mm}$ Reset Covariance: $P_k(0)=P_0$
\State $\hspace{5mm}$ \textbf{For} $j=1$ to $M$:
\State $\hspace{10mm}$ \textbf{For system} $i=1,\dots,N$:
\State $\hspace{15mm}$ Choose $e_t$ and find $u_t^{(i)}=K_k^{(i)}x_t^{(i)}+e_t$
\State $\hspace{15mm}$ Collect $x_{t+1}^{(i)}$ by applying $u_t^{(i)}$ to the system
\State $\hspace{15mm}$ Update $\hat{\theta}_k^{(i)}(j)$ using RLS
\State $\hspace{15mm}$ $t=t+1$
\State $\hspace{5mm}$ \textbf{For system} $i=1,\dots,N$:
\State $\hspace{10mm}$ Find symmetric $\hat{H}_k^{(i)}$ corresponding to $\hat{\theta}_k^{(i)}$
\State $\hspace{10mm}$ Policy update: $K_{k+1}^{(i)}=-\hat{H}_{k(22)}^{{(i)}^{-1}}\hat{H}^{(i)}_{k(21)}$
\State $\hspace{10mm}$ Initialize parameters $\hat{\theta}_{k+1}^{(i)}(0)=\hat{\theta}_k^{(i)}(M)$
\State $\hspace{5mm}$ $k=k+1$
\end{algorithmic}
\label{alg:1}
\end{algorithm}
\noindent We briefly explain the steps of the algorithm:
$\hat{\theta}_k^{(i)}$ is the estimate of $H^{(i)}$ as in \eqref{eq:quad}.
In the sequel, $\theta^{*(i)}_k$ denotes the parameters of $H^{(i)}$ obtained using the true system parameters.
$K_k^{(i)}$ denotes the controller estimate.
The counter $t$ keeps track of the number of collected data while $k$ designates the iteration count on the parameters estimate.
Note that these counters are never reset to zero.
$P_k(j)$ is the covariance matrix reset to some constant $P_0$ at each iteration to revitalize the gain.
Each RLS estimation interval includes $M$ time-steps.
The value of $M$ is dependent on the number of unknown parameters in $\hat{\theta}_k^{(i)}$ and also the desired accuracy.
The control signal is PE at each iteration of the RLS and $e_t$ is the excitation component which is assumed to be the same for all agents.
After convergence of RLS, the controller for each agent is updated based on \eqref{eq:update}.
The estimation parameters are reinitialized from the final value of the previous iteration such that $\hat{\theta}_{k+1}(0) = \hat{\theta}_{k}(M)$.
The reader is referred to Chapter 3 of \cite{goodwin2014adaptive} for exact steps of RLS.
\vspace{2mm}
\begin{theorem}
Assume that for all $i=1,\dots,N$, the pair $(A,B)$ is a controllable and $K_0^{(i)}$ is stabilizing with a PE signal $\phi_t^{(i)}$.
Then there exists $M<\infty$ such that \Cref{alg:1} generates a sequence $\{ K_k^{(i)}\}$ with $\lim_{k\rightarrow\infty} \|K_k^{(i)} - K^{*}\| = 0$, where $K^{*}=\text{LQR}(A,B,\bar{Q},\bar{R})$.
\end{theorem}
\vspace{2mm}
\begin{proof}
From \eqref{eq:reward},
\begin{align*}
r^{(i)}_t = y_t^{(i)\top} H^{(i)} y_t^{(i)} - \gamma y_{t+1}^{(i)\top} H^{(i)} y_{t+1}^{(i)}.
\end{align*}
Also from \cref{sec:DQfun},
\begin{align*}
r_t^{(i)} &= x_t^{(i)\top} \bar{Q} x_t^{(i)} + u_t^{(i)\top} \bar{R} u_t^{(i)} \\
&\hspace{15mm} + \sum_{k=1}^{d_i} \big( x_t^{(i)}-x_t^{(j_k)} \big)^{\top} \bar{Q} \big( x_t^{(i)}-x_t^{(j_k)} \big),
\end{align*}
and, \footnotesize
\begin{equation*}
\begin{aligned}
y_t^{(i)\top} & H^{(i)} y_t^{(i)} \\
&= \big[x_t^{(i)\top}\ u_t^{(i)\top} \big]
\begin{bmatrix}
\bar{Q}+\gamma A^{\top}P^{(i)}A & \gamma A^{\top}P^{(i)}B \\
\gamma B^{\top}P^{(i)}A & \bar{R}+\gamma B^{\top}P^{(i)}B
\end{bmatrix}
\begin{bmatrix}
x_t^{(i)} \\ u_t^{(i)}
\end{bmatrix} .\\
\end{aligned}
\end{equation*} \normalsize
Consequently, we obtain,
\begin{equation}
\begin{aligned}
\xi_t^{(i)} = \bar{z}_t^{(i)\top} \hat{\theta}^{(i)}_k
\end{aligned}
\label{eq:RLSagent}
\end{equation}
where,
\begin{equation}
\begin{aligned}
\xi_t^{(i)} = x_t^{(i)\top}& \bar{Q} x_t^{(i)} + u_t^{(i)\top} \bar{R} u_t^{(i)} \\
&+ \sum_{k=1}^{d_i} \big( x_{t+1}^{(i)}-x_{t+1}^{(j)} \big)^{\top} \bar{Q} \big( x_{t+1}^{(i)}-x_{t+1}^{(j)} \big).
\end{aligned}
\label{eq:introCoupling}
\end{equation}
Hence the distributed nature of the problem narrows down to a particular distributed form of RLS.
We stack the $N$ equations of the form \eqref{eq:RLSagent} for all agents into vector form as,
\begin{align}
\underbrace{\begin{bmatrix}
\xi^{(1)}_t \\ \vdots \\ \xi^{(N)}_t
\end{bmatrix}}_{\Xi_t} =
\underbrace{\begin{bmatrix}
\bar{z}_t^{(1)^{\top}} & & \\
& \ddots & \\
& & \bar{z}_t^{(N)^{\top}}
\end{bmatrix}}_{Z_t}
\underbrace{\begin{bmatrix}
\hat{\theta}_k^{(1)} \\ \vdots \\ \hat{\theta}_k^{(N)}
\end{bmatrix}}_{\hat{\Theta}_k} .
\label{eq:RLS}
\end{align}
Based on the definition of PE in \eqref{eq:PE}, it is straightforward to show that the matrix $Z_t$ is PE if $\bar{z}_t^{(i)}$ is PE for all $i$.
This results in the convergence of \cref{eq:RLS} to some $\Theta^*$ for large enough $M$.\footnote{Parameter estimation for the multi-output system is an straightforward extension of the scalar case and is discussed in Chapter 3.8 of \cite{goodwin2014adaptive}.}
From Theorem 5.1 in \cite{bradtke1994incremental},
\begin{equation}
\begin{aligned}
\lim_{k\rightarrow\infty} \| \hat{\theta}^{(i)}_k - \theta^{*(i)}_k \| = 0, \quad
\lim_{k\rightarrow\infty} \| \theta^{*(i)}_k - \theta^{*(i)}_{k-1} \| = 0.
\end{aligned}
\label{eq:conv0}
\end{equation}
However, the convergence of $\hat{\theta}_k^{(i)}$ for all $i$ to one single value is non-trivial due to the interdependency in RLS.
We will show that for a connected network of agents,
\begin{align*}
\lim_{t\rightarrow \infty}\ \| x_t^{(i)} - x_t^{(j)} \| = 0,
\end{align*}
for any $i$ and $j$.
Note that according to \eqref{eq:introCoupling}, if a node is disconnected from the graph it can be individually examined as in the centralized case.
Recall that for $\ell = i, j$,
\begin{align*}
x_{t+1}^{(\ell)} = Ax_t^{(\ell)} + Bu_t^{(\ell)} = (A-BK_k^{(\ell)})x_t^{(\ell)} + Be_t.
\end{align*}
As such,
\begin{align}
x_{t+1}^{(i)} - x_{t+1}^{(j)} = (A-BK_k^{(i)}) (x_t^{(i)} - x_t^{(j)}) + B \Delta K_k x_t^{(j)}
\label{eq:difference}
\end{align}
where $\Delta K_k = K_k^{(i)} - K_k^{(j)}$.
Then if we show that $\|\Delta K_k\|\rightarrow 0$ as $k\rightarrow \infty$ we obtain,
\begin{align}
\| x_{t+1}^{(i)} - x_{t+1}^{(j)} \| = \| (A-BK_k^{(i)})^{t+1} (x_0^{(i)} - x_0^{(j)}) \| \rightarrow 0,
\label{eq:final}
\end{align}
given that the policy iteration algorithm leads to a more stabilizing controller $K_k^{(i)}$ as $k$ increases \cite{bradtke1994incremental}.
Then,
\begin{equation}
\begin{aligned}
\| K_k^{(i)} &- K_k^{(j)} \| \\
&= \| \hat{H}_{{k-1}_{(22)}}^{{(j)}^{-1}}\hat{H}^{(j)}_{{k-1}_{(21)}} - \hat{H}_{{k-1}_{(22)}}^{{(i)}^{-1}}\hat{H}^{(i)}_{{k-1}_{(21)}} \| \\
&= \| \hat{H}_{{k-1}_{(22)}}^{{(j)}^{-1}} \Big( (\hat{H}^{(j)}_{{k-1}_{(21)}} - \hat{H}^{(i)}_{{k-1}_{(21)}})\\
&\hspace{10mm} + (\hat{H}^{(i)}_{{k-1}_{(22)}} - \hat{H}^{(j)}_{{k-1}_{(22)}}) \hat{H}_{{k-1}_{(22)}}^{{(i)}^{-1}} \hat{H}^{(i)}_{{k-1}_{(21)}} \Big) \| .
\end{aligned}
\label{eq:HH}
\end{equation}
Since $\hat{H}_{22}$ and $\hat{H}_{21}$ contain only a subset of elements in $\hat{\theta}$,
\begin{equation}
\begin{aligned}
\| \hat{H}^{(j)}_{{k-1}_{(21)}} - \hat{H}^{(i)}_{{k-1}_{(21)}} \| &\leq \| \hat{\theta}_{k-1}^{(j)} - \hat{\theta}_{k-1}^{(i)} \|, \\
\| \hat{H}^{(j)}_{{k-1}_{(22)}} - \hat{H}^{(i)}_{{k-1}_{(22)}} \| &\leq \| \hat{\theta}_{k-1}^{(j)} - \hat{\theta}_{k-1}^{(i)} \|.
\end{aligned}
\label{eq:ineq}
\end{equation}
Hence equations \eqref{eq:HH} and \eqref{eq:ineq} lead to,
\begin{equation}
\begin{aligned}
\| & K_k^{(i)} - K_k^{(j)} \| \\
&\leq \| \hat{H}_{{k-1}_{(22)}}^{{(j)}^{-1}} \|\ \| \hat{\theta}_{k-1}^{(j)} - \hat{\theta}_{k-1}^{(i)} \|\ \| 1 + \hat{H}_{{k-1}_{(22)}}^{{(i)}^{-1}} \hat{H}^{(i)}_{{k-1}_{(21)}} \| \\
&\leq \kappa_0 \| \hat{\theta}_{k-1}^{(j)} - \hat{\theta}_{k-1}^{(i)} \| ,
\end{aligned}
\label{eq:uBound}
\end{equation}
where we have used the fact that the estimated parameters are bounded and $\kappa_0 > 0$ is a constant such that,
\begin{align*}
\| \hat{H}_{{k-1}_{(22)}}^{{(j)}^{-1}} \| ~.~ \| 1 + \hat{H}_{{k-1}_{(22)}}^{{(i)}^{-1}} \hat{H}^{(i)}_{{k-1}_{(21)}} \| \leq \kappa_0.
\end{align*}
From Lemma 5.2 in \cite{bradtke1994incremental},
\begin{align*}
\| \theta_k^{*(\ell)} - \hat{\theta}_k^{(\ell)} \| \leq \epsilon_M \big( \| \theta_k^{*(\ell)} - \theta_{k-1}^{*(\ell)}\| + \| \theta_{k-1}^{*(\ell)} - \hat{\theta}_{k-1}^{(\ell)} \| \big),
\end{align*}
which for large enough $M$ results in,
\begin{equation}
\begin{aligned}
\label{eq:conv1}
&\| \theta_k^{*(i)} - \hat{\theta}_k^{(i)} \| + \| \theta_k^{*(j)} - \hat{\theta}_k^{(j)} \| \\
&\hspace{13mm} \leq \epsilon_M \big( \| \theta_k^{*(i)} - \theta_{k-1}^{*(i)}\| + \| \theta_{k-1}^{*(i)} - \hat{\theta}_{k-1}^{(i)} \| \\
&\hspace{19.5mm} + \| \theta_k^{*(j)} - \theta_{k-1}^{*(j)}\| + \| \theta_{k-1}^{*(j)} - \hat{\theta}_{k-1}^{(j)} \| \big).
\end{aligned}
\end{equation}
Using triangle inequality on the left side of this inequality,
\begin{equation}
\begin{aligned}
\label{eq:conv2}
&\Big| \| \hat{\theta}_k^{(i)} - \hat{\theta}_k^{(j)} \| - \| \theta_k^{*(i)} - \theta_k^{*(j)} \| \Big| \\
&\hspace{23mm}\leq \big\| \big( \hat{\theta}_k^{(i)} - \hat{\theta}_k^{(j)} \big) - \big( \theta_k^{*(i)} - \theta_k^{*(j)} \big) \big \| \\
&\hspace{23mm}\leq \| \theta_k^{*(i)} - \hat{\theta}_k^{(i)} \|\ +\ \| \theta_k^{*(j)} - \hat{\theta}_k^{(j)} \|.
\end{aligned}
\end{equation}
Then, from \eqref{eq:conv1} and \eqref{eq:conv2} and for large $k$,
\begin{align*}
\| \hat{\theta}_k^{(i)} - \hat{\theta}_k^{(j)} \| &\leq \epsilon_M \Big( \| \theta_k^{*(i)} - \theta_{k-1}^{*(i)}\| + \| \theta_{k-1}^{*(i)} - \hat{\theta}_{k-1}^{(i)} \| \\
&\hspace{7mm} +\| \theta_k^{*(j)} - \theta_{k-1}^{*(j)}\| + \| \theta_{k-1}^{*(j)} - \hat{\theta}_{k-1}^{(j)} \| \Big) \\
&\hspace{7mm} + \| \theta_k^{*(i)} - \theta_k^{*(j)} \|.
\end{align*}
Hence using the result in \eqref{eq:conv0},
\begin{align*}
\| \hat{\theta}_k^{(i)} - \hat{\theta}_k^{(j)} \| \rightarrow 0,
\end{align*}
and plugging this into \eqref{eq:uBound},
\begin{align}
\|\Delta K_k\| = \| K_k^{(i)} - K_k^{(j)} \| \rightarrow 0,
\end{align}
Hence,
\begin{align*}
\| x_{t+1}^{(i)} - x_{t+1}^{(j)} \| \rightarrow 0.
\end{align*}
This implies that based on \eqref{eq:introCoupling}, for identical systems the algorithm moves towards $N$ decoupled $Q$-learning algorithms for each agent.
Thus, although the provided data is from an interconnected system, each controller converges to its optimal value, i.e.,
\begin{align*}
\lim_{k\rightarrow\infty} \|K_k^{(i)} - K^{*}\| = 0, \qquad \text{for} \quad i=1,2,\dots,N .
\end{align*}
\end{proof}
\begin{remark}
In \Cref{alg:1}, we have assumed that the exploration signal, $e_t$, is equal for every agent at each time step.
This is a valid assumption as long as $Z_t$ in \eqref{eq:RLS} is PE so that RLS is assured to converge.
Another option would be to choose the excitation signals $e_t^{(i)}$ and $e_t^{(j)}$ in a way that,
\begin{align*}
e_t^{(i)} - e_t^{(j)} = -\Delta K_k x_t^{(j)}.
\end{align*}
Hence, not only the input to the RLS is PE, the difference cancels out $\Delta K_k x_t^{(j)}$ in \eqref{eq:difference}.
However, this setup is more challenging to implement, particularly for large-scale systems.
\end{remark}
\subsection{Computational Saving}
\label{sec:comSaving}
The computational saving resulting from using the distributed $Q$-learning algorithm is significant, since for a large system, the design of the LQR controller with the computational complexity of solving ARE of order $\mathcal{O}(n^3)$, can be prohibitively expensive.
The main computational burden of \Cref{alg:1} comes from RLS where the complexity of its implementation is $\mathcal{O}(\gamma^2)$ with $\gamma$ parameters to learn.
Assuming that the system contains $N$ agents each having $n$ states and $m$ inputs, the computational complexity of the centralized $Q$-learning is obtained by,
\begin{align*}
\mathcal{O}\Big( \big( \frac{(Nn+Nm)(Nn+Nm+1)}{2} \big)^2 \Big) \approx \mathcal{O} \big( N^4(n+m)^4 \big),
\end{align*}
while for the distributed case the code performs $N$ repetitions of the same RLS leading to the complexity bound,
\begin{align*}
\mathcal{O}\Big( N \big( \frac{(n+m)(n+m+1)}{2} \big)^2 \Big) \approx \mathcal{O} \big( N(n+m)^4 \big).
\end{align*}
Hence the complexity reduction is,
\begin{align*}
\frac{N^4(n+m)^4 - N(n+m)^4}{N^4(n+m)^4} \times 100 = \frac{N^3-1}{N^3} \times 100 \% ,
\end{align*}
which is substantial for large $N$.
\Cref{tab:save} compares the computational saving for some values of $N$.
\begin{table}[h]
\centering
\resizebox{0.49\textwidth}{!}{
\begin{tabular}{|c||c|c|c|c|c|}
\hline
N & 2 & 3 & 5 & 8 & 100 \\ [0.5ex]
\hline
Saving (\%) & 87.5 & 96.29 & 99.2 & 99.8 & 99.99 \\ [0.5ex]
\hline
\end{tabular}
}
\caption{Approximate computational saving of the distributed $Q$-learning compared to the centralized case in \cite{bradtke1994incremental}.}
\label{tab:save}
\end{table}
\section{Example}
\label{sec:example}
In this section, we provide an example to show the efficiency of the distributed $Q$-learning algorithm for a set of identical communicating UAVs.
We consider the autonomous flight of a network of six interconnected Unmanned Aerial Vehicles (UAVs) which are set to perform a common task such as geographical data collection or putting out a wildfire.
To cover the whole targeted area, these UAVs are programmed to move in parallel and in order for minimal signal transmissions, each UAV only communicates with its closest neighbor in the network as depicted in \Cref{fig:1}.
\begin{figure}[H]
\centering
\includegraphics[scale=0.19]{uav}
\caption{A group of identical firefighting UAVs maneuvering in parallel aiming to extinguish a blaze (Aerial view of the forest fire - \emph{Photo Credit: Alex Punker, Bigstock}).}
\label{fig:1}
\end{figure}
\noindent The discrete-time dynamics of UAVs is considered by mini-aircraft linear parameters that can be found in \cite{hung1982multivariable}.
We assume $\bar{Q}=\textsc{I}_5$ and $\bar{R} = \textsc{I}_3$.
We will show the results of the distributed policy iteration for $N=6$, $n=5$, and $m=3$ and compare the computational performance with the centralized case.
For the distributed algorithm we consider $M=50$ and the exploration signal $e_t$ is generated from a normal distribution.
\Cref{fig:2} shows the results of simulations regarding the controller error norm.
\begin{figure}[H]
\centering
\includegraphics[scale=0.16]{mainPlot}
\caption{Performance of the distributed Q-learning algorithm.
The plot demonstrates the norm of the error between the LQR optimal controller of each subsystem and the estimate of the algorithm at each iteration $k$.}
\label{fig:2}
\end{figure}
\noindent A comparison between the computational performance of the centralized and distributed methods is also provided in \Cref{fig:3}.
For scaling purposes, $M$ and $e_t$ are re-adjusted for each $N$.\footnote{The main reason for this is that $M$ needs to be modified since $N$ is proportionally related to the centralized system dimensions $Nn$ and $Nm$.}
\begin{figure}[H]
\centering
\includegraphics[scale=0.16]{comparePlot}
\caption{Computational performance of distributed and centralized algorithm for different number of interacting agents.}
\label{fig:3}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
In this paper we examined a data-guided approach for the control of large-scale interconnected identical systems with decoupled dynamics; it is assumed that the interconnection is reflected in the cost function for the control design problem.
We leveraged a distributed $Q$-learning as a policy iteration method. In this direction, it is shown that the proposed distributed algorithm converges to each agent's individual optimal controller, which could have been obtained by running a centralized $Q$-learning algorithm.
The significance of the resulting computational savings are also discussed.
There are a number of directions to pursue as future works.
First, the observation in this paper can be further extended to more elaborate cost structure, highlighting the trade-off between local and global optimality in large-scale distributed systems.
This can be achieved if other types of interconnections such as dynamics or feedback coupling as well as consensus through the $Q$-function are adopted for the analysis.\footnote{A similar approach has been examined in \cite{kar2013cal}.}
Another line of work is to consider other types of data-guided distributed control mechanisms for structures such as layering or systems with switching dynamics.
\bibliographystyle{ieeetr}
|
\section{Introduction}
At the frontiers of new physics beyond standard model (SM)
natural or TeV-scale supersymmetry (SUSY) offers us a grand unification (GUT) of SM gauge coupling constants \cite{unif1,unif2,unif3,unif4}.
Such natural SUSY hosts a lot of SUSY particles which can be directly detected at the particle collider LHC
or dark matter direct detection facilities such as Xenon-1T.
Meanwhile, embedding these TeV-scale SUSY models into the ultraviolet completions - SUSY GUT-
may solve the long-standing issues such as SM flavor puzzle and neutrino masses.
Nowadays experimental data seems to oppose the minimal supersymmetric standard model (MSSM)
either from the bottom or top viewpoint.
In the former one,
the $125$ GeV Higgs mass \cite{Higgsmass1, Higgsmass2} requires either large mixing effect or soft masses of order $10$ TeV for the stop scalars \cite{MSSMHiggs1, MSSMHiggs2, MSSMHiggs3}.
When the mixing effects among generations are significant,
the constraints from flavor violation tend to require the SUSY mass order far above the weak scale.
Moreover, the direct detection limits of dark matter \cite{MSSMDark1,MSSMDark2} impose rather strong pressure on the scenario of neutralino dark matter.
In the later perspective,
the minimal SUSY $\rm{SU}(5)$ referring to MSSM is significantly constrained by the proton decay \cite{protonreivew}.
It requires the color-triplet Higgs mass of order GUT scale, which together with unification,
leads to the MSSM mass spectrum at least of order $100$ TeV \cite{Murayama, Hisano,Bajc}.
Therefore, it is of great interest to explore the MSSM with rational extensions that can resurrect the natural SUSY once again.
Along this direction, there are at least two simple examples -the next-to-minimal supersymmetric standard model (NMSSM) \cite{NMSSM}
and the MSSM with vectorlike (VL) fermions (VMSSM) \cite{0807.3055} -
which are consistent with unification \cite{Zheng1, Zheng2}.
While these extensions provide natural explanations of Higgs mass and dark matter in the low energy region,
it is unclear what the statuses of them are in the ultraviolet energy region.
This is the main focus of this study.
In this paper,
we discuss the embeddings of MSSM, NMSSM and VMSSM into realistic grand unification (GUT) \cite{GG, SO101, SO102}.
In each case, both the SUSY $\rm{SU}(5)$ \cite{SUSYSU51,SUSYSU52, SUSYSU53} and SUSY $\rm{SO}(10)$ \cite{SUSYSO10} representations will be explored.
In these SUSY GUTs, we discuss the GUT-scale superpotential
\begin{eqnarray}{\label{W}}
W=W_{Y}+W_{\text{SB}},
\end{eqnarray}
with following features:
\begin{itemize}
\item $W_Y$ and $W_{\text{SB}}$ are both renormalizable.
\item All of vacuum expectation values (vevs) are dynamically generated from $W_{\text{SB}}$.
\item All of SM matters and extended matters obtain their masses via the Higgs mechanism in $W_Y$.
\end{itemize}
Since $W_Y$ is fixed by the SM and the extra matters such as the singlet $N$ or VL fermions at TeV scale,
it is crucial to find suitable content of $W_{\text{SB}}$ that achieves the breaking of gauge group
$G_{\text{GUT}}\rightarrow G_{\text{SM}}$.
In Sec.II, we explore the embedding of MSSM into realistic SUSY GUTs,
where useful conventions and notation will be introduced.
The analysis on the MSSM is of great use to guide us towards the embeddings of NMSSM and VMSSM.
Sec.III and Sec.IV is devoted to study the embeddings of NMSSM and VMSSM into realistic SUSY GUTs, respectively.
Finally, we conclude in Sec.V.
\section{The Benchmark Model: MSSM}
In the minimal $\rm{SU}(5)$ of Standard Model (SM) gauge group,
the SM fermions of each generation are assigned as $1$, $\bar{5}$ ($\psi$) and $10$ ($\Phi$) of $\rm{SU}(5)$ for right-hand neutrino $N_{R}$,
$L$ and down quark $\bar{d}$, and $Q$, up quark $\bar{u}$ and $\bar{e}$, respectively;
whereas in the $\rm{SO}(10)$ representation,
the SM fermions of each generation are embedded into a $16$-dimensional representation,
which decomposes as $16=1+\bar{5}+10$ under the $\rm{SU}(5)$.
\subsection{\rm{SU}(5)}
The Yukawa superpotenail $W_Y$ in Eq.(\ref{W}) contains two parts
\begin{eqnarray}{\label{Yukawa5}}
W^{\rm{SU}(5)}_{Y}=W^{\rm{SU}(5)}_{Y_{f}}+W^{\rm{SU}(5)}_{Y_{\nu}}
\end{eqnarray}
which refers to SM fermions without neutrinos and neutrinos, respectively.
According to the product $\bar{5}\times 10=5+45$ and $10\times10=\bar{5}_{s}+\bar{45}_{a}+\bar{50}_{s}$,
where subscript ``s'' and ``a" refers to symmetric and anti-symmetric, respectively.
With the Higgs representations composed of $5$, $\bar{5}$ and $\bar{45}$ \cite{GJ},
$W^{\rm{SU}(5)}_{Y_{f}}$ is given by,
\begin{eqnarray}{\label{Yukawa5f}}
W^{\rm{SU}(5)}_{Y_{f}}&=&Y_{d}^{ij}\psi_{ai}(\bar{5})\bar{H}_{b}(\bar{5})\Phi^{ab}_{j}(10)\nonumber\\
&+&Y_{45}^{ij}\psi_{ai}(\bar{5})\bar{H}^{a}_{bc}(\bar{45})\Phi^{bc}_{j}(10)\nonumber\\
&+&Y_{u}^{ij}\epsilon_{abcde}\Phi^{ab}_{i}(10)H^{e}(5)\Phi^{cd}_{j}(10)
\end{eqnarray}
where $a, b, c$, etc denote the $\rm{SU}(5)$ indexes, $i, j$ are the generation indexes,
and $Y_{u,d, 45}$ are Yukawa matrixes.
Note, $\Phi^{ab}$ is an anti-symmetric field and $H^{a}_{bc}=-H^{a}_{cb}$.
The reason to include $\bar{45}$ is clear in the SM fermion mass matrixes as derived from Eq.(\ref{Yukawa5f}):
\begin{eqnarray}{\label{massmatrix1}}
M_{u}&=&Y_{u}^{ij}\upsilon^{5}_{u}, \nonumber\\
M_{d}&=&Y_{d}^{ij}\upsilon^{\bar{5}}_{d}+Y^{ij}_{45}\upsilon^{\bar{45}}_{d}, \nonumber\\
M_{e}&=&Y_{d}^{ij}\upsilon^{\bar{5}}_{d}-3Y^{ij}_{45}\upsilon^{\bar{45}}_{d}.
\end{eqnarray}
where $\upsilon^{5}_{u}$ and $\upsilon^{\bar{5}}_{d}$ is vev of Higgs doublet $H_{u}$ and $H_{d}$ in $5$ and $\bar{5}$, respectively;
$\upsilon_{45}$ is the vev of doublet $\sigma_{d}$ in $\bar{45}$,
which is defined as $\left<H^{b5}_{a}\right>=\upsilon_{45}(1,1,1,-3)_{\text{diag}}$ for $a,b=1-4$.
Without the $\bar{45}$, $M_{d}=M_{e}$ at the GUT scale for all of three generations.
While such mass relation is viable for the third generation,
they should be avoided for the first two generations.
Adding $\bar{45}$ can tune the incorrect mass relations to the desired ones,
\begin{eqnarray}{\label{app}}
m_{d}\simeq 3 m_{e},~m_{s}\simeq m_{\mu}/3,~m_{b}\simeq m_{\tau},
\end{eqnarray}
The GUT-scale mass relations in Eq.(\ref{app}) strongly constrain the Yukawa matrixes $Y_{u,d, 45}$.
For example, some specific choices on $Y_{u,d, 45}$ in the Georgi-Jarlskog scheme \cite{GJ} lead to a stable $b$ quark.
In order to solve the SM flavor issue, we choose the Fritzsch scheme \cite{Fritzsch},
in which $Y_{u,d, 45}$ take the following forms
\begin{eqnarray}{\label{Fritzsch}}
Y_{u}^{ij}\upsilon_{u}=
\left(%
\begin{array}{ccc}
0 & A_{u} &0 \\
A_{u} & 0 & B_{u} \\
0 & B_{u} & C_{u}\\
\end{array}%
\right),\nonumber\\
Y_{d}^{ij}\upsilon_{d}=
\left(%
\begin{array}{ccc}
0 & A_{d} & 0 \\
A_{d} & 0 & B_{d} \\
0 & B_{d} & C_{d}\\
\end{array}%
\right),\nonumber \\
Y_{45}^{ij}\upsilon_{45}=
\left(%
\begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & D_{d} \\
0 & D_{d} & 0 \\
\end{array}%
\right),
\end{eqnarray}
where there are small mass hierarchies $C_{f}>>B_{f}>>A_{f}$ with $f=\{u,d,e\}$ in Eq.(\ref{Fritzsch}) so as to address the SM flavor mass hierarchies.
Substituting Eq.(\ref{Fritzsch}) into Eq.(\ref{app}) implies that there is a fine tuning between $B_{d}$ and $D_{d}$,
\begin{eqnarray}{\label{tuning}}
D_{d}=(1\pm \frac{2\sqrt{3}}{3})B_{d}.
\end{eqnarray}
With this fine-tuning solution, the diagonalizations of matrixes in Eq.(\ref{Fritzsch}) yield the CKM matrix
\begin{eqnarray}{\label{s}}
\mid V_{\text{CKM}}\mid \simeq
\left(%
\begin{array}{ccc}
0.974 & 0.227 & 0.004 \\
0.227 & 0.970 & 0.042 \\
0.008 & 0.042 & 0.999 \\
\end{array}%
\right),
\end{eqnarray}
which is in good agreement with experimental data.
For $W_{Y_{\nu}}$ responsible neutrino masses we take a simple form as follows,
\begin{eqnarray}{\label{Yukawan}}
W^{\rm{SU}(5)}_{Y_{\nu}}&=&Y_{N}^{ij}N_{Ri}(1)\psi_{aj}(\bar{5})H^{a}(5)\nonumber\\
&+& Y_{S}^{ij}S(1) \cdot N_{Ri}(1) N_{Rj}(1)+\text{H.c}
\end{eqnarray}
where $S$ is a singlet of $G_{\text{SM}}=SU(3)_{c}\times SU(2)_{L}\times U(1)_{Y}$, with $\langle S\rangle$ of order GUT scale.
In Eq.(\ref{Yukawan}),
one finds the neutrino Dirac mass $M_{\nu}=Y^{ij}_{N}\upsilon^{5}_{u}$ and
the right-hand neutrino mass $M_{N_{R}}=Y^{ij}_{S}\left<S\right>$,
which results in the left-hand neutrino masses in terms of type-I \cite{typeI1,typeI2,typeI3,typeI4} seesaw mechanism,
\begin{eqnarray}{\label{neutrino}}
m_{\nu}=M^{T}_{\nu}M_{N_{R}}^{-1}M_{\nu}=Y_{N}^{T}Y^{-1}_{S}Y_{N}\frac{\upsilon^{5}_{u}}{\langle S\rangle}.
\end{eqnarray}
Given $\upsilon^{5}_{u}\sim 10^2$ GeV, $\langle S\rangle \sim 10^{15}-10^{16}$ GeV and Yukawa couplings of order unity,
the neutrino mass is of order $\sim 10^{-2}-10^{-3}$ eV.
Similar to $Y_{u,d,e}$, $Y_{S}$ and $Y_{N}$ in Eq.(\ref{Yukawan}) are also constrained
by the fit to neutrino mixings as described by PMNS matrix $U_{\text{PMNS}}=U_{e}^{-1}U_{\nu}^{-1}$,
where $U_{\nu}$ and $U_{e}$ are defined to diagonalize mass matrix $m_{\nu}$ and $M_e$, respectively.
After we have established on a benchmark solution\footnote{It is of special interest to examine whether there is viable solution to the input parameters in the case that all mass matrixes such as $M_{u,d,e, N_{R}}$ are assigned the Fritzsch form.}
to the input parameters at the GUT scale which can explain the SM flavor issue and neutrino masses,
we turn to the structure of $W_{\text{SB}}$.
$1)$. In order to obtain light neutrino masses, $\langle S\rangle$ of order GUT scale is required.
$2)$. In order to break the $\rm{SU}(5)$ we introduce $75(Z)$.
With such $75$, we can add a $50$ and $\bar{50}$ to achieve the doublet-triplet splitting \cite{0007254} for $5$ and $\bar{5}$.
$3)$. In order to gain nonzero vev $\upsilon^{\bar{45}}_{d}$, we include another $75(Z')$ with a vev of the GUT scale.
The reason for this is that neither $1$ nor $24$ with large vev is favored by the product $H(5)\times \bar{H}(\bar{5})=1+24$.
$4)$, Due to the singlet $S$ there is an unsafe operator
\begin{eqnarray}{\label{unsafe5}}
W_{\text{unsafe}}=SH(5)\bar{H}(\bar{5}),
\end{eqnarray}
which must be eliminated.
Shown in Table.\ref{MSSMc} is the $Z_{2}\times Z'_{2}$ parity,
which can eliminate the unsafe operator in Eq.(\ref{unsafe5}).
Under this parity, $W_{\text{SB}}$ reads as,
\begin{eqnarray}{\label{SSBM5}}
W^{\rm{SU}(5)}_{\text{SB}}&=& \frac{M_{Z}}{2} Z^{2} +\frac{M_{Z'}}{2} Z'^{2}+\frac{M_{S}}{2}S^{2}+\frac{M_{S'}}{2}S'^{2}\nonumber\\
&+&\frac{\lambda_{Z}}{3} Z^{3}+\frac{\lambda_{S}}{3}S^{3}\nonumber\\
&+&\lambda_{1}SZ^{2}+\lambda_{2}SZ'^{2}+ \lambda_{3}ZZ'^{2}+\lambda_{4}SS'^{2}\nonumber\\
&+&\bar{H}(\bar{45}) \left(M_{45}+\lambda_{5}S\right)H(45) \nonumber\\
&+&\lambda_{6}S' \bar{H}(\bar{50})H(50) \nonumber\\
&+& \lambda_{7} Z \bar{H}(\bar{5})H(50)+\lambda_{8}Z' H(5)\bar{H}(\bar{50})\nonumber\\
&+& \lambda_{9}Z H(45)\bar{H}(\bar{5})+\text{H.c}.
\end{eqnarray}
According to the $F$ terms in Eq.(\ref{SSBM5}),
the nonzero singlet vevs $ \langle(1,1,1)_{Z}\rangle=Z$, $\langle(1,1,1)_{Z'}\rangle=Z'$, $\langle(1,1,1)_{S}\rangle=S$ and $\langle(1,1,1)_{S'}\rangle=S'$ are given by,
\begin{eqnarray}{\label{vevs}}
\frac{S}{M_{Z}}&=&-\frac{1}{2\lambda_{4}}\eta_{3}= -\frac{1}{2\lambda_{2}}\left(-2\lambda_{3}\frac{Z}{M_{Z}}+\eta_{1}\right), \nonumber\\
\frac{Z'}{M_{Z}}&=&\frac{1}{\sqrt{\lambda_{S}}}\left(\frac{-Z}{M_{Z}}\right)^{\frac{1}{2}}\left[1+\lambda_{Z}+\frac{\lambda_{1}}{\lambda_{2}}\left(-2\lambda_{3}\frac{Z}{M_{Z}}+\eta_{1}\right)\right]^{\frac{1}{2}} \nonumber\\
\frac{Z}{M_{Z}}&=&\frac{-b\pm \sqrt{b^{2}-4ac}}{2a},
\end{eqnarray}
where
\begin{eqnarray}{\label{abc}}
a&=&\lambda_{1}+\frac{\lambda^{2}_{3}}{\lambda^{2}_{2}}\lambda_{S}-\frac{\lambda_{2}\lambda_{Z}}{\lambda_{S}}+2\frac{\lambda_{1}\lambda_{3}}{\lambda_{S}}\nonumber\\
b&=&-\frac{\lambda_{3}}{\lambda_{2}}-\frac{\lambda_{3}\lambda_{S}}{\lambda^{2}_{2}}\eta-\frac{\lambda_{2}}{\lambda_{S}}-\frac{\lambda_{1}}{\lambda_{S}}\eta\nonumber\\
c&=&\frac{1}{2\lambda_{2}}\eta_{1}\eta_{2}+\frac{\lambda_{S}}{4\lambda^{2}_{2}}\eta_{1}^{2}+\lambda_{4}\left(\frac{S'}{M_{Z}}\right)^{2}
\end{eqnarray}
and $\eta_{1}=M_{Z'}/M_{Z}$, $\eta_{2}=M_{S}/M_{Z}$, $\eta_{3}=M_{S'}/M_{Z}$.
A few comments are in order regarding the parity assignments.
Firstly, the $Z_2$ parity eliminates the unsafe operator in Eq.(\ref{unsafe5}).
Secondly, without $Z'_2$ large $\mu$ term for the doublets in $5$ and $\bar{5}$
would be induced by the mixings with $45$ and $\bar{45}$.
Instead, imposing $Z'_2$ forbids the operator $Z' H(5)\bar{H}(\bar{45})$,
which then keeps the doublets in $5$ and $\bar{5}$ light.
Finally, due to the last line in Eq.(\ref{SSBM5}) which is consistent with $Z_{2}\times Z'_{2}$,
the effective operator $\bar{H}(\bar{5})H(45)Z'^{2}/M_{Z}$ is produced after integrating $Z$.
Thus, the effective superpotential for the doublets in $45$ and $\bar{45}$ at the leading order is given by,
\begin{eqnarray}{\label{eff1}}
W_{\text{eff}}\sim \left(M_{45}+\lambda_{5}\langle S\rangle\right)\sigma_{u}\sigma_{d}+ \lambda_{3}H_{d}\sigma_{u}Z'^{2}/M_{Z},
\end{eqnarray}
where corrections due to those mixings among singlets of $Z$, $Z'$ and $S$ have been neglected.
From Eq.(\ref{eff1}) we obtain the vev
\begin{eqnarray}{\label{v45}}
\sigma_{d}\sim \lambda_{3}\frac{Z'^{2}}{(M_{45}+\lambda_{4}\langle S\rangle)M_{Z}}\upsilon_{d}.
\end{eqnarray}
Given $\lambda_{3}\sim 0.1$, $\upsilon_{d} \sim 10$ GeV,
we have $\sigma_{d}\sim 100$ MeV.
A rational Yukawa texture such as $Y_{45}\sim (0.01,0.1, 1)$ for the three generations then reproduces the mass relations in Eq.(\ref{app}).
\begin{table}
\tiny{
\begin{center}
\begin{tabular}{|c|c|c|c||c|c|c||c|c|c|c|c|c|c|}
\hline
Field & \ $N_{R}$&\ $\psi$ & \ $\Phi$ & \ $5$ & \ $\bar{5}$ & \ $\bar{45}$ & \ $1(S)$& \ $1(S')$& \ $75(Z)$ &\ $75(Z')$ & $45$ & \ $50$ & $\bar{50}$ \\
\hline
$Z_{2}$ &+ & + & - & + & - & - & + & + & + & - & - & - & - \\
$Z'_{2}$ &+ & + & + & + & + & + & + & - & + & - & + & + & - \\
\hline
\end{tabular}
\caption{$Z_{2}\times Z'_{2}$ parity assignments in the case of embedding MSSM into $\rm{SU}(5)$,
which are consistent with the superpotentials in Eq.(\ref{Yukawa5}) and Eq.(\ref{SSBM5}). }
\label{MSSMc}
\end{center}}
\end{table}
\subsection{\rm{SO}(10)}
Unlike the case of $\rm{SU(5)}$,
the input parameters in the $\rm{SO(10)}$ which control the SM fermions masses $M_{u,d,e}$
and neutrino masses $m_{\nu}$ are tied to each other.
The main reason for this is that the MSSM matters of each generation are contained in a single $16(\phi)$.
Here, we give a brief review on the embedding of MSSM into $\rm{SO}(10)$.
According to the product $16\times 16= 10_{s}+120_{a}+126_{s}$,
the Higgs representation responsible for SM fermion masses can be composed of $10$, $120$ and $\bar{126}$.
By following previous discussions on $\rm{SU}(5)$ \cite{GN1, GN2}, the simplest choice is to introduce $10$ and $\bar{126}$ \cite{9209215}, with Yukawa superpotential given by,
\begin{eqnarray}{\label{Yukawa10}}
W^{\rm{SO}(10)}_{Y}=\phi_{i}(16)[Y^{ij}_{10}H(10)+Y^{ij}_{126}\bar{H}(\bar{126})]\phi_{j}(16),
\end{eqnarray}
where matrixes $Y_{10}$ and $Y_{126}$ are both symmetric.
From Eq.(\ref{Yukawa10}) we have
\begin{eqnarray}{\label{massmatrix2}}
M_{u}&=&Y_{10}^{ij}\upsilon^{10}_{u}+Y_{126}^{ij}\upsilon^{\bar{126}}_{u}, \nonumber\\
M_{d}&=&Y_{10}^{ij}\upsilon^{10}_{d}+Y_{126}^{ij}\upsilon^{\bar{126}}_{d}, \nonumber\\
M_{e}&=&Y_{10}^{ij}\upsilon^{10}_{d}-3Y_{126}^{ij}\upsilon^{\bar{126}}_{d},
\end{eqnarray}
for SM quarks and electrons, and
\begin{eqnarray}{\label{massmatrix3}}
M_{N_{R}}&=&Y_{126}^{ij}\upsilon^{\bar{126}}_{s} \nonumber\\
M_{\nu}&=&Y_{10}^{ij}\upsilon^{10}_{u}-3Y_{126}^{ij}\upsilon^{\bar{126}}_{d},\nonumber\\
M_{L}&=&Y_{126}^{ij}\upsilon^{\bar{126}}_{L}
\end{eqnarray}
for neutrinos,
where $\upsilon^{10}_{u,d}$ refer to the doublet vevs in $H(10)$,
and $\upsilon^{\bar{126}}_{s, u, d, L}$ denote the singelt (s), doublet (u, d) and triplet (L) vevs in $\bar{H}(\bar{126})$, respectively.
Therefore, the neutrino mass arises from both type-I \cite{typeI1,typeI2,typeI3,typeI4} and type-II \cite{typeII1,typeII2,typeII3} contributions,
\begin{eqnarray}{\label{seesaw}}
m_{\nu}=M_{L}-M^{T}_{\nu}M_{N_{R}}^{-1}M_{\nu}.
\end{eqnarray}
Benchmark solutions to the input parameters in Eq.(\ref{massmatrix2}) and Eq.(\ref{massmatrix3})
have been extensively studied \cite{9209215, SO10fit1, SO10fit2, SO10fit3, SO10fit4},
which demonstrate that the fit to SM flavor masses and mixings is viable.
But the construction of $W_{\text{SB}}$ responsible for the breaking of $\rm{SO}(10)\rightarrow G_{\text{SM}}$ is challenging \cite{SO10Mp,SO10Mp1,SO10Mp2, SO10Mp3},
since some of triplet fields in $126$ and $\bar{126}$ obtain mass of order $\upsilon^{\bar{126}}_{s}\sim 10^{13}$ GeV,
which spoils the perturbative unification.
Attempts to solve this issue involve adding $54$ \cite{1805.10631} to $W_{\text{SB}}$
or adding $120$ \cite{1805.05776} to $W_{Y}$ in Eq.(\ref{Yukawa10}).
We employ the solution of modifying $W_{\text{SB}}$ \cite{1805.10631},
where the Higgs fields are composed of $210(Y)$, $H(126)$, $\bar{H}(\bar{126})$ and $54(X)$,
and $W_{\text{SB}}$ takes the form\footnote{We follow the notation in \cite{0405300} for Yukawa coupling constants.}
\begin{eqnarray}{\label{SSBM10}}
W^{\text{SO}(10)}_{\text{SB}}&=& \frac{M_{Y}}{2} Y^{2} +\frac{M_{X}}{2}X^{2}+ M_{126} H(126)\bar{H}(\bar{126})\nonumber\\
&+&\lambda_{1} Y^{3}+\lambda_{2} Y H(126)\bar{H}(\bar{126}) \nonumber\\
&+& Y\left[\lambda_{3}H(126)H(10)+\lambda_{4}\bar{H}(\bar{126})H(10)\right]\nonumber\\
&+&X\left[\lambda_{8}X^{2}+\lambda_{10}Y^{2}+\lambda_{11}H^{2}(126)\right.\ \nonumber\\
&+&\left.\ \lambda_{12}\bar{H}^{2}(\bar{126})+ \lambda_{13}H^{2}(10)\right]+ \text{H.c}.
\end{eqnarray}
Here are a few comments about $W_{\text{SB}}$ in order.
$1)$. Under the notation of $SU(4)\times SU(2)_{L}\times SU(2)_{R}$
the SM singlet vevs $Y_{1}=\langle(1,1,1)_{Y}\rangle$, $Y_{2}=\langle(15,1,1)_{Y}\rangle$ and
$Y_{3}=\langle(15,1,3)_{Y}\rangle$ in $Y$ and $X=\langle(1,1,1)_{X}\rangle$ in $X$
are responsible for the breaking $\rm{SO}(10)\rightarrow G_{\text{SM}}\times U(1)_{B-L}$.
$2)$. The SM singlet vevs $\upsilon^{126}_{s}=\langle(\overline{10},1,3)_{126}\rangle$ and
$\upsilon^{\bar{126}}_{s}=\langle(10,1,3)_{\bar{126}}\rangle$
result in the breaking of $G_{\text{SM}}\times U(1)_{B-L}\rightarrow G_{\text{SM}}$.
In the limit $\upsilon^{126}_{s}=\upsilon^{\bar{126}}_{s}<<Y_{i}, X$ the vevs are given by \cite{1805.10631},
\begin{eqnarray}{\label{vevs3}}
\frac{Y_1}{Y_2}&=&\frac{1}{2\sqrt{3}}\eta^2,\nonumber\\
\frac{M_Y}{Y_2}&=&-\frac{\lambda_1}{5\sqrt{2}}\left(3+\eta^2\right),\nonumber\\
\frac{X}{Y_2}&=&\frac{\lambda_1}{\lambda_{10}\sqrt{30}}\left(-2+\eta^2\right),\nonumber\\
\frac{M_{126}}{Y_2}&=&-\frac{\lambda_2}{120}\left(6\sqrt{2}+12\eta+\sqrt{2}\eta^2\right),\nonumber\\
\frac{M_X}{Y_2}&=&\frac{-1}{20\sqrt{2}\lambda_1\lambda_{10}}\left[\left(5\lambda_{10}^3+2\lambda_{1}^{2}\lambda_{8} \right)
\eta^2+20\lambda_{10}^3-4\lambda_{1}^{2}\lambda_{8}\right]\nonumber\\
\end{eqnarray}
with $\eta=Y_{3}/Y_2$.
\section{NMSSM}
With the embedding of MSSM into realistic SUSY GUTs as a benchmark,
in this section we analyze the NMSSM.
According to the starting points in the Introduction,
a viable embedding should satisfy two constraints:
\begin{itemize}
\item The mass of $N$ should be of order TeV scale.
\item The vev of $N$ should be of order TeV scale.
\end{itemize}
Both of them may be spoiled by a few dangerous mixings between $N$ and Higgs fields which contain singlet vev of order GUT scale.
The key point is whether there is suitable symmetry to avoid such mixings.
\subsection{\rm{SU}(5)}
In this situation, $W_Y$ in Eq.(\ref{Yukawa5}) should be extended to include $NH_{u}H_{d}+\frac{\kappa}{3}N^{3}+\text{H.c}$ in the NNSSM,
which means that\footnote{Operator $NH(5)\bar{H}(\bar{5})$ contributes to Yukawa interaction $NH_{c}\bar{H}_{c}$ beyond MSSM,
with $H_c$ and $\bar{H}_c$ being the color-triplet Higgs fields.
However, it does not affect proton decay at all,
as the singlet $N$ mass is always far larger than proton mass.},
\begin{eqnarray}{\label{NMSSM5}}
\delta W_{Y}=\lambda NH(5)\bar{H}(\bar{5})+\frac{\kappa}{3}N^{3}.
\end{eqnarray}
Eq.(\ref{NMSSM5}) do not affect the fit to SM flavor masses and mixings in Sec.IIA.
Nevertheless, compared to MSSM, $W_{\text{SB}}$ is allowed to contain superpotential terms
\begin{eqnarray}{\label{SSBNM5}}
W_{\text{unsafe}}= N Z^{2}+NZ'^{2}+NS^{2}+N^{2}S.
\end{eqnarray}
These new terms in Eq.(\ref{SSBNM5}) yield corrections to the $F$-terms of $Z$, $Z'$, $S$
such as $F_{S}=F^{\text{MSSM}}_{S}+2\lambda_{N}SN$,
which can be adjusted to the case of MSSM by e.g. $\left<N\right>=0$.
Even so, the singlet vevs $\left<Z\right>$, $\left<Z'\right>$ and $\left<S\right>$ still lead to either large $N$ mass or large mixing.
In order to avoid all of mixing terms in Eq.(\ref{SSBNM5}), we need to impose new parity.
The first observation is that an odd $N$ under a $Z_2$ parity as shown in Table.\ref{MSSMc} excludes the first three terms in Eq.(\ref{SSBNM5}).
But the last term therein still remains \footnote{A economic solution to keep light $N$
is adding another singlet $S'=1$ which is even under the $Z_2$.
With such $S'$ the $W_{\text{SB}}$ is further extended by,
\begin{eqnarray}{\label{S'}}
\delta W_{\text{SB}}(S')=\frac{1}{2}M_{S'}S'^{2}+m_{SS'}SS'+N^{2}S'.
\end{eqnarray}
from which $F_{S'}=M_{S'}S'+m_{SS'}S+N^{2}$,
and $\langle S'\rangle=-\langle S\rangle$ if $M_{S'}=m_{SS'}$.
The two different contributions to $N$ mass cancel each other, leaving us a light $N$.
Unfortunately, neither Abelian or $Z_N$ parity can ensure $M_{S'}=m_{SS'}$.}.
Similar result holds for $Z_N$ or an Abelian symmetry.
\subsection{\rm{SO}(10)}
Similar to the embedding of NMSSM into $\rm{SU}(5)$,
$W_Y$ in Eq.(\ref{Yukawa10}) is modified by,
\begin{eqnarray}{\label{NMSSM52}}
\delta W_{Y}=\lambda NH^{2}(10)+\frac{\kappa}{3}N^{3}.
\end{eqnarray}
Instead of Eq.(\ref{SSBNM5}), $W_\text{SB}$ in Eq.(\ref{SSBM10}) is allowed by gauge invariance to contain
\begin{eqnarray}{\label{NMSSM102}}
W_{\text{unsafe}}&=&N Y^{2}+NH(126)\bar{H}(\bar{126})+N X^{2}+XH^{2}(10),\nonumber\\
\end{eqnarray}
in which $N$ mixes with the SM singlets of $Y$, $126$, $\bar{126}$ and $X$.
Thus, all of Yukawa couplings in Eq.(\ref{NMSSM102}) have to be extremely small.
What kind of parity allows Yukawa superpotential in Eq.(\ref{NMSSM52})
but eliminates that in Eq.(\ref{NMSSM102}) simultaneously?
The first observation is that a $Z_2$ parity does not work.
Since Eq.(\ref{NMSSM102}) would imply $N$ an odd field,
which contradicts with the Yukawa superpotential in Eq.(\ref{NMSSM52}).
Similar results hold for any $Z_N$ parity.
Because the rational assignment $n_{Y}=0$ as required by successful symmetry breaking,
demonstrates that in order to allow the Yukawa superpotential in Eq.(\ref{NMSSM52}), $n_{10}=n_{\bar{126}}=n_{126}=N/2$.
Accordingly, $n_{N}=0$ from $N10(H)10(H)$, which implies that some of terms in Eq.(\ref{NMSSM102}) are still allowed.
To conclude, in our setup embedding NMSSM into the minimal $\rm{SO}(10)$ is not viable.\\
\section{VMSSM}
Let us proceed to discuss the embedding of VMSSM into SUSY $\rm{SU(5)}$ and $\rm{SO}(10)$.
The VL fermions with mass of order TeV scale can be composed of $5$ with $\bar{5}$, $10$ with $\bar{10}$ in the $\rm{SU}(5)$
or $16$ with $\bar{16}$ in the $\rm{SO}(10)$ respectively \cite{Zheng1, Zheng2}.
A realistic embedding should satisfy the following constraints.
\begin{itemize}
\item The vev of Higgs field $\rho$ responsible for the VL fermion masses should be of order TeV scale.
\item The VL fermions are prevented from directly coupling to the Higgs fields triggering high-scale gauge symmetry breaking.
\end{itemize}
Violating the first constraint is likely to occur because
either $\rho=\{1,24, 75\}$ or $\rho=\{1, 45,210\}$
may directly couple to $S$,$S'$ $Z$, $Z'$ in the case of $\rm{SU}(5)$ or $X$ and $Y$ in the case of $\rm{SO}(10)$, respectively,
which tends to yield $\langle\rho\rangle$ of GUT scale.
In contrast, $\langle\rho\rangle$ of order TeV scale can be only realized by the effective operator such as
\begin{eqnarray}{\label{vvev1}}
W_{\text{eff}}(\rho)=\frac{M_{\rho}}{2}\rho^{2}+\frac{\rho\cdot A\cdot B\cdots}{M^{n}_{\text{U}}}
\end{eqnarray}
where $A$, $B$, $\cdots$ refer to $H(5)$, $\bar{H}(\bar{5})$, $S$, $S'$, $Z$, $Z'$ in the $\rm{SU}(5)$,
or $H(126)$, $\bar{H}(\bar{126})$, $X$ and $Y$ in the $\rm{SO}(10)$, respectively,
with $M_{U}$ denoting the GUT scale.
\subsection{\rm{SU}(5)}
For the VL fermions of $5(\Sigma)$ and $\bar{5}(\bar{\Sigma})$,
$W_Y$ in Eq.(\ref{Yukawa5}) is extended by,
\begin{eqnarray}{\label{VMSSM5}}
\delta W_{Y}=\rho\bar{\Sigma}(\bar{5})\Sigma(5)+\text{H.c},
\end{eqnarray}
where $\rho=\{1,24\}$ of $\rm{SU}(5)$.
The reason for adding $\rho$ is that either singlet vev $S$ or $S'$ in the Sec.IIA is too large to provide a VL mass of order TeV.
In this case, the unsafe superpotential at least includes
\begin{eqnarray}{\label{VMSSMu}}
W_{\text{unsafe}}\supset S\Sigma\bar{\Sigma},
\end{eqnarray}
which can be excluded by imposing the $Z_2$ parity assignments as shown in Table.\ref{VMSSMc1}.
$W_{\text{unsafe}}$ in Eq.(\ref{VMSSMu}) can also contain the following terms depending on $\rho$,
\begin{eqnarray}{\label{VMSSMu2}}
W_{\text{unsafe}}=
\left\{
\begin{array}{lcl}
\rho Z^{2}+\rho Z'^{2}+S\rho Z'+\rho^{2}S, ~\rho=24\\
\rho Z^{2}+\rho Z'^{2}+\rho^{2}S+\rho S^{2}, ~~~\rho=1\\
\end{array} \right.
\end{eqnarray}
\begin{table}
\tiny{
\begin{center}
\begin{tabular}{|c|c|c|c|c|c||c|c|c|c||c|c|c|c|c|c|c|c|}
\hline
Field & $N_{R}$&\ $\psi$ & \ $\Phi$ & $\Sigma$ & $\bar{\Sigma}$ & \ $5$ & \ $\bar{5}$ & \ $\bar{45}$ & $\rho$ & \ $1(S)$ & $1(S')$ &\ $75(Z)$ & $75(Z')$ & $45$ & \ $50$ & $\bar{50}$ \\
\hline
$Z_{2}$ &+ & + & - & + & - & + & - & - & - & + & + & + & - & -& - & - \\
\hline
\end{tabular}
\caption{$Z_{2}$ parity assignments for embedding NMSSM into $\rm{SU}(5)$,
which is consistent with the superpotential in Eq.(\ref{VMSSM5}).
Yet, it is unable to exclude all the unsafe structures in Eq.(\ref{VMSSMu2}). }
\label{VMSSMc1}
\end{center}}
\end{table}
Besides the unsafe operators in Eq.(\ref{VMSSMu2}),
there are also no suitable Feynman graphs to generate the desired effective operator with correct mass order in Eq.(\ref{vvev1}).
In principle, the form of effective operator in Eq.(\ref{vvev1}) can be divided as follows:
\begin{itemize}
\item It is of type $\rho AB\cdots /M^{2}_{\text{U}}$, with $A$ and $B$ referring to $S$, $Z$ or $Z'$.
In this situation, one obtains $\langle\rho\rangle \sim$ GUT scale given all of $A$, $B$, $\cdots$ are of order GUT scale.
\item It looks like $\rho AB\cdots H(5)/M_{\text{U}}$ or $\rho AB\cdots\bar{H}(\bar{5})/M_{\text{U}}$,
which contain only a doublet vev.
Such operator contributes to $\langle\rho\rangle\sim$ TeV for $M_{\rho}\sim$ GUT scale.
Unfortunately, there is no suitable intermediator.
\item It is of form $\rho AB\cdots H(5) \bar{H}(\bar{5})\cdots/M^{2}_{\text{U}}$ or $\rho AB\cdots \bar{H}(\bar{5})H(45)\cdots/M^{2}_{\text{U}}$,
where there are at least two doublet vevs.
In this case, $\langle\rho\rangle$ is less than $\mathcal{O}(1)$ eV.
\end{itemize}
In summary, since there is no appropriate vev scale, the embedding of VMSSM into $\rm{SU}(5)$ is not viable.
Similar result can be found for VL $10$ and $\bar{10}$, where $\rho=\{1,24, 75\}$ of $\rm{SU}(5)$.
\\
\\
\subsection{\rm{SO}(10)}
As shown in \cite{Zheng2},
the non-minimal extension through the $16$-dimensional VL fermions remains consistent with the $\rm{SO}(10)$ unification.
In this model, $W_Y$ is modified by,
\begin{eqnarray}{\label{VMSSM10}}
\delta W_{Y}=\rho \Delta(16)\bar{\Delta}(\bar{16}),
\end{eqnarray}
where $\rho=\{1, 45, 210 \}$ of $\rm{SO}(10)$.
Similar to Sec.IVA,
the main challenge to the embedding is the generation of singlet vev $\langle\rho\rangle\sim$ TeV scale for a $\rho$ mass of GUT scale.
What differs from the discussions in Sec.IVA is the existence of the intermediate mass scales
$\upsilon^{126}_{s}$ and $\upsilon^{\bar{126}}_{s}$ in the $\rm{SO}(10)$,
which is critical to solve the problem.
\begin{figure}
\centering
\includegraphics[width=5cm,height=4cm]{vmssm.eps}
\centering
\caption{Super Feynman graph for the generation of higher-dimensional effective operator in the case of $\rho=210$.}
\label{vmssm}
\end{figure}
We firstly consider $\rho=210$.
We add Higgs fields $54(V)$ and $54(U)$ to $W_{\text{SB}}$,
with the $Z_{2}$ parity assignments as shown in Table.\ref{VMSSMc2}.
The $Z_2$ parity excludes the unsafe operator
\begin{eqnarray}{\label{unsafev}}
W_{\text{unsafe}}=Y\Delta(16)\bar{\Delta}(\bar{16}),
\end{eqnarray}
and simultaneously allows Yukawa interactions
\begin{eqnarray}{\label{SSBM10V1}}
\delta W_{\text{SB}}&=&\frac{M_{\rho}}{2}\rho^{2}+\frac{M_{V}}{2}V^{2}+\frac{M_{U}}{2}U^{2}\nonumber\\
&+&\lambda'_{10}\rho UY+\lambda''_{10}\rho VY \nonumber\\
&+&\lambda'_{8}UVX+\lambda''_{8}U^{2}X+\lambda'''_{8}V^{2}X.
\end{eqnarray}
In terms of Eq.(\ref{SSBM10V1}) and Eq.(\ref{SSBM10}) the effective operator for $\rho$ is given by,
\begin{eqnarray}{\label{operator}}
W_{\text{eff}}&\simeq& \frac{M_{\rho}}{2}\rho^{2}+ \rho(\lambda'_{10}U+\lambda''_{10}V)\left[\frac{H(126)\bar{H}(\bar{126})}{M_{Y}}+\cdots\right]\nonumber\\
&+&\rho\frac{\lambda'_{10} U H^{3}(126)\bar{H}(\bar{126})}{M_{Y}M_{X}M_{V}}+\rho\frac{\lambda''_{10} V H^{3}(126)\bar{H}(\bar{126})}{M_{Y}M_{X}M_{U}}\nonumber\\ &+& \mathcal{O}(\rho^{3})
\end{eqnarray}
For calculating an effective superpotential in the infrared region from those in the ultraviolet region,
integrating out heavy chiral superfields in the Feynman graphs is equivalent to solving the nonlinear equations of $F$ terms related to these heavy chiral superfileds.
The leading-order terms with coefficient $\lambda'_{10}U+\lambda''_{10}V=\delta F_{Y}$ in Eq.(\ref{operator}) are obtained after integrating out superfield $Y$ for $\lambda_1$ and $\lambda_{10}$ less than unity.
Similarly, the next-leading order operators therein are induced by further integrating out $X$,
referring to which the Feynman graph is shown in Fig.\ref{vmssm}.
Note, we have used the mass term in Eq.(\ref{operator}) representing those quadratic terms,
and neglected the higher-order terms.
Apart from the $F$-term contributions in Eq.(\ref{operator}),
the potential for singlet component in $\rho$ also contains soft SUSY-breaking terms such as $A_{\rho}\rho^{3}+\text{H.c}$.
The scale of $A_{\rho}$ depends on the details of SUSY breaking.
It can be neglected in gauge mediation, but be of order $\sim 1$ TeV in the other scenarios.
In the later case, it is easily to verify that the $F$-term contribution still dominates over those soft terms.
In what follows, we simply ignore those soft terms for the estimate of $\langle\rho\rangle$.
The leading-order contribution in Eq.(\ref{operator}) has to be eliminated if one wants to obtain singlet vev $\langle\rho\rangle\sim$ TeV scale.
Alternatively, the coefficient $\delta F_{Y}$ must be suppressed without much fine tuning.
There is a dynamical realization for this purpose.
From Eq.(\ref{SSBM10V1}),
one finds that the SUSY vacuum described by the vevs in Eq.(\ref{vevs3}) remains
only if the following constraints
\begin{eqnarray}{\label{vevs4}}
0&=&\lambda'_{8}UV+\lambda''_{8}U^{2}+\lambda'''_{8}V^{2},\nonumber\\
0&=&\lambda'_{10}U+\lambda''_{10}V,\nonumber\\
0&=&M_{V}V+\frac{X}{2\sqrt{15}}\left(\lambda'_{8}U+2\lambda'''_{8}V\right),\nonumber\\
0&=&M_{U}U+\frac{X}{2\sqrt{15}}\left(\lambda'_{8}V+2\lambda''_{8}U\right),
\end{eqnarray}
are satisfied.
Given GUT-scale mass parameters $M_U$, $M_V$, $\langle U\rangle$ and $\langle V\rangle$,
there are indeed rational solutions to Eq.(\ref{vevs4}),
under which the leading-order operator in Eq.(\ref{operator}) vanishes due to $\delta F_{Y}=0$ (the second formula in Eq.(\ref{vevs4})).
Therefore, the next leading-order contribution in Eq.(\ref{operator}) dominates
the effective superpotential for $\rho $ below GUT scale as long as $M_{U}\neq M_{V}$,
\begin{eqnarray}{\label{vvev3}}
W_{\text{eff}}\sim M_{\rho}\rho^{2}+\rho \frac{V H^{3}(126)\bar{H}(\bar{126})}{M^{3}_{U}},
\end{eqnarray}
which contributes to a nonzero vev:
\begin{eqnarray}{\label{rho}}
\langle\rho\rangle \sim \frac{\langle V\rangle}{M_{\rho}} \frac{\left(\upsilon^{126}_{s}\right)^{4}}{M^{3}_{U}}.
\end{eqnarray}
Given singelt vev $\upsilon^{126}_{s}\simeq \upsilon^{\bar{126}}_{s}\sim 10^{13}$ GeV fixed by fit to SM flavor masses \cite{1805.10631},
and $M_{\rho}\sim M_{U}\sim \langle V\rangle \sim 10^{16}$ GeV,
we have $\langle\rho\rangle \sim 1-10$ TeV.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c||c|c|c||c|c|c|c|c|c|}
\hline
Field & \ $\phi$&\ $\Delta$ & \ $\bar{\Delta}$ & \ $10$ & $\bar{126}$ & $\rho$ & \ $X$ & \ $126$&\ $54(V)$ & $54(U)$ & $Y$ \\
\hline
$Z_{2}$ & + & + & - & + & + & - & + & + & - & - & + \\
\hline
\end{tabular}
\caption{$Z_{2}$ parity assignments in the VMSSM model with $\rho=210$,
which are consistent with the superpotentials in Eq.(\ref{Yukawa10}),Eq.(\ref{SSBM10}), Eq.(\ref{VMSSM10}) and Eq.(\ref{SSBM10V1}),
and simultaneously avoid the unsafe operator in Eq.(\ref{unsafev}). }
\label{VMSSMc2}
\end{center}
\end{table}
The analysis for $\rho=210$ can shed light on other cases such as $\rho=\{1,45\}$.
For $\rho=1$ we can naively choose $Z_2$ odd fields $U=V=210$ in Fig.\ref{vmssm}.
However, an unsafe operator $V(U)\Delta(16)\bar{\Delta}(\bar{16})$ appears again.
For $\rho=45$, one may choose $U=45$ and $V=45$ or $54$,
which is unfavored by an operator similar to $\rho=1$.
\section{Conclusion}
According to the observed Higgs mass at the LHC and the dark matter direct detection limits,
the conventional MSSM - the simplest natural SUSY that is consistent with unification - is under more pressure than ever.
Such stress can be greatly relaxed in the extended MSSM models such as NMSSM and VMSSM which retain the unification and are still simple.
In this paper, following the assumptions that $W_Y$ is fixed by the SM matter content and its TeV-scale extension,
and they receive their masses from $W_{\text{SB}}$ through Higgs mechanism,
we have studied the embeddings of these three models into SUSY GUTs.
First of all, we discussed the MSSM, where the realistic $\rm{SU}(5)$ and $\rm{SO}(10)$ realizations
serve as benchmark solutions to the SM flavor issue and neutrino masses.
Then, we utilize the benchmark MSSM as guidance to the embedding of NMSSM and VMSSM.
We have found that the embedding of NMSSM is not viable
due to a large mount of mixings between the singlet $N$ and the Higgs fields responsible for the GUT symmetry breaking.
But the problem can be evaded in the VMSSM,
because the Higgs field $\rho$ which provides $16$-dimensional VL mass of order TeV scale
can avoid the same problems the singlet $N$ encounters due to the intermediate mass scale in the $126$-dimensional Higgs.
{\it Acknowledgments}.
The author thanks the anonymous referee for useful suggestions.
This work is supported by the National Natural Science Foundation of China under Grant No.11775039
and the Fundamental Research Funds for the Central Universities with project No. cqu2017hbrc1B05 at Chongqing University.
|
\section{Introduction}
Transfer learning or domain adaption aims at digging potential information in auxiliary source domain to assist the learning task in target domain, where bare labeled data with prior knowledge exist \cite{Survey}. Without the help of related source domain data, the learning tasks like image classification or recognition would fail with insufficient pre-existing labeled data. For most big data problems, the labeled data are highly required but always not enough as labeling process would be quite tedious and laborious. Therefore, having a better use of auxiliary source domain data by transfer learning methods has attracted researchers' attention.
It should be noted direct application of labeled source domain data to a new scene of target domain would result in poor performance due to the semantic gap between the two domains, even they are representing the same objects \cite{Heterogeneous2011zhu}\cite{Learning2012Lixin}. The semantic gap can be resulted from different acquisition conditions(illumination or view angle) and the use of different cameras or sensors. Transfer learning methods are proposed to overcome this distribution divergence or feature bias \cite{Translated2009Dai}\cite{Liu2017IJCAI}\cite{Balanced2017Jingdong}. Traditionally, these transfer learning methods would adopt linear or non-linear transformation with kernel function to learn a common subspace on which the gap is bridged \cite{Learning2017Yugang}. Recent advancement has proven that the features learnt on such common subspace are inefficient. Therefore, deep learning based model has been introduced due to its power on high level feature representation.
Current deep learning based transfer learning topics include two research branches, what knowledge to transfer and how to transfer knowledge \cite{Gopalan2015visual}. For what knowledge to transfer, researchers mainly concentrate on instance-based transfer learning and parameter transfer approaches. Instance-based transfer learning methods assume that only certain parts of the source data can be reused for learning in the target domain by re-weighting \cite{Liu2016ICML}. As for parameter transfer approaches, people mainly try to find the pivot parameters in deep network to transfer to accelerate the transfer process. For how to transfer knowledge, different deep networks are introduced to complete the transfer learning process. However, for both research areas, the right correspondence of layers is ignored.
For what knowledge to transfer problem, the transferred content might even be negative or wrong. A fundamental problem for current transfer learning work should be negative transfer\cite{Distant2017Tan}. If the knowledge from the source domain to target domain is transferred to wrong layers, the transferred knowledge is quite error-prone. With the wrong prior information added, bad effect can be generated on target domain data. For how to transfer knowledge problem, as the two deep networks for the source domain data and the target domain data need to have the same number of layers, the two models could not be optimal at the same time. This situation is especially important for cross-resolution heterogeneous transfer. For data with different resolutions, The data with higher resolution might need more max-pooling layers than the data with lower resolution, and more neural network layers are needed. Based on the above observation and assumption, we propose a novel research topic, where to transfer. In this work, the number of layers for two domains does not need to be the same, and optimal matching of layers will be found by the newly proposed objective function. With the best parameters from the source domain data transferred to the right layer of the target domain, the performance of the target domain learning task can be improved.
The proposed work is named Deep Transfer Learning by Exploring where to Transfer(DT-LET), which is based on Stacked Auto-Encoders\cite{Auto2018Zhuang}. A detailed flowchart is shown in Fig. \ref{fig:1}. The main contributions are concluded as follows.
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{figure_floww.pdf}
\caption{The flowchart of the proposed DT-LET framework. The two neural networks are first trained by the co-occurrence data $C_s$ and $C_t$. It is found the best layer matching to match the corresponding layers by CCA. After network training, the common subspace is found and the training data $D_S^l$ is transferred to such space to train SVM classifier, to classify $D_T$.}
\label{fig:1}
\end{figure}
\begin{itemize}
\item This paper for the first time introduces the where to transfer problem. The deep networks from the source domain and the target domain no longer need to be with the same parameter settings, and the cross-layer transfer learning is proposed in this paper.
\item We propose a new principle for finding the correspondence between neural networks in the source domain and in the target domain by defining new unified objctive loss function. By optimizing this objective function, the best setting of two deep networks as well as the correspondence relationship can be figured out.
\end{itemize}
\section{Related Work}
Deep learning intends to learn nonlinear representation of raw data to reveal the hidden features \cite{Deep2016long}. However, a large number of labeled data are required to avoid over-fitting during the feature learning process. To achieve this goal, transfer learning has been introduced to augment the data with prior knowledge. By aligning data from different domains to high-level correlation space, the data information on different domains can be shared. To find this correlation space, many deep transfer learning frameworks have been proposed in recent years. The main motivation is to bridge the semantic gap between the two deep neural networks of the source domain and the target domain. However, due to the complexity of transfer learning, some transfer mechanisms still lack satisfying interpreting. Based on this consideration, quite a few interesting ideas have been generated. To solve how to determine which domain to be source or target problem, Fabio et al. \cite{Autodial} propose to automatically align domains for the source and target domain. To boost the transfer efficiency and find extra profit during the transfer process, deep mutual learning \cite{deepmutual} has been proposed to transfer knowledge bidirectionally. The function of each layer in transfer learning is explored in \cite{CactusNets}. The transfer learning with unequal classes and data are experimented in \cite{optimal} and \cite{investigating} respectively. However, all the above works still just explain what knowledge to transfer and how to transfer knowledge problems. They still ignore interpreting the matching mechanisms between layers of deep networks of the source domain and the target domain. For this problem, We also name it as DT-LET: Deep Transfer Learning by Exploring Where to Transfer. For this work, we adopt stacked denoising autoencoder(SDA) as the baseline deep network for transfer learning.
Glorot et al. for the first time employ stacked denoising to learn homogeneous features based on joint space for sentiment classification \cite{glorot2010}. The computation complexity is further reduced by Chen et al. by the proposing of Marginalized Stacked Denoising Autoencoder(mSDA) \cite{Chen2012Marginalized}. In this work, some characteristics of word vector are set to zero in the equations of expectation to optimize the representation. Still by matching the marginal as well as conditional distribution, Zhang et al. and Zhuang et al. also develop SDA based homogeneous transfer learning framework \cite{Supervised2015zhuang}\cite{Deep2015zhang}. For heterogeneous case, Zhou et al. \cite{Hybrid2014zhou} propose an extension of mSDA to bridge the semantic gap by finding the cross-domain corresponding instances in advance. Google brain team in recent time introduces generative adversarial network to SDA and propose the Wasserstein Auto-Encoders \cite{Wasserstein} to generate samples of better quality on target domain. It can be found SDA is with quite high potential, and our work also chooses SDA as the basic neural network for the where to transfer problem.
\section{Deep Mapping Mechanism}
The general framework of such deep mapping mechanism can be summarized as three steps, network setting up, correlation maximization, and layer matching. We would like to first introduce the deep mapping mechanism by defining the variables.
The samples in the source domain are denoted as $D_S = \{I_i^s\}_{i=1}^{n_s}$, in which the labeled data in the source domain is further denoted as $D_S^l = \{X_i^s, Y_i^s\}_{i=1}^{n_l}$, they are used to supervise the classification process. In the target domain, the samples are demoted as $D_T = \{I_i^t\}_{i=1}^{n_t}$. The co-occurrence data \cite{Learning2016Yang}(the data in the source domain and the target domain belonging to the same classes but with no prior label information) in the source domain are denoted as $C_S = \{C_i^s\}_{i=1}^{n_c}$, in target domain are denoted as $C_T = \{C_i^t\}_{i=1}^{n_c}$. They are further jointly represented by $D^C = \{C_i^S, C_i^T\}_i^{n_c}$, which are used to supervise the transfer learning process. The parameters of deep network in the source domain are denoted by $\Theta^S = \{W^s, b^s\}$, and $\Theta^T = \{W^t, b^t\}$ in the target domain.
The matching of layers is denoted by ${R_{s,t}} = \{ r_{{i_1},{j_1}}^1,r_{{i_2},{j_2}}^2,...,r_{{i_a},{j_b}}^m\}$, in which a represents the total number of layers for the source domain data, and b represents the total layers for the target domain data. $m$ is the total number of matching layers. We define here, if ${\rm{m}} = \min \{ a - 1,b - 1\} $(as the first layer is the original layer which will not be used to transfer, m is compared with a-1 or b-1 instead of a or b), we define the transfer process as full rank transfer learning; else if ${\rm{m}} < \min \{ a - 1,b - 1\} $, we define this case as non-full rank transfer learning.
The common subspace is represented by $\Omega$ and the final classifier is represented by $\Psi$. The labeled data $D_S^l$ from the source domain are used to predict the label of $D_T$ by applying $\Psi(\Omega(D_T))$.
\subsection{Network setting up}
The stacked auto-encoder (SAE) is first employed in the source domain and the target domain to get the hidden feature representation $H^S$ and $H^T$ of original data as shown in eq. (\ref{eq:1}) and eq. (\ref{eq:2}).
\begin{equation}
\begin{aligned}
H^S(n+1) = f(W^S(n)\times H^S(n)+b^S(n)), n>1; \\
H^S(n) = f(W^S(n)\times C_S+b^S(n)), n=1.\label{eq:1}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
H^T(n+1) = f(W^T(n)\times H^T(n)+b^T(n)), n>1; \\
H^T(n) = f(W^T(n)\times C_T+b^T(n)), n=1.\label{eq:2}
\end{aligned}
\end{equation}
Here $W^S$ and $b^S$ are parameters from neural network $\Theta^S$, $W^T$ and $b^T$ are parameters from neural network $\Theta^T$. $H^S(n)$ and $H^T(n)$ mean the $n$th hidden layers in the source domain and in the target domain respectively. The two neural networks are first initialized by above functions.
\subsection{Correlation maximization}
To set up the initial relationship of the two neural networks, we resort to Canonical Correlation Analysis (CCA) which can maximize the correlation between two domains \cite{Canonical2004hardoon}. A multi-layer correlation model based on the above deep networks is further constructed. Both the $C_S$ and the $C_T$ are projected by CCA to a common subspace $\Omega$ on which a uniformed representation is generated. Such projection matrices obtained by CCA are denoted as $V^S(n)$ and $V^T(n)$. To find optimal neural networks in the source domain and in the target domain, we have two general objectives: to minimize the reconstruction error of neural networks of the source domain and the target domain, and to maximize the correlation between the two neural networks. To achieve the second objective, we need further on one hand find the best layer matching, on the other hand maximize the correlation between corresponding layers. To achieve this goal, we can minimize the final objective function
\begin{equation}
L({R_{s,t}}) = \frac{{{L_s}({\theta ^S}) + {L_T}({\theta ^T})}}{{P({V^S},{V^T})}}, \label{eq:3}
\end{equation}
in this function, the objective function is defined as L, and $L({R_{s,t}})$ is in corresponding with different matching of $R_{s,t}$. We would like to generate the best matching by finding the minimum $L({R_{s,t}})$. In $L({R_{s,t}})$, ${L_s}({\Theta^S})$ and ${L_T}({\Theta^T})$ represent the reconstruction errors of data in the source domain and the target domain, which are defined as follows:
\begin{equation}
\begin{aligned}
&{L_S}({\theta ^S}) = [\frac{1}{n_s}\sum\limits_{i = 1}^{n_s} {(\frac{1}{2}||{h_{{W^S},{b^S}}}(C_i^s) - X_i^s|{|^2})} ]\\
& + \frac{\lambda }{2}\sum\limits_{l = 1}^{{n^S} - 1} {\sum\limits_{j = 1}^{n_l^S} {\sum\limits_{k = 1}^{n_{l + 1}^S} {{{(W_{kj}^{S(l)})}^2}} } } \label{eq:4}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
&{L_T}({\theta ^T}) = [\frac{1}{n_t}\sum\limits_{i = 1}^{m_t} {(\frac{1}{2}||{h_{{W^T},{b^T}}}(C_i^t) - C_i^t|{|^2})} ]\\
& + \frac{\lambda }{2}\sum\limits_{l = 1}^{{n^T} - 1} {\sum\limits_{j = 1}^{n_l^T} {\sum\limits_{k = 1}^{n_{l + 1}^T} {{{(W_{kj}^{T(l)})}^2}} } },\label{eq:5}
\end{aligned}
\end{equation}
The third term $P({V_s},{V_t})$ represents the domain divergence after projection by CCA which we want to maximize. The definition for this term is in eq. (\ref{eq:6})
\begin{equation}
P({V^S},{V^T}) = \sum\limits_{l = 2}^{{n^S} - 1} {\frac{{{V^{S{{(l)}^T}}}\sum\nolimits_{ST} {{V^{T(l)}}} }}{{\sqrt {{V^{S{{(l)}^T}}}\sum\nolimits_{SS} {{V^{S(l)}}} } \sqrt {{V^{T{{(l)}^T}}}\sum\nolimits_{TT} {{V^{T(l)}}} } }}}, \label{eq:6}
\end{equation}
where $\sum\nolimits_{ST} = {H^{S(l)}}{H^{T{{(l)}^T}}}$ , $\sum\nolimits_{SS} = {H^{S(l)}}{H^{S{{(l)}^T}}}$, $\sum\nolimits_{TT} = {H^{T(l)}}{H^{T{{(l)}^T}}}$,
By minimizing eq. (\ref{eq:3}), we can collectively train the two neural networks $\theta^T = \{W^T, b^T\}$ and $\theta^S = \{W^S, b^S\}$.
\subsection{Layer matching}
After constructing the multiple layers of the networks by eq. (\ref{eq:3}), we need to further find the best matching for layers after construction of neural networks. As different layer matching would generate different function loss value $L$ in eq. (\ref{eq:3}), we further define the objective function for layer matching as
\begin{equation}
{R_{s,t}} = \arg \min L. \label{eq:7}
\end{equation}
As the solution for eq. (\ref{eq:3}) should be NP-hard, we would like to solve the problem exhaustively. It is also found for regular data, stack auto encoder layers deeper than 5 cannot generate better results, we suppose the layer number on both domains to be lower than 5 here.
\section{Model Training}
Here we would like to first optimize the eq. \ref{eq:3}. As the equation is not joint convex with all the parameters $\theta^S$, $\theta^T$, $V_s$, and $V_t$, and the two parameters $\theta^S$ and $\theta^T$ are not related with $V^S$ and $v^T$, we would like to introduce two-step iteration optimization.
\subsection{Step.1: Updating ${V^S, V^T}$ with fixed ${\Theta^S, \Theta^T}$}
In Eq. (\ref{eq:3}), the optimization of ${V^S, V^T}$ is just related to the dominator term. The optimization of each layer ${V^S(l_1), V^T(l_2)}$(suppose the layer $1_1$ on source domain is in corresponding with layer $l_2$ on target domain) can be formulated as
\begin{equation}
\mathop {\max }\limits_{{V^{S(l_1)}},{V^{T(l_2)}}} \frac{{{V^{S{{(l_1)}^T}}}\sum\nolimits_{ST} {{V^{T(l_2)}}} }}{{\sqrt {{V^{S{{(l_1)}^T}}}\sum\nolimits_{SS} {{V^{S(l_1)}}} } \sqrt {{V^{T{{(l_2)}^T}}}\sum\nolimits_{TT} {{V^{T(l_2)}}} } }}\label{eq:8}
\end{equation}
As ${{V^{S{{(l_1)}^T}}}\sum\nolimits_{SS} {{V^{S(l_1)}}} }=1$ and ${{V^{T{{(l_2)}^T}}}\sum\nolimits_{TT} {{V^{T(l_2)}}} }=1$ \cite{Canonical2004hardoon}, we can rewrite eq. (\ref{eq:8}) as
\begin{equation}
\begin{array}{l}
\max {V^{S{{(l_1)}^T}}}\sum\nolimits_{ST} {{V^{T(l_2)}}} ,\\
s.t. {V^{S{{(l_1)}^T}}}\sum\nolimits_{SS} {{V^{S(l_1)}}} = 1, {V^{T{{(l_2)}^T}}}\sum\nolimits_{TT} {{V^{T(l_2)}}} = 1 \label{eq:9}
\end{array}
\end{equation}
This is a typical constrained problem which can be formulated as a series of unconstrained minimization problems, and be easily solved by Lagrangian multiplier.
\subsection{Step.2: Updating ${\Theta^S, \Theta^T}$ with fixed ${V^S, V^T}$}
As $\Theta^S$ and $\Theta^T$ are mutual independent and with the same form, we here just demonstrate the solution of $\Theta^S$ on the source domain (the solution of $\Theta^T$ can be derived similarly). Actually the objective division operation is with the same function with subtraction operation and we reformulate the objective function as
\begin{equation}
\mathop {\min }\limits_{{\theta ^S}} \phi ({\theta ^S}) = {L_S}({\theta^S}) - \Gamma ({V^S},{V^T})\label{eq:20}
\end{equation}
Here we apply the gradient descent method to adjust the parameter as
\begin{equation}
\begin{aligned}
&{W^{S(l_1)}} = {W^{S(l_1)}} - {\mu ^S}\frac{{\partial \phi }}{{\partial {W^{S(l_1)}}}} \\
&= \frac{{\partial {L_S}({\theta^S})}}{{\partial {W^{S(l_1)}}}} - \frac{{\partial {\Gamma}({V^S},{V^T})}}{{\partial {W^{S(l_1)}}}}\\
&= \frac{{({\alpha ^{S(l_1 + 1)}} - {\beta ^{S(l_1 + 1)}} + {\omega _l}{\gamma ^{S(l_1 + 1)}}) \times {H^{S(l_1)}}}}{{{n_c} + {\lambda ^S}{W^{S(l_1)}}}}\label{eq:21}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
&{b^{S(l_1)}} = {b^{S(l_1)}} - {\mu ^S}\frac{{\partial \phi }}{{\partial {b^{S(l_1)}}}} \\
&= \frac{{\partial {L_S}({\theta^S})}}{{\partial {b^{S(l_1)}}}} - \frac{{\partial \Gamma ({V^S},{V^T})}}{{\partial {b^{S(l_1)}}}} = \\
&\frac{{({\alpha ^{S(l_1 + 1)}} - {\beta ^{S(l_1 + 1)}} + {\omega _l}{\gamma ^{S(l_1 + 1)}})}}{{{n_c}}},\label{eq:22}
\end{aligned}
\end{equation}
in which
\begin{equation}
{\alpha ^{S(l_1)}} = \left\{ {\begin{array}{*{20}{c}}
{ - ({D_S^l} - {H^{S(l_1)}}) \cdot {H^{S(l_1)}} \cdot (1 - {H^{S(l_1)}}),l = {n^S}}\\
{{W^{S{{(l_1)}^T}}}{\alpha ^{S(l_1 + 1)}} \cdot {H^{S(l_1)}} \cdot (1 - {H^{S(l_1)}})},\\
l = 2,...,{n^S} - 1
\end{array}} \right.
\end{equation}
\begin{equation}
{\beta ^{S(l_1)}} = \left\{ {\begin{array}{*{20}{c}}
{0,\qquad l = {n^S}}\\
{{H^{T(l_2)}}{V^{T(l_2)}}{V^{S{{(l_1)}^T}}} \cdot {H^{S(l_1)}} \cdot (1 - {H^{S(l_1)}})},\\
l = 2,...,{n^S} - 1
\end{array}} \right.
\end{equation}
\begin{equation}
{\gamma ^{S(l_1)}} = \left\{ {\begin{array}{*{20}{c}}
{0,l = {n^S}}\\
{{H^{S(l_1)}}{V^{S(l_1)}}{V^{S{{(l_1)}^T}}} \cdot {H^{S(l_1)}} \cdot (1 - {H^{S(l_1)}})},\\
l = 2,...,{n^S} - 1
\end{array}} \right..
\end{equation}
The operator $\cdot$ here stands for the dot product. The same optimization process works for $\Theta^T$ on the target domain.
\floatname{algorithm}{algorithm}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{CJK*}{UTF8}{gkai}
\begin{algorithm}
\caption{Deep Mapping Model Training}\label{alg:1}
\begin{algorithmic}[1]
\Require $D^C = \{C_i^s, C_i^t\}_i^{n_c}$,
\Require $\lambda^S = 1, \lambda^T = 1$, $\mu^S = 0.5, \mu^T = 0.5$
\Ensure $\Theta(W^S, b^S), \Theta(W^T, b^T), V^S, V^T$
\Function{NetworkSetUp}{}
\State Initialize $\Theta(W^S, b^S), \Theta(W^T, b^T)\gets Random Num$
\Repeat
\For {{$l=1,2,...,n^S$}}
\State ${V^S} \gets \arg \min L({\omega _l},{V^{S(l)}})$
\EndFor
\For {{$l=1,2,...,n^T$}}
\State ${V^T} \gets \arg \min L({\omega _l},{V^{T(l)}})$
\EndFor
\State ${\theta ^S} = \arg \min \phi ({\theta ^S}), {\theta ^T} = \arg \min \phi ({\theta ^T})$
\Until {Convergence}
\EndFunction
\Function{LayerMatching}{}
\State Initialize $R_{s,t} \gets Random Matching$
\State Initialize $m, n \gets 0$
\If {$m<5$}
\If {$n<A_5^5/5$}
\State $R^m_{s,t} = \max {\rm{\{ r}}_n^m{\rm{,r}}_{n + 1}^m{\rm{\} }}$
\State n = n+1
\EndIf
\State $R_{s,t} = \max {\rm{\{ R}}_{s,t}^m{\rm{,r}}_{s,t}^{m + 1}{\rm{\} }}$
\State m = m+1
\EndIf
\EndFunction
\end{algorithmic}
\end{algorithm}
\end{CJK*}
After these two optimizations for each layer, the two whole networks (the source domain network and the target domain network) are further fine-tuned by the back-propagation process. The forward and backward propagations will iterate until convergence.
\subsection{Optimization of $R_{s,t}$}
We finally get the minimized $R_{s,t}$ by the above procedures. Take the above layer matching as an example, in $R_{s,t}$, layer $l_1$ in source domain is in corresponding with
$l_2$ in target domain. As we define the both network for no more than 5 layers(including the original data layers, which will not be used to transfer), the theoretically maximum number of combination should be no more than $A_5^5$(as some layers can be vacancy with no matching). However, in our experiments, we heuristically find the number of matching layers should be in direct proportion to the resolution of images. This observation would save a lot of training time.
The training process is finally summarized in Alg. (\ref{alg:1}), where $r^m_{i,j}$ is generally written as $r^m_n$.
\subsection{Classification on common semantic subspace}
The final classification is performed on the common subspace $\Omega$. The target domain data $D_T$ and the labeled $D_S^l$ are both projected to the common subspace $\Omega$ by the correlation coefficients $V^S(n_S)$ and $V^T(n_T)$. The standard SVM algorithm is applied on $\Omega$. The classifier $\Psi$ is trained by $D_S^l$. This trained classifier $\Psi$ is applied to $D^T$ as $\Psi({D^T}*{V^T})$.
\section{Experiments}
We carry out our DT-LET framework on two cross-domain recognition tasks, handwritten digit recognition, and text-to-image classification.
\subsection{Experimental dataset descriptions}
\textbf{\emph{Handwritten digit recognition:}}For this task, we mainly conduct the experiment on Multi Features Dataset collected from UCI machine learning repository. This dataset consists of features of handwritten numerals (0-9, in total 10 classes) extracted from a collection of Dutch utility maps. 6 features exist for each numeral and we choose the most popular features 216-D profile correlations and 240-D pixel averages in 2*3 windows to complete the transfer learning based recognition task.
\textbf{\emph{Text-to-image classification:}}For this task, we make use of NUS-WIDE dataset. In our experiment, the images in this dataset are represented with 500-D visual features and annotated with 1000-D text tags from Flickr. 10 categories of instances are included in this classification task, which are birds, building, cars, cat, dog, fish, flowers, horses, mountain, and plane.
\subsection{Comparative methods and evaluation}
As the proposed ET-LET framework mainly have four components, deep learning, CCA, layer matching, and SVM classifier, we first select 3 baseline methods, CCA-SVM \cite{Canonical2004hardoon}, Kernelized-CCA-SVM(KCCA-SVM)\cite{KCCA}, and Deep-CCA-SVM(DCCA-SVM)\cite{Chen2012Marginalized} as baseline comparison methods. We also conduct experiment just without layer matching(the number of layers are the same on the source and the target domains) while all the other parameters are the same with the proposed ET-LET, and we name this framework NoneDT-LET. The final comparison method is the most representative duft-tDTNs method\cite{duft}, which should be up to now heterogenous transfer learning method with best performance.
For the deep network based method, the DCCA-SVM, duft-tDTNs, NoneDT-LET are all with 4 layers for the source domain and the target domain data, as we find more or less layer would generate worse performance.
At last, for the evaluation metric, we select the classification accuracies on the target domain data over the 2 pairs of datasets.
\subsection{Task 1: Handwritten digit recognition}
In the first experiment, we conduct our study for handwritten digit recognition. The source domain data are the 240-D pixel averages in 2*3 windows feature, while the target domain data are the 216-D profile correlations feature. As there are 10 classes in total, we complete 45 ($C_{10}^2$) binary classification tasks, for each category, the accuracy is the average accuracy of 9 binary classification tasks. We use 60\% data as co-occurrence data to complete the transfer learning process and find the common subspace, 20\% labeled samples on source domain as the training samples, and the rest samples on target domain as the testing samples to complete the classification process. The experiments are repeated for 100 times with 100 sets of randomly chosen training and testing data to avoid data bias \cite{Beyond2012Tati}. The final accuracy is the average accuracy of the 100 repeated experiments. This data setting applies for all four methods under comparison.
For the deep network, the numbers of neurons of 4 layer networks are 240-170-100-30 for source domain data and 216-154-92-30 for target domain data, this setting works for the all comparison methods. For the proposed DT-LET, we find the best two layer matching with lowest loss after 20 iterations are $r_{4,3}^2$ and $r_{5,4}^3$. The numbers of neurons for $r_{4,3}^2$ are 240-170-100-30 for source domain data and 216-123-30 for target domain data. The average objective function loss of the all 45 binary classification tasks for these two are 0.856 and 0.832 respectively. The numbers of neurons for $r_{5,4}^3$ are 240-185-130-75-30 for source domain data and 216-154-92-30 for target domain data. The one-against-one SVM classification is applied for final classification. The average classification accuracies of 10 categories are shown in Tab. \ref{tab:1}. The matching correlation is detailed in Fig. \ref{fig:2}.
\begin{table*}[t]\footnotesize
\begin{center}
\caption{Classification Accuracy Results on Multi Feature Dataset. (The best performance is emphasized by boldface.)} \label{tab:1}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
numeral & CCA-SVM & KCCA-SVM & DCCA-SVM & duft-tDTNs & NoneDT-LET & DT-LET($r_{5,4}^3$) & DT-LET($r_{4,3}^2$)
\\
\hline
0 & 0.750 & 0.804 & 0.961 & 0.972 & 0.983 & \textbf{0.989} & 0.984\\
1 & 0.740 & 0.767 & 0.943 & 0.956 & 0.964 & 0.976 & \textbf{0.982}\\
2 & 0.780 & 0.812 & 0.955 & 0.972 & 0.979 & 0.980 & \textbf{0.989}\\
3 & 0.748 & 0.790 & 0.945 & 0.956 & 0.966 & \textbf{0.976} & 0.975\\
4 & 0.752 & 0.799 & 0.956 & 0.969 & 0.980 & \textbf{0.987 }& 0.983\\
5 & 0.728 & 0.762 & 0.938 & 0.949 & 0.958 & 0.971 & \textbf{0.977}\\
6 & 0.755 & 0.770 & 0.958 & 0.966 & 0.978 & \textbf{0.988} & 0.986\\
7 & 0.775 & 0.797 & 0.962 & 0.968 & 0.978 & 0.975 & \textbf{0.985}\\
8 & 0.764 & 0.793 & 0.948 & 0.954 & 0.965 & 0.968 & \textbf{0.975}\\
9 & 0.754 & 0.781 & 0.944 & 0.958 & 0.970 & \textbf{0.976} & 0.961\\
\hline
\end{tabular}
\end{center}
\end{table*}
As can be found in Tab. \ref{tab:1}, the best performances have been highlighted, which all exist in DT-LET framework. However, the best performances for different categories do not exist in the framework with same layer matching. Overall, $r_{5,4}^3$ and $r_{4,3}^2$ should be the best two layer matchings compared with other settings. Based on these results, we heuristically get the conclusion that the best layer matching ratio(5/4, 4/3) is generally in direct proportion to the dimension ratio of original data(240/216). However, more matched layers do not guarantee better performance as the classification results for number ``1", ``2", ``5", ``7", ``8" of DT-LET ($r_{4,3}^2$) with 2 layer matchings perform better than DT-LET($r_{5,4}^3$) with 3 layer matchings.
\begin{figure*}[!ht]
\centering
\includegraphics[width=0.7\linewidth]{figure_comp1.pdf}
\caption{The comparison of different layer matching setting for different frameworks on Multi Feature Dataset.}
\label{fig:2}
\end{figure*}
\subsection{Task 2: Text-to-image classification}
In the second experiment, we conduct our study for Text-to-image classification. The source domain data are the 1000-D text feature, while the target domain data are the 500-D image feature. As there are 10 classes in total, we complete 45 ($C_{10}^2$) binary classification tasks. We still use 60\% data as co-occurrence data\cite{Learning2016Yang}, 20\% labeled samples on source domain as the training samples, and the rest samples on target domain as the testing samples. The same data setting as Task 1 applies for all four methods under comparison.
For the deep network, the numbers of neurons of 4 layer networks are 1000-750-500-200 for source domain data and 500-400-300-200 for target domain data, this setting works for the all comparison methods. For the proposed DT-LET, we find the best two layer matchings with lowest loss after 20 iterations are $r_{5,3}^2$, $r_{5,4}^3$ and $r_{5,4}^2$(non-full rank). The average objective function loss of 45 binary classification tasks for these two layer matchings are 3.231, 3.443 and 3.368. The numbers of neurons for $r_{5,3}^2$ are 1000-800-600--400-200 for source domain data and 500-350-200 for target domain data. The numbers of neurons for both $r_{5,4}^3$ and $r_{5,4}^2$ are 1000-750-500-200 for source domain data and 500-400-300-200 for target domain data. As matching principle would also influence the performance of transfer learning, we present two $r_{5,3}^2$ with different matching principles as shown in Fig. \ref{fig:3}(the average objective function loss for the two different matching principles are 3.231 and 3.455), in which all the detailed layer matching principles are described.
\begin{table*}[t]\small
\begin{center}
\caption{Classification Accuracy Results on NUS-WIDE Dataset. (The best performance is emphasized by boldface.)} \label{tab:2}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}*{categories}& \multirow{2}*{CCA-SVM}& \multirow{2}*{KCCA-SVM}& \multirow{2}*{ DCCA-SVM}& \multirow{2}*{duft-tDTNs}& \multirow{2}*{NoneDT-LET}& \multicolumn{4}{c|}{DT-LET}\\
\cline{7-10}
& & & & & & $r_{5,3}^2$(1) & $r_{5,3}^2$(2) & $r_{5,4}^2$& $r_{5,4}^3$\\
\hline
birds & 0.690 & 0.723 & 0.784 & 0.770 & 0.796 & 0.825 & 0.830 &\textbf{ 0.848 } & 0.825\\
building & 0.706 & 0.741 & 0.810 & 0.783 & 0.816 & 0.881 & 0.838 & 0.881 & \textbf{0.891}\\
cars & 0.702 & 0.731 & 0.803 & 0.773 & 0.812 & 0.832 & 0.827 & \textbf{0.867} & 0.853\\
cat & 0.692 & 0.731 & 0.797 & 0.766 & 0.806 & 0.868 & \textbf{0.873} & 0.859 & \textbf{0.873}\\
dog & 0.687 & 0.726 & 0.798 & 0.765 & 0.805 &0.847 & 0.847 & \textbf{0.863} & 0.823\\
fish & 0.674 & 0.713 & 0.773 & 0.752 & 0.781 & 0.848 & 0.834 & \textbf{0.852} & 0.839\\
flowers & 0.698 & 0.733 & 0.799 & 0.783 & 0.805 & 0.863 & 0.844 & 0.844 & \textbf{0.875}\\
horses & 0.700 & 0.736 & 0.802 & 0.775 & 0.808 & \textbf{0.841 }& 0.812 & \textbf{0.841} & 0.831\\
mountain & 0.717 & 0.748 & 0.816 & 0.786 & 0.827 & 0.825 & 0.813 & 0.821 & \textbf{0.831}\\
plane & 0.716 & 0.747 & 0.824 & 0.787 & 0.828 & 0.810 & \textbf{0.832} & \textbf{0.832} & 0.825\\
average & 0.698 & 0.733 & 0.801 & 0.774 & 0.808 & 0.844 & 0.833 & \textbf{0.851} & 0.847 \\
\hline
\end{tabular}
\end{center}
\end{table*}
For this task, as the overall accuracies are generally lower than task 1, we would like to compare more different settings for this cross-layers matching task. We first verify the effectiveness of DT-LET framework. Compared with the comparison methods, the accuracy of DT-lET framework is generally with around 85\% accuracy while the comparison methods are generally with no more than 80\%. This observation generates the conclusion that finding the appropriate layer matching is essential. The second comparison is between the full rank and non-full rank framework. As can be found in the table, actually $r_{5,4}^2$ is with the highest overall accuracy, although the other non-full rank DT-LETs do not perform quite well. This observation gives us a hint that full rank transfer is not always best as negative transfer would degrade the performance. However, the full rank transfer is generally good, although not optimal. The third comparison is between the same transfers with different matching principles. We present two $r_{5,3}^2$ with different matching principles, and we find the performances vary. The case 1 performs better than case 2. This result tell us continuous transfer might be better than discrete transfer: as for case 1, the transfer is in the last two layers of both domains, and in case 2, the transfer is conducted in layer 3 and layer 5 of the source domain data.
By comparing specific objects, we can find the objects with large semantic difference with other categories are with higher accuracy. For the objects which are hard to classify and with low accuracy, like ``birds" and ``plane", the accuracies are always low even the DT-LET is introduced. This observation proves the conclusion that DT-LET can only be used to improve the transfer process, which helps with the following classification process; while the classification accuracy is still based on the semantic difference of data of different categories.
We also have to point out the relationship between the average objective function loss and the classification accuracy is not strictly positive correlated. Overall $r_{5,4}^2$ is with the highest classification accuracy while its average objective function loss is not lowest. Based on this observation, we have to point out, the lowest average objective function loss can only generate the best transfer leaning result with optimal common subspace. On such common subspace, the data projected from target domain are classified. These classification results are also influenced by the classifier as well as training samples projected randomly from the source domain. Therefore, we conclude as follows. We can just guarantee a good classification performance after getting the optimal transfer learning result, while the classification accuracy is also influenced by the classification settings.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{figure_comp2.pdf}
\caption{The comparison of different layer matching setting for different frameworks on NUS-WIDE Dataset.}
\label{fig:3}
\end{figure}
\subsection{Parameter sensitivity}
In this section, we study the effect of different parameters in our networks. We have to point out the even the layer matching is random, the last layer of the two neural networks from the source domain and the target domain must be correlated to construct the common subspace. Actually, the number of neurons at last layer would also affect the final classification result. For the last layer, we take experiments on Multi Feature Dataset as an example. The result is shown in Tab. \ref{tab:7}.
From this figure, it can be noted when the number of neuron is 30, the performance is the best. Therefore in our former experiments, 30 neurons are used. The conclusion can also be drawn that more neurons are not always better. Based on this observation, The number of layers in Task 1 is set as 30, and in Task 2 as 200.
\begin{table}[t]\small
\begin{center}
\caption{Effects of the number of neurons at the last layer} \label{tab:7}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
layer matching & 10 & 20 & 30 & 40 & 50\\
\hline
$r_{5,4}^3$ & 0.9082& 0.9543& 0.9786& 0.9771& 0.9653 \\
\hline
$r_{4,3}^2$ & 0.8853& 0.9677& 0.9797& 0.9713& 0.9522 \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Conclusion}
In this paper, we propose a novel framework, referred to as Deep Transfer Learning by Exploring where to Transfer(DT-LET), for hand writing digit recognition and text-to-image classification. In the proposed model, we find the best matching with lowest loss value. After the transfer, the final correlated common subspace on which classifier is applied. Experimental results support the effectiveness of the proposed framework.
As the current framework is only suitable for binary classification, extending it to multi-class classification is our future work. We would also propose more robust model to solve this ``where to transfer" problem in the future.
\bibliographystyle{aaai}
|
\section{Introduction}
\label{sec:introduction}
Textual information is often represented through structured documents which have an inherent 2D structure. This is even more so the case with the advent of new types of media and communications such as presentations, websites, blogs and formatted notebooks.
In such documents, the layout, positioning, and sizing might be crucial to understand its semantic content and provide a strong guidance to the human perception.
NLP addresses the task of processing and understanding natural language texts through sub-tasks like language modeling, classification, information extraction, summarization, translation, and question answering among others. NLP methods typically operate on serialized text, which is a 1D sequence of characters. Such methods have been proven very successful for various tasks on unformatted text (e.g.~books, reviews, news articles, short text snippets).
However, when processing structured and formatted documents in which the relation between words is impacted not only by the sequential order, but also by the document layout, NLP can result in significant shortcomings.
Computer vision algorithms, on the other hand, are designed to exploit 2D information in the visual domain. Images are commonly processed with convolutional neural networks \cite{lecun1998gradient, krizhevsky2012imagenet, DBLP:journals/corr/RenHG015, pinheiro2016learningRefineObjectSegments} (or likes) that preserve and exploit the 2D correlation between neighboring pixels.
While it is feasible to convert structured documents into images and then apply computer vision algorithms, this approach is not optimal for understanding their semantics as it is driven mostly by the visual content and not by the textual content. As a result, a machine learning model would first need to extract the text from the image followed by learning the semantics. This purely visual approach requires a more complex model and significantly larger training data compared to text-based approaches.
We propose a novel paradigm for processing and understanding structured documents. Instead of serializing a document into a 1D text, the proposed method, named \emph{chargrid}, preserves the spatial structure of the document by representing it as a sparse 2D grid of characters.
\sugg{We then formulate the document understanding task as instance-level semantic segmentation on \emph{chargrid}. More precisely, the model predicts a segmentation mask with pixel-level labels and object bounding boxes to group multiple instances of the same class.}
We apply the chargrid paradigm on an information extraction task from invoices and demonstrate that this method is superior to both, state-of-the-art NLP algorithms as well as computer vision algorithms.
The rest of the document is organized in three parts: we first introduce the \textit{chargrid} paradigm; we then describe a specific application of information extraction from documents; finally, we present experimental results and conclusion.
\section{Related Work}
\label{sec:relatedwork}
NLP focuses on understanding natural language text through tasks like classification \cite{DBLP:journals/corr/Kim14f}, translation \cite{DBLP:journals/corr/BahdanauCB14}, summarization \cite{DBLP:journals/corr/RushCW15}, and named entity recognition \cite{lample2016neural}. Such methods expect unformatted text as input and, therefore, do not assume any intrinsic 2D structure.
Document Analysis, on the other hand, deals largely with problems such as recognizing printed/hand-written characters from a variety of documents \cite{NIPS2008_3449}, processing document images for document localization \cite{8269957}, binarization \cite{abs-1708-03276}, and layout segmentation \cite{7333914}. As a result, it does not focus on understanding the character- and/or word-level semantics of the document the same way as NLP.
Within computer vision, problems such as scene text detection and recognition \cite{DBLP:journals/corr/GoodfellowBIAS13}, semantic segmentation \cite{badrinarayanan2017segnet} as well as object detection \cite{gupta2014learning, RetinaNet} can be considered as related problems to ours, but applied on a different domain, i.e., processing natural images instead of documents as input.
The closest to our work is \citet{yang2017learning} that performs pixel-wise layout segmentation on a structured document, using sentence embeddings as additional input to an encoder-decoder network architecture. For each pixel inside the area of a sentence, the sentence embedding is appended to the visual feature embedding at the last layer of the decoder. The authors show that the layout segmentation accuracy can be significantly improved when using the textual features.
Another related work is \citet{palm2017cloudscan} which extracts key-value information from structured documents (invoices) using a recurrent neural network (RNN). Their work addresses the problem of document understanding, however, the RNN operates on serialized 1D text.
Combining approaches from computer vision, NLP, and document analysis, our work is the first to systematically address the task of understanding 2D documents the same way as NLP while still retaining the 2D structure in structured documents.
\section{Document Understanding with Chargrid}
A human observer comprehends a document by understanding the semantic content of characters, words, paragraphs, and layout components. We encapsulate all such tasks under the common umbrella of \textit{document understanding}. Therefore, we can formulate this problem as an instance segmentation task of characters on the page. In the following sections, we describe a new approach for solving that task.
\subsection{Chargrid}
\label{sec:chargrid}
\textit{Chargrid} is a novel representation of a document that preserves its 2D layout.
A chargrid can be constructed from \textit{character boxes}, i.e.,~bounding boxes that each surround a single character somewhere on a given document page. This positional information can come from an optical character recognition (OCR) engine, or can be directly extracted from the layout information in the document as provided by, e.g., PDF or HTML. The coordinate space of a character box is defined by page height $H$ and width $W$, and is usually measured in units of pixels.
The complete text of a document page can thus be represented as a set of tuples $\mathcal{D} = \{(c_k, \mathbf{b}_k)\,|\,k=0,...,n\}$, where $c_k$ denotes the $k$-th character in the page and $\mathbf{b}_k$ the associated character box of the $k$-th character, which is formalized by the top-left pixel position, width and height, thus $\mathbf{b}_k = (x_k, y_k, w_k, h_k)$.
We can now construct the chargrid $g\in \mathbb{N}^{H \times W}$ of the original document page, and its \textit{character-pixel} $g_{ij}$ from the set $\mathcal{D}$ with
\begin{equation}
\label{eq:chargrid}
g_{ij} =
\begin{cases}
E(c_k) & \text{if}\quad (i,j)\prec\mathbf{b}_k \\
0 &
\end{cases}
\end{equation}
where $\prec$ means 'overlaps with', and
where each point $(i,j)$ corresponds to some pixel in the original document page pixel coordinate space defined by $(H,W)$.
$E(c_k)$ is some encoding of the character in the $k$-th character box, i.e.~the value of character $c_k$ may be mapped to a specific integer index. For instance, we may map the alphabet (or any character of interest) to non-zero indices $\{a,b,c,...\} \rightarrow \{1,2,3,...\}$.
Note that we assume that character boxes cannot overlap, i.e.~each character on a document page occupies a unique region on the page. In practice, it may happen that the corners and edges of a character box may overlap with other closeby characters. We solve such corner cases by assigning the character-pixel to the box that has the closest box center.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/Chargrid_fig1b.pdf}
\end{center}
\caption{Example for a document page (left) and corresponding chargrid representation $g$ (right).}
\label{fig:chargrid}
\end{figure}
In other words, the chargrid representation is constructed as follows: for each character $c_k$ at location $\mathbf{b}_k$, the area covered by that character is filled with some constant index value $E(c_k)$. All remaining character-pixels corresponding to empty regions on the original document page are initialized with 0. Figure~\ref{fig:chargrid} visualizes the chargrid representation of an example input document.
The advantage of the new chargrid representation is twofold: (i) we directly encode a character by a single scalar value rather than by a granular collection of grayscale pixels as is the case for images, thus making it easy
for the subsequent document analysis algorithms to understand the document, and (ii), because the group of pixels that belonged to a given character are now all mapped to the same constant value, we can significantly downsample the chargrid representation without loss of any information.
For instance, if the smallest character occupied a $10\times10$ pixel region in the original document, we can downsample the chargrid representation by a factor of $10\times10$.
This significantly reduces the computational time of subsequent processing steps, such as training a machine learning model on this data representation.
We point out that a character can occupy a region spanning several character pixels. This is in contrast to traditional NLP where each character is represented by exactly one token. In our representation, therefore, the larger a given character, the more character-pixels represent that single character.
We do not find this to be a problem, instead, it even helps since it implicitly encodes additional information, for example, the font size, that would otherwise not be available.
Before the chargrid representation is used as input to, e.g.,~a neural network (see Sec.~\ref{sec:networkarch}), we apply 1-hot encoding to the chargrid $g$. Thus, the original chargrid representation $g \in \mathbb{N}^{H \times W}$ becomes a vector representation $\tilde{g} \in \mathbb{R}^{H \times W \times N_C}$, where $N_C$ denotes the number of characters in the vocabulary including a padding/background character (in our case this is mapped to 0) and an unknown character (all characters that are not mapped by our encoding $E$ will be mapped to this character).
\sugg{We note that similar to using characters for constructing the chargrid, one can also use words to construct a \textit{wordgrid} in the same way. In that case, rather than using 1-hot encoding, one may use a word embedding like word2vec or GloVe. While the construction of a wordgrid seems straight-forward, we have not experimented with it in the present work as our dataset contains too many unusual words and spans multiple languages (see Sec.~\ref{sec:invoivoice_analysis}).}
\subsection{Network Architecture}
\label{sec:networkarch}
\begin{figure*}
\begin{center}
\includegraphics[width=1.\textwidth]{figures/chargrid_networ5.pdf}
\end{center}
\caption{Network architecture for document understanding, the \textit{chargrid-net}. Each convolutional block in the network is represented as a box. The height of a box is a proxy for feature map resolution while the width is a proxy for the number of output channels. $C$ corresponds to the number of 'base' channels, which in turns corresponds to the number of output channels in the first encoder block. $d$ denotes dilation rate.}
\label{fig:networkarch}
\end{figure*}
We use the 1-hot encoded chargrid representation $\tilde{g}$ as input to a fully convolutional neural network to perform semantic segmentation on the chargrid and predict a class label for each character-pixel on the document. As there can be multiple and an unknown number of instances of the same class, we further perform instance segmentation. This means, in addition to predicting a segmentation mask, we may also predict bounding boxes using the techniques from object detection. This allows the model to assign characters from the same segmentation class to distinct instances.
Our model is described in Figure~\ref{fig:networkarch}. It is comprised of two main parts: The encoder network and the decoder network. The decoder network is further made up of two branches: The segmentation branch and the bounding box regression branch.
The encoder boils down to a VGG-type network \cite{simonyan2014very} with dilated convolutions \cite{YuKoltun2016}, batch normalization \cite{ioffe2015batch}, and spatial dropout \cite{tompson2015efficient}. Essentially, the encoder consists of five blocks where each block consists of three $3\times3$ convolutions (which themselves are made of convolution, batch normalization, ReLU activation) followed by spatial dropout at the end of a block. The first convolution in a block is a stride-2 convolution to downsample the input to that block. Whenever we downsample, we increase the number of output channels $C$ of each convolution by a factor of two.
We have found stride-2 convolutions to yield slightly better results compared to max pooling. In block four and five of the encoder, we do not apply any downsampling, and we leave the number of channels at $512$ (the first block has $C=64$ channels). We use dilated convolutions in block three, four, five with rates $d=2,4,8$, respectively.
The decoder for semantic segmentation and for bounding box regression are both made of convolutional blocks which essentially reverse the downsampling of the encoder via stride-2 transposed convolutions.
Each block first concatenates features from the encoder via lateral connections followed by $1\times1$ convolutions \cite{UNET, pinheiro2016learningRefineObjectSegments}. Subsequently, we upsample via a $3\times3$ stride-2 transposed convolution. This is followed by two $3\times3$ convolutions. Note that whenever we upsample, we decrease the number of channels by a factor of two.
The two decoder branches are identical in architecture up to the last convolutional block. The decoder for semantic segmentation has an additional convolutional layer without batch normalization, but with bias and with softmax activation. The number of output channels of the last convolution corresponds to the number of classes.
Together with the encoder, the decoder for the bounding box regression task forms a one-stage detector which makes use of focal loss \cite{RetinaNet}.
We also make use of the anchor box representation and corresponding bounding box regression targets as discussed in \citet{FasterRCNN}. The anchor-box representation allows us to handle bounding boxes that vary widely with respect to size and aspect ratios. Moreover, the anchor-box representation allows us to detect boxes of different classes.
The number of output channels are $2N_a$ for the box mask (foreground versus background) and $4N_a$ for the four box coordinates, where $N_a$ is the number of anchors per pixel.
The weights of all layers are initialized following \citet{DelvingDeepIntoRectifiers}, except for the last ones, which are initialized with a small constant value $1e-3$ for stabilization purposes.
In total, we have three equally contributing loss terms:
\begin{equation}
\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{seg}} + \mathcal{L}_{\mathrm{boxmask}} + \mathcal{L}_{\mathrm{boxcoord}},
\end{equation}
where $\mathcal{L}_{\mathrm{seg}}$ is the cross entropy loss for segmentation, e.g.~\cite{UNET}, $\mathcal{L}_{\mathrm{boxmask}}$ is the binary cross entropy loss for box masks, and $\mathcal{L}_{\mathrm{boxcoord}}$ is the Huber loss for box coordinate regression \cite{FasterRCNN}. Both cross entropy loss terms are augmented following the focal loss idea \cite{RetinaNet}. We also make use of aggressive static class weighting in both cross entropy loss terms mainly to counter the strong class imbalance between irrelevant and "easy" pixels (the "background" class) versus relevant and "hard" pixels (all other classes; see Sec.~\ref{sec:impl-details} for further details).
We refer to the network depicted in Figure~\ref{fig:networkarch} as the \textit{chargrid-net}.
\section{Information Extraction from Invoices}
\label{sec:invoivoice_analysis}
As a concrete example for understanding structured 2D documents, we extract key-value information from invoices. We make no assumption on the format of the invoice, the country of origin (and consequently taxes, date formats, amount formats, currency etc.) or the language. In addition to that, real-world invoices often contain incomplete sentences, nouns, and abbreviations.
We want our model to parse an invoice and to extract 5 \textit{header fields} (i.e., Invoice Number, Invoice Date, Invoice Amount, Vendor Name and Vendor Address) as well as the list of product items purchased, referred to as the \textit{line-items}. Line-items include details for each item such as Line-item Description, Line-item Quantity and Line-item Amount. Together with the background class, this yields 9 classes and each character on the invoice is associated to exactly one class. We note that while header fields may only appear once on an invoice (are unique), line-items may occur in multiple instances.
To extract the values for each field, we collect all characters that are classified as belonging to the corresponding class. For line-items, we further group the characters by the predicted item bounding boxes.
\subsection{Data}
\begin{figure}
\centering
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/vendor.png}
\end{subfigure}
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/lang.png}
\end{subfigure}
\caption{The left image shows a histogram of number of vendors over the number of contributing invoices in the dataset. Most vendors appear only once in the dataset. The right image shows a distribution over languages, illustrating the diversity of the invoice data.}
\label{fig:vendors_language}
\end{figure}
Our invoice dataset consists of 12k scanned sample invoices from a large variety of different vendors and languages. We assign 10k samples for training, 1k for validation and 1k for test on which we report our results (see Sec.~\ref{sec:results}). We ensure that vendors that appear in one set do not appear in any other set. This is more restrictive than necessary but gives a good estimate on how well the model generalizes to unseen invoice layouts.
Figure~\ref{fig:vendors_language} shows the distribution over vendors and languages. From the vendor distribution, it can be seen that most vendors contribute only one invoice to the dataset with at most 6 invoices coming from a single vendor. From the language distribution, it can be seen that while English is the predominant language, there are large representations from French, Spanish, Norwegian, and German.
For all invoices, we collected manual annotations with bounding boxes around the fields of interest. Considerable efforts were spent to ensure that the labels are correct. In particular, each invoice was analyzed by three annotators plus a reviewer. Over 35k invoices were investigated and finally only clean set of 12k invoices were selected. Furthermore, detailed instructions were given to the annotators for each field. Figure~\ref{fig:spatial_distribution} visualizes the locations of annotated boxes on invoices for Invoice Amount and Line-item Quantity. It can be seen that while some regions are denser, the occurrences are spread widely over the whole invoice thereby illustrating the diversity in the invoice format.
\subsection{Implementation Details}
\label{sec:impl-details}
The invoices were processed with an open-source engine for OCR, \citet{tesseract}, to extract the text from the documents.
We limit the number of distinct characters in our encoding $E$ (in Eq.~\ref{eq:chargrid}) to the most frequent ones, which in our present case means $N_C=54$ different characters including the background/padding and unknown character. Thus, the 1-hot encoded chargrid $\tilde{g}$ has $54$ channels.
Each page of an invoice is processed independently.
Once we created the chargrid representation of a document page, we downsample to a fixed resolution using nearest neighbor interpolation
to ensure that all chargrid representations have the same resolution for training (and inference). Note that this downsampling operation is still performed in the token space (i.e.,~before 1-hot encoding).
To further ensure that no tokens in the chargrid representation are lost due to nearest neighbor interpolation, we downsample to twice the target resolution. This resolution is determined by the smallest characters in the training set.
A second downsampling step, now in the 1-hot encoding to the final target resolution of 336x256 in our case, is performed using bilinear interpolation and fed into the network.
Note that we could have also first performed 1-hot encoding on the document and applied bilinear interpolation directly to the target resolution. Computationally, however, this two-stage downsampling is more efficient.
We handle landscape documents by simply squeezing all input pages into our target resolution using interpolation (similar to image re-sizing). We have found that this approach is not harmful and for simplicity stick to it.
Line-items can occur in an unknown number of distinct instances. Therefore, we require instance segmentation of characters on the document. To accomplish this, the model
is trained to predict bounding boxes that span across the entire row of one instance of a line-item, while the segmentation mask classifies those characters belonging to given column classes (such as, e.g.,~Line-item Quantity, or Line-item Description) of that line-item instance.
We implemented our model in TensorFlow 1.4.
We use SGD with momentum $\beta=0.9$ and learning rate $\alpha=0.05$. We used weight decay of $\lambda=10^{-4}$, and spatial dropout with probability $P=0.1$.
We perform random cropping of the chargrid for data augmentation (that is we pad by 16 character pixels in each direction after downsampling and then crop with a random offset in range $(16,16)$).
We use aggressive class weighting in the cross-entropy loss for semantic segmentation and for the bounding box mask. We have found this to be more effective than the focal loss (which can bee seen as a form of dynamic class weighting). We implement class weighting following \cite{paszke2016enet} with a constant of $c=1.04$.
In early stages of our experiments not yet using class weights, we noted poor performance on the bounding box regression task as well as the segmentation task.
The distribution of line-items on invoices in our dataset reveals that around $50\%$ of all invoices only contain less than three line-items. We found that repeating those invoices with more than three line-items during training more often than those with only few line-items significantly boosts bounding box regression accuracy. With a mini batch size of $7$, each model took around $360k$ iterations and 3 days to fully converge on a single Nvidia Tesla V100.
\subsection{Evaluation Measure}
\label{sec:metric}
For evaluating our model, we would like to measure how much work would be saved by using the extraction system, compared to performing the field extraction manually.
To capture this, we use a measure similar to the word error rate \cite{prabhavalkar2017minimum} used in speech recognition or translation tasks.
For a given field, we count the number of insertions, deletions, and modifications of the predicted instances (pooled across the entire test set) to match the ground truth instances. Evaluations are made on the string level. We compute this measure as follows:
\[1 - \frac{\#[\text{\footnotesize insertions}]+\#[\text{\footnotesize deletions}]+\#[\text{\footnotesize modifications}]}{N}\]
where $N$ is the total number of instances occurring in the ground truth of the entire test set.
This measure can be negative, meaning that it would be less work to perform the extraction manually. In our present case, the error caused by the OCR engine does not affect this measure, because the same errors are present in the prediction and in the ground truth and are not considered a mismatch.
\begin{figure}
\centering
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/amt.png}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/item_qty.png}
\end{subfigure}
\caption{Spatial distribution of Invoice Amount (left) and Line-item Quantity (right) over the invoice. This depicts the variation in the invoice layouts contained in our dataset.}
\label{fig:spatial_distribution}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=1\textwidth]{figures/chargrid_examples_squeezed_adjusted.pdf}
\end{center}
\caption{Sample invoices and their corresponding network predictions: the top row shows the chargrid input, the bottom row shows the predicted segmentation mask with overlaid bounding box predictions. Our model is able to handle a large diversity of layouts. The encoding of the characters on the chargrid has been scrambled to preserve privacy.}
\label{fig:sampleinvoices}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{figures/chargrid_error_examples_adjusted.pdf}
\end{center}
\caption{\sugg{In example 1, chargrid-net attributes unrelated rows of text to the line-item's description. Moreover, some character-pixels are mis-classified such that the predicted header field above the line-item is incorrect. Thus, we observe errors in both, the segmentation mask and the bounding box predictions. In example 2, the model predicts the segmentation mask mostly correct, but it fails to predict boxes for two of the four line-items. In example 3, the model fails to separate adjacent multi-row line-items from each other correctly. Please note that the ground truth is debatable in some cases; for example, it may be hard to decide which information belongs to a line-item's description and which does not.}}
\label{fig:errorexamples}
\end{figure}
\section{Experiments and Results}
\label{sec:results}
Figure~\ref{fig:sampleinvoices} shows some sample predictions. It can be seen that the model can successfully extract key-value information on sample invoices despite significant diversity and complexity in the invoice layouts.
We show the quantitative results in Table~\ref{tab:results}.
We compare the proposed model, chargrid-net (Sec.~\ref{sec:networkarch}), against four other models. The first one is a \textit{sequential} model based purely on text. It is a stack of bi-directional GRUs \cite{DBLP:journals/corr/ChoMGBSB14} taking a sequence of characters as input, and producing a sequence of labels as output. This model is our implementation of \citet{palm2017cloudscan} and serves as a baseline comparison against a more traditional NLP approach which is based on sequential input.
\sugg{We note that we have also experimented with a extension of this model that along with the characters also takes as input the position of each character on the document, however, with negligible returns. Therefore, we stick to this simpler model.}
The second \textit{image-only} model is identical to chargrid-net (Figure~\ref{fig:networkarch}), except that we directly take the original image of the document page as network input rather than the chargrid $\tilde{g}$. This model serves as a baseline comparison to a purely image-based approach using directly the raw pixel information as input. We note that we downsample the image to the same input resolution as the chargrid representation, that is 336x256.
The third and fourth models are both a \textit{hybrid} between the chargrid-net, and the image-only model. In both models, we replicate the encoder: one encoder for the chargrid input $\tilde{g}$, and one encoder for the image of the document page. Information from the two encoders is concatenated in the decoder: whenever a lateral connection from the encoder of the original chargrid-net is concatenated in a decoder block, we now concatenate lateral connections from the two encoders in a decoder block.
We distinguish two configurations of the hybrid model: model \textit{chargrid-hybrid-C64} where chargrid and image encoders both have $C=64$ base channels, and model \textit{chargrid-hybrid-C32} where chargrid and image encoders have $C=32$ base channels while the decoder still retains $C=64$ base channels. The latter model is used to assess the influence on the number of encoder parameters since the hybrid model effectively has more parameters due to the two encoders.
It can be seen that compared to the purely text-based approach, the chargrid-net performs equivalently on
single-instance single-word fields where the 2D structure is not as important. Examples of single-instance, single-word fields are 'Invoice Number', 'Invoice Amount', and 'Invoice Date'. The extraction of these fields could essentially also be easily tackled with named entity recognition approaches based on serialized text \cite{gillick2015multilingual, lample2016neural}.
\begin{center}
\begin{table*}[htb]
{\small
\hfill{}
\begin{tabular}{|l|l|c|c|c|c|c|c|c|}
\hline
Model/Field & \vtop{\hbox{\strut Invoice}\hbox{\strut Number}} & \vtop{\hbox{\strut Invoice}\hbox{\strut Amount}} & \vtop{\hbox{\strut Invoice}\hbox{\strut Date}} & \vtop{\hbox{\strut Vendor}\hbox{\strut Name}} & \vtop{\hbox{\strut Vendor}\hbox{\strut Address}} & \vtop{\hbox{\strut Line-item}\hbox{\strut Description}} & \vtop{\hbox{\strut Line-item}\hbox{\strut Quantity}} & \vtop{\hbox{\strut Line-item}\hbox{\strut Amount}} \\ \hline
sequential & 80.98\% & 79.13\% & 83.98\% & 28.97\% & 16.94\% & -0.01\% & -0.18\% & 0.22\% \\ \hline
image-only & 47.79\% & 68.91\% & 45.67\% & 19.68\% & 13.99\% & 49.50\% & 46.79\% & 63.49\% \\ \hline
chargrid-net & 80.48\% & 80.74\% & 83.78\% & 36.00\%& 39.13\%& 52.80\% & 65.20\% & 65.57\% \\ \hline
chargrid-hybrid-C32 & 74.85\% & 77.93\% & 80.40\% & 32.00\% & 31.48\% & 46.27\% & 64.04\% & 63.25\% \\ \hline
chargrid-hybrid-C64 & 82.49\% & 80.14\% & 84.28\% & 34.27\% & 36.83\% & 48.81\% & 64.59\% & 64.53\% \\ \hline
\end{tabular}
\hfill{}
\caption{Accuracy measure (c.f.~Sec.~\ref{sec:metric}) for an 8-class information extraction problem on invoices. The proposed chargrid models perform consistently well on all extracted fields compared to sequential and image models.}
\label{tab:results}
}\end{table*}
\end{center}
The chargrid-net, however, significantly outperforms the sequential model on multi-instance or multi-word fields where 2D relationships between text entities are important.
Examples of such fields are Line-item Description, Line-item Quantity and Line-item Amount, which are all grouped as \textit{sub-fields} to a specific line-item. Each of those fields may span a varying number of rows per line-item, see Figure~\ref{fig:sampleinvoices}.
The sequential model fails to correctly identify those line-item fields which is manifested in a negative accuracy measure (Sec.~\ref{sec:metric}). This implies that it is better to perform manual extraction over using automatic extraction. This is understandable since the line-item fields have a strong 2D structure that the sequential approach is not designed to capture.
In comparison with the image-only model, chargrid-net still performs much better. This is especially true for smaller fields which need to be \emph{read} to be accurately localized. On the other hand, for larger fields like Line-item Description, which can be localized by only vision, the gap is much smaller.
The values for the hybrid models are a bit more interesting. One could expect that combining two complementary inputs such as the chargrid representation and image - one capturing the content and the other capturing, e.g.~table delineations - would boost the accuracy.
It turns out, however, that at least in the case of our invoice dataset, the additional image encoder does not bring additional benefits.
Model chargrid-hybrid-C64, where the chargrid encoder branch has the same number of base channels $C=64$ as the original chargrid-net, is essentially as accurate as chargrid-net.
Model chargrid-hybrid-C32, where the image and chargrid encoder combined have the same number of channels (that is the chargrid encoder branch only has $C=32$ base channels), the accuracy is significantly reduced.
We conclude that at least in our present extraction problem, most of the discriminative information comes from the chargrid encoder branch and thus from the chargrid representation.
\sugg{Error examples of the chargrid-net model are depicted in Figure~\ref{fig:errorexamples}. One frequent error-type is that the model may fail to disentangle line-items, if they have a peculiar structure. While this is also challenging for human experts, other erroneous predictions are observed in samples for which the ground truth annotations are debatable.}
\section{Discussion}
We introduce a new way of modeling documents by using a character grid as document representation. The chargrid allows models to capture 2D relationships between characters, words, and larger units of text. The idea of the chargrid paradigm is inspired by the human perception which is heavily guided by 2D shapes and structures for understanding this type of documents.
Therefore, the chargrid allows to encode the positioning, size and alignment for textual components in a meaningful manner.
While the chargrid paradigm could be applied to various kinds of NLP tasks, we demonstrate its potential on an information extraction task from invoices. We train a deep neural network with an encoder-decoder architecture and we show that the network computes accurate segmentation masks and bounding boxes, which pinpoint the relevant information on the invoice.
We evaluate the accuracy of the model and compare it to state-of-the-art NLP and computer vision approaches. While those baseline models achieve accurate predictions for individual fields, only the chargrid performs well on all information extraction tasks. Some fields such as Invoice Number are relatively easy to detect for a model that operates on serialized text, as discriminative keywords are commonly preceding the words to be extracted. However, more complex extraction tasks (e.g. Vendor Address or Line-item Quantity) cannot be performed accurately as they require to exploit both, textual components and 2D layout structures. The traditional image-only computer vision model yields accurate predictions only for large visual columns (such as Line-item Amount) and it fails to locate extractions that require to understand the text.
\sugg{However, compared to traditional sequential neural NLP models, the benefits in accuracy come at a larger computational cost compared. Even though the proposed model is fully convolutional and parallelizes very well on a single (or even multiple) GPU(s), the added complexity introduced by using a 2D data representation can significantly increase the total data dimensionality. In our current use case, a chargrid-net training requires up to three days until full convergence, whereas our sequential model converges after only a few hours.}
Comparing chargrid to semantic segmentation on natural images, one should note that character-pixels, unlike pixels of a gray-scale image, are categorical. This requires the character-pixels to be encoded as 1-hot. This yields a highly sparse data representation. While such representation is very common for NLP problems, it is new for segmentation networks that have previously only been applied to image segmentation.
\section{Conclusion}
\sugg{Chargrid is a generic representation for 2D text. Using this as a base, one could solve any task on a 2D text such as document classification, named-entity recognition, information extraction, parts-of-speech tagging etc. Furthermore, chargrid is highly beneficial for scenarios where text and natural images are blended. One could imagine performing all the above NLP tasks also on such inputs. This work demonstrates the advantages of chargrid for an information extraction task but we believe that it is only the first step towards incorporating the 2D document structure into document understanding tasks. Some follow-up direction could be solving other NLP tasks on structured documents using chargrid or experimenting with other computer vision algorithms on chargrid. Furthermore, it may be interesting to use word embeddings rather than 1-hot encoded characters, i.e.~a wordgrid, as 2D text input representation.}
\section*{Acknowledgments}
In acknowledgement of the partnership between Nvidia and SAP this publication has been prepared using Nvidia DGX-1. We thank Zbigniew Jerzak and Markus Noga for their support and fruitful discussions.
|
\section{Introduction}
Top quark studies, that are currently driven by the Large Hadron
Collider (LHC) experiments ATLAS and CMS, play a major role in
deciphering the fundamental interactions. At the LHC top quarks are
mostly produced in pairs through strong interactions, but they can
also be created individually in single-top production via electroweak
interactions. Thus, depending on the production mode, the top quark
allows for different tests of the underlying forces. Both experiments
concentrate on the measurements of top quark properties like for
example the top quark mass $(m_t)$, the top quark width $(\Gamma_t)$,
the top quark charge $(Q_t)$, the total and differential cross
sections as well as the top quark spin correlations and the top quark
charge asymmetry $(A^{C}_{t\bar{t}})$, including differential top quark
charge asymmetries. High in the LHC program is the determination of
the top quark couplings to gauge bosons and the Standard Model (SM)
Higgs boson. Due to the large top quark mass various new physics
scenarios introduce modifications within the top quark sector. Some
examples include heavy new particles decaying into top quark pairs,
flavour changing neutral currents, anomalous missing transverse
momentum, same-sign top pair production or charged Higgs
production. Such new physics models can be tested by precise
measurements of top quark pairs, that are abundantly produced at the
LHC. Furthermore, top quark production, also with additional $b$- or
light jet(s), constitutes dominant irreducible backgrounds to many of
the searches for new physics processes. Thus, it is vital to
understand the properties and the characteristics of top production
and decay mechanisms. The level of precision available on the theory
side can have a huge impact on whether we can actually see the effects
of new physics.
As a result of the large collision energy at the LHC also more
exclusive final states, like for example $t\bar{t}\gamma$, have
started to be accessible \cite{Aad:2015uwa,Sirunyan:2017iyh,
Aaboud:2017era}. Even though the cross section for the
$t\bar{t}\gamma$ production process at the LHC is much smaller than
the cross section for the production of the top quark pair alone, the
former can provide key information on the strength and the structure
of the top quark coupling to the photon. Thus, it can for example
substantially constrain anomalous top quark couplings at the LHC,
see e.g. Ref.~\cite{Bylund:2016phk,Schulze:2016qas}. Regardless of the
applications, whether these are measurements within the SM or outside
of this framework, precise theoretical predictions are compulsory to
carry out such measurements.
To increase the theoretical precision for $pp\to t\bar{t}\gamma$
higher order corrections in QCD should be consistently
included. Moreover, the most accurate description of top quark decay
chains has to be incorporated. Appropriate calculations have recently
been made available. Specifically, a complete description of top quark
pair production in association with a hard photon in the di-lepton top
quark decay channel has been provided in
Ref. \cite{Bevilacqua:2018woc}. The calculations include factorisable
and non-factorisable contributions at NLO in QCD, that imply a cross
talk between production and decays of top quarks which require going
beyond the so-called Narrow Width Approximation (NWA). Specifically,
they include double resonant, single resonant and non-resonant Feynman
diagrams with respect to the top quark and $W$ gauge boson,
interferences among them as well as finite-width effects of $t$ and
$W$. With a fairly inclusive selection of cuts on the final states,
which are two $b$-jets, a hard photon, two charged leptons and the
missing transverse momentum, $p_T^{miss}$, the full $pp$ cross section
for a fixed renormalisation and factorisation scale choice receives
negative and moderate NLO QCD corrections of $10\%$. An assessment of
the uncertainties of theoretical origin left us with a $14\%$
theoretical error. Inclusion of a kinematic dependent scale, that
captures some parts of the unknown higher order effects, has improved
the situation yielding positive and small NLO corrections of
$2.5\%$. In this case the theoretical uncertainties resulting from
scale variations have been estimated at the level of $6\%$ only.
The impact of higher order corrections on differential distributions,
however, is much larger. For some observables, incidentally
important in searches for new physics, shape distortions of more than
$100\%$ have been observed. As expected for specific phase space
regions also the theoretical errors have increased substantially.
Improvement of the accuracy of theoretical predictions also at the
differential level to a few percent can be obtained by including
the next order in the perturbative expansion in $\alpha_s$. However, going
beyond NLO even for on-shell $t\bar{t}\gamma$ production seems to be a
formidable task at present. Inclusion of non-factorisable QCD
contributions at NNLO is simply difficult to imagine. Instead, a ratio
of cross sections can be studied since it may be significantly more stable
against radiative corrections and scale variations than the cross
sections themselves. Moreover, a ratio may reduce other theoretical
uncertainties like for example those stemming from parton distribution
functions. To this end a process that is under excellent
theoretical control must to be employed in the denominator of the
ratio. The $t\bar{t}$ production process, albeit in the same decay
channel, seems to be the best candidate for the job due to its large
cross section and similar behaviour with regard to radiative
corrections \cite{Denner:2010jp,Bevilacqua:2010qb, Denner:2012yc}.
Consequently, the following cross section ratio
\begin{equation}
{\cal R} = \frac{\sigma_{t\bar{t}\gamma}}{\sigma_{t\bar{t}}} \,,
\end{equation}
represents an interesting quantity to search for deviations from the SM
theory at the LHC. Moreover, a set of differential cross section
ratios can be constructed to look for any shape deviations from
those predicted within the SM
\begin{equation}
{\cal R}_X=\left(\frac{d\sigma_{t\bar{t}\gamma}}{dX}\right)
\left(\frac{d\sigma_{t\bar{t}}}{dX}\right)^{-1} \,.
\end{equation}
Here $X$ stands for the particular observable under consideration,
e.g. the invariant mass of two charged leptons, $m_{\ell \ell}$, the
invariant mass of two $b$-jets, $m_{b\bar{b}}$, etc. Since for a
realistic analysis specific cuts on top quark decay products need to
be imposed, a reliable description of top quark decays is mandatory for
both processes $pp \to t\bar{t}\gamma$ and $pp\to t\bar{t}$. In order
to avoid the introduction of additional unnecessary theoretical
uncertainties to the construction of the cross section ratio, the same
level of accuracy in the modelling of top quark decays must to be
employed in the numerator and denominator of ${\cal R}$ and ${\cal
R}_{X}$. Besides the modelling of top quark decays, where the
incorporation of radiative corrections is mandatory, a proper
renormalisation and factorisation scale choice has to be carefully
investigated. The scale choice should play an even greater role when
various differential cross section ratios are constructed. For the
latter phase space regions away from those dominated by
double resonant top quark contributions, which are sensitive to
non-factorizable QCD corrections, would be probed as well.
The purpose of this paper is twofold. First, we would like to provide
a systematic analysis of the two processes $pp\to t\bar{t}\gamma$ and
$pp\to t\bar{t}$ in the di-lepton top quark decay channel and extract
the most accurate NLO prediction for the total cross section
ratio. Such precise theoretical results can be used in comparisons
with the LHC data. The second goal of the paper is to examine whether
differential cross section ratios have enhanced predictive power for
new physics searches, by investigating possible correlations between
the two processes in various phase space regions in the quest of
reducing theoretical errors. Calculations for both processes will be
carried out with the same input parameters, parton distribution
functions (PDFs), jet algorithm and the same set of inclusive cuts up
to the cuts on the hard photon, which are present only in
the case of $t\bar{t}\gamma$ production. Finally, for both
processes factorisation and renormalisation scales will be
set to a common fixed value, whereas for a dynamical scale choice
scales as similar as possible will be selected. Cross section ratios
calculated in this way are free of additional and undesired theoretical
uncertainties that are introduced when different input parameters are
employed in the numerator and the denominator of ${\cal R}$. The size
of such additional theoretical uncertainties, however, must be
estimated. In various experimental analyses different Monte Carlo (MC)
programs are employed to provide theoretical predictions for $
t\bar{t}$ and $t\bar{t}\gamma$ production. Such general purpose MC
frameworks are often used by experimental collaborations with a default
set up, among others with a different scale choice and parton
distribution functions for $pp\to t\bar{t}$ and $pp\to
t\bar{t}\gamma$. Thus, in the paper we will quantify the impact of the
additional theoretical uncertainties coming from different theoretical
inputs. Finally, the stability of the cross section ratio with respect to
the transverse momentum cut on the hard photon will be examined. To
this end, theoretical predictions for $ t\bar{t}\gamma$ production
will be evaluated for two different values of the $p_{T,\gamma}$ cut.
The article is organised as follows. In Section \ref{section:2} the
\textsc{Helac-NLO} computational framework and input parameters used
in our studies are described. In Section \ref{section:3} the
normalised differential cross sections for off-shell $t\bar{t}\gamma$ and
$t\bar{t}$ production are provided in order to study the correlation
of the two processes. The results given there are used to understand
how the theoretical errors on the cross section ratios should be
estimated. NLO predictions for absolute cross sections are presented
in Section \ref{section:4} together with the theoretical uncertainties
from the scale dependence. Additionally, results with different parton
distribution functions are shown in Section \ref{section:4}. They are
calculated to estimate the size of theoretical uncertainties that come
from the parametrisation of parton distribution functions. In Section
\ref{section:5} we provide results for NLO cross section
ratios. Theoretical uncertainties are also discussed
there. Theoretical predictions for the differential cross section
ratios and their theoretical uncertainties are exhibited and discussed
in Section \ref{section:6}. Finally, in Section \ref{section:7} our
conclusions are laid out.
\section{Computational Framework and Input Parameters}
\label{section:2}
All the LO and NLO results for $e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b
\bar{b} \,\gamma$ and $e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b}$
production, which are presented in this paper, have been obtained with
the help of the \textsc{Helac-NLO} MC framework
\cite{Bevilacqua:2011xh}. The package comprises \textsc{Helac-1Loop}
\cite{vanHameren:2009dr} with \textsc{CutTools} \cite{Ossola:2007ax}
for the virtual corrections and \textsc{Helac-Dipoles}
\cite{Czakon:2009ss,Bevilacqua:2013iha} for the real emission
part. The \textsc{Helac-Dipoles} software deals with singularities
from soft or collinear parton emissions that are isolated via
subtraction methods for NLO QCD calculations. Specifically, the
commonly used Catani-Seymour dipole subtraction
\cite{Catani:1996vz,Catani:2002hc,Czakon:2009ss} and the so-called
Nagy-Soper subtraction scheme \cite{Bevilacqua:2013iha} are both
implemented in the \textsc{Helac-Dipoles} program and used in our
simulations. The integration over the phase space has been achieved
with the help of \textsc{Kaleu} \cite{vanHameren:2010gg}. For
unstable top quarks the complex mass scheme is utilised
\cite{Denner:1999gp,Denner:2005fg}. At the one loop level the
appearance of $\Gamma_t\ne 0$ in the propagator requires the
evaluation of scalar integrals with complex masses, which is supported
by the \textsc{OneLOop} program \cite{vanHameren:2010cp}.
Further details of these calculations can be found in our earlier work
on $pp\to t\bar{t}$ \cite{Bevilacqua:2010qb}, $pp\to t\bar{t}j$
\cite{Bevilacqua:2015qha,Bevilacqua:2016jfk} and $pp\to
t\bar{t}\gamma$ \cite{Bevilacqua:2018woc} where complete top quark
off-shell effects have also been consistently taken into account at
the NLO level in QCD. Specifically, in each case all resonant and
non-resonant Feynman diagrams, interferences and finite width effects
of the top quark as well as $W$ gauge bosons have been included. The
methods developed there have been straightforwardly adapted in the
current studies and, therefore, do not need a recollection. We refer
the interested readers to previously published results. In the
calculations of cross sections for $e^+ \nu_e \, \mu^- \bar{\nu}_\mu
\, b \bar{b} \,\gamma$ and $e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b
\bar{b}$ we employ the following SM parameters
\begin{equation}
\begin{array}{lcl}
G_F=1.166378 \cdot 10^{-5} ~{\rm GeV}^{-2} \,,
& \quad\quad \quad\quad & m_{t}=173.2 ~{\rm GeV} \,,\\[0.2cm]
m_{W}=80.385 ~{\rm GeV} \,,
&& \Gamma_{W} = 2.0988 ~{\rm GeV} \,,\\[0.2cm]
m_{Z}=91.1876 ~{\rm GeV}\,,
& & \Gamma_{Z} = 2.50782 ~{\rm GeV} \,,\\[0.2cm]
\Gamma_{t}^{\rm LO} = 1.47848 ~{\rm GeV} \,, & &
\Gamma_{t}^{\rm NLO} = 1.35159 ~{\rm GeV}\,.\\
\end{array}
\end{equation}
All other particles including bottom quarks are treated as
massless. Since leptonic $W$ gauge boson decays do not receive NLO QCD
corrections, to account for some higher order effects the NLO QCD
values for the gauge boson widths are used everywhere, i.e. for LO and
NLO matrix elements. The electromagnetic coupling $\alpha$ is
calculated from the Fermi constant $G_F$ in the $G_\mu$-scheme via
\begin{equation}
\alpha_{G_\mu}=\frac{\sqrt{2}}{\pi} \,G_F \,m_W^2
\left(1-\frac{m_W^2}{m_Z^2}\right)\,.
\end{equation}
For the emission of the isolated photon, however, $\alpha_{\rm QED} =
1/137$ is used instead. The running of the strong coupling constant
$\alpha_s$ with two-loop (one-loop) accuracy at NLO (LO) is provided
by the LHAPDF interface \cite{Buckley:2014ana}. The number of active
flavours is set to $N_F = 5$, however, contributions induced by the
bottom-quark parton density are neglected due to their numerical
insignificance. Following recommendations of PDF4LHC
\cite{Butterworth:2015oua} for the usage of parton distribution
functions (PDFs) suitable for applications at the LHC Run II we employ
CT14 \cite{Dulat:2015mca}, which is our default choice, MMHT14
\cite{Harland-Lang:2014zoa} and NNPDF3.0 \cite{Ball:2014uwa} PDFs.
Our calculation, like any fixed-order calculations, contains a
residual dependence on the renormalisation $(\mu_R)$ and the
factorisation scales $(\mu_F)$ arising from the truncation of the
perturbative expansion in $\alpha_s$. As a consequence, all
observables depend on the values of $\mu_R$ and $\mu_F$ that are
provided as input parameters. The theoretical uncertainty of the
total cross section, associated with neglected higher order terms in
the perturbative expansion, can be estimated by varying $\mu_R$ and
$\mu_F$ in $\alpha_s$ and in the PDFs. We assume that $\mu_R$ and
$\mu_F$ are set to a common value $\mu_R = \mu_F = \mu_0$. However,
the scale dependence is evaluated by varying $\mu_R$ and $\mu_F$
independently in the range
\begin{equation}
\label{s1}
\frac{1}{2} \, \mu_0 \le \mu_R\,,\mu_F \le 2 \,
\mu_0\,,
\end{equation}
with the additional condition
\begin{equation}
\label{s2}
\frac{1}{2} \le \frac{\mu_R}{\mu_F} \le 2 \,.
\end{equation}
In practice, such restrictions are equivalent to evaluating the
following scale variations
\begin{equation}
\label{s3}
\left(\frac{\mu_R}{\mu_0},\frac{\mu_F}{\mu_0}\right) = \left\{
(2,1),(0.5,1),(1,2),(1,1),(1,0.5),(2,2),(0.5,0.5)
\right\}\,.
\end{equation}
The final error is estimated from the envelope of the
resulting cross sections. For the central value of the scale, $\mu_0$,
we consider the fixed scale (the phase-space independent scale choice)
$\mu_0= m_{t}/2$ and the dynamic scale (the phase-space dependent
scale choice) $\mu_0=H_T/4$. The latter is defined on an
event-by-event basis according to
\begin{equation}
H_T=p_{T, \,e^+}+p_{T, \,\mu^-} +p_{T}^{miss}+ p_{T,\,b_1} +
p_{T,\, b_2} \,,
\end{equation}
where $p^{miss}_T$ denotes missing transverse momentum and
$p_{T,\,b_1}$, $p_{T,\, b_2}$ are transverse momenta of the two
$b$-jets. In the case of $pp \to t\bar{t}\gamma$ the transverse
momentum of the hard photon is also included into the definition of
$H_T$. Jets are constructed from final-state partons with
pseudo-rapidity $|\eta| <5$ with the help of the infrared safe {\it
anti}$-k_T$ jet algorithm \cite{Cacciari:2008gp} with the separation
parameter $R=0.4$. For $e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b
\bar{b}$ production exactly two $b$-jets, two charged leptons and
missing transverse momentum are required. Additionally, for the $ e^+
\nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b} \,\gamma$ production process
an isolated hard photon is requested. The latter is defined with
$p_{T,\gamma}>25$ GeV (our default transverse momentum cut on the hard
photon) and $|y_\gamma|<2.5$
\cite{Sirunyan:2017iyh,Aaboud:2017era}. To examine the stability of
our theoretical predictions at NLO in QCD we also present results
for the higher value of the $p_{T,\gamma}$ cut, namely for $p_{T,\gamma}>50$
GeV. To ensure infrared safety we use the photon isolation prescription
described in Ref. \cite{Frixione:1998jh} that is based on a modified
cone approach. The photon isolation condition is implemented in the
same way for quarks and gluons. For each parton $i$ we evaluate the
distance in the rapidity-azimuthal angle plane between this parton and
the photon, according to
\begin{equation}
\Delta
R_{\gamma i}=\sqrt{\Delta y_{\gamma i}^2+\Delta \phi_{\gamma i}^2
}=\sqrt{(y_\gamma-y_i)^2+(\phi_\gamma-\phi_i)^2}\,.
\\[0.2cm]
\end{equation}
We reject the event unless the following condition is fulfilled
\begin{equation}
\sum_{i} E_{T,\,i} \, \Theta(R - R_{\gamma i}) \le E_{T,\,\gamma} \left(
\frac{1-\cos(R)}{1-\cos(R_{\gamma j})}
\right)\,,\\[0.2cm]
\end{equation}
where $R\le R_{\gamma j}=0.4$ and $i$ runs over all partons. Moreover,
$E_{T,\,i}$ is the transverse energy of the parton $i$ and
$E_{T,\,\gamma}$ is the transverse energy of the photon. We apply all
other selection criteria to jets if and only if their separation from
the photon exceeds $R_{\gamma j}$. A jet reconstructed inside the cone
size $R_{\gamma j}$ is not subjected to any cuts. All final states
have to fulfil the subsequent selection criteria that mimic as closely
as possible the ATLAS and the CMS detector acceptances
\cite{Sirunyan:2017iyh,Aaboud:2017era}
\begin{equation}
\begin{array}{lclcl}
p_{T,\,\ell}>30 ~{\rm GeV} & \quad \quad \quad \quad
& p_{T,\,b}>40 ~{\rm GeV} &\quad \quad \quad \quad& p^{miss}_{T} >20
~{\rm GeV} \\[0.2cm]
|y_\ell|<2.5 && |y_b|<2.5 && \Delta R_{\ell \gamma}>0.4 \\[0.2cm]
\Delta R_{\ell b} > 0.4 && \Delta R_{bb}>0.4 &&
\Delta R_{\ell \ell} > 0.4\,.
\end{array}
\end{equation}
We set no restriction on the kinematics of the extra (non $b$-)jet.
\section{Differential Cross Sections at NLO in QCD}
\label{section:3}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{hist_rap_b12.pdf}
\includegraphics[width=0.49\textwidth]{hist_dr_bb.pdf}
\includegraphics[width=0.48\textwidth]{hist_rap_l12.pdf}
\includegraphics[width=0.49\textwidth]{hist_dr_ll.pdf}
\end{center}
\caption{\it Comparison of the normalised NLO differential cross
sections for $pp \to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b}
\,\gamma+X$ and $pp \to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b}+X$
at the LHC with $\sqrt{s}=13$ TeV. We present: the averaged rapidity
of the $b$-jet $(y_b)$, the distance in the azimuthal angle rapidity plane
between two $b$-jets ($\Delta R_{bb}$), the averaged rapidity of the
charged lepton $(y_\ell)$ as well as the distance in the azimuthal angle rapidity
plane between two charged leptons ($\Delta R_{\ell\ell}$). Results for
two different values of the transverse momentum cut on the hard
photon are shown. The NLO CT14 PDF set is employed and $\mu_R = \mu_F
= \mu_0 = m_t/2 $ is used.}
\label{fig:1}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{hist_pt_b12.pdf}
\includegraphics[width=0.49\textwidth]{hist_m_bb.pdf}
\includegraphics[width=0.49\textwidth]{hist_pt_l12.pdf}
\includegraphics[width=0.49\textwidth]{hist_m_ll.pdf}
\end{center}
\caption{\it Comparison of the normalised NLO differential cross
sections for $pp \to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b}
\,\gamma+X$ and $pp \to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b
\bar{b}+X$ at the LHC with $\sqrt{s}=13$ TeV. The following
distributions are shown: the averaged transverse momentum of the
$b$-jet $(p_{T,b})$, the invariant mass of two $b$-jets
$(m_{bb})$, the averaged transverse momentum of the charged
lepton $(p_{T,\ell})$ and the invariant mass of two charged leptons
$(m_{\ell\ell})$. Results for two different values of the transverse
momentum cut on the hard photon are shown. The NLO CT14 PDF set is
employed and $\mu_R = \mu_F = \mu_0 = m_t/2 $ is used.}
\label{fig:2}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{hist_m_tt.pdf}
\includegraphics[width=0.49\textwidth]{hist_pt_t12.pdf}
\includegraphics[width=0.49\textwidth]{hist_rap_t12.pdf}
\end{center}
\caption{\it
Comparison of the normalised NLO differential cross sections for $pp
\to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b} \,\gamma+X$ and $pp
\to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b}+X$ at the LHC with
$\sqrt{s}=13$ TeV. The top quark kinematics is shown. Specifically,
the invariant mass of the reconstructed $t\bar{t}$ system $(m_{t\bar{t}})$ as well
as the averaged transverse momentum $(p_{T,t})$ and rapidity $(y_t)$
of the top quark are depicted. Results for two different values of the
transverse momentum cut on the hard photon are shown. The NLO CT14 PDF
set is employed and $\mu_R = \mu_F = \mu_0 = m_t/2 $ is used.}
\label{fig:3}
\end{figure}
In this Section we present results for differential cross section
distributions for both processes: $pp \to e^+ \nu_e \, \mu^-
\bar{\nu}_\mu \, b \bar{b} \,\gamma +X$ at ${\cal O}(\alpha_s^3
\alpha^5)$ and $pp \to e^+ \nu_e \, \mu^- \bar{\nu}_\mu \, b \bar{b}
+X$ at ${\cal O}(\alpha_s^3 \alpha^4)$. They are obtained for the LHC
Run II energy of ${\sqrt{s} = 13}$ TeV. For brevity, we will refer to
these reactions as $pp\to t\bar{t}\gamma$ and $pp\to t\bar{t}$. To
understand similarities and potential differences between the two
production processes, it is helpful to identify the dominant partonic
subprocesses. In both cases the most important production mechanism is
via scattering of two gluons. With our selection of cuts, the gg
channel contributes $79\%$ ($88\%$) to the LO $pp\to t\bar{t}\gamma$
($pp\to t\bar{t}$) cross section while the $q\bar{q}+\bar{q}q$
channels account for $21\%$ ($12\%$). The dominance of the $gg$
production process in both cases suggests that $pp \to t\bar{t}$ and
$pp\to t\bar{t}\gamma$ should show similar features in the kinematics
of the final states, i.e. two charged leptons, the missing transverse
momentum and two $b$-jets. All differential cross sections that are
presented in the following have been obtained for the CT14 PDF
set. For both production processes we use the kinematic-independent
factorisation and renormalisation scales $\mu_R=\mu_F=\mu_0$ with the
central value $\mu_0=m_t/2$ rather than simply $\mu_0=m_t$. Even
though the mass of the heaviest particle appearing in the process
seems to be a more natural option, the $\mu_0=m_t/2$ scale choice is
very well motivated by the fact that $pp\to t\bar{t}$ at the LHC is
dominated by $t$-channel gluon fusion, which favours smaller values of
the scale. Additionally, effects beyond NLO that include soft-gluon
resummation for the hadronic cross-section at next-to-leading
logarithmic accuracy are smaller for $\mu_0=m_t/2$ than for
$\mu_0=m_t$ \cite{Bonciani:1998vc, Czakon:2016dgf} as we have
explicitly checked with the help of the \textsc{Top++} program
\cite{Czakon:2011xx}. From the QCD point of view both processes $pp
\to t\bar{t}$ and $pp \to t\bar{t}\gamma$ are similar, which
motivates our scale choice for $pp \to t\bar{t}\gamma$ as well.
We start with a collection of angular cross section distributions
that are given in Figure \ref{fig:1}. Specifically, we present the
averaged rapidity distribution of the $b$-jet and the charged lepton
as well as the separation in the rapidity-azimuthal angle plane
between the two $b$-jets, $\Delta R_{bb}$, and between the two charged
leptons, $\Delta R_{\ell\ell}$. All differential cross section
distributions are normalised to the corresponding absolute cross
sections to illustrate shape similarities and differences between the
two processes. For the $pp\to t\bar{t}$ process the two $b$-jets are
emitted in the central regions (in rapidity) of the detector. This is
a consequence of the dominance of the $gg$ production channel, which
favours emissions of jets at smaller rapidities. Charged leptons are
also produced mostly centrally. Moreover, both charged leptons
$\ell^\pm$ and $b$-jets are preferably produced in the back-to-back
configurations. Hereby, $b$-jets come more often from top quark
decays rather than from the $g\to b\bar{b}$ splitting. The latter
configuration would manifest itself in the enhancement in the lower
values of $\Delta R_{bb}$. Singularities stemming from the collinear
$g\to b\bar{b}$ splitting are, however, screened off by the
(effective) invariant mass cut of $m_{b\bar{b}} \gtrsim 16$ GeV. The
latter is implied once the $\Delta R_{bb}$ separation between the two
$b$-jets of $0.4$ is introduced by the jet algorithm together with the
requirement of having both $b$-jets with transverse momentum larger
than $40$ GeV. For the two charged leptons the situation is rather
simplified due to the fact that we simulate decays of the weak bosons
to different lepton generations only, thus, virtual photon
singularities stemming from collinear $\gamma \to \ell^+ \ell^-$
decays are avoided. As might be observed in Figure \ref{fig:1} we can
not see large shape differences in dimensionless observables, when the
emission of the additional hard photon is included. This is in line
with our expectation that the $t\bar{t}$ and $ t\bar{t}\gamma$ production
processes are similar from the QCD point of view. All the kinematical
features described above are insensitive to the $p_{T,\gamma}$ cut as
can be additionally observed in Figure \ref{fig:1} since results for
two cases $p_{T,\gamma} > 25$ GeV and $p_{T,\gamma} > 50$ GeV are
plotted.
In the next step, we consider dimensionful observables like for example
the averaged transverse momentum of the $b$-jet, the averaged
transverse momentum of the charged lepton as well as the invariant
mass of the two charged leptons and the two $b$-jets. They are
collected in Figure \ref{fig:2}. Again shapes of all
observables are not affected by the hard photon emissions. In the case
of $pp \to t\bar{t}\gamma$ all plotted spectra are slightly harder.
However, this is a consequence of the additional $p_{T, \, \gamma}$
cut that effectively sets higher transverse momentum thresholds on the
whole $t\bar{t}$ system, thus, consequently on all top quark decay
products. Overall, for $t\bar{t}$ and $t\bar{t}\gamma$ production
similarities in the jet activity and the way charged leptons are
produced could be observed.
Subsequently, we turn our attention to the common underlying
$t\bar{t}$ kinematics. To this end, in Figure \ref{fig:3} we depict
the invariant mass of the $t\bar{t}$ system as well as the averaged
transverse momentum and rapidity of the top quark. We note here, that
top quarks are reconstructed from their decay products assuming exact
reconstruction of the $W$ gauge boson. Specifically, we have defined
$p(t) = p(b)+p(e^+)+ p(\nu_e)$ and $p(\,\bar{t}\,)=p(\bar{b})+p(\mu^-)
+p(\bar{\nu}_\mu)$, where $b$ and $\bar{b}$ denotes the $b$-jets. We
could observe harder spectra for the averaged transverse momentum of
the top quark and for the invariant mass of the $t\bar{t}$ system in
the case of the $ t\bar{t}\gamma$ production process as compared to
the corresponding distributions for $t\bar{t}$ production. Since we
consider the whole reconstructed top quark system, not only its decay
products separately, the higher transverse momentum threshold set by
the $p_{T,\gamma}$ cut is more pronounced here. Moreover, for both
processes the top quarks are predominantly produced in the central
rapidity regions and in the back-to-back configuration.
To summarise this part, let us repeat that as anticipated both
$t\bar{t}\gamma$ and $t\bar{t}$ production processes are highly
correlated. This fact will be exploited in the next section when the
theoretical error for the $t\bar{t}\gamma$ and $t\bar{t}$ cross section
ratio will be estimated. Additionally, conclusions drawn here are
independent of the $p_{T,\gamma}$ cut. Furthermore,
they are not modified when the dynamical scale choice
($\mu_R=\mu_F=\mu_0=H_T/4$) is used instead for both processes or when
different PDF sets (MMHT14 or NNPDF3.0) are employed.
\section{Absolute Cross Sections at NLO in QCD}
\label{section:4}
\begin{table}[t!]
\begin{center}
\begin{tabular}{c||c c||c}
&&&\\
PDF set, $\mu_R=\mu_F=\mu_0$ &
$\sigma^{\rm NLO}_{e^+\nu_e \mu^- \bar{\nu}_\mu
b\bar{b}}$ [fb]
& $\sigma^{\rm NLO}_{e^+\nu_e \mu^- \bar{\nu}_\mu b\bar{b}\gamma}$
[fb]
&$\sigma^{\rm NLO}_{e^+\nu_e \mu^- \bar{\nu}_\mu b\bar{b}\gamma}$
[fb] \\[0.2cm]
&&$p_{T,\gamma} > 25$ GeV& $p_{T,\gamma} > 50$ GeV\\[0.2cm]
\hline \hline
&&&\\
CT14, $\mu_0=m_t/2$ & $1629.4^{\,\,\,+18.4\,(1\%)}_{-144.7\,(9\%)}$
& $7.436^{+0.074\,\,\,\,(1\%)}_{-1.034 \,(14\%)}$
& $3.081^{+0.050\,\,\,\,(2\%)}_{-0.514 \,(17\%)}$ \\[0.2cm]
CT14, $\mu_0=H_T/4$ & $1620.5^{\,\,\,+21.6\,\,(1\%)}_{-118.8\,\,(7\%)}$
&$7.496^{+0.099\,\,\,(1\%)}_{-0.457\,\,\,(6\%)}$
& $3.125^{+0.040\,\,\,(1\%)}_{-0.142\,\,\,(4\%)}$ \\[0.2cm]
\hline \hline
&&&\\
MMHT14, $\mu_0=m_t/2$ & $1650.5^{\,\,\, +17.0 \,(1\%)}_{-152.7\, (9\%)}$
& $7.490^{+0.080\,\,\,\,(1\%)}_{-1.081\, (14\%)}$
& $3.093^{+0.053\,\,\,\,(2\%)}_{-0.535\, (17\%)}$ \\[0.2cm]
NNPDF3.0, $\mu_0=m_t/2$& $1695.0^{ \, \, \,+18.4 \,(1\%)}_{-153.3\,
(9\%)}$& $7.718^{+0.078\,\,\,\, (1\%)}_{-1.102 \, (14\%)}$
& $3.195^{+0.054\,\,\,\,(2\%)}_{-0.550\,(17\%)}$\\[0.2cm]
\end{tabular}
\end{center}
\caption{\label{tab:3} \it NLO cross sections for $pp\to e^+\nu_e
\mu^- \bar{\nu}_\mu b\bar{b}+X$ and $pp\to e^+\nu_e \mu^-
\bar{\nu}_\mu b\bar{b}\gamma+X$ at the LHC with $\sqrt{s}=13$
TeV. Also included are theoretical errors as obtained from the scale
variation. In the case of $pp \to e^+\nu_e \mu^- \bar{\nu}_\mu
b\bar{b}\gamma+X$ results for two different values of the
$p_{T,\gamma}$ cut are given. Various PDF sets are employed.}
\end{table}
In this Section we present predictions for $pp\to e^+\nu_e
\mu^-\bar{\nu}_\mu b\bar{b}$ and $pp\to e^+\nu_e \mu^-\bar{\nu}_\mu
b\bar{b}\gamma$ at the LHC with $\sqrt{s} = 13$ TeV. NLO QCD cross
sections are shown in Table \ref{tab:3} together with their
theoretical errors from the scale dependence. Results are presented
for the following two values of the transverse momentum cut on the
hard photon $p_{T,\,\gamma} > 25$ GeV and $p_{T,\,\gamma} >50$
GeV. The default CT14 PDF set is employed together with two additional
PDF sets, namely MMHT14 and NNPDF3.0. Moreover, the following two
scale choices, $\mu_0=m_t/2$ and $\mu_0=H_T/4$, are studied. In the
first step we examine results that we have obtained for the CT14 PDF
set. Looking at the total cross sections, which are mostly influenced
by final state production relatively close to the $t\bar{t}$
threshold, both scale choices are in equally good shape since the
results agree well within the corresponding theoretical
errors. However, the size of the theoretical uncertainties, especially
in the case of $t\bar{t}\gamma$ production, does depend on the scale
choice. The latter finding tells us that the absolute cross sections
for $t\bar{t}$ and $t\bar{t}\gamma$ production in the di-lepton top
quark decay channel with the selection of cuts that we have imposed
are not as inclusive observables as one would expect. Specifically,
for $\mu_0=m_t/2$ the NLO theoretical uncertainties for the $pp\to
t\bar{t}\gamma$ process are of the order of $14\%$ for $p_{T,\gamma}>
25$ GeV and $17\%$ for $p_{T,\gamma}> 50$ GeV. In the case of
$t\bar{t}$ production theoretical uncertainties, as
obtained from Eq.~\eqref{s3}, of the order of
$9\%$ have been estimated. For the dynamical scale choice in each case
the theoretical uncertainties are well below $10\%$. Specifically, our
judicious dynamical scale choice has allowed us to obtain $7\%$ for
$t\bar{t}$ production and $6\%$ for $t\bar{t}\gamma$ production with
$p_{T,\gamma}> 25$ GeV. In the latter case an increase of the
transverse momentum cut to $50$ GeV has resulted in the smaller
theoretical error of $4\%$. These facts suggest that the proposed
dynamical scale efficiently describes the multi-scale kinematics of
the process. Let us note at this point, that should we
instead vary $\mu_R$ and $\mu_F$ simultaneously, up and down by a
factor of $2$ around $\mu_0$, the uncertainties would remain
unchanged. This is due to the fact that the scale variation is driven
solely by the changes in $\mu_R$, see
Ref.~\cite{Bevilacqua:2018woc}.
Before discussing results for other PDF sets let us remind the reader
in this place that in the case of on-shell $t\bar{t}$ and
$t\bar{t}\gamma$ production for stable top quarks the size of the
theoretical error as obtained from the scale dependence is not
substantially reduced when the dynamical scale choice is used instead
of the fixed one, of course as long as this scale is properly
selected. To better outline this conclusion, we show the NLO QCD
results for on-shell $t\bar{t}$ and $t\bar{t}\gamma$ production at the
LHC, that we denote with a special index {\it ``(on-shell)''} to
distinguish them from the results with top quark and $W$ gauge boson
decays and off-shell effects included. Results are
generated with the same input parameters as given in Section
\ref{section:2}. For the on-shell $t\bar{t}$ sample we do not apply
any kinematical cuts, while in the case of $t\bar{t}\gamma$ we apply
cuts on the hard photon only, following the same criteria used for the
off-shell case. Namely, we ask for $p_{T,\,\gamma}> 25$ GeV,
$|y_\gamma|<2.5$ and $R_{\gamma j}=0.4$ in the photon isolation
condition. In the LO-like configurations the latter condition is
translated to the simpler $\Delta R_{\gamma j} > 0.4$ cut where $j$
stands for all light partons including bottom quarks. Additionally,
we present these results for the following two scale choices
$\mu_R=\mu_F=\mu_0=m_t/2$ and $\mu_R=\mu_F= \mu_0=E_T/4$. The
dynamical scale choice, which is defined as
\begin{equation} E_T=\sqrt{p_T^2(t)+m_t^2} +
\sqrt{p_T^2(\bar{t\,})+m_t^2}\,,
\end{equation}
is similar to our previous choice $\mu_0=H_T/4$\footnote{For example
for $\mu_R=\mu_F=\mu_0$ set to $\mu_0=E_T/4$ we have obtained the
following results for $e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}$ and
$e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}\gamma$ production
\begin{equation}
\begin{split}
\sigma^{\rm NLO}_{e^+\nu_e \mu^-\bar{\nu}_\mu
b\bar{b}} \, (\mu_0=E_T/4,{\rm CT14}) &= {1628.4}^{+ 19.7 \, (1\%)}
_{-69.9 \, (4\%)}\, {\rm fb}\,,\\[0.2cm]
\sigma^{\rm NLO}_{e^+\nu_e \mu^-\bar{\nu}_\mu
b\bar{b}\gamma} \,(\mu_0=E_T/4,{\rm CT14}) &= {7.524}^{+ 0.106 \, (1\%)}
_{-0.393 \, (5\%)} \, {\rm fb}\,.\\[0.2cm]
\end{split}
\end{equation}
In the latter case the $p_{T,\gamma} > 25$ GeV cut has been applied
on the hard photon. }. For obvious reasons the latter can not be
applied for the on-shell $t\bar{t}$ and $t\bar{t}\gamma$
production. Once more in the case of $pp \to t\bar{t}\gamma$ the
transverse momentum of the hard photon, $p_{T,\gamma}$, has been
added to the definition of $E_T$. Our results for top quark pair
production can be summed up as
\begin{equation}
\begin{split}
\sigma^{\rm NLO \, (on-shell)}_{t\bar{t}}(\mu_0=m_t/2, {\rm CT14})
&= 797.07^{+65.88 \, (\,\,\,8\%)}_{-82.41 \, (10\%)} \, {\rm pb} \,,\\[0.2cm]
\sigma^{\rm NLO \, (on-shell)}_{t\bar{t}}(\mu_0=E_T/4, {\rm CT14})
&= 770.11^{+74.61 \, (10\%)}_{-83.92\, (11\%)} \, {\rm pb} \,. \\[0.2cm]
\end{split}
\end{equation}
For $pp\to t\bar{t}\gamma$, on the other hand, we have obtained
%
\begin{equation}
\begin{split}
\sigma^{\rm NLO\, (on-shell)}_{t\bar{t}\gamma}(\mu_0=m_t/2, {\rm CT14},
p_{T,\gamma} > 25 ~{\rm GeV})
&= 2.035^{+0.137 \, (\,\,\,7\%)}_{-0.211 \, (10\%)} \, {\rm pb} \,, \\[0.2cm]
\sigma^{\rm NLO\, (on-shell)}_{t\bar{t}\gamma }(\mu_0=E_T/4, {\rm CT14},
p_{T,\gamma} > 25 ~{\rm GeV})
&= 1.901^{+0.209\, (11\%)}_{-0.227\, (12\%)} \, {\rm pb} \,. \\[0.2cm]
\end{split}
\end{equation}
The theoretical uncertainties for the on-shell $t\bar{t}$ and
$t\bar{t}\gamma$ production process are at the level of $10\%-11\%$
for $pp\to t\bar{t}$ and $10\%-12\%$ for $pp\to t\bar{t}\gamma$.
The theoretical uncertainties as obtained from the scale dependence of
the studied cross sections are, however, not the only source of
systematic uncertainties. Another source of theoretical uncertainties
is associated with the parameterisation of PDFs. Thus, we have given
in Table \ref{tab:3} NLO results for two additional PDF sets MMHT14
and NNPDF3.0 for $\mu_0=m_t/2$. In this way the various theoretical
assumptions that enter into the parameterisation of the PDFs, which are
difficult to quantify within a given scheme, are assessed. When
comparing CT14 results for $\sigma^{\rm NLO}_{t\bar{t}\gamma}$ and
$\sigma^{\rm NLO}_{t\bar{t}}$ with the corresponding numbers for
MMHT14 and NNPDF3.0 we observe that the PDF uncertainties for NLO
cross sections are of the order of $1\%$ for MMHT14 for both
production processes. In the case of the NNPDF3.0 set they are at the
level of $4\%$. Taken very conservatively as the maximum of MMHT14 and
NNPDF3.0 results PDF uncertainties for $t\bar{t}$ and $t\bar{t}\gamma$
are estimated to be of the order of $4\%$. We have also
performed the individual estimates of PDF systematics. We have
followed the prescription of each PDF collaboration in order to
provide the $68\%$ confidence level (C.L.) PDF uncertainties. Both
CT14 and MMHT14 include a central set and error
sets in the Hessian representation. More precisely, there are $2N= 56$
and $2N = 50$ eigenvector PDF members for CT14 and MMHT14
respectively that we have employed to the asymmetric expression for
PDF uncertainties as described for example in
Ref~\cite{Buckley:2014ana}. Let us note at this point that the CT14
errors are rescaled by a factor $1/1.645$ since they are originally
provided only at $90\%$ C.L. On the other hand for the NNPDF3.0 PDF
set, which uses the Monte Carlo sampling method in conjunction with
neural networks, PDF uncertainties are obtained using the replicas
method, see e.g. Ref~\cite{Buckley:2014ana}. In this case a set of
$N=100$ Monte Carlo PDF members has been used to extract PDF
uncertainties from the the NNPDF3.0 PDF set. Our findings for the
$e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}\gamma$ production process can be
summarised as follows
\begin{equation}
\begin{gathered}
\sigma_{e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}\gamma}^{\rm NLO}(\mu_0=m_t/2,
{\rm CT14}, p_{T,\,\gamma} > 25 \,{\rm GeV}) = 7.436^{+0.220 \,
(3\%)}_{-0.235\, (3\%)}
\, {\rm fb} \,,\\[0.2cm]
\sigma_{e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}\gamma}^{\rm NLO}(\mu_0=m_t/2,
{\rm MMHT14}, p_{T,\,\gamma} > 25 \,{\rm GeV}) = 7.490^{+0.160\,
(2\%)}_{-0.143\, (2\%)} \, {\rm fb} \,,\\[0.2cm]
\sigma_{e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}\gamma}^{\rm NLO}(\mu_0=m_t/2,
{\rm NNPDF3.0}, p_{T,\,\gamma} > 25 \,{\rm GeV}) = 7.718 \pm 0.106 \, (1\%)\,
{\rm fb} \,.
\end{gathered}
\end{equation}
On the other hand for $e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}$
production we have obtained
\begin{equation}
\begin{gathered}
\sigma_{e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}}^{\rm NLO} \, (\mu_0=m_t/2,
{\rm CT14}) = 1629.4^{+44.6 \, (3\%)}_{-49.3\,(3\%)} \, {\rm fb} \,,\\[0.2cm]
\sigma_{e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}}^{\rm NLO} \, (\mu_0=m_t/2,
{\rm MMHT14}) = 1650.5^{+35.1\, (2\%)}_{-33.1\, (2\%)} \, {\rm fb} \,,\\[0.2cm]
\sigma_{e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}}^{\rm NLO} \, (\mu_0=m_t/2,
{\rm NNPDF3.0}) = 1695.0 \pm 26.0 \, (1.5\%)\, {\rm fb} \,.
\end{gathered}
\end{equation}
Overall, we can observe that the size of the (internal) PDF
uncertainties for each PDF set separately is of the order of
$1\%-3\%$, therefore, smaller than the difference between various PDF
sets, which is at the level of $1\%-4\%$. Nevertheless, for the total
$e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}$ and $e^+\nu_e
\mu^-\bar{\nu}_\mu b\bar{b}\gamma$ cross sections the PDF
uncertainties are below the theoretical uncertainties due to scale
dependence, which remain the dominant source of the theoretical
systematics.
\section{Cross Section Ratios at NLO in QCD}
\label{section:5}
In the following, we study the cross section ratios. Our main goal
here is to verify whether even further improvement in the accuracy of
theoretical predictions can be obtained. More precisely we would like
to see if theoretical uncertainties below $10\%$ can be obtained for
the fixed scale choice. On the other hand, in the case of the
dynamical scale choice, that has been adopted for these studies, we
would like to determine whether a few percent precision, i.e.
comparable accuracy to that of NNLO calculations for $
t\bar{t}$ production \cite{Baernreuther:2012ws,Czakon:2013goa}, might
be achieved. To this end results for ${\cal R}=\sigma_{
t\bar{t}\gamma}/\sigma_{t\bar{t}}$ cross sections ratio for the $p_{T,
\gamma}$ cut of $p_{T, \gamma} > 25$ GeV and $p_{T, \gamma} > 50$ GeV
are provided. They are constructed with the help of the absolute cross
sections that are collected in Table \ref{tab:3}. The theoretical
error for the cross section ratio is estimated by calculating
\begin{equation}
{\cal R}= \frac{\sigma^{\rm
NLO}_{t\bar{t}\gamma} \, (\mu_1)}{\sigma^{\rm NLO}_{
t\bar{t}} \, (\mu_2)}\,,
\end{equation}
where $\mu_1=\mu_2=\mu_0$ and due to correlation of $pp\to
t\bar{t}\gamma$ and $pp\to t\bar{t}$ only the following combinations
are considered, see e.g. Ref.~\cite{Bevilacqua:2014qfa}
\begin{equation}
\label{s4}
\left(\frac{\mu_1}{\mu_0}, \frac{\mu_2}{\mu_0}\right) = \left\{(2, 2),
(0.5, 0.5)\right\}\,.
\end{equation}
Specifically, we use values of cross sections from Table \ref{tab:3}
also for the scale dependence. The latter have been estimated with the
help of Eq.~\eqref{s3}. Nevertheless, since the scale variation in
total cross sections is driven solely by the changes in $\mu_R$,
see Ref.~\cite{Bevilacqua:2018woc}, this is equivalent
to employing Eq.~\eqref{s4}. For $p_{T, \gamma} > 25$ GeV we have
obtained the following results for ${\cal R}$ at NLO in QCD
\begin{equation}
\begin{split}
{\cal R} \left( \mu_0 = m_t/2, {\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &= (4.56 \pm 0.25) \cdot 10^{-3} \, (5\%) \,,\\[0.2cm]
{\cal R}\left(\mu_0 =H_T/4,{\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &= ( 4.62 \pm 0.06) \cdot 10^{-3} \, (1\%) \,, \\[0.2cm]
\end{split}
\end{equation}
while for $p_{T, \gamma} > 50$ GeV our findings can be summarised as
follows
\begin{equation}
\begin{split}
{\cal R}(\mu_0=m_t/2,{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV} ) &= (1.89\pm 0.16) \cdot
10^{-3} \, (8\%)\,,\\[0.2cm]
{\cal R}(\mu_0=H_T/4,{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV} ) &= (1.93 \pm 0.06)\cdot 10^{-3} \,
(3\%)\,. \\[0.2cm]
\end{split}
\end{equation}
The observed change in the value of ${\cal R}$ for the
scale variation is truly asymmetric. Specifically, one of the
estimated values is always below the quoted precision. Thus, similarly
as for absolute cross sections, theoretical errors, which are provided
as well, are taken very conservatively as a maximum of these two
results. Ratio results for our default $p_{T, \gamma}$ cut of $25$
GeV for the two different scale choices are in perfect agreement
within theoretical errors that are provided. This outcome is not
affected by a higher value of the $p_{T, \gamma}$ cut, albeit, the
absolute value of the ratio is smaller in the latter case. We notice
that for $\mu_0=m_t/2$ theoretical uncertainties from the scale
variation are, indeed, below $10\%$, i.e. they are at the level of
$5\%$ and $8\%$ respectively for the $p_{T, \gamma}$ cut of $25$ GeV
and $50$ GeV. For $\mu_0=H_T/4$, however, theoretical errors are
substantially reduced down to $1\%$ and $3\%$. Such precision is
comparable to the precision one would rather expect from NNLO QCD
results for top quark pair production. Thus, the ratio of
$t\bar{t}\gamma$ and $t\bar{t}$ cross sections represents a very
precise observable to be used at the LHC. One of the possible
applications might be the measurement of the strength and the
structure of the $t$-$\bar{t}$-$\gamma$ vertex in $t\bar{t}\gamma$
production. The latter could shed some light on possible new physics
that can reveal itself only once sufficiently precise theoretical
predictions are available.
Once again, if on-shell $t\bar{t}$ and $t\bar{t}\gamma$ production is
employed to construct the cross section ratio
\begin{equation}
{\cal R}^{\rm
on-shell}=\frac{\sigma^{\rm NLO \, {\rm
(on-shell)}}_{t\bar{t}\gamma}}{\sigma^{\rm NLO\, {\rm (on-shell)}}_{
t\bar{t}}} \,,
\end{equation}
no substantial reduction in the theoretical uncertainties
could be observed when replacing $\mu_R=\mu_F=\mu_0=m_t/2$ with
$\mu_R=\mu_F=\mu_0=E_T/4$. Indeed, we can write
\begin{equation}
\begin{split}
{\cal R}^{\rm on-shell} \left( \mu_0 = m_t/2, {\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &=
\left( 2.55 \pm 0.04\right) \cdot 10^{-3} \, (2\%)\,,
\\[0.2cm]
{\cal R}^{\rm on-shell}\left(\mu_0 =E_T/4,{\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &=
\left( 2.47 \pm 0.03 \right) \cdot 10^{-3} \, (1\%)\,.
\\[0.2cm]
\end{split}
\end{equation}
It is worth mentioning at this point that, the theoretical error for
${\cal R}^{\rm on-shell}$ as calculated from the scale dependence is
at the $2\%$ level already for the fixed scale choice. From the
experimental point of view, however, measurements in the phase space
regions defined by the specific selection cuts that simulate as
closely as possible detector response are more appropriate, simply
because such measurements do not introduce additional and unnecessary
uncertainties due to model-dependent extrapolations to parton level
$t$ and $\bar{t}$ objects and to phase-space regions outside the
detector sensitivity. Having on-shell results at hand we can also
study the impact of top quark decays on the cross section ratio. We
note that the central value of ${\cal R}^{\rm on-shell}$ is
smaller by a factor of $1.8$ when comparing to ${\cal R}$. The cuts
on the final state decay products in conjunction with hard photon
emission from $b$-jets and charged leptons modify the ratio
substantially. Since the set of selection cuts is different in both
cases there is no particular reason why one would expect ${\cal
R}^{\rm on-shell}$ and ${\cal R}$ to be equal.
To assess the PDF uncertainties, we have recalculated the
${\cal R}$ observable for two different PDF sets, namely MMHT14 and
NNPDF3.0 with $\mu_0=m_t/2$. For $p_{T,\gamma} > 25$ GeV theoretical
predictions for ${\cal R}$ are given by
\begin{equation}
\begin{split}
{\cal R}\left(\mu_0=m_t/2,{\rm MMHT14}, p_{T,\gamma} > 25 \, {\rm
GeV}\right)
&=(4.54\pm 0.26) \cdot
10^{-3} \,(6\%)\,,\\[0.2cm]
{\cal R}\left(\mu_0=m_t/2,{\rm NNPDF 3.0}, p_{T,\gamma} > 25 \, {\rm
GeV}\right) & = (4.55 \pm 0.26) \cdot
10^{-3}\, (6\%)\,,\\[0.2cm]
\end{split}
\end{equation}
while for $p_{T,\gamma} > 50$ GeV we have found instead
\begin{equation}
\begin{split}
{\cal R}(\mu_0=m_t/2,{\rm MMHT14}, p_{T,\gamma} > 50 \, {\rm
GeV}) &= (1.87 \pm 0.17 )\cdot 10^{-3}
\, (9\%) \,,\\[0.2cm]
{\cal R}\left(\mu_0=m_t/2,{\rm NNPDF 3.0}, p_{T,\gamma} > 50 \, {\rm
GeV}\right) & = (1.88 \pm 0.17 ) \cdot
10^{-3}\, (9\%)\,.\\[0.2cm]
\end{split}
\end{equation}
We have estimated the size of the PDF uncertainties to be
$\pm \, 0.02 \cdot 10^{-3}$ independently of the $p_{T,\gamma}$ cut.
Thus, they are below $0.5\%$ for $p_{T,\gamma} > 25$ GeV and of the
order of $1\%$ for $p_{T,\gamma} > 50$ GeV. For our best
NLO QCD predictions for the ${\cal R}$ observable with $\mu_0=H_T/4$
theoretical uncertainties due to scale dependence are $\pm \, 0.06
\cdot 10^{-3}$, thus a factor of $3$ larger than the PDF
uncertainties. We have additionally computed internal PDF
uncertainties for the cross section ratios. We have obtained $\pm \,
0.02\cdot 10^{-3}$ for MMHT2014 and NNPDF3.0 as well as $\pm \, 0.04\cdot
10^{-3}$ for CT14. In the latter case PDF uncertainties are of the
order of $1\%-2\%$ depending on the $p_{T,\gamma}$ cut, however, still
a factor of $1.5$ smaller than theoretical errors from the scale
dependence for $\mu_0=H_T/4$. Finally, our best NLO QCD predictions
for the ${\cal R}$ observable at the LHC with $\sqrt{s}=13$ TeV are
given by
\begin{equation}
\begin{split}
{\cal R}\left(\mu_0 =H_T/4,{\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &= \left( 4.62 \pm 0.06\, {\rm [scales]} \pm 0.04 \,
{\rm [PDFs]} \right) \cdot 10^{-3}
\\[0.2cm]
{\cal R}(\mu_0=H_T/4,{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV} ) &= (1.93 \pm 0.06 \, {\rm [scales]}\pm 0.04 \, {\rm [PDFs]})
\cdot 10^{-3} \,, \\[0.2cm]
\end{split}
\end{equation}
where we have included theoretical errors both from the scale
dependence and from the PDFs. For the latter the internal PDF
uncertainties as obtained for the default CT14 PDF set are
quoted. Likewise for the ${\cal R}$ observable the dominant source of
theoretical systematics is associated with the scale dependence.
We would like to note here that a meaningful theoretical error on
${\cal R}$ coming from the scale variation can be calculated for the
first time only at NLO in QCD. At LO theoretical predictions for
${\cal R}$ for the fixed scale choice and for the CT14 PDF set with
$p_{T,\gamma} > 25$ GeV are given by
\begin{equation}
{\cal R}(\mu_0=m_t/2, \text{CT14}, p_{T,\gamma} > 25 \, {\rm GeV})
= \left(4.94 \pm 0.08 \right)\cdot 10^{-3} \, (2\%)\,.
\end{equation}
The scale variation of ${\cal R}$ at LO is much smaller than at NLO.
In the latter case we have obtained $5\%$ instead. Since at LO we
generate $pp\to t\bar{t}\gamma$ at ${\cal O}(\alpha_s^2\alpha^5)$ and
$pp\to t\bar{t}$ at ${\cal O}(\alpha_s^2\alpha^4)$ we have the same
order in $\alpha_s$ for both production processes and the dependence
on $\alpha_s(\mu_R)$ cancels out in the cross section ratio. The only
source of the scale dependence comes from variations in PDFs. The
latter, however, also largely cancels out in the cross section
ratio. The dependence on $\mu_R$ is introduced for the
first time at NLO due to the virtual and the real
corrections. Specifically, different one loop and real emission
structures in both processes give us a handle on
$\alpha_s(\mu_R)$. Additionally, the NLO predictions depend on the
renormalisation scale through logarithms of $\mu_R$, which appear in
both the virtual and the real emission contributions. Thus, the LO
error is truly underestimated and only the NLO theoretical error
should be considered as reliable.
In the next step, we would like to study the effect of various
settings in the numerator and denominator of the ${\cal R}$ observable
on the cross section ratio. In many experimental studies various MC
programs are employed usually with the default scale choice
implemented in a given program. Thus, one should assess the size of
the additional theoretical uncertainties due to the mismatch to see if
they are substantial or can be simply ignored. To this end for the
CT14 PDF set we calculate cross section ratios assuming different
scale choices in the numerator ($\mu_1$) and in the denominator
($\mu_2$) of the $\cal R$ observable. Specifically, we set $\mu_1$ to
the fixed scale choice $m_t/2 $ and $\mu_2$ to the dynamical scale
choice $H_T/4$ and vice versa. With the $p_{T, \gamma}> 25$ GeV cut we
have obtained the following results at NLO in QCD
\begin{equation}
\begin{split}
{\cal R}\left(
\frac{\mu_1= m_t/2}{\mu_2 = H_T/4}, \text{CT14}, p_{T,\gamma} > 25 \,{\rm
GeV}
\right)
&=(4.59 \pm 0.33)\cdot 10^{-3} \, (7\%)\,,
\\[0.2cm]
{\cal
R}\left(
\frac{\mu_1= H_T/4}{\mu_2=m_t/2}, \text{CT14}, p_{T,\gamma} > 25 \,{\rm
GeV}
\right)
&= (4.60 \pm 0.14)\cdot 10^{-3} \, (3\%)\,.
\\[0.2cm]
\end{split}
\end{equation}
In the case of $p_{T, \gamma}> 50$ GeV cut our NLO QCD findings can
be summarised as
\begin{equation}
\begin{split}
{\cal
R}\left( \frac{\mu_1=m_t/2}{\mu_2=H_T/4}, \text{CT14}, p_{T,\gamma} > 50 \,{\rm
GeV}\right)
&=( 1.90\pm 0.19 )\cdot 10^{-3} \, (10\%)\,,
\\[0.2cm]
{\cal
R}\left(
\frac{\mu_1=H_T/4}{\mu_2=m_t/2}, \text{CT14}, p_{T,\gamma} > 50 \,{\rm
GeV}
\right)
&= ( 1.92 \pm 0.09)\cdot 10^{-3} \, (5\%)\,.
\\[0.2cm]
\end{split}
\end{equation}
Even though in each case the central value of the cross section ratio
has not been changed, we observe an increase of the theoretical error
due to scale dependence. For the $p_{T,\gamma}$ cut of $25$ GeV ($50$
GeV) the following increase of the relative error can be quoted: for
$\mu_1=m_t/2$ the rise from $5\%$ ($8\%$) to $7\%$ ($10\%$) and for
$\mu_1=H_T/4$ from $1\%$ ($3\%$) to $3\%$ ($5\%$). Therefore, in order
to have $t\bar{t}\gamma$ production under excellent theoretical
control the same scale choice should be employed for the generation of
both processes $pp\to t\bar{t}\gamma$ and $pp\to t\bar{t}$. We can
also study the impact of using various PDF sets for the ${\cal R}$
observable. In that case our NLO QCD predictions for $p_{T,\gamma}>
25$ GeV are given by
\begin{equation}
\begin{split}
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm CT14}{\rm MMHT14}, p_{T,\gamma} > 25 \,{\rm
GeV}
\right) &= \left( 4.50 \pm 0.23 \right) \cdot 10^{-3} \, (5\%) \,, \\[0.2cm]
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm MMHT14}{\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV}
\right)&= \left( 4.60 \pm 0.28\right) \cdot 10^{-3} \, (6\%)
\,. \\[0.2cm]
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm CT14}{\rm NNPDF3.0}, p_{T,\gamma} > 25 \,{\rm
GeV}
\right) &= \left( 4.39 \pm 0.23 \right) \cdot 10^{-3} \, (5\%) \,, \\[0.2cm]
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm NNPDF3.0}{\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV}
\right)&= \left( 4.74 \pm 0.28 \right) \cdot 10^{-3} \, (6\%) \,. \\[0.2cm]
\end{split}
\end{equation}
For $p_{T,\gamma}> 50$ GeV have obtained instead
\begin{equation}
\begin{split}
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm CT14}{\rm MMHT14}, p_{T,\gamma} > 50 \,{\rm
GeV}
\right)&= \left( 1.87 \pm 0.15 \right) \cdot 10^{-3} \, (8\%) \,, \\[0.2cm]
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm MMHT14}{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV}
\right)&= \left( 1.90 \pm 0.17 \right) \cdot 10^{-3} \, (9\%)
\,. \\[0.2cm]
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm CT14}{\rm NNPDF3.0}, p_{T,\gamma} > 50 \,{\rm
GeV}
\right)&= \left( 1.82 \pm 0.15 \right) \cdot 10^{-3} \, (8\%) \,, \\[0.2cm]
{\cal R}\left(
\mu_0=m_t/2, \frac{\rm NNPDF3.0}{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV}
\right)&= \left( 1.96 \pm 0.18 \right) \cdot 10^{-3} \, (9\%) \,. \\[0.2cm]
\end{split}
\end{equation}
Although the choice of various PDF sets in the cross section ratio is
not theoretically very well motivated it does not affect the
estimation of the theoretical errors. Overall, unlike the theoretical
uncertainties due to the different scale choice, additional undesired
PDF uncertainties are negligible. Let us note here, that the issue of
choosing the same value of $\mu_0$ for both production processes is
going to play a crucial role when various differential cross section
ratios will be constructed.
To summarise this part of the paper, we discuss the stability of the
cross section ratio against higher order corrections. To
this end we also comment on the size of NLO QCD corrections to the
absolute $t\bar{t}$ and $t\bar{t}\gamma$ cross section with complete
top quark off-shell effects included. With $\mu_R=\mu_F=\mu_0$ set to
$\mu_0=m_t/2$ and for the CT14 PDF set the full $pp\to t\bar{t} $
cross section receives negative and small NLO corrections of
$3\%$. For $pp\to t\bar{t}\gamma$ we have obtained negative and
moderate NLO corrections of $10\%$ $(13\%)$ for the $p_{T, \gamma}$
cut of $25$ GeV ($50$ GeV). The size of NLO QCD corrections to cross
section ratio ${\cal R}$ is similar to the $pp\to t\bar{t}\gamma$
case. Specifically, NLO QCD corrections are also negative and of the
order of $8\%$ and $11\%$ depending on the $p_{T,\gamma}$ cut. For the
dynamical scale choice $\mu_0=H_T/2$ the NLO QCD corrections to $pp\to
t\bar{t}$ are positive and very small of the order of $0.6\%$
only. For the absolute $pp\to t\bar{t}\gamma$ cross section they are
also positive and small, however, of the order of $2\%$ and $5\%$ for
$p_{T, \gamma}>25$ GeV and $p_{T,\gamma}>50$ GeV respectively. The
size of NLO QCD corrections to the ${\cal R}$ observable as evaluated
with $\mu_0=H_T/4$ follows the same pattern as for the
$t\bar{t} \gamma$ production process. Thus, the cross section ratio
is very stable against higher order corrections and behaves similarly
as the absolute $t\bar{t}\gamma$ cross section when NLO QCD
corrections are incorporated.
\section{Differential Cross Section Ratios at NLO in QCD}
\label{section:6}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.49\textwidth]{hist_ratio_m_bb_25gev.pdf}
\includegraphics[width=0.49\textwidth]{hist_ratio_m_bb_50gev.pdf}
\end{center}
\caption{\it Differential cross section distributions as a function of
the invariant mass of two $b$-jets for the $pp\to e^+ \nu_e
\mu^-\bar{\nu}_\mu b\bar{b}\gamma +X$ process at the LHC run II with
$\sqrt{s}=13$ TeV. The upper plots show absolute NLO predictions for
$p_{T,\gamma}> 25$ GeV (left panel) and $p_{T,\gamma}> 50$ GeV (right
panel) together with the corresponding uncertainty bands resulting from
scale variations. Renormalisation and factorisation scales are set to
the common value $\mu_R=\mu_F=\mu_0$ where $\mu_0=m_t/2$ and
$H_T/4$. The CT14 PDF set is employed. The lower panels display
differential cross section ratios together with their uncertainty
bands. In the first case (middle panel), the same fixed (dynamical)
scale choice is employed in the numerator and the denominator of the
cross section ratio. In the second case (bottom panel), different
scale choices in the numerator and in the denominator have been
assumed.}
\label{fig:ratio1}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.49\textwidth]{hist_ratio_m_ll_25gev.pdf}
\includegraphics[width=0.49\textwidth]{hist_ratio_m_ll_50gev.pdf}
\end{center}
\caption{\it Differential cross section distributions as a function of
the invariant mass of two charged leptons for the $pp\to e^+ \nu_e
\mu^-\bar{\nu}_\mu b\bar{b}\gamma +X$ process at the LHC run II with
$\sqrt{s}=13$ TeV. The upper plots show absolute NLO predictions for
$p_{T,\gamma}> 25$ GeV (left panel) and $p_{T,\gamma}> 50$ GeV (right
panel) together with the corresponding uncertainty bands resulting from
scale variations. Renormalisation and factorisation scales are set to
the common value $\mu_R=\mu_F=\mu_0$ where $\mu_0=m_t/2$ and
$H_T/4$. The CT14 PDF set is employed. The lower panels display
differential cross section ratios together with their uncertainty
bands. In the first case (middle panel), the same fixed (dynamical)
scale choice is employed in the numerator and the denominator of the
cross section ratio. In the second case (bottom panel), different
scale choices in the numerator and in the denominator have been
assumed.}
\label{fig:ratio2}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.49\textwidth]{hist_ratio_dphi_ll_25gev.pdf}
\includegraphics[width=0.49\textwidth]{hist_ratio_dphi_ll_50gev.pdf}
\end{center}
\caption{\it Differential cross section distributions as a function of
$\Delta \phi_{\ell \ell}$ for the $pp\to e^+ \nu_e \mu^-\bar{\nu}_\mu
b\bar{b}\gamma +X$ process at the LHC run II with $\sqrt{s}=13$ TeV.
The upper plots show absolute NLO predictions for $p_{T,\gamma}> 25$
GeV (left panel) and $p_{T,\gamma}> 50$ GeV (right panel) together
with the corresponding uncertainty bands resulting from scale variations.
Renormalisation and factorisation scales are set to the common value
$\mu_R=\mu_F=\mu_0$ where $\mu_0=m_t/2$ and $H_T/4$. The CT14
PDF set is employed. The lower panels display differential cross section
ratios together with their uncertainty bands. In the first case
(middle panel), the same fixed (dynamical) scale choice is employed in
the numerator and the denominator of the cross section ratio. In the
second case (bottom panel), different scale choices in the numerator
and in the denominator have been assumed.}
\label{fig:ratio3}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{hist_ratio_pt_l1_25gev.pdf}
\includegraphics[width=0.49\textwidth]{hist_ratio_pt_l1_50gev.pdf}
\end{center}
\caption{\it Differential cross section distributions as a function of
the transverse momentum of the hardest charged lepton for the $pp\to
e^+ \nu_e \mu^-\bar{\nu}_\mu b\bar{b}\gamma +X$ process at the LHC run
II with $\sqrt{s}=13$ TeV. The upper plots show absolute NLO
predictions for $p_{T,\gamma}> 25$ GeV (left panel) and $p_{T,\gamma}>
50$ GeV (right panel) together with the corresponding uncertainty bands
resulting from scale variations. Renormalisation and factorisation
scales are set to the common value $\mu_R=\mu_F=\mu_0$ where
$\mu_0=m_t/2$ and $H_T/4$. The CT14 PDF sets are employed. The lower
panels display differential cross section ratios together with their
uncertainty bands. In the first case (middle panel), the same fixed
(dynamical) scale choice is employed in the numerator and the
denominator of the cross section ratio. In the second case (bottom
panel), different scale choices in the numerator and in the denominator
have been assumed.}
\label{fig:ratio4}
\end{figure}
In the following we present results for differential cross section
ratios defined according to
\begin{equation} {\cal R}_X=\left(\frac{d\sigma^{\rm
NLO}_{t\bar{t}\gamma} \left(\mu_1\right)}{dX}\right)
\left(\frac{d\sigma^{\rm NLO}_{ t\bar{t}} \left(\mu_2\right)
}{dX}\right)^{-1} \,,
\end{equation}
where $X$ stands for the observable that is under scrutiny. In Figure
\ref{fig:ratio1} we present differential cross section distributions as
a function of the invariant mass of two $b$-jets. Thus, in that case
we have $X=m_{bb}$ and ${\cal R}_{m_{bb}}$. The upper plots show
absolute NLO predictions for the $e^+ \nu_e \mu^-\bar{\nu}_\mu
b\bar{b}\gamma +X$ production process at the LHC at the centre-of-mass
energy of $\sqrt{s}=13$ TeV. Results are given for
$\mu_R=\mu_F=\mu_0$, where $\mu_0=m_t/2$ or $\mu_0=H_T/4$, for the
CT14 PDF set and for two different values of the $p_{T,\gamma}$ cut,
i.e. $p_{T,\gamma}> 25$ GeV and $p_{T,\gamma}> 50$ GeV. Also provided
are corresponding uncertainty bands resulting from scale
variations. The lower panels display differential cross section ratios
together with their uncertainty bands. In the first case, that is
presented in the middle panel, we have employed $\mu_1=\mu_2=\mu_0$,
where $\mu_0=m_t/2$ or $\mu_0=H_T/4$. In the second case, which is
shown in the bottom panel, $\mu_1\ne \mu_2$ has been assumed. For that
case two options are investigated, $(\mu_1=m_t/2)/(\mu_2=H_T/4)$ and
$(\mu_1=H_T/4)/(\mu_2=m_t/2)$. For the fixed scale choice for both
values of the $p_{T,\gamma}$ cut we have observed that the theoretical
error due to scale dependence for the absolute differential cross
section is in the range of $40\%-45\%$ towards the end of the $m_{bb}$
spectrum. On the other hand, for $\mu_0=H_T/4$ theoretical
uncertainties up to only $6\%-7\%$ have been estimated in the same
region. The situation is substantially changed when the cross section
ratio, ${\cal R}_{m_{bb}}$, is studied instead. In the case of
$\mu_1=\mu_2=\mu_0=m_t/2$ a reduction almost by a factor of $2$ can be
noticed. Indeed, theoretical uncertainties of the order of $20\%$ are
estimated at the end of the $m_{bb}$ spectrum. This finding is also
independent of the $p_{T,\gamma}$ cut. For $\mu_1=\mu_2=\mu_0=H_T/4$,
however, one can acquire theoretical uncertainties of the order of
$1\%-2\%$ in the whole plotted range. This shows that the
differential cross section ratio is also a very precise observable to
be studied together with the total cross section ratio at the LHC to
constrain physics beyond the SM for example via constraining anomalous
top quark couplings. It can also be used to extract the electric
charge of the top quark or the top quark charge asymmetry with a very
high precision. When the scales $\mu_1$ and $\mu_2$ are chosen
independently, however, the size of theoretical uncertainties has
increased dramatically up to $30\%-40\%$, as can be clearly seen in
Figure \ref{fig:ratio1}. Thus, choosing the same $\mu_0$ in the
numerator and denominator of ${\cal R}_{m_{bb}}$ is essential for
building a high precision observable, otherwise the theoretical errors
are drastically overestimated.
Similar conclusions can be drawn for the differential cross section
distribution as a function of the invariant mass of two charged
leptons. This observable is plotted in Figure \ref{fig:ratio2} again
for $p_{T,\gamma} > 25$ GeV and $p_{T,\gamma} > 50$ GeV. The
advantage of this observable in comparison to the invariant mass of
two $b$-jets lies, however, in the fact that measurements of lepton
kinematic observables are particularly precise at the LHC due to the
excellent lepton energy resolution of the ATLAS and CMS
detectors. Moreover, the reconstruction of the top quarks is not
required to construct $m_{\ell\ell}$. For $m_{bb}$, on the other
hand, good $b$-jet tagging efficiency and low light jet misstag rate
is mandatory. For the cross section ratio, ${\cal R}_{m_{\ell\ell}}$,
with the dynamical scale choice $\mu_1=\mu_2=H_T/4$ (the fixed scale
choice $\mu_1=\mu_2=m_t/2$) the theoretical uncertainties of the order
of $1\%-4\%$ ($20\%-25\%$) have been estimated. These should be
compared to uncertainties of up to $10\%$ ($50\%$) for the absolute
differential cross section. Again, our findings mildly depend on the
$p_{T,\gamma}$ cut. When different scales are applied to the numerator
and the denominator of ${\cal R}_{m_{\ell\ell}}$ theoretical
uncertainties in the tail of the distribution have increased up to
$35\%-40\%$ for $\mu_1=m_t/2$ and $\mu_2=H_T/4$ whereas in the case of
$\mu_1=H_T/4$ and $\mu_2=m_t/2$ they are in the range of $50\%-60\%$.
In Figure \ref{fig:ratio3} the differential cross section distribution
as a function of the difference in azimuthal angle between the two
charged leptons, $\Delta \phi_{\ell
\ell}=|\phi_{\ell_1}-\phi_{\ell_2}|$, is presented. The $\Delta
\phi_{\ell \ell}$ observable is also measured very precisely at the
LHC by both the ATLAS and CMS collaborations. It can be used for example
to construct the leptonic charge asymmetry, $A^{\ell\ell}_C$, which is
sensitive to signals of numerous beyond the SM scenarios, where among
others new heavy states might be produced. In general, angular
distributions of charged leptons are of huge importance since they
reflect spin correlations of the top quark pair and can be employed to
probe the $CP$ numbers of such new states, see
e.g. \cite{Frederix:2007gi,Bernreuther:2010ny,Bernreuther:2013aga,
Cacciapaglia:2015eqa,Godbole:2015bda}. For the fixed (dynamical)
scale choice theoretical uncertainties for
$\Delta \phi_{\ell \ell}$ in the region given by $\Delta
\phi_{\ell\ell} \gtrsim 2.5$ are of the order of $40\%-50\%$
($15\%-20\%$) depending on the transverse
momentum cut on the hard photon. When the cross section ratio ${\cal
R}_{\Delta \phi_{\ell \ell}}$ is investigated instead theoretical
errors of $20\%-30\%$ for $\mu_0=m_t/2$ and $2\%-3\%$ for
$\mu_0=H_T/4$ can be estimated in that region. Overall, for $\Delta
\phi_{\ell\ell} \in \left( 0,\pi\right)$ with $\mu_0=m_t/2$ or with
$\mu_0=H_T/4$ theoretical uncertainties for ${\cal R}_{\Delta
\phi_{\ell \ell}}$ are in the range $1\%-30\%$, $1\%-6\%$
respectively. As in the previous cases when $\mu_1\ne \mu_2$ is set
instead substantial overestimation of the theoretical uncertainties
can be observed for ${\cal R}_{\Delta \phi_{\ell \ell}}$. For example
for $\Delta \phi_{\ell\ell} \gtrsim 2.5$ an increase from $20\%-30\%$
up to $30\%-40\%$ has been procured once $\mu_1=m_t/2$ and
$\mu_2=H_T/4$ have been assumed, while for $\mu_1=H_T/4$ and
$\mu_2=m_t/2$ we have obtained a change from $2\%-3\%$ to $15\%$.
Finally, in Figure \ref{fig:ratio4} the differential cross section as a
function of the transverse momentum of the hardest charged lepton is
shown. This observable is also sensitive to effects of possible new
physics beyond the SM, see e.g. \cite{Melnikov:2011ta}. Among
others, it can be used to test exotic physics scenarios where top like
quarks with the electric charge of $Q_t = - 4/3$ might be
produced. For the absolute $t\bar{t}\gamma$ cross section theoretical
uncertainties are up to $30\%-45\%$ for $\mu_0=m_t/2$ and up to $8\%$
for $\mu_0=H_T/4$. Once the cross section ratio, ${\cal
R}_{p_{T,\ell_1}}$, is investigated theoretical uncertainties have
been substantially reduced down to $20\%-30\%$ for the fixed scale
choice and to $4\%-5\%$ for the phase-space dependent scale choice. If
we assume different scales in ${\cal R}_{p_{T,\ell_1}}$,
i.e. $\mu_1=m_t/2$ and $\mu_2=H_T/4$, theoretical errors
comparable to these quoted for the absolute $t\bar{t}\gamma$ cross
sections with $\mu_0=m_t/2$ have been evaluated. On the other hand,
setting $\mu_1=H_T/4$ and $\mu_2=m_t/2$, has resulted in theoretical
uncertainties for ${\cal R}_{p_{T,\ell_1}}$ maximally up to $15\%$.
To summarise this part of the paper, the theoretical uncertainties due
to scale dependence of the order of $1\%-6\%$ have been obtained for
the studied cross section ratios if $\mu_1=\mu_2=\mu_0=H_T/4$ has been
employed to construct ${\cal R}_X$, where $X$ stands for
$X=m_{b\bar{b}}, m_{\ell\ell}, \Delta \phi_{\ell \ell},
p_{T,\ell_1}$. Let us mention at this point that also for
differential cross section ratios the PDF uncertainties are smaller
than the theoretical uncertainties due to scale dependence. The latter
remains the dominant source of theoretical uncertainties.
\section{Conclusions}
\label{section:7}
The purpose of the paper is to obtain more precise theoretical
predictions for $t\bar{t}\gamma$ production in the di-lepton top quark
decay channel for the LHC Run II energy of $\sqrt{s}=13$ TeV without the
need of including terms beyond NLO in the perturbation expansion in
$\alpha_s$. To this end cross section ratios ${\cal R}=\sigma^{\rm
NLO}_{pp\to t\bar{t}\gamma}(\mu_1)/\sigma^{\rm NLO}_{pp \to
t\bar{t}}(\mu_2)$ have been studied. Fully realistic NLO computations
for $t\bar{t}$ and $t\bar{t}\gamma$ production have been
employed. They are based on LO and NLO matrix elements for
$e^+\nu_e\mu^-\bar{\nu}_\mu b\bar{b}$ and $e^+\nu_e\mu^-\bar{\nu}_\mu
b\bar{b}\gamma$ production processes that include all resonant and
non-resonant top quark and $W$ gauge boson Feynman diagrams,
interferences, and off-shell effects of $t$ and $W$. Various
renormalisation and factorisation scale choices and parton density
functions have been examined to assess their impact on the cross
section ratio. Our best NLO QCD predictions for the ${\cal R}$
observable have been obtained for $\mu_1=\mu_2=\mu_0=H_T/4$ and
can be summarised as follows
\begin{equation}
\begin{split}
{\cal R}\left(\mu_0 =H_T/4,{\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &= \left( 4.62 \pm 0.06\, {\rm [scales]} \pm 0.04 \,
{\rm [PDFs]} \right) \cdot 10^{-3}
\\[0.2cm]
{\cal R}(\mu_0=H_T/4,{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV} ) &= (1.93 \pm 0.06 \, {\rm [scales]}\pm 0.04 \, {\rm [PDFs]})
\cdot 10^{-3} \,. \\[0.2cm]
\end{split}
\end{equation}
The theoretical uncertainties due to scale dependence
have been estimated to be above $1\%$ for $p_{T,\gamma}> 25$ GeV and
of the order of $3\%$ for $p_{T,\gamma}> 50$ GeV. The theoretical
uncertainties due to various PDF parameterisations, on the other hand,
are $0.5\%$ and $1\%$ respectively. When the internal PDF
uncertainties are extracted from the CT14 PDF set following the
prescription of the CT14 PDF collaboration then they are slightly
higher, namely below $1\%$ and of the order of $2\%$ for
$p_{T,\gamma}> 25$ GeV and $p_{T,\gamma}> 50$ GeV
respectively. The latter are quoted for our best theoretical
predictions. Nevertheless, they are still smaller than theoretical errors
from the scale dependence. Regardless of which uncertainty we will
assign for PDF uncertainties such small theoretical
uncertainties are normally available only in the case of top quark
pair production at NNLO in QCD. Thus, the cross section ratio has
proven to be a very precise observable that should be measured at the
LHC. There are many possible applications, including, but not limited
to, precise measurements of the top quark charge as well as searches
for new physics effects that can reveal themselves only when a few
percent precision on the theory side is available. For the fixed
scale choice, which is still commonly used in experimental analyses,
our finding for $\mu_1=\mu_2=\mu_0=m_t/2$ are given by
\begin{equation}
\begin{split}
{\cal R} \left( \mu_0 = m_t/2, {\rm CT14}, p_{T,\gamma} > 25 \,{\rm
GeV} \right) &= (4.56\pm 0.25 \, {\rm [scales]} \pm 0.04\, {\rm
[PDFs]}) \cdot 10^{-3} \,,\\[0.2cm]
{\cal R}(\mu_0=m_t/2,{\rm CT14}, p_{T,\gamma} > 50 \,{\rm
GeV} ) &= (1.89\pm 0.16\, {\rm [scales]} \pm 0.04 \, {\rm [PDFs]}) \cdot
10^{-3} \,.\\[0.2cm]
\end{split}
\end{equation}
Also in this case theoretical errors due to scale dependence of the
order of a few percent, $5\%$ for $p_{T,\gamma}> 25$ GeV and $8\%$
for $p_{T,\gamma}> 50$ GeV, have been estimated. We have also shown
that such high precision can only be obtained if $\mu_1$ and $\mu_2$
are set to a common scale. Otherwise, theoretical uncertainties from
the scale dependence are overestimated. We have argued on the
similarity of the two processes, that using the same scale for both is
well justified.
Subsequently, we have turned our attention to differential cross
section distributions. Four observables have been presented at the
differential level for the $t\bar{t}\gamma$ production process at the
LHC. Specifically, we have shown the invariant mass of two $b$-jets
($m_{bb}$), the invariant mass of two charged leptons
($m_{\ell\ell}$), the difference in azimuthal angle between two
charged leptons ($\Delta \phi_{\ell\ell}$) and the transverse momentum
of the hardest charged lepton ($p_{T,\ell_1}$). Afterwards, we have
calculated differential cross section ratios for these observables
according to
\begin{equation}
{\cal
R}_X=\left(\frac{d\sigma^{\rm NLO}_{t\bar{t}\gamma} (\mu_1)}{dX}\right)
\left(\frac{d\sigma^{\rm NLO}_{t\bar{t}} (\mu_2)}{dX}\right)^{-1} \,,
\end{equation}
where $X=m_{bb}, m_{\ell\ell}, \Delta \phi_{\ell\ell}$ and
$p_{T,\ell_1}$. A clear conclusion could be drawn from our
considerations. For observables that we have presented, which are also
important for beyond the SM physics searches, the most precise
predictions for ${\cal R}_X$ have been obtained for
$\mu_1=\mu_2=\mu_0$. Especially interesting conclusions have been
reached for the case of the dynamical scale choice,
i.e. $\mu_0=H_T/4$. For all observables that have been investigated,
theoretical uncertainties due to scale dependence, which
are the dominant source of theoretical systematics also at the
differential level, are in the $1\%-6\%$ range. These findings
are independent of the transverse momentum cut on the isolated hard
photon. Such precise theoretical predictions at the differential
level should be now employed to indirectly search for new physics at
the LHC. When different scale choices $\mu_1\ne \mu_2$ for ${\cal
R}_X$ have been assumed instead, theoretical uncertainties have been
dramatically overestimated. Thus, care must be taken to ensure that
$\mu_1$ and $\mu_2$ are as similar as possible when building the
${\cal R}_X$ observables to be used in experimental studies. Based on
our studies we advocate for the $H_T$ based scale choice for $\mu_1$
and $\mu_2$ in ${\cal R}_X$. Definitely, mixing dynamical and fixed
scales in $t\bar{t}\gamma$ and $t\bar{t}$ production introduces
additional and unnecessary theoretical uncertainties that should be
avoided.
As a further matter let us note here that, from the experimental point
of view measurements in the fiducial phase space, which is the phase
space defined by the specific selection cuts that simulate detector
response as closely as possible, are the most appropriate for new
physics searches in the top quark sector and for precision
measurements of top quark properties within the SM theory of particle
physics. The reason being that such measurements do not introduce
additional and unnecessary systematic uncertainties due to
model-dependent extrapolations to parton level $t$ and $\bar{t}$
objects and to phase-space regions outside the detector
sensitivity. Therefore, our theoretical predictions for observables
like ${\cal R}_{m_{\ell\ell}}$, ${\cal R}_{\Delta\phi_{\ell\ell}}$ and
${\cal R}_{p_{T,\ell_1}}$ might be directly compared with experimental
data at the fiducial level that are collected at the LHC by the ATLAS
and CMS experimental collaborations. These leptonic
observables should be marginally affected by parton shower
effects. Firstly, such effects are milder if the underlying computations
are NLO accurate since the fixed order contribution already includes
part of the radiation simulated by the shower. Secondly, leptonic
observables are rather stable against shower effects, see
e.g. Ref.~\cite{Heinrich:2017bqp}, as compared for example
to observables, which are build out of light- and
$b$-jets. Nevertheless, due to the fact that both processes are very
similar from the QCD point of view such effects should cancel to a
large extent in cross section ratios. Regardless, let us mention at
this point that recently a new method for matching NLO QCD
calculations to parton shower programs has been proposed
\cite{Jezo:2015aia}, which incorporates top quark finite width effects
in the shower approach together with all interference
effects. Specifically, it allows for a consistent treatment of
resonances in the \textsc{Powheg} framework by preserving the mass of
top-quark resonances near their peak. First results for the simplest
case of the $e^+ \nu_e \mu^-\bar{\nu} b\bar{b}$ production process
have been already presented in Ref.~\cite{Jezo:2016ujg}. Until now,
however, this method has not been applied to the more complex process
$e^+ \nu_e \mu^-\bar{\nu} b\bar{b} \gamma$. Once available it can be
used for studying $\sigma_{e^+ \nu_e \mu^-\bar{\nu} b\bar{b} \gamma}/
\sigma_{e^+ \nu_e \mu^-\bar{\nu} b\bar{b}}$.
Finally, let us also add that we have not investigated the impact of
NLO electroweak corrections on $\sigma_{e^+ \nu_e \mu^-\bar{\nu}
b\bar{b} \gamma}/\sigma_{e^+ \nu_e \mu^-\bar{\nu} b\bar{b}}$. However,
based on results for the simpler $e^+ \nu_e \mu^-\bar{\nu} b\bar{b}$
production process at the LHC with $\sqrt{s}=13$ TeV
\cite{Denner:2016jyo}, the ratio observable is not expected to be very
sensitive to such effects. For the integrated cross section, NLO
electroweak corrections to off-shell top anti-top production with
leptonic decays are below $1\%$. In the case of differential cross
sections such effects are small as well. Specifically, for $\Delta
\phi_{\ell\ell}$ and $p_{T\, \ell}$, which are studied in
Ref.~\cite{Denner:2016jyo}, NLO electroweak corrections are
respectively below $1\%$ and $2\%$ (in the plotted range, i.e. up to
$200$ GeV for $p_{T,\, \ell}$). For the latter case they increase up
to $7\%$ when $p_{T,\, \ell} \approx 800$ GeV. Nevertheless, their
size is substantially smaller than the size of NLO QCD corrections in
the same regions. Furthermore, we expect that NLO electroweak effects
will be additionally minimised once ${\cal R}$ and ${\cal R}_X$ ratios
are constructed. For the more complicated process $e^+ \nu_e
\mu^-\bar{\nu} b\bar{b} \gamma$ NLO electroweak calculations are not
available in the literature. Up to now only NLO electroweak
corrections to the on-shell $t\bar{t}\gamma$ production process have
been evaluated \cite{Duan:2016qlc}. Nevertheless, for the integrated
$t\bar{t}\gamma$ cross section at the LHC with $\sqrt{s}=13$ TeV and
for the $p_{T,\, \ell} > 50$ GeV cut these corrections are below
$2\%$. For various differential distributions presented in
Ref.~\cite{Duan:2016qlc}, like for example $p_{T, \, t}$ and $p_{T,\,
\gamma}$, these effects are well below $10\%$ even in the tails of
these distributions. Based on that and the fact that further reduction
is foreseen once ${\cal R}$ and ${\cal R}_X$ are built up out of the
integrated and differential cross sections for the $e^+ \nu_e
\mu^-\bar{\nu} b\bar{b} \gamma$ and $e^+ \nu_e \mu^-\bar{\nu}
b\bar{b}$ production processes, we expect that our results for ${\cal
R}$ and ${\cal R}_X$ will not be changed substantially.
On the technical side let us mention that all our results have been
generated with the help of the \textsc{Helac-NLO} MC framework. The
final results are available (upon request) as Ntuple files
\cite{Bern:2013zja}. In detail, they are stored in the form of
modified Les Houches \cite{Alwall:2006yp} and ROOT event files
\cite{Antcheva:2009zz} that might be directly used for experimental
studies at the LHC.
\section*{Acknowledgements}
The work of M.W. and T.W. was supported in part by the DFG Research
Grant {\it "Top-Quarks under the LHCs Magnifying Glass: From Process
Modelling to Parameter Extraction"}.
The work of H.B.H. has received funding from the European Research
Council (ERC) under the European Union's Horizon 2020 research and
innovation programme (grant agreement No 772099). Additionally, he has
been supported by Rutherford Grant ST/M004104/1.
The research of G.B. was supported by grant K 125105 of the National
Research, Development and Innovation Office in Hungary.
Simulations were performed with computing resources that are granted
by RWTH Aachen University under project {\tt rwth0211}.
|
\section{Introduction}
Euclidean Ramsey theory typically seeks for a given pattern, such as vertices of a square, an arithmetic progression, etc., in a single partition class determined by an arbitrary (or only measurable) coloring of the Euclidean space. Stronger results than the mere coloring theorems are the so-called density theorems, which establish existence of the pattern inside an arbitrary measurable subset of positive density. The appropriate notion of density for this purpose is the \emph{upper Banach density}, defined as
\begin{equation}\label{eq:defofdensity}
\overline{\delta}(A) := \limsup_{N\to\infty} \sup_{x\in\mathbb{R}^d} \frac{|A \cap (x+[0,N]^d)|}{N^d}
\end{equation}
for any measurable $A\subseteq\mathbb{R}^d$. Here and in what follows, $|B|$ denotes the Lebesgue measure of a measurable set $B\subseteq\mathbb{R}^d$.
An interesting class of density results tries to find congruent copies of all sufficiently large dilates of a given pattern. There is always a critical dimension $d_{\textup{min}}$ below which positive statements cannot hold. Since many dimension-related issues are still unresolved, one is often content with proving the claim when $d$ is sufficiently large. An initial result of this type starts with the simplest possible pattern, a pair of points in $\mathbb{R}^d$ for $d\geq 2$, and it was established independently by Bourgain \cite{B86:roth}, Falconer and Marstrand \cite{FM86:dist}, and Furstenberg, Katznelson, and Weiss \cite{FKW90:dist}. Moreover, Bourgain \cite{B86:roth} generalized it to non-degenerate $k$-point patterns, also viewed as vertices of $(k-1)$-dimensional simplices, in $\mathbb{R}^d$ for $d\geq k$.
More recently, Lyall and Magyar \cite{LM16:prod} initiated the consideration of product-type patterns. They proved that, for fixed $a_1,a_2>0$, a positive density subset of $\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}$, $d_1,d_2\geq 2$, contains vertices of a rectangle,
\[ (x_1,x_2), \quad (x_1,x_2+s_2), \quad (x_1+s_1,x_2), \quad (x_1+s_1,x_2+s_2), \]
with $x_1,s_1\in\mathbb{R}^{d_1}$, $x_2,s_2\in\mathbb{R}^{d_2}$, $\|s_1\|_{\ell^2}=\lambda a_1$, and $\|s_2\|_{\ell^2}=\lambda a_2$, for all sufficiently large $\lambda>0$. We write $\|v\|_{\ell^2}$ for the Euclidean norm of a vector $v=(v_1,v_2,\ldots,v_d)\in\mathbb{R}^d$, since later we will also consider more general $\ell^p$-norms, $1<p<\infty$, defined as
\[ \|v\|_{\ell^p} := \big( \sum_{i=1}^{d} |v_i|^p \big)^{1/p}. \]
The particular case $a_1=a_2=1$ corresponds to the search for squares. In the same paper the authors proceed to Cartesian products of two general non-degenerate simplices.
As our first result, we establish a different generalization, replacing (vertices of) rectangles with (vertices of) higher-dimensional rectangular solids. Let $d_1,d_2,\ldots,d_n$ be positive integers. In what follows, a \emph{box} will be a pattern consisting of $2^n$ points in $\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\dots\times\mathbb{R}^{d_n}$ of the form
\begin{equation}\label{eq:box1}
(x_1 + k_1 s_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n), \quad k_1,k_2,\dots,k_n\in\{0,1\}
\end{equation}
for any $x_j,s_j\in\mathbb{R}^{d_j}$, $s_j\neq\mathbf{0}$, $j=1,2,\ldots,n$.
\begin{theorem}\label{thm:main1}
Fix numbers $a_1,a_2,\ldots,a_n>0$. For any positive integers $d_1,d_2,\dots,d_n\geq 5$ and any measurable set $A\subseteq\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\dots\times\mathbb{R}^{d_n}$ with $\overline{\delta}(A)>0$ one can find $\lambda_0>0$ with the property that for any real number $\lambda\geq\lambda_0$ the set $A$ contains a box \eqref{eq:box1} with $x_j,s_j\in\mathbb{R}^{d_j}$ and $\|s_j\|_{\ell^2} = \lambda a_j$, $j=1,2,\ldots,n$.
\end{theorem}
The possibility of generalizing the aforementioned results of Lyall and Magyar to $n$-fold products with $n\geq 3$ was announced by the same authors in \cite{LM16:prod}. The present paper and the more recent preprint by Lyall and Magyar \cite{LM19:hypergraphs} achieve this goal for boxes independently of each other and using quite different approaches.
In fact, \cite{LM19:hypergraphs} proves a sharp variant of Theorem~\ref{thm:main1} above, in which the assumptions $d_j\geq 5$ are relaxed to $d_j\geq 2$. It is clearly necessary to assume $d_j\geq 2$: if we had $d_1=1$, then the set of all points with the first coordinate from
\[ \bigcup_{k\in\mathbb{Z}}\big[(k-1/10)a_1,(k+1/10)a_1\big] \]
would be a counterexample, since it would contain no boxes associated with half-integer values of $\lambda$.
The approach pursued in this paper is in the spirit of the paper by Cook, Magyar, and Pramanik \cite{CMP15:roth}, and the same method will allow us to handle certain enlarged patterns we are about to discuss.
\smallskip
Bourgain \cite{B86:roth} also constructed a measurable set $A\in\mathbb{R}^d$ with $\overline{\delta}(A)>0$ such that lengths $\|s\|_{\ell^2}$ of gaps $s$ for all $3$-term arithmetic progressions
\[ x,\ x+s,\ x+2s \]
inside $A$ omit an unbounded set of positive values. This prevents us from having the most obvious candidate for a density theorem for $3$-term arithmetic progressions. On the other hand, Cook, Magyar, and Pramanik \cite{CMP15:roth} showed that the corresponding density theorem still holds if one is allowed to measure sizes of gaps $s$ in the $\ell^p$-norms for $1<p<\infty$, $p\neq 2$.
Our second result is a common generalization of Theorem~\ref{thm:main1} above and Theorem~2.1 from \cite{CMP15:roth}.
Consider $n$ additional points in $\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\dots\times\mathbb{R}^{d_n}$,
\begin{equation}\label{eq:box2}
(x_1 + 2 s_1, x_2, \dots, x_n),\quad (x_1, x_2 + 2 s_2, \dots, x_n),\quad \ldots,\quad (x_1, x_2, \dots, x_n + 2 s_n)
\end{equation}
for given $x_j,s_j\in\mathbb{R}^{d_j}$, $s_j\neq\mathbf{0}$, $j=1,2,\ldots,n$. The union of \eqref{eq:box1} and \eqref{eq:box2} will be called a \emph{$3$AP-extended box}: it has a $3$-term arithmetic progression attached to each edge coming from a fixed vertex of the box. From the aforementioned observation of Bourgain we know that an analogue of Theorem~\ref{thm:main1} for the $3$AP-extended boxes is not possible, so one has to give up on the Euclidean norm.
\begin{theorem}\label{thm:main2}
Fix numbers $a_1,a_2,\ldots,a_n>0$ and an exponent $1<p<\infty$, $p\neq 2$. There exists a dimensional threshold $d_{\textup{min}}$ such that for any positive integers $d_1,d_2,\dots,$ $d_n\geq d_{\textup{min}}$ and any measurable set $A\subseteq\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\dots\times\mathbb{R}^{d_n}$ with $\overline{\delta}(A)>0$ one can find $\lambda_0>0$ with the property that for any real number $\lambda\geq\lambda_0$ the set $A$ contains a $3$AP-extended box \eqref{eq:box1}$\cup$\eqref{eq:box2} with $x_j,s_j\in\mathbb{R}^{d_j}$ and $\|s_j\|_{\ell^p} = \lambda a_j$, $j=1,2,\ldots,n$.
\end{theorem}
The pattern consisting of points \eqref{eq:box1} and \eqref{eq:box2} can also be viewed as a subset of the grid
\[ (x_1 + k_1 s_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n), \quad k_1,k_2,\dots,k_n\in\{0,1,2\}, \]
consisting of $3^n$ points. At the moment we are not able to prove a result analogous to Theorem~\ref{thm:main2} for this grid. Larger grids bring further complications: one should first handle longer arithmetic progressions and it is known that additional restrictions on the values of $p$ are needed; see the remarks in \cite{DKR17:corner}.
The same approach will enable a further generalization of Theorems~\ref{thm:main1} and \ref{thm:main2}. The present authors and Rimani\'{c} \cite{DKR17:corner} have raised the generality of the result by Cook, Magyar, and Pramanik \cite{CMP15:roth} from $3$-term arithmetic progressions to \emph{corners}, which are triples of points in $\mathbb{R}^d\times\mathbb{R}^d$ of the form
\[ (x,y),\ (x+s,y),\ (x,y+s) \]
for $x,y,s\in\mathbb{R}^d$, $s\neq\mathbf{0}$. A \emph{corner-extended box} will be a pattern in $(\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\dots\times\mathbb{R}^{d_n})^2$ consisting of $2^n$ points forming a box,
\begin{equation}\label{eq:box3}
(x_1 + k_1 s_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1, y_2, \dots, y_n), \quad k_1,k_2,\dots,k_n\in\{0,1\},
\end{equation}
and $n$ additional points completing corners with $n$ of its edges,
\begin{equation}\label{eq:box4}
(x_1, x_2, \dots, x_n,\,y_1 + s_1, y_2, \dots, y_n),\quad \ldots,\quad (x_1, x_2, \dots, x_n,\,y_1, y_2, \dots, y_n + s_n),
\end{equation}
where $x_j,y_j,s_j\in\mathbb{R}^{d_j}$, $s_j\neq\mathbf{0}$, $j=1,2,\ldots,n$. The following result is a common generalization of Theorem~\ref{thm:main1} above and Theorem~1.2 from \cite{DKR17:corner}.
\begin{theorem}\label{thm:main3}
Fix numbers $a_1,a_2,\ldots,a_n>0$ and an exponent $1<p<\infty$, $p\neq 2$. There exists a dimensional threshold $d_{\textup{min}}$ such that for any positive integers $d_1,d_2,\dots,$ $d_n\geq d_{\textup{min}}$ and any measurable set $A\subseteq(\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\dots\times\mathbb{R}^{d_n})^2$ with $\overline{\delta}(A)>0$ one can find $\lambda_0>0$ with the property that for any real number $\lambda\geq\lambda_0$ the set $A$ contains a corner-extended box \eqref{eq:box3}$\cup$\eqref{eq:box4} with $x_j,y_j,s_j\in\mathbb{R}^{d_j}$ and $\|s_j\|_{\ell^p} = \lambda a_j$, $j=1,2,\ldots,n$.
\end{theorem}
Theorem~\ref{thm:main3} implies Theorem~\ref{thm:main2}, as can be seen by considering the skew projections $(x_j,y_j)\mapsto y_j-x_j$; see \cite{DKR17:corner} for details. Consequently, it is still necessary to assume $p\neq 2$, while the endpoint cases $p=1$ and $p=\infty$ clearly do not allow any nontrivial results; see the comments in \cite{CMP15:roth}.
Let us emphasize that the dimensional threshold $d_{\textup{min}}$ in Theorems~\ref{thm:main2} and \ref{thm:main3} depends on the exponent $p$. Careful analysis of the arguments below can give $d_{\textup{min}}=O(p)$ for each fixed $n$, but we do not even have counterexamples to the possibility $d_{\textup{min}}=2$. The optimal value of $d_{\textup{min}}$ is still far from understood, even in the case $n=1$ studied in \cite{CMP15:roth} and \cite{DKR17:corner}.
Lyall and Magyar \cite{LM18:graphs} also worked on the Euclidean embedding of all large dilates of a fixed distance graph. Their results do not include Theorem~\ref{thm:main1}, since the boxes (or even rectangles) are simultaneously ``too rigid'' and ``too degenerate;'' compare with the definition of a \emph{proper $k$-degenerate distance graph} from \cite{LM18:graphs}. They also clearly do not overlap with Theorems~\ref{thm:main2} and \ref{thm:main3}, simply because these theorems fail in the Euclidean metric.
\smallskip
As we have already mentioned, our method of approach is based on the paper by Cook, Magyar, and Pramanik \cite{CMP15:roth}. This method reduces Theorems~\ref{thm:main1}--\ref{thm:main3} to boundedness of certain multilinear singular integral operators. In order to obtain bounds for these operators we invoke the main result from the recent paper by Thiele and one of the present authors \cite{DT18}, which in turn uses techniques gradually developed in a series of papers including \cite{K12:tp}, \cite{D15:L4}, \cite{DKR17:corner}, \cite{DKST17:nvea}, and \cite{DKT16:splx}.
In Section~\ref{sec:scheme} we list the main ingredients of the proofs in the form of several propositions and we explain how they imply the three theorems. Section~\ref{sec:combinatorics} establishes the propositions that belong to the combinatorial part of the proof, by either invoking \cite{CMP15:roth}, or performing necessary modifications. Section~\ref{sec:analysis} establishes the propositions dealing with singular integral operators, which constitute the analytical part of the proof.
\section{Scheme of the proofs}
\label{sec:scheme}
We have already explained how Theorem~\ref{thm:main2} can be derived from Theorem~\ref{thm:main3}, so in this section we give outlines of proofs of Theorems~\ref{thm:main1} and \ref{thm:main3}. In complete analogy with the steps from \cite{CMP15:roth}, they will be reduced to Propositions~\ref{prop:1mainterm}--\ref{prop:3multiscale} below.
If $A,B\colon\mathcal{D}\to[0,\infty)$ are two functions or functionals for which there exists a finite constant $C$ depending on a set of parameters $P$ such that $A(x)\leq CB(x)$ for each $x\in\mathcal{D}$, then we write
\[ A(x)\lesssim_P B(x). \]
If both $A(x)\lesssim_P B(x)$ and $B(x)\lesssim_P A(x)$, then we write
\[ A(x)\sim_P B(x). \]
The parameters in $P$ that are understood throughout the text will be omitted from this notation. In particular, it will always be understood that all constants implicit in the notation $\lesssim$ and $\sim$ depend on the fixed vector of positive numbers
\[ \mathbf{a} := (a_1,a_2,\ldots,a_n), \]
which determines the shape of the patterns (the aspect ratios of the boxes), and the exponent $p$, which is relevant to the proof of Theorem~\ref{thm:main3} only.
Characteristic function (i.e.\@ the indicator function) of a set $E$ will be written as $\mathbbm{1}_E$.
Let us write $g_t$ for an $\textup{L}^1$-normalized dilate of a function $g\colon\mathbb{R}^d\to\mathbb{C}$ by a factor $t>0$, i.e.,
\begin{equation}\label{eq:funcdilate}
g_t(s) := t^{-d}g(t^{-1}s)
\end{equation}
for each $s\in\mathbb{R}^d$.
The Fourier transform of $\textup{L}^1$ functions is normalized as
\[ \widehat{g}(\xi) := \int_{\mathbb{R}^d} g(s) e^{-2\pi i s\cdot\xi} \,\textup{d}s, \]
where $s\cdot\xi$ stands the standard scalar product of vectors $s$ and $\xi$ in $\mathbb{R}^d$.
If $\sigma$ is a measure on Borel subsets of $\mathbb{R}^d$, then we define its dilate by $t>0$ as another measure $\sigma_t$ given as
\begin{equation}\label{eq:measdilate}
\sigma_{t}(E):=\sigma(t^{-1}E)
\end{equation}
for each Borel set $E\subseteq\mathbb{R}^d$. A consequence of a linear change of variables is
\[ \int_{\mathbb{R}^d} f(s) \,\textup{d}\sigma_{t}(s) = \int_{\mathbb{R}^d} f(ts) \,\textup{d}\sigma(s) \]
for any measurable function $f\colon\mathbb{R}^d\to\mathbb{C}$ such that the above integrals exist.
Notation \eqref{eq:funcdilate} and \eqref{eq:measdilate} is mutually consistent when $\sigma$ is absolutely continuous with respect to the Lebesgue measure with density $g$. Occasionally we will need an $\textup{L}^p$-normalized dilate of $g\colon\mathbb{R}^d\to\mathbb{C}$ by $t>0$, for a more general exponent $1\leq p<\infty$, which will be denoted $\textup{D}_t^{p}g$ and defined as
\begin{equation}\label{eq:funcdilatep}
\textup{D}_t^{p}g(s) := t^{-d/p}g(t^{-1}s).
\end{equation}
Let us fix an exponent $1<p<\infty$; it will simply be $p=2$ in relation with Theorem~\ref{thm:main1}, while the proof of Theorem~\ref{thm:main3} will assume $p\neq 2$. We introduce a measure $\sigma^{d,p}$ on Borel subsets of $\mathbb{R}^{d}$ in the Dirac $\delta$ notation as
\[ \sigma^{d,p}(s) := \delta\big(1-\|s\|_{\ell^p}^p\big) \]
or, less formally and abusing the integral representation for the Fourier transform, as
\[ \sigma^{d,p}(s) = \int_{\mathbb{R}} e^{-2\pi i u (1-\|s\|_{\ell^p}^p)} \,\textup{d}u. \]
Its dilate $\sigma^{d,p}_{\lambda}$ by $\lambda>0$ is clearly supported on the $\textup{C}^1$ surface $\{s\in\mathbb{R}^d : \|s\|_{\ell^p} = \lambda\}$.
Let us also fix a Schwartz function $\psi\colon\mathbb{R}\to[0,1]$ such that $\widehat{\psi}\geq 0$, $\widehat{\psi}$ is supported in $[-4,4]$, $\widehat{\psi}(1)>0$, and $\psi(0)=1$. For instance, we can take a $\textup{C}^\infty$ function $\rho\colon\mathbb{R}\to[0,\infty)$ such that $\rho>0$ on $[-1,1]$, $\rho=0$ outside $[-2,2]$, and $\rho$ has integral $1$; then we can simply set $\psi=|\widehat{\rho}|^2$. Any constants implicit in the notation $\lesssim$ and $\sim$ will also be understood to depend on $\psi$. Furthermore, for any $\varepsilon>0$ we introduce a function $\omega^{d,p,\varepsilon}\colon\mathbb{R}^d\to\mathbb{C}$ by the formula
\[ \omega^{d,p,\varepsilon}(s) := \int_{\mathbb{R}} e^{-2\pi i u (1-\|s\|_{\ell^p}^p)} \psi(\varepsilon u) \,\textup{d}u
= \varepsilon^{-1}\widehat{\psi}\big(\varepsilon^{-1}(1-\|s\|_{\ell^p}^p)\big). \]
It was shown in Lemma~4.1 of \cite{CMP15:roth} that
\[ \int_{\mathbb{R}^d} \omega^{d,p,\varepsilon}(s) \,\textup{d}s \sim_{d,p} 1 \]
for all $0<\varepsilon<1/10d$. Thus, for such $\varepsilon$ we set
\begin{equation}\label{eq:omegaquotient}
c(d,p,\varepsilon) := \frac{\int_{\mathbb{R}^d} \omega^{d,p,\varepsilon}(s) \,\textup{d}s}{\int_{\mathbb{R}^d} \omega^{d,p,1}(s) \,\textup{d}s} \sim_{d,p} 1
\end{equation}
and then $k^{d,p,\varepsilon}\colon\mathbb{R}^d\to\mathbb{R}$ defined by
\begin{equation}\label{eq:defofk}
k^{d,p,\varepsilon}(s) := \omega^{d,p,\varepsilon}(s) - c(d,p,\varepsilon) \omega^{d,p,1}(s)
\end{equation}
has integral equal to $0$.
We introduce the number $D=d_1+\cdots+d_n$, so that
\[ \mathbb{R}^{D}\cong\mathbb{R}^{d_1}\times\dots\times\mathbb{R}^{d_n}. \]
Throughout the proofs we will use the shorthand notation
\[ \mathbf{x} := (x_1, \dots, x_n),\quad \mathbf{y} := (y_1, \dots, y_n),\quad \mathbf{s} := (s_1, \dots, s_n) \]
and we view $\mathbf{x}$, $\mathbf{y}$, and $\mathbf{s}$ as vectors from $\mathbb{R}^{D}$.
It will also be convenient to adopt some derived notation, such as
\[ \textup{d}\mathbf{x} := \textup{d}x_1 \cdots \textup{d}x_n, \quad
\textup{d}\sigma^{p}_{\lambda\mathbf{a}}(\mathbf{s}) := \textup{d}\sigma^{d_1,p}_{\lambda a_1}(s_1) \cdots \textup{d}\sigma^{d_n,p}_{\lambda a_n}(s_n). \]
In the same spirit we define
\[ \omega^{p,\varepsilon}_{\lambda\mathbf{a}}(\mathbf{s}) := \omega^{d_1,p,\varepsilon}_{\lambda a_1}(s_1) \cdots \omega^{d_n,p,\varepsilon}_{\lambda a_n}(s_n). \]
Let us also write $\mathbf{k}=(k_1,\ldots,k_n)\in\{0,1\}^n$ and denote
\[ (\mathcal{F}f)(\mathbf{x},\mathbf{s}) := \prod_{\mathbf{k}\in\{0,1\}^n} f(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n) \]
for a function $f\colon\mathbb{R}^D\to[0,1]$ and
\begin{align*}
(\widetilde{\mathcal{F}}f)(\mathbf{x},\mathbf{y},\mathbf{s}) :=
& \Big( \prod_{\mathbf{k}\in\{0,1\}^n} f(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n,\,y_1, \dots, y_n) \Big) \\
& f(x_1, \dots, x_n,\,y_1 + s_1, \dots, y_n) \cdots f(x_1, \dots, x_n,\,y_1, \dots, y_n + s_n)
\end{align*}
for a function $f\colon(\mathbb{R}^{D})^2\to[0,1]$.
The most important objects are the \emph{pattern-counting forms}, defined as follows. For a ``scale'' $\lambda>0$ we set
\[ \mathcal{N}^{p}_{\lambda}(f) := \int_{(\mathbb{R}^{D})^2} (\mathcal{F}f)(\mathbf{x},\mathbf{s}) \,\textup{d}\sigma^{p}_{\lambda\mathbf{a}}(\mathbf{s}) \,\textup{d}\mathbf{x} \]
and
\[ \widetilde{\mathcal{N}}^{p}_{\lambda}(f) := \int_{(\mathbb{R}^{D})^3} (\widetilde{\mathcal{F}}f)(\mathbf{x},\mathbf{y},\mathbf{s}) \,\textup{d}\sigma^{p}_{\lambda\mathbf{a}}(\mathbf{s}) \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y}. \]
The name comes from the fact that if $\mathcal{N}^{p}_{\lambda}(\mathbbm{1}_A)>0$ (resp.\@ $\widetilde{\mathcal{N}}^{p}_{\lambda}(\mathbbm{1}_A)>0$), then $A$ contains a box \eqref{eq:box1} (resp.\@ a corner-extended box \eqref{eq:box3}$\cup$\eqref{eq:box4}) with $\|s_j\|_{\ell^p} = \lambda a_j$, $j=1,2,\ldots,n$.
We will also need their smoothened versions, defined for $\varepsilon>0$ as
\[ \mathcal{M}^{p,\varepsilon}_{\lambda}(f) := \int_{(\mathbb{R}^{D})^2} (\mathcal{F}f)(\mathbf{x},\mathbf{s}) \omega^{p,\varepsilon}_{\lambda\mathbf{a}}(\mathbf{s}) \,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x} \]
and
\[ \widetilde{\mathcal{M}}^{p,\varepsilon}_{\lambda}(f) := \int_{(\mathbb{R}^{D})^3} (\widetilde{\mathcal{F}}f)(\mathbf{x},\mathbf{y},\mathbf{s}) \omega^{p,\varepsilon}_{\lambda\mathbf{a}}(\mathbf{s}) \,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y}. \]
By the standard approximation of identity arguments,
\begin{equation}\label{eq:limits}
\lim_{\varepsilon\to0^+} \mathcal{M}^{p,\varepsilon}_{\lambda}(f) = \mathcal{N}^{p}_{\lambda}(f), \quad
\lim_{\varepsilon\to0^+} \widetilde{\mathcal{M}}^{p,\varepsilon}_{\lambda}(f) = \widetilde{\mathcal{N}}^{p}_{\lambda}(f)
\end{equation}
for functions $f$ as above.
Finally, we denote
\[ \mathcal{E}^{p,\varepsilon}_{\lambda}(f) := \mathcal{M}^{p,\varepsilon}_{\lambda}(f) - b(p,\varepsilon) \mathcal{M}^{p,1}_{\lambda}(f) \]
and
\[ \widetilde{\mathcal{E}}^{p,\varepsilon}_{\lambda}(f) := \widetilde{\mathcal{M}}^{p,\varepsilon}_{\lambda}(f) - b(p,\varepsilon) \widetilde{\mathcal{M}}^{p,1}_{\lambda}(f), \]
where we recall that the numbers $c(d_j,p,\varepsilon)$ come from \eqref{eq:omegaquotient} and use the shorthand notation
\[ b(p,\varepsilon) := c(d_1,p,\varepsilon) \cdots c(d_n,p,\varepsilon). \]
Here are the three main propositions needed in the proofs.
\begin{proposition}\label{prop:1mainterm}
Suppose that $1<p<\infty$ and that $\delta,\lambda,N$ are real numbers such that $0<\delta\leq 1$ and $0<\lambda\leq N$.
\begin{itemize}
\item[(a)] If $f\colon\mathbb{R}^{D}\to[0,1]$ is a measurable function supported in $[0,N]^D$ and satisfying\linebreak $\int_{[0,N]^D} f \geq \delta N^D$, then
\[ \mathcal{M}^{p,1}_{\lambda}(f) \gtrsim_{D,\delta} N^D. \]
\item[(b)] If $f\colon\mathbb{R}^{2D}\to[0,1]$ is a measurable function supported in $[0,N]^{2D}$ and satisfying $\int_{[0,N]^{2D}} f \geq \delta N^{2D}$, then
\[ \widetilde{\mathcal{M}}^{p,1}_{\lambda}(f) \gtrsim_{D,\delta} N^{2D}. \]
\end{itemize}
\end{proposition}
\begin{proposition}\label{prop:2difference}
Suppose that $\varepsilon,\lambda,N$ are real numbers such that $0<\varepsilon<1$ and $0<\lambda\leq N$.
\begin{itemize}
\item[(a)] If $d_j\geq 5$ for $j=1,2,\ldots,n$ and if $f\colon\mathbb{R}^{D}\to[0,1]$ is a measurable function supported in $[0,N]^D$, then
\[ \big| \mathcal{N}^{2}_{\lambda}(f) - \mathcal{M}^{2,\varepsilon}_{\lambda}(f) \big| \lesssim_D \varepsilon^{1/4} N^D. \]
\item[(b)] Additionally, take $1<p<\infty$, $p\neq 2$. If each $d_j$ is sufficiently large for $j=1,2,\ldots,n$ and if $f\colon\mathbb{R}^{2D}\to[0,1]$ is a measurable function supported in $[0,N]^{2D}$, then
\[ \big| \widetilde{\mathcal{N}}^{p}_{\lambda}(f) - \widetilde{\mathcal{M}}^{p,\varepsilon}_{\lambda}(f) \big| \lesssim_D \varepsilon N^{2D}. \]
\end{itemize}
\end{proposition}
\begin{proposition}\label{prop:3multiscale}
Suppose that $1<p<\infty$, $0<\varepsilon<1/10D$, and that $\lambda_1,\lambda_2,\ldots,\lambda_M$ are positive numbers such that $\lambda_{m+1}/\lambda_{m}\geq 2$ for $m=1,2,\ldots,M-1$.
\begin{itemize}
\item[(a)] If $f\colon\mathbb{R}^{D}\to[0,1]$ is a measurable function supported in $[0,N]^D$, then
\[ \sum_{m=1}^{M} |\mathcal{E}^{p,\varepsilon}_{\lambda_m}(f)| \lesssim_{D,\varepsilon} N^D. \]
\item[(b)] If $f\colon\mathbb{R}^{2D}\to[0,1]$ is a measurable function supported in $[0,N]^{2D}$, then
\[ \Big( \sum_{m=1}^{M} |\widetilde{\mathcal{E}}^{p,\varepsilon}_{\lambda_m}(f)|^2 \Big)^{1/2} \lesssim_{D,\varepsilon} N^{2D}. \]
\end{itemize}
\end{proposition}
Proofs of Propositions~\ref{prop:1mainterm}--\ref{prop:3multiscale} are postponed to the later sections. Now we show how they imply Theorems~\ref{thm:main1} and \ref{thm:main3}.
\begin{proof}[Proof of Theorem~\ref{thm:main1}]
We argue by contradiction and suppose that there is a set $A$ with strictly positive upper Banach density $\overline{\delta}(A)$ for which the claim does not hold: there exists a sequence $(\lambda_m)_{m=1}^{\infty}$ such that $\lim_{m\to\infty}\lambda_m=\infty$ and that, for each $m$, the set $A$ contains no boxes \eqref{eq:box1} with $\|s_j\|_{\ell^2} = \lambda_m a_j$, $j=1,2,\ldots,n$. By omitting some terms we can achieve $\lambda_{m+1}/\lambda_{m}\geq 2$ for each $m$. Fix an arbitrary positive integer $M$.
For $\delta:=\overline{\delta}(A)/2>0$, by the definition of the upper Banach density \eqref{eq:defofdensity}, we can find $N\geq\lambda_M$ and $\mathbf{x}\in\mathbb{R}^D$ such that
\[ \big| A\cap(\mathbf{x}+[0,N]^D) \big| \geq \delta N^D. \]
Denote $A':=(-\mathbf{x}+A)\cap[0,N]^D$ and $f=\mathbbm{1}_{A'}$, so that $A'$ is now a measurable subset of $[0,N]^D$ satisfying
\begin{equation}\label{eq:thmgrtdelta}
\int_{[0,N]^D}f(\mathbf{x})\,\textup{d}\mathbf{x} = |A'| \geq \delta N^D
\end{equation}
and it still has no boxes \eqref{eq:box1} of the previously described sizes determined by $\lambda$.
Consequently,
\begin{equation}\label{eq:thmaux1}
\mathcal{N}^{2}_{\lambda_m}(f)=0
\end{equation}
for $m=1,2,\ldots,M$. Because of condition \eqref{eq:thmgrtdelta}, we can apply part (a) of Proposition~\ref{prop:1mainterm} and get
\begin{equation}\label{eq:thmaux2}
\mathcal{M}^{2,1}_{\lambda_m}(f) \gtrsim_{D,\delta} N^D,
\end{equation}
again for each $m=1,2,\ldots,M$. Moreover, \eqref{eq:thmaux1}, part (a) of Proposition~\ref{prop:2difference} and \eqref{eq:thmaux2} together give
\begin{equation}\label{eq:thmaux3}
\mathcal{M}^{2,\varepsilon}_{\lambda_m}(f) = \big| \mathcal{N}^{2}_{\lambda_m}(f) - \mathcal{M}^{2,\varepsilon}_{\lambda_m}(f) \big|
\lesssim_D \varepsilon^{1/4} N^D \lesssim_{D,\delta} \varepsilon^{1/4} \mathcal{M}^{2,1}_{\lambda_m}(f).
\end{equation}
By \eqref{eq:thmaux3} and \eqref{eq:omegaquotient}, for a sufficiently small $\varepsilon$ depending on the dimensions and $\delta$, we have
\[ \mathcal{M}^{2,\varepsilon}_{\lambda_m}(f) \leq \frac{1}{2} b(2,\varepsilon) \mathcal{M}^{2,1}_{\lambda_m}(f), \]
so,
\[ |\mathcal{E}^{2,\varepsilon}_{\lambda_m}(f)| = b(2,\varepsilon) \mathcal{M}^{2,1}_{\lambda_m}(f) - \mathcal{M}^{2,\varepsilon}_{\lambda_m}(f)
\geq \frac{1}{2} b(2,\varepsilon) \mathcal{M}^{2,1}_{\lambda_m}(f). \]
By \eqref{eq:omegaquotient} and \eqref{eq:thmaux2} again, we conclude
\begin{equation}\label{eq:thmaux4}
|\mathcal{E}^{2,\varepsilon}_{\lambda_m}(f)| \gtrsim_{D,\delta} N^D
\end{equation}
for each $m=1,2,\ldots,M$. Summing the lower bound \eqref{eq:thmaux4} in $m$ gives
\begin{equation}\label{eq:thmaux5}
\sum_{m=1}^{M} |\mathcal{E}^{2,\varepsilon}_{\lambda_m}(f)| \gtrsim_{D,\delta} M N^D.
\end{equation}
Finally, combining \eqref{eq:thmaux5} with Proposition~\ref{prop:3multiscale} yields $M\lesssim_{D,\delta}1$, which contradicts the fact that $M$ could have been chosen arbitrarily large.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main3}]
The same outline also applies here. The only difference is that part (b) of Proposition~\ref{prop:2difference} only holds for sufficiently large dimensions $d_j$ depending on $p$.
The reader can also consult the corresponding proofs of Theorem~2.2 in \cite{CMP15:roth} and Theorem~1.2 in \cite{DKR17:corner}.
\end{proof}
\section{Combinatorial results}
\label{sec:combinatorics}
The following lemma is needed in the proof of Proposition~\ref{prop:1mainterm} in the same way in which Bourgain's version of Roth's theorem for compact abelian groups \cite{B86:roth} is needed in the analogous proposition in \cite{CMP15:roth}.
\begin{lemma}\label{lm:1combroth}
Suppose $0<\delta\leq 1$.
\begin{itemize}
\item[(a)] If $f\colon\mathbb{R}^{D}\to[0,1]$ is a measurable function supported in $[0,N]^D$ and satisfying\linebreak $\int_{[0,N]^D} f \geq \delta N^D$, then
\[ \int_{([0,1]^{D})^2} (\mathcal{F}f)(\mathbf{x},\mathbf{s}) \,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x} \gtrsim_{D,\delta} 1. \]
\item[(b)] If $f\colon\mathbb{R}^{2D}\to[0,1]$ is a measurable function supported in $[0,N]^{2D}$ and satisfying $\int_{[0,N]^{2D}} f \geq \delta N^{2D}$, then
\[ \int_{([0,1]^{D})^3} (\widetilde{\mathcal{F}}f)(\mathbf{x},\mathbf{y},\mathbf{s}) \,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y} \gtrsim_{D,\delta} 1. \]
\end{itemize}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lm:1combroth}]
Both parts of the lemma are shown using multidimensional Szemer\'{e}di's theorem of Furstenberg and Katznelson \cite{FK78:msz}.
By this result, for any dimension $n$ and any number $0<\beta\leq 1$ there exists a positive integer $m_{n,\beta}$ such that for each positive integer $m\geq m_{n,\beta}$ one has the following.
\begin{itemize}
\item Each subset $S\subseteq\{0,1,\ldots,m-1\}^n$ of cardinality at least $\beta m^n$ contains (vertices of) an $n$-dimensional cube,
\[ (i_1 + k_1 l,\,i_2 + k_2 l,\,\dots,\,i_n + k_n l), \quad k_1,k_2,\dots,k_n\in\{0,1\} \]
for some $i_1,i_2,\ldots,i_n,l\in\mathbb{Z}$ with $l\neq 0$.
\item Each subset $S\subseteq\{0,1,\ldots,m-1\}^{2n}$ of cardinality at least $\beta m^{2n}$ contains a $2n$-dimensional corner-extended cube,
\begin{align*}
(i_1 + k_1 l,\,i_2 + k_2 l,\,\dots,\,i_n + k_n l,\,j_1, j_2, \dots, j_n), \quad k_1,k_2,\dots,k_n\in\{0,1\}, \\
(i_1, i_2, \dots, i_n,\,j_1 + l, j_2, \dots, j_n),\quad \ldots,\quad (i_1, i_2, \dots, i_n,\,j_1, j_2, \dots, j_n + l)
\end{align*}
for some $i_1,i_2,\ldots,i_n,j_1,j_2,\ldots,j_n,l\in\mathbb{Z}$ with $l\neq 0$.
\end{itemize}
Then one applies the averaging trick of Varnavides \cite{Var59:dens}, in the same way it was done in the proof of Lemma~3.2 in \cite{DKR17:corner} for the particular case of the three-point corners.
\end{proof}
Indeed, multidimensional Szemer\'{e}di's theorem applies to any finite pattern on the integer lattice, not only to boxes and corner-extended boxes, so Lemma~\ref{lm:1combroth} can be generalized easily. The reasons why we restrict our attention to very special patterns lie in the rigidity of other auxiliary results, most notably Proposition~\ref{prop:3multiscale} above and Theorem~\ref{thm:singular} from Section~\ref{sec:analysis}.
\begin{proof}[Proof of Proposition~\ref{prop:1mainterm}]
The proposition is shown by cutting the Euclidean space into cubes, the scaled copies of $([0,1]^{D})^2$ or $([0,1]^{D})^3$, and applying Lemma~\ref{lm:1combroth} on each of them. For details the reader can consult the proof of Proposition~2.1 in \cite{CMP15:roth}.
\end{proof}
Now we turn to the proof of the second proposition. We will need the Euclidean version of the notion of the Gowers norms, so let us begin by setting
\[ (\Delta_{h}g)(s) := g(s) \overline{g(s+h)} \]
for $s,h\in\mathbb{R}^d$ and a function $g\colon\mathbb{R}^d\to\mathbb{C}$. If such $g$ is also measurable, then its \emph{Gowers uniformity norm of degree $k$} is defined as
\begin{align}
\|g\|_{\textup{U}^k(\mathbb{R}^d)} & := \Big( \int_{(\mathbb{R}^d)^{k+1}} (\Delta_{h_k} \cdots \Delta_{h_1} g)(s) \,\textup{d}s \,\textup{d}h_1 \cdots \,\textup{d}h_k \Big)^{2^{-k}} \nonumber \\
& = \Big( \int_{(\mathbb{R}^d)^{k-1}} \Big| \int_{\mathbb{R}^d} (\Delta_{h_{k-1}} \cdots \Delta_{h_1} g)(s) \,\textup{d}s \Big|^2 \,\textup{d}h_1 \cdots \,\textup{d}h_{k-1} \Big)^{2^{-k}} \label{eq:gowersnorm}
\end{align}
We will only need the norms $\|\cdot\|_{\textup{U}^2(\mathbb{R}^d)}$ and $\|\cdot\|_{\textup{U}^3(\mathbb{R}^d)}$. The Gowers norms scale properly with respect to the $\textup{L}^1$-normalized dilations of the function. In particular,
\begin{equation}\label{eq:gowersscaling}
\|g_t\|_{\textup{U}^2(\mathbb{R}^d)} = t^{-d/4} \|g\|_{\textup{U}^2(\mathbb{R}^d)}, \quad
\|g_t\|_{\textup{U}^3(\mathbb{R}^d)} = t^{-d/2} \|g\|_{\textup{U}^3(\mathbb{R}^d)},
\end{equation}
as is shown by an easy change of variables of integration.
\begin{lemma}\label{lm:2combcs}
Suppose that $\lambda$ and $N$ are real numbers such that $0<\lambda\leq N$.
\begin{itemize}
\item[(a)] If $f_1,f_2\colon\mathbb{R}^{d}\to[0,1]$ are measurable functions supported in $[0,N]^d$, and $g\colon\mathbb{R}^{d}\to\mathbb{R}$ is a measurable function supported in $[-3\lambda,3\lambda]^d$, then
\begin{equation}\label{eq:lemmanewa}
\Big| \int_{(\mathbb{R}^{d})^2} f_1(x) f_2(x+s) g(s) \,\textup{d}s \,\textup{d}x \Big| \lesssim_d N^{d} \lambda^{d/4} \|g\|_{\textup{U}^2(\mathbb{R}^d)}.
\end{equation}
\item[(b)] If $f_1,f_2,f_3\colon\mathbb{R}^{2d}\to[0,1]$ are measurable functions supported in $[0,N]^{2d}$ and $g\colon\mathbb{R}^{d}\to\mathbb{R}$ is a measurable function supported in $[-3\lambda,3\lambda]^d$, then
\begin{equation}\label{eq:lemmanewb}
\Big| \int_{(\mathbb{R}^{d})^3} f_1(x,y) f_2(x+s,y) f_3(x,y+s) g(s) \,\textup{d}s \,\textup{d}x \,\textup{d}y \Big| \lesssim_d N^{2d} \lambda^{d/2} \|g\|_{\textup{U}^3(\mathbb{R}^d)}.
\end{equation}
\end{itemize}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lm:2combcs}]
This lemma will be shown in a way similar to the proof of Lemma 4.2 in \cite{CMP15:roth}.
Proof of (a).
The left hand side of \eqref{eq:lemmanewa} can be rewritten as
\[ \Big| \int_{[0,N]^d} f_1(x) \Big( \int_{\mathbb{R}^d} f_2(x+s) g(s) \,\textup{d}s \Big) \,\textup{d}x \Big|, \]
so the Cauchy--Schwarz inequality in $x$ bounds it by
\[ N^{d/2} \bigg( \int_{(\mathbb{R}^d)^3} f_2(x+s) f_2(x+s') g(s) g(s') \,\textup{d}s \,\textup{d}s' \,\textup{d}x \bigg)^{1/2}. \]
Substituting $y=x+s$, $h=s'-s$, applying the Cauchy--Schwarz inequality again, and recognizing the $\textup{U}^2$-norm from \eqref{eq:gowersnorm}, we get
\[ N^{d/2} \|g\|_{\textup{U}^2(\mathbb{R}^d)} \bigg( \int_{[-6\lambda,6\lambda]^d} \Big( \int_{[0,N]^d} f_2(y) f_2(y+h) \,\textup{d}y \Big)^2 \,\textup{d}h \bigg)^{1/4}. \]
The expression within the outer parentheses is clearly at most a constant times $N^{2d}\lambda^{d}$.
Proof of (b).
Using the Cauchy--Schwarz inequality we first bound the left hand side of \eqref{eq:lemmanewb} by
\[ N^d \bigg( \int_{(\mathbb{R}^{d})^4} f_2(x+s,y) f_2(x+s',y) f_3(x,y+s) f_3(x,y+s') g(s) g(s') \,\textup{d}s \,\textup{d}s' \,\textup{d}x \,\textup{d}y \bigg)^{1/2}. \]
Then we substitute $h=s'-s$ and perform the shift $x\mapsto x-s$ to transform the integral inside parentheses into
\[ \int_{(\mathbb{R}^{d})^4} f_2(x,y) f_2(x+h,y) f_3(x-s,y+s) f_3(x-s,y+s+h) g(s) g(s+h) \,\textup{d}x \,\textup{d}y \,\textup{d}s \,\textup{d}h. \]
Another application of the Cauchy--Schwarz inequality controls the left hand side of \eqref{eq:lemmanewb} with
\begin{align*}
N^{3d/2} \lambda^{d/4} \bigg( \int_{(\mathbb{R}^{d})^5} & f_3(x-s,y+s) f_3(x-s',y+s') f_3(x-s,y+s+h) f_3(x-s',y+s'+h) \\
& g(s) g(s') g(s+h) g(s'+h) \,\textup{d}s \,\textup{d}s' \,\textup{d}x \,\textup{d}y \,\textup{d}h \bigg)^{1/4}.
\end{align*}
It remains to substitute $h'=s'-s$, shift $x\mapsto x+s$, $y\mapsto y-s$, use the Cauchy--Schwarz inequality one more time, and finally recognize $\|g\|_{\textup{U}^3(\mathbb{R}^d)}$ from \eqref{eq:gowersnorm}.
\end{proof}
\begin{lemma}\label{lm:3combgowers}
Suppose that $\eta$ and $\varepsilon$ are real numbers such that $0<\eta<\varepsilon<1$.
\begin{itemize}
\item[(a)] For $d\geq 5$ we have
\[ \big\| \omega^{d,2,\eta} - \omega^{d,2,\varepsilon} \big\|_{\textup{U}^2(\mathbb{R}^d)} \lesssim_{d} \varepsilon^{1/4}. \]
\item[(b)] For $1<p<\infty$, $p\neq 2$ and sufficiently large $d$ depending on $p$ we have
\[ \big\| \omega^{d,p,\eta} - \omega^{d,p,\varepsilon} \big\|_{\textup{U}^3(\mathbb{R}^d)} \lesssim_{p,d} \varepsilon. \]
\end{itemize}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lm:3combgowers}]
Part (b) was already established in \cite{CMP15:roth}; it is Lemma~2.4 of that paper.
We will show part (a) using the same lines of proof, but we need to use to our advantage the fact that we only need the $\textup{U}^2$-norm and get more concrete decay in $\varepsilon$ as $\varepsilon\to0^+$.
Let $\varphi\colon\mathbb{R}\to[0,\infty)$ be a compactly supported $\textup{C}^\infty$ function that is constantly equal to $1$ on $[-3,3]$ and set
\[ \Phi = \underbrace{\varphi\otimes\cdots\otimes\varphi}_{d}. \]
All constants are assumed to depend on $\varphi$ without further mention.
Observe that $\omega^{d,2,\eta}$ and $\omega^{d,2,\varepsilon}$ are supported on $[-3,3]^d$, so
\[ \omega^{d,2,\eta}(s) - \omega^{d,2,\varepsilon}(s)
= \int_{\mathbb{R}} ( \psi(\eta u) - \psi(\varepsilon u) ) \Phi(s) e^{2\pi i u (\|s\|_{\ell^2}^2-1)} \,\textup{d}u. \]
Integral version of the triangle inequality for the Gowers norm and the tensor product splitting of the exponential give
\begin{equation}\label{eq:u2aux1}
\big\| \omega^{d,2,\eta} - \omega^{d,2,\varepsilon} \big\|_{\textup{U}^2(\mathbb{R}^d)}
\leq \int_{\mathbb{R}} | \psi(\eta u) - \psi(\varepsilon u) | \big\| \varphi(s) e^{2\pi i u s^2} \big\|_{\textup{U}^2_{s}(\mathbb{R})}^d \,\textup{d}u.
\end{equation}
By definition of the Gowers norm \eqref{eq:gowersnorm},
\[ \big\| \varphi(s) e^{2\pi i u s^2} \big\|_{\textup{U}^2_{s}(\mathbb{R})}^4
= \int_{[-3,3]} \Big| \int_{\mathbb{R}} \varphi(s) \varphi(s+h) e^{-4 \pi i u h s} \,\textup{d}s \Big|^2 \,\textup{d}h, \]
which is certainly bounded by a constant, for each $u\in\mathbb{R}$. However, for $|u|\geq 1$ we get a better estimate by splitting the outer domain of integration into $|h|\leq|u|^{-1}$ and $|u|^{-1}<|h|\leq 3$. The first part of the integral is clearly at most a constant times $|u|^{-1}$. Integration by parts in the second part gives
\[ \Big| \int_{\mathbb{R}} \varphi(s) \varphi(s+h) e^{-4 \pi i u h s} u h \,\textup{d}s \Big| \lesssim 1, \]
so that
\[ \int_{\{|u|^{-1}<|h|\leq3\}} \Big| \int_{\mathbb{R}} \varphi(s) \varphi(s+h) e^{-4 \pi i u h s} \,\textup{d}s \Big|^2 \,\textup{d}h
\lesssim |u|^{-2} \int_{\{|u|^{-1}<|h|\leq3\}} h^{-2} \,\textup{d}h \lesssim |u|^{-1}. \]
From these we conclude
\begin{equation}\label{eq:u2aux2}
\big\| \varphi(s) e^{2\pi i u s^2} \big\|_{\textup{U}^2_{s}(\mathbb{R})} \lesssim \min\{1,|u|^{-1/4}\}.
\end{equation}
Now we combine \eqref{eq:u2aux1} and \eqref{eq:u2aux2} into a single estimate
\[ \big\| \omega^{d,2,\eta} - \omega^{d,2,\varepsilon} \big\|_{\textup{U}^2(\mathbb{R}^d)}
\lesssim \int_{\mathbb{R}} | \psi(\eta u) - \psi(\varepsilon u) | \min\{1,|u|^{-d/4}\} \,\textup{d}u. \]
This time we split the domain of integration into three parts: $|u|\leq 1$, $1<|u|\leq\varepsilon^{-1}$, and $|u|>\varepsilon^{-1}$.
We bound the corresponding integrals respectively as
\[ \int_{\{|u|\leq 1\}} \|\psi'\|_{\textup{L}^\infty(\mathbb{R})} \varepsilon |u| \,\textup{d}u \lesssim \varepsilon \lesssim \varepsilon^{1/4}, \]
\[ \int_{\{1<|u|\leq\varepsilon^{-1}\}} \|\psi'\|_{\textup{L}^\infty(\mathbb{R})} \varepsilon |u|^{1-d/4} \,\textup{d}u \lesssim_d \varepsilon^{d/4-1} \lesssim \varepsilon^{1/4}, \]
and
\[ \int_{\{|u|>\varepsilon^{-1}\}} 2\|\psi\|_{\textup{L}^\infty(\mathbb{R})} |u|^{-d/4} \,\textup{d}u \lesssim_d \varepsilon^{d/4-1} \lesssim \varepsilon^{1/4}. \]
In the last display we needed $d>4$ for the convergence of the improper integral and also to have $d/4-1\geq 1/4>0$.
This allows us to conclude the desired inequality.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:2difference}]
Both parts of the proposition are shown in exactly the same way, using Lemmata~\ref{lm:2combcs} and \ref{lm:3combgowers}, so we only elaborate on the proof of part (a).
Because of \eqref{eq:limits} it is enough to bound the difference
\begin{equation}\label{eq:diffaux}
\mathcal{M}^{2,\eta}_{\lambda}(f)-\mathcal{M}^{2,\varepsilon}_{\lambda}(f)
\end{equation}
for all $0<\eta<\varepsilon<1$, with a constant independent of $\eta$. The difference of the corresponding cutoff functions can be expanded as
\[ \omega^{p,\eta}_{\lambda\mathbf{a}}(\mathbf{s}) - \omega^{p,\varepsilon}_{\lambda\mathbf{a}}(\mathbf{s})
= \sum_{j=1}^{n} \tau^{(j)}_{\lambda}(\mathbf{s}), \]
where
\[ \tau^{(j)}(\mathbf{s}) := \Big( \prod_{i=1}^{j-1} \omega^{d_i,p,\eta}_{a_i}(s_i) \Big)
\big( \omega^{d_j,p,\eta}_{a_j}(s_j) - \omega^{d_j,p,\varepsilon}_{a_j}(s_j) \big)
\Big( \prod_{i=j+1}^{n} \omega^{d_i,p,\varepsilon}_{a_i}(s_i) \Big). \]
This decomposes \eqref{eq:diffaux} into $n$ pieces, $\sum_{j=1}^{n}\mathcal{P}^{(j)}_{\lambda}(f)$, where
\[ \mathcal{P}^{(j)}_{\lambda}(f) := \int_{(\mathbb{R}^{D})^2} (\mathcal{F}f)(\mathbf{x},\mathbf{s}) \tau^{(j)}_{\lambda}(\mathbf{s}) \,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x}. \]
Without loss of generality we will estimate the piece $\mathcal{P}^{(1)}_{\lambda}(f)$. Part (a) of Lemma~\ref{lm:3combgowers} and \eqref{eq:gowersscaling} give
\[ \big\| \| \tau^{(j)}_{\lambda}(\mathbf{s}) \|_{\textup{U}^2_{s_1}(\mathbb{R}^{d_1})} \big\|_{\textup{L}^{\infty}_{s_2,\ldots,s_n}(\mathbb{R}^{D-d_1})}
\lesssim_D \lambda^{-d_{1}/4} \varepsilon^{1/4} \lambda^{-(d_2+\cdots+d_n)} = \varepsilon^{1/4} \lambda^{-D+3d_1/4}. \]
Then we observe that $(\mathcal{F}f)(\mathbf{x},\mathbf{s})$ can, for fixed $x_2,\ldots,x_n$, $s_2,\ldots,s_n$, be written in the form $f_1(x_1) f_2(x_1+s_1)$ from Lemma~\ref{lm:2combcs}, so we obtain
\[ \Big| \int_{(\mathbb{R}^{d_1})^2} (\mathcal{F}f)(\mathbf{x},\mathbf{s}) \tau^{(j)}_{\lambda}(\mathbf{s}) \,\textup{d}s_1 \,\textup{d}x_1 \Big|
\leq N^{d_1} \lambda^{d_{1}/4} \varepsilon^{1/4} \lambda^{-D+3d_{1}/4} = \varepsilon^{1/4} N^{d_1} \lambda^{-D+d_1}. \]
Integrating in $x_j\in[0,N]^{d_j}$ and $s_j\in[-3\lambda,3\lambda]^{d_j}$, $j=1,2,\ldots,n$ we finally get
\[ \big|\mathcal{P}^{(1)}_{\lambda}(f)\big| \lesssim_{D} \varepsilon^{1/4} N^{d_1} \lambda^{-D+d_1} N^{D-d_1} \lambda^{D-d_1} = \varepsilon^{1/4} N^D, \]
which completes the proof.
\end{proof}
\section{Analytical results}
\label{sec:analysis}
The main ingredient in the proof of Proposition~\ref{prop:3multiscale} is an estimate for multilinear singular integral forms. We formulate it as a separate theorem.
\begin{theorem}\label{thm:singular}\ \
\begin{itemize}
\item[(a)]
Suppose that $K\colon\mathbb{R}^{d_1}\times\cdots\times\mathbb{R}^{d_n}\to\mathbb{C}$ is a bounded compactly supported function and that its Fourier transform satisfies the standard symbol estimates (cf.\@ \cite{S93:habook}),
\begin{equation}\label{eq:symbolest}
|\widehat{K}(\xi)| \leq C_{\kappa} \|\xi\|_{\ell^2}^{-|\kappa|}
\end{equation}
for any multi-index $\kappa$. Then we have the inequality
\begin{equation}\label{est:thm10a}
\Big| \int_{(\mathbb{R}^{D})^2} K(\mathbf{s}) \prod_{\mathbf{k}\in\{0,1\}^n} F_{\mathbf{k}}(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n) \,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x} \Big| \lesssim \prod_{\mathbf{k}\in\{0,1\}^n} \|F_{\mathbf{k}}\|_{\textup{L}^{2^n}}
\end{equation}
with the implicit constant depending only on $(C_{\kappa})_\kappa$ and the dimensions $d_i$.
\item[(b)]
Suppose that $K\colon\mathbb{R}^{d_1}\times\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\times\cdots\times\mathbb{R}^{d_n}\to\mathbb{C}$ is a bounded compactly supported function such that its Fourier transform satisfies the standard symbol estimates \eqref{eq:symbolest} for any multi-index $\kappa$. Then we have the inequality
\begin{align*}\nonumber
\bigg| \int_{\mathbb{R}^{3D+d_1}}
& K(s_1,s'_1,s_2,\ldots,s_n)
\Big( \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \\
& F_{\widetilde{\mathbf{k}}}(x_1 + l_1 s_1 + l_2 s'_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1 + l_3 s_1 + l_4 s'_1,\, y_2,\dots, y_n) \Big) \\
& \,\textup{d}s_1 \,\textup{d}s'_1 \,\textup{d}s_2 \cdots \,\textup{d}s_n \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y} \bigg|
\lesssim \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \|F_{\widetilde{\mathbf{k}}}\|_{\textup{L}^{2^{n+1}}},
\end{align*}
with the implicit constant depending only on $(C_{\kappa})_\kappa$ and the dimensions $d_i$.
\end{itemize}
\end{theorem}
Note that the implicit constants in both inequalities claimed by Theorem~\ref{thm:singular} depend only on the implicit constants from the symbol estimates \eqref{eq:symbolest} and the dimensions, the latter being regarded as fixed throughout the text. Also observe that part (b) of the theorem specialized to $n=1$ coincides with Theorem~1.3 from \cite{DKR17:corner}, the main analytic result of that paper.
Once Theorem~\ref{thm:singular} is established, it is easy to complete the proof of Proposition~\ref{prop:3multiscale}. Let us elaborate on that argument and postpone the proof of the theorem to the second half of this section.
\begin{proof}[Proof of Proposition~\ref{prop:3multiscale}]
For the proof of part (a), note that the cutoff function appearing in $\mathcal{E}^{p,\varepsilon}_{\lambda_m}(f)$ is
\[ \omega^{p,\varepsilon}_{\lambda_m \mathbf{a}}(\mathbf{s}) - b(p,\varepsilon) \omega^{p,1}_{\lambda_m \mathbf{a}}(\mathbf{s})
= \prod_{j=1}^{n} \omega^{d_j,p,\varepsilon}_{\lambda_m a_j}(s_j) - \prod_{j=1}^{n} c(d_j,p,\varepsilon) \omega^{d_j,p,1}_{\lambda_m a_j}(s_j). \]
Recalling the introduction of $k^{d,p,\varepsilon}$ in \eqref{eq:defofk}, we can rewrite it as
\begin{equation}\label{eq:ekernel}
\sum_{j=1}^{n} \Big(\prod_{i=1}^{j-1} \omega^{d_i,p,\varepsilon}_{\lambda_m a_i}(s_i)\Big) k^{d_j,p,\varepsilon}_{\lambda_m a_j}(s_j) \Big(\prod_{i=j+1}^{n} c(d_i,p,\varepsilon) \omega^{d_i,p,1}_{\lambda_m a_i}(s_i)\Big).
\end{equation}
Note that each summand in \eqref{eq:ekernel} is of the form
\begin{equation}\label{eq:ekernel2}
\varphi^{(1)}_{\lambda_m}(s_1) \varphi^{(2)}_{\lambda_m}(s_2) \dots \varphi^{(n)}_{\lambda_m}(s_n),
\end{equation}
where $\varphi^{(j)}$, $j=1,2,\ldots,n$, are $\textup{C}^1$ functions and one of them has integral equal to $0$, while the others are nonnegative.
Take arbitrary signs $\alpha_m\in\{-1,1\}$, $m=1,2,\ldots,M$. By a standard computation (see \cite{CMP15:roth}) the kernels
\[ K(\mathbf{s}) := \sum_{m=1}^{M} \alpha_m \varphi^{(1)}_{\lambda_m}(s_1) \varphi^{(2)}_{\lambda_m}(s_2) \dots \varphi^{(n)}_{\lambda_m}(s_n) \]
satisfy the conditions from Theorem~\ref{thm:singular}, with constants $C_\kappa$ independent of the numbers $M$, $\lambda_1,\ldots,\lambda_M$ and signs $\alpha_1,\ldots,\alpha_M$, but we allow dependencies on the dimensions (i.e.\@ $D$), on $\varepsilon$, on the numbers $a_1,\ldots,a_n$, and on the exponent $p$.
Applying part (a) of Theorem~\ref{thm:singular} to those kernels and $F_\mathbf{k}=f$, we obtain
\[ \Big| \sum_{m=1}^{M} \alpha_m \mathcal{E}^{p,\varepsilon}_{\lambda_m}(f) \Big| \lesssim_{D,\varepsilon} \|f\|_{\textup{L}^{2^n}}^{2^n} \leq N^D. \]
It remains to choose the signs $\alpha_m$ appropriately, so that the left hand side becomes\linebreak $\sum_{m=1}^{M} |\mathcal{E}^{p,\varepsilon}_{\lambda_m}(f)|$.
In the proof of part (b) we begin with the same splitting \eqref{eq:ekernel} into summands of the form \eqref{eq:ekernel2}.
Since the notation has become symmetric in $j$, without loss of generality we can suppose $\int_{\mathbb{R}^{d_1}} \varphi^{(1)}=0$, i.e.\@ the cancellation comes from the variable $s_1$.
Gathering inside parentheses all factors containing that variable, the corresponding part of $\widetilde{\mathcal{E}}^{p,\varepsilon}_{\lambda_m}(f)$ can be rewritten as
{\allowdisplaybreaks\begin{align*}
\int_{\mathbb{R}^{3D-d_1}}
& \bigg( \int_{\mathbb{R}^{d_1}} \Big( \prod_{(k_2,\ldots,k_n)\in\{0,1\}^{n-1}} f(x_1 + s_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1, \dots, y_n) \Big) \\
& \qquad\quad f(x_1, \dots, x_n,\,y_1 + s_1, \dots, y_n) \varphi^{(1)}_{\lambda_m}(s_1) \,\textup{d}s_1 \bigg) \\
& \Big( \prod_{(k_2,\ldots,k_n)\in\{0,1\}^{n-1}} f(x_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1, \dots, y_n) \Big) \\
& \qquad\quad f(x_1, \dots, x_n,\,y_1, y_2 + s_2, \dots, y_n) \cdots f(x_1, \dots, x_n,\,y_1, y_2,\dots, y_n + s_n) \\
& \varphi^{(2)}_{\lambda_m}(s_2) \dots \varphi^{(n)}_{\lambda_m}(s_n) \,\textup{d}s_2 \cdots \,\textup{d}s_n \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y}.
\end{align*}}
By the Cauchy--Schwarz inequality its square is bounded with
\begin{equation}\label{eq:propaux1}
\mathcal{A}_{\lambda_m}(f) \mathcal{B}_{\lambda_m}(f),
\end{equation}
where
\begin{align*}
\mathcal{A}_{\lambda}(f) := \int_{\mathbb{R}^{3D+d_1}}
& \Big( \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \\
& f_{\widetilde{\mathbf{k}}}(x_1 + l_1 s_1 + l_2 s'_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1 + l_3 s_1 + l_4 s'_1,\, y_2,\dots, y_n) \Big) \\
& \varphi^{(1)}_{\lambda}(s_1) \varphi^{(1)}_{\lambda}(s'_1) \varphi^{(2)}_{\lambda}(s_2) \dots \varphi^{(n)}_{\lambda}(s_n) \,\textup{d}s_1 \,\textup{d}s'_1 \,\textup{d}s_2 \cdots \,\textup{d}s_n \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y}
\end{align*}
and
\begin{align*}
\mathcal{B}_{\lambda}(f) := \int_{\mathbb{R}^{3D-d_1}}
& \Big( \prod_{(k_2,\ldots,k_n)\in\{0,1\}^{n-1}} \mathbbm{1}_{[0,N]^{2D}}(x_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1, \dots, y_n) \Big) \\
& \mathbbm{1}_{[0,N]^{2D}}(x_1, \dots, x_n,\,y_1, y_2 + s_2, \dots, y_n) \\
& \cdots \mathbbm{1}_{[0,N]^{2D}}(x_1, \dots, x_n,\,y_1, y_2,\dots, y_n + s_n) \\
& \varphi^{(2)}_{\lambda}(s_2) \dots \varphi^{(n)}_{\lambda}(s_n) \,\textup{d}s_2 \cdots \,\textup{d}s_n \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{y}.
\end{align*}
The functions $f_{\widetilde{\mathbf{k}}}$ appearing in the definition of $\mathcal{A}_{\lambda}(f)$ are all supported in $[0,N]^{2D}$ and taking values in $[0,1]$; some of them are equal to $f$, while the others are artificially inserted into the expression as $\mathbbm{1}_{[0,2N]^{2D}}$.
Clearly,
\begin{equation}\label{eq:propaux2}
\mathcal{B}_{\lambda_m}(f)
\leq \Big( \prod_{j=2}^{n} \int_{\mathbb{R}^{d_j}} \varphi^{(j)}_{\lambda_m} \Big) \Big( \prod_{j=1}^{n} \int_{\mathbb{R}^{d_j}} \mathbbm{1}_{[0,N]^{d_j}} \Big)^2
\lesssim_{D,\varepsilon} N^{2D}.
\end{equation}
On the other hand, part (b) of Theorem~\ref{thm:singular} applied with the kernel
\[ K(\mathbf{s}) := \sum_{m=1}^{M} \varphi^{(1)}_{\lambda_m}(s_1) \varphi^{(1)}_{\lambda_m}(s'_1) \varphi^{(2)}_{\lambda_m}(s_2) \dots \varphi^{(n)}_{\lambda_m}(s_n) \]
and the functions $F_{\widetilde{\mathbf{k}}}=f_{\widetilde{\mathbf{k}}}$ yields
\begin{equation}\label{eq:propaux3}
\Big| \sum_{m=1}^{M} \mathcal{A}_{\lambda_m}(f) \Big|
\lesssim_{D,\varepsilon} N^{2D}.
\end{equation}
Summing the products \eqref{eq:propaux1} in $m=1,2,\ldots,M$, from \eqref{eq:propaux2} and \eqref{eq:propaux3} we finally conclude
\[ \sum_{m=1}^{M} |\widetilde{\mathcal{E}}^{p,\varepsilon}_{\lambda_m}(f)|^2 \lesssim_{D,\varepsilon} N^{4D}, \]
as claimed.
\end{proof}
We finalize the paper with the proof of the remaining analytical result.
As we will soon see, the case when all dimensions $d_i$ are equal will be an easy consequence of the main result from the paper by Thiele and one of the present authors \cite{DT18}.
We will spend just a slight additional effort to reduce the case of possibly different dimensions $d_i$ to the very same result.
An alternative to this addition could be considering appropriate lower-dimensional sections of the sets and functions appearing throughout the paper.
\begin{proof}[Proof of Theorem~\ref{thm:singular}]
Proof of (a).
First we consider the case of equal dimensions, i.e.\@ $d_1 = d_2 =\cdots = d_n$, and we write them simply as $d$.
Relabeling $x_i$ to $x_i^0$ in the left hand-side of \eqref{est:thm10a} and changing variables $x_i^0 + s_i = x_i^1$
for $i=1,\ldots,n$, we see that we need to show
\begin{align*}
\Big| \int_{\mathbb{R}^{2dn}} \prod_{\mathbf{k}\in \{0,1\}^n} F_{\mathbf{k}} (\Pi_\mathbf{k} x) K(\Pi x) \,\textup{d}x \Big |
\lesssim_{(C_{\kappa})_\kappa} \prod_{\mathbf{k}\in\{0,1\}^n} \|F_{\mathbf{k}}\|_{\textup{L}^{2^n}(\mathbb{R}^{nd})},
\end{align*}
where $x=(x_1^0,\ldots, x_n^0,x_1^1,\ldots x_n^1)\in (\mathbb{R}^d)^{2n}$ and $\Pi_\mathbf{k},\Pi : (\mathbb{R}^{d})^{2n} \rightarrow (\mathbb{R}^{d})^n$ are linear operators given by
\begin{align*}
\Pi_\mathbf{k}x := (x_1^{k_1},\ldots, x_n^{k_n}),\quad \Pi x := (x_1^1-x_1^0,\ldots, x_n^1-x_n^0).
\end{align*}
This estimate can be recognized as one of the singular Brascamp--Lieb inequalities from the main theorem of \cite{DT18}, which establishes the claim in the case of equal dimensions.
The general case of different dimensions in \eqref{est:thm10a} will be deduced from the case of equal dimensions as follows. By approximating $K$ with smooth compactly supported functions in $\textup{L}^1$ and applying H\"older's inequality, we may assume that $K$ is a smooth compactly supported function on $\mathbb{R}^D \cong \mathbb{R}^{d_1}\times \cdots \times \mathbb{R}^{d_n}$, satisfying the symbol estimates \eqref{eq:symbolest}. We set
\[d:=\max_{1\leq i \leq n} d_i+1.\]
(We have added $1$ for technical reasons, so that $d-d_i>0$ for all $1\leq i \leq n$.)
First we define a new function $\widetilde{K}$ on $\mathbb{R}^{nd}$, whose integration in a certain direction gives $K$.
To achieve this take a smooth compactly supported function $\varphi$ on $\mathbb{R}^{nd-D}\cong \mathbb{R}^{d-d_1}\times \cdots \times \mathbb{R}^{d-d_n}$ satisfying $\varphi(\mathbf{0})=1$. Then we define $\widetilde{K}$ by setting
\begin{align}\label{eq:defKtilde}
\widehat{\widetilde{K}}(\xi,\widetilde{\xi}) := \widehat{K}(\xi) \varphi \Big ( \frac{\widetilde{\xi}}{\|\xi\|} \Big ).
\end{align}
for $\mathbf{0}\neq \xi\in \mathbb{R}^{D}$, $\widetilde{\xi}\in \mathbb{R}^{nd-D}$.
Observe that we have
\begin{align*}
\widehat{\widetilde{K}}(\xi,0) = \widehat{K}(\xi)
\end{align*}
or, passing to the spatial side,
\begin{align*}
K(\mathbf{s}) = \int_{\mathbb{R}^{nd-D}} \widetilde{K}(\mathbf{s},\mathbf{\widetilde{s}}) \,\textup{d}\mathbf{\widetilde{s}}.
\end{align*}
Moreover, the function $\widehat{\widetilde{K}}$ satisfies the symbol estimates
\begin{equation*}
\big|\widehat{\widetilde{K}}(\xi,\widetilde{\xi})\big| \leq \widetilde{C}_{\kappa} \|(\xi,\widetilde{\xi})\|_{\ell^2}^{-|\kappa|}
\end{equation*}
for all multi-indices $\kappa$ and all $(\xi,\widetilde{\xi})\neq 0$, with $\widetilde{C}_{\kappa}$ depending only on $C_\kappa$.
Assuming the estimate \eqref{est:thm10a} in the case of equal dimensions,
let us plug in the kernel $\widetilde{K}$ and the functions $\widetilde{F}_{\mathbf{k}}: (\mathbb{R}^d)^n\rightarrow \mathbb{C}$ defined by
\begin{align*}
\widetilde{F}_{\mathbf{k}}(\mathbf{z},\mathbf{\widetilde{z}}) := F_{\mathbf{k}}(\mathbf{z}) (\textup{D}_{\lambda}^{2^n}\varphi_1\otimes \cdots \otimes \textup{D}_{\lambda}^{2^n}\varphi_n)(\mathbf{\widetilde{z}}),
\end{align*}
where $\mathbf{z}\in \mathbb{R}^{D},\,\mathbf{\widetilde{z}}\in \mathbb{R}^{nd-D}$, $F_{\mathbf{k}}:\mathbb{R}^{D} \rightarrow \mathbb{C}$, $\varphi_i$ is a Schwartz function on $\mathbb{R}^{d-d_i}$ and $\textup{D}_{\lambda}^{2^n} \varphi_i$ is defined by \eqref{eq:funcdilatep}. Then
the form in question becomes
\begin{align}\nonumber
\int_{\mathbb{R}^{2nd}} \widetilde{K}(\mathbf{s},\mathbf{\widetilde{s}})\, \Big (&\, \prod_{\mathbf{k}\in\{0,1\}^n} F_{\mathbf{k}}(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n) \\
& \textup{D}_{\lambda}^{2^n}\varphi_1(\widetilde{x}_1 + k_1 \widetilde{s}_1)\cdots \textup{D}_{\lambda}^{2^n}\varphi_n(\widetilde{x}_n + k_n \widetilde{s}_n) \Big )
\,\textup{d}\mathbf{s}\,\textup{d}\mathbf{\widetilde{s}} \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{\widetilde{x}}, \label{eq:lhs}
\end{align}
where $\mathbf{\widetilde{s}} = (\widetilde{s}_1,\ldots , \widetilde{s}_n),\, \mathbf{\widetilde{x}} = (\widetilde{x}_1,\ldots , \widetilde{x}_n)\in \mathbb{R}^{nd-D}$.
From the case of equal dimensions we know that the last display is bounded by
\begin{align}
\prod_{\mathbf{k}\in \{0,1\}^n} \|\widetilde{F}_{\mathbf{k}}\|_{\textup{L}^{2^n}(\mathbb{R}^{nd})} = \|\varphi_1\otimes \cdots \otimes \varphi_n\|^{2^n}_{\textup{L}^{2^n}(\mathbb{R}^{nd-D})}\prod_{\mathbf{k}\in \{0,1\}^n} \|F_{\mathbf{k}}\|_{\textup{L}^{2^n}(\mathbb{R}^{D})} \label{eq:boundconstant}
\end{align}
times a constant depending only on $(C_\kappa)_\kappa$.
On the other hand, \eqref{eq:lhs} equals
\begin{align}\nonumber
\int_{\mathbb{R}^{2nd}} & \widetilde{K}(\mathbf{s},\mathbf{\widetilde{s}}) \Big( \prod_{i=1}^n \lambda^{-(d-d_i)}\varphi_i(\lambda^{-1}\widetilde{x}_i)^{2^{n-1}}\varphi_i(\lambda^{-1}(\widetilde{x}_i + \widetilde{s}_i))^{2^{n-1}} \Big ) \\
\label{eq:lhs2}
& \prod_{\mathbf{k}\in\{0,1\}^n} F_{\mathbf{k}}(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n) \,\textup{d}\mathbf{s}\,\textup{d}\mathbf{\widetilde{s}} \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{\widetilde{x}} .
\end{align}
Integrating in $\mathbf{\widetilde{x}}$ we obtain
\begin{align}\label{eq:analytic1}
\int_{\mathbb{R}^{nd+D}} \widetilde{K}(\mathbf{s},\mathbf{\widetilde{s}}) \Phi(\lambda^{-1}\mathbf{\widetilde{s}}) \prod_{\mathbf{k}\in\{0,1\}^n} F_{\mathbf{k}}(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n)
\,\textup{d}\mathbf{s}\,\textup{d}\mathbf{\widetilde{s}} \,\textup{d}\mathbf{x},
\end{align}
where we have set
\[\Phi:=\big( \varphi_1^{2^{n-1}}\ast \widetilde{\varphi}_1^{\,2^{n-1}} \big )\otimes \cdots \otimes \big( \varphi_n^{\, 2^{n-1}}\ast \widetilde{\varphi}_n^{2^{n-1}}\big ) \]
and $\widetilde{\varphi}_i(s):=\varphi_i(-s).$
Integrating in $\mathbf{\widetilde{s}}$ and taking the limit as $\lambda\rightarrow \infty$, \eqref{eq:analytic1} becomes, up to a constant,
\begin{align}\label{eq:analytic2}
\int_{(\mathbb{R}^D)^2}{K}(\mathbf{s}) \prod_{\mathbf{k}\in\{0,1\}^n} F_{\mathbf{k}}(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n)
\,\textup{d}\mathbf{s} \,\textup{d}\mathbf{x}
\end{align}
and we know that it is bounded by \eqref{eq:boundconstant}, as desired.
To justify passage to the limit we observe that the difference of \eqref{eq:analytic1} and \eqref{eq:analytic2} equals
\begin{align*}
\int_{(\mathbb{R}^D)^2} \Big ( \int_{\mathbb{R}^{nd-D}} \widetilde{K}(\mathbf{s},\mathbf{\widetilde{s}}) \big(\Phi(\lambda^{-1}\mathbf{\widetilde{s}}) - 1\big) \,\textup{d}\mathbf{\widetilde{s}}\Big ) \prod_{\mathbf{k}\in\{0,1\}^n} F_{\mathbf{k}}(x_1 + k_1 s_1,\,\dots,\,x_n + k_n s_n)
\, \textup{d}\mathbf{s}\, \,\textup{d}\mathbf{x},
\end{align*}
which tends to zero as $\lambda\rightarrow \infty$.
Indeed, this follows by applying
H\"older's inequality in $\mathbf{x}$, which bounds the last display by
\begin{align*}
\Big( \int_{\mathbb{R}^D} \Big | \int_{\mathbb{R}^{nd-D}} \widetilde{K}(\mathbf{s},\mathbf{\widetilde{s}}) \big(\Phi(\lambda^{-1}\mathbf{\widetilde{s}}) - 1\big) \,\textup{d}\mathbf{\widetilde{s}} \, \Big | \,\textup{d}\mathbf{s} \Big ) \prod_{\mathbf{k}\in \{0,1\}^n} \|{F}_{\mathbf{k}}\|_{\textup{L}^{2^n}(\mathbb{R}^{D})} .
\end{align*}
We note that the expression in the bracket
tends to zero as $\lambda \rightarrow \infty$, as desired.
Proof of (b).
Note that it suffices to show the bound
\begin{align}\nonumber
\bigg| \int_{\mathbb{R}^{2D+2d_1}}
& K(\mathbf{s},s'_1)
\Big( \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \\ \nonumber
& F_{\widetilde{\mathbf{k}}}(x_1 + l_1 s_1 + l_2 s'_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1 + l_3 s_1 + l_4 s'_1) \Big) \\ \label{est:thm10b}
& \,\textup{d}\mathbf{s} \,\textup{d}s'_1 \,\textup{d}\mathbf{x} \,\textup{d}y_1\bigg|
\lesssim_{(C_{\kappa})_\kappa} \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \|F_{\widetilde{\mathbf{k}}}\|_{\textup{L}^{2^{n+1}}(\mathbb{R}^{(n+1)d})}
\end{align}
for functions $F_{\widetilde{\mathbf{k}}}:\mathbb{R}^{D+d_1}\rightarrow \mathbb{C}$, where $\mathbf{s}=(s_1,\ldots, s_n)\in \mathbb{R}^{D}$. Indeed,
part (b) of Theorem \ref{thm:singular} then follows by Fubini, applying the estimate \eqref{est:thm10b} and H\"older's inequality in $y_2,\ldots , y_n$.
To show \eqref{est:thm10b} we again first consider the case of equal dimensions, $d_1= \cdots =d_n$. Once again, we write them simply as $d$.
Changing variables $x_1+y_1+s_1 = x_{n+1}^0$, $x_1+y_1+s_1'=x_{n+1}^1$, we obtain
{\allowdisplaybreaks\begin{align*}
\int_{\mathbb{R}^{2d(n+1)}}
& K(x_{n+1}^0-x_1-y_1,s_2,\ldots,s_n,x_{n+1}^1-x_1-y_1) \Big(\,\prod_{k=(k_2,\ldots,k_n)\in \{0,1\}^{n-1}} \\
& F_{(k,e_1)}(x_{n+1}^0-y_1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,y_1)\\
& F_{(k,e_2)}(x_{n+1}^1-y_1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,y_1)\\
& F_{(k,e_3)}(x_1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,x_{n+1}^0-x_1)\\
& F_{(k,e_4)}(x_1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,x_{n+1}^1-x_1) \Big) \,\textup{d}x_{n+1}^0 \,\textup{d}x_{n+1}^1\,\textup{d}s_2 \cdots \,\textup{d}s_n \,\textup{d}\mathbf{x} \,\textup{d}y_1,
\end{align*}}
where $e_i$ are standard unit vectors in $\mathbb{R}^4$.
Shearing the functions $F_{(k,e_i)}$ we see that it suffices to show estimate an estimate for the form
{\allowdisplaybreaks\begin{align*}
\int_{\mathbb{R}^{2d(n+1)}}
&K(x_{n+1}^0-x_1-y_1,s_2,\ldots,s_n,x_{n+1}^1-x_1-y_1) \Big( \, \prod_{k=(k_2,\ldots,k_n)\in \{0,1\}^{n-1}} \\
& F_{(k,e_1)}(x_{n+1}^0, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,y_1)\\
& F_{(k,e_2)}(x_{n+1}^1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1)\\
& F_{(k,e_3)}(x_1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,x_{n+1}^0)\\
& F_{(k,e_4)}(x_1, x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,x_{n+1}^1) \Big) \,\textup{d}x_{n+1}^0 \,\textup{d}x_{n+1}^1 \,\textup{d}s_2 \cdots \,\textup{d}s_n \,\textup{d}\mathbf{x} \,\textup{d}y_1.
\end{align*}}
We relabel $y_1$ into $x_1^1$, $x_i$ into $x_i^0$ for $1\leq i \leq n$, and change variables $x_i^0 + s_i = x_i^1$ for $2\leq i \leq n$.
Then we see that it suffices to show the estimate
\begin{align*}
\Big| \int_{\mathbb{R}^{2d(n+1)}} \prod_{\mathbf{k}\in \{0,1\}^{n+1}} F_{\mathbf{k}} (\Pi_\mathbf{k} x) K(\Pi x) \,\textup{d}x \Big |
\lesssim_{(C_{\kappa})_\kappa} \prod_{\mathbf{k}\in\{0,1\}^n} \|F_{\mathbf{k}}\|_{\textup{L}^{2^{n+1}}(\mathbb{R}^{(n+1)d})}
\end{align*}
for
$x=(x_1^0,\ldots, x_{n+1}^0,x_1^1,\ldots x_{n+1}^1)\in (\mathbb{R}^d)^{n+1}$ and linear operators $\Pi_\mathbf{k},\Pi : (\mathbb{R}^{d})^{2(n+1)} \rightarrow (\mathbb{R}^{d})^{n+1}$ given by
\begin{align*}
\Pi_\mathbf{k}x := (x_1^{k_1},\ldots, x_{n+1}^{k_{n+1}}),\quad \Pi x := (x_{n+1}^0-x_1^1-x_1^0, x_2^1-x_2^0,\ldots, x_{n}^1-x_{n}^0, x_{n+1}^1-x_1^1-x_1^0).
\end{align*}
This estimate again follows from the main result in \cite{DT18}.
To finish the proof of \eqref{est:thm10b} it remains to deduce the case of different dimensions from the case of equal dimensions. This follows similarly as in part (a) and we only sketch the necessary modifications.
Let $d$ be defined as in the proof of part (a) of this theorem.
Assuming the estimate \eqref{est:thm10b} in the case of equal dimensions,
let us plug in the kernel $\widetilde{K}$ defined on $(\mathbb{R}^d)^{n+1}$ as in \eqref{eq:defKtilde} and the functions $F_{\mathbf{\widetilde{k}}}\colon (\mathbb{R}^d)^{n+1}\rightarrow \mathbb{C}$ given by
\begin{align*}
\widetilde{F}_{\mathbf{\widetilde{k}}}(\mathbf{z},\mathbf{\widetilde{z}}) := F_{\mathbf{\widetilde{k}}}(\mathbf{z}) (\textup{D}_{\lambda}^{2^{n+1}}\varphi_1\otimes \cdots \otimes \textup{D}_{\lambda}^{2^{n+1}}\varphi_{n} \otimes \textup{D}_{\lambda}^{2^{n+1}}\varphi_{1})(\mathbf{\widetilde{z}}),
\end{align*}
where $\mathbf{z}\in \mathbb{R}^{D+d_1},\,\mathbf{\widetilde{z}}\in \mathbb{R}^{(n+1)d-D-d_1}$, $F_{\mathbf{\widetilde{k}}}:\mathbb{R}^{D+d_1}\rightarrow \mathbb{C}$, $\varphi_i$ is a Schwartz function on $\mathbb{R}^{d-d_i}$ for $1\leq i \leq n$, and $\textup{D}_{\lambda}^{2^{n+1}} \varphi_i$ was defined in \eqref{eq:funcdilatep}. Then the form in \eqref{est:thm10b} becomes, analogously to the display \eqref{eq:lhs2} in part (a),
{\allowdisplaybreaks\begin{align}
\int_{\mathbb{R}^{2d(n+1)}} \nonumber
& \widetilde{K}(\mathbf{s}, s'_1)\,
\lambda^{-(d-d_1)}\varphi_1(\lambda^{-1}\widetilde{x}_1)^{2^{n}}\varphi_1(\lambda^{-1}(\widetilde{x}_1 + \widetilde{s}_1))^{2^{n-1}} \varphi_1(\lambda^{-1}(\widetilde{x}_1 + \widetilde{s}'_1))^{2^{n-1}} \\ \nonumber
& \Big( \prod_{i=2}^n \lambda^{-(d-d_i)}\varphi_i(\lambda^{-1}\widetilde{x}_i)^{2^{n}}\varphi_i(\lambda^{-1}(\widetilde{x}_i + \widetilde{s}_i))^{2^{n}} \Big ) \\ \nonumber
&\lambda^{-(d-d_1)}\varphi_{1}(\lambda^{-1}\widetilde{y}_1)^{2^{n}}\varphi_{1}(\lambda^{-1}(\widetilde{y}_1+ \widetilde{s}_{1}))^{2^{n-1}} \varphi_{1}(\lambda^{-1}(\widetilde{y}_1 + \widetilde{s}'_{1}))^{2^{n-1}} \\ \nonumber
&\Big (\, \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \\ \nonumber
&F_{\widetilde{\mathbf{k}}}(x_1 + l_1 s_1 + l_2 s'_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1 + l_3 s_1 + l_4 s'_1)\, \Big ) \\ \label{form:partb}
& \,\textup{d}\mathbf{s}\,\textup{d}s'_1 \,\textup{d}\mathbf{\widetilde{s}} \,\textup{d}\widetilde{s}'_1 \,\textup{d}\mathbf{x} \,\textup{d}\mathbf{\widetilde{x}} \,\textup{d}y_1 \,\textup{d}\widetilde{y_1},
\end{align}}
where $\mathbf{\widetilde{s}} = (\widetilde{s}_1,\ldots, \widetilde{s}_n)\in \mathbb{R}^{nd-D}, \,\widetilde{s}'_1\in \mathbb{R}^{d-d_1},\, \widetilde{y}_1\in \mathbb{R}^{d-d_1}$.
From the case of equal dimensions we know that it is bounded by a constant times
\begin{align*}
\prod_{\mathbf{\widetilde{k}}} \|\widetilde{F}_{\mathbf{\widetilde{k}}}\|_{\textup{L}^{2^{n+1}}(\mathbb{R}^{(n+1)d})}
= \|\varphi_1\otimes \cdots \otimes \varphi_n \otimes \varphi_1 \|^{2^{n+1}}_{\textup{L}^{2^{n+1}}(\mathbb{R}^{(n+1)d-D-d_1})}
\prod_{\mathbf{\widetilde{k}}} \|F_{\mathbf{k}}\|_{\textup{L}^{2^{n+1}}(\mathbb{R}^{D+d_1})}.
\end{align*}
On the other hand, integrating in $\mathbf{\widetilde{x}}$ and $\widetilde{y}_1$ gives that the form \eqref{form:partb} equals
\begin{align*}
\int_{\mathbb{R}^{2d(n+1)}}
& \widetilde{K}(\mathbf{s}, s'_1)\,
\Phi_1(\lambda^{-1}\widetilde{s_1},\lambda^{-1}\widetilde{s}_1') ^2 \Phi (\lambda^{-1}(\widetilde{s}_2,\ldots,\widetilde{s}_n))
\, \Big(\, \prod_{\substack{\widetilde{\mathbf{k}}=(k_2,\ldots,k_n,l_1,l_2,l_3,l_4)\in\{0,1\}^{n+3}\\ l_1+l_2+l_3+l_4=1}} \\
& \!\!\! F_{\widetilde{\mathbf{k}}}(x_1 + l_1 s_1 + l_2 s'_1,\,x_2 + k_2 s_2,\,\dots,\,x_n + k_n s_n,\,y_1 + l_3 s_1 + l_4 s'_1)\, \Big) \,\textup{d}\mathbf{s}\,\textup{d}s'_1 \,\textup{d}\mathbf{\widetilde{s}} \,\textup{d}\widetilde{s}'_1 \,\textup{d}\mathbf{x} \,\textup{d}y_1,
\end{align*}
where
\[ \Phi_1(\widetilde{s}_1,\widetilde{s}'_1) := \int_{\mathbb{R}^{d-d_1}} \varphi_i(u)^{2^{n}}\varphi_i(u + \widetilde{s}_1)^{2^{n-1}} \varphi_1(u + \widetilde{s}'_1)^{2^{n-1}} \,\textup{d}u \]
and
\[ \Phi:= \big ( \varphi_2^{2^{n}}\ast \widetilde{\varphi}_2^{\, 2^{n}} \big )\otimes \cdots \otimes \big ( \varphi_n^{2^{n}}\ast \widetilde{\varphi}_n^{\, 2^{n}} \big ), \]
where $\widetilde{\varphi}_i$ is defined as in (a).
Taking the limit as $\lambda \rightarrow \infty$ and integrating in $\mathbf{\widetilde{s}}, \widetilde{s}_1'$, similarly as in the proof of part (a), we recover the form on the left hand-side of \eqref{est:thm10b}.
\end{proof}
\section*{Acknowledgments}
The authors thank Christoph Thiele for inspiring discussions aided by the bilateral DAAD-MZO grant \emph{Multilinear singular integrals and applications}.
The authors are also grateful to the anonymous referee for useful suggestions.
V. K. was supported in part by the Croatian Science Foundation under the project UIP-2017-05-4129 (MUNHANAP).
|
\section{Introduction}
A \textit{virtual reality} (VR) video is generated by capturing a scene of interest \textcolor{black}{from} \textcolor{black}{all directions} at the same time using omnidirectional cameras. A user wearing a VR headset can freely watch the scene of interest in any viewing direction at any time, hence enjoying \textcolor{black}{immersive} viewing experience. VR has vast applications in entertainment, education, medicine, etc. It is predicted that the global market of VR related products will reach 30 billion \textcolor{black}{US dollars} by 2020 \textcolor{black}{\cite{111}}.
Transmitting an entire 360 VR video which is of a much larger size than a traditional video brings a heavy burden to wireless networks.
To improve transmission efficiency and avoid view switch delay, a 360 VR video is divided into smaller rectangular segments of the same size, referred to as tiles\textcolor{black}{. The set of tiles} covering a user's current \textit{field-of-view} (FoV) and the FoVs that may be watched shortly should be transmitted simultaneously.
In \cite{unicast-liu,unicast-one}, the authors consider tiled 360 VR video transmission in single-user wireless networks, and optimize video encoding parameters to maximize the utility which reflects the quality of the received 360 VR video.
The optimization problems are discrete, and are solved by exhaustive search.
In \cite{multicast-nojoint,multicast-olcoding}, the authors consider 360 VR video transmission in multi-user wireless networks, and exploit multicast opportunities for serving concurrent viewers to improve transmission efficiency.
In particular,
\cite{multicast-nojoint} adopts the tiling technique, and optimizes the modulation and coding level of each tile to maximize the video quality. It is a discrete optimization problem of a large size, and the proposed heuristic algorithm may not provide desirable performance and complexity.
\textcolor{black}{Different from \cite{unicast-liu,unicast-one,multicast-nojoint}, \cite{multicast-olcoding} considers online encoding} to encode only the required area to be transmitted. \textcolor{black}{In addition, \cite{multicast-olcoding}} proposes a dynamic multicast mechanism which is adaptive to the channel conditions of all users. The online encoding in \cite{multicast-olcoding} has higher coding efficiency (at the cost of the increase of implementation complexity). \textcolor{black}{However}, the transmission scheme in \cite{multicast-olcoding} is not optimization based, and may not guarantee promising performance.
Note that the solutions in \cite{multicast-nojoint,multicast-olcoding} do not provide many design insights for 360 VR video multicast in wireless networks. It is still not known how the \textcolor{black}{required} FoVs and channel conditions of all users affect optimal resource allocation.
In this letter, we would like to address the above \textcolor{black}{issues}. We \textcolor{black}{consider optimal multicast of tiled 360 VR video from one \textcolor{black}{server (base station or access point)}} to multiple users. Specifically, we divide the 360 VR video into tiles. For each user, we deliver a set of tiles that cover the user's current FoV and the FoVs that may be watched shortly. In order to make use of
multicast opportunities and avoid redundant transmissions, we partition the set of tiles to be transmitted to users into disjoint subsets and multicast different subsets of tiles to different groups of users.
\textcolor{black}{We consider random viewing directions and \textcolor{black}{random} channel conditions, and \textcolor{black}{adopt} time division multiple access (TDMA).}
\textcolor{black}{For given video quality, we optimize the transmission time and power allocation to minimize the average transmission energy under the transmission time allocation constraints and the transmission rate constraint. For given transmission energy budget, we optimize the transmission time and power allocation as well as the encoding rate of each tile to maximize the received video quality under the maximum transmission energy constraint.}
These two optimization problems are challenging non-convex problems.
For each non-convex optimization problem, by analyzing optimality properties, we \textcolor{black}{successfully} transform it into an equivalent convex problem and obtain a globally optimal closed-form solution using KKT conditions.
\textcolor{black}{The derived solutions reveal important design insights for multicast of tiled 360 VR video.}
To the best of our knowledge, these important design insights have never been explicitly explored and analytically verified in existing literature.
Finally, numerical results demonstrate the advantage of the proposed optimal solutions.
\section{System Model}
\begin{figure}[t]
\vspace*{-0.1cm}
\begin{center}
\includegraphics[width=2.8in]{Cui_WCL2018-0604_fig1.eps
\end{center}
\vspace*{-0.35cm}
\caption{\small{\textcolor{black}{Illustration} of tiled 360 VR video multicast with $K=2$, $N_h\times\ N_v=8\times4$, $V_h\times V_v=8\times4$. A circle represents a viewing direction.}}
\label{fig:system-model}
\vspace*{-0.55cm}
\end{figure}
As illustrated in Fig.~\ref{fig:system-model}, we consider downlink transmission of a 360 VR video from a single-antenna \textcolor{black}{server (base station or access point)} to $K~(\geq1)$ single-antenna users. Let $\mathcal{K}\triangleq\{1,\ldots,K\}$ denote the set of user indices.
For ease of implementation, \textcolor{black}{as illustrated in Fig.~\ref{fig:system-model},} we consider $N_h\times N_v$ viewing directions, where $N_h$ and $N_v$ represent the numbers of horizontal and vertical viewing directions,
and can be arbitrarily large.
The $(n_h,n_v)$-th viewing direction refers to the viewing direction in the $n_h$-th row and \textcolor{black}{$n_v$-th} column. When a VR user is interested in one viewing direction, he can \textcolor{black}{view} a rectangular FoV of size $F_h\times F_v$ with the viewing direction as its center.
To improve transmission efficiency, we consider tiling. In particular, the 360 VR video is divided into $V_h\times V_v$ rectangular segments of the same size, referred to as tiles, where $V_h$ and $V_v$ represent the numbers of segments in each row and each column, respectively.
Each tile is encoded into $L$ versions with $L$ different quality levels and encoding rates.
\textcolor{black}{The $l$-th version of each tile has the $l$-th lowest quality with encoding rate $D_l$. Note that $D_1<D_2<\cdots<D_L$. For fairness, we assume the tiles transmitted to all users are of the same quality level and let $D$ denote the corresponding encoding rate.
A VR user may watch one FoV at sometime, and freely \textcolor{black}{switch} to another FoV after a while. To avoid view switch delay, for each user, the set of tiles that cover its current FoV and the FoVs that may be watched shortly will be delivered.}
$K$ users \textcolor{black}{randomly} select their viewing directions. Let
$\mathbf{X}_k\in\mathcal X$ denote the random viewing direction of user $k$, where \textcolor{black}{$\mathcal{X}\triangleq\{(n_h,n_v)|n_h=1,\ldots,N_h,n_v=1,\ldots,N_v\}$} represents the set of all possible viewing directions of each user.
Let $\mathbf X\triangleq (\mathbf{X}_k)_{k\in\mathcal K}\in \mathcal X^K$ denote the random system viewing direction state.
For given $\mathbf{X}$, let $\Phi_k(\mathbf{X})$ denote the set of tiles that need to be transmitted to user $k$, and let
$\Phi(\mathbf{X})\triangleq\cup_{k\in\mathcal{K}}\Phi_k(\mathbf{X})$ denote the set of tiles that need to be transmitted to all users.
In order to make use of multicasting opportunities and avoid redundant transmissions, we divide $\Phi(\mathbf{X})$ into $I(\mathbf{X})$ disjoint non-empty sets $\mathcal{S}_i(\mathbf{X})$, \textcolor{black}{$i\in\mathcal{I}(\mathbf{X})\triangleq\{1,\ldots,I(\mathbf{X})\}$.}
For all $i,j\in I(\mathbf{X}),$ $i\neq j$, $\mathcal{S}_i(\mathbf{X})$ and $\mathcal{S}_j(\mathbf{X})$ are for different groups of users.
Let $S_i(\mathbf{X})$ denote the number of tiles in set $\mathcal{S}_i(\mathbf{X})$.
Let $\mathcal K_{i}(\mathbf X)$ and $K_{i}(\mathbf{X})$ denote the set and the number of users that need to receive the tiles in $\mathcal{S}_i(\mathbf{X})$.
Without
loss of generality, we refer to the transmission of the tiles in $\mathcal{S}_i(\mathbf{X})$ to the users in $\mathcal{K}_{i}(\mathbf{X})$ as multicast, although both unicast ($K_i(\mathbf{X})=1$) and multicast ($K_i(\mathbf{X})>1$) may happen.
\textcolor{black}{We consider a discrete time narrow band TDMA\footnote{\textcolor{black}{\textcolor{black}{Note that TDMA is analytically tractable and has applications in WiFi systems. In addition, the} multicast transmission scheme and the optimization framework \textcolor{black}{for a TDMA system} can be extended to an OFDMA system.}} system with bandwidth $B$ (in Hz), and study one time frame of duration $T$ (in seconds). Assume block fading, i.e., each channel state (over bandwidth $B$) does not change within the considered time frame.} \textcolor{black}{In addition, we assume that the channel states of $K$ users are random.} Let $H_k\in \mathcal H$ denote the random channel state of user $k$, representing the power of the channel between user $k$ and \textcolor{black}{the server}, where $\mathcal H$ denotes the finite channel state space.
Let $\mathbf H\triangleq (H_k)_{k\in\mathcal K}\in \mathcal H^{K}$ denote the random system channel state.
The random system state consists of the random system viewing direction state $\mathbf X$ and the random system channel state $\mathbf H$, denoted by $(\mathbf X, \mathbf H)\in \mathcal X^K\times \mathcal H^K$.\footnote{\textcolor{black}{Note that the analysis and optimization in this letter do not require the independence between $\mathbf{X}$ and $\mathbf{H}$.}} \textcolor{black}{The probability that the random system state $(\mathbf X,\mathbf H)$ takes the value $(\mathbf x,\mathbf h)\in \mathcal X^K\times \mathcal H^K$ is denoted as
$\Pr[(\mathbf X,\mathbf H)=(\mathbf x, \mathbf h) ]$,}
\textcolor{black}{where $\mathbf x\triangleq (\mathbf{x}_k)_{k\in\mathcal K}$, $\mathbf{x}_k\in \mathcal X$, $\mathbf h \triangleq (h_k)_{k\in\mathcal K}$, and $h_k\in\mathcal{H}$.}
We assume that \textcolor{black}{the server} is aware of the system state $(\mathbf X,\mathbf H)$.
The time allocated to transmit the tiles in $\mathcal{S}_i(\mathbf{X})$ is denoted by $t_{i}(\mathbf{X},\mathbf{H})$.
Thus, we have the following transmission time allocation constraints:
\begin{align}
&t_i(\mathbf{X},\mathbf{H})\geq0,~i\in\mathcal{I}(\mathbf{X}),\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K,\label{const:tdma-time0}\\
&\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}t_{i}(\mathbf{X},\mathbf{H})\leq T,~\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K.\label{const:tdma-time}
\end{align}
Let $p_{i}(\mathbf{X},\mathbf{H})\geq0$ denote the transmission power of the symbols for the tiles in $\mathcal{S}_i(\mathbf{X})$.
Thus,
the transmission energy is:
\begin{align}\label{eqn:tdma-total energy}
E(\mathbf{t}(\mathbf{X},\mathbf{H}),\mathbf{p}(\mathbf{X},\mathbf{H}))=\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}t_{i}(\mathbf{X},\mathbf{H})p_{i}(\mathbf{X},\mathbf{H}),\nonumber\\
~\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K,
\end{align}
where
$\mathbf{t}(\mathbf{X},\mathbf{H})\triangleq\left(t_{i}(\mathbf{X},\mathbf{H})\right)_{i\in\mathcal{I}(\mathbf{X})}$ and $\mathbf{p}(\mathbf{X},\mathbf{H})\triangleq\left(p_{i}(\mathbf{X},\mathbf{H})\right)_{i\in\mathcal{I}(\mathbf{X})}$.
To obtain design insights, we consider capacity achieving code.
To guarantee that all users in $\mathcal{K}_i(\mathbf{X})$ can successfully receive the tiles in $\mathcal{S}_i(\mathbf{X})$, we have the following transmission rate constraint:
\begin{align}
t_{i}(\mathbf{X},\mathbf{H})B \log_2 \left(1 + \frac{p_{i}(\mathbf{X},\mathbf{H}) H_{k}}{n_0}\right)\geq S_{i}(\mathbf{X})DT,\nonumber\\
~k\in\mathcal{K}_i(\mathbf{X}),i\in\mathcal{I}(\mathbf{X}),\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K,\label{const:tdma-size}
\end{align}
where $n_0$ is the power of the complex additive white Gaussian noise at each receiver. Denote $\mathbf{t}\triangleq\left(\mathbf{t}(\mathbf{X},\mathbf{H})\right)_{(\mathbf{X},\mathbf{H})\in\mathcal{X}^K\times\mathcal{H}^K}$ and
$\mathbf{p}\triangleq\left(\mathbf{p}(\mathbf{X},\mathbf{H})\right)_{(\mathbf{X},\mathbf{H})\in\mathcal{X}^K\times\mathcal{H}^K}$.
\section{\textcolor{black}{Average Transmission Energy Minimization}}\label{sec:energy}
\subsection{Problem Formulation}
Given the video quality (i.e., encoding rate of each tile $D\in\{D_1,\ldots,D_L\}$),
we would like to minimize the \textcolor{black}{average transmission energy} under the transmission time allocation constraints and the transmission rate constraint.
\begin{Prob} [Average Transmission Energy Minimization]\label{P1}
\begin{align}
\bar{E}^{\star}\triangleq\min_{\mathbf{t},\mathbf{p}}~&\mathbb{E}[E(\mathbf{t}(\mathbf{X},\mathbf{H}),\mathbf{p}(\mathbf{X},\mathbf{H}))]\nonumber\\
\mathrm{s.t.} ~~&\eqref{const:tdma-time0},~\eqref{const:tdma-time},~\eqref{const:tdma-size},\nonumber
\end{align}
where the expectation $\mathbb{E}$ is taken over the random system state $(\mathbf{X},\mathbf{H})\in\mathcal{X}^K\times\mathcal{H}^K$, \textcolor{black}{i.e.,
$\mathbb{E}[E(\mathbf{t}(\mathbf{X},\mathbf{H}),\mathbf{p}(\mathbf{X},\mathbf{H}))]=\sum\nolimits_{(\mathbf{x},\mathbf{h})\in\mathcal{X}^K\times\mathcal{H}^K}\Pr[(\mathbf X,\mathbf H)=(\mathbf x, \mathbf h)]E(\mathbf{t}(\mathbf{x},\mathbf{h}),\mathbf{p}(\mathbf{x},\mathbf{h}))$.}
\textcolor{black}{Let $\mathbf{t}^{\star}_{\text{e}}$ and $\mathbf{p}^{\star}_{\text{e}}$ denote an optimal solution.}
\end{Prob}
Note that the objective function of Problem~\ref{P1} and the constraint in $\eqref{const:tdma-size}$ are non-convex. Thus, Problem~\ref{P1} is non-convex.
\textcolor{black}{In general, it is challenging to obtain a globally optimal solution of a non-convex problem.}
\subsection{Optimal Solution}
In this part, we obtain a globally optimal solution of the non-convex Problem~\ref{P1}.
\textcolor{black}{Note that Problem~\ref{P1} can be \textcolor{black}{decomposed} into subproblems, \textcolor{black}{one for each
$(\mathbf{X},\mathbf{H})\in\mathcal{X}^K\times\mathcal{H}^K$.}
\begin{align}&\emph{Subproblem~1 (\text{Subproblem of Problem~\ref{P1}):}}\nonumber\\
&E^{\star}(\mathbf{X},\mathbf{H})\triangleq\min_{\mathbf{t}(\mathbf{X},\mathbf{H}),\mathbf{p}(\mathbf{X},\mathbf{H})}~E(\mathbf{t}(\mathbf{X},\mathbf{H}),\mathbf{p}(\mathbf{X},\mathbf{H}))\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\mathrm{s.t.} ~~~\eqref{const:tdma-time0},~\eqref{const:tdma-time},~\eqref{const:tdma-size}.\nonumber
\end{align}}
~Denote $H_{i,\text{min}}(\mathbf{X},\mathbf{H})\triangleq\min\limits_{k\in\mathcal{K}_{i}(\mathbf{X})}H_k(\mathbf{X},\mathbf{H})$.
To solve \textcolor{black}{Subproblem~1}, we first analyze its optimality properties.
\begin{Lem}[Optimality Properties of \textcolor{black}{Subproblem~1}]\label{lem:opt structure}
The optimal solution of \textcolor{black}{Subproblem~1} satisfies:
\begin{align}
p_{\text{e},i}^{\star}(\mathbf{X},\mathbf{H})=\frac{n_0}{H_{i,\text{min}}(\mathbf{X},\mathbf{H})}\left(2^{\frac{S_{i}(\mathbf{X})DT}{Bt^{\star}_{\text{e},i}(\mathbf{X},\mathbf{H})}}-1\right),~i\in\mathcal{I}(\mathbf{X}).\label{eqn:tdma-p1}
\end{align}
\end{Lem}
\textcolor{black}{\begin{proof}[Proof (sketch)]
Suppose an optimal solution of Subproblem~1 is $((t^{\star}_{\text{e},i}(\mathbf{X},\mathbf{H}))_{i\in\mathcal{I}(\mathbf{X})},(p^{\dagger}_{\text{e},i}(\mathbf{X},\mathbf{H}))_{i\in\mathcal{I}(\mathbf{X})})$, and $(p^{\dagger}_{\text{e},i}(\mathbf{X},\mathbf{H}))_{i\in\mathcal{I}(\mathbf{X})}\neq(p^{\star}_{\text{e},i}(\mathbf{X},\mathbf{H}))_{i\in\mathcal{I}(\mathbf{X})}$, \textcolor{black}{where $(p^{\star}_{\text{e},i}(\mathbf{X},\mathbf{H}))_{i\in\mathcal{I}(\mathbf{X})}$} is given by \eqref{eqn:tdma-p1}. Then, there exists $j\in\mathcal{I}(\mathbf{X})$ such that $p^{\dagger}_{\text{e},j}(\mathbf{X},\mathbf{H})<p^{\star}_{\text{e},j}(\mathbf{X},\mathbf{H})$.
\textcolor{black}{By} \eqref{eqn:tdma-p1}, we can show that $p^{\dagger}_{\text{e},j}(\mathbf{X},\mathbf{H})$ does not satisfy \eqref{const:tdma-size}, which \textcolor{black}{contradicts with the assumption}.
Therefore, \textcolor{black}{by contradiction, we can prove} Lemma~\ref{lem:opt structure}.
\end{proof}}
\textcolor{black}{Note that $\frac{S_{i}(\mathbf{X})DT}{t^{\star}_{\text{e},i}(\mathbf{X},\mathbf{H})}$ in \eqref{eqn:tdma-p1} represents the transmission rate (in bit/s) for the tiles in $\mathcal{S}_i(\mathbf{X})$ for given $(\mathbf{X},\mathbf{H})$.}
By Lemma~\ref{lem:opt structure}, we can eliminate $\mathbf{p}$ and transform \textcolor{black}{Subproblem~1} into an equivalent problem:
\begin{Prob} [Equivalent Problem of \textcolor{black}{Subproblem~1}]\label{ESP1}
\begin{align}
\min_{\textcolor{black}{\mathbf{t}(\mathbf{X},\mathbf{H})}}~&\textcolor{black}{\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}\frac{n_0t_{i}(\mathbf{X},\mathbf{H})}{H_{i,\text{min}}(\mathbf{X},\mathbf{H})}\left(2^{\frac{S_{i}(\mathbf{X})DT}{Bt_{i}(\mathbf{X},\mathbf{H})}}-1\right)}\nonumber\\
\mathrm{s.t.}~~&\eqref{const:tdma-time0},~\eqref{const:tdma-time}.\nonumber
\end{align}
\end{Prob}
We can easily verify that Problem~\ref{ESP1} is convex and the Slater's condition is satisfied, implying that strong duality holds. Thus, Problem~\ref{ESP1} can be solved using KKT conditions \textcolor{black}{as in \cite[pp.~243-246]{cvx}}. Based on the optimal solution of Problem~\ref{ESP1} and Lemma~\ref{lem:opt structure}, we have the following result.
\begin{Lem}[Optimal Solution of Problem~\ref{P1}]\label{lem:opt solution}
\begin{align}
t_{\text{e},i}^{\star}(\mathbf{X},\mathbf{H})=&\frac{S_{i}(\mathbf{X})DT\ln{2}}{B\left(W\left(\frac{\lambda^{\star}(\mathbf{X},\mathbf{H})H_{i,\text{min}}(\mathbf{X},\mathbf{H})}{n_0e}-\frac{1}{e}\right)+1\right)},\nonumber\\
&~i\in\mathcal{I}(\mathbf{X}),\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K,\label{eqn:tdma-opt solution1}\\
p_{\text{e},i}^{\star}(\mathbf{X},\mathbf{H})=&\frac{n_0\left(e^{W\left(\frac{\lambda^{\star}(\mathbf{X},\mathbf{H})H_{i,\text{min}}(\mathbf{X},\mathbf{H})}{n_0e}-\frac{1}{e}\right)+1}-1\right)}{H_{i,\min}(\mathbf{X},\mathbf{H})},\nonumber\\
&~i\in\mathcal{I}(\mathbf{X}),\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K,\label{eqn:tdma-opt solution11}
\end{align}
where $W(\cdot)$ denotes the Lambert function, $e$ is a mathematical constant that is the base of the natural logarithm, and the Lagrangian multiplier $\lambda^{\star}(\mathbf{X},\mathbf{H})$ satisfies
\begin{align}\label{eqn:tdma-lambda1}
\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}\frac{S_{i}(\mathbf{X})D\ln{2}}{B\left(W\left(\frac{\lambda^{\star}(\mathbf{X},\mathbf{H})H_{i,\text{min}}(\mathbf{X},\mathbf{H})}{n_0e}-\frac{1}{e}\right)+1\right)}=1,\nonumber\\
~\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K.
\end{align}
\end{Lem}
Note that $\lambda^{\star}(\mathbf{X},\mathbf{H})$ in \eqref{eqn:tdma-lambda1} can be easily obtained using the bisection method. Thus, we can obtain a globally optimal solution of Problem~\ref{P1} efficiently.
\textcolor{black}{
By Lemma~\ref{lem:opt structure} and Lemma~\ref{lem:opt solution}, we have the following result.}
\begin{Lem}[Minimum Transmission Energy]\label{lem:bound}
\textcolor{black}{If $H_{1}=\cdots=H_{K}\triangleq \widetilde{H}$, then
\begin{align}
E^{\star}(\mathbf{X},\mathbf{H})=\frac{n_0T}{\widetilde{H}}\left(2^{\frac{D\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})}{B}}-1\right).\label{pro:energy}
\end{align}
Furthermore,} for all $(\mathbf{X},\mathbf{H})\in\mathcal{X}^K\times\mathcal{H}^K$,
\begin{align}\label{eqn:bound}
&\frac{n_0T}{\max \mathcal{H}}\left(2^{\frac{D\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})}{B}}-1\right)\leq E^{\star}(\mathbf{X},\mathbf{H})\nonumber\\
&\leq\frac{n_0T}{\min\mathcal{H}}\left(2^{\frac{D\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})}{B}}-1\right).
\end{align}
\end{Lem}
\textcolor{black}{\begin{proof}[Proof (sketch)]
When $H_{1}=\cdots=H_{K}=\widetilde{H}$,
by Lemma~\ref{lem:opt solution} and \eqref{eqn:tdma-total energy}, we \textcolor{black}{can prove \eqref{pro:energy}.}
For given $\mathbf{X}$, we can show that \textcolor{black}{any} optimal solution of Subproblem~1 with arbitrary $\mathbf{H}$ \textcolor{black}{is a feasible solution} of Subproblem~1 with $H_{1}=\cdots=H_{K}=\max\mathcal{H}$, \textcolor{black}{and any optimal solution of Subproblem~1 with $H_{1}=\cdots=H_{K}=\min\mathcal{H}$ is a feasible solution of Subproblem~1 with arbitrary $\mathbf{H}$.} Thus, by \eqref{pro:energy}, we \textcolor{black}{can prove \eqref{eqn:bound}.}
\end{proof}}
\textcolor{black}{Lemma~\ref{lem:bound} indicates that approximately, the minimum transmission energy $E^{\star}(\mathbf{X},\mathbf{H})$ increases \textcolor{black}{exponentially} with \textcolor{black}{the total number of tiles that need to be transmitted, i.e., $\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})$, and is inversely proportional to the channel \textcolor{black}{powers}, i.e., $H_k,k\in\mathcal{K}$.}} Note that $\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})$ reflects the concentration of the viewing directions of all users.
A smaller value of $\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})$ means closer viewing directions of all users.
\section{Received Video Quality Maximization}\label{sec:quality}
\subsection{Problem Formulation}
Let $E_{\text{limit}}$ denote the transmission energy budget of the system. Consider the maximum transmission energy constraint:
\begin{align}\label{const:tdma-energy constraint}
\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}t_{i}(\mathbf{X},\mathbf{H})p_{i}(\mathbf{X},\mathbf{H})\leq E_{\text{limit}},~\mathbf{X}\in\mathcal{X}^K,\mathbf{H}\in\mathcal{H}^K.
\end{align}
To guarantee user experience, the encoding rate should not change as frequently as the viewing directions and channel states, and should remain constant within a certain time \textcolor{black}{duration}.
Given the \textcolor{black}{transmission energy budget} $E_{\text{limit}}$, we would like to maximize the received video quality (i.e., encoding rate of each tile $D$) under the maximum transmission energy constraint.
\begin{Prob} [Received Video Quality Maximization]\label{P2}
\begin{align}
D_{\text{q}}^{\star}&\triangleq\max_{D,\mathbf{t},\mathbf{p}}D\nonumber\\
\mathrm{s.t.} ~~\quad
&\eqref{const:tdma-time0},~\eqref{const:tdma-time},~\eqref{const:tdma-size},~\eqref{const:tdma-energy constraint}.\nonumber
\end{align}
Let $D_{\text{q}}^{\star}$, $\mathbf{t}_{\text{q}}^{\star}$ and $\mathbf{p}_{\text{q}}^{\star}$ denote an optimal solution.
\end{Prob}
Note that the actual achievable video quality is $\max\limits_{l\in\{1,\ldots,L\}}\{D_l|D_l\leq D^{\star}_{\text{q}}\}$.
The constraints in $\eqref{const:tdma-size}$ and \eqref{const:tdma-energy constraint} are non-convex. Thus, Problem~\ref{P2} is non-convex.
\subsection{Optimal Solution}
In this part, we obtain a globally optimal solution of the non-convex Problem~\ref{P2}. First, we transform it into an equivalent convex problem.
Let $\mathbf{H}_{\text{min}}\triangleq(H_k)_{k\in\mathcal{K}}$ with $H_k=\min\mathcal{H}$.
\begin{Prob} [Equivalent Problem of Problem~\ref{P2}]\label{EP2}
\begin{align}
&\min_{\mathbf{X}\in\mathcal{X}^K}D_{\text{q}}^{\star}(\mathbf{X},\mathbf{H}_{\min})\nonumber
\end{align}
where $D_{\text{q}}^{\star}(\mathbf{X},\mathbf{H}_{\min})$ is given by the following subproblem.
\end{Prob}
\begin{Prob} [Subproblem of Problem~\ref{EP2}]\label{SEP2}
For all $\mathbf X\in \mathcal X^K$,
\begin{align}
&D^{\star}_{\text{q}}(\mathbf{X},\mathbf{H}_{\min})\triangleq\max_{D,\mathbf{t}(\mathbf{X},\mathbf{H}_{\min})}D\nonumber\\
&\mathrm{s.t.} \quad \eqref{const:tdma-time0},~\eqref{const:tdma-time},\nonumber\\
& \sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}\frac{n_0t_{i}(\mathbf{X},\mathbf{H}_{\min})}{\min\mathcal{H}}\left(2^{\frac{S_{i}(\mathbf{X})DT}{Bt_{i}(\mathbf{X},\mathbf{H}_{\min})}}-1\right)\leq E_{\text{limit}}.\label{const:tdma-energy const}
\end{align}
\end{Prob}
By carefully exploring structural properties of Problem~\ref{P2}, we have the following result.
\begin{Lem}\label{lem:equivalent}
\textcolor{black}{The optimal values of Problem~\ref{P2} and Problem~\ref{EP2} are equivalent, i.e., $D^{\star}_{\text{q}}=\min\limits_{\mathbf{X}\in\mathcal{X}^K}D^{\star}_{\text{q}}(\mathbf{X},\mathbf{H}_{\min})$.}
Furthermore,
\begin{align}
p_{\text{q},i}^{\star}(\mathbf{X},\mathbf{H}_{\min})=\frac{n_0}{\min\mathcal{H}}\left(2^{\frac{S_{i}(\mathbf{X})D_{\text{q}}^{\star}(\mathbf{X},\mathbf{H}_{\min})T}{Bt^{\star}_{\text{q},i}(\mathbf{X},\mathbf{H}_{\min})}}-1\right), \nonumber\\
i\in\mathcal{I}(\mathbf{X}),\mathbf{X}\in\mathcal{X}^K.\label{eqn:tdma-p2}
\end{align}
\end{Lem}
\textcolor{black}{Note that Lemma~\ref{lem:equivalent} can be proved in a similar way to Lemma~\ref{lem:opt structure}.}
We can easily verify that Problem~\ref{SEP2} is convex and the Slater's condition is satisfied, implying that strong duality holds. Thus, Problem~\ref{SEP2} can be solved using KKT conditions \textcolor{black}{as in \cite[pp.~243-246]{cvx}}. Based on the optimal \textcolor{black}{solution} of Problem~\ref{SEP2} and Lemma~\ref{lem:equivalent}, we have the following result.
\begin{Lem}[Optimal Solution of Problem~\ref{P2}]\label{lem:quality opt structure}
\begin{align}
&D_{\text{q}}^{\star}=\frac{B\ln(\frac{E_{\text{limit}}\min\mathcal{H}}{n_0T}+1)}{\ln{2}\max\limits_{\mathbf{X}\in\mathcal{X}^K}\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_{i}(\mathbf{X})},\label{eqn:tdma-opt D}
\end{align}
and $t_{\text{q},i}^{\star}(\mathbf{X},\mathbf{H})$ and $p_{\text{q},i}^{\star}(\mathbf{X},\mathbf{H})$ are given by $t_{\text{e},i}^{\star}(\mathbf{X},\mathbf{H})$ in \eqref{eqn:tdma-opt solution1} and $p_{\text{e},i}^{\star}(\mathbf{X},\mathbf{H})$ in \eqref{eqn:tdma-opt solution11}, respectively, with $D=D_{\text{q}}^{\star}$.
\end{Lem}
\textcolor{black}{Lemma~\ref{lem:quality opt structure} indicates that $D_{\text{q}}^{\star}$ is affected by \textcolor{black}{the smallest channel power $\min\mathcal{H}$ among all channel powers} instead of $H_i,i\in\mathcal{I}(\mathbf{X})$, and is inversely proportional to $\max\limits_{\mathbf{X}\in\mathcal{X}^K}\sum\nolimits_{i\in\mathcal{I}(\mathbf{X})}S_i(\mathbf{X})$ which represents the maximum number of tiles that need to be transmitted for all viewing directions.}
\section{Simulation}\label{sec:simulation}
\textcolor{black}{In this section, we compare the proposed solutions in Section~\ref{sec:energy} and Section~\ref{sec:quality} with two baselines using numerical results.
\textcolor{black}{Baseline 1 considers serving each user separately using unicast in an optimal way, similar to the proposed optimal solutions. Baseline~2 considers multicast as in this letter but with equal time allocation for each transmitted tile and optimal power allocation based on the equal time allocation.}
In the simulation, we use Kvazaar as the 360 VR video encoder and video sequence $Boxing$ as the video source \cite{boxing}. We set $F_h=F_v=100^{\circ}$, \textcolor{black}{$N_h\times N_v=36\times2$}, and \textcolor{black}{$V_h\times V_v=30\times15$}. To avoid view switch delay, we transmit extra $15^{\circ}$ in the four directions of each requested FoV.
Set $B =10$ MHz, $n_0 = 10^{-9}$ W, $T=0.1$ s,
$\mathcal{H}=\{0.5d,1.5d\}$, \textcolor{black}{$\Pr[H_k=0.5d]=0.5$, and $\Pr[H_k=1.5d]=0.5$} for all $k\in\mathcal{K}$, where $d$ reflects the path loss.
For ease of exposition, we consider Zipf distribution for the $N_h\times N_v$ viewing directions. In particular, suppose viewing direction $(n_h,n_v)$ is of rank\footnote{\textcolor{black}{Note that $(n_h,n_v)$ \textcolor{black}{can be an
arbitrary rank}, and the ranks of the viewing directions do not influence the trends of the curves in Fig.~\ref{fig:energy} (a), but the Zipf exponent $\gamma$ does. In addition, the ranks of the viewing directions do not influence the curves in Fig.~\ref{fig:energy}~(b), as the solution of Problem~\ref{P2} depends only on $\mathcal{X}^K\times\mathcal{H}^K$.}} $(n_h-1)N_v+n_v$ and $\Pr[\mathbf{X}_k=(n_h,n_v)]=\frac{((n_h-1)N_v+n_v)^{-\gamma}}{\sum\nolimits_{i=1,\ldots,N_hN_v}i^{-\gamma}}$, where $\gamma$ is the Zipf exponent. Note that a smaller $\gamma$ indicates a longer tail.}
\textcolor{black}{Fig.~\ref{fig:energy} (a) illustrates the average transmission energy versus the Zipf exponent $\gamma$.
\textcolor{black}{\textcolor{black}{We can see that the} average transmission energy of the two multicast schemes decreases with $\gamma$,} \textcolor{black}{as multicast opportunities increases with $\gamma$.}
Fig.~\ref{fig:energy} (b) illustrates the received video quality (i.e., encoding rate of each tile) versus the path loss $d$. \textcolor{black}{\textcolor{black}{We can observe that the} received video quality of each scheme increases with the channel powers.}
From Fig.~\ref{fig:energy}, we can \textcolor{black}{also} observe that the proposed optimal solutions outperform the two baselines. Specifically, the gains of the proposed optimal solutions over Baseline~1 arise from the fact that the proposed solutions utilize multicast.
The gains of the proposed optimal solutions over Baseline~2 are attributed by the fact that the proposed solutions carefully allocate both transmission time and transmission power.}
\begin{figure}[t]
\vspace*{-0.65cm}
\begin{center}
\subfigure[\small{Average transmission energy versus $\gamma$. $K=3$, $D=30561$ bit/s, $d=1\times10^{-6}$.}]
{\resizebox{4.11cm}{!}{\includegraphics{Cui_WCL2018-0604_fig2.eps}}
\subfigure[\small{Encoding rate of each tile versus $d$. $K=4$, $\gamma=0.8$, $E_{\text{limit}}=0.1$ J.}]
{\resizebox{4.17cm}{!}{\includegraphics{Cui_WCL2018-0604_fig3.eps}}
\vspace*{-0.42cm}
\end{center}
\caption{\small{Performance comparison.}}
\label{fig:energy}
\vspace*{-0.55cm}
\end{figure}
\section{Conclusion}
In this letter, we \textcolor{black}{studied} optimal multicast of tiled 360 VR video
using TDMA.
We \textcolor{black}{considered} random
viewing directions and random channel conditions\textcolor{black}{. We} \textcolor{black}{formulated} two non-convex
optimization problems, i.e., the minimization of the average
transmission energy for given video
quality, and the maximization of the received video quality
for given
\textcolor{black}{transmission energy budget}. We \textcolor{black}{obtained} globally optimal closed-form solutions of the two challenging non-convex problems and \textcolor{black}{revealed} important design insights for tiled 360 VR multicast. Finally, numerical results
\textcolor{black}{demonstrated} the advantage of the proposed optimal solutions.
\bibliographystyle{IEEEtran}
|
\section{The CMS Collaboration \label{app:collab}}\begin{sloppypar}\hyphenpenalty=5000\widowpenalty=500\clubpenalty=5000\input{HIN-18-006-authorlist.tex}\end{sloppypar}
\end{document}
|
\section{Introduction}\label{sec:intro}
We consider the distributed optimization problem in which a network of $n$ agents are connected through an undirected graph. The global objective is to minimize the average of $n$ functions $f_i:\R^d\rightarrow \R$ that are local to each agent $i=1,\dots, n$. While each agent has access to its local variable $x_i$, it must exchange information with its neighbors to arrive at the global minimizer
\begin{equation}\label{eq:distrop-problem}
x^\star \defeq \argmin_{x\in \R^d} f(x),
\quad\text{where }f(x) \defeq \frac{1}{n}\sum_{i=1}^n f_i(x).
\end{equation}
Distributed optimization is an active area of research and has attracted significant attention in recent years due to its applications in distributed machine learning, distributed estimation, and resource allocation~\cite{forero2010consensus,johansson2008distributed,predd2009collaborative,ling2010decentralized}. To solve~\eqref{eq:distrop-problem}, many gradient-based algorithms have been proposed involving gradient computation steps in conjunction with gossip steps, i.e., local averaging operations~\cite{DGD, EXTRA, NIDS, DIGing, ExactDiffusion, AugDGM}. Such algorithms have much structural variety in how they perform gossip, evaluate the gradient, and update their local state variables. Furthermore, changing the order of these communication and computation steps results in different algorithms. As an example, consider the NIDS~\cite{NIDS} algorithm update:
\begin{align*}
x_i^0 &\in\R^d\text{ arbitrary,}\quad x_i^1 = x_i^0 - \alpha \df_i(x_i^0) \tag{NIDS}\\
x_i^{k+2} &= \sum_{j=1}^n \widetilde{w}_{ij} \bigl(2 x_j^{k+1} - x_j^k - \alpha\df_j(x_j^{k+1}) + \alpha\df_j(x_j^k)\bigr)
\end{align*}
where $\{\widetilde{w}_{ij}\}\in\R^{n\times n}$ is a gossip matrix, $\alpha\in\R$ the stepsize, and $x_i^k\in\R^d$ the state of agent~$i$ at iteration~$k$. Each update involves two previous iterates as well as a difference of gradients, similar to so-called ``gradient tracking" algorithms~\cite{DIGing}. In contrast, Exact Diffusion~\cite{ExactDiffusion} uses two state variables, which are updated with a gradient step followed by a gossip step:
\begin{equation}\label{eq:ExactDiffusion}\tag{Exact Diffusion}
\begin{aligned}
x_{1i}^{k+1} &= x_{2i}^{k}-\alpha\,\df_i (x_{2i}^k) \\
x_{2i}^{k+1} &= \sum_{j=1}^n w_{ij}\,(x_{1j}^{k+1} - x_{1j}^k + x_{2j}^k).
\end{aligned}
\end{equation}
While their update equations appear different, we show in Section~\ref{sec:algorithms} that NIDS and Exact Diffusion are actually equivalent, which exemplifies the need to better understand the taxonomy of distributed algorithms. To this end, we develop a canonical form that parameterizes the following broad class of first-order distributed algorithms.
\paragraph{Algorithm properties.}\label{alg_class} \textit{We consider the class of distributed algorithms that satisfy the following properties.}
\begin{enumerate}[itemsep=0mm,topsep=1mm,label=P\arabic*.,ref=\textup{P\arabic*}]
\item \textit{Each agent has a local state of dimension at most~$2d$. Since $\df_i:\R^d\to\R$ by assumption, this allows each agent to store up to two past iterates.}\label{P1}
\item \textit{At each iteration, each agent may do each of the following once in any order:}\label{P2}
\begin{enumerate}[itemsep=0mm,topsep=0mm,label=(\alph*)]
\item \textit{communicate any number of local variables with its immediate neighbors simultaneously,}
\item \textit{compute its local gradient at a single point, and}
\item \textit{update its local state.}
\end{enumerate}
\item \textit{The local state updates are linear, time-invariant, deterministic, and homogeneous across all agents and dimensions of the objective function.}\label{P3}
\end{enumerate}
While NIDS and Exact Diffusion belong to the class of algorithms described by \ref{P1}--\ref{P3}, not all algorithms do. Exceptions include accelerated~\cite{QuLi_accelerated,DHB}, proximal~\cite{PG-EXTRA}, stochastic~\cite{DGST}, and asynchronous~\cite{AsynDGM} variants.
\subsection{Canonical form}
Similar to the controllable canonical form for linear time-invariant systems, we want our canonical form to have existence and uniqueness properties. That is, any algorithm satisfying \ref{P1}--\ref{P3} should be equivalent to the canonical form through appropriate selection of parameters, with a one-to-one correspondence between algorithms and parameter selections. The canonical form should also be minimal in that it uses the fewest number of parameters possible to express all algorithms in the considered class.
In the canonical form, we use the graph Laplacian matrix $L\in\R^{n\times n}$ to model local communication. This is equivalent to using a gossip matrix $W$ with $W\defeq I-L$. We make the following assumption throughout the paper.
\begin{assumption}\label{assumption:graph} The Laplacian matrix $L\in \R^{n\times n}$ is symmetric, positive semidefinite, and satisfies $L\1=0$, where $\1$ is the all-ones vector. The zero eigenvalue of~$L$ is distinct, implying that the underlying graph is connected.
\end{assumption}
\noindent
Our canonical form satisfies properties \ref{P1}--\ref{P3}, and is parameterized by five real scalars: $\alpha,\zeta_0,\zeta_1,\zeta_2,\zeta_3$.
\vspace{-3mm}
\begin{algorithm}[H]
\small\caption{}\label{alg:canonical}
\begin{algorithmic}
\State\textbf{Initialization:}~Let $L\in\R^{n\times n}$ be a Laplacian matrix. Each agent $i\in\{1,\ldots,n\}$ chooses $x_i^0 \in \R^d$ and $w_i^0\in\R^d$.
\For {iteration $k=0,1,2,\ldots$}
\For {agent $i\in\{1,\ldots,n\}$}
\State \textbf{Local communication}
\State $~~\hphantom{w_i^{k+1}}\mathllap{v_{1i}^k} = \sum_{j=1}^n L_{ij}\,x_j^k \hfill \text{(C.1)}$
\State $~~\hphantom{w_i^{k+1}}\mathllap{v_{2i}^k} = \sum_{j=1}^n L_{ij}\,w_j^k \quad \text{(only required if } \zeta_2 \ne 0) \hfill \text{(C.2)}$
\State \textbf{Local gradient computation}
\State $~~\hphantom{w_i^{k+1}}\mathllap{y_i^k} = x_i^k - \zeta_3\,v_{1i}^k \hfill \text{(C.3)}$
\State $~~\hphantom{w_i^{k+1}}\mathllap{u_i^k} = \df_i(y_i^k) \hfill \text{(C.4)}$
\State \textbf{Local state update}
\State $~~\hphantom{w_i^{k+1}}\mathllap{x_i^{k+1}} = x_i^k + \zeta_0\,w_i^k - \alpha\,u_i^k - \zeta_1 v_{1i}^k + \zeta_2 v_{2i}^k \hfill \text{(C.5)}$
\State $~~\hphantom{w_i^{k+1}}\mathllap{w_i^{k+1}} = w_i^k - v_{1i}^k \hfill \text{(C.6)}$
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\vspace{-4mm}
The algorithm requires agent $i$ to store two local state variables, $x_i^k$ and $w_i^k$. Agents first communicate their states with neighbors using the Laplacian matrix; note that $w_i^k$ need not be transmitted if $\zeta_2=0$ since the result is unused in that case. Agent $i$ then evaluates its local gradient at $y_i^k$, and updates its states using the results of the communication and computation. The sequence $\{y_i^k\}$ is then agent $i$'s estimate of the optimizer~$x^\star$.
\paragraph{Technical conditions.} \textit{The parameters $\alpha,\zeta_0,\zeta_1,\zeta_2,\zeta_3$ are constant scalars that satisfy the following:}
\begin{enumerate}[itemsep=0mm,topsep=1mm,label=T\arabic*.,ref=\textup{T\arabic*}]
\item $\alpha \ne 0$.\label{T1}
\item \textit{The linear system $\alpha u = (\zeta_0 I + \zeta_2 L) w$ has a solution $w\in\R^n$ for all $u\in\R^n$ that satisfies $\1^\tp u = 0$.} \label{T2}
\item \textit{Either $\zeta_0=0$ or the algorithm is initialized such that $\sum_{i=1}^n w_{i}^0=0$. An easy way to satisfy this condition is for every agent to use the initialization $w_i^0=0$.} \label{T3}
\end{enumerate}
\subsection{Main results}
We present the two main results of the paper, which relate the distributed algorithm class to our canonical form.
\begin{defn}[Optimal fixed point]\label{def:opt_fixed_pt}
The canonical form has an \emph{optimal fixed point} if there exists a fixed point $(v_{1i}^\star,v_{2i}^\star,y_i^\star,u_i^\star,x_i^\star,w_i^\star)$ for $i\in\{1,\dots,n\}$ such that ${y_i^\star = x^\star}$ for all $i\in\{1,\ldots,n\}$, where $x^\star$ is defined in~\eqref{eq:distrop-problem}.
\end{defn}
Our first main result states that the technical conditions are \emph{sufficient} for the canonical form to have an optimal fixed point; we prove the result in Appendix~\ref{sec:fixedpointproof}.
\begin{thm}[Sufficiency]\label{thm:fixed_point}
If technical conditions~\ref{T1}--\ref{T3} are satisfied, then the canonical form has a fixed point. Moreover, all fixed points are optimal fixed points.
\end{thm}
Note that the technical conditions are sufficient for the canonical form to have an optimal fixed point, but do \textit{not} guarantee convergence to the fixed point.
Our second main result is that the technical conditions are \textit{necessary} for convergence to an optimal fixed point. The result also characterizes the existence and uniqueness properties of our canonical form. We prove the result in Appendix~\ref{sec:realizationproof}.
\begin{thm}[Necessity]\label{thm:realization}
Consider a distributed algorithm that satisfies properties~\ref{P1}--\ref{P3} and converges to an optimal fixed point. Then there exist parameters $\alpha,\zeta_0,\zeta_1,\zeta_2,\zeta_3$ for which the algorithm is equivalent to our canonical form. Furthermore, the parameters are unique, satisfy technical conditions~\ref{T1}--\ref{T3}, and constitute a minimal parameterization of all such algorithms.
\end{thm}
To prove the main results, we develop the notion of a transfer function for distributed algorithms in Section~\ref{sec:transferfunction} and provide associated necessary conditions that hold when the algorithm converges to an optimal fixed point. We then show how to put distributed algorithms into canonical form through an example in Section~\ref{sec:algorithms}.
\section{Transfer function interpretation}\label{sec:transferfunction}
In this section, we show how to represent distributed algorithms as dynamical systems and obtain corresponding transfer functions. We derive necessary conditions on the poles and zeros of the transfer function of any distributed algorithm that solves~\eqref{eq:distrop-problem}. These conditions are used to prove Theorem~\ref{thm:realization}, and also give rise to an impossibility result stating that no algorithm with a single state can solve the distributed optimization problem.
The general dynamical system model for a distributed optimization algorithm satisfying~\ref{P1}--\ref{P3} is given by\footnote{The iterations~\eqref{eq:alg} are similar to the polynomial linear protocol~\cite{PLP} used to study the dynamic average consensus problem.}
\begin{subequations}\label{eq:alg}
\begin{align}
\bmat{\xi_i^{k+1} \\ y_i^k} &= \biggl(\bmat{A_0 & B_0 \\ C_0 & D_0}\otimes I_d\biggr) \bmat{\xi_i^k \\ u_i^k} \nonumber\\
&\qquad + \biggl(\,\bmat{A_1 & B_1 \\ C_1 & D_1}\otimes I_d\biggr)\sum_{j=1}^n L_{ij} \bmat{\xi_j^k \\ u_j^k} \\
u_i^k &= \df_i(y_i^k)
\end{align}
\end{subequations}
for $i\in\{1,\ldots,n\}$ where $\xi_i^k \in \R^{2d}$ is the state, $u_i^k \in \R^d$ the input, and $y_i^k \in \R^d$ the output of agent~$i$ at iteration~$k$. The model~\eqref{eq:alg} over-parameterizes all algorithms in our class of interest. To simplify notation, we set $d=1$ throughout the rest of the section, though our results hold for general $d\in\mathbb{N}$.
We can write~\eqref{eq:alg} in vectorized form by concatenating the states of the agents as $\xi^k \defeq \bmat{ (\xi_1^{k})^\tp & \cdots & (\xi_n^{k})^\tp}^\tp$ and similarly for $u^k$ and $y^k$. The iterations are then
\begin{subequations}\label{eq:alg_vec}
\begin{align}
\xi^{k+1} &= A(L) \xi^k + B(L) u^k \\
y^k &= C(L) \xi^k + D(L) u^k \\
u^k &= \df(y^k)
\end{align}
\end{subequations}
where the $i\textsuperscript{th}$ component of $\df(y^k)$ is $\df_i(y_i^k)$, and the system matrices are:
\begin{align*}
A(L) &\defeq I_n\otimes A_0 + L\otimes A_1, & \!\!
B(L) &\defeq I_n\otimes B_0 + L\otimes B_1, \\
C(L) &\defeq I_n\otimes C_0 + L\otimes C_1, & \!\!
D(L) &\defeq I_n\otimes D_0 + L\otimes D_1.
\end{align*}
By Assumption~\ref{assumption:graph}, we may write $L = V \Lambda V^\tp$ where $V=\bmat{v_1 & \cdots & v_n}\in\R^{n\times n}$ is an orthogonal matrix of eigenvectors and $\Lambda=\diag(\lambda_1,\ldots,\lambda_n)$ is a diagonal matrix of eigenvalues with $0 = \lambda_1 < \lambda_2 \le \dots \le \lambda_n$. For $\ell=1,\dots,n$ we define the state, input, and output in direction $v_\ell$ as follows:
\begin{equation}\label{eq:coord_transform}
\bar{\xi}_\ell^k \defeq (v_\ell^\tp\otimes I_2) \xi^k \qquad
\bar{u}_\ell^k \defeq v_\ell^\tp u^k \qquad
\bar{y}_\ell^k \defeq v_\ell^\tp y^k
\end{equation}
In these new coordinates, \eqref{eq:alg_vec} is equivalent to the $n$ decoupled single-input single-output systems
\begin{subequations}\label{eq:sep_sys}
\begin{align}
\bar{\xi}_\ell^{k+1} &= (A_0+\lambda_\ell A_1) \bar{\xi}_\ell^k + (B_0+\lambda_\ell B_1) \bar{u}_\ell^k \\
\bar{y}_\ell^k &= (C_0+\lambda_\ell C_1) \bar{\xi}_\ell^k + (D_0+\lambda_\ell D_1) \bar{u}_\ell^k \\
\bar{u}_\ell^k &= v_\ell^\tp \df\bigl( v_1\, \bar{y}_1^k + \cdots + v_n\, \bar{y}_n^k \bigr)
\end{align}
\end{subequations}
for all $\ell \in\{1,\ldots,n\}$ where $\bar{\xi}_\ell^0\in\R^2$. We refer to~\eqref{eq:sep_sys} as the \emph{separated system}, and the corresponding transfer functions are given by
\begin{align}\label{eq:tf_sep}
G_{\lambda_\ell}(z) &=
(C_0+\lambda_\ell C_1) \bigl( z I - (A_0+\lambda_\ell A_1)\bigr)^{-1} (B_0+\lambda_\ell B_1)\nonumber \\
& \qquad +(D_0+D_1\lambda_\ell).
\end{align}
Note that $\ell$ indexes the $n$ directions in the eigenspace of~$L$ as opposed to the $n$ agents, which are indexed by $i$.
Given the generic transfer function $G_{\lambda_\ell}(z)$, imposing that algorithm~\eqref{eq:alg} has an optimal fixed point leads to properties that the transfer functions must satisfy. Here, an optimal fixed point is any point $(\xi_i^\star,y_i^\star,u_i^\star)$ such that $y_i^\star=x^\star$ solves~\eqref{eq:distrop-problem}. We summarize these in the following lemma; see Appendix~\ref{sec:tfproof} for a proof.
\begin{lem}\label{lem:tf}
Suppose algorithm~\eqref{eq:alg} converges to an optimal fixed point. Then the transfer function~\eqref{eq:tf_sep} satisfies the following properties.
\begin{enumerate}
\item $G_{\lambda_1}(z)$ has a pole at $z=1$ and is marginally stable.
\item $G_{\lambda_\ell}(z)$ has a zero at $z=1$ and is strictly stable for all $\ell = 2,\dots,n$.
\end{enumerate}
\end{lem}
The requirements on the transfer function in Lemma~\ref{lem:tf} imply that no single-state algorithm of the form~\eqref{eq:alg} can solve the distributed optimization problem. This provides an explanation as to why the Distributed Gradient Descent algorithm must use a diminishing stepsize~\cite{DGD}. We therefore restrict ourselves to algorithms with two states (i.e., $\xi_i^k\in\R^{2d}$) since these are the simplest algorithms that can achieve linear convergence rates.
\begin{cor}\label{cor:onestate}
No algorithm satisfying properties~\ref{P2} and~\ref{P3} where each agent has a local state of dimension~$d$ can solve the distributed optimization problem~\eqref{eq:distrop-problem}.
\end{cor}
\paragraph{Proof of Corollary~\ref{cor:onestate}.}
Such an algorithm can be written in the form~\eqref{eq:alg} with $A_i,B_i,C_i,D_i\in\R$ for $i=1,2$ (since the state is dimension $d$) and $D_0=D_1=0$. The corresponding transfer function then has the form
\begin{align*}
G_{\lambda_{\ell}}(z) = \frac{(C_0+\lambda_{\ell} C_1) (B_0+\lambda_{\ell} B_1)}{z-(A_0+\lambda_{\ell} A_1)}.
\end{align*}
Suppose the algorithm solves the distributed optimization problem~\eqref{eq:distrop-problem}. Then from Lemma~\ref{lem:tf}, $G_{\lambda_0}(z)$ must have a pole at $z=1$, and $G_{\lambda_\ell}(z)$ must have a zero at $z=1$ for all $\ell=2,\dots,n$. The first condition implies that $A_0=1$ with $B_0 C_0 \neq 0$, while the second condition implies that either $B_0=B_1=0$ or $C_0=C_1=0$. These conditions contradict each other, implying that the algorithm does \emph{not} solve the distributed optimization problem. \qedhere
\section{Putting algorithms in canonical form}\label{sec:algorithms}
We now outline a procedure to put any distributed optimization algorithm in our class of interest in canonical form. Note that the canonical form has the form~\eqref{eq:alg} with
\begin{align}\label{eq:realization}
\begin{aligned}
\stsp{A_0}{B_0}{C_0}{D_0} &= \left[\begin{array}{cc|c} 1 & \zeta_0 & -\alpha \\ 0 & 1 & 0 \\\hlinet 1 & 0 & 0 \end{array}\right]\quad\text{and} \\
\stsp{A_1}{B_1}{C_1}{D_1} &= \left[\begin{array}{cc|c} -\zeta_1 & \zeta_2 & 0 \\ -1 & 0 & 0 \\\hlinet -\zeta_3 & 0 & 0 \end{array}\right]
\end{aligned}
\end{align}
and the corresponding transfer function~\eqref{eq:tf_sep} is
\begin{align}\label{eq:tf_canonical}
G^\text{CF}_{\lambda_{\ell}}(z) = -\frac{\alpha\,(1-\zeta_3\lambda_\ell) (z-1)}{(z-1)(z-1+\zeta_1 \lambda_{\ell}) + \lambda_{\ell}\,(\zeta_0 + \zeta_2 \lambda_{\ell})}.
\end{align}
We can then apply the following procedure to put any distributed algorithm satisfying properties \ref{P1}--\ref{P3} into our canonical form. We summarize the results for several known algorithms in Table~\ref{table::comparison}.
\begin{enumerate}[itemsep=-0.8mm,topsep=0.8mm]
\item Write the algorithm in the form~\eqref{eq:alg} to obtain the system matrices $(A_0,B_0,C_0,D_0)$ and $(A_1,B_1,C_1,D_1)$.
\item Compute the transfer function $G_{\lambda_{\ell}}(z)$ using~\eqref{eq:tf_sep}.
\item Obtain parameters $\alpha,\zeta_0,\zeta_1,\zeta_2,\zeta_3$ by comparing coefficients of the transfer function $G_{\lambda_{\ell}}(z)$ with that of the canonical form $G^\text{CF}_{\lambda_{\ell}}(z)$ in~\eqref{eq:tf_canonical}.
\end{enumerate}
\begin{table}[htb]
\caption{Parameters of distributed optimization algorithms in canonical form using $W = I-L$.}\label{table::comparison}
\centering
\renewcommand{\arraystretch}{1.1
\begin{tabular}{p{3.7cm}cccc}
& $\zeta_0$ & $\zeta_1$ & $\zeta_2$ & $\zeta_3$ \\ \hline
EXTRA~\cite{EXTRA}
& $\frac{1}{2}$ & $1$ & $0$ & $0$ \\
NIDS~\cite{NIDS}
& $\frac{1}{2}$ & $1$ & $0$ & $\frac{1}{2}$ \\
Exact Diffusion~\cite{ExactDiffusion}
& $\frac{1}{2}$ & $1$ & $0$ & $\frac{1}{2}$ \\
DIGing~\cite{DIGing,QuLi}
& $0$ & $2$ & $1$ & $0$ \\
AsynDGM\tablefootnote{The similar algorithm AugDGM~\cite{AugDGM} does not fit in our framework because it requires two sequential rounds of communication per iteration.}~\cite{AsynDGM}
& $0$ & $2$ & $1$ & $1$ \\
Jakoveti\'{c}~\cite{Unification} ($\mathcal{B}=\beta I$)
& $\alpha\beta$ & $2$ & $1$ & $0$ \\
Jakoveti\'{c}~\cite{Unification} ($\mathcal{B}=\beta W$)
& $\alpha\beta$ & $2$ & $1-\alpha\beta$ & $0$ \\
\end{tabular}\vspace{-6mm}
\end{table}
\noindent Since the parameters of our canonical form are unique, we can compare various algorithms by putting them in this form. For example, from Table~\ref{table::comparison} we observe that NIDS and Exact Diffusion have the same parameters and are therefore equivalent. Similarly, for the algorithm of Jakoveti\'{c}, choosing $\mathcal{B}=0$ recovers DIGing, and $\mathcal{B}=\frac{1}{\alpha} W$ with $W\mapsto\frac{1}{2}(I+W)$ recovers EXTRA, as noted in~\cite{Unification}.
\begin{rem}
To provide an additional degree of freedom, we can relate the gossip matrix to the Laplacian matrix by $W = I-\mu L$ where $\mu\in\R$ is an additional scalar parameter. This is equivalent to choosing the gossip matrix as a convex combination of $I-L$ and the identity, i.e., $W = (1-\mu) I + \mu (I-L)$.
\end{rem}
\begin{rem}[Uniqueness of realization]\label{rem:non_unique_realization}
The canonical form~\eqref{eq:realization} can be reliably obtained from an algorithm by computing its transfer function~\eqref{eq:tf_canonical} and then extracting coefficients. In contrast, working with realizations directly can be problematic because even if coordinates are chosen such that $(A_0,B_0,C_0,D_0)$ has the desired form, this does not ensure uniqueness of $(A_1,B_1,C_1,D_1)$. For example, one can easily check that the following realization has the same transfer function as~\eqref{eq:realization}:
\begin{align*}
\stsp{A_0}{B_0}{C_0}{D_0} &= \left[\begin{array}{cc|c} 1 & \zeta_0 & -\alpha \\ 0 & 1 & 0 \\\hlinet 1 & 0 & 0 \end{array}\right] \quad\text{and} \\
\stsp{A_1}{B_1}{C_1}{D_1} &= \left[\begin{array}{cc|c} -\zeta_1 & \zeta_2 & \alpha\zeta_3 \\ -1 & 0 & 0 \\\hlinet 0 & 0 & 0 \end{array}\right].
\end{align*}
The equivalence class of realizations for a doubly-indexed transfer function (i.e. with $z$ and $\lambda$) is more involved than the singly-indexed case and a broader discussion on this topic is outside the scope of the present work; seminal references include~\cite{fornasini,roesser}.
\end{rem}
\subsection{Example: NIDS}
To illustrate our approach, we return to the NIDS algorithm from Section~\ref{sec:intro}. As suggested by the authors of~\cite{NIDS}, we choose $\widetilde{W}\defeq (I+W)/2$ where the gossip matrix~$W$ is related to the Laplacian matrix~$L$ by $W = I-L$.
First, we write NIDS in the form~\eqref{eq:alg} as follows:
\begin{equation*}
\begin{aligned}
\bmat{x^{k+2} \\ x^{k+1} \\ \df(y^k)} &= \bmat{2I-L & \!\frac{1}{2} L-I & \!\alpha(I-\frac{1}{2} L) \\ I & 0 & 0 \\ 0 & 0 & 0} \!\! \bmat{x^{k+1} \\ x^k \\ \df(y^{k-1})} \\
&\qquad + \bmat{\alpha(\frac{1}{2} L - I) \\ 0 \\ I} \df(y^k) \\
y^k &= \bmat{I & 0 & 0} \bmat{x^{k+1} \\ x^k \\ \df(y^{k-1})}
\end{aligned}
\end{equation*}
This is equivalent to~\eqref{eq:alg} with state-space matrices
\begin{align*}
&\stsp{A_0}{B_0}{C_0}{D_0} = \left[\begin{array}{ccc|c} 2 & -1 & \alpha & -\alpha \\ 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 \\\hlinet 1 & 0 & 0 & 0 \end{array}\right] \qquad\text{and} \\
&\stsp{A_1}{B_1}{C_1}{D_1} = \left[\begin{array}{ccc|c} -1 & \frac{1}{2} & -\frac{\alpha}{2} & \frac{\alpha}{2} \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\\hlinet 0 & 0 & 0 & 0 \end{array}\right],
\end{align*}
which have associated transfer function~\eqref{eq:tf_sep} given by
\begin{align*}
G_{\lambda_{\ell}}(z) = -\frac{\alpha\,(1-\frac{1}{2}\lambda_{\ell})(z-1)}{(z-1) (z-1+\lambda_{\ell}) + \frac{1}{2}\lambda_{\ell}}.
\end{align*}
Comparing coefficients with that of $G_{\lambda_{\ell}}^\text{CF}(z)$ in~\eqref{eq:tf_canonical}, we find that the parameters of the canonical form are $(\zeta_0,\zeta_1,\zeta_2,\zeta_3) = (\frac{1}{2},1,0,\frac{1}{2})$.
\section{Conclusion}\label{sec:conclusion}
In this paper, we derived a canonical form for a large class of first-order distributed algorithms that fulfills existence and uniqueness properties, and features a minimal parameterization. We provided sufficient conditions for an algorithm in canonical form to have a fixed point corresponding to the solution of the distributed optimization problem, as well as necessary conditions for convergence to such a fixed point.
Combined with analysis tools such as those described in~\cite{sundararajan_allerton,lessard16,iqcadmm_ICML} and references therein, this work provides a first step toward a principled and automated methodology for the analysis and design of distributed optimization algorithms.
\bibliographystyle{abbrv}
{\small |
\section{Introduction}
In the second half of the 19th century mathematicians became interested in the normal subgroups of classical groups, at first over finite fields and then over infinite fields. The case of finite fields was of special interest because normal subgroups here yielded concrete examples of finite nonabelian simple groups, which were important for Galois theory. In 1870, C. Jordan \cite{jordan} showed that the normal subgroups of the general linear group $\operatorname{GL}_n(\mathbb{F}_p)$, where $\mathbb{F}_p$ is a finite field of prime order $p$, are precisely the overgroups of the elementary subgroup $\operatorname{E}_n(\mathbb{F}_p)$ and the subgroups of $\operatorname{Center}(\operatorname{GL}_n(\mathbb{F}_p))$. In 1900, L. E. Dickson \cite{dickson, dickson_book} extended Jordan's result to all fields and further to all classical groups. In the period 1948-1963, J. Dieudonne \cite{dieudonne_book_1,dieudonne_1,dieudonne_2,dieudonne_book_2} generalized Dickson's results to division rings.
In 1961, W. Klingenberg \cite{klingenberg_1} proved that if $R$ is a commutative local ring, $n\geq 3$, and $H$ is a subgroup of $\operatorname{GL}_n(R)$, then
\begin{equation}
H\text{ is normal}~\Leftrightarrow~ \exists \text{ ideal }I:~\operatorname{E}_n(R,I)\subseteq H\subseteq \operatorname{C}_n(R,I)\end{equation}
where $E_n(R,I)$ denotes the elementary subgroup of level $I$ and $\operatorname{C}_n(R,I)$ the full congruence subgroup of level $I$. If $R$ is a field or division ring one gets back the results of Dickson resp. Dieudonne. Klingenberg proved analog results for the symplectic and the orthogonal groups \cite{klingenberg_2,klingenberg_3}. In \cite{klingenberg_4}, Klingenberg tried to generalize (1) to arbitrary (not necessarily commutative) local rings $R$.
In 1964, H. Bass \cite{bass} noticed that Klingenberg had made a mistake. What Klingenberg really had shown in \cite{klingenberg_4} was that if $R$ is a local ring, $n\geq 3$, and $H$ is a subgroup of $\operatorname{GL}_n(R)$, then
\begin{equation}
H\text{ is E-normal}~\Leftrightarrow~ \exists \text{ ideal }I:~\operatorname{E}_n(R,I)\subseteq H\subseteq \operatorname{C}_n(R,I).\end{equation}
Recall that a subgroup of a classical group $G$ is called {\it E-normal} if it is normalized by the elementary subgroup of $G$. Further Bass showed that (2) holds true if $R$ is an arbitrary ring and $n$ is large enough with respect to the stable rank of $R$, including the case that $R$ is semilocal and $n\geq 3$. From that moment on the focus of research changed. It was now on the E-normal structure of classical and classical-like groups instead of the normal structure. Amazingly, it seems that today it is still unknown what the normal subgroups of the general linear group over a commutative ring are (while a description of the E-normal subgroups is available, see the next paragraph).
In 1972, J. S. Wilson \cite{wilson} proved that (2) holds if $R$ is commutative and $n\geq 4$. In 1973, I. Z. Golubchik \cite{golubchik} added the case $n=3$. In 1981, L. Vaserstein \cite{vaserstein} showed that (2) holds if $R$ is almost commutative (i.e. module finite over its center) and $n\geq 3$.
A natural question is if similar results hold for other classical groups. In 1969, A.
Bak \cite{bak_2} defined hyperbolic unitary groups $\operatorname{U}_{2n}(R,\Lambda)$ which include as special cases the classical Chevalley groups of type $C_m$ and $D_m$. He proved that if $H$ is a subgroup of $\operatorname{U}_{2n}(R,\Lambda)$, then
\begin{align}
H \text{ is E-normal}~\Leftrightarrow~ \exists\text{ form ideal } (I,\Gamma):\operatorname{EU}_{2n}((R,\Lambda),(I,\Gamma)) \subseteq H\subseteq \operatorname{CU}_{2n}((R,\Lambda),(I,\Gamma))
\end{align}
if $n$ is large enough with respect to the Bass-Serre dimension of $R$. In 1995, L. N. Vaserstein and H. You \cite{vaserstein-you} gave an incorrect proof of (3) for the case that $R$ is almost commutative and $n\geq 3$. The proof can be repaired when $2$ is invertible in $R$. In 2012, H. You \cite{you} proved that (3) holds if $R$ is commutative and $n\geq 4$ and one year later H. You and X. Zhou \cite{you-zhou} added the case that $n=3$. In 2017, the author \cite{preusser} proved that (3) holds if $R$ is almost commmutative and $n\geq 3$.
In 2005, V. A. Petrov \cite{petrov} defined odd unitary groups $\operatorname{U}(\mathcal{V})$. These groups are very general, they include as special cases all the classical Chevalley groups, Bak's hyperbolic unitary groups and the Hermitian groups of G. Tang \cite{tang}. Recently Bak and the author \cite{bak-preusser} studied a subclass of Petrov's odd unitary groups, big enough to contain all the classical Chevalley groups, and classified the E-normal subgroups of the members of this subclass. Also recently W. Yu, Y. Li and H. Liu \cite{yu-li-liu} studied the E-normal subgroups of odd unitary groups under a $\Lambda$-stable rank condition.
In this paper we investigate the E-normal subgroups of Petrov's odd unitary groups. The main result is the following: Let $\mathcal{V}$ be an odd quadratic space of Witt index $\geq 3$ over a commutative ring with pseudoinvolution. Choose three mutually orthogonal hyperbolic pairs $(e_1,e_{-1}),(e_2,e_{-2}),(e_3,e_{-3})$ in $\mathcal{V}$ and let $\mathcal{V}_0$ be the orthogonal complement of the odd quadratic space spanned by $e_1,e_2,e_3,e_{-3},e_{-2},e_{-1}$. Then the odd unitary group $\operatorname{U}(\mathcal{V})$ is isomorphic to the odd hyperbolic unitary group $\operatorname{U}_{6}(\mathcal{V}_0)$ see \cite[p. 4756]{petrov}. We prove that if $H$ is a subgroup of $\operatorname{U}_{6}(\mathcal{V}_0)$, then
\begin{equation}
H \text{ is E-normal}~\Leftrightarrow~\exists\text{ odd form ideal } \mathfrak{I}_0:\operatorname{EU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)\subseteq H\subseteq \operatorname{CU}_{6}(\mathcal{V}_0,\mathfrak{I}_0).
\end{equation}
A priori, the notion of E-normality depends on the choice of the hyperbolic pairs $(e_1,e_{-1}),(e_2,e_{-2}),(e_3,e_{-3})$. Petrov has shown that it depends only on the choice of one hyperbolic pair, see \cite[Section 7]{petrov}.
Furthermore we correct a couple of small mistakes that have been made in \cite{petrov} and later were repeated in \cite{yu-tang} and \cite{yu-li-liu}. See the remarks in Sections 3 and 4.
The rest of the paper is organized as follows. In Section 2 we recall some standard notation
which will be used throughout the paper. In Section 3 we recall the definitions of Petrov's odd unitary groups and some important subgroups. In Section 4 we recall the definitions of the odd hyperbolic unitary groups and some important subgroups. In Section 5 we prove (4).
\section{Notation}
$\mathbb{N}$ denotes the set of positive integers and $\mathbb{N}_0$ the set of nonnegative integers. If $G$ is a group and $g,h\in G$, we let $^hg:=hgh^{-1}$ and $[g,h]:=ghg^{-1}h^{-1}$. By a ring we will always mean an associative ring with $1$ such that $1\neq 0$. Ideal will mean two-sided ideal. If $x_1,\dots,x_n$ are elements of a ring $R$, then we denote by $I(x_1,\dots,x_n)$ the ideal of $R$ generated by the $x_i$'s. The set of all $n\times n$ matrices over a ring $R$ is denoted by $\operatorname{M}_n(R)$.
\section{Petrov's odd unitary groups}
\subsection{Odd unitary groups}
Throughout Section 3, $R$ denotes a ring and $~\bar{}~$ a pseudoinvolution on $R$. Recall that a {\it pseudoinvolution} on $R$ is an additive map $~\bar{}: R\rightarrow R$, $x\mapsto \bar x$ such that $\overline{xy}=\bar y \bar 1^{-1}\bar x$ and $\bar{\bar{x}}=x$ for any $x,y\in R$.
\begin{definition}
Let $V$ be a right $R$-module. A biadditive map $B:V\times V\rightarrow R$ such that $B(vx,wy)=\bar x \bar 1^{-1}B(v,w) y$ for any $v,w\in V$ and $x,y\in R$ is called a {\it sesquilinear form on $V$}. A sesquilinear form $B$ on $V$ is called an {\it anti-Hermitian form on $V$} if $B(v,w)=-\overline{B(w,v)}$ for any $v,w\in V$. If $B$ is an anti-Hermitian form on $V$, then the pair $(V,B)$ is called an {\it anti-Hermitian space}.
\end{definition}
\begin{definition}
Let $(V,B)$ be an anti-Hermitian space. The {\it Heisenberg group $\mathfrak{H}$ of $(V,B)$} is defined as follows. As a set $\mathfrak{H}=V\times R$. The composition is given by
\[(v,x)\+(w,y)=(v+w,x+y+B(v,w)).\]
The inverse $\overset{.}{-} (v,x)$ of an element $(v,x)\in\mathfrak{H}$ is given by $\overset{.}{-}(v,x)=(-v,-x+B(v,v))$.
\end{definition}
\begin{definition}
Let $(V,B)$ be an anti-Hermitian space. Define a right action $\leftharpoonup$ of the multiplicative monoid $(R,\cdot)$ on $\mathfrak{H}$ by
\[(v,x)\leftharpoonup y=(vy,\bar y\bar 1^{-1}xy).\]
This action has the property that $(\alpha\+\beta)\leftharpoonup y=\alpha\leftharpoonup y\+ \beta\leftharpoonup y$ for any $\alpha, \beta\in \mathfrak{H}$ and $y\in R$. \\
\end{definition}
\begin{definition}
Let $(V,B)$ be an anti-Hermitian space. The {\it trace map} $\operatorname{tr}:\mathfrak{H}\rightarrow R$ is defined by
\[\operatorname{tr}(v,x)=x-\bar x-B(v,v).\]
One checks easily that $\operatorname{tr}$ is a group homomorphism from $\mathfrak{H}$ to $(R,+)$.
\end{definition}
\begin{definition}
Let $(V,B)$ be an anti-Hermitian space. Set
\begin{align*}
\mathfrak{L}_{\min}&:=\{(0,x+\bar x)\mid x\in R\},\\
\mathfrak{L}_{\max}&:=\ker\operatorname{tr}.
\end{align*}
A subgroup $\mathfrak{L}$ of $\mathfrak{H}$ that is stable under $\leftharpoonup$ and lies between $\mathfrak{L}_{\min}$ and $\mathfrak{L}_{\max}$ is called an {\it odd form parameter}. Since $\mathfrak{L}_{\min}$ and $\mathfrak{L}_{\max}$ are subgroups of $\mathfrak{H}$ that are stable under $\leftharpoonup$, they are the smallest resp. the largest odd form parameter.
\end{definition}
One checks easily that $\mathfrak{L}_{\min}$ contains the commutator subgroup of $\mathfrak{H}$. Hence any odd form parameter is a normal subgroup of $\mathfrak{H}$.
\begin{definition}\label{defoqs}
Let $(V,B)$ be an anti-Hermitian space. Define the map
\begin{align*}
Q:V&\rightarrow \mathfrak{H}\\
v&\mapsto (v,0)
\end{align*}
Obviously
\begin{align*}
&Q(vx)=Q(v)\leftharpoonup x~\forall v\in V, x\in R\text{ and}\\
&Q(v+w)=Q(v)\+Q(w)\overset{.}{-}(0,B(v,w))~\forall v,w\in V.
\end{align*}
If $\mathfrak{L}$ is an odd form parameter, then the map $Q_\mathfrak{L}:V\rightarrow \mathfrak{H}/\mathfrak{L}$ induced by $Q$ is called the {\it odd quadratic form on $V$ defined by $\mathfrak{L}$}. The triple $\mathcal{V}:=(V, B, Q_\mathfrak{L})$ is called an {\it odd quadratic space}.
\end{definition}
\begin{remark}
In \cite{petrov} the odd quadratic form $Q_\mathfrak{L}$ is not explicitly mentioned, but it appears implicitly in the definition of an odd unitary group. Petrov calls the pair $q=(B,\mathfrak{L})$ an odd quadratic form and the pair $(V,q)$ and odd quadratic space.
\end{remark}
Until the end of Section 4, $\mathcal{V}=(V, B, Q_\mathfrak{L})$ denotes an odd quadratic space.
\begin{definition}
The group
\[\operatorname{U}(\mathcal{V})=\{\sigma\in \operatorname{GL}(V)\mid B(\sigma v,\sigma w)=B(v,w)\land Q_\mathfrak{L}(\sigma v)=Q_\mathfrak{L}(v)~\forall v,w\in V\}\]
is called the {\it odd unitary group of $\mathcal{V}$}.
\end{definition}
In the following we will denote the inverse of an element $\sigma\in \operatorname{U}(\mathcal{V})$ by $\widetilde \sigma$ instead of $\sigma^{-1}$ in order to get nicer-looking formulas.
\subsection{Eichler-Siegel-Dickson transvections}
\begin{definition}
Two elements $u,v\in V$ are called {\it orthogonal} if $B(u,v)=0$. An element $w\in V$ is called {\it isotropic} if $Q_\mathfrak{L}(w)=0$.
\end{definition}
\begin{definition}
Let $u\in V$ be an isotropic element and $(v,x)\in \mathfrak{L}$ such that $u$ and $v$ are orthogonal. Then the endomorphism $T_{uv}(a)$ of $V$ mapping
\[w\mapsto w+u\bar 1^{-1}(B(v,w)+xB(u,w))+vB(u,w)\]
is called an {\it Eichler-Siegel-Dickson transvection} or an {\it ESD transvection}.
\end{definition}
\begin{lemma}\label{lemesd}
The ESD transvections lie in $\operatorname{U}(\mathcal{V})$. Furthermore
\begin{align*}
&T_{uv_1}(x)T_{uv_2}(y)=T_{u,v_1+v_2}(x+y+B(v_1,v_2))\text{ and}\\
&^\sigma\!T_{u,v}(x)=T_{\sigma u,\sigma v}(x)\text{ for any }\sigma \in \operatorname{U}(\mathcal{V}).
\end{align*}
\end{lemma}
\begin{proof}
See \cite[Section 4]{petrov}.
\end{proof}
\subsection{Principal congruence subgroups}
The submodule $V^{ev}:=\{v\in V\mid (v,x)\in \mathfrak{L}\text{ for some }x\in R\}$ of $V$ is called the {\it even part of $V$}. We denote the Heisenberg group of the anti-Hermitian space $(V^{ev}, B|_{V^{ev}})$ by $\mathfrak{H}^{ev}$. It is easy to see that any $\sigma \in \operatorname{U}(\mathcal{V})$ restricts to an automorphism of $V^{ev}$ (in fact any $\sigma \in \operatorname{U}(\mathcal{V})$ has the stronger property that $(\sigma v,x)\in \mathfrak{L}$ for any $(v,x)\in\mathfrak{L}$).
\begin{definition}\label{defrelfp}
Let $I$ be an ideal of $R$ such that $\bar I=I$. Set
\begin{align*}
\mathfrak{M}_{\min}&:=\mathfrak{L}\leftharpoonup I\+\{(0,x+\bar x)\mid x\in I\},\\
\mathfrak{M}_{\max}&:=\{(v,x)\in \mathfrak{L}\mid x\in I, B(v,w)\in I~\forall w\in V^{ev}\}.
\end{align*}
A subgroup $\mathfrak{M}$ of $\mathfrak{H}$ that is stable under $\leftharpoonup$ and lies between $\mathfrak{M}_{\min}$ and $\mathfrak{M}_{\max}$ is called an {\it relative odd form parameter of level $I$}. Since $\mathfrak{M}_{\min}$ and $\mathfrak{M}_{\max}$ are subgroups of $\mathfrak{H}$ that are stable under $\leftharpoonup$, they are the smallest resp. the largest relative odd form parameter of level $I$. If $\mathfrak{M}$ is a relative odd form parameter of level $I$, then the pair $\mathfrak{I}=(I,\mathfrak{M})$ is called an {\it odd form ideal}.
\end{definition}
\begin{remark}
In \cite{petrov}, $\mathfrak{M}_{\max}$ was defined as $\mathfrak{M}_{\max}=\{(u, a)\in\mathfrak{L} \mid a\in I, B_0(u, v)\in I\text{ for any } v\in V_0\}$. Obviously the zeros should be removed. Further $V$ should be replaced by $V^{ev}$ because otherwise it is not clear that the level of an E-normal subgroup as defined in Definition \ref{deflevel} is an odd form ideal. In an e-mail to the author Petrov confirmed that the definition of $\mathfrak{M}_{\max}$ in Definition \ref{defrelfp} above is correct. Unfortunately, Petrov's definition of $\mathfrak{M}_{\max}$ was copied in \cite{yu-tang} and \cite{yu-li-liu}. In \cite{yu-li-liu} this led to an unnatural definition of the level of an E-normal subgroup and as a consequence to a very strange looking main result.
\end{remark}
One checks easily that any relative odd form parameter $\mathfrak{M}$ is a normal subgroup of $\mathfrak{H}^{ev}$. The map $Q_{\mathfrak{M}}:V^{ev}\rightarrow \mathfrak{H}^{ev}/\mathfrak{M}$ induced by the map $Q$ defined in Definition \ref{defoqs} is called the {\it relative odd quadratic form on $V^{ev}$ defined by $\mathfrak{M}$}.
\begin{definition}
Let $\mathfrak{I}=(I,\mathfrak{M})$ be an odd form ideal. The subgroup
\[\operatorname{U}(\mathcal{V}, \mathfrak{I})=\{\sigma\in\operatorname{U}(\mathcal{V})\mid Q_{\mathfrak{M}}(\sigma v)=Q_{\mathfrak{M}}(v)~\forall v\in V^{ev}\}\]
of $\operatorname{U}(\mathcal{V})$ is called the {\it principal congruence subgroup of level $\mathfrak{I}$}. The normalizer of $\operatorname{U}(\mathcal{V}, \mathfrak{I})$ in $\operatorname{U}(\mathcal{V})$ is called the {\it normalized principal congruence subgroup of level $\mathfrak{I}$} and is denoted by $\operatorname{NU}(\mathcal{V}, \mathfrak{I})$.
\end{definition}
\begin{remark}
In \cite{petrov}, the principal congruence subgroup of level $\mathfrak{I}$ was defined as the subgroup of $\operatorname{GL}(V)$ consisting of all elements $g$ such that $(gv - v, (v - gv, v)_q)$ lies in $\mathfrak{M}$ for every $v\in V$. Obviously $\operatorname{GL}(V)$ should be replaced by $\operatorname{U}(V,q)$. Further ``for every $v\in V$" should be replaced by ``for every $v\in V^{ev}$" (Petrov agreed on that). The reason for that is, that otherwise it is not clear that in the hyperbolic case the relative elementary subgroup of a certain level is contained in the principal congruence subgroup of the same level.
\end{remark}
\begin{definition}\label{defaction}
Let $\sigma\in \operatorname{U}(\mathcal{V})$. If $\mathfrak{I}=(I,\mathfrak{M})$ is an odd form ideal, set $^{\sigma}\mathfrak{I}:=(I,{}^{\sigma}\mathfrak{M})$ where $^{\sigma}\mathfrak{M}:=\{(\sigma v,x)\mid (v,x)\in \mathfrak{M}\}$. This defines a left action of $\operatorname{U}(\mathcal{V})$ on the set of all form ideals.
\end{definition}
\begin{proposition}\label{proptild}
Let $\sigma\in \operatorname{U}(\mathcal{V})$ and $\mathfrak{I}$ an be an odd form ideal. Then $^{\sigma}\!\operatorname{U}(\mathcal{V}, \mathfrak{I})=\operatorname{U}(\mathcal{V}, {}^{\sigma}\mathfrak{I})$.
\end{proposition}
\begin{proof}
Clearly it suffices to show that $^{\sigma}\!\operatorname{U}(\mathcal{V}, \mathfrak{I})\subseteq \operatorname{U}(\mathcal{V}, {}^{\sigma}\mathfrak{I})$. Let $\tau\in\operatorname{U}(\mathcal{V}, \mathfrak{I})$ and $v\in V^{ev}$. Then $\widetilde\sigma v\in V^{ev}$. Since $\tau$ preserves $Q_\mathfrak{M}$, we have
\begin{align*}
(\tau\widetilde\sigma v-\widetilde\sigma v,B(\widetilde\sigma v-\tau\widetilde\sigma v,\widetilde\sigma v))\in\mathfrak{M}.
\end{align*}
Since $\sigma$ preserves $B$, it follows that
\begin{align*}
(\tau\widetilde\sigma v-\widetilde\sigma v,B(v-\sigma\tau\widetilde\sigma v,v))\in\mathfrak{M}
\end{align*}
and hence
\begin{align*}
(\sigma\tau\widetilde\sigma v-v,B(v-\sigma\tau\widetilde\sigma v,v))\in{}^{\sigma}\mathfrak{M}.
\end{align*}
Thus $\sigma\tau\widetilde\sigma $ preserves $Q_{^{\sigma}\mathfrak{M}}$ and therefore lies in $\operatorname{U}(\mathcal{V},{}^{\sigma}\mathfrak{I})$.
\end{proof}
\begin{corollary}
Let $\sigma\in \operatorname{U}(\mathcal{V})$ and $\mathfrak{I}$ an be an odd form ideal. Then $^{\sigma}\!\operatorname{NU}(\mathcal{V}, \mathfrak{I})=\operatorname{NU}(\mathcal{V}, {}^{\sigma}\mathfrak{I})$.
\end{corollary}
\section{The hyperbolic case}
In this section $R$ denotes a ring with pseudoinvolution $~\bar{}~$.
\subsection{Odd hyperbolic unitary groups}
\begin{definition}
Let $\mathcal{V}=(V, B, Q_\mathfrak{L})$ and $\mathcal{V}'=(V', B', Q_{\mathfrak{L}'})$ denote odd quadratic spaces. Their {\it orthogonal sum} is the odd quadratic space
\[\mathcal{V}\operp \mathcal{V}':=(V\oplus V', B+B', Q_{\mathfrak{L}+{\mathfrak{L}}'})\]
where
\[B+B'(v+v',w+w')=B(v,w)+B'(v',w')\]
and
\[\mathfrak{L}+{\mathfrak{L}}'=\{(v+v',x+x')\mid (v,x)\in\mathfrak{L}, (v',x')\in {\mathfrak{L}}'\}.\]
\end{definition}
\begin{definition}
Let $H$ be the free right $R$-module with basis $\{e_1, e_{-1}\}$. Let $B$ be the anti-Hermitian form on $H$ defined by $B_H(e_1, e_{-1})=1$ and $B_H(e_1, e_1)=0=B(e_{-1}, e_{-1})$. Further set $\mathfrak{L}_H:=\{(e_1x_1+e_{-1}x_{-1},\bar x_1\bar 1^{-1}x_{-1}+y+\bar y)\mid x_1,x_{-1},y\in R\}$. Then $\mathcal{H}:=(H, B_H, \mathfrak{L}_H)$ is an odd quadratic space which is called the {\it hyperbolic plane}. For any $n\in \mathbb{N}$ we denote by $\mathcal{H}^n$ the orthogonal sum of $n$ copies of the hyperbolic plane $\mathcal{H}$.
\end{definition}
\begin{definition}
An {\it odd hyperbolic unitary space of rank $n$} is an orthogonal sum $\mathcal{H}^n\operp \mathcal{V}_0$ where $\mathcal{V}_0=(V_0, B_0, Q_{\mathfrak{L}_0})$ is some odd quadratic space. The odd unitary group $\operatorname{U}(\mathcal{H}^n\operp \mathcal{V}_0)$ is called an {\it odd hyperbolic unitary group} and is denoted by $\operatorname{U}_{2n}(\mathcal{V}_0)$.
\end{definition}
\begin{definition}\label{defwit}
Let $\mathcal{V}=(V, B, Q_\mathfrak{L})$ be an odd quadratic space. A {\it hyperbolic pair} is a pair $(v,w)$ where $v$ and $w$ are isotropic elements of $V$ such that $B(v,w)=1$. The greatest $n\in \mathbb{N}_0$ satisfying the condition that there exist $n$ mutually orthogonal hyperbolic pairs in $V$ is called the Witt index of $\mathcal{V}$ and is denoted by $\operatorname{ind} \mathcal{V}$. It is easy to see that if $1\leq n\leq \operatorname{ind} \mathcal{V}$, then $\operatorname{U}(\mathcal{V})$ is isomorphic to an odd hyperbolic unitary group $\operatorname{U}_{2n}(\mathcal{V}_0)$, cf. \cite[Section 5]{petrov}.
\end{definition}
Until the end of Section 4, $n\in \mathbb{N}$ denotes a nonnegative integer and $\mathcal{V}_0=(V_0, B_0, Q_{\mathfrak{L}_0})$ an odd quadratic space. We set $\mathcal{V}:=\mathcal{H}^n\operp \mathcal{V}_0$ and write $\mathcal{H}^n=(V_h, B_h, Q_{\mathfrak{L}_h})$ and $\mathcal{V}=(V, B, Q_\mathfrak{L})$ (hence $V=V_h\oplus V_0$, $B=B_h+B_0$ and $\mathfrak{L}=\mathfrak{L}_{h}+\mathfrak{L}_0$). We denote the basis of $V_h$ coming from the bases of the orthogonal summands by $e_1,\dots, e_n, e_{-n}, \dots, e_{-1}$. Further we set $\Theta_+:=\{1,\dots,n\}$, $\Theta_{-}:=\{-n,\dots,-1\}$ and $\Theta:=\Theta_+\cup \Theta_-$. We define an order $<$ on $\Theta$ by $1<\dots<n<-n<\dots<-1$. We set $\epsilon_i:=\bar 1^{-1}$ if $i\in\Theta_+$ and $\epsilon_i:=-1$ if $i\in \Theta_-$.\\
If $v\in V$, then there are uniquely determined $v_h\in V_h$ and $v_0\in V_0$ such that $v=v_h+v_0$. Moreover, there are uniquely determined $v_i\in R~(i\in \Theta)$ such that $v_h=\sum\limits_{i\in\theta}e_iv_i$. Define a sesquilinear form $F_h$ on $V_h$ by
\[F_h(v_h,w_h)=\sum\limits_{i\in\Theta_+}\bar v_i\bar 1^{-1}w_{-i}\quad\forall v_h,w_h\in V_h.\]
Then
\[B_{h}(v_h,w_h)=F_h(v_h,w_h)-\overline{F_h(w_h,v_h)}\quad\forall v_h,w_h\in V_h\]
and
\[\mathfrak{L}_{h}=\{(v_h,F_h(v_h,v_h)+x+\bar x)\mid v_h\in V_h,x\in R\}.\]
For a $\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0)$ define the linear maps
\begin{align*}
\sigma^{hh}:V_h&\rightarrow V_h,&\sigma^{h0}:V_h&\rightarrow V_0,&\sigma^{0h}:V_0&\rightarrow V_h,&\sigma^{00}:V_0&\rightarrow V_0.\\
v_h&\mapsto (\sigma v_h)_h &v_h&\mapsto(\sigma v_h)_0&v_0&\mapsto (\sigma v_0)_h &v_0&\mapsto (\sigma v_0)_0
\end{align*}
Since $V_h$ is a free right $R$-module of rank $2n$, we can identify $\sigma^{hh}$ with a matrix in $\operatorname{M}_{2n}(R)$. Instead of $\sigma_{ij}^{hh}$ we write just $\sigma_{ij}$ for the entry of $\sigma^{hh}$ at position $(i,j)$ (which equals $(\sigma e_j)_i$).\\
\begin{lemma}\label{lemodd}
Let $\sigma \in \operatorname{GL}(V)$. Then $\sigma\in U_{2n}(\mathcal{V}_0)$ iff Conditions (i)-(vi) below are satisfied.
\begin{enumerate}[(i)]
\item $\widetilde\sigma_{ij}=-\epsilon_i\bar\sigma_{-j,-i}\epsilon_{-j}$ for any $i,j\in\Theta$.
\item $(\widetilde\sigma^{0h}v_0)_{i}=-\epsilon_iB_0(\sigma^{h0}e_{-i},v_0)$ for any $v_0\in V_0$ and $i\in\Theta$.
\item $B_0(v_0,\widetilde\sigma^{h0}e_i)=\overline{(\sigma v_0)_{-i}}\epsilon_{-i}$ for any $v_0\in V_0$ and $i\in\Theta$.
\item $B_0(v_0,\widetilde\sigma^{00}w_0)=B_0(\sigma^{00}v_0,w_0)$ for any $v_0,w_0\in V_0$.
\item $Q(\sigma e_i)\in \mathfrak{L}$ for any $i\in\Theta$.
\item $Q(\sigma v_0)\overset{.}{-} Q(v_0)\in \mathfrak{L}$ for any $v_0\in V_0$.
\end{enumerate}
\end{lemma}
\begin{proof}
($\Rightarrow$) Suppose that $\sigma \in U_{2n}(\mathcal{V}_0)$. Then
\begin{align*}
\widetilde\sigma_{ij}=-\epsilon_iB(e_{-i},\widetilde\sigma e_j)=-\epsilon_iB(\sigma e_{-i},e_j)=-\epsilon_i\bar\sigma_{-j,-i}\epsilon_{-j}
\end{align*}
for any $i,j\in \Theta$ since $\sigma$ preserves $B$. Hence (i) holds. Conditions (ii)-(iv) can be shown similarly. Conditions (v) and (vi) are satisfied since $\sigma$ preserves $Q_\mathfrak{L}$ and $Q(e_i)\in \mathfrak{L}_h\subseteq \mathfrak{L}$ for any $i\in \Theta$.\\
($\Leftarrow$) Suppose $\sigma$ satisfies (i)-(vi). It is an easy exercise to deduce from Conditions (i)-(iv) that $B(\sigma v,\sigma w)=B(v,w)$ for any $v,w\in V$ (hint: replace $w$ by $\widetilde\sigma w'$). It remains to show that $Q(\sigma v)\overset{.}{-} Q(v)\in \mathfrak{L}$ for any $v\in V$. A straightforward computation shows that
\begin{align}
&Q(\sigma v)\overset{.}{-} Q(v)\notag\\
=&(\plus\limits_{i=1}^{-1}(Q(\sigma e_i)\overset{.}{-} Q(e_i))\leftharpoonup v_i)\+Q(\sigma v_0)\overset{.}{-} Q(v_0)\notag\\
&\+(0,\sum\limits_{i<j}\bar v_i\epsilon_i(\sigma_{-i,j}-\delta_{-i,j})v_{-i})+\overline{\sum\limits_{i<j}\bar v_i\epsilon_i(\sigma_{-i,j}-\delta_{-i,j})v_{-i}})\notag\\
&\+(0,\sum\limits_{i\in\Theta}\bar v_i\epsilon_i(\sigma^{0h} v_0)_{-i}+\overline{\sum\limits_{i\in\Theta}\bar v_i\epsilon_i(\sigma^{0h} v_0)_{-i}}).
\end{align}
It follows from (v) and (vi) that $Q(\sigma v)\overset{.}{-} Q(v)\in \mathfrak{L}$.
\end{proof}
\begin{definition}
Let $\mathfrak{H}_0$ be the Heisenberg group of the anti-Hermitian space $(V_0, B_0)$. Define the map
\begin{align*}
Q^0:V&\rightarrow \mathfrak{H}_0\\
v&\mapsto (v_0,-F_h(v_h, v_h))
\end{align*}
One checks easily that.
\begin{align*}
&Q^0(vx)=Q^0(v)\leftharpoonup x~\forall v\in V, x\in R\text{ and}
\\
&Q^0(v+w)=Q^0(v)\+Q^0(w)\overset{.}{-}(0,B(v,w)+F_h(w_h,v_h)+\overline{F_h(w_h,v_h)})~\forall v,w\in V.
\end{align*}
\end{definition}
\begin{lemma}\label{lemtec}
Let $v,w\in V$. Then $Q(v)\equiv Q(w)\bmod \mathfrak{L}~\Leftrightarrow~ Q^0(v)\equiv Q^0(w)\bmod \mathfrak{L}_0$.
\end{lemma}
\begin{proof}
Clearly
\begin{align*}
&Q(v)\overset{.}{-} Q(w)\in \mathfrak{L}\\
\Leftrightarrow~&(v-w,B(w-v,w))\in \mathfrak{L}\\
\Leftrightarrow~&(v-w,B(w-v,w))\overset{.}{-} (v_h-w_h,F_H(v_h-w_h,v_h-w_h))\in \mathfrak{L}\\
\Leftrightarrow~& Q^0(v-w)\+(0,B(w-v,w))\in \mathfrak{L}\\
\Leftrightarrow~& Q^0(v-w)\+(0,B(w-v,w))\in \mathfrak{L}_0
\end{align*}
(note that $(v_h-w_h,F_H(v_h-w_h,v_h-w_h))\in \mathfrak{L}_h\subseteq \mathfrak{L}$). One checks easily that
\begin{align*}
&Q^0(v-w)\+(0,B(w-v,w))\overset{.}{-} (Q^0(v)\overset{.}{-} Q^0(w))\\
=&(0,-F_h(w_h,w_h-v_h)-\overline{F_h(w_h,w_h-v_h)})\in(\mathfrak{L}_0)_{\min}\subseteq\mathfrak{L}_0.
\end{align*}
The assertion of the lemma follows.
\end{proof}
\begin{proposition}\label{propodd}
Let $\sigma \in \operatorname{GL}(V)$. Then $\sigma\in U_{2n}(\mathcal{V}_0)$ iff Conditions (i)-(vi) below are satisfied.
\begin{enumerate}[(i)]
\item $\widetilde\sigma_{ij}=-\epsilon_i\bar\sigma_{-j,-i}\epsilon_{-j}$ for any $i,j\in\Theta$.
\item $(\widetilde\sigma^{0h}v_0)_{i}=-\epsilon_iB_0(\sigma^{h0}e_{-i},v_0)$ for any $v_0\in V_0$ and $i\in\Theta$.
\item $B_0(v_0,\widetilde\sigma^{h0}e_i)=\overline{(\sigma v_0)_{-i}}\epsilon_{-i}$ for any $v_0\in V_0$ and $i\in\Theta$.
\item $B_0(v_0,\widetilde\sigma^{00}w_0)=B_0(\sigma^{00}v_0,w_0)$ for any $v_0,w_0\in V_0$.
\item $Q^0(\sigma e_i)\in \mathfrak{L}_0$ for any $i\in\Theta$.
\item $Q^0(\sigma v_0)\overset{.}{-} Q^0(v_0)\in \mathfrak{L}_0$ for any $v_0\in V_0$.
\end{enumerate}
\end{proposition}
\begin{proof}
Follows from the Lemmas \ref{lemodd} and \ref{lemtec}.
\end{proof}
\subsection{The elementary subgroup and relative elementary subgroups}
\begin{definition}\label{defel}
If $i,j\in \Theta$, $i\neq\pm j$ and $x\in R$, then the ESD transvection \[T_{ij}(x):=T_{e_{-j},-e_ix\epsilon_j}(0)\]
is called an {\it elementary short root transvection}. If $i\in \Theta$ and $(v_0,x)\in \mathfrak{L}_0$, then the ESD transvection \[T_{i}(v_0,x):=T_{e_{i},v\epsilon_{-i}}(\bar \epsilon_{-i}\bar 1^{-1}x\epsilon_{-i})\]
is called an {\it elementary extra short root transvection}.
If an element of $\operatorname{U}_{2n}(\mathcal{V}_0)$ is an elementary short or extra short root transvection, then it is called an {\it elementary transvection}. The subgroup of $\operatorname{U}_{2n}(\mathcal{V}_0)$ generated by all elementary transvections is called the {\it elementary subgroup} of $\operatorname{U}_{2n}(\mathcal{V}_0)$ and is denoted by $\operatorname{EU}_{2n}(\mathcal{V}_0)$.
\end{definition}
One checks easily that an elementary short root transvection $T_{ij}(x)$ maps
\begin{align*}
e_k\mapsto e_k~(k\neq -i,j), \quad e_{-i}\mapsto e_{-i}+e_{-j}\epsilon_{-j}\bar x\epsilon_i, \quad e_j\mapsto e_j+e_ix, \quad w_0\mapsto w_0~(w_0\in V_0).
\end{align*}
Hence
\begin{align*}
T_{ij}^{hh}(x)=e+e^{ij}x+e^{-j,-i}\epsilon_{-j}\bar x\epsilon_i,\quad T_{ij}^{h0}(x)=0=T_{ij}^{0h}(x),\quad T_{ij}^{00}(x)=id_{V_0}
\end{align*}
where $e$ denotes the identity matrix in $M_{2n}(R)$ and $e^{pq}$ the matrix in $M_{2n}(R)$ with a $1$ at position $(p,q)$ and zeros elsewhere.\\
One checks easily that an elementary extra short root transvection $T_{i}(v_0,x)$ maps
\begin{align*}
e_k\mapsto e_k~(k\neq -i), \quad e_{-i}\mapsto e_{-i}+e_i\epsilon_{i} x-v_0,\quad w_0\mapsto w_0-e_i\epsilon_iB_0(v_0,w_0)~(w_0\in V_0).
\end{align*}
Hence
\begin{align*}
&T_{i}^{hh}(v_0,x)=e+e^{i,-i}\epsilon_ix,\quad T_{i}^{h0}(v_0,x) w_h=-v_0w_{-i}~(w_h\in V_h),\\
&T_{i}^{0h}(v_0,x)w_0=-e_i\epsilon_i B_0(v_0,w_0)~(w_0\in V_0), \quad T_{i}^{00}(v_0,x)=id_{V_0}.
\end{align*}
\begin{lemma}\label{39}
The following relations hold for the elementary transvections.
\begin{align*}
&T_{ij}(x)=T_{-j,-i}(\epsilon_{-j}\bar x\epsilon_{i}), \tag{R0}\\
&T_{ij}(x)T_{ij}(y)=T_{ij}(x+y), \tag{R1}\\
&T_{i}(v_0,x)T_{i}(w_0,y)=T_{i}((v_0,x)\+(w_0,y)), \tag{R2}\\
&[T_{ij}(x),T_{hk}(y)]=id_V \text{ if } h\neq j,-i \text{ and } k\neq i,-j, \tag{R3}\\
&[T_{i}(v_0,x),T_{jk}(y)]=id_V \text{ if } i\neq -j,k,\tag{R4}\\
&[T_{ij}(x),T_{jk}(y)]=T_{ik}(xy) \text{ if } i\neq\pm k, \tag{R5}\\
&[T_{i}(v_0,x),T_{j}(w_0,y)]=T_{i,-j}(\epsilon_iB_0(v_0,w_0)) \text{ if } i\neq\pm j, \tag{R6}\\
&[T_{i}(v_0,x),T_{i}(w_0,y)]=T_{i}(0,B_0(v_0,w_0)+\overline{B_0(v_0,w_0)}), \tag{R7}\\
&[T_{i}(v_0,x),T_{-i,j}(y)]=T_{ij}(\epsilon_ixy)T_{-j}((v_0,-\bar x)\leftharpoonup y)\text{ and}\tag{R8}\\
&[T_{ij}(x),T_{j,-i}(y)]=T_{i}(0,-\epsilon_{-i}\bar 1xy-\overline{\epsilon_{-i}\bar 1xy}). \tag{R9}
\end{align*}
\end{lemma}
\begin{proof}
Straightforward computation.
\end{proof}
\begin{definition}\label{42}
If $i,j\in\Theta$ such that $i\neq\pm j$, then the element
\begin{align*}
P_{ij}:=T_{ij}(1)T_{ji}(-1)T_{ij}(1)
\end{align*}
of $\operatorname{EU}_{2n}(\mathcal{V}_0)$ is called an {\it elementary permutation}.
One checks easily that
\begin{align*}
&P^{hh}_{ij}=e-e^{ii}-e^{jj}-e^{-i,-i}-e^{-j,-j}+e^{ij}-e^{ji}+e^{-i,-j}\epsilon_{-i}\epsilon_{-j}^{-1}-e^{-j,-i}\epsilon_{-j}\epsilon_{-i}^{-1},\\
&P^{h0}_{ij}=0=P^{0h}_{ij},\quad P_{ij}^{00}=id_{V_0}.
\end{align*}
It is easy to show that $\widetilde P_{ij}=P_{ji}$.
\end{definition}
\begin{lemma}\label{43}
Let $i,j,k\in\Theta$ such that $i\neq \pm j$ and $k\neq \pm i,\pm j$. Let $x\in R$ and $(v_0,x)\in \mathfrak{L}_0$. Then
\begin{enumerate}[(i)]
\item $^{P_{ki}}T_{ij}(x)=T_{kj}(x)$,
\item $^{P_{kj}}T_{ij}(x)=T_{ik}(x)$ and
\item $^{P_{-k,-i}}T_{i}(v_0,x)=T_{k}(v_0,x)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Follows from the relations in Lemma \ref{39}.
\end{proof}
\begin{definition}
Let $\mathfrak{I}_0=(I,\mathfrak{M}_0)$ denote an odd form ideal on $\mathcal{V}_0$. An elementary short root transvection $T_{ij}(x)$ is called {\it $\mathfrak{I}_0$-elementary} if $x\in I$. An elementary extra short root transvection $T_{i}(\xi)$ is called {\it $\mathfrak{I}_0$-elementary} if $\xi\in \mathfrak{M}_0$. If an element of $U_{2n}(\mathcal{V}_0)$ is an $\mathfrak{I}_0$-elementary short or extra short root transvection, then it is called an {\it $\mathfrak{I}_0$-elementary transvection}. The subgroup $\operatorname{EU}_{2n}(\mathfrak{I}_0)$ of $\operatorname{EU}_{2n}(\mathcal{V}_0)$ generated by the $\mathfrak{I}_0$-elementary transvections is called the {\it preelementary subgroup of level $\mathfrak{I}_0$}. Its normal closure $\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ in $\operatorname{EU}_{2n}(\mathcal{V}_0)$ is called the {\it elementary subgroup of level $\mathfrak{I}_0$}.
\end{definition}
\subsection{Congruence subgroups}
If $\mathfrak{I}_0=(I,\mathfrak{M}_0)$ is an odd form ideal on $\mathcal{V}_0$, then $\mathfrak{I}:=(I,\mathfrak{M})$ where $\mathfrak{M}=\mathfrak{L}_h\leftharpoonup I\+\mathfrak{M}_0$ is an odd form ideal on $\mathcal{V}$. Conversely, if $\mathfrak{I}=(I,\mathfrak{M})$ is an odd form ideal on $\mathcal{V}$, then $\mathfrak{I}_0:=(I,\mathfrak{M}_0)$ where $\mathfrak{M}_0=\{Q^0(v)\+(0,x)\mid (v,x)\in M\}$ is an odd form ideal on $\mathcal{V}_0$. This yields a 1-1 corespondence between odd form ideals on $\mathcal{V}_0$ and odd form ideals on $\mathcal{V}$.\\
In this subsection $\mathfrak{I}_0=(I,\mathfrak{M}_0)$ denotes an form ideal on $\mathcal{V}_0$ and $\mathfrak{I}=(I,\mathfrak{M})$ the corresponding odd form ideal on $\mathcal{V}$. We denote the principal congruence subgroup $\operatorname{U}(\mathcal{V},\mathfrak{I})$ by $\operatorname{U}_{2n}(\mathcal{V}_0, \mathfrak{I}_0)$ and the normalized principal congruence subgroup $\operatorname{NU}(\mathcal{V},\mathfrak{I})$ by $\operatorname{NU}_{2n}(\mathcal{V}_0, \mathfrak{I}_0)$. The even part of $V_0$ is denoted by $V^{ev}_0$.
\begin{lemma}\label{lemoddrel}
Let $\sigma \in \operatorname{U}_{2n}(\mathcal{V}_0)$. Then $\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ iff Conditions (i) and (ii) below are satisfied.
\begin{enumerate}[(i)]
\item $Q(\sigma e_i)\overset{.}{-} Q(e_i)\in \mathfrak{M}$ for any $i\in\Theta$.
\item $Q(\sigma v_0)\overset{.}{-} Q(v_0)\in \mathfrak{M}$ for any $v_0\in V^{ev}_0$.
\end{enumerate}
\end{lemma}
\begin{proof}
($\Rightarrow$) Suppose that $\sigma \in U_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$. Then clearly $Q(\sigma e_i)\overset{.}{-} Q(e_i)\in \mathfrak{M}$ for any $i\in\Theta$ and $Q(\sigma v_0)\overset{.}{-} Q(v_0)\in \mathfrak{M}$ for any $v_0\in V^{ev}_0$ since $\sigma$ preserves $Q_\mathfrak{M}$. \\
($\Leftarrow$) Suppose that Conditions (i) and (ii) are satisfied. If $i\in \Theta$, then $Q(\sigma e_i)\overset{.}{-} Q(e_i)=(\sigma e_i-e_i,x)$ for some $x\in R$. Since $Q(\sigma e_i)\overset{.}{-} Q(e_i)\in \mathfrak{M}\subseteq\mathfrak{M}_{\max}$, it follows that $B(e_j,\sigma e_i-e_i)\in I$ for any $j\in \Theta$. But one checks easily that $B(e_j,\sigma e_i-e_i)=-\epsilon_{-j}^{-1}(\sigma_{-j,i}-\delta_{-j,i})$. Thus $\sigma^{hh}\equiv e\bmod I$. If $v_0\in V^{ev}_0$, then $Q(\sigma v_0)\overset{.}{-} Q(v_0)=(\sigma v_0-v_0,x)$ for some $x\in R$. Since $Q(\sigma v_0)\overset{.}{-} Q(v_0)\in \mathfrak{M}\subseteq\mathfrak{M}_{\max}$, it follows that $B(e_j,\sigma v_0-v_0)\in I$ for any $j\in \Theta$. But one checks easily that $B(e_j,\sigma v_0-v_0)=-\epsilon_{-j}^{-1}(\sigma^{0h}v_0)_{-j}$ and therefore $(\sigma^{0h}v_0)_{-j}\in I$. Now it follows from Equation (5) in the proof of Lemma \ref{lemodd} that $Q(\sigma v)\overset{.}{-} Q(v)\in \mathfrak{M}$ for any $v\in V^{ev}$. Thus $\sigma\in U_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$.
\end{proof}
\begin{lemma}\label{lemtecrel}
Let $v,w\in V$ such that $v_h-w_h\in V_hI$ (i.e. $v_i-w_i\in I$ for any $i\in \Theta$). Then
\[Q(v)\equiv Q(w)\bmod \mathfrak{M}~\Leftrightarrow~Q^0(v)\equiv Q^0(w)\bmod \mathfrak{M}_0.\]
\end{lemma}
\begin{proof}
See the proof of Lemma \ref{lemtec}.
\end{proof}
\begin{proposition}\label{propoddrel}
Let $\sigma \in \operatorname{U}_{2n}(\mathcal{V}_0)$. Then $\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ iff the Conditions (i)-(iii) below are satisfied.
\begin{enumerate}[(i)]
\item $\sigma^{hh}\equiv e\bmod I$.
\item $Q^0(\sigma e_i)\in \mathfrak{M}_0$ for any $i\in\Theta$.
\item $Q^0(\sigma v_0)\overset{.}{-} Q^0(v_0)\in \mathfrak{M}_0$ for any $v_0\in V^{ev}_0$.
\end{enumerate}
\end{proposition}
\begin{proof}
Follows from the Lemmas \ref{lemoddrel} and \ref{lemtecrel}.
\end{proof}
\begin{corollary}\label{cor1}
$\operatorname{EU}_{2n}(\mathfrak{I}_0)\subseteq \operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$
\end{corollary}
\begin{proof}
One checks easily that any $\mathfrak{I}_0$-elementary transvection satisfies Conditions (i)-(iii) in Proposition \ref{propoddrel}.
\end{proof}
\begin{corollary}\label{corobv}
Let $\mathfrak{N}_0\subseteq \mathfrak{L}_0$ be another relative odd form parameter of level $I$. Then
\[\operatorname{U}_{2n}(\mathcal{V}_0,(I,\mathfrak{M}_0))=\operatorname{U}_{2n}(\mathcal{V}_0,(I,\mathfrak{N}_0))~\Leftrightarrow~\mathfrak{M}_0=\mathfrak{N}_0.\]
\end{corollary}
\begin{proof}
Suppose that $\operatorname{U}_{2n}(\mathcal{V}_0,(I,\mathfrak{M}_0))=\operatorname{U}_{2n}(\mathcal{V}_0,(I,\mathfrak{N}_0))$. Let $(v_0,x)\in \mathfrak{M}_0$. Then $T_{-1}(v_0,x)\in \operatorname{U}_{2n}(\mathcal{V}_0,$ $(I,\mathfrak{M}_0))=\operatorname{U}_{2n}(\mathcal{V}_0,(I,\mathfrak{N}_0))$ by the previous corollary. It follows from Proposition \ref{propoddrel} that $(-v_0,x)=Q^0(T_{-1}(v_0,x)e_1)\in \mathfrak{N}_0$. Hence $(v_0,x)=(-v_0,x)\leftharpoonup(-1)\in\mathfrak{N}_0$ and therefore we have shown $\mathfrak{M}_0\subseteq \mathfrak{N}_0$. Analogously one can show that $\mathfrak{N}_0\subseteq \mathfrak{M}_0$.
\end{proof}
By Definition \ref{defaction} we have a left action of $\operatorname{U}_{2n}(\mathcal{V}_0)$ on the set of all form ideals on $\mathcal{V}$. This action induces a left action of $\operatorname{U}_{2n}(\mathcal{V}_0)$ on the set of all form ideals on $\mathcal{V}_0$, given by $(\sigma,\mathfrak{I}_0)\mapsto {}^{\sigma}\mathfrak{I}_0:=(I,{}^{\sigma}\mathfrak{M}_0)$ where \[{}^{\sigma}\mathfrak{M}_0=\{Q^0(\sigma v_0)\+(0,x)\mid (v_0,x)\in \mathfrak{M}_0\}.\]
\begin{proposition}\label{proptildhyp}
$\operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)=\{\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0)\mid {}^{\sigma}\mathfrak{I}_0=\mathfrak{I}_0\}$.
\end{proposition}
\begin{proof}
Follows from Proposition \ref{proptild} and Corollary \ref{corobv}.
\end{proof}
\begin{remark}
On page 4760 in \cite{petrov}, Petrov defined the set $\tilde \operatorname{U}(V,q):=\{g\in \operatorname{U}(V,q)\mid (g w,a)\in \mathfrak{I}~\forall (w,a)\in\mathfrak{I} \}$ and claimed that it is a group. But is is not clear that $\tilde \operatorname{U}(V,q)$ is closed under taking inverses. In an e-mail to the author, Petrov wrote that he ``probably missed the condition that every $g$ from $\tilde \operatorname{U}(V,q)$ must preserve $U$ (that is $gU=U$)" where $U$ is the submodule of $V$ generated by $VI$ and all elements $v\in V$ such that $(v, x)\in \mathfrak{M}$ for some $x\in I$. After making this correction, $\tilde \operatorname{U}(V,q)$ conincides with the normalized principal subgroup $\operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ (follows from Proposition \ref{proptildhyp} above).
\end{remark}
\begin{corollary}\label{cor2}
$\operatorname{EU}_{2n}(\mathcal{V}_0)\subseteq \operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$.
\end{corollary}
\begin{proof}
Let $\sigma$ be an elementary transvection. Then $Q^0(\sigma v_0)=Q^0(v_0)$ for any $v_0\in V_0$ (see the two paragraphs after Definition \ref{defel}). Hence ${}^{\sigma}\mathfrak{I}_0=\mathfrak{I}_0$.
\end{proof}
\begin{corollary}
$\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)\subseteq \operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$.
\end{corollary}
\begin{proof}
Follows from Corollaries \ref{cor1} and \ref{cor2}.
\end{proof}
\begin{definition}
The subgroup
\[\operatorname{CU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0):=\{\sigma\in \operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)\mid [\sigma,\operatorname{EU}_{2n}(\mathcal{V}_0)]\subseteq \operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)\}\]
of $\operatorname{U}_{2n}(V_0)$ is called the {\it full congruence subgroup of level $\mathfrak{I}_0$}.
\end{definition}
\subsection{The standard commutator formula}
In this subsection we assume that $R$ is commutative and $n\geq 3$. Further $\mathfrak{I}_0$ denotes an odd form ideal on $\mathcal{V}_0$.
\begin{theorem}[V. A. Petrov, 2005]\label{thmscf}
$\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ is a normal subgroup of $\operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ and the standard commutator formula
\begin{align*}
[\operatorname{CU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0),\operatorname{EU}_{2n}(\mathcal{V}_0)]
=[\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0),\operatorname{EU}_{2n}(\mathcal{V}_0)]
=\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)
\end{align*}
holds. In particular from the absolute case $\mathfrak{I}_0=(R,\mathfrak{L}_0)$ it follows that $\operatorname{EU}_{2n}(\mathcal{V}_0)$ is perfect and normal in $\operatorname{U}_{2n}(\mathcal{V}_0)$
\end{theorem}
\begin{proof}
Follows from \cite[Proposition 4 and Corollary on page 4765]{petrov} and \cite[Lemma 1]{stepanov} (note that in \cite{petrov} the full congruence subgroup is defined a little differently).
\end{proof}
\begin{corollary}\label{165}
If $\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0)$, then
\[^{\sigma}\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)=\operatorname{EU}_{2n}(\mathcal{V}_0,{}^{\sigma}\mathfrak{I}_0)\]
and
\[^{\sigma}\operatorname{CU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)=\operatorname{CU}_{2n}(\mathcal{V}_0,{}^{\sigma}\mathfrak{I}_0).\]
Further $\operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ is the normalizer of $\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ and of $\operatorname{CU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$.
\end{corollary}
\begin{proof}
Follows from Proposition \ref{proptild} and the standard commutator formula in Theorem \ref{thmscf}.
\end{proof}
\begin{definition}\label{deflevel}
A subgroup $H$ of $\operatorname{U}_{2n}(\mathcal{V}_0)$ is called {\it E-normal} iff it is normalized by the elementary subgroup $\operatorname{EU}_{2n}(\mathcal{V}_0)$. Suppose $H$ is an E-normal subgroup of $\operatorname{U}_{2n}(\mathcal{V}_0)$. Set
\[I:=\{x\in R\mid T_{ij}(x)\in H \text{ for some }i,j\in\Theta\}\]
and
\[\mathfrak{M}_0:=\{(v_0,y)\in \mathfrak{L}_0\mid T_{i}(v_0,y)\in H \text{ for some }i\in\Theta\}.\]
Then $\mathfrak{I}_0:=(I,\mathfrak{M}_0)$ is an odd form ideal on $\mathcal{V}_0$ such that $\operatorname{EU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)\subseteq H$. It is called the {\it level of $H$} and $H$ is called an {\it E-normal subgroup of level $\mathfrak{I}_0$}.
\end{definition}
\begin{corollary}\label{cornu}
Let $H$ be an E-normal subgroup of level $\mathfrak{I}_0$. Then for any $\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0)$, $^{\sigma}H$ is an E-normal subgroup of level ${}^{\sigma}\mathfrak{I}_0$. It follows that $H\subseteq \operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$.
\end{corollary}
\begin{proof}
Follows from the previous corollary and Proposition \ref{proptildhyp}.
\end{proof}
\subsection{The case $\mathfrak{M}_0=(\mathfrak{M}_0)_{\max}$}
In this subsection we assume that $R$ is commutative and $n\geq 3$. Further $(\mathfrak{I}_0)_{\max}=(I,(\mathfrak{M}_0)_{\max})$ denotes an odd form ideal on $\mathcal{V}_0$ with maximal relative odd form parameter. Proposition \ref{proptildhyp} implies that $\operatorname{NU}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})=\operatorname{U}_{2n}(\mathcal{V}_0)$, i.e. $\operatorname{U}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$ is normal in $\operatorname{U}_{2n}(\mathcal{V}_0)$.
\begin{lemma}\label{lemumax}
Let $\sigma \in \operatorname{U}_{2n}(\mathcal{V}_0)$. Then $\sigma\in \operatorname{U}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$ iff Conditions (i)-(iv) below are satisfied.
\begin{enumerate}[(i)]
\item $\sigma^{hh}\equiv e\bmod I$.
\item $(\sigma^{0h}v_0)_i \in I$ for any $v_0\in V^{ev}_0$ and $i\in\Theta$.
\item $B_0(\sigma^{h0}e_i,v_0)\in I$ for any $v_0\in V^{ev}_0$ and $i\in\Theta$.
\item $B_0((\sigma^{00}-id_{V_0})v_0,w_0)\in I$ for any $v_0,w_0\in V^{ev}_0$.
\end{enumerate}
\end{lemma}
\begin{proof}
Follows from Proposition \ref{propoddrel}.
\end{proof}
We leave it to the reader to deduce the next lemma from the last one.
\begin{lemma}\label{lemgen}
Let $\sigma, \tau\in \operatorname{U}_{2n}(\mathcal{V}_0)$. Then $\sigma\equiv \tau \bmod \operatorname{U}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$ iff Conditions (i)-(iv) below are satisfied.
\begin{enumerate}[(i)]
\item $\sigma^{hh}\equiv \tau^{hh}\bmod I$.
\item $((\sigma^{0h}-\tau^{0h})v_0)_i \in I$ for any $v_0\in V^{ev}_0$ and $i\in\Theta$.
\item $B_0((\sigma^{h0}-\tau^{h0})e_i,v_0)\in I$ for any $v_0\in V^{ev}_0$ and $i\in\Theta$.
\item $B_0((\sigma^{00}-\tau^{00})v_0,w_0)\in I$ for any $v_0,w_0\in V^{ev}_0$.
\end{enumerate}
\end{lemma}
\begin{proposition}\label{propcumax}
Let $\sigma\in\operatorname{U}_{2n}(\mathcal{V}_0)$. Then $\sigma\in \operatorname{CU}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$ iff Conditions (i)-(v) below are satisfied.
\begin{enumerate}[(i)]
\item $\sigma_{ij}\in I$ for any $i,j\in \Theta$ such that $i\neq j$.
\item $\sigma_{ii}-\sigma_{jj}\in I$ for any $i,j\in \Theta$ such that $i\neq j$.
\item $(\sigma^{0h}v_0)_i \in I$ for any $v_0\in V^{ev}_0$ and $i\in\Theta$.
\item $B_0(\sigma^{h0}e_i,v_0)\in I$ for any $v_0\in V^{ev}_0$ and $i\in\Theta$.
\item $B_0((\sigma^{00}-id_{V_0}\sigma_{ii})v_0,w_0)\in I$ for any $v_0,w_0\in V^{ev}_0$ and $i\in \Theta$.
\end{enumerate}
\end{proposition}
\begin{proof}
($\Rightarrow$) Suppose that $\sigma\in \operatorname{CU}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$. Let $\tau\in \operatorname{EU}_{2n}(\mathcal{V}_0)$. Then $[\sigma,\tau]\in \operatorname{U}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$ by the definition of $\operatorname{CU}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$. Hence $\sigma\tau\equiv \tau\sigma \bmod \operatorname{U}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$. By letting $\tau$ vary over all elementary transvections $T_{ij}(1)~(i,j\in\Theta,i\neq\pm j)$ and $T_{i}(v_0,y)~(i\in\Theta,(v_0,y)\in\mathfrak{L}_0)$ and applying Lemma \ref{lemgen} one gets Conditions (i)-(v).\\
($\Leftarrow$) Suppose that Conditions (i)-(v) are satisfied. Then it follows from Lemma \ref{lemgen} that $[\sigma,\tau]\in \operatorname{U}_{2n}(\mathcal{V}_0,$ $(\mathfrak{I}_0)_{\max})$ for any elementary transvection $\tau$. Since $\operatorname{U}_{2n}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$ is normal, it follows that $\sigma \in \operatorname{CU}_{2n}(\mathcal{V}_0,$ $(\mathfrak{I}_0)_{\max})$.
\end{proof}
The following lemma will be used in the proof of the Sandwich Classification Theorem \ref{thmsct}.
\begin{lemma}\label{lemcu}
Let $\mathfrak{I}_0=(I,\mathfrak{M}_0)$ be an odd form ideal. Let $\sigma\in \operatorname{NU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)\cap \operatorname{CU}_{2n}(\mathcal{V}_0, (\mathfrak{I}_0)_{\max})$ such that $Q^0(\sigma e_i)\in\mathfrak{M}_0$ for any $i\in\Theta$ and further $Q^0([\sigma,\tau] e_i)\in\mathfrak{M}_0$ for any elementary transvection $\tau$ and $i\in\Theta$. Then $\sigma \in\operatorname{CU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$.
\end{lemma}
\begin{proof}
Let $\tau$ be an elementary transvection and set $\xi:=[\sigma,\tau]$. In order to prove that $\sigma \in\operatorname{CU}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ it suffices to show that $\xi\in \operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ (note that $\operatorname{U}_{2n}(\mathcal{V}_0,\mathfrak{I}_0)$ is E-normal by Corollary \ref{cor2}). In view of Proposition \ref{propoddrel} it suffices to show that $Q^0(\xi v_0)\overset{.}{-} Q^0(v_0)\in \mathfrak{M}_0$ for any $v_0\in V^{ev}_0$.\\
\\
\underline{Case 1} Suppose that $\tau=T_{ij}(x)$. A straightforward computation shows that
\[Q^0(\xi v_0)\overset{.}{-} Q^0(v_0)=Q^0(\sigma e_i)\leftharpoonup a\+Q^0(\sigma e_{-j})\leftharpoonup b\+(0,c\+\bar c)\]
for some $a,b,c\in I$. Hence $Q^0(\xi v_0)\overset{.}{-} Q^0(v_0)\in\mathfrak{M}_0$.\\
\\
\underline{Case 2} Suppose that $\tau=T_{i}(w_0,x)$. A straightforward computation shows that
\[Q^0(\xi v_0)\overset{.}{-} Q^0(v_0)=Q^0(\sigma e_i)\leftharpoonup a\+(Q^0(\sigma w_0)\overset{.}{-} Q^0(w_0))\leftharpoonup b\+(w_0,x)\leftharpoonup b\+(0,c\+\bar c)\]
for some $a\in R$ and $b,c\in I$. Hence $Q^0(\xi v_0)\overset{.}{-} Q^0(v_0)\in\mathfrak{M}_0$.\\
\end{proof}
\section{Sandwich classification of E-normal subgroups}
In this section $R$ denotes a commutative ring with pseudoinvolution $~\bar{}~$. Let $\mathcal{V}$ denote an odd quadratic space over $(R,~\bar{}~)$ such that $\operatorname{ind} \mathcal{V}\geq 3$. As mentioned in Definition \ref{defwit}, $\operatorname{U}(\mathcal{V})$ is isomorphic to an odd hyperbolic unitary group $\operatorname{U}_{6}(\mathcal{V}_0)$ where $\mathcal{V}_0$ is some odd quadratic space. The main goal of this section is to classify the E-normal subgroups of $\operatorname{U}_{6}(\mathcal{V}_0)$. We start with some easy-to-check lemmas which will be used in the proof of Theorem \ref{thmm}.
\begin{lemma}\label{pre}
Let $G$ be a group and $a,b,c\in G$. Then ${}^{b^{-1}}[a,bc]=[b^{-1},a][a,c]$.
\end{lemma}
The following lemma will be used in the proof of Theorem \ref{thmm} without explicit reference.
\begin{lemma}\label{pre2}
Let $G$ be a group, $E$ a subgroup and $a\in G$. Suppose that $b\in G$ is a product of $n$ elements of the form $^{\epsilon}a^{\pm 1}$ where $\epsilon\in E$. Then for any $\epsilon'\in E$
\begin{enumerate}[(i)]
\item $^{\epsilon'}b$ is a product of $n$ elements of the form $^{\epsilon}a^{\pm 1}$ and
\item $[\epsilon',b]$ is a product of $2n$ elements of the form $^{\epsilon}a^{\pm 1}$.
\end{enumerate}
\end{lemma}
\begin{lemma}\label{lemesdspec}
Let $v\in V$ be an isotropic element such that $v_{-1}=0$. Then
\[T_{e_1v}(0)=T_1(Q^0(v)\leftharpoonup(-1)\+(0,\bar 1 v_1+\overline{\bar 1 v_1}))\prod\limits_{i=2}^{-2}T_{i,-1}(v_i)\in \operatorname{EU}_{6}(\mathcal{V}_0).\]
\end{lemma}
\begin{lemma}\label{last}
Let $(v_0,y)\in \mathfrak{L}_0$ and $x_1,\dots,x_n\in R$. Then
\[(v_0,y)\leftharpoonup \sum\limits_{i=1}^nx_i=(\plus\limits_{i=1}^n (v_0,y)\leftharpoonup x_i)\+(0,\sum\limits_{i>j} \bar x_i\bar 1^{-1}yx_j+\overline{\bar x_i\bar 1^{-1}yx_j}).\]
\end{lemma}
\begin{lemma}\label{new}
Let $\sigma\in U_{2n}(\mathcal{V}_0)$ and $i,j\in \Theta$ such that $i\neq \pm j$. Then
\[Q^0(^{P_{ij}}\!\sigma e_i)=
\begin{cases}
Q^0(\sigma e_j), \text{ if } i,j\in \Theta_+ \text{ or }i,j\in \Theta_-,\\
Q^0(\sigma e_j)\+(0,-(\bar{\sigma}_{ij}\sigma_{-i,j}+\bar{\sigma}_{-j,j}\sigma_{jj}+\overline{\bar{\sigma}_{ij}\sigma_{-i,j}+\bar{\sigma}_{-j,j}\sigma_{jj}}~)), \text{ if } i\in \Theta_+,j\in\Theta_-,\\
Q^0(\sigma e_j)\+(0,-(\bar{\sigma}_{-i,j}\sigma_{ij}+\bar{\sigma}_{jj}\sigma_{-j,j}+\overline{\bar{\sigma}_{-i,j}\sigma_{ij}+\bar{\sigma}_{jj}\sigma_{-j,j}}~)), \text{ if } i\in \Theta_-,j\in\Theta_+.
\end{cases}
\]
\end{lemma}
\begin{definition}
Let $\sigma\in \operatorname{U}_{6}(\mathcal{V}_0)$. Then an element $\zeta\in \operatorname{U}_{6}(\mathcal{V}_0)$ is called an {\it elementary $\sigma$-conjugate} if $\zeta={}^{\tau}\sigma$ or $\zeta={}^{\tau}\widetilde\sigma$ for some $\tau\in \operatorname{EU}_{6}(\mathcal{V}_0)$.
\end{definition}
\begin{theorem}\label{thmm}
Let $\sigma\in \operatorname{U}_{6}(\mathcal{V}_0)$ and $i,j,k,l\in\Theta$ such that $k\neq\pm l$ and $i\neq \pm j$. Further let $v_0\in V_0^{ev}$. Then
\begin{enumerate}[(i)]
\item $T_{kl}(\sigma_{ij})$ is a product of $160$ elementary $\sigma$-conjugates,
\item $T_{kl}(\sigma_{i,-i})$ is a product of $320$ elementary $\sigma$-conjugates,
\item $T_{kl}((\sigma^{0h}v_0)_i)$ is a product of $480$ elementary $\sigma$-conjugates,
\item $T_{kl}(B_0(v_0,\sigma^{h0}e_j))$ is a product of $480$ elementary $\sigma$-conjugates,
\item $T_{kl}(\sigma_{ii}-\sigma_{jj})$ is a product of $480$ elementary $\sigma$-conjugates,
\item $T_{kl}(\sigma_{ii}-\sigma_{-i,-i})$ is a product of $960$ elementary $\sigma$-conjugates,
\item $T_{k}(Q^0(\sigma e_j))$ is a product of $10564$ elementary $\sigma$-conjugates and
\item $T_{k}((\sigma_{00}-id_{V_0}\sigma_{jj})v_0,x)$ is a product of $34092$ elementary $\sigma$-conjugates for some $x\in R$.
\end{enumerate}
\end{theorem}
\begin{proof}(i) Let $x\in R$. In Step 1 we show that $T_{kl}(x\bar\sigma_{23}\sigma_{21})$ is a product of $16$ elementary $\sigma$-conjugates. In Step 2 we show that $T_{kl}(x\bar\sigma_{23}\sigma_{2,-1})$ is a product of $16$ elementary $\sigma$-conjugates. In Step 3 we show that $T_{kl}(x\bar\sigma_{23}\sigma_{22})$ is a product of $32$ elementary $\sigma$-conjugates. In Step 4 we use Steps 1-3 in order to prove (i).\\
\\
{\bf Step 1.} Set $\tau:=T_{1,-2}(\bar 1^{-1}\bar\sigma_{23}\sigma_{23})T_{3,-2}(-\bar 1^{-1}\bar\sigma_{23}\sigma_{21})T_{3,-1}(\bar 1^{-2}\bar\sigma_{23}\sigma_{22})T_{3}(0,\bar\sigma_{22}\sigma_{21}+\overline{\bar\sigma_{22}\sigma_{21}})$ and $\xi:={}^{\sigma}\widetilde\tau$. One checks easily that $(\sigma\widetilde\tau v)_2=(\sigma v)_2$ for any $v\in V$. Further $\widetilde\tau\widetilde\sigma e_{-2}=\widetilde\sigma e_{-2}$. It follows that $(\xi v)_2=v_2$ for any $v\in V$ and further $\xi e_{-2}=e_{-2}$ (note that this implies that $(\widetilde\xi v)_2=v_2$ for any $v\in V$ and further $\widetilde\xi e_{-2}=e_{-2}$). Set
\begin{align*}
\zeta:={}^{\widetilde\tau}[T_{-2,-1}(1),[\tau,\sigma]]={}^{\widetilde\tau}[T_{-2,-1}(1),\tau\xi]=[\widetilde\tau,T_{-2,-1}(1)][T_{-2,-1}(1),\xi],
\end{align*}
the last equality by Lemma \ref{pre}. It follows from the relations in Lemma \ref{39} that
\[[\widetilde\tau,T_{-2,-1}(1)]=T_{3,-1}(\bar 1^{-1}\bar\sigma_{23}\sigma_{21})T_{1}(\alpha)\]
for some $\alpha\in \mathfrak{L}_0$. On the other hand one checks easily that
\[[T_{-2,-1}(1),\xi]=T_{-2}(\beta)T_{12}(\xi_{11}-1)T_{32}(\xi_{31})T_{-3,2}(\xi_{-3,1})T_{-1,2}(\xi_{-1,1})\]
where $\beta=Q^0(\xi e_1)\leftharpoonup (-1)\overset{.}{-} (0,\xi_{-2,1}+\bar\xi_{-2,1})$. Hence \[\zeta=T_{3,-1}(\bar 1^{-1}\bar\sigma_{23}\sigma_{21})T_{1}(\alpha)T_{-2}(\beta)T_{12}(\xi_{11}-1)T_{32}(\xi_{31})T_{-3,2}(\xi_{-3,1})T_{-1,2}(\xi_{-1,1}).\]
It follows from the relations in Lemma \ref{39} that $[T_{-1,3}(-\bar 1 x),[T_{-2,3}(1),\zeta]]=T_{-2,3}(x\bar\sigma_{23}\sigma_{21})$ for any $x\in R$. Hence we have shown
\[[T_{-1,3}(-\bar 1x),[T_{-2,3}(1),{}^{\widetilde\tau}[T_{-2,-1}(1),[\tau,\sigma]]]]=T_{-2,3}(x\bar\sigma_{23}\sigma_{21}).\]
This implies that $T_{-2,3}(x\bar\sigma_{23}\sigma_{21})$ is a product of $16$ elementary $\sigma$-conjugates. It follows from Lemma \ref{43} that $T_{kl}(x\bar\sigma_{23}\sigma_{21})$ is a product of $16$ elementary $\sigma$-conjugates.\\
\\
{\bf Step 2.} Step 2 can be done similarly. Namely one can show that
\[[T_{-1,3}(-\bar 1 x),[T_{12}(1),{}^{\widetilde\tau}[T_{-2,1}(1),[\tau,\sigma]]]]=T_{-1,2}(x\bar\sigma_{23}\sigma_{2,-1})\]
for any $x\in R$ where $\tau=T_{21}(\bar 1^{-1}\bar\sigma_{23}\sigma_{23})T_{31}(-\bar 1^{-1}\bar\sigma_{23}\sigma_{22})T_{3,-2}(\bar 1^{-1}\bar\sigma_{23}\sigma_{2,-1})T_{3}(0,-\bar\sigma_{22}\sigma_{2,-1}-\overline{\bar\sigma_{22}\sigma_{2,-1}})$.\\
\\
{\bf Step 3.} Set $\tau:=T_{21}(-\bar 1^{-1}\bar\sigma_{22}\sigma_{23})T_{31}(\bar 1^{-1}\bar\sigma_{22}\sigma_{22})T_{2,-3}(\bar 1^{-1}\bar\sigma_{22}\sigma_{2,-1})T_{2}(0,-\bar\sigma_{23}\sigma_{2,-1}-\overline{\bar\sigma_{23}\sigma_{2,-1}})$ and $\xi:={}^{\sigma}\widetilde\tau$. One checks easily that $(\sigma\widetilde\tau v)_2=(\sigma v)_2$ for any $v\in V$. Further $\widetilde\tau\widetilde\sigma e_{-2}=\widetilde\sigma e_{-2}$. It follows that $(\xi v)_2=v_2$ for any $v\in V$ and further $\xi e_{-2}=e_{-2}$. Set
\begin{align*}
\zeta:={}^{\widetilde\tau}[T_{32}(1),[\tau,\sigma]]={}^{\widetilde\tau}[T_{32}(1),\tau\xi]=[\widetilde\tau,T_{32}(1)][T_{32}(1),\xi].
\end{align*}
It follows from the relations in Lemma \ref{39} that $\eta:=[\widetilde\tau,T_{32}(1)]=T_{31}(-\bar 1^{-1}\bar\sigma_{22}\sigma_{23})T_{3}(\alpha)T_{3,-2}(a)$ for some $\alpha\in\mathfrak{L}_0$ and $a\in R$. On the other hand one checks easily that $\theta:=[T_{32}(1),\xi]=T_{-2}(\beta)\prod\limits_{i\neq \pm 2}T_{i2}(x_i)$ for some $\beta\in\mathfrak{L}_0$ and $x_1, x_3,x_{-3}, x_{-1}\in R$. Set
\[\chi:={}^{\widetilde\eta }[T_{12}(1),\zeta]={}^{\widetilde\eta }[T_{12}(1),\eta\theta]=[\widetilde\eta ,T_{12}(1)][T_{12}(1),\theta].\]
It follows from the relations in Lemma \ref{39} that $[\widetilde\eta ,T_{12}(1)]=T_{32}(\bar 1^{-1}\bar\sigma_{22}\sigma_{23})T_{3}(\gamma)T_{3,-1}(b)$ for some $\gamma\in\mathfrak{L}_0$ and $b\in R$ and $[T_{12}(1),\theta]=T_{-2}(\delta)$ for some $\delta\in \mathfrak{L}_0$. Hence $\chi=T_{32}(\bar 1^{-1}\bar\sigma_{22}\sigma_{23})T_{3}(\gamma)T_{3,-1}(b)T_{-2}(\delta)$. Therefore
\[[T_{-2,3}(\overline{(\bar 1)^3}\bar x),[T_{2,-1}(1),\chi]]=T_{-2,-1}(-\overline{(\bar 1)^3}\bar 1^{-1}\bar x\bar\sigma_{22}\sigma_{23})=T_{12}(x\bar\sigma_{23}\sigma_{22})\]
for any $x\in R$, again by the relations in Lemma \ref{39}. Hence we have shown
\[[T_{-2,3}(\overline{(\bar 1)^3}\bar x),[T_{2,-1}(1),{}^{\widetilde\eta }[T_{12}(1),{}^{\widetilde\tau}[T_{32}(1),[\tau,\sigma]]]]]=T_{12}(x\bar\sigma_{23}\sigma_{22}).\]
This implies that $T_{12}(x\bar\sigma_{23}\sigma_{22})$ is a product of $32$ elementary $\sigma$-conjugates. It follows from Lemma \ref{43} that $T_{kl}(x\bar\sigma_{23}\sigma_{22})$ is a product of $32$ elementary $\sigma$-conjugates.\\
\\
{\bf Step 4.} Set $I:=I(\bar\sigma_{23}\sigma_{21},\overline{\bar\sigma_{23}\sigma_{2,-1}})$, $J:=I(\bar\sigma_{23}\sigma_{21},\overline{\bar\sigma_{23}\sigma_{2,-1}},\bar\sigma_{23}\sigma_{22})$ and
$\tau:=[\widetilde\sigma ,T_{12}(-\bar\sigma_{23})]$. One checks easily that
\begin{align*}
\tau_{11}&=1-\widetilde\sigma_{11}\bar\sigma_{23}\sigma_{21}+\bar 1^{-1}\widetilde\sigma_{1,-2}\sigma_{23}\sigma_{-1,1}\\
&=1-\widetilde\sigma_{11}\underbrace{\bar\sigma_{23}\sigma_{21}}_{\in I}-\bar 1^{-3}\underbrace{\bar\sigma_{2,-1}\sigma_{23}}_{\in I}\sigma_{-1,1},
\end{align*}
the last equality by Proposition \ref{propodd}, and
\begin{align*}
\tau_{12}&=-\widetilde\sigma_{11}\bar\sigma_{23}\sigma_{22}+\bar 1^{-1}\widetilde\sigma_{1,-2}\sigma_{23}\sigma_{-1,2}+\tau_{11}\bar\sigma_{23}\\
&=-\widetilde\sigma_{11}\underbrace{\bar\sigma_{23}\sigma_{22}}_{\in J}-\bar 1^{-3}\underbrace{\bar\sigma_{2,-1}\sigma_{23}}_{\in J}\sigma_{-1,2}+\tau_{11}\bar\sigma_{23},
\end{align*}
again the last equality by Proposition \ref{propodd}.
Hence $\tau_{11}\equiv 1 \bmod I$ and $\tau_{12}\equiv \bar\sigma_{23}\bmod J$ (note that $I\subseteq J$). Set
$\zeta:={}^{P_{13}P_{21}}\tau$. Then $\zeta_{22}=\tau_{11}$ and $\zeta_{23}=\tau_{12}$ and hence $\bar\zeta_{23}\zeta_{22}\equiv \sigma_{23}\bmod I+\bar J$. Applying Step 3 above to $\zeta$, we get that $T_{kl}(\bar\zeta_{23}\zeta_{22})$ is a product of $32$ elementary $\zeta$-conjugates. Since any elementary $\zeta$-conjugate is a product of $2$ elementary $\sigma$-conjugates, it follows that $T_{kl}(\bar\zeta_{23}\zeta_{22})$ is a product of $64$ elementary $\sigma$-conjugates. Thus, by Steps 1-3, $T_{kl}(\sigma_{23})$ is a product of $64+16+16+16+16+32=160$ elementary $\sigma$-conjugates. Since one can bring $\sigma_{ij}$ to position $(3,2)$ by conjugating elementary permutations, the assertion of (i) follows.\\
\\
(ii) One checks easily that $({}^{T_{ji}(1)}\sigma)_{j,-i}=\sigma_{i,-i}+\sigma_{j,-i}$. Applying (i) to $^{T_{ji}(1)}\sigma$ we get that $T_{kl}(\sigma_{i,-i}+\sigma_{j,-i})$ is a product of $160$ elementary $\sigma$-conjugates (note that any elementary $^{T_{ji}(1)}\sigma$-conjugate is also an elementary $\sigma$-conjugate). Applying (i) to $\sigma$ we get that $T_{kl}(\sigma_{j,-i})$ is a product of $160$ elementary $\sigma$-conjugates. It follows that $T_{kl}(\sigma_{i,-i})=T_{kl}(\sigma_{i,-i}+\sigma_{j,-i})T_{ji}(-\sigma_{j,-i})$ is a product of $320$ elementary $\sigma$-conjugates.\\
\\
(iii) Choose an $x\in R$ such that $(v_0,x)\in \mathfrak{L}_0$. One checks easily that $({}^{T_{-j}(v_0,x)}\sigma)_{ij}=\sigma_{ij}-\sigma_{i,-j}\epsilon_{-j}\bar x+(\sigma^{0h}v_0)_i$. Applying (i) to $^{T_{-j}(v_0,x)}\sigma$ we get that $T_{kl}(\sigma_{ij}-\sigma_{i,-j}\epsilon_{-j}\bar x+(\sigma^{0h}v_0)_i)$ is a product of $160$ elementary $\sigma$-conjugates. Applying (i) to $\sigma$ we get that $T_{kl}(-\sigma_{ij})$ and $T_{kl}(\sigma_{i,-j}\epsilon_{-j}\bar x)$ each are a product of $160$ elementary $\sigma$-conjugates. It follows that $T_{kl}((\sigma^{0h}v_0)_i)=T_{kl}(\sigma_{ij}-\sigma_{i,-j}\epsilon_{-j}\bar x+(\sigma^{0h}v_0)_i)T_{kl}(-\sigma_{ij})T_{kl}(\sigma_{i,-j}\epsilon_{-j}\bar x)$ is a product of $480$ elementary $\sigma$-conjugates.\\
\\
(iv) Choose an $x\in R$ such that $(v_0,x)\in \mathfrak{L}_0$. One checks easily that $({}^{T_{i}(v_0,x)}\sigma)_{ij}=-\epsilon_iB_0(v_0,\sigma^{h0}e_j)+\sigma_{ij}+\epsilon_ix\sigma_{-i,j}$. Applying (i) to $^{T_{i}(-1)}\sigma$ we get that $T_{kl}(B_0(v_0,\sigma^{h0}e_j)-\epsilon_i^{-1}\sigma_{ij}-x\sigma_{-i,j})$ is a product of $160$ elementary $\sigma$-conjugates. Applying (i) to $\sigma$ we get that $T_{kl}(\epsilon_i^{-1}\sigma_{ij})$ and $T_{kl}(x\sigma_{-i,j})$ each are a product of $160$ elementary $\sigma$-conjugates. It follows that $T_{kl}(B_0(v_0,\sigma^{h0}e_j))=T_{kl}(B_0(v_0,\sigma^{h0}e_j)-\epsilon_i^{-1}\sigma_{ij}-x\sigma_{-i,j})T_{kl}(\epsilon_i^{-1}\sigma_{ij})T_{kl}(x\sigma_{-i,j})$ is a product of $480$ elementary $\sigma$-conjugates.\\
\\
(v) One checks easily that $({}^{T_{ji}(1)}\sigma)_{ji}$ equals $\sigma_{ii}-\sigma_{jj}+\sigma_{ji}-\sigma_{ij}$. Applying (i) to $^{T_{ji}(1)}\sigma$ we get that $T_{kl}(\sigma_{ii}-\sigma_{jj}+\sigma_{ji}-\sigma_{ij})$ is a product of $160$ elementary $\sigma$-conjugates. Applying (i) to $\sigma$ we get that $T_{kl}(\sigma_{ij})$ and $T_{kl}(-\sigma_{ji})$ each are a product of $160$ elementary $\sigma$-conjugates. It follows that $T_{kl}(\sigma_{ii}-\sigma_{jj})=T_{kl}(\sigma_{ii}-\sigma_{jj}+\sigma_{ji}-\sigma_{ij})T_{kl}(\sigma_{ij})T_{kl}(-\sigma_{ji})$ is a product of $480$ elementary $\sigma$-conjugates.\\
\\
(vi) Follows from (v) since $T_{kl}(\sigma_{ii}-\sigma_{-i,-i})=T_{kl}(\sigma_{ii}-\sigma_{jj})T_{kl}(\sigma_{jj}-\sigma_{-i,-i})$.\\
\\
(vii) Set $m:=160$. In Step 1 we show that if $x\in R$, then $T_{k}(Q^0(\sigma e_1)\leftharpoonup\sigma_{11}x)$ is a product of $25m+4$ elementary $\sigma$-conjugates. In Step 2 we use Step 1 in order to prove (vii).\\
\\
{\bf Step 1.} Set $u:=e_{-2}\widetilde\sigma_{-1,-1}-e_{-1}\widetilde\sigma_{-1,-2}\in V_h$ and $v:=\widetilde\sigma u$. Then clearly $v_{-1}=0$. Further $v$ is isotropic since $u$ is isotropic and $\widetilde\sigma $ is unitary. By Lemma \ref{lemesdspec} we have $T_{e_1v}(0)\in \operatorname{EU}_6(\mathcal{V}_0)$. By Lemma \ref{lemesd}, $\widetilde{T_{e_1v}(0)}=T_{e_1,-v}(0)$ since $B(v,v)=B(\widetilde\sigma u,\widetilde\sigma u)=B(u,u)=0$. Set $\xi:={}^{\sigma}T_{e_1,-v}(0)=T_{\sigma e_1,-u}(0)$. One checks easily that
\begin{align*}
\xi^{hh}=
\arraycolsep=4pt\def1.5}\left(\begin{array}{ccc|ccc{1.5}\left(\begin{array}{ccc|ccc}
1-\sigma_{11}\sigma_{21}&\sigma_{11}\sigma_{11}&0&0&0&0\\
-\sigma_{21}\sigma_{21}&1+\sigma_{21}\sigma_{11}&0&0&0&0\\
-\sigma_{31}\sigma_{21}&\sigma_{31}\sigma_{11}&1&0&0&0\\
\hline-\sigma_{-3,1}\sigma_{21}&\sigma_{-3,1}\sigma_{11}&0&1&0&0\\
\bar\beta&\alpha&\overline{\sigma_{-3,1}\sigma_{11}}&-\bar 1^{-1}\overline{\sigma_{31}\sigma_{11}}&1-\bar 1^{-1}\overline{\sigma_{21}\sigma_{11}}&-\bar 1^{-1}\overline{\sigma_{11}\sigma_{11}}\\
\gamma&\beta&-\overline{\sigma_{-3,1}\sigma_{21}}&\bar 1^{-1}\overline{\sigma_{31}\sigma_{21}}&\bar 1^{-1}\overline{\sigma_{21}\sigma_{21}}&1+\bar 1^{-1}\overline{\sigma_{11}\sigma_{21}}
\end{array}\right)
\end{align*}
where $\alpha=\sigma_{-2,1}\sigma_{11}+\overline{\sigma_{-2,1}\sigma_{11}}$, $\beta=\sigma_{-1,1}\sigma_{11}-\overline{\sigma_{-2,1}\sigma_{11}}$ and $\gamma=-\sigma_{-1,1}\sigma_{21}-\overline{\sigma_{-1,1}\sigma_{21}}$. Further
\[\xi^{h0}(e_j)=
\begin{cases}
-\sigma^{h0}e_1\sigma_{21},&\text{ if }j=1,\\
\sigma^{h0}e_1\sigma_{11},&\text{ if }j=2,\\
0,&\text{ if }j\neq 1,2,
\end{cases},
\quad \xi^{0h}(w_0)=-uB_0(\sigma^{h0}e_1, w_0)~\text{ and }~\xi^{00}=id_{V_0}.\]
Set \[\tau:=T_{-3,1}(\sigma_{-3,1}\sigma_{21})T_{-3,2}(-\sigma_{-3,1}\sigma_{11}).\]
It follows from (i) that $\tau$ is a product of $2m$ elementary $\sigma$-conjugates. Then
\begin{align*}
(\xi\tau)^{hh}=
\arraycolsep=10pt\def1.5}\left(\begin{array}{ccc|ccc{1.5}\left(\begin{array}{ccc|ccc}
*&\sigma_{11}\sigma_{11}&0&0&0&0\\
*&1+\sigma_{21}\sigma_{11}&0&0&0&0\\
*&\sigma_{31}\sigma_{11}&1&0&0&0\\
\hline0&0&0&1&0&0\\
*&\alpha+\delta&0&*&*&*\\
*&\beta-\epsilon&0&*&*&*
\end{array}\right)
\end{align*}
where $\delta=\bar 1^{-1}\overline{\sigma_{31}\sigma_{11}}\sigma_{-3,1}\sigma_{11}$, $\epsilon=\bar 1^{-1}\overline{\sigma_{31}\sigma_{21}}\sigma_{-3,1}\sigma_{11}$ and a $*$ stands for an entry we are not interested in very much. Further $(\xi\tau)^{h0}=\xi^{h0}$, $(\xi\tau)^{0h}=\xi^{0h}$ and $(\xi\tau)^{00}=\xi^{00}$. Let $x\in R$ and set
\begin{align*}
\zeta:=&{}^{T_{e_1,-v}(0)}[T_{2,-3}(x),[T_{e_1v}(0),\sigma]\tau]\\
=&{}^{T_{e_1,-v}(0)}[T_{2,-3}(x),T_{e_1v}(0)\xi\tau]\\
=&[T_{e_1,-v}(0),T_{2,-3}(x)][T_{2,-3}(x),\xi\tau].
\end{align*}
Clearly $\zeta$ is a product of $4m+4$ elementary $\sigma$-conjugates. It follows from the relations in Lemma \ref{39} that
\begin{align*}
&[T_{e_1,-v}(0),T_{2,-3}(x)]\\
=&T_{1}(0,x\bar v_{-2}v_{-3}+\overline{x\bar v_{-2}v_{-3}}))T_{2,-1}(xv_{-3})T_{1,-3}(\bar 1^{-1} x\bar v_{-2})\\
=&T_{1}(0,a+\bar a)T_{1,-2}(b)T_{1,-3}(x(\sigma_{11}\sigma_{22}-\sigma_{12}\sigma_{21}))
\end{align*}
for some $a,b\in I(\sigma_{21},\sigma_{23})$. With a little effort one can show that
\[[T_{2,-3}(x),\xi\tau]=T_{1,-3}(-x\sigma_{11}\sigma_{11})T_{2,-3}(-x\sigma_{21}\sigma_{11})T_{-2,-3}(-x(\alpha+\delta))T_{-1,-3}(-x(\beta-\epsilon))T_{3}(\chi)\]
where $\chi=Q^0(\sigma e_1)\leftharpoonup\sigma_{11}x\+(0,c+\bar c)$ for some $c\in I(\sigma_{31},\sigma_{-2,1})$. Hence
\begin{align*}
\zeta=&T_{1}(0,a+\bar a)T_{1,-2}(b)T_{1,-3}(x(\sigma_{11}(\sigma_{22}-\sigma_{11})-\sigma_{12}\sigma_{21}))\cdot\\
&\cdot T_{2,-3}(-x\sigma_{21}\sigma_{11})T_{-2,-3}(-x(\alpha+\delta))T_{-1,-3}(-x(\beta-\epsilon))T_{3}(\chi).
\end{align*}
It follows from (i), (ii), (v) and relation (R9) in Lemma \ref{39} that $T_{3}(\chi)$ is a product of $4m+4+4m+2m+4m+m+3m+3m=21m+4$ elementary $\sigma$-conjugates. By (i) and relation (S5) in Lemma \ref{39}, $T_{3}(0,-c-\bar c)$ is a product of $4m$ elementary $\sigma$-conjugates. Hence $T_{3}(Q^0(\sigma e_1)\leftharpoonup\sigma_{11}x)=T_{3}(\chi)T_{3}(0,-c-\bar c)$ is a product of $25m+4$ elementary $\sigma$-conjugates. It follows from Lemma \ref{43} that $T_{k}(Q^0(\sigma e_1)\leftharpoonup\sigma_{11}x)$ is a product of $25m+4$ elementary $\sigma$-conjugates.\\
\\
{\bf Step 2.} Clearly
\[1=(\widetilde\sigma\sigma e_1)_1=(\widetilde\sigma ^{hh}\sigma ^{hh}e_1)_1+\underbrace{(\widetilde\sigma ^{0h}\sigma^{h0}e_1)_1}_{x:=}=
\sum\limits_{s\in\Theta}\widetilde\sigma _{1s}\sigma_{s1}+x.\]
It follows from Lemma \ref{last} that
\begin{align*}
&T_{k}(Q^0(\sigma e_1))\\
=&T_{k}(Q^0(\sigma e_1)\leftharpoonup(\sum\limits_{s\in\Theta}\widetilde\sigma _{1s}\sigma_{s1}+x))\\
=&T_{k}(\underbrace{\plus\limits_{s=1}^{-1} Q^0(\sigma e_1)\leftharpoonup\widetilde\sigma _{1s}\sigma_{s1}}_{A:=}\+\underbrace{Q^0(\sigma e_1)\leftharpoonup x}_{B:=}\+\underbrace{(0,y+\bar y)}_{C:=})\\
=&T_k(A)T_k(B)T_k(C)
\end{align*}
for some $y\in I(F_h(\sigma e_1))$ (note that $Q^0(\sigma e_1)\in \mathfrak{L}_0$ by Proposition \ref{propodd}). By Step 1, $T_{k}(Q^0(\sigma e_1)\leftharpoonup\widetilde\sigma _{11}\sigma_{11})$ is a product of $25m+4$ elementary $\sigma$-conjugates. By (i) and relation (R8) in Lemma \ref{39}, $T_k(Q^0(\sigma e_1)\leftharpoonup\widetilde\sigma _{1s}\sigma_{s1})$ is a product of $3m$ elementary $\sigma$-conjugates if $s\neq \pm 1$. By (ii) and relation (R8), $T_k(Q^0(\sigma e_1)\leftharpoonup\widetilde\sigma _{1,-1}\sigma_{-1,1})$ is a product of $6m$ elementary $\sigma$-conjugates. Hence $T_k(A)$ is a product of $43m+4$ elementary $\sigma$-conjugates. By (iv) and relation (R8), $T_k(B)$ is a product of $9m$ elementary $\sigma$-conjugates (note that $\sigma^{h0}e_1\in V^{ev}_0$). Since $F_h(\sigma e_1)=-\sum\limits_{r=1}^3\bar\sigma_{r1}\bar 1^{-1}\sigma_{-r,1}$, it follows from (i), (ii) and relation (R9) in Lemma \ref{39} that $T_k(C)$ is a product of $4m+2m+2m=8m$ elementary $\sigma$-conjugates. Hence $T_{k}(Q^0(\sigma e_1))$ is a product of $60m+4=9604$ elementary $\sigma$-conjugates. The assertion of (vii) follows now from Lemma \ref{new}.\\
\\
(viii) Choose an $x\in R$ such that $(v_0,x)\in \mathfrak{L}_0$. Set $\xi:={}^{T_{-j}(v_0,x)}\sigma$. One checks easily that \[\xi^{h0}e_j=\sigma^{00}v_0-v_0\sigma_{jj}+\sigma^{h0}e_j-\sigma^{h0}e_{-j}\epsilon_{-j}\bar x+v_0\sigma_{j,-j}\epsilon_{-j}\bar x-v_0(\sigma^{0h}v_0)_j.\]
By applying (vii) to $\xi$ we get that $T_{k}(Q^0(\xi e_j))=T_{k}(\xi^{h0}e_j, a)$ is a product of $10564$ elementary $\sigma$-conjugates where $a$ is some ring element. By applying (vii) to $\sigma$ we get that $T_{k}(Q^0(\sigma e_j)\leftharpoonup(-1))=T_k(-\sigma^{h0} e_j,b)$ and $T_{k}(Q^0(\sigma e_{-j})\leftharpoonup \epsilon_{-j}\bar x)=T_{k}(\sigma^{h0}e_{-j}\epsilon_{-j}\bar x,c)$ each are a product of $10564$ elementary $\sigma$-conjugates where $b$ and $c$ are some ring elements. By (ii) and relation (R8) in Lemma \ref{39}, $T_{k}(-v_0\sigma_{j,-j}\epsilon_{-j}\bar x,d)$ is a product of $3\cdot 320=960$ elementary $\sigma$-conjugates where $d$ is some ring element. By (iii) and relation (R8) in Lemma \ref{39}, $T_{k}(v_0(\sigma^{0h}v_0)_j,f)$ is a product of $3\cdot 480=1440$ elementary $\sigma$-conjugates where $f$ is some ring element. It follows that
\[T_{k}(\xi^{h0}e_j, a)T_k(-\sigma^{h0} e_j,b)T_{k}(\sigma^{h0}e_{-j}\epsilon_{-j}\bar x,c)T_{k}(-v_0\sigma_{j,-j}\epsilon_{-j}\bar x,d)T_{k}(v_0(\sigma^{0h}v_0)_j,f)=T_k(\sigma^{00}v_0-v_0\sigma_{jj},x')\] is a product of $3\cdot 10564+960+1440=34092$ elementary $\sigma$-conjugates where $x'$ is some ring element.
\end{proof}
We can now deduce our main result from Theorems \ref{thmscf} and \ref{thmm}.
\begin{SCT}\label{thmsct}
Let $H$ be a subgroup of $\operatorname{U}_{6}(\mathcal{V}_0)$. Then $H$ is E-normal iff
\begin{equation}
\operatorname{EU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)\subseteq H\subseteq \operatorname{CU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)
\end{equation}
for some odd form ideal $\mathfrak{I}_0$. Furthermore, $\mathfrak{I}_0$ is uniquely determined, namely it is the level of $H$.
\end{SCT}
\begin{proof}
First suppose that $H$ is E-normal. Let $\mathfrak{I}_0$ denote the level of $H$. Then $\operatorname{EU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)\subseteq H$. It remains to show that $H\subseteq \operatorname{CU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)$. Let $\sigma\in H$. By Corollary \ref{cornu} we have $\sigma\in \operatorname{NU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)$. Theorem \ref{thmm} and Proposition \ref{propcumax} imply that $\sigma\in \operatorname{CU}_{6}(\mathcal{V}_0,(\mathfrak{I}_0)_{\max})$. Further Theorem \ref{thmm} implies that $Q^0(\sigma e_i)\in\mathfrak{M}_0$ for any $i\in\Theta$ and $Q^0([\sigma,\tau] e_i)\in\mathfrak{M}_0$ for any elementary transvection $\tau$ and $i\in\Theta$. It follows from Lemma \ref{lemcu} that $\sigma\in \operatorname{CU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)$. Thus we have shown that (6) holds. Conversely, suppose that (6) holds for some odd form ideal $\mathfrak{I}_0$. Then it follows from the standard commutator formula in Theorem \ref{thmscf} that $H$ is E-normal. The uniqueness of $\mathfrak{I}_0$ follows from the standard commutator formula and the easy-to-check fact that $\operatorname{EU}_{6}(\mathcal{V}_0,\mathfrak{I}_0)\subseteq \operatorname{U}_{6}(\mathcal{V}_0,\mathfrak{J}_0)~\Leftrightarrow~\mathfrak{I}_0\subseteq \mathfrak{J}_0$.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
The planar interface of two dissimilar materials can guide the propagation of electromagnetic surface waves. A variety of different types of electromagnetic surface wave have been identified, with the type depending upon whether the partnering materials are isotropic or anisotropic, dissipative or nondissipative, homogeneous or nonhomogeneous, and so forth \c{ESW_book}. The surface-plasmon-polariton (SPP) wave \c{Pitarke,Maier}~---~which is guided by the planar interface of a metal
and a dielectric material~---~is the most familiar type of electromagnetic surface wave, being widely exploited in optical sensing applications \c{Homola_book,AZLe}. The Dyakonov surface wave \c{Dyakonov88}~---~which is guided by the planar interface of an isotropic dielectric material and an anisotropic dielectric material~---~has also been widely reported upon \c{Walker98,Takayama_exp}. Dyakonov waves are promising for applications in optical communications \c{DSWreview}.
As compared to achiral materials, chiral materials \c{Beltrami}~---~with their inherent magnetoelectric coupling~---~offer wider opportunities for surface-wave propagation. However,
whereas surface waves guided by planar interfaces involving achiral partnering materials have been comprehensively studied, there have been only a few studies of the surface waves supported by chiral materials \c{Pattanayak,Engheta,Fantino,Pellegrini}. And these few studies have largely focussed on isotropic partnering materials and nondissipative chiral materials. In the present paper we address this issue by
investigating surface waves supported by a chiral material with the effects of dissipation and anisotropy of the partnering
materials being taken into account. In particular, a novel type of chiral material that can simultaneously support attenuation and amplification of plane waves \c{ML_JO}, depending upon circular polarization state, is considered. Our theoretical and numerical studies
are based on the canonical boundary-value problem for surface-wave propagation \c{ESW_book}.
In the following, the permittivity and permeability of free space are written as $\eps_{\scriptscriptstyle 0}$ and $\mu_{\scriptscriptstyle 0}$, respectively. The free-space wavenumber is $k_{\scriptscriptstyle 0} = \omega \sqrt{\eps_{\scriptscriptstyle 0} \mu_{\scriptscriptstyle 0}}$, where $\omega$ is the angular frequency. The operators $\mbox{Re} \left\{ \. \right\}$
and $\mbox{Im} \left\{ \. \right\}$ deliver the
real and imaginary parts of complex-valued quantities, and $i = \sqrt{-1}$.
Single underlining signifies a 3 vector while double underlining signifies a 3$\times$3 dyadic.
The triad of unit vectors aligned with the Cartesian axes are denoted as $\left\{ \hat{\#u}_x, \hat{\#u}_y, \hat{\#u}_z \right\}$.
\section{Isotropic chiral material/isotropic dielectric material interface} \l{iso_sec}
Let us consider the canonical boundary-value problem for surface waves guided by the planar interface of an isotropic chiral material
and an isotropic dielectric material. Both partnering materials are homogeneous.
The isotropic chiral material, labeled $\mathcal{A}$, fills
the half-space $z>0$ and is characterized by the frequency-domain Tellegen constitutive relations \c{Beltrami}
\begin{equation} \l{cr}
\left.
\begin{array}{l}
\#D (\#r) = \eps_{\scriptscriptstyle 0} \varepsilon_\mathcal{A} \#E (\#r) + i \sqrt{\eps_{\scriptscriptstyle 0} \mu_{\scriptscriptstyle 0}} \xi_\mathcal{A} \#H (\#r) \,\\ [5pt]
\#B (\#r)= - i \sqrt{\eps_{\scriptscriptstyle 0} \mu_{\scriptscriptstyle 0}} \xi_\mathcal{A} \#E (\#r) + \mu_{\scriptscriptstyle 0} \mu_\mathcal{A} \#H (\#r) \,
\end{array}
\right\}\, \qquad z>0 .
\end{equation}
The relative permittivity scalar $\varepsilon_{\cal A}$, the relative permeability scalar $\mu_{\cal A}$, and the relative
chirality pseudoscalar
$\xi_{\cal A}$ are frequency dependent and complex valued, per the principle
of causality embodied by the Kramers--Kronig relations \cite{2lines}.
The isotropic dielectric material, labeled $\mathcal{B}$, fills
the half-space $z<0$ and is characterized by the relative permittivity
$\varepsilon_\mathcal{B}$.
\subsection{Theory}
The electromagnetic field phasors in the partnering materials ${\cal A}$ and ${\cal B}$ are represented by
\begin{equation} \label{planewave}
\left.\begin{array}{l}
\#E_{\,\ell} (\#r)= \#{\mathcal E}_{\,\ell} \,\exp\left({i\#k_{\,\ell} \cdot\#r}\right) \\[4pt]
\#H_{\,\ell}(\#r)= \#{\mathcal H}_{\,\ell} \,\exp\left({i\#k_{\,\ell} \cdot\#r}\right)
\end{array}\right\}\,, \qquad \ell \in\left\{ {\cal A}, {\cal B} \right\}\,.
\end{equation}
The amplitude vectors $ \#{\mathcal E}_{\,\ell} $ and $ \#{\mathcal H}_{\,\ell}$ have complex-valued components, and so does
the wave vector
$\#k_{\,\ell}$. The field phasors (and the wave vector) can vary with angular frequency $\omega$. Without loss of generality, we consider the surface-wave propagation parallel to $\hat{\#u}_x$ in the $xy$ plane; i.e., $\hat{\#u}_y\cdot\#k_{\,\ell}\equiv 0$.
In the half-space $z>0$, the Maxwell curl postulates yield
\begin{equation} \l{MP_A}
\left.
\begin{array}{l}
\#k_{\, {\cal A}} \times \#{\mathcal E}_{\,{\cal A}} - \omega \left( - i \sqrt{\eps_{\scriptscriptstyle 0} \mu_{\scriptscriptstyle 0}} \xi_{{\cal A}} \#{\mathcal E}_{\,{\cal A}} +
\mu_{\scriptscriptstyle 0} \mu_\mathcal{A} \#{\mathcal H}_{\,{\cal A}} \right) = \#0
\vspace{4pt} \\
\#k_{\, {\cal A}} \times \#{\mathcal H}_{\,{\cal A}} + \omega \left( \eps_{\scriptscriptstyle 0} \varepsilon_{{\cal A}} \#{\mathcal E}_{\,{\cal A}}
+ i \sqrt{\eps_{\scriptscriptstyle 0} \mu_{\scriptscriptstyle 0}} \xi_{{\cal A}} \#{\mathcal H}_{\,{\cal A}}
\right) = \#0
\end{array}
\right\}\,,
\end{equation}
where the wave vector
\begin{equation} \l{wvA}
\#k_{\, {\cal A}} = k_{\scriptscriptstyle 0} \left( q \, \hat{\#u}_x + i \alpha_{\cal A} \hat{\#u}_z \, \right)
\end{equation}
and ${\rm Re}\left\{\alpha_{\cal A}\right\}>0$ for surface-wave propagation.
On combining Eqs.~\r{MP_A} and Eq.~\r{wvA}, a biquadratic dispersion relation
emerges for $\alpha_{\cal A}$. The two $\alpha_{\cal A}$ roots with non-negative real parts are identified as
\begin{equation} \l{a_decay_const}
\left.
\begin{array}{l}
\alpha_{{\cal A} 1} = \sqrt{q^2 - \kappa_R^2 } \vspace{8pt}\\
\alpha_{{\cal A} 2} =
\sqrt{q^2 - \kappa_L^2 }
\end{array}
\right\}\,,
\end{equation}
with the complex-valued scalars
\begin{equation}
\left.
\begin{array}{l}
\kappa_R= \sqrt{\varepsilon_{\cal A} \mu_{\cal A} } + \xi_{\cal A} \vspace{6pt}\\
\kappa_L= \sqrt{\varepsilon_{\cal A} \mu_{\cal A} } - \xi_{\cal A} \end{array} \right\}
\end{equation}
being associated with the relative wave numbers for right and left circularly-polarized light in an unbounded chiral medium \c{Beltrami}.
Accordingly the field-phasor amplitudes are given as
\begin{equation}
\left.
\begin{array}{l} \l{A_p}
\#{\mathcal E}_{\,{\cal A}} = A_{{\cal A} 1}\, \#{\mathcal E}_{\,{\cal A} 1} + A_{{\cal A} 2} \,\#{\mathcal E}_{\,{\cal A} 2} \vspace{6pt}\\
\#{\mathcal H}_{\,{\cal A}} = \displaystyle{\sqrt{\frac{\eps_{\scriptscriptstyle 0}}{\mu_{\scriptscriptstyle 0}}} \sqrt{\frac{\varepsilon_{\cal A}}{\mu_{\cal A}}
} \left( A_{{\cal A} 1} \,\#{\mathcal H}_{\,{\cal A} 1} + A_{{\cal A} 2} \,\#{\mathcal H}_{\,{\cal A} 2} \right) }
\end{array}
\right\},
\end{equation}
where the vectors
\begin{equation}
\left.
\begin{array}{l} \l{A_p2}
\#{\mathcal E}_{\,{\cal A} 1} = \alpha_{{\cal A} 1} \, \hat{\#u}_x
+ \kappa_R \, \hat{\#u}_y + i q \, \hat{\#u}_z \vspace{6pt}\\
\#{\mathcal E}_{\,{\cal A} 2} =
- \alpha_{{\cal A} 2} \, \hat{\#u}_x
+ \kappa_L \, \hat{\#u}_y - i q \, \hat{\#u}_z
\vspace{6pt}\\
\#{\mathcal H}_{\,{\cal A} 1} = \displaystyle{
-i \alpha_{{\cal A} 1} \, \hat{\#u}_x
- i \kappa_R \, \hat{\#u}_y +
q \, \hat{\#u}_z
}
\vspace{6pt}\\
\#{\mathcal H}_{\,{\cal A} 2} =
\displaystyle{
-i \alpha_{{\cal A} 2} \, \hat{\#u}_x +
i \kappa_L \, \hat{\#u}_y +
q \, \hat{\#u}_z
}
\end{array}
\right\}\,.
\end{equation}
In the half-space $z<0$, the Maxwell curl postulates yield
\begin{equation}
\l{MP_B}
\left.
\begin{array}{l}
\#k_{\, {\cal B}} \times \#{\mathcal E}_{\,{\cal B}} - \omega \mu_{\scriptscriptstyle 0} \#{\mathcal H}_{\,{\cal B}} = \#0\vspace{4pt} \\
\#k_{\, {\cal B}} \times \#{\mathcal H}_{\,{\cal B}} + \omega \eps_{\scriptscriptstyle 0} \varepsilon_{\cal B} \#{\mathcal E}_{\,{\cal B}} = \#0
\end{array}
\right\}\,,
\end{equation}
where the wave vector
\begin{equation} \l{kB}
\#k_{\, {\cal B}} = k_{\scriptscriptstyle 0} \left( q \, \hat{\#u}_x - i \alpha_{\cal B} \hat{\#u}_z \, \right),
\end{equation}
and the scalar
\begin{equation} \l{b_decay_const}
\alpha_{\cal B} = \sqrt{q^2 - \varepsilon_{\cal B} }
\end{equation}
satisfies the inequality $\mbox{Re} \left\{ \alpha_{\cal B} \right\} > 0$ for surface-wave propagation. Hence the field-phasor amplitudes are given as
\begin{equation} \l{B_p}
\left.
\begin{array}{l}
\#{\mathcal E}_{\,{\cal B}} = A_{{\cal B} 1} \,\hat{\#u}_y + A_{{\cal B} 2} \,\left( i \alpha_{\cal B} \hat{\#u}_x + q \hat{\#u}_z \right) \vspace{6pt}\\
\#{\mathcal H}_{\,{\cal B}} = \displaystyle{\sqrt{\frac{\eps_{\scriptscriptstyle 0}}{ \mu_{\scriptscriptstyle 0}}} \left[
A_{{\cal B} 1}\, \left( i \alpha_{\cal B} \hat{\#u}_x + q \hat{\#u}_z \right) - A_{{\cal B} 2} \, \varepsilon_{\cal B}\hat{\#u}_y \, \right] }
\end{array}
\right\}.
\end{equation}
The scalars
$A_{{\cal A} 1} $ and $A_{{\cal A} 2} $ in Eqs.~\r{A_p}, and $A_{{\cal B} 1} $ and $A_{{\cal B} 2} $ in Eqs.~\r{B_p}, as well as
the relative wave number $q$,
are determined by enforcing boundary conditions
across the planar interface $z=0$, as follows.
The continuity of tangential components of the electric and magnetic field
phasors across the planar interface $z=0$ imposes
the four conditions \c{Chen}
\begin{equation} \l{4cond}
\left.
\begin{array}{l}
\alpha_{{\cal A} 1} A_{{\cal A} 1} - \alpha_{{\cal A} 2} A_{{\cal A} 2} = i \alpha_{\cal B} A_{{\cal B} 2} \vspace{6pt} \\
\kappa_R A_{{\cal A} 1} + \kappa_L A_{{\cal A} 2} = A_{{\cal B} 1} \vspace{6pt} \\
\displaystyle{-
\sqrt{\varepsilon_{\cal A}}
\left( \alpha_{{\cal A} 1} A_{{\cal A} 1}
+ \alpha_{{\cal A} 2}
A_{{\cal A} 2} \right) = \alpha_{{\cal B} } \sqrt{\mu_{\cal A}} A_{{\cal B} 1}}
\vspace{6pt} \\
\displaystyle{
\sqrt{\varepsilon_{\cal A} }
\left( \kappa_R A_{{\cal A} 1}
- \kappa_L
A_{{\cal A} 2} \right)
= - i \varepsilon_{{\cal B}} \sqrt{\mu_{\cal A}} A_{{\cal B} 2}}
\end{array}
\right\}.
\end{equation}
The four conditions \r{4cond} may be represented compactly as
\begin{equation} \l{M0}
\left[ M \right] \. \left[ \begin{array}{c}
A_{{\cal A} 1} \\
A_{{\cal A} 2} \\
A_{{\cal B} 1} \\
A_{{\cal B} 2}
\end{array}
\right] = \left[ \begin{array}{c}
0 \\
0 \\
0 \\
0
\end{array}
\right],
\end{equation}
wherein the 4$\times$4 matrix $\left[ M \right]$ must be singular for surface-wave propagation.
The dispersion equation $\det \left[ M \right] = 0$ reduces to the equation
\begin{eqnarray} \l{DE}
&&
2 \sqrt{\varepsilon_{\cal A} \mu_{\cal A}} \left( \alpha_{{\cal A} 1} \alpha_{{\cal A} 2} \varepsilon_{\cal B} + \kappa_L \kappa_R \alpha^2_{{\cal B}} \right)
+ \alpha_{{\cal B}} \left( \kappa_L \alpha_{{\cal A} 1} + \kappa_R \alpha_{{\cal A} 2} \right) \left( \varepsilon_{\cal A} + \varepsilon_{\cal B} \mu_{\cal A} \right) = 0,
\end{eqnarray}
from which $q$ may be extracted, generally by numerical means. Once $q$ is known, relative values of the
four
scalars $A_{{\cal A} 1,2} $ and $A_{{\cal B} 1,2} $ can be determined from Eq.~\r{M0} by straightforward algebraic manipulations.
\subsection{Numerical studies} \label{ns1}
For our numerical studies we fix partnering material ${\cal A}$ by selecting
the relative constitutive parameters
$\varepsilon_{\cal A} = 2.6724 - 0.0007 i $, $\xi_{\cal A} = 0.0652 + 0.0005 i$ and $\mu_A = 0.9642 + 0.0001 i $. These parameters prescribe a
homogenized composite material that arises from blending together a realistic isotropic chiral material with an active dielectric material, namely a rhodamine mixture \c{ML_JO}. The constitutive parameters of the homogenized composite material were estimated using the Bruggeman formalism \c{MAEH}. For material ${\cal A}$, the relative wave numbers are $\kappa_L = 1.5401 - 0.0006 i $ and $\kappa_R =1.6705 + 0.0004 i$. Therefore, if material ${\cal A}$ were unbounded it would simultaneously support the amplification of left circularly-polarized plane waves and the attenuation of right circularly-polarized plane waves \c{ML_JO}.
Parenthetically, in a similar vein, anisotropic dielectric materials can be engineered that amplify plane waves of
one
linearly-polarized state but attenuate plane waves of the other linearly-polarized state \c{ML_PRA}.
Let us begin by considering the case wherein partnering material ${\cal B}$ is characterized by a real-valued relative permittivity with $\varepsilon_{\cal B} < 0$, That is, material ${\cal B}$ behaves like an idealized plasmonic material
and the corresponding surface waves are akin to Fano waves \c{ESW_book,Fano}. The dispersion equation
\r{DE}
then yields one (complex-valued) $q$ solution. The real and imaginary parts of the relative wave number $q$ are plotted
against relative permittivity $\varepsilon_{\cal B}$ in Fig.~\ref{Fig1}. While the real part of $q$ remains almost constant as $\varepsilon_{\cal B}$ is increased, the imaginary part of $q$ undergoes a much more dramatic change. In particular, $\mbox{Im} \left\{ q \right\}$ is positive valued for $\varepsilon_{\cal B} < -21$ but is negative valued for $\varepsilon_{\cal B} > -21$. Therefore, for sufficiently small values of
$\varepsilon_{\cal B}$ the surface wave is attenuated as it propagates whereas for larger values of
$\varepsilon_{\cal B}$ the surface wave is amplified as it propagates.
We explore this behaviour further, in a more realistic setting, by supposing that partnering material ${\cal B}$
exhibits a small degree of dissipation. That is, we fix $\mbox{Re} \left\{ \varepsilon_{\cal B} \right\} = - 10$ and consider $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\} > 0 $.
Accordingly, material ${\cal B}$ behaves like a realistic plasmonic material and the corresponding surface waves are akin to SPP waves \c{BGprA,BGprB}.
As for the case represented in Fig.~\ref{Fig1}, the dispersion equation
\r{DE}
delivers a single (complex-valued) $q$ solution. The real and imaginary parts of the relative wave number $q$ are plotted
against the imaginary part of the relative permittivity $\varepsilon_{\cal B}$ in Fig.~\ref{Fig2}.
The real part of $q$ is almost independent of $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\}$ over the range considered in Fig.~\ref{Fig2}.
In contrast, for $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\} < 0.005$ we have $\mbox{Im} \left\{ q \right\} < 0$
whereas for $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\} > 0.005$ we have $\mbox{Im} \left\{ q \right\} > 0$. Therefore,
when the degree of dissipation exhibited by material ${\cal B}$ is sufficiently small the surface wave is amplified, but
when the degree of dissipation exhibited by material ${\cal B}$ is larger the surface wave is attenuated.
Next we turn to the case where $\mbox{Re} \left\{ \varepsilon_{\cal B} \right\} > 0$.
Specifically, let $\mbox{Re} \left\{ \varepsilon_{\cal B} \right\} = 10$.
Notice that in this case
partnering material ${\cal B}$ does not behave like a plasmonic material and the corresponding surface waves
are not at all akin to
SPP waves. In Fig.~\ref{Fig3}, the real and imaginary parts of $q$ are plotted against
$\mbox{Im} \left\{ \varepsilon_{\cal B} \right\}$, for the single (complex-valued) solution emerging from the dispersion equation
\r{DE} for $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\} > 0.05$. The real part of $q$ is almost independent of $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\}$ whereas the imaginary part of $q$ increases approximately linearly as $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\}$ increases.
Since $\mbox{Im} \left\{ q \right\} > 0 $, the corresponding surface wave is attenuated over the range of values of $\mbox{Im} \left\{ \varepsilon_{\cal B} \right\}$ considered. If $ 0 \leq \mbox{Im} \left\{ \varepsilon_{\cal B} \right\} < 0.05$ then no solutions emerge from the dispersion equation
\r{DE}. In particular, no solutions are found when material ${\cal B}$ is nondissipative and $\varepsilon_{\cal B} > 0$, a result that is consistent with an
earlier study \c{Fantino}.
\section{Isotropic chiral material/anisotropic dielectric material interface} \l{aniso_sec}
Now we extend the canonical boundary-value problem presented in Sec.~\ref{iso_sec} by replacing
the
isotropic dielectric material, i.e., partnering material ${\cal B}$, with an anisotropic dielectric material. Partnering material ${\cal A}$ is still taken to be an isotropic chiral material, per the Tellegen constitutive relations \r{cr}.
To be specific, partnering material ${\cal B}$ is taken to be a uniaxial dielectric material characterized by the relative permittivity dyadic \c{Chen}
\begin{equation}
\=\varepsilon_\mathcal{B} = \varepsilon_\mathcal{B}^{\rm s} \=I + \left(
\varepsilon_\mathcal{B}^{\rm t} - \varepsilon_\mathcal{B}^{\rm s} \right) \,
\hat{\#{u}}_{\rm } \, \hat{\#{u}}_{\rm }\,, \l{Ch4_eps_uniaxial}
\end{equation}
where $\=I=\hat{\#u}_x\ux+\hat{\#u}_y\uy+\hat{\#u}_z\uz$ is the identity dyadic.
The
optic axis of material ${\cal B}$ lies in the $xy$ plane,
oriented at angle $\psi$
with respect to the direction of surface-wave propagation; i.e,
\begin{equation}
\hat{\#{u}}_{\rm } = \cos \psi \, \hat{\#u}_x + \sin \psi \, \hat{\#u}_y.
\end{equation}
The dyadic $\=\varepsilon_\mathcal{B}$ has two eigenvalues: $\varepsilon_\mathcal{B}^{\rm s}$ which governs the propagation of \textit{ordinary}
plane waves and $\varepsilon_\mathcal{B}^{\rm t}$ which, together with $\varepsilon_\mathcal{B}^{\rm s}$ ,
governs the propagation of \textit{extraordinary}
plane waves \cite{BW}.
\subsection{Theory}
In the half-space $z>0$, Eqs.~\r{MP_A}--\r{A_p2} continue to hold.
In the half-space $z<0$, the Maxwell curl postulates yield
\begin{equation} \l{xMP_A}
\left.
\begin{array}{l}
\#k_{\, {\cal B}} \times \#{\mathcal E}_{\,{\cal B}} = \omega \mu_{\scriptscriptstyle 0} \#{\mathcal H}_{\,{\cal B}} \vspace{4pt} \\
\#k_{\, {\cal B}} \times \#{\mathcal H}_{\,{\cal B}} = - \omega \eps_{\scriptscriptstyle 0} \=\varepsilon_{\,{\cal B}} \. \#{\mathcal E}_{\,{\cal B}}
\end{array}
\right\}\,,
\end{equation}
where the wave vector $\#k_{\, {\cal B}}$ has the form given in Eq.~\r{kB}.
The combination of Eqs.~\r{xMP_A} and Eq.~\r{kB} yields a biquadratic dispersion relation
for $\alpha_{\cal B}$.
For surface-wave progation the two $\alpha_{\cal B}$ roots with non-negative real parts are prescribed; these are
\begin{equation} \l{xa_decay_const}
\left.
\begin{array}{l}
\alpha_{{\cal B} 1} = \sqrt{q^2 -\varepsilon_{\cal B}^s} \vspace{8pt}\\
\alpha_{{\cal B} 2} = \displaystyle{
\sqrt{ \varepsilon^t_{\cal B} \left[
q^2 \left( \frac{ \cos^2 \psi}{\varepsilon^s_{\cal B}} +
\frac{\sin^2 \psi}{\varepsilon^t_{\cal B}} \right) -1 \right]}}
\end{array}
\right\}\,.
\end{equation}
Hence the field phasor amplitudes for $z<0$ are
\begin{equation}
\left.
\begin{array}{l} \l{B_p2}
\#{\mathcal E}_{\,{\cal B}} = A_{{\cal B} 1}\, \#{\mathcal E}_{\,{\cal B} 1} + A_{{\cal B} 2} \,\#{\mathcal E}_{\,{\cal B} 2} \vspace{6pt}\\
\#{\mathcal H}_{\,{\cal B}} = \displaystyle{\sqrt{\frac{\eps_{\scriptscriptstyle 0}}{\mu_{\scriptscriptstyle 0}}} \left( A_{{\cal B} 1} \,\#{\mathcal H}_{\,{\cal B} 1} + A_{{\cal B} 2} \,\#{\mathcal H}_{\,{\cal B} 2} \right) }
\end{array}
\right\},
\end{equation}
where the vectors
\begin{equation}
\left.
\begin{array}{l}
\#{\mathcal E}_{\,{\cal B} 1} = i \alpha_{{\cal B} 1} \sin \psi \, \hat{\#u}_x
- i \alpha_{{\cal B} 1} \cos \psi \, \hat{\#u}_y + q \sin \psi \, \hat{\#u}_z \vspace{6pt}\\
\#{\mathcal E}_{\,{\cal B} 2} = \alpha^2_{{\cal B} 1} \cos \psi \, \hat{\#u}_x
- \varepsilon^s_{\cal B} \sin \psi \, \hat{\#u}_y - i q \alpha_{{\cal B} 2} \cos \psi \, \hat{\#u}_z \vspace{6pt}\\
\#{\mathcal H}_{\,{\cal B} 1} = \displaystyle{
\alpha^2_{{\cal B} 1} \cos \psi \, \hat{\#u}_x
- \varepsilon^s_{\cal B} \sin \psi \, \hat{\#u}_y -
i \alpha_{{\cal B} 1} q \cos \psi \, \hat{\#u}_z
}
\vspace{6pt}\\
\#{\mathcal H}_{\,{\cal B} 2} = \displaystyle{ - i \alpha_{{\cal B} 2} \varepsilon^s_{\cal B} \sin \psi \, \hat{\#u}_x
+ i \alpha_{{\cal B} 2} \varepsilon^s_{\cal B} \cos \psi \, \hat{\#u}_y -
\varepsilon^s_{\cal B} q \sin \psi \, \hat{\#u}_z
}
\end{array}
\right\}\,.
\end{equation}
As in Sec.~\ref{iso_sec}, the scalars
$A_{{\cal A} 1} $ and $A_{{\cal A} 2} $ in Eqs.~\r{A_p} , and $A_{{\cal B} 1} $ and $A_{{\cal B} 2} $ in Eqs.~\r{B_p2}, as well as
the wave number $q$,
are determined by enforcing boundary conditions
across the planar interface $z=0$. To this end,
the continuity of tangential components of the electric and magnetic field
phasors across the planar interface $z=0$ yields
the four conditions \c{Chen}
\begin{equation} \l{4cond2}
\left.
\begin{array}{l}
\alpha_{{\cal A} 1} A_{{\cal A} 1} - \alpha_{{\cal A} 2} A_{{\cal A} 2} = i \alpha_{{\cal B} 1} \sin \psi A_{{\cal B} 1}
+ \alpha^2_{{\cal B} 1} \cos \psi A_{{\cal B} 2}
\vspace{6pt} \\
\kappa_R A_{{\cal A} 1} + \kappa_L A_{{\cal A} 2} =
- i \alpha_{{\cal B} 1} \cos \psi A_{{\cal B} 1}
- \varepsilon^{\rm s}_{{\cal B} } \sin \psi A_{{\cal B} 2}
\vspace{6pt} \\
\displaystyle{- i
\sqrt{\frac{\varepsilon_{\cal A}}{\mu_{\cal A}}}
\left( \alpha_{{\cal A} 1} A_{{\cal A} 1}
+ \alpha_{{\cal A} 2}
A_{{\cal A} 2} \right) = \sqrt{\frac{\eps_{\scriptscriptstyle 0}}{\mu_{\scriptscriptstyle 0}}} \left( \alpha^2_{{\cal B} 1} \cos \psi A_{{\cal B} 1}
- i \alpha_{{\cal B} 2} \varepsilon^{\rm s}_{{\cal B} } \sin \psi A_{{\cal B} 2} \right)
}
\vspace{6pt} \\
\displaystyle{-i
\sqrt{\frac{\varepsilon_{\cal A}}{\mu_{\cal A}}}
\left( \kappa_R A_{{\cal A} 1}
- \kappa_L
A_{{\cal A} 2} \right)
= \sqrt{\frac{\eps_{\scriptscriptstyle 0}}{\mu_{\scriptscriptstyle 0}}} \left( - \varepsilon^{\rm s}_{{\cal B} } \sin \psi A_{{\cal B} 1}
+ i \alpha_{{\cal B} 2} \varepsilon^{\rm s}_{{\cal B} } \cos \psi A_{{\cal B} 2} \right) }
\end{array}
\right\},
\end{equation}
which are represented compactly as
\begin{equation} \l{N0}
\left[ N \right] \. \left[ \begin{array}{c}
A_{{\cal A} 1} \\
A_{{\cal A} 2} \\
A_{{\cal B} 1} \\
A_{{\cal B} 2}
\end{array}
\right] = \left[ \begin{array}{c}
0 \\
0 \\
0 \\
0
\end{array}
\right].
\end{equation}
The 4$\times$4 matrix $\left[ N \right]$ must be singular for surface-wave propagation.
The dispersion equation $\det \left[ N \right] = 0$ reduces to the equation
\begin{eqnarray} \l{DE2}
&&
\alpha_{{\cal B} 1} \left[ 2 \sqrt{\varepsilon_{\cal A} \mu_{\cal A}} \left( \kappa_L \kappa_R \alpha^3_{{\cal B} 1} + \varepsilon^s_{\cal B} \alpha_{{\cal A} 1} \alpha_{{\cal A} 2}
\alpha_{{\cal B} 2} \right) + \alpha_{{\cal B} 1} \left( \kappa_L \alpha_{{\cal A} 1} + \kappa_R \alpha_{{\cal A} 2} \right) \left( \varepsilon_{\cal A} \alpha_{{\cal B} 1} + \varepsilon^s_{\cal B}
\mu_{\cal A} \alpha_{{\cal B} 2} \right) \right] \cos^2 \psi \nonumber \\ && -\varepsilon^s_{\cal B} \left\{ \left[ 2 \sqrt{\varepsilon_{\cal A} \mu_{\cal A}} \left( \kappa_L \kappa_R \alpha_{{\cal B} 1} \alpha_{{\cal B} 2} + \varepsilon^s_{\cal B} \alpha_{{\cal A} 1}
\alpha_{{\cal A} 2} \right) + \left( \kappa_L \alpha_{{\cal A} 1} + \kappa_R \alpha_{{\cal A} 2} \right) \left( \varepsilon_{\cal A} \alpha_{{\cal B} 1} + \varepsilon^s_{\cal B} \mu_{\cal A} \alpha_{{\cal B} 2}
\right) \right] \sin^2 \psi \right. \nonumber \\ && \left.
- \sqrt{\varepsilon_{\cal A} \mu_{\cal A}} \left( \kappa_L \alpha_{{\cal A} 1} - \kappa_R \alpha_{{\cal A} 2} \right) \alpha_{{\cal B} 1} \left( \alpha_{{\cal B} 1} - \alpha_{{\cal B} 2} \right) \sin 2 \psi
\right\} = 0.
\end{eqnarray}
After extracting $q$ from Eq.~\r{DE2}, generally by numerical means, the relative values of the
four
scalars $A_{{\cal A} 1,2} $ and $A_{{\cal B} 1,2} $ can be determined from Eq.~\r{N0} by straightforward algebraic manipulations.
\subsection{Numerical studies}\label{ns2}
As in Sec.~\ref{ns1}, we fix partnering material ${\cal A}$ by selecting
the relative constitutive parameters
$\varepsilon_{\cal A} = 2.6724 - 0.0007 i $, $\xi_{\cal A} = 0.0652 + 0.0005 i$ and $\mu_A = 0.9642 + 0.0001 i $.
Let us start with the case where
the relative permittivity parameters of material ${\cal B}$, namely $\varepsilon^s_{\cal B}$ and $\varepsilon^t_{\cal B}$,
have negative-valued real parts and positive-valued imaginary parts.
Specifically, let $\varepsilon^s_{\cal B} = -12 + 0.0045 i $ and $\varepsilon^t_{\cal B}= - 8 + 0.005i$.
Thus, material ${\cal B}$ behaves like an anisotropic plasmonic material that exhibits a modest degree of dissipation. The corresponding surface waves are akin to SPP waves.
In Fig.~\ref{Fig4}, the real and imaginary parts of $q$ are plotted against orientation angle $\psi$
for the solitary (complex-valued) $q$ solution emerging from the dispersion equation
\r{DE2}. The real part of $q$ is almost independent of $\psi$. In contrast, the imaginary part of $q$ oscillates as $\psi$ increases such that $\mbox{Im} \left\{ q \right\} > 0$ for $\psi \in \left( 0^\circ, 64^\circ \right) \cup
\left( 170^\circ, 180^\circ \right)$ and $\mbox{Im} \left\{ q \right\} < 0$
for $\psi \in \left[ 64^\circ, 170^\circ \right]$. Therefore, for certain orientations of material ${\cal B}$ the surface wave is attenuated as it propagates while for other orientations
it is amplified as it propagates.
Furthermore, at two specific orientations of material ${\cal B}$ the surface wave propagates with neither attenuation nor amplification.
A similar simultaneous amplification/attenuation phenomenon has been reported for SPP waves guided by the planar interface of a metal and a uniaxial dielectric material \c{ML_SPP_JO}.
Lastly, we turn to the case where
the relative permittivity parameters of material ${\cal B}$
have positive-valued real parts and positive-valued imaginary parts.
Specifically, let $\varepsilon^s_{\cal B} = 2 + i $ and $\varepsilon^t_{\cal B}= 3 + 1.5i$.
Thus, material ${\cal B}$ behaves like an anisotropic dielectric material that exhibits dissipation. The corresponding surface waves are akin to Dyakonov waves \c{Dyakonov88,Takayama_exp}.
Unlike the case represented in Fig.~\ref{Fig4}, the dispersion equation \r{DE2} here yields two, one or no $q$ solutions, depending upon the value of $\psi$.
The real and imaginary parts of $q$ are plotted against $\psi$ for all solutions in Fig.~\ref{Fig5}.
Specifically, we found two $q$ solutions for $23^\circ < \psi < 76^\circ$, one $q$ solution
for $14^\circ < \psi < 23^\circ$ and $76^\circ < \psi < 90^\circ$, and no $q$ solution at all for $0^\circ < \psi < 14^\circ$. The existence of two solutions signifies a marked difference from the case for conventional Dyakonov waves \c{ESW_book}.
In the case represented in Fig.~\ref{Fig5} all surface-wave solutions are attenuated regardless of the orientation angle $\psi$, which contrasts with
a simultaneous amplification/attenuation phenomenon that has been reported for Dyakonov waves guided by the planar interface of an isotropic dielectric material and a uniaxial dielectric material \c{ML_Dyakonov}. Also, the existence of two surface waves at each orientation for $23^\circ < \psi < 76^\circ$ in Fig.~\ref{Fig5} contrasts with the usual case for Dyakonov waves in which the dispersion equation admits only one solution for each propagation direction \c{DSWreview}.
\section{Closing remarks}\label{dcr}
Electromagnetic surface waves supported by an isotropic chiral material were investigated numerically via the associated canonical boundary-value problem. In the case where
the partnering material was an isotropic dielectric material, surface waves akin to SPP waves were excited;
while in the case where the partnering material was a uniaxial dielectric material,
surface waves akin to Dyakonov waves and SPP waves were excited. By selecting an isotropic chiral material capable of simultaneously supporting attenuation and amplification of plane waves, depending upon circular polarization state, the amplitudes of the SPP-like waves
could be either amplified or attenuated, depending upon the relative permittivity of the isotropic dielectric partnering material
or the direction of propagation relative to the optic axis of the uniaxial dielectric partnering material.
Therefore, wider opportunities for surface-wave propagation are supported by isotropic chiral materials, as compared to achiral materials in analogous scenarios.
\vspace{10mm}
\noindent {\bf Acknowledgment} JN is supported by a Deans' Vacation Scholarship (University of Edinburgh).
|
\section{Introduction}
Relativistic collision of heavy nuclei, by varying center of mass collision energy ({{$\sqrt{s_{_{\mathrm{NN}}}}$}}), allows us to map the quantum chromodynamics (QCD) phase diagram characterized by temperature $T$ and baryonic chemical potential $\mu_{B}$ ~\cite{Mohanty:2009vb, SN0493,SN0598, starNP10, starNP,starNC, starNK,BraunMunzinger:2007zz}. Lattice QCD calculations suggest a smooth crossover transition from hadronic matter (a state of confined quarks and gluons) to a deconfined state of quarks and gluons for $\mu_{B}=0$ at certain finite $T$~\cite{Aoki:2006we}. Also at larger value of $\mu_{B}$, theoretical QCD model predictions of a first-order phase boundary~\cite{Ejiri:2008xt, Bowman:2008kc} and existence of the QCD critical point~\cite{qcp, qcp1} are the motivations for carrying out high energy heavy-ion collision with varying beam energies. Higher order cumulants of conserved quantities such as the net-baryon, net-charge, and net-strangeness numbers, as a function of beam energy, are expected to show non-monotonic behaviour near the critical point~\cite{Gupta:2011wh,stephanovmom}. Near the QCD critical point, the third order cumulant (or skewness) and fourth order cumulant (or kurtosis) of net-baryon are expected to be negative valued~\cite{Asakawa:2009aj, Stephanov:2011pb} while the sixth order cumulant of both net-baryon and net-charge distribution is expected to be negative near a crossover phase transition~\cite{negc6:friman}. Using heavy-ion collider facility at RHIC, STAR experiment has performed measurements on higher order moments of net-proton~\cite{starNP10, starNP}, net-charge~\cite{starNC} and net-kaon~\cite{starNK} multiplicity distributions. PHENIX experiment at RHIC also has carried out measurements of moments of net-charge distributions~\cite{phenixNC}. At the Large Hadron Collider (LHC), ALICE experiment is also studying higher order moments to characterize the nature and order of the QCD phase transition~\cite{alice:fluct1, alice:meanpt, alice:np}.
Higher order cumulants of net-baryon and net-charge fluctuations are predicted to be sensitive to the nature of the QCD transition (crossover and chiral). In particular, ratios of the sixth to second and eighth to second order cumulants of the net-baryon number fluctuations change rapidly and the sixth order cumulants of both net-baryon and net-charge are predicted to remain negative if the system formed in such collisions freeze-out in the proximity of crossover region of the QCD phase diagram~\cite{negc6:friman, negc6:cheng}. The STAR experiment reported a preliminary result on centrality dependence of $C_{6}/C_{2}$ of net-proton and net-charge distribution in \mbox{Au$+$Au}\xspace collisions at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV~\cite{Esha:2017yeh}. $C_{6}/C_{2}$ of net-proton distribution exhibits $-ve$ values systematically across peripheral ($70-80\%$) to most central ($0-5\%$) collisions and net-charge results shows $-ve$ values only for most central ($0-5\%$) collisions. Further STAR experiment has reported preliminary results that reveal a non-monotonic variation of ratio of fourth order cumulant to second order cumulant of net-proton distributions ($C_{4}/C_{2}$) with beam energy~\cite{luo:NP}. This trend in measurement is similar to as expected, if the system traverses in the vicinity of a critical point in the QCD phase diagram. However due to finite size and time effects the signals corresponding to phase transition and critical point as discussed above are expected to be small. A high statistics run as part of the second phase of the Beam Energy Scan program at RHIC is envisaged to make a statistically more accurate measurement of $C_{4}/C_{2}$~\cite{SN0493, SN0598}. \\
In this work, via Monte Carlo simulation, we investigated the effect of limited statistics on the values of higher order cumulants up to $C_{7}$ within two statistical distribution models. In the first model, net-proton distribution is taken as a Skellam distribution (assuming proton and antiproton distributions are Poissonian) where the input parameters are mean ($C_{1}$) of experimentally measured proton and antiproton distributions from STAR experiment. In the second model, net-proton is obtained from the difference of binomially distributed protons and antiprotons, where the experimental value of mean ($C_{1}$) and variance ($C_{2}$) of proton and antiproton are used to characterize the input binomial distributions. We have performed this study for two top energies ({{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$ and $62.4$~GeV) of heavy-ion (Au) collision at RHIC, the energy region where signals of crossover transition are predicted to manifest. However the conclusions of the work remain valid for both higher and lower beam energies. In this study, we see that the estimated higher order cumulants, especially in the case of relatively small event sample size, randomly fluctuate around their true values, saturate and approach their true values with larger event sample size (quantitative description of smaller and larger sample size varies with different aspects of simulation such as input models, input parameters etc. and depends on order of cumulants, hence the discussions are left to the later sections). The value of $C_{6}$ of net-proton as a function of sample size in certain cases is observed to be $-ve$ in both the collision energies due to limited statistics of sample used and approaches the $+ve$ true value as the sample size is increased. The statistical errors which should reflect the consistency of estimated results with their true values, are also studied. We investigate the robustness of commonly followed methods for statistical error estimation such as delta theorem~\cite{adasgupta:book}, bootstrap method~\cite{bootstrap, bootstrap1}, and sub-group method (used in some experiments)~\cite{Abdelwahab:2014yha, Acharya:2017cpf}. Finally we obtain the exclusion curves for minimum statistics required to estimate the cumulants of net-proton distribution with a precision of $5\%$ and $10\%$ (relative to their true values) for various scenarios within the ambit of the models studied here.
We organize the content of this work as follows. In the following section, we describe mathematical prescription of relationships between moments and cumulants of various orders, along with discussions on models, input parameters, and cumulants calculation. A brief discussion on error estimation methods is also included in this section. In section-III, the results on sample size dependence of various order net-proton cumulants and studies on statistical error estimation using various methods are presented. Interpretation and discussion on the significance of these results in the context of precision measurement of higher order cumulants for the experimental search of phase transition and critical point signals are carried out in the Section-IV. Finally, in Section-V we summarize the findings of this study.
\section{Methodology}
We have performed Monte Carlo simulations to study the effect of limited statistics on higher order cumulants. In the simulation, two statistical models are used to generate net-proton distributions with experimental inputs from two collision energies. In the first case, net-proton distribution is generated from the difference of binomially distributed protons and antiprotons characterized by mean ($1^{st}$ order cumulant) and variance ($2^{nd}$ order cumulant) of proton and antiproton measured in STAR experiment~\cite{starNP}. The proton and antiproton distributions are independently (no correlations) generated. In the second case, net-proton follows a Skellam distribution and experimental values of mean of proton and antiproton are used as input parameters for the simulation.
\subsection{Cumulants Calculation}
The higher order fluctuations are often expressed in terms of cumulants or moments.
Cumulants and central moments are related to each other in the following way.
Let the deviation of $N$ (any entry in the data sample) from its mean value ($\mu=<N>$, referred as $1^{st}$ raw moment) be defined by
\begin{equation}
\delta N=N-<N>.
\end{equation}
Any $r^{th}$ order central moment is defined as:
\begin{equation}
\mu_r = \,< (\delta N)^r >,
\end{equation}
$i.e.$ $1^{st}$ order central moment turns out to be zero. $\mu _1=\,< (\delta N) >=0$.
The Cumulants of a given data sample could be written in terms of moments as follows.
\begin{eqnarray}
C_1 &=&\mu \\
C_2 &=&\mu_2 \\
C_3 &=& \mu_3 \\
C_n(n>3) &=& \mu_n - \sum\limits_{m = 2}^{n - 2}
{\left(
\begin{array}{l}
n - 1 \\
m - 1 \\
\end{array} \right)C_m } \mu_{n - m}
\end{eqnarray}
For two independent variables $X$ and $Y$, the cumulants of probability distribution of their sum ($X+Y$), are just addition of cumulants of individual distributions for $X$ and $Y$, $i.e.$ $C_{n,X+Y}=C_{n,X}+C_{n,Y}$ for $n^{th}$ order cumulant. For distribution of difference in $X$ and $Y$, even order cumulants are addition of the individual cumulants, while odd order cumulants are obtained by subtracting the individual cumulants. Explicit expressions for cumulants in terms of central moments are mentioned in Appendix~\ref{sec:appA}.
\subsubsection*{Binomial and Poissonian Model}
In heavy-ion collisions, several physics processes could lead to the correlated production of protons and antiprotons. For simplicity, we do not include any physics motivated correlations between proton and antiproton multiplicity distributions. In the first case, we take protons and antiprotons generated independently each following a Poisson distribution. The probability distribution of a Poisson distributed discrete random variable $X$ is given as follows.
\begin{equation}
p(k;\lambda)=\mathrm{Pr}(X=k)=\frac{\lambda^{k}e^{-\lambda}}{k!},
\end{equation}
where $\lambda$ is the mean, $e$ is the Euler's number. The resultant distribution of difference in two Poisson distributed variables is a Skellam distribution.
The probability mass function for the Skellam distribution is given by,
\begin{equation}
p(k;\mu_{1},\mu_{2})=e^{-(\mu_{1}+\mu_{2})}\,\Big(\,\frac{\mu_{1}}{\mu_{2}}\,\Big)^{\,k/2}\,\,I_{k} (2\sqrt{\mu_{1}\mu_{2}}),
\end{equation}
where $\mu_{1}$ and $\mu_{2}$ are the mean of Poisson distributions and $I_{k}(z)$ is the modified Bessel function of the first kind. This representation with above-specified inputs is referred to \emph{Poissonian} model in the text that follows. For simulation, experimental values of $C_{1}$ for proton and antiproton distributions are taken as the input parameters $\mu_{1}$ and $\mu_{2}$ respectively. \\ \\
Similarly, in the \emph{Binomial} model, we consider protons and antiprotons are independently generated each following a binomial distribution. The probability mass function of a binomially distributed random variable $k$ is given as follows.
\begin{equation}
B(k;n,p)=\frac{n!}{k!\,(n-k)!}\,p^{k}\,(1-p)^{k},
\end{equation}
where $n$ is the number of trials, $p$ is probability of success and $k$ is the number of sucesses.
Two parameters that characterize the binomial distribution are determined from relations such as $\mu=np$, $\sigma^{2}=np(1-p)$ where $\mu$ and $\sigma^{2}$ are the $1^{st}$ order ($C_{1}$) and $2^{nd}$ order cumulant ($C_{2}$) of proton or antiproton distributions measured in the experiment. Mathematical expressions for cumulants of Binomial and Skellam distributions are mentioned in Appendix~\ref{sec:appB}.
\\ \\
In this study experimental data used as inputs for simulation are (anti) proton cumulants in a kinematic region of $|y|<0.5$ and $0.4<p_{T}\mathrm{(GeV/c)}<0.8$ for most central (0-5\%) collisions measured by STAR collaboration~\cite{starNP}. Input parameters set from two energies $\sqrt{s_{\mathrm{NN}}}(\mathrm{GeV})=62.4, 200$ are considered for both the models. In all the cases studied, the maximum statistics is $10^{10}$ events. We study the effect of the size of an event sample on net-proton cumulants up to $7^{th}$ order for both the Binomial and Poisson models. The results and detailed discussion occur in the later sections.
\subsection{Error Estimation}
The higher order cumulants are sensitive to the shape of the distribution specially the tails, hence estimating the statistical uncertainties correctly is crucial. Further, in higher order cumulants analysis, the concern in the context of statistical error estimation has gone beyond the basic necessity of constraining an experimentally observed parameter in certain numerical range, to the affair of finding and testing of an accurate and reliable method for error estimation itself. We discuss here three such methods to estimate the statistical errors on cumulants: Delta theorem method~\cite{adasgupta:book, Luo:2011tp}, Bootstrap method~\cite{bootstrap, bootstrap1} and Sub-group method ~\cite{Abdelwahab:2014yha, Acharya:2017cpf}. Errors obtained are further subjected to a Monte Carlo verification procedure to test the reliability of the methods.
\subsubsection*{Delta theorem method}
Using delta theorem, a concise form of standard error propagation method, analytical formulae for statistical errors on cumulants and moments exist in the literature~\cite{Luo:2011tp, Luo:2014rea}. Delta theorem method has been quite extensively used in estimating statistical errors in the analysis of higher order cumulants in heavy-ion collision experiments~\cite{starNP10, starNP, starNC, starNK, phenixNC}. The error estimation method uses the Central Limit Theorem (CLT) - an important theorem on convergence of probability. One of the most common statements of the CLT is as follows.\\
\emph{Central Limit Theorem}: Suppose ${X_{1},X_{2}, .., X_{n}}$ is a collection of random variables that are independent and identically distributed with mean E$[X_{i}]=\mu$ and Var$[X_{i}]=\sigma^{2}<\infty$, then as $n$ approaches infinity, the random variable $\sqrt{n}(S_{n}-\mu)$ converges to a normal distribution $N(0,\sigma^{2})$, where $S_{n}=(X_{1}+X_{2}+..+X_{n})/n$. In other words, for large value of $n$, $S_{n}$ approximately follows a normal distribution $N(\mu,\sigma^{2}/n)$.\\ \\
Given the knowledge of the approximate distribution of a statistic itself, delta theorem gives a prescription to approximate the distribution of a transformation of the statistic in large samples~\cite{adasgupta:book}.
The statement of delta theorem is mentioned below.\\ \\
\textbf{Delta theorem :} Let $T_{n}$ be a sequence of statistics such that
\begin{equation}
\sqrt{n}(T_{n}-\theta)\,\, \lim_{\rightarrow} \,\, N(0,\sigma^{2}(\theta)), \, \, \sigma(\theta)>0. \\
\end{equation}
Let $g$ be a real function which atleast is differentiable at $\theta$ with $g^{'}(\theta)\neq 0$. Then \\
\begin{equation}
\sqrt{n}(g(T_{n})-g(\theta))\,\, \lim_{\rightarrow} \,\, N(0,[g^{'}(\theta)]^{2}\sigma^{2}(\theta)).
\end{equation}
Error on cumulants upto $8^{th}$ order expressed in terms of central moments are mentioned in Sec.~\ref{sec:appC} of Appendix~\ref{sec:appC}.
Here, it is important to note that, the estimation of error on moments calculated from a sample requires the knowledge of corresponding parameters ($e.g.$ $\sigma$ in the statement of CLT and $\theta$ in the definition of delta theorem) of the population, which is never within the scope of comprehension in the experiments. Hence the moments calculated from the sample, are used as estimators for corresponding parameters of the population.
\subsubsection*{Bootstrap Method}
The bootstrap method finds the error on the estimators in an efficient Monte Carlo way by forming bootstrap samples without involving the complexities of standard error propagation method. It makes use of random selection of elements with replacement from the original sample, to construct bootstrap samples over which the sampling variance of an estimator is calculated~\cite{bootstrap, bootstrap1}.\\ \\
To start with, let $X$ be a random sample representing the experimental dataset drawn randomly from an unknown parent distribution. Let $\hat{e}$ be the estimator of a statistic (such as mean or variance), on which we intend to find the standard error. The sequence of steps followed to estimate standard error using the bootstrap method is as follows. \\ \\
1. Given a parent sample of size $n$, construct $B$ number of independent bootstrap samples $X^{*}_{1}$, $X^{*}_{2}$, $X^{*}_{3}$, ..., $X^{*}_{B}$, each consisting of $n$ data points randomly drawn with replacement from the parent sample. \\
2. Evaluate the estimator in each bootstrap sample,
\begin{equation}
\hat{e}^{*}_{b}=\hat{e}(X^{*}_{b}) \qquad b=1,2,3, ..., B.
\end{equation}
3. The sampling variance of the estimator is given as follows.
\begin{equation}
Var(\hat{e}) = \frac{1}{B-1}\sum_{b=1}^{B}\Big(\hat{e}^{*}_{b} - \bar{\hat{e}}\Big)^{2},
\end{equation}
where $\bar{\hat{e}}^{*} = \frac{1}{B}\sum_{b=1}^{B}(\hat{e}^{*}_{b})$\\
The sufficient enough value of $B$ for an accurate estimation of error within the bootstrap method varies from case to case depending upon the initial sample size. However, in general, the larger value of $B$ estimates the error better. We find the error on cumulants using the bootstrap method and the method is subjected to verification procedure to test the robustness. To determine the value of $B$ and maximize stability of estimated error with respect to the number of bootstrap samples, we further employ a qualitative test as discussed in results section~\ref{sec:Results}.
\subsubsection*{Sub-group Method}
Random replication of samples could also be done by dividing the initial data sample into many groups. To be more precise, without replacement of entries, subgroup method divides the data sample into sub-samples, evaluates the estimator over each sub-sample and estimates the sampling variance on the estimator. We mention the steps involved in the sub-group method as follows. \\ \\
Given the data sample $X$ and estimator of interest be denoted by $\hat{e}$,
1. Construct $S$ number of sub-samples $X^{*}_{1}$, $X^{*}_{2}$, $X^{*}_{3}$, ..., $X^{*}_{S}$, each consisting of $n = \frac{N}{S}$ number of data points randomly drawn from the parent sample without replacement, where $N$ is the size of parent sample. \\
2. Calculate the value of estimator for each sub-sample,
\begin{equation}
\hat{e}^{*}_{s} = \hat{e}(X^{*}_{s}) \qquad s = 1,2,3, ..., S.
\end{equation}
3. Variance on the estimator is calculated as follows.
\begin{equation}
Var(\hat{e}) = \frac{1}{S(S-1)}\sum_{s=1}^{S}\Big(\hat{e}^{*}_{s} - \bar{\hat{e}}\Big)^{2},
\label{eq:subgrp_var}
\end{equation}
where $\bar{\hat{e}}^{*} = \frac{1}{S}\sum_{s=1}^{S}(\hat{e}^{*}_{s})$. The extra factor of $1/S$ in Eq.\ref{eq:subgrp_var} is used to scale back the variance on estimator from sub-sample level to original data sample level to account for the statistics mismatch.\\
We obtain the error on cumulants using the sub-group method for different sub-sample sizes and compared to the corresponding results from bootstrap and delta-theorem method. Gaussian behaviour of estimators from the sub-samples in sub-group method is tested by subjecting to a verification procedure.
\section{Results\label{sec:Results}}
\begin{figure*}[hbtp]
\centering
\includegraphics[scale=0.3]{bandplot_skellam_200gev_new.png}
\caption{(Color online) Cumulants of net-proton distribution upto $6^{th}$ order are shown as function of sample size. The net-proton distribution is a Skellam distribution with input parameters $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of mean ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Blue solid circle markers represent the calculated cumulant values from simulated net-proton distribution. The solid red line represents the true value of cumulants obtained using analytical formula for a Skellam distribution. The magenta band represents $\pm1\sigma$ statistical error (from delta theorem method) about the true value of cumulants.}
\label{fig:band_sklm_200G}
\end{figure*}
To find the effect of limited statistics in the measurement of higher order cumulants, we calculate cumulants of various order of simulated net-proton distributions within two models namely, Poissonian model and Binomial model. Statistical errors are estimated using delta theorem method, bootstrap method and sub-group method. We present the results on calculating back the cumulants values viz-a-viz the true values as a function of event statistics in the first part of this section and results on error estimation methods are investigated in the later part.
\subsubsection{\label{CumulantResults}Dependence of cumulants on sample size}
\begin{figure*}[hbtp]
\centering
\includegraphics[scale=0.3]{bandplot_skellam_62gev.png}
\caption{(Color online) Cumulants of net-proton distribution upto $6^{th}$ order are shown as function of sample size. The net-proton distribution is a Skellam distribution with input parameters $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Blue solid circle markers represent the calculated cumulant values from simulated net-proton distribution. The solid red line represents the true value of cumulants obtained using analytical formula for a Skellam distribution. The magenta band represents $\pm1\sigma$ statistical error (from delta theorem method) about the true value of cumulants.}
\label{fig:band_sklm_62G}
\end{figure*}
Net-proton following a Skellam distribution is generated with input parameters $\mu_{1}=5.66$ and $\mu_{2}=4.11$ which are efficiency corrected mean of proton and antiproton distributions respectively, in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV in STAR experiment~\cite{starNP}. As shown in Fig.~\ref{fig:band_sklm_200G}, values of cumulants (upto $7^{th}$ order) randomly fluctuate for the smaller value of sample statistics and quality of agreement with true values improves by increasing number of events in a sample. However, in most of the cases, as shown in Fig.~\ref{fig:skell_chisq_200G} the cumulants agree with their true values within $\pm 1\sigma$ statistical error while in very few cases cumulants fluctuate beyond $\pm 1\sigma$ error. To be precise, value of cumulants deviate in a range of $0.005\sigma$ to $2.1\sigma$. The fact that cumulants in some cases lie beyond $\pm 1\sigma$ error could be accounted to the small yet significant probability outside $\pm 1\sigma$ region of a Gaussian distribution. In few cases, $C_{5}$, $C_{6}$ and $C_{7}$ show negative values. Note that, sign of cumulants have physical significance in the field of studying QCD phase diagram. The statistical error is estimated using the delta theorem method. The true values of cumulants are calculated using the analytical formula for Skellam distribution. \\ \\
\begin{figure}[hbtp]
\includegraphics[scale=0.45]{skell_chisq_200_new.png}
\caption{(Color online) Sample size dependence of standard deviation (Modulus deviation of cumulants from their true values scaled with $1\sigma$ statistical error) of cumulants up to $7^{th}$ order in the Poissonian model with input parameters from proton and antiproton distribution in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV.}
\label{fig:skell_chisq_200G}
\end{figure}
The central values of $C_{6}$ for sample size $10^{5}$ events and $10^{7}$ events are found to be $-116.656$ and $-2.201$, which is interesting due to their $-ve$ signs. Though the theoretical prediction for a crossover/chiral phase transition suggests a $-ve$ value of $C_{6}$ to be a possible hint for such a transition~\cite{negc6:friman}, we find that the $-ve$ sign of $C_{6}$ (also $-ve$ sign of $C_{5}$ and $C_{7}$) in our simulation results is due to insufficient statistics of the sample set used. In some cases large values of $C_{4}$ are also observed. In the simulation, Skellam distribution of net-proton is generated using mean ($C_{1}$) of experimentally obtained proton and antiproton distributions. Hence, we would like to emphasize that in order to extract information related to unique physics features through the higher order cumulant studies, the first step is to ensure that the sample has sufficient statistics.
\begin{figure}[hbtp]
\includegraphics[scale=0.45]{skell_chisq_62.png}
\caption{(Color online) Sample size dependence of standard deviation (Modulus deviation of cumulants from their true values scaled with $1\sigma$ statistical error) of cumulants up to $7^{th}$ order in the Poissonian model with input parameters from proton and antiproton distribution in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV.}
\label{fig:skell_chisq_62G}
\end{figure}
We have performed similar study by taking inputs from another RHIC collision energy $i.e.$ for {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV~\cite{starNP} (where the difference in mean values of proton and antiproton are larger than that for $200$~GeV) as shown in Fig.~\ref{fig:band_sklm_62G}. As seen in the case of {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV, cumulants calculated in simulation randomly fluctuate around the true value and with increase in sample size better agreement between cumulants with their true values is observed. While the central values of $C_{6}$ and $C_{7}$ pick up $- ve$ values for sample sizes of both $10^{5}$ and $10^{7}$ events, the $\pm1\sigma$ statistical error makes the values consistent with true values. Figure ~\ref{fig:skell_chisq_62G} shows that cumulants lie in a range of $0.04\sigma$ to $2.5\sigma$.
\begin{figure*}[hbtp]
\centering
\includegraphics[scale=0.3]{bandplot_bino_200gev.png}
\caption{(Color online) Cumulants of net-proton distribution upto $6^{th}$ order are shown as function of sample size. The net-proton distribution is difference of binomially distributed proton and antiproton with input parameters $\mu_{1}$, $\sigma_{1}^{2}$ (for proton) and $\mu_{2}$, $\sigma_{2}^{2}$ (for antiproton) same as corresponding experimental value in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Blue solid circle markers represent the calculated cumulant values from simulated net-proton distribution. The solid red line represents the true value of cumulants obtained using analytical formula for a binomial distribution. The magenta band represents $\pm1\sigma$ statistical error (from delta theorem method) about the true value of cumulants.}
\label{fig:band_bino_200G}
\end{figure*}
\begin{figure*}[hbtp]
\centering
\includegraphics[scale=0.3]{bandplot_bino_62gev.png}
\caption{(Color online) Cumulants of net-proton distribution upto $7^{th}$ order are shown as function of sample size. The net-proton distribution is difference of binomially distributed proton and antiproton with input parameters $\mu_{1}$, $\sigma_{1}^{2}$ (for proton) and $\mu_{2}$, $\sigma_{2}^{2}$ (for antiproton) same as corresponding experimental value in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV. Blue solid circle markers represent the calculated cumulant values from simulated net-proton distribution. The solid red line represents the true value of cumulants obtained using analytical formula for a binomial distribution. The magenta band represents $\pm1\sigma$ statistical error (from delta theorem method) about the true value of cumulants.}
\label{fig:band_bino_62G}
\end{figure*}
In the binomial model, net-proton distribution is generated from difference of binomially distributed proton and antiproton where input parameters $\mu_{1}$, $\sigma_{1}^{2}$ is the efficiency corrected mean and variance of proton distribution and $\mu_{2}$, $\sigma_{2}^{2}$ is the efficiency corrected mean and variance of antiproton distribution obtained in the experiment. The results from Binomial model are shown in Fig.~\ref{fig:band_bino_200G} and ~\ref{fig:band_bino_62G}. For {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV we show results for only up to $6^{th}$ order cumulants\footnote{True value of $7^{th}$ order cumulant for {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV in Binomial model is 0.14. The relative error ($|\frac{C_{meas.}-C_{true}}{C_{true}}|$) takes up large value because of vanishingly small value of the denominator.}. For both {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV and $62.4$~GeV, qualitative behaviour of cumulants as a function of sample size is similar as seen in Poissonian model.
\begin{figure}[hbtp]
\includegraphics[scale=0.45]{bino_chisq_200.png}
\caption{(Color online) Sample size dependence of standard deviation (Modulus deviation of cumulants from their true values scaled with $1\sigma$ statistical error) of cumulants up to $6^{th}$ order in the binomial model with input parameters from proton and antiproton distribution in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV.}
\label{fig:bino_chisq_200G}
\end{figure}
In the binomial model, for {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV, $-ve$ values are found for $C_{4}$ (for sample size $10^{3}$), $C_{5}$ (for sample size $10^{4}$ and $10^{5}$) and $C_{6}$ (for sample size $10^{5}$). For {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV, $C_{6}$ shows $+ve$ values across different sample sizes in this model.
\begin{figure}[hbtp]
\includegraphics[scale=0.45]{bino_chisq_62.png}
\caption{(Color online) Sample size dependence of standard deviation (Modulus deviation of cumulants from their true values scaled with $1\sigma$ statistical error) of cumulants up to $7^{th}$ order in the binomial model with input parameters from proton and antiproton distribution in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV.}
\label{fig:bino_chisq_62G}
\end{figure}
Figure~\ref{fig:bino_chisq_200G} and ~\ref{fig:bino_chisq_62G} show cumulants agree with their true values with absolute standard deviation varying in a range of $0.07\sigma$ to $2.2\sigma$ for {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV and $0.04\sigma$ to $2.3\sigma$ for {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV, respectively. As pointed out in the text before, settling down on proper statistically sound sample size is very important before interpretation of magnitude or sign of experimentally measured the higher order cumulants.\\
As can be seen from the above results and discussions and the fact that any experiment in laboratory will always have a finite sample size, proper estimation of statistical error is very crucial. In next part of current section, we will investigate the results from the study of different methods used to estimate the statistical error on various order cumulants.
\subsubsection{\label{sec:ErrorResults}Comparison of methods of error estimation}
Figure~\ref{fig:compdelta_skellam200G_size6} shows the relative deviation of errors in percentage obtained from bootstrap and sub-group methods from those obtained in delta theorem for Skellam distributed net-proton with a sample size of $10^{6}$ events. The input parameters for Skellam distribution are taken from experimental data for most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. In bootstrap method errors are calculated for 3 different cases with $(i)\,100$, $(ii)\,1000$ and $(iii)\,10000$ numbers of bootstrap samples. Similarly, in the sub-group method errors are obtained with $(i)\,100$, $(ii)\,1000$ and $(iii)\,10000$ numbers of sub-samples. We observe in Fig.~\ref{fig:compdelta_skellam200G_size6} that the errors from sub-group and bootstrap methods are in good agreement with those from delta theorem for lower order cumulants. With increase in the order of cumulants, the relative percentage deviation also increases. It is important to note that the sub-group method with sub-samples $s=10000$ shows a maximum deviation of nearly $150$\% w.r.t. to delta theorem method. Errors obtained from the bootstrap method for $B=10000$ show close agreement with the error obtained from the delta theorem.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.45]{compare_delta_skellam200gevsize6.png}
\caption{(Color online) Relative percentage deviation of errors obtained in different methods (Sub-group and Bootstrap) from those obtained in delta theorem method is shown for different cumulants upto $8^{th}$ order for a Skellam distribution of sample size $10^{6}$ generated using input parameters $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Black triangle, red solid circle and magenta solid square markers correspond to cases of $(i)\,100$, $(ii)\,1000$ and $(iii)\,10000$ number of sub-samples respectively in the sub-group method. Open blue circle, green star and purple cross symbols represent bootstrap method with number of bootstrap samples corresponding to cases $(i)$,$(ii)$ and $(iii)$, respectively.}
\label{fig:compdelta_skellam200G_size6}
\end{figure}
\begin{figure*}[ht!]
\includegraphics[scale=0.27]{verify_error_skellam200gev_size6.png}
\caption{(Color online) Verification of Gaussian (probability distribution) behaviour of errors obtained in different methods in a sample of size $10^{6}$. Values of cumulants $C_{4}$, $C_{6}$ and $C_{8}$, shown in blue points are calculated for 10000 samples generated by employing bootstrap sampling method from a Skellam distributed parent sample (of size $10^{6}$) characterized by $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Red lines in left most column panel figures represent bootstrap error on cumulants in parent sample. Black lines in middle column figures represent delta method error on cumulants. Magenta lines in right most column figures represent error on cumulants from sub-group method.}
\label{fig:ver_err_skellam200G_size6}
\end{figure*}
\begin{figure*}[ht!]
\centering
\includegraphics[scale=0.28]{verify_error_skellam200gev_size8.png}
\caption{(Color online) Verification of Gaussian (probability distribution) behaviour of errors obtained in different methods in a sample of size $10^{8}$. Values of cumulants $C_{4}$, $C_{6}$ and $C_{8}$, shown in blue points are calculated for 10000 samples generated by employing bootstrap sampling method from a Skellam distributed parent sample (of size $10^{8}$) characterized by $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Red lines in left most column panel figures represent bootstrap error on cumulants in parent sample. Black lines in middle column figures represent delta method error on cumulants. Magenta lines in right most column figures represent error on cumulants from sub-group method.}
\label{fig:ver_err_skellam200G_size8}
\end{figure*}
To test the Gaussian (probability distribution) behaviour of errors, in Fig.~\ref{fig:ver_err_skellam200G_size6}, we show the values $C_{4}$, $C_{6}$ and $C_{8}$ calculated for $10000$ samples constructed from the same parent sample as used in Fig.~\ref{fig:compdelta_skellam200G_size6}. The replicated samples (each of size $10^{6}$), used in this study are constructed by random selection of entries from parent sample with replacement. We find that within $\pm1\sigma$ of bootstrap error $68$\% of the sampling distribution lies for all the cumulants shown. Delta theorem method also satisfies the Gaussian behaviour as $67.9$\%, $69.2$\% and $68.6$\% of the total number of samples lie within $\pm1\sigma$ delta method error for $C_{4}$, $C_{6}$ and $C_{8}$ respectively. For $C_{6}$ and $C_{8}$ sub-group method is seen to over estimate the errors.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.45]{compare_delta_skellam200gevsize8.png}
\caption{(Color online) Relative percentage deviation of errors obtained in different methods (Sub-group and Bootstrap) from those obtained in delta theorem method is shown for different cumulants upto $8^{th}$ order for a Skellam distribution of sample size $10^{8}$ generated using input parameters $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Black triangle, red solid circle and magenta solid square markers correspond to cases of $(i)\,100$, $(ii)\,1000$ and $(iii)\,10000$ number of sub-samples respectively in the sub-group method. Open blue circle, green star and purple cross symbols represent bootstrap method with number of bootstrap samples corresponding to cases $(i)$,$(ii)$ and $(iii)$, respectively.}
\label{fig:compdelta_skellam200G_size8}
\end{figure}
Figure~\ref{fig:compdelta_skellam200G_size8} shows the relative percentage deviation of errors from sub-group and bootstrap methods from those obtained using delta theorem method for different order of cumulants (up to $8^{th}$ order) in Skellam distributed net-proton sample of size $10^{8}$ events. Only by changing the parent sample size to $10^{8}$, we observe an overall better agreement between errors from different methods compared to the case with a parent sample size of $10^{6}$ events. In the current case, the sub-group method with the number of sub-samples, $s=10000$ shows better agreement (agrees nearly within $1$\%) with the delta theorem. The bootstrap method with $B=10000$ also shows agreement consistently within $1$\% with delta theorem method across all order of cumulants shown. Values $C_{4}$, $C_{6}$ and $C_{8}$ calculated for $10000$ samples with $\pm1\sigma$ errors on them calculated from different methods are shown in Fig.~\ref{fig:ver_err_skellam200G_size8}. For all order cumulants shown, $\pm1\sigma$ error from bootstrap and delta theorem method include $68$\% of sampling distributions of cumulants. In this case, out of $10000$ samples $63$\% for $C_{4}$, $58$\% for $C_{6}$ and $65$\% for $C_{8}$ fall within $\pm1\sigma$ error from sub-group method. From above studies, we observe that sub-group method does not satisfy the Gaussian behaviour of cumulants across sampling distributions.
\section{\label{sec:discussions}Discussions}
In the context of measurement of higher order cumulants, we discuss below few points on the results mentioned in the previous section.
Cumulant values in both Poissonian and binomial model with input parameters taken from measured event-by-event distributions of proton and antiprotons for Au+Au collisions at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$ and $62.4$~GeV show statistical randomness as a function of sample size, saturates and approaches their true values towards the large value of sample size.
Higher order cumulants which characterize the subtle details of the shape of a statistical distribution are predicted to carry signals of phase transition of matter created in heavy-ion collisions. The $-ve$ sign of $C_{6}$ which is attributed to the physics of nature of phase transition in QCD phase diagram, is observed in our study for a statistically generated distribution where none of the above physics processes are included. Hence, this observation could only be attributed to the cumulants obtained from a sample with insufficient number of events. Hence obtaining a sample with sufficient statistics is very crucial for the experiments looking for signals of phase transition and critical point. \\
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.3]{skellam200exclusion.png}
\caption{(Color online) Exclusion limits on event statistics for measurement of cumulants of net-proton (Skellam distributed). Net-proton distribution is characterized by $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Case - $(i)$: Red bottom line corresponds to minimum statistics needed for performing measurement of various order cumulants up to $7^{th}$ order in 10\% relative deviation from the true value. Case - $(ii)$: Black top line corresponds to the 5\% relative deviation from the true value. In the measurement of cumulants, filled regions below red and black lines are excluded for requiring better event statistics as described in cases $(i)$ and $(ii)$ respectively.}
\label{fig:skellam200G_exclusion}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.3]{skellam62exclusion.png}
\caption{(Color online) Exclusion limits on event statistics for measurement of cumulants of net-proton (Skellam distributed). Net-proton distribution is characterized by $\mu_{1}$ (and $\mu_{2}$) taken from experimental value of means ($C_{1}$) of proton (and antiproton) distributions in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV. Case - $(i)$: Red bottom line corresponds to minimum statistics needed for performing measurement of various order cumulants up to $7^{th}$ order in 10\% relative deviation from the true value. Case - $(ii)$: Black top line corresponds to the 5\% relative deviation from the true value. In the measurement of cumulants, filled regions below red and black lines are excluded for requiring better event statistics as described in cases $(i)$ and $(ii)$ respectively.}
\label{fig:skellam62G_exclusion}
\end{figure}
In order to guide the experiments on what could be sufficient statistics using our model (Poissonian and Binomial) based study, we have obtained the exclusion limits on statistics of data sample for precise measurement of cumulants with two different input parameters. We assume that the signal (due to phase transition or QCD critical point) to be at the level of 5\% or 10\% above the statistical level for each order of cumulant (for simplicity). Figure~\ref{fig:skellam200G_exclusion} shows the exclusion curves on sample statistics for various order cumulants in the Poissonian model with input parameters fixed from measured proton and antiproton distributions in Au+Au collisions at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. The curves (black and red) represent the minimum statistics needed for measurement of cumulants with certain required precision ($5$\% and $10$\% relative deviation from their true values). This also translates into the minimum statistics needed to observe a signal of strength of 5\% or 10\% above the statistical baseline. The statistics regions lying above the curves are desirable for measurement of cumulants for a particular order, while regions below the curve are excluded by the requirement of the above-mentioned precision. For example, estimation of $C_{4}$ within an accuracy of $5$\% and $10$\% require a minimum statistics of $1.2$ million and $0.3$ million events respectively. Similarly for $C_{6}$, accuracy of $5$\% and $10$\% need minimum statistics of $1.5$ billion and $588$ million events respectively. In Fig.~\ref{fig:skellam62G_exclusion}, we show similar exclusion curve for Poissonian model with input parameters of {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV. For measurement of $C_{4}$, $1.6$ million (for $5$\%) and $0.4$ million (for $10$\%) minimum event statistics is needed. For $C_{6}$, accuracy of $5$\% and $10$\% require a minimum statistics of $1.5$ billion and $684$ million events.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.3]{bino200exclusion.png}
\caption{(Color online) Exclusion limits on event statistics for measurement of cumulants of net-proton. The net-proton distribution is difference of binomially distributed proton and antiproton with input parameters $\mu_{1}$, $\sigma_{1}^{2}$ (for proton) and $\mu_{2}$, $\sigma_{2}^{2}$ (for antiproton) same as corresponding experimental values in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV. Case - $(i)$: Red bottom line corresponds to minimum statistics needed for performing measurement of various order cumulants up to $6^{th}$ order in 10\% relative deviation from the true value. Case - $(ii)$: Black top line corresponds to the 5\% relative deviation from the true value. In the measurement of cumulants, filled regions below red and black lines are excluded for requiring better event statistics as described in cases $(i)$ and $(ii)$ respectively.}
\label{fig:bino200G_exclusion}
\end{figure}
Figure~\ref{fig:bino200G_exclusion} and Fig.~\ref{fig:bino62G_exclusion} show the exclusion curves for minimum statistics required for precision measurement of cumulants in a binomial model for input parameters for both {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$ and $62.4$~GeV. In the case of {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=200$~GeV for measurement of $C_{4}$ event sample with size of $1.3$ million (for $5$\% accuracy) and $0.3$ million (for $10$\% accuracy) and for measurement of $C_{6}$ event sample with size of $1.8$ billion (for $5$\% accuracy) and $552$ million (for $10$\% accuracy) are required. Similarly in the case of {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV, measurement of $C_{4}$ requires event sample with size of $1.2$ million (for $5$\% accuracy) and $0.2$ million (for $10$\% accuracy) and measurement of $C_{6}$ requires sample size of $3.6$ billion (for $5$\% accuracy) and $1.6$ billion (for $10$\% accuracy). We observe that, for higher ($C_{5}$ and higher order) cumulants, the binomial model needs larger statistics compared to Poissonian model to measure cumulants within similar degrees of precision. This is important to notice that, the required event statistics reported above is for a single centrality bin (say $e.g.$ 0-5\%). So, in one collision energy for precise measurement of higher order cumulants such as $C_{4}$ and $C_{6}$, the total number of events (in full centrality classes) would be at least $1$ order of magnitude higher than those quoted above for a single centrality bin. This size of statistics, in fact, is very large compared to the currently available experimental data statistics.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.3]{bino62exclusion.png}
\caption{(Color online) Exclusion limits on event statistics for measurement of cumulants of net-proton. The net-proton distribution is difference of binomially distributed proton and antiproton with input parameters $\mu_{1}$, $\sigma_{1}^{2}$ (for proton) and $\mu_{2}$, $\sigma_{2}^{2}$ (for antiproton) same as corresponding experimental values in most central (0-5\%) Au$+$Au collision at {{$\sqrt{s_{_{\mathrm{NN}}}}$}}$=62.4$~GeV. Case - $(i)$: Red bottom line corresponds to minimum statistics needed for performing measurement of various order cumulants up to $7^{th}$ order in 10\% relative deviation from the true value. Case - $(ii)$: Black top line corresponds to the 5\% relative deviation from the true value. In the measurement of cumulants, filled regions below red and black lines are excluded for requiring better event statistics as described in cases $(i)$ and $(ii)$ respectively.}
\label{fig:bino62G_exclusion}
\end{figure}
\\Though the statistics of data sample required are very large for measuring higher cumulants within $5$\% and $10$\% of their true values, within $\pm1\sigma$ statistical error most of the cumulants agree with their true values for the events statistics available for current experimental data. However, a certain degree of caution should be taken when interpreting the energy dependence, centrality dependence and sign of the higher order cumulants. Till the experiments run for longer period and accumulate enough statistics for these measurements, one should obtain the statistical errors carefully and look for systematic variations across centrality/energy for physical interpretation.
\section{\label{sec:summary}Summary}
Higher order cumulants of distributions of conserved quantities are predicted to carry signals of phase transitions. The higher order cumulants are sensitive to the shape of the distributions and their values are driven by tails of the distributions. Hence the biggest challenges in such measurements are two fold (a) limited statistics and (b) correct estimation of statistical errors. We tried to address both these issues in this work. We study via a Monte Carlo simulation method, the effect of limited data statistics on the accuracy of the estimation of different order cumulants in two different models and for two different experimental input conditions. We observe that the estimated values of the higher order cumulants have a strong dependence on sample size and central values for cumulants randomly fluctuate around their true values at low event statistics, saturate and approach their true values for the large size of event sample. $C_{6}$ of net-proton distribution is predicted to show negative values if the system formed in heavy-ion collisions freezes-out close to a crossover transition. In this simulation, which has no phase transition effects incorporated, $C_{6}$ of the net-proton distribution (characterized only by lowest order cumulants of proton and antiproton distribution from experiment) shows the $-ve$ sign for some values of sample statistics. The positive point however is that, irrespective of the sample size the estimated values lies within $\pm1\sigma$ statistical error (obtained using Delta theorem method) on the cumulants from their true values, in most of the cases (for rest few cases agreements are observed within $\pm2.5\sigma$). In this work, we attribute this behaviour ($-ve$ sign of $C_{6}$) to purely statistical effect because simulated net-proton distribution does not include any physics of phase transitions. We have discussed the methods for accurate estimation of statistical error. We have studied three different methods namely, delta theorem method, the bootstrap method and sub-group method. We have verified the Gaussian probability distribution nature of the errors. We find bootstrap method is a robust method to estimate statistical error followed by the method using the delta theorem. Finally we have presented the exclusion limits on the minimum statistics of data sample needed for estimation of cumulants of different order in order to detect a signal (related to phase transition or critical point) of $5$\% and $10$\% above the statistical baseline using different models and input parameters.
\section*{\textbf{ACKNOWLEDGEMENTS}}
We thank Sourendu Gupta for useful discussions on error estimation methods for statistical distributions. We also would like to thank Nu Xu and Zhangbu Xu for helpful discussions on experimental aspects related to the measurement of higher cumulants with limited statistics. B.M. acknowledges the financial support from J C Bose National Fellowship of DST, Government of India. D.M. and A.P. acknowledge the financial support from Department of Atomic Energy, Government of India.
|
\section{Introduction}\label{SK1}
The question of how eigenvalues change when the domain is slightly perturbed is a classical problem
probably going at least as far back as Rayleigh~\cite{Rayleigh1877}, who studied eigenvalues and domain perturbation
in connection with acoustics as early as in the nineteenth century.
The approach given in this article owes to results by Hadamard~\cite{Hadamard1968}, who in the early 20th century
studied perturbations of domains with smooth boundary,
where the perturbed domain~$\Omega_{\epsilon}$ is represented by~$x_{\nu} = \epsilon h(x')$
where~$x' \in \partial \Omega_0$,~$x_{\nu}$ is the signed distance to the boundary ($x_{\nu} < 0$ for~$x \in \Omega_0$),~$h$ is a smooth
function,~$\epsilon$ is a small parameter, and~$\Omega_0$ is the reference domain. Hadamard's result
for the first eigenvalue of the Dirichlet-Laplacian is given by
\begin{equation}\label{A4a}
\Lambda(\Omega_{\epsilon}) = \Lambda(\Omega_0) -
\epsilon \int_{\partial \Omega_0} h |\nabla \vp|^2 dS
+ o(\epsilon),
\end{equation}
where $dS$ is the surface measure on~$\partial \Omega_0$ and~$\vp$ is an eigenfunction corresponding
to~$\Lambda(\Omega_0)$ such that~$\| \vp \|_{L^2(\Omega_0)} = 1$.
It is worth noting, that the problem of how eigenvalues change when the domain is perturbed, is closely
related to shape optimization. We refer here
to Henrot~\cite{Henrot2006}, and Soko\l{}owski and Zol\'esio~\cite{Sokolowski1992}, and references found therein.
The aim of this article is to find minimal assumptions on the smoothness of the boundary when the
Hadamard formula is still valid.
A large quantity of studies of the Dirichlet problem already exists in the literature; see, for instance,
Grinfeld~\cite{Grinfeld2010}, Henrot~\cite{Henrot2006}, Kozlov~\cite{Kozlov2013,Kozlov2006},
Kozlov and Nazarov~\cite{Kozlov2012}, and references found therein. In this article,
we present an asymptotic formula of Hadamard type for perturbations in the case when the
domains are of class~$C^1$ or~$C^{1,\alpha}$, respectively.
The first class of domains is optimal for validity of the Hadamard formula. For the second class of domains,
we give an optimal estimate of the remainder term.
Let~$\Oe$ and~$\Ot$ be bounded domains in~$\mathbf{R}^n$ with boundaries~$\Ge$ and~$\Gt$, respectively.
We consider the spectral problems
\begin{equation}
\label{eq:maineq}
\left\{
\begin{aligned}
-\Delta u &= \Llm[{}] u & & \mbox{in } \Oe,\\
u &= 0 & & \mbox{on } \Ge
\end{aligned} \right.
\end{equation}
and
\begin{equation}
\label{eq:maineq2}
\left\{
\begin{aligned}
-\Delta v &= \Lum[{}] v & & \mbox{in } \Ot,\\
v &= 0 & & \mbox{on } \Gt.
\end{aligned} \right.
\end{equation}
In the case when the boundary is nonsmooth, we consider the corresponding weak formulation of the problem
on the Hilbert spaces~$H^1_0(\Oe)$ and~$H^1_0(\Ot)$ with the usual inner product.
Note though, that the techniques used are applicable to a wider class of partial differential operators.
In particular to uniformly elliptic operators of second order.
It is known that if the two domains are close enough, both problems have the same number of
eigenvalues in a small enough neighborhood of~$\Llm[{}]$ as the multiplicity of the eigenvalue.
This means that for a fixed eigenvalue~$\Llm[{}]$ of~(\ref{eq:maineq}) of multiplicity~$m$,
there are precisely~$m$ eigenvalues~$\Lum[{}]$ of~(\ref{eq:maineq2}) (counting multiplicity) near~$\Llm[{}]$.
This is a consequence of
the continuous dependence of eigenvalues on the domain;
see, e.g., Kato~\cite{Kato1966} (Sections~IV.3 and~V.3) or Henrot~\cite{Henrot2006} and references therein.
The explicit result in terms of quantities used in this article can be found in Kozlov~\cite{Kozlov2013} (Proposition~3).
We will denote by~$\Xj[k] \subset H^1_0(\Oe)$ the eigenspace corresponding to the eigenvalue~$\Llm[k]$ and denote
the dimension of~$\Xj[k]$ by~$\Jm[k]$.
For our results, we will characterize how close the two domains are in the sense
of the Hausdorff distance between the sets~$\Oe$ and~$\Ot$, i.e.,
\begin{equation}
\label{eq:d_def}
d = \max \{ \sup_{x \in \Oe} \inf_{y \in \Ot} |x-y|, \; \sup_{y \in \Ot} \inf_{x \in \Oe} |x-y| \}.
\end{equation}
We do not assume that one domain is a subdomain of the other. It should be noted however,
that the abstract result presented below in Section~\ref{s:prelim} permits a more general type of proximity quantity
for the two domains.
We consider three cases of regularity of the boundary~$\Ge$, namely $C^{1,\alpha}$, $C^1$,
and Lipschitz boundaries. Let us first consider the Lipschitz case.
Then there exists a positive constant~$M$ such that the boundary $\Ge$ can be covered by
a finite number of cylinders $\Ck$,~$k=1,2,\ldots,N$, where there exists orthogonal coordinate systems in which
\[
\Ck \cap \Oe= \Ck \cap \{ y = (y', y_n) \,:\, y_n > h_k^{(1)}(y')\},
\]
where the center of $\Ck$ is at the origin
and
\[
\Ck = \{ (y',y_n)\,:\, y' \in B'_{r_k}(0) ,\, |y_n| < \delta_k \}.
\]
Here,~$B'_{k}=B'_{r_k}(0)$ is the $(n-1)$-dimensional ball of radius $r_k$ and with the center $0$.
We assume that~$h_k^{(1)}(0) = 0$ and that~$h_k^{(1)}$ are Lipschitz functions, i.e.,
\[
|h_k^{(1)}(y') - h_k^{(1)}(z')| \leq M |y' - z'|.
\]
This class of domains defined by a constant~$M$ and cylinders~$\Ck$,~$k=1,2,\ldots,N$, will be denoted in what follows by
${\mathcal L}={\mathcal L}(M,\C_1,\ldots,\C_N)$.
We assume that~$\Ot$ is close to~$\Oe$ in the sense that~$\Ot \in {\mathcal L}(M,\C_1,\ldots,\C_N)$ can be described by
\[
\Ck \cap \Ot = \Ck \cap \{ y = (y', y_n) \,:\, y_n > h_k^{(2)}(y')\},
\]
where~$h_k^{(2)}$ are also Lipschitz continuous with Lipschitz constant~$M$ and
\[
\widehat{d} = \max_{k=1,2,\ldots,N} \sup \{ | h_k^{(1)}(y') - h_k^{(2)}(y')| \,:\, y'\in B'_k \}
\]
is assumed to be small. One can show that there exists positive constants~$c_1$ and~$c_2$,
such that~$c_1 \widehat{d} \leq d \leq c_2 \widehat{d}$.
The case when $\Oe$ is a $C^1$- or $C^{1,\alpha}$-domain is defined analogously, with the following additional
assumptions.
\noindent {\bf $\bm{C^1}$-assumption.} We assume that $h_k^{(1)} \in C^{1}(\overline{B'_k})$ such that
\begin{equation}\label{K2a}
h_k^{(1)}(0) = \partial_{x_i} h_k^{(1)}(0) = 0, \quad i=1,2,\ldots,n-1,
\end{equation}
and
\begin{equation}
\label{eq:grad_hk_c1}
| \nabla ( h_k^{(1)} - h_k^{(2)}) | = o(1), \quad \mbox{as } d \rightarrow 0.
\end{equation}
\noindent {\bf $\bm{C^{1,\alpha}}$-assumption.} We assume that~$h_k^{(1)} \in C^{1,\alpha}(\overline{B'_k})$
and that~(\ref{K2a}) holds. Furthermore,
\begin{equation}
\label{eq:grad_hk_holder}
| \nabla ( h_k^{(1)} - h_k^{(2)}) | \leq C d^{\alpha}.
\end{equation}
Note that $h_k^{(2)}$ are only assumed to be Lipschitz continuous in both cases and satisfy~(\ref{eq:grad_hk_c1})
or~(\ref{eq:grad_hk_holder}), respectively.
Let us define the function~$\sigma$ on the surface $\Ge$ in the case of the $C^1$-assumption on the boundary.
Let $\nu = \nu(P)$ be the unit outward normal to the boundary~$\Ge$ at the point $P$.
For $P \in \Ge$ we introduce the number $\sigma = \sigma(P)$. It is the smallest (in the absolute value) root
of the equation
\begin{equation}
\label{KK1a}
P + \sigma \nu = Q \in \Gt,
\end{equation}
where~$Q$ is the nearest point on~$\Gt$ to~$P$ lying on the line passing through~$P$ with the direction~$\nu$.
Clearly~$\sigma$ is positive if~$Q$ is outside~$\Oe$ and negative if~$Q$ is inside.
One can verify that this function is Lipschitz continuous on~$\Ge$ with Lipschitz constant
depending on~$M$ and~$\C_1,\ldots,\C_N$.
\begin{theorem}
\label{i:t:holder_boundary}
Suppose that~$\Oe$ is a~$C^{1,\alpha}$-domain with~$0 < \alpha < 1$ and that~$\Ot$ is as described above.
Then
\begin{equation}
\label{i:eq:holder_main_boundary}
\begin{aligned}
\Lum[k] - \Llm[m] = {} &
\kk + O(d^{1+\alpha})
\end{aligned}
\end{equation}
for every~$k = 1,2,\ldots,\Jm[m]$.
Here~$\kk[{}] = \kk$ is an eigenvalue of the problem
\begin{equation}
\label{eq:T2_tk_intro}
\kk[{}] \int_{\Oe} \nabla \vp \cdot \nabla \vs \, dx =
\int_{\Ge} \sigma \,
\nabla \vp \cdot \nabla \vs
\, dS
\quad \mbox{for all } \vs \in \Xj[m],
\end{equation}
where~$\vp \in \Xj[m]$. Moreover, $\kk[1],\kk[2],\ldots,\kk[\Jm]$ in~{\rm(}\ref{i:eq:holder_main_boundary}{\rm)}
run through all eigenvalues of~{\rm(}\ref{eq:T2_tk_intro}{\rm)} counting their multiplicities.
\end{theorem}
\noindent
Observe that~(\ref{eq:T2_tk_intro}) can be phrased as a spectral problem on the Hilbert space~$\Xj[m]$ by using the
Riesz representation theorem of the operator on the right-hand side. Here $dS$ is the surface measure.
\begin{theorem}
\label{i:t:C1_boundary}
Suppose that~$\Oe$ is a~$C^{1}$-domain, and~$\Ot$ is as described above.
Then
\begin{equation}
\label{i:eq:C1_main_boundary}
\begin{aligned}
\Lambda_k(\Ot) - \Lambda_m(\Oe) = \kappa_k + o(d)
\end{aligned}
\end{equation}
for every~$k = 1,2,\ldots,J_m$.
\end{theorem}
\noindent
If~$\Oe$ and~$\Ot$ are two Lipschitz domains, then we can prove the estimate
\begin{equation}
\label{JiJ}
|\Lambda_k(\Ot) - \Lambda_m(\Oe)| \leq C d
\end{equation}
for every~$k = 1,2,\ldots,J_m$.
We note that if~$u$ satisfies (\ref{eq:maineq}) for a Lipschitz domain~$\Oe$,
then the normal derivative~$\partial_{\nu} u|_{\Ge} \in L^2(\Ge)$ and so the integral in the
right-hand side of~(\ref{eq:T2_tk_intro}) is well defined. Despite this fact,
we have demonstrated in Section~\ref{SA4a} that the formula presented in~(\ref{i:eq:holder_main_boundary})
does not hold for Lipschitz domains and their small perturbations. Indeed, we show in Section~\ref{SA4a}
that the remainder~$O(d^{1+\alpha})$ is optimal for~$C^{1,\alpha}$-perturbations.
The result similar to Theorem \ref{i:t:holder_boundary}
for the Neumann problem is proved in~\cite{Kozlov2016} and an analogue of Theorem \ref{i:t:C1_boundary}
for the Neumann problem, albeit with a more complicated expression for the leading term~$\kappa_k$, is
derived in~\cite{Thim2015}.
\bigskip
The paper is organized as follows. In Section~\ref{s:prelim} we present an abstract approach
developed in \cite{Kozlov2013}, which concerns the asymptotics of eigenvalues of unbounded
operators when their domains are changed. As a result of this approach, a theorem on asymptotics
of eigenvalues is presented. It involves different terms and these terms are estimated in the remaining
parts of the paper. The analysis is based on the theory of elliptic boundary value problems in
Lipschitz, $C^1$- and~$C^{1,\alpha}$-domains that was developed in particular in the works of Dahlberg,
Fabes, Jodeit, Kenig and Rivi\'{e}re. In the last section, we present an example which demonstrates the
sharpness of Theorems~\ref{i:t:holder_boundary} and~\ref{i:t:C1_boundary}.
\section{Preliminary Results and Definitions}
\label{s:prelim}
Here we present an abstract result from \cite{Kozlov2013}, which will play an important
role in the proofs of Theorems~\ref{i:t:holder_boundary} and~\ref{i:t:C1_boundary}.
We suppose that~$\Oe$ and~$\Ot$ are open Lipschitz domains and that~$\Dd$ is an open ball such
that~$\Oe \subset \Dd$ and~$\Ot \subset \Dd$. Put~$\He = H^1_0(\Oe)$,~$\Ht = H^1_0(\Ot)$ and finally~$\H = H^1_0(\Dd)$.
We extend functions by zero outside their respective domains.
We let~$\ip{\cdot}{\cdot}$ and~$\Lip{\cdot}{\cdot}$ denote the inner products on~$\H$ given by
\[
\ip{u}{v} = \int_{\Dd} \nabla u \cdot \nabla v \, dx
\qquad \mbox{and} \qquad
\Lip{u}{v} = \int_{\Dd} u \, v \, dx,
\]
respectively. Moreover, let~$\| \, \cdot \, \|$ and~$| \, \cdot \, |$ be the norms induced by~$\ip{\cdot}{\cdot}$
and~$\Lip{\cdot}{\cdot}$, respectively. We will consider the spectral problems
\begin{equation}
\label{eq:maineig}
\ip{\vp}{v} = \lambda \Lip{\vp}{v} \quad \mbox{for every~$v \in \He$}
\end{equation}
and
\begin{equation}
\label{eq:perturbeig}
\ip{U}{V} = \mu \Lip{U}{V} \quad \mbox{for every~$V \in \Ht$.}
\end{equation}
We enumerate the eigenvalues~$\Llm[k] = \lm[k]$,~$k=1,2,\ldots$, of~(\ref{eq:maineq}) increasingly
according to~$0 < \lm[1] < \lm[2] < \cdots$.
Let~$\Xj[k] \subset \He$ be the eigenspace corresponding to the eigenvalue~$\lm[k]$ and denote
the dimension of~$\Xj[k]$ by~$\Jm[k]$.
In this article we study eigenvalues of~(\ref{eq:perturbeig}) located in a neighborhood of~$\lm$,
where~$m$ is fixed.
Note that it is known that there are precisely~$\Jm[m]$ eigenvalues of~(\ref{eq:maineq2}) near~$\lm$;
see, e.g., Lemma~3.1 in~\cite{Kozlov2013}.
We denote these eigenvalues by~$\um[1],\um[2],\ldots,\um[{\Jm[m]}]$.
Let~$\Sj \colon \H \rightarrow \Hj$ be the orthogonal projectors with image~$\Hj$ for~$j=1,2$.
For simplicity, we denote~$\St$ by~$\S$. Furthermore, put~$\T v = v - \S v$ for~$v \in \He$.
We will employ results from \cite{Kozlov2013}, where an abstract framework for domain dependence
of Dirichlet eigenvalues was presented. To measure the closeness of two domains, the norm difference between
the projectors~$\Se$ and~$\St$ is used:
\begin{equation}
\label{eq:sigma}
| (\Se - \St) u |^2 \leq \eps \| u \|^2, \quad u \in \H.
\end{equation}
The best constant~$\eps$ measures the proximity of the
spaces~$\He$ and~$\Ht$, and therefore the closeness of~$\Oe$ and~$\Ot$.
One can show that~$\eps \approx d$ in our case.
To formulate an abstract asymptotic result, we will use the solution~$\Vp \in \Ht$ to the problem
\begin{equation}
\label{eq:def_Vp}
\ip{\Vp}{w} = \ip{\vp}{w} - \lm \Lip{\vp}{w} \quad \mbox{ for every~$w \in \Ht$}.
\end{equation}
Theorem~1 in \cite{Kozlov2013} can be reformulated as follows.
\begin{theorem}
\label{t:kozlov2013}
The asymptotic formula
\begin{equation}\label{M28s}
\um[k]^{-1} = \lm[m]^{-1} + \tk + O(\rr + |\tk|\eps), \quad k=1,2,\ldots,J_m,
\end{equation}
holds, where
\begin{equation}
\label{eq:rho}
\rr = \max_{\vp \in \Xj[m], \; \| \vp \| = 1} \bigl( |\T \vp |^2 + |\Vp|^2 + \eps \| \Vp \|^2 \bigr).
\end{equation}
Here~$\tau = \tk$ is an eigenvalue of the problem
\begin{equation}
\label{eq:main_asymp_D}
\lm^{-1} \bigl( \ip{\Vp}{\Vs} - \ip{\T\vp}{\T\vs} - \ip{\Vp}{\vs} - \ip{\vp}{\Vs} \bigr) = \tau \ip{\S \vp}{\S \vs},
\end{equation}
where~$\vp \in \Xj[m]$ and~$\tau_1,\tau_2,\ldots,\tau_{j_m}$ run through all eigenvalues
of~{\rm(}\ref{eq:main_asymp_D}{\rm)} counted with multiplicity.
\end{theorem}
\section{Estimates for the functions $\varphi$, $\Psi_\varphi$ and $T\varphi$}
\label{s:def_lipdom}
In this section we give estimates and some representations for the terms appearing in (\ref{eq:rho})
and (\ref{eq:main_asymp_D}). All of them are valid for Lipschitz domains.
We will use the following notation.
Let~$\Om$ be a Lipschitz domain.
The truncated cones~$\Gamma(x')$ at~$x' \in \partial \Om$ are given by, e.g.,
\[
\Gamma(x') = \{ x \in \Om \sa |x - x'| < 2 \mbox{dist}(x, \partial \Om) \}
\]
and the non-tangential maximal function is defined on the boundary~$\partial \Om$ by
\[
\N(u)(x') = \max_{k=1,2,\ldots,N} \sup \{ |u(x)| \sa x \in \Gamma(x') \cap {\mathcal C}_k \}
\]
and
$$
\N(\nabla u)(x') = \max_{k=1,2,\ldots,N} \sup \{ |\nabla u(x)| \sa x \in \Gamma(x') \cap {\mathcal C}_k \}
$$
We refer to Kenig~\cite{Kenig1994} for further details.
\noindent We will use the short-hand notation~$\nablat u$ for the tangential gradient.
From the definition of $\Vp$ in~(\ref{eq:def_Vp}), it follows that $\Vp$ is harmonic in
the domains~$\Oe \cap \Ot$ and~$\Ot \setminus \overline{\Oe}$. We will use the representation
\begin{equation}
\label{eq:rep_Vp}
\Vp = \vp - \Rvp,
\end{equation}
where~$\Rvp \in \Ht$ solves the equation~$\ip{\Rvp}{w} = \lm \Lip{\vp}{w}$ for all~$w \in \Ht$, or, equivalently,
\[
-\Delta \Rvp =\lm \vp \;\;\mbox{in $\Ot$}\;\;\mbox{and}\;\; \Rvp = 0\;\;\mbox{on $\Gt$.}
\]
We represent $\Rvp$ as $\Rvp = \Rvpo + \Rvpt$, where
\[
-\Delta \Rvpo = \lm \vp \;\;\mbox{in $\Dd$}\;\;\mbox{and}\;\; \Rvpo = 0 \;\;\mbox{on $\partial \Dd$}
\]
and
\[
\Delta \Rvpt = 0 \;\;\mbox{in $\Ot$}\;\;\mbox{and}\;\; \Rvpt = -\Rvpo \;\;\mbox{on $\Gt$.}
\]
Since $\vp \in H^1(\Dd)$ it follows that $\Rvpo \in H^3(\Dd)$ and
\begin{equation}
\label{M7a}
\| \Rvpo \|_{H^3(\Dd)} \leq C \| \vp \|_{H^1(\Dd)}.
\end{equation}
Moreover, from, e.g., Theorems 1 and~3 in Jerison and Kenig~\cite{jerisonkenig81N}, we obtain that
\begin{equation}
\label{eq:N_Rvpo}
\int_{\Gt} \left( | N(\Rvpt) |^2 + | N(\nabla \Rvpt) |^2 \right) dS \leq
C \| \Rvpo \|_{H^{1}(\Gt)}^2 \leq C\|\vp\|_{H^1(\Dd)}^2.
\end{equation}
From the representation (\ref{eq:rep_Vp}) one can see that the normal derivative of the
function~$\Vp$ has a jump on $\Ge$ since the function $\Rvp$ is of class $H^3$ in a
neighborhood of~$\Ge$.
\newcommand{\Lb}{\ensuremath{L^2}}
To make the notation more compact, we let
\[
\Oet = \Oe \cap \Ot, \;\; \Get = \partial \Oet \;\; \mbox{and} \;\; \Lb = L^2(\Get).
\]
Then the boundary~$\Get$ is given by the relation~$y_n=h^{(12)}_k(y')$ in each cylinder~${\mathcal C}_k$,
where~$h^{(12)}_k(y')=\max (h^{(1)}_k(y'),h^{(2)}_k(y'))$. One can verify that this function
also satisfies~(\ref{eq:grad_hk_c1}) in the~$C^1$-case and~(\ref{eq:grad_hk_holder}) in
the~$C^{1,\alpha}$-case. In what follows we shall use the short-hand notation~$h_1$,~$h_2$ and~$h_{12}$
for the functions~$h^{(1)}_k$,~$h^{(2)}_k$ and~$h^{(12)}_k$, respectively.
Let~$\nu = \nu(P)$ denote the outwards unit normal vector at the point~$P$ to the boundary $\Ge$ ($\Gt$).
The normal derivative~$\ndiv u$ at~$P \in \Ge$ ($P \in \Gt$) is defined by~$\ndiv u = \nabla u \cdot \nu$.
We can now formulate the following lemmas concerning the inner products in~(\ref{eq:main_asymp_D}).
\begin{lemma}
\label{l:VpVs}
Let~$\Oe$ and~$\Ot$ be Lipschitz domains from~${\mathcal L}(M,\C_1,\ldots,\C_N)$.
Then
\begin{equation}
\label{eq:est_VpVs}
\left| \ip{\Vp}{\Vs} - \int_{\Ot \setminus \Oe} \nabla \Rvp \cdot \nabla \Rvs \, dx \right|
\leq
C \| \Psi_\varphi \|_{\Lb}
\| \nabla_\tau \Psi_\psi \|_{\Lb}
\end{equation}
for all~$\vp,\vs \in \Xj[m]$.
\end{lemma}
\begin{proof}
Since $\varphi$ and $\psi$ vanish outside $\Oe$, we have
\[
\ip{\Vp}{\Vs}
= \int_{\Ot \setminus \Oe} \nabla \Rvp \cdot \nabla \Rvs \, dx
+ \int_{\Oet} \nabla \Vp \cdot \nabla \Vs \, dx.
\]
Using the relation
\[
\int_{\Oet} \nabla \Vp \cdot \nabla \Vs \, dx
=
\int_{\Get} \Vp \ndiv \Vs \, dS
\]
and the inequality
\begin{equation}
\label{M19a}
\| \ndiv \Vs \|_{\Lb}
\leq C
\| \nablat \Vs \|_{\Lb},
\end{equation}
we obtain that
\[
\left| \int_{\Oet} \nabla \Vp \cdot \nabla \Vs \, dx \right|
\leq C \| \Vp \|_{\Lb} \| \nablat \Vs \|_{\Lb},
\]
which in turn shows that~(\ref{eq:est_VpVs}) holds.
\end{proof}
\begin{lemma}
\label{l:vpVs}
Suppose that~$\Oe$ and~$\Ot$ are the same domains as in Lemma \ref{l:VpVs}.
Then
\begin{equation}
\label{eq:est_vpVs}
\bigl| \ip{\vp}{\Vs} \bigr| \leq C
\| \vp \|_{\Lb}
\| \nablat \Vs \|_{\Lb}
\end{equation}
for all~$\vp,\vs \in \Xj[m]$.
\end{lemma}
\begin{proof}
Since $\Vs$ is harmonic in $\Oet$, we have
\[
\ip{\vp}{\Vs} = \int_{\Get} \vp \ndiv \Vs \, dS.
\]
Using (\ref{M19a}) we arrive at (\ref{eq:est_vpVs}).
\end{proof}
\noindent Note that Lemma~\ref{l:vpVs} immediately implies that
\begin{equation}
\label{eq:est_Vpvs}
\bigl| \ip{\Vp}{\vs} \bigr| \leq C
\| \vs \|_{\Lb}
\| \nablat \Vp \|_{\Lb}
\end{equation}
for all~$\vp,\vs \in \Xj[m]$.
\begin{lemma}
\label{l:est_T}
Suppose that~$\Oe$ and~$\Ot$ are the same domains as in Lemma \ref{l:VpVs}.
Then
\begin{equation}
\label{eq:est_T}
\left| \ip{\T\vp}{\T\vs} - \int_{\Oe \setminus \Ot} \nabla \vp \cdot \nabla \vs \, dx \right|
\leq C \| \vs \|_{\Lb} \| \nablat \vp \|_{\Lb}
\end{equation}
for all~$\vp,\vs \in \Xj[m]$.
\end{lemma}
\begin{proof}
Let~$\Pvp = \T \vp = \vp - \S \vp$ and $\Psi = T\vs$. Since $\S \vp \in H_2$ and $\Phi$ is orthogonal to~$\Ht$,
we have $\Phi = \vp$ in $\Oe \setminus \Ot$ and that~$\Phi$ is harmonic in $\Oet$. Therefore
\[
\int_{\Oe} \nabla \Phi \cdot \nabla \Psi \, dx =
\int_{\Oe \setminus \Ot} \nabla\Phi \cdot \nabla \Psi \, dx
+ \int_{\Oet} \nabla\Phi \cdot \nabla \Psi \, dx.
\]
Since
\begin{equation*}
\int_{\Oet} \nabla \Pvp \cdot \nabla \Psi \, dx =
\int_{\Get} \Psi \ndiv \Pvp \, dS,
\end{equation*}
we can use~(\ref{M19a}), which implies that~(\ref{eq:est_T}) holds.
\end{proof}
\begin{lemma}
\label{LM12a}
Let $U\in H^1(\Oet)$ be harmonic in $\Oet$. Then
\begin{equation}\label{M8}
\int_{\Oet}|U|^2dx\leq C\int_{\Get}|U|^2dS.
\end{equation}
\end{lemma}
\begin{proof} We start from the estimate
\begin{equation}
\label{M8a}
\int_{\Get}N(U)^2 \, dS \leq C \int_{\Get} |U|^2 \, dS,
\end{equation}
which is true because $U$ is harmonic in $\Oet$. Let
\[
\Omega^\delta=\{ x\in \Oet\; :\; \dist (x,\Get)<\delta\}.
\]
Now, using the Caccioppoli-inequality
\begin{equation*}
\int_{\Oet \setminus \Omega^{\delta}} |\nabla U|^2 \, dx
\leq C \delta^{-2} \int_{\Omega^{\delta}} |U|^2 \, dx,
\end{equation*}
we obtain that
\begin{equation}
\label{M8b}
\int_{\Oet \setminus \Omega^{2\delta}} |U|^2 \, dx
\leq C \delta^{-2} \int_{\Omega^{2\delta}} |U|^2 \, dx,
\end{equation}
which together with the fact that the integral in the right-hand side of~(\ref{M8b}) is estimated by a constant times
the left-hand side of~(\ref{M8a}), provided that~$\delta$ is sufficiently small independently of~$d$, proves
that inequality~(\ref{M8}) holds.
\end{proof}
Let us now prove the assertion from Section~\ref{s:prelim} that we can choose~$\eps$ in~(\ref{eq:sigma})
as the small parameter~$\eps = d$.
\begin{lemma}
\label{l:close}
Let~\/$\Oe$ and~\/$\Ot$ be Lipschitz domains from ${\mathcal L}(M,\C_1,\ldots,\C_N)$.
Then
\begin{equation}
\label{eq:close}
\int_{\Dd} |\Se u - \St u|^2 \, dx \leq C d \, \| u \|_{H^1(\Dd)}.
\end{equation}
for all~$u \in H^1_0(\Dd)$.
\end{lemma}
\begin{proof}
Let~$w = \Se u - \St u$. Then~$w \in H_0^1(D)$ vanishes outside~$\Oe \cup \Ot$ and satisfies~$\Delta w = 0$ in~$\Oet$.
We start by proving that
\begin{equation}
\label{eq:est_w_cap}
\int_{\Oet} |w|^2 \, dx \leq C d \, \| w \|_{H^1(\Dd)}^2.
\end{equation}
By Lemma~\ref{LM12a}, the left-hand side of~(\ref{eq:est_w_cap}) is bounded by the squared~$L^2$-norm
of~$w$ along the boundary~$\Get$. Since~$w$ vanishes outside~$\Oe \cup \Oe$,
it is clear that
\[
|w(x)| \leq \int_{g_{12}}^{h_{12}} |\partial_s w(x',s)| \, ds, \quad x \in \Get,
\]
where~$h_{12} = \max(h_1,h_2)$ and~$g_{12} = \min(h_1,h_2)$. Applying H\"older's inequality, we obtain that
\[
w^2 \leq C d \int_{g_{12}}^{h_{12}} |\nabla w(x',s)|^2 \, ds,
\]
which implies that~(\ref{eq:est_w_cap}) holds.
Let us now prove an inequality of Poincar\'{e}-Friedrichs type:
\begin{equation}
\label{eq:PF_w}
\int_{A} w^2 \, dx \leq C d^2 \int_{A} |\nabla w|^2 \, dx,
\end{equation}
where~$A = (\Oe \setminus \Ot) \cup (\Ot \setminus \Oe)$.
Indeed, since~$w = 0$ outside~$\Oe \cup \Ot$, we obtain that
\[
\begin{aligned}
\int_{g_{12}}^{h_{12}} w^2(x', x_n) \, dx_n &=
2 \int_{g_{12}}^{h_{12}} \int_{g_{12}}^{x_n} w(x', s) \partial_s w(x',s) ds \, dx_n \\
&\leq C \int_{g_{12}}^{h_{12}} \left( \int_{g_{12}}^{x_n} w(x', s)^2 \, ds\right)^{1/2}
\left( \int_{g_{12}}^{x_n} |\nabla w(x',s)|^2 \, ds \right)^{1/2} \, dx_n \\
&\leq C d \left( \int_{g_{12}}^{h_{12}} w(x', s)^2 \, ds\right)^{1/2}
\left( \int_{g_{12}}^{h_{12}} |\nabla w(x',s)|^2 \, ds \right)^{1/2}, \\
\end{aligned}
\]
where we used H\"older's inequality and the fact that~$|h_{12} - g_{12}| \leq C d$. This implies that
\[
\int_{B_{r_k}'(0)} \int_{g_{12}}^{h_{12}} w^2(x', x_n) \, dx_n dx' \leq
C d^2 \int_{B_{r_k}'(0)} \int_{g_{12}}^{h_{12}} |\nabla w(x', x_n)|^2 \, dx_n \, dx',
\]
which proves that~(\ref{eq:PF_w}) holds.
\end{proof}
\begin{lemma}
\label{L21a}
Let~$\vp \in \Xj[m]$ with~$\| \vp \|_{H^1(\Oe)} = 1$. Then
\begin{equation}
\label{M20}
\int_{\Get} \vp^2 dS \leq C d^2\;\;\mbox{and}\;\; \int_{\Get} |\nabla\vp|^2 dS \leq C.
\end{equation}
Moreover,
\begin{equation}
\label{M21}
\int_{\Oe \setminus \Ot} |\nabla\vp|^2 dx \leq C d.
\end{equation}
\end{lemma}
\begin{proof}
We represent~$\vp$ as~$\varphi = U + V$, where
\[
-\Delta U = \lm \vp \;\; \mbox{in $\Dd$ and} \;\; U = 0 \;\; \mbox{on $\partial \Dd$}
\]
and
\[
\Delta V = 0 \;\; \mbox{on $\Oe$ and} \;\; V = -U \;\; \mbox{on $\Ge$.}
\]
Since $\vp \in H^1(\Oe)$, we have that~$U \in H^3(D)$ and that its norm is bounded
by~$C\|\vp\|_{H^1(\Oe)}$. Using an embedding theorem for anisotropic spaces
(see Section~10 of~\cite{BIN}), we get~$U \in L^{\infty,{\bf 2}}({\mathcal C}_k)$
and~$|\nabla U|\in L^{\infty,{\bf 2}}({\mathcal C}_k)$,
where~${\bf 2}=(2,\ldots,2)$ is an~$(n-1)$-vector, and~${\mathcal C}_k$, $k=1,\ldots,N$, are
the same cylinders as in Section~\ref{SK1}. In the local coordinates $(y',s)$,
this means that
\begin{equation}
\label{M20a}
\int_{B'_{r_k}(0)} \sup_{|s| < \delta_k} |U(y',s)|^2 \, dy' < \infty
\end{equation}
and
\begin{equation}
\label{M20b}
\int_{B'_{r_k}(0)} \sup_{|s| < \delta_k} |\nabla U(y',s)|^2 \, dy' < \infty.
\end{equation}
Moreover, the above integrals are estimated by~$C\| \vp \|^2_{H^1(\Oe)}$.
Furthermore, since~$V$ and~$\nablat V$ belong to $L^2(\partial\Oe)$, we have that
\begin{equation}
\label{M20c}
\int_{\Ge} (N(V)^2 + N(\nabla V)^2) dS \;\; \mbox{is bounded and $\leq C\| \vp \|^2_{H^1(\Oe)}$}.
\end{equation}
Inequalities~(\ref{M20b}) and~(\ref{M20c}) imply
\begin{equation}
\label{M21a}
\int_{B'_{r_k}(0)} \sup_{|s| < \delta_k} |\nabla \vp(y',s)|^2 \, dy' < \infty.
\end{equation}
Therefore
\[
\begin{aligned}
\int_{B'_{r_k}(0)} \vp^2(y',h_{12}(y')) \, dy'
&= 2\int_{B'_{r_k}(0)} \int_{h_1}^{h_{12}} |\partial_s \vp(y',s)| \, |\vp(y',s)| ds \, dy'\\
&\leq C d^2 \int_{B'_{r_k}(0)} \sup_{|s| < \delta_k} |\nabla \vp(y',s)|^2 \, dy'
\end{aligned}
\]
and the right hand side is bounded by~$Cd^2 \| \vp \|_{H^1(\Oe)}$ because of~(\ref{M20b})
and~(\ref{M20c}). This proves the first inequality in~(\ref{M20}).
The second inequality in~(\ref{M20}) and~(\ref{M21}) follows from~(\ref{M21a}).
\end{proof}
\noindent The next lemma asserts that similar estimates are valid for the function $\Rvp$.
\begin{lemma}
\label{L21b}
Let~$\vp \in \Xj[m]$ with~$\| \vp \|_{H^1(\Oe)} = 1$. Then
\begin{equation}
\label{M20k}
\int_{\Get} \Rvp^2 \, dS \leq Cd^2 \;\; \mbox{and} \;\; \int_{\Get}|\nabla \Rvp|^2 \, dS \leq C.
\end{equation}
Moreover,
\begin{equation}
\label{M21l}
\int_{\Ot \setminus \Oe} |\nabla \Rvp|^2 \, dx \leq C d.
\end{equation}
\end{lemma}
\begin{proof}
We represent $\Rvp$ as $\Rvp = U + V$, where~$U$ is the same as in the proof of the previous lemma, and
\[
\Delta V=0 \;\; \mbox{in $\Ot$ and} \;\; V = -U \;\; \mbox{on $\Gt$.}
\]
The remaining part of the proof is analogous with the proof of Lemma~\ref{L21a}.
\end{proof}
\begin{corollary} Let~$\vp \in \Xj[m]$ with~$\| \vp \|_{L^2(\Oe)} = 1$. Then
\begin{equation}
\label{M12c}
\rho=O(d^2)\,.
\end{equation}
\end{corollary}
\begin{proof}
Since $\T\vp$ is harmonic in~$\Oet$, Lemma~\ref{LM12a} implies that
\[
\| \T \vp \|_{L^2(\Oet)}^2 \leq C \int_{\Get} \vp^2 \, dS,
\]
where we used the relation~$\T\vp = \vp$ in~$\Oe \setminus \Ot$.
By~(\ref{M20}) we get
\begin{equation}
\label{M21aa}
\| \T \vp \|_{L^2(\Oet)}^2 \leq C d^2.
\end{equation}
From (\ref{M21}) we derive
\[
\int_{\Oe \setminus \Ot} \vp^2 \, dx = O(d^3),
\]
which together with~(\ref{M21aa}) gives
\begin{equation}
\label{M21ab}
\| \T \vp \|_{L^2(\Oe)}^2 \leq C d^2.
\end{equation}
Since $\Vp = \vp - \Rvp$, using Lemmas~\ref{L21a} and~\ref{L21b}, we get
\begin{equation}
\label{M21ac}
\| \Vp \|_{L^2(\Ot)}^2 \leq C d^2,
\end{equation}
and from~(\ref{eq:est_VpVs}) and Lemma~\ref{L21b}, we obtain
\begin{equation}
\label{QQq}
\int_{\Ot}|\nabla \Vp|^2 \, dx \leq Cd.
\end{equation}
By the definition of~$\rr$ in~(\ref{eq:rho}), estimates~(\ref{M21ab}),~(\ref{M21ac}) and~(\ref{QQq}),
and finally Lemma~\ref{l:close}, we obtain that~(\ref{M12c}) holds.
\end{proof}
Another consequence of the above analysis is the following
estimate for the eigenvalues for Lipschitz domains.
\begin{corollary} Let~$\Oe$ and $\Ot$ be Lipschitz domains from~${\mathcal L}(M,\C_1,\ldots,\C_N)$.
Then estimate {\rm (\ref{JiJ})} holds.
\end{corollary}
\begin{proof}
By~(\ref{eq:est_vpVs}) and~(\ref{eq:est_Vpvs}) combined with the first formula in~(\ref{M20})
and the last formula in~(\ref{M20k}), we get
\[
|\ip{\vp}{\Vs}| + |\ip{\vs}{\Vp}| = O(d).
\]
From~(\ref{eq:est_T}) and~(\ref{M20}), it follows that
\[
|\ip{\T\vp}{\T\vs}| = O(d)
\]
and
\[
|\ip{\S\vp}{\S\vs} - \ip{\vp}{\vs}| = O(d).
\]
Taking into account~(\ref{QQq}), we get $\tk = O(d)$,
which together with~(\ref{M12c}) leads to~(\ref{JiJ}).
\end{proof}
\section{Two lemmas}
In the case that the boundary~$\Ge$ of~$\Oe$ is of class~$C^1$ (or~$C^{1,\alpha}$), it is more convenient to use
another equivalent (for small $d$) description of the closeness of~$\Oe$ and~$\Ot$:
there exists a positive constant~$\delta_0$ such that for every~$P \in \Ge$,
in a cartesian coordinate system~$y=(y',y_n)$ with the center at~$P$ and the tangent plane to
the boundary~$\Ge$ at~$P$ given by $y_n = 0$, the domain is given by
\[
\Oe \cap {\mathcal C}_{\delta_0,\delta_0}(P) =
\{ y \in {\mathcal C}_{\delta_0,\delta_0}(P) \,:\, y_n > h_1(y'), \, y' \in B'_{\delta_0}(P) \},
\]
where
\begin{equation}
\label{M31a}
{\mathcal C}_{\delta,\sigma} = \{ y \, : \, y' \in B'_{\delta}(P), \, |y_n| < \sigma \}.
\end{equation}
If $\sigma=\delta$ we shall use the notation
${\mathcal C}_{\delta}$ for ${\mathcal C}_{\delta,\delta}$. The function $h_1 \in C^1$ satisfies
\begin{equation}\label{M23a}
h_1(0) = 0 \;\; \mbox{and} \;\; \nabla_{y'}h_1(0) = 0.
\end{equation}
The analogous representation is valid for $\Ot$ (certainly in this case we do not assume (\ref{M23a})) and
\begin{equation}
\label{M23b}
\sup_{B'_{\delta_0}(P)} |h_2 - h_1| = O(d) \;\; \mbox{and} \;\;
\sup_{B'_{\delta_0}(P)} |\nabla_{y'}(h_2 - h_1)| = o(1).
\end{equation}
In the case of a~$C^{1,\alpha}$-perturbation, we assume that~$h_1$ is of class $C^{1,\alpha}$ and the
relations in~(\ref{M23b}) are replaced by
\begin{equation}
\label{M23c}
\sup_{B'_{\delta_0}(P)} |h_2 - h_1| = O(d) \;\; \mbox{and} \;\;
\sup_{B'_{\delta_0}(P)} |\nabla_{y'}(h_2 - h_1)| = O(d^\alpha).
\end{equation}
As before we use the notation $h_{12}=\max (h_1,h_2)$, which defines the domain~$\Oet$ in the local
coordinates in~${\mathcal C}_{\delta_0,\delta_0}(P)$.
In what follows we shall use the following embedding assertion.
If~$u\in H^1(\Dd)$ and the cylinder~${\mathcal C}_{\delta,\sigma}$ given by~(\ref{M31a})
in a certain cartesian coordinate system belongs to~$\Dd$, then, by Theorem~10.2 in~\cite{BIN},
\begin{equation}
\label{M31b}
\int_{B'_\delta} \sup_{|y_n| < \sigma} |u(y',y_n)|^2 \, dy' \leq C \| u \|^2_{H^1(\Dd)}.
\end{equation}
\begin{lemma}
\label{l:Phi}
Let~$\Oe$ and~$\Ot$ satisfy the~$C^1$-assumption.
Then the solution to~$\Delta U = f$ in~$\Oe$ and~$U = g$ on~$\Ge$,~$f\in L^2(\Oe)$ and~$g \in L^2(\Ge)$, satisfies
\begin{equation}
\label{eq:Phi}
\begin{aligned}
\int_{\Get \cap {\mathcal C}_\delta(P) } |U|^2 dS {} & \leq
C \Big( \int_{\Ge \cap {\mathcal C}_{\delta}(P)} |g|^2 \, dS
+\int_{\Oe\cap {\mathcal C}_{\delta}(P)} |f|^2 \, dx\Big)\\
&+ C d \Big(\int_{\Oe} |f|^2 \, dx+\int_{\Ge} |g|^2 \, dS\Big),
\end{aligned}
\end{equation}
where $P \in\Ge$, $\delta \leq \delta_0$ and~$C$ is a constant independent of~$\delta$,~$f$ and~$g$.
\end{lemma}
\begin{proof}
Let $f_\delta = f$ in~${\mathcal C}_{\delta}$ and $0$ otherwise and similarly, let~$g_\delta=g$
on~$\Ge \cap {\mathcal C}_{\delta}$ and zero on the remaining part of the boundary.
Let~$U_\delta$ solve the problem~$\Delta U_\delta = f_\delta$ in~$\Oe$ and~$U_\delta = g_\delta$ on~$\Ge$.
We represent~$U_\delta$ as $U_\delta = U_\delta^1 + U_\delta^2$,
where~$\Delta U_\delta^1 =f_\delta$ in~$\Dd$, $U_\delta^1 = 0$ on $\partial \Dd$
and~$\Delta U_\delta^2 = 0$ in~$\Oe$, $U_\delta = g_\delta - U_\delta^1$ on $\Ge$. Then
\begin{equation}
\label{M4a}
U_\delta^1 \in H^2(\Dd) \;\; \mbox{and} \;\; \| U_\delta^1 \|_{H^2(\Dd)} \leq C\| f_\delta \|_{L^2(\Oe)}.
\end{equation}
Then, according to, e.g.,~\cite{Kenig1994},~$U_\delta^2$ satisfies
\begin{equation}
\label{M4b}
\int_{\Ge} |N(U_\delta^2)|^2 \, dS \leq C \Big( \int_{\Ge } |g_\delta|^2 \, dS + \int_{\Oe} |f_\delta|^2 \, dx \Big)
\end{equation}
and it follows from~(\ref{M4a}) and~(\ref{M4b}) that
\begin{equation}
\label{eq:N_Phz_est}
\int_{\Get} |U_\delta|^2 \, dS \leq
C \Big(\int_{\Ge} |g_\delta|^2 \, dS + \int_{\Oe} |f_\delta|^2 \, dx \Big).
\end{equation}
Let~$U_r$ be the solution to~$ \Delta U_r = f - f_\delta$ in~$\Oe$
and~$U_r = g - g_\delta$ on~$\Ge$.
Since~$U_r = 0$ on~$\Ge \cap {\mathcal C}_{\delta}$, we have (compare with the proof of Lemma \ref{L21a})
\begin{equation}
\label{eq:N_Phr_est}
\begin{aligned}
\int_{\Get \cap {\mathcal C}_{\delta}} | U_r |^2 \, dS {} & \leq
C \int_{\Ge \cap {\mathcal C}_{\delta}} \int_{h_1}^{h_{12}}
|\nabla U_r(x',s)| \, |U_r(x',s)| \, ds \, dS(x')\\
{} & \leq C d \int_{\Ge \cap {\mathcal C}_{\delta}} \int_{h_1}^{h_{12}} |\nabla U_r(x',s)|^2 \, ds \, dS(x').
\end{aligned}
\end{equation}
This implies that
\[
\int_{\Get \cap {\mathcal C}_{\delta}} | U_r |^2 \, dS \leq C d \int_{\Oe \setminus \Ot} |\nabla U_r|^2 \, dx.
\]
Now applying the estimate
\[
\int_{\Oe} |\nabla U_r |^2 \, dx \leq C \Big( \int_{\Oe} |f|^2 \, dx + \int_{\Ge} |g|^2 \, dS(x') \Big)
\]
and using~(\ref{eq:N_Phz_est}), we arrive at~(\ref{eq:Phi}).
\end{proof}
The estimates in Lemmas~\ref{l:VpVs}, \ref{l:vpVs} and \ref{l:est_T} contain the
terms~$\nablat \Vp$ and~$\nablat \vp$.
In the next lemma we present estimates for such forms in terms of the small parameter~$d$.
\begin{lemma}
\label{LM28a}
Let~$\vp \in \Xj[m]$ with~$||\varphi||_{H^1(\Oe)}=1$. We have
\begin{equation}
\label{M25a}
\int_{\Get} |\nablat \vp|^2 \, dS = o(1) \;\; \mbox{and} \;\; \int_{\Get} |\nablat \Vp|^2 \, dS = o(1)
\end{equation}
in the~$C^1$-case and
\begin{equation}
\label{M25b}
\int_{\Get} |\nablat \vp|^2 \, dS = O(d^{2\alpha}) \;\; \mbox{and} \;\; \int_{\Get} |\nablat \Vp|^2 \, dS = O(d^{2\alpha})
\end{equation}
in the~$C^{1,\alpha}$-case.
\end{lemma}
\begin{proof}
Consider a cylinder~${\mathcal C}_{\delta}(P)$ for a certain~$\delta=O(d)$ and~$P \in \Ge$.
Then the directions $y_j$,~$j=1,\ldots,n-1$, are tangent to~$\Ge$ at~$P$.
We choose a direction~$y_j$ for a certain~$j=1,\ldots,n-1$, and put~$u_j=\partial_{y_j} \vp$.
Then
\[
-\Delta u_j = \lm u_j \;\; \mbox{in $\Oe$ and} \;\; u_j = g := \partial_{y_j}\vp \;\; \mbox{on $\Ge$.}
\]
The function~$g$ belongs to~$L^2(\partial\Oe)$ and since~$\partial_{y_j} \vp(y',h_1(y')) = 0$, we have
\[
g(y',h_1(y')) = -\partial_{y_n}\vp(y',h_1(y')) \partial_{y_j} h_1(y').
\]
Using that $\partial_{y_n}\vp(y',h_1(y')) \in L^2(\Ge)$, we obtain that
\[
\int_{B'_\delta} g^2 \, dy' = o(1) \int_{B'_\delta} |\nabla\vp|^2 \, dS.
\]
Noting that
\[
\int_{\Oe \cap {\mathcal C}_{\delta}(P)} \vp^2 \, dx \leq
Cd^2 \int_{\Oe \cap {\mathcal C}_{\delta}} |\nabla\vp|^2 \, dx
\]
and applying Lemma \ref{l:Phi}, we get
\[
\int_{\Get \cap {\mathcal C}_{\delta}(P)} u_j^2 \, dS = o(1) \int_{B'_\delta} |\nabla\vp|^2 \, dS
+ O\Big( d^2 \int_{\Oe \cap {\mathcal C}_{\delta}} |\nabla\vp|^2 \, dx \Big).
\]
This implies the first inequality in (\ref{M25a}).
In the~$C^{1,\alpha}$-case, we have
\[
\int_{B'_\delta} g^2 \, dy' \leq C \delta^{2\alpha} \int_{B'_\delta} |\nabla\vp|^2 \, dS.
\]
Repeating the above proof for the~$C^1$-case, we arrive at the first inequality of~(\ref{M25b}).
Since~$\Vp = \vp - \Rvp$, where~$-\Delta \Rvp = \lm \vp$ in~$\Ot$ and~$\Rvp = 0$ on~$\Gt$, and
the required inequality for~$\vp$ is proved already, it is sufficient to prove the
inequalities for~$\Rvp$ only. However, since the boundary value problem for~$\Rvp$ is similar
to that for~$\vp$, the proof can be carried out in the same way as for~$\vp$.
\end{proof}
\section{Proof of Theorems \protect\ref{i:t:holder_boundary} and \protect\ref{i:t:C1_boundary}}
By~(\ref{eq:est_vpVs}) and~(\ref{eq:est_Vpvs}) combined with Lemma~\ref{LM28a}, we obtain that
\[
|\ip{\vp}{\Vs}|+|\ip{\vs}{\Vp}| = \left\{ \begin{array}{ll}
o(d) & \mbox{for the~$C^1$-case}\\
O(d^{1+\alpha}) & \mbox{for the~$C^{1,\alpha}$-case}
\end{array} \right. \;\; \mbox{as $d\to 0$.}
\]
Moreover, due to~(\ref{eq:est_T}) and Lemma~\ref{LM28a},
\[
|\ip{\S\vp}{\S\vs} - \ip{\vp}{\vs}| = \left\{ \begin{array}{ll}
o(d) & \mbox{for the~$C^1$-case}\\
O(d^{1+\alpha}) & \mbox{for the~$C^{1,\alpha}$-case}
\end{array} \right. \;\; \mbox{as $d\to 0$.}
\]
Having in mind these relations and~(\ref{M12c}), we can write formula~(\ref{M28s}) as
\begin{equation}
\label{M28t}
\um[k]^{-1} = \lm[m]^{-1} + \tk + \left\{ \begin{array}{ll}
o(d) & \mbox{for the~$C^1$-case}\\
O(d^{1+\alpha}) & \mbox{for the~$C^{1,\alpha}$-case}
\end{array}\right., \; \; k = 1,2,\ldots,\Jm,
\end{equation}
where~$\tau = \tk$ is an eigenvalue of the problem
\begin{equation}
\label{M28u}
\lm[m]^{-1} \Big(\int_{\Ot \setminus \Oe} \nabla \Rvp \cdot \nabla \Rvs \, dx
- \int_{\Oe \setminus \Ot} \nabla \vp \cdot \nabla \vs \, dx \Big)
= \tau \ip{\vp}{\vs},
\end{equation}
where~$\vp \in \Xj[m]$ and~$\tau_1,\ldots,\tau_{J_m}$ run through all eigenvalues of~(\ref{M28u})
counted with multiplicity.
To obtain results for integrals over the domains~$\Oe \setminus \Ot$ and~$\Ot \setminus \Oe$
in the case of~$C^1$ or~$C^{1,\alpha}$ boundaries
phrased in terms of boundary integrals, the following two Lemmas will be helpful.
\begin{lemma}
\label{l:lip_boundaryint}
Let~$\Oe$ and~$\Ot$ satisfy the~$C^1$-assumption from Section~\ref{SK1}.
Suppose that~$g \in L^2(\Ge)$, $f \in L^2(\Oe)$ and that~$U$ solves~$\Delta U = f$ in~$\Oe$ and~$U = g$ on~$\Ge$.
Then
\begin{equation}
\label{eq:lip_boundaryint}
\int_{\Oe \setminus \Ot} |U|^2 \, dx = -\int_{\Ge} \sigma_{-} |g|^2 \, dS + o(d),
\end{equation}
as~$d \rightarrow 0$. Here,~$\sigma_{-} = \max (0,-\sigma)$, where~$\sigma$ is the function defined by~{\rm (\ref{KK1a})}.
If, in addition,~$\Oe$ and~$\Ot$ satisfy the~$C^{1,\alpha}$-assumption from Section~\ref{SK1},~$g \in C^{0,\alpha}(\Ge)$
and~$f\in L^\infty(\Oe)$, then~{\rm(}\ref{eq:lip_boundaryint}{\rm)} is valid with~$o(d)$ replaced by~$O(d^{1+\alpha})$.
\end{lemma}
\begin{proof}
We represent~$U$ as~$U = u + v$, where
\[
\Delta u = f\;\; \mbox{in $\Dd$ and} \;\; u = 0 \;\; \mbox{on $\partial \Dd$}
\]
and
\[
\Delta v = 0\;\; \mbox{in $\Oe$ and} \;\; v = g - u \;\; \mbox{on $\Ge$.}
\]
Then~$u \in H^2(\Dd)$ with~$\|u\|_{H^2(\Dd)} \leq C\|f \|_{L^2(\Oe)}$ and
\begin{equation}
\label{M5a}
\int_{\Ge} \N(v)^2 \, dS \leq C( \|g\|^2_{L^2(\Ge)} + \|f\|^2_{L^1(\Oe)}).
\end{equation}
These estimates lead to
\begin{equation*}
\begin{aligned}
\int_{B'_\delta} \int_{h_1(y')}^{h_{12}(y')} |U(y',y_n)|^2 \, dy
& {} \leq C d \int_{B'_\delta} \sup_{h_1 < y_n < h_{12}} U^2(y',y_n) \, dy'\\
& {} \leq C d \, ( \|g\|^2_{L^2(\Ge)} + \|f\|^2_{L^2(\Oe)} ),
\end{aligned}
\end{equation*}
which implies that
\begin{equation}
\label{M30a}
\int_{\Oe \setminus \Ot} |U|^2 \, dx \leq C d \, ( \|g\|^2_{L^2(\Ge)} + \|f\|^2_{L^2(\Oe)} ).
\end{equation}
Let us choose an approximating sequence~$g_j \in H^1(\Ge)$, $j=1,2\ldots$, of the function~$g$,
i.e., we assume that $\epsilon_j := \| g - g_j \|_{L^2(\Ge)} \to 0$ when~$j \to \infty$.
If we denote by~$U_j$ the solution to
\[
\Delta U_j = f \;\; \mbox{in $\Oe$ and} \;\; U_j = g_j \;\; \mbox{on $\Ge$,}
\]
then by~(\ref{M30a}),
\begin{equation}
\label{feb23b}
\int_{\Oe \setminus \Ot} |U-U_j|^2 \, dx \leq C d \, \epsilon_j^2.
\end{equation}
We represent~$U_j$ as~$U_j = u + v_j$, where~$\Delta v_j = 0$ in $\Oe$ and $v_j = g_j - u$ on~$\Ge$. Then
\[
\int_{B'_\delta} \sup_{h_1 < y_n < h_{12}} |\partial_{y_n} U_j(y',y_n)|^2 \, dy'
\leq C(\|g_j\|^2_{H^1(\Ge)} + \|f\|^2_{L^2(\Oe)})
\]
since~$g_j \in H^1(\Ge)$ and~$u \in H^2(\Oe)$ (see (\ref{M31b})).
Observing that
\[
\begin{aligned}
\int_{B'_{\delta}} \int_{h_1}^{h_{12}} U_j^2(y',y_n) \, dy_n \, dy'
= {} & \int_{B'_{\delta}} \int_{h_1}^{h_{12}} U_j^2(y',h_1(y')) \, dy_n \, dy'\\
& + 2 \int_{B'_{\delta}} \int_{h_1}^{h_{12}} \int_{h_1}^{y_n} \partial_s U_j(y',s) \,U_j(y',s) \, ds \, dy_n \, dy',
\end{aligned}
\]
we thus obtain that
\begin{equation}
\label{feb23a}
\Big| \int_{B'_{\delta}} \int_{h_1(y')}^{h_{12}(y') } (U_j^2(y',y_n) - g_j^2(y')) \, dy_n \, dy' \Big|
\leq Cd^2 \int_{B'_{\delta}} \widehat{N}(U_j) \widehat{N}(\partial_{y_n} U_j) \, dy',
\end{equation}
where
\[
\widehat{N}(w)(y') = \sup_{|y_n| < \delta} |w(y',y_n)|.
\]
From~(\ref{feb23a}) it follows that
\begin{equation}
\label{M31c}
\Big| \int_{B'_{\delta}} \int_{h_1(y')}^{h_{12}(y')} U_j^2(y',y_n) \, dy_n dy' -
\int_{B'_{\delta}} \bigl( h_2(y') - h_1(y') \bigr) g_j^2(y') \, dy' \Big| \leq C_j d^2
\end{equation}
with a certain constant depending on~$j$ but independent of~$d$.
Now, since
\[
\nu(y') = \frac{(\nabla h_1(y'),-1)}{\sqrt{1+|\nabla h_1(y')|^2}},
\]
we have (due to the~$C^1$-assumption)
\[
\sigma(y') = \frac{h_1(y') - h_2(y')}{\sqrt{1+|\nabla h_1(y')|^2}} + o(d) \;\; \mbox{as $d\to 0$,}
\]
which together with~(\ref{M31c}) implies that
\[
\Big| \int_{B'_\delta} \int_{h_1(y')}^{h_{12}(y')} U_j^2(y',y_n) \, dy_n \, dy' -
\int_{B'_\delta} \sigma(x') |g_j(x')|^2 \, dS(x') \Big| = o(d),
\]
where~$dS(x') = \sqrt{1+|\nabla h_1(y')|^2} \, dy'$.
This together with~(\ref{feb23b}) implies that~(\ref{eq:lip_boundaryint}) holds.
We now consider the~$C^{1,\alpha}$-case. Indeed, the solution (weak solution) then
belongs to~$C^{1,\alpha}(\Oe)$ and
\[
\| U \|_{C^{1,\alpha}(\Oe)} \leq C ( \| g\|_{C^{1,\alpha}(\Ge)} + \|f\|_{L^\infty(\Oe)} ) \mbox{;}
\]
see for instance Agmon et al.~\cite{Agmon1959}.
In this case, we can write
\begin{equation*}
\Big| \int_{B'_{\delta}} \int_{h_1(y')}^{h_{12}(y')} U^2(y',y_n) \, dy_n \, dy' -
\int_{B'_{\delta}} \int_{h_1(y')}^{h_{12}(y')} U^2(y',h_1(y')) \, dy_n \, dy' \Big| \leq C d^{1+\alpha}
\end{equation*}
because of the H\"older continuity of~$U$. Furthermore, it is clear that
\[
\sigma (y') = \frac{h_1(y') - h_2(y')}{\sqrt{1+|\nabla h_1(y')|^2}} + O(d^{1+\alpha}) \;\; \mbox{as $d\to 0$,}
\]
by the~$C^{1,\alpha}$-assumption. This proves that~(\ref{eq:lip_boundaryint}) holds with the remainder~$o(d)$ replaced
by~$O(d^{1+\alpha})$.
\end{proof}
\noindent Analogously with Lemma~\ref{l:lip_boundaryint}, we can prove the following result.
\begin{lemma}
\label{l:lip_boundaryint2}
Let~$\Oe$ and~$\Ot$ be the same domains as in the previous lemma.
Suppose that~$g \in L^2(\partial \Ot)$, $f\in L^2(\Ot)$ and that~$U \in H^1(\Ot)$ solves~$\Delta U = f$ in~$\Ot$ and~$U = g$ on~$\partial \Ot$.
Then
\begin{equation}
\label{M28a}
\int_{\Ot \setminus \Oe} |U|^2 \, dx = \int_{\Ge \cap \Ot} \sigma_{+} \, |U|^2 \, dS + o(d)\;\;\mbox{as~$d \rightarrow 0$}
\end{equation}
in the~$C^1$-case, where~$\sigma_{+} = \max(\sigma, 0)$. In the~$C^{1,\alpha}$-case, we assume in addition
that~$g\in C^{0,\alpha}(\Gt)$ and~$f\in L^\infty(\Ot)$, and the remainder~$o(d)$ can be replaced by~$O(d^{1+\alpha})$
in~{\rm(}\ref{M28a}{\rm)}.
\end{lemma}
\section{The end of the proof of Theorems \protect\ref{i:t:holder_boundary} and \protect\ref{i:t:C1_boundary}}
The function~$U = \partial_{x_j}\vp$ satisfies
\[
\Delta U = f := -\lm \partial_{x_j} \vp \;\; \mbox{in $\Oe$ and}
\;\; U = g := \partial_{x_j}\vp \;\; \mbox{on $\Ge$.}
\]
Clearly, $f \in L^2(\Oe)$ and~$g \in L^2(\Ge)$ in the~$C^1$-case,
and in the~$C^{1,\alpha}$-case, we have~$f \in C^{0,\alpha}(\Oe)$ and~$g \in C^{0,\alpha}(\Ge)$.
Applying Lemma~\ref{l:lip_boundaryint}, we obtain
\begin{equation}
\label{M31k}
\int_{\Oe \setminus \Ot} \partial_{x_j} \vp \, \partial_{x_j} \vs \, dx =
-\int_{\Ge} \sigma_{-} \, \partial_{x_j} \vp \, \partial_{x_j} \vs \, dS + o(d)
\end{equation}
in the~$C^1$-case and with the remainder~$o(d)$ replaced by~$O(d^{1+\alpha})$ in the~$C^{1,\alpha}$-case.
We note that in order to apply Lemma~\ref{l:lip_boundaryint},
it is useful to use the relation
\begin{equation}
\label{M31n}
\partial_{x_j} \vp \, \partial_{x_j} \vs = \frac{1}{2} \bigl( \partial_{x_j}\vp + \partial_{x_j}\vs\bigr)^2 -
\frac{1}{2} \partial_{x_j}\vp^2 - \frac{1}{2}\partial_{x_j}\vs^2
\end{equation}
since formula~(\ref{eq:lip_boundaryint}) only contains squares of functions.
Equality~(\ref{M31k}) implies that
\begin{equation}
\label{M31l}
\int_{\Oe \setminus \Ot} \nabla\vp \cdot \nabla\vs \, dx = -\int_{\Ge} \sigma_{-} \nabla\vp \cdot \nabla\vs \, dS + o(d)
\end{equation}
for the~$C^1$-case and with the remainder replaced by~$O(d^{1+\alpha})$ in the~$C^{1,\alpha}$-case.
Furthermore, the function~$U = \partial_{x_j} \Rvp$ satisfies
\[
\Delta U = f := -\lm \partial_{x_j} \vp \;\; \mbox{in~$\Ot$ and} \;\; U = g := \partial_{x_j} \Rvp \;\; \mbox{on~$\Gt$.}
\]
Applying Lemma \ref{l:lip_boundaryint2}, we get
\begin{equation}
\label{M31m}
\int_{\Oe \setminus \Ot} \partial_{x_j}\Rvp \, \partial_{x_j} \Rvs \, dx =
\int_{\Ge} \sigma_{+} \, \partial_{x_j} \Rvp \, \partial_{x_j} \Rvs \, dS + o(d)
\end{equation}
in the~$C^1$-case, and in the~$C^{1,\alpha}$-case, the remainder~$o(d)$ is replaced by~$O(d^{1+\alpha})$.
Using that~$\partial_{x_j} \Rvp = \partial_{x_j} \vp - \partial_{x_j}\Vp$ and that
\[
\int_{\Get} |\nabla \Vp|^2 \, dS = o(1) \;\; \mbox{as $d\to 0$}
\]
in the~$C^1$-case and~$O(d^{2\alpha})$ in the~$C^{1,\alpha}$-case, we obtain
\begin{equation}
\label{M31r}
\int_{\Oe \setminus \Ot} \partial_{x_j} \Rvp \, \partial_{x_j} \Rvs \, dx =
\int_{\Ge} \sigma_{+} \, \partial_{x_j} \vp \, \partial_{x_j}\vs \, dS + o(d)
\end{equation}
in the~$C^1$-case, and with the remainder replaced by~$O(d^{1+\alpha})$ in the~$C^{1,\alpha}$-case.
Therefore
\begin{equation}
\label{M31p}
\int_{\Ot \setminus \Oe} \nabla \Rvp \cdot \nabla \Rvs \, dx = \int_{\Ge} \sigma_{+} \, \nabla\vp \cdot \nabla\vs \, dS + o(d)
\end{equation}
in the~$C^1$-case, and with the remainder replaced by~$O(d^{1+\alpha})$ in the~$C^{1,\alpha}$-case.
Now, applying formulas~(\ref{M31l}) and~(\ref{M31p}) to~(\ref{M28u}), we prove
Theorems~\ref{i:t:holder_boundary} and~\ref{i:t:C1_boundary}.
\section{Counterexamples}\label{SA4a}
Let~$\Oe$ be the domain in $\mathbf{R}^2$ given by
\[
\Oe = \{ (x,y) \,:\, 0 < x < T,\, 0 < y < R \}.
\]
The domain~$\Ot$ is given by
\[
\Ot = \{ (x,y) \,:\, 0 < x < T, \, d\eta(x/\delta) < y < R \},
\]
where~$\eta$ is a positive, periodic, $C^1$-function such that~$\eta(X+1)=\eta(X)$.
We assume that~$\delta$ and~$d$ are small parameters, $d \leq \delta$, and that $T/\delta = N$,
where~$N$ is a large integer. We will consider three cases:~$\delta = d$ is a Lipschitz perturbation,
if~$\delta = o(d)$ we are dealing with a~$C^1$-perturbation and if~$\delta = d^{1-\alpha}$,
the perturbation is of the class~$C^{1,\alpha}$.
Consider the eigenvalue problems
\[
-\Delta \vp = \lm \vp \;\; \mbox{in $\Oe$ and $\vp(x,0) = \vp(x,R) = 0$ for $0 < x < T$.}
\]
In addition, we assume that the function is periodic in $x$, i.e.
\[
\vp(0,y) = \vp(T,y) \;\; \mbox{and} \;\; \partial_x \vp(0,y) = \partial_x\vp(T,y) \;\; \mbox{for $0 < y < R$.}
\]
We are interested in a perturbation of the first eigenvalue~$\lm=\lambda_1$.
Separating the variables, one can easily find that
\[
\lambda_1 = \frac{\pi^2}{R^2} \;\; \mbox{and} \;\; \vp = \varphi_1 = \frac{\sqrt{2}}{\sqrt{RT}}\sin\Big(\frac{\pi}{R}y\Big),
\]
where $\varphi_1$ is a corresponding eigenfunction normalized so that~$\|\varphi_1\|_{L^2(\Oe)} = 1$.
Clearly,~$\lambda_1$ is a simple eigenvalue.
The perturbed problem is the following:
\[
-\Delta u = \mu u \;\; \mbox{in $\Ot$ and $u(x,d \eta(x/\delta)) = u(x,R)=0$ for $0 < x < T$.}
\]
Furthermore, one must add periodicity conditions on the parts of the boundary where~$x=0$ or~$x=T$.
In this case,~$\Ot \subset \Oe$, and therefore~$\Vp = 0$ and the formula in~(\ref{eq:main_asymp_D})
is reduced to
\[
\tau_1 = -\frac{1}{\lambda_1} \int_{\Oe} |\nabla\Phi|^2 \, dx \, dy
\]
and~(\ref{M28s}) becomes
\[
\mu_1^{-1} = \lambda_1^{-1} + \tau_1 + O(d^2).
\]
Here $\Phi = \varphi_1$ in $\Oe \setminus \Ot$ and
\[
\Delta \Phi = 0 \;\; \mbox{in $\Ot$ and} \;\; \Phi(x,R) = 0,\,
\Phi(x,d\eta(x/\delta)) = \varphi_1(x,d\eta(x/\delta)),\; 0 < x < T.
\]
Moreover, the periodicity condition must be satisfied on the parts of the boundary where~$x=0$ or~$x=T$.
Since
\begin{equation}
\label{A3a}
\int_{\Oe} |\nabla\Phi|^2 \, dx\, dy = \int_{\Oe \setminus \Ot} |\nabla\varphi_1|^2 \, dx\, dy
+ \int_{\Ot} |\nabla\Phi|^2 \, dx \, dy
\end{equation}
and the first term in the right-hand side corresponds to the Hadamard term in~(\ref{A4a}) and
has the order~$O(\delta)$, it suffices to analyze the second term in~(\ref{A3a}).
We will construct the function~$\Phi$ in~$\Ot$ using the following representation:
\[
\Phi(x,y) = d w_0(x,y) + d V_0(X,Y) + {\mathcal R}(x,y),
\]
where~$X=x/\delta$ and~$Y=y/\delta$. The function~$w_0$ solves the following boundary value problem in $\Oe$:
\[
\Delta w_0 = 0 \;\; \mbox{in $\Oe$ and} \;\; w_0(x,R) = 0, \;\; w_0(x,0) = c_0,
\]
where~$c_0$ is a constant to be determined later. As usual, periodicity conditions are valid on the vertical
parts of the boundary.
The solution to this problem is given by
\[
w_0 = \frac{c_0(R-y)}{R}.
\]
In order to write the boundary value problem for~$V_0$, we note that
\[
\varphi_1(x,y) = d C_1 \eta(X) + O(d^3) \;\;\; \mbox{when $y=d\eta(x/\delta)$}, \;\;
C_1 = \frac{\sqrt{2}}{\sqrt{RT}}\frac{\pi }{R}.
\]
The function~$V_0$ is periodic with respect to~$X$ with the period~$1$ and solves the problem
\begin{equation}
\label{A3b}
\Delta_{X,Y} V_0 = 0 \;\;\mbox{in ${\mathcal D}$ and} \;\; V_0(X,d\eta(X)/\delta) = C_1\eta(X)-c_0,
\end{equation}
where
\[
{\mathcal D}=\{ (X,Y) \,:\, \eta(X) < Y < \infty, \, 0 < X < 1 \}.
\]
Furthermore, it is assumed that~$V_0$ decays exponentially as~$Y \to \infty$.
As it is known, the problem~(\ref{A3b}) has a unique solution~$V_0$ in the class of periodic
solutions subject to, e.g.,
\[
\int_{{\mathcal D}} |\nabla V_0|^2 \, dX \, dY + \int_0^1 \int_{\eta(X)}^1 |V_0|^2 \, dY \, dX < \infty,
\]
and this solution admits the asymptotic representation
\[
V_0(X,Y) = q + O(e^{-\pi Y}) \;\; \mbox{as $y \to \infty$.}
\]
To calculate the coefficient~$q$, one can introduce a special solution~${\mathcal V}(X,Y)$
which solves the to~(\ref{A3b}) corresponding homogeneous problem and has the asymptotics
\[
{\mathcal V}(X,Y) = Y + c_2 + O(e^{-\pi Y}).
\]
By applying the maximum principle we obtain that~${\mathcal V} > 0$ inside~${\mathcal D}$.
If~$\delta=d$, then by Hopf's lemma
\[
\partial_{\nu} {\mathcal V}(X,\eta(X)) < 0 \;\; \mbox{for $0 \leq X \leq 1$.}
\]
In the case when~$\delta = o(d)$ ($\delta = O(d^{1-\alpha})$),
${\mathcal V} = \widetilde{\mathcal V} + o(d)$ (${\mathcal V} = \widetilde{\mathcal V} + O(d^\alpha)$),
where~$\widetilde{\mathcal V}$ solves the homogeneous problem
\[
\Delta_{X,Y} \widetilde{\mathcal V} = 0 \;\; \mbox{in $\widetilde{\mathcal D}$ and} \;\;
\widetilde{\mathcal{V}} (X,0) = 0,
\; \; \mbox{where~$\widetilde{\mathcal D} = (0,1) \times (0,\infty)$.}
\]
By Hopf's lemma,~$\partial_Y \widetilde{\mathcal V}(X,0) > 0$ for~$0 \leq X \leq 1$.
Now, let~$\mathcal{D}_c$ be the domain~$\mathcal{D}$ truncated at~$Y = c$ so that~$\eta(X) < Y < c$.
Since~$V_0$ and~$\mathcal{V}$ are harmonic in~$\mathcal{D}$, it is clear that
\[
\begin{aligned}
\int_{0}^{1} \mathcal{V}(X, c) \partial_{Y} V_0(X,c) \, dX
= {} &
\int_{0}^{1} \partial_{Y} \mathcal{V}(X,c) V_0(X,c) \, dX\\
& + \int_0^1 \partial_\nu {\mathcal V}(X,\eta(X))(C_1\eta(X) - c_0)\sqrt{1+\eta'(X)^2} \, dX.
\end{aligned}
\]
Using the asymptotic representations above and letting~$c \to \infty$,
we obtain that
\[
q = -\int_0^1 \partial_\nu {\mathcal V}(X,\eta(X))(C_1\eta(X)-c_0)\sqrt{1+\eta'(X)^2} \, dX.
\]
The constant~$c_0$ is sought from the relation $q = 0$, which guarantees the exponential decay of the
function~$V_0$.
Finally, the remainder~${\mathcal R}$ satisfies
\[
\Delta {\mathcal R} = 0 \;\; \mbox{in $\Ot$,} \;\; {\mathcal R}(x,R) = O(e^{-\pi R/\delta}) \;\;
\mbox{and} \;\; {\mathcal R}(x,d\eta(x/\delta)) = O(d^3)
\]
and the differentiation of the boundary conditions will give a factor~$1/\delta$.
This leads to the estimate
\[
\int_{\Ot}|\nabla {\mathcal R}|^2 \, dx\, dy = O(d^2).
\]
Therefore
\[
\begin{aligned}
\int_{\Ot} |\nabla\Phi|^2 \, dx \, dy = {} & \int_0^T \int_{d\eta(x/\delta)}^R
\Big(\Big| \frac{d}{\delta} \partial_{X} V_0(X,Y) \Big |^2+|d (w_0)'_y + \frac{d}{\delta} \partial_{Y} V_0(X,Y)|^2\Big) dy \, dx \\
& + O(d^2).
\end{aligned}
\]
Since
\[
\begin{aligned}
\int_0^T \int_{d\eta(x/\delta)}^R \Big|\frac{d}{\delta} \partial_X V_0(X,Y)\Big |^2 dy \, dx = {} &
T \frac{d^2}{\delta} \int_0^1 \int_{d\eta(X)/\delta}^\infty |\partial_X V_0(X,Y)|^2 dY \, dX\\
& + O(d^2)
\end{aligned}
\]
and
\[
\int_0^T \int_{d\eta(x/\delta)}^R |d (w_0)'_y + \frac{d}{\delta}\partial_Y V_0(X,Y)|^2 dy \, dx =
O\Big( \frac{d^2}{\delta} \Big) + O(d^2),
\]
we see that an additional term in the Hadamard formula will appear if~$\delta=d$ and that
the remainders~$o(d)$ and~$O(d^{1+\alpha})$, respectively, are sharp in
theorems~\ref{i:t:holder_boundary} and~\ref{i:t:C1_boundary}.
\def\bibname{References}
|
\section{Introduction}
\thispagestyle{empty}
Podcast audiences have more than doubled in size over the last decade, bringing with them demand for more frequent releases and an expanded scope of content \cite{WinNT}. In parallel, the barrier to entry for hosting a podcast has been significantly lowered, with reduced costs for audio recording technology allowing non-experts to engage in discussion of topics such as politics or entertainment that were previously reserved for more traditionally-accredited and financially-supported individuals \cite{baumgartner2010myfacetube, kang2012differences}. In this combination, podcasts have demonstrated tremendous potential as a gauge for social opinion.
Nonetheless, the volume of podcast content produced today necessarily precludes large-scale topical analysis. In particular, it is currently difficult to track mentions of noteworthy properties across multiple podcast channels. As such, NER, the identification of qualitatively significant word phrases such as people and organizations \cite{chieu2002named}, is a well-needed focus in the area of podcast analysis. NER is a key step necessary to make higher level inferences such as measuring sentiment, identifying emotions associated with properties, or building predictive models for property-level response variables such as revenue based on podcast data.
Although NER systems for formal, written language perform accurately \cite{lin2009phrase, ratinov2009design}, there remains substantial room for improvement in evolving communication mediums where traditional linguistic structures are used less consistently (e.g. social media posts and informal conversations). Entity recognition faces further challenges in the domain of human speech, where an intermediary step to transform audio into computer-readable text entirely removes orthographic features, which are often used to highlight entities in writing (e.g. captitalization, punctuation).
To address the aforementioned challenges for entity detection in podcasts, we propose a two-stage NER system and evaluate it in the context of film title detection. We note that detection of film titles is one of the more niche and ambitious tasks within this research domain, given that new films are released each week and most NER systems rely on large volumes of training data that can be costly to obtain. Given this scope, we believe that a successful film title detection approach has promise to transfer to other entity classes which are traditionally more stable over time.
In this paper, we begin by briefly reviewing existing NER systems used in informal language domains such as human speech, highlighting their limitations in the context of film title detection. In Section 3, we discuss our data sources, external dependencies and our pre-processing approach. In Section 4, we propose and detail the two-step candidate identification and entity classification procedure that lies at the crux of this paper. Finally, in Section 5, we evaluate our proposed method and compare it against three baselines.
\section{Related Work}
Although there exists plenty of research on NER for traditional noun phrases such as people, locations, and organization names, little has been done for niche entities such as movies and books. The challenge in the latter is that these properties evolve in much shorter time intervals (i.e. new movies are released every week).
Prior research has used supervised machine learning to recognize entities from audio using acoustic features. For example, the speech recognition model proposed by \cite{hatmi2013incorporating} uses constrained maximum likelihood linear regression to simultaneously predict the most probable sequence of words and entity classes within an audio waveform. However, this approach requires a large set of token-labeled and time-synchronized training transcripts. Such a dependency is prohibitive in the dynamically changing landscape for film titles.
Focusing less on acoustic signal, \cite{chowdhury2013simple} feeds word tokens and linguistic features from audio transcripts into a Conditional Random Field (CRF) model to detect people, locations, organizations, and geo-political entities. While this approach works quite well, it is not suited for unstable entity classes and proves quite vulnerable to word/phrase transcription errors.
Given the limitations of previous research in the audio analysis domain, we look for inspiration in another challenging medium - social media. Notably, language in social media typically suffers from inconsistent spelling, poor grammar, and a dearth of orthographic features \cite{eisenstein2013bad}. Given the current state of automatic audio transcription tools, the computer-readable text generated from podcast audio often encounters the same inconsistencies. Moreover, language in both social media and human speech requires normalization, as people often invoke non-standard tokens and phrases within these domains \cite{liu2012broad}.
In a relatively recent study, \cite{Ashwini} propose a system that identifies entity candidates in social media by matching token sequences in tweets to phrases in a gazetteer. To address issues with precision, they train a classifier using a combination of orthographic, n-gram, and syntactic features to determine whether entity candidates are indeed true entities. Importantly, their system does not require constant re-training, as the gazetteer may be updated with new elements in real time. While their results demonstrate promise for a variety of entity types, their system still relies on capitalization, special characters, and syntactic features to achieve the desired performance. As mentioned above, these features are critically absent in transcripts of podcast audio and thus motivate our work.
\section{Data and Pre-processing}
We collect 20 film-related podcasts from various publicly available channels (see Table 1), including National Public Radio, SlashFilm, Screen Junkies and Looper. We listen to each podcast and manually note film properties mentioned within each one to serve as the ground truth. The quality of labels is evaluated by the Cohen's Kappa inter-annotator agreement (0.63). The podcasts are of similar length (10 $\pm{5}$ minutes) except those from SlashFilm (100 minutes). The complete distribution of entities within our dataset and estimated transcription errors can be found in Table 1.
\begin{table}[]
\centering
\caption{Podcast Label Distribution and Average Word Error Rates}
\begin{tabular}{@{}lllllll@{}}
\toprule
{Channels} & {\#1} & {\#2} & {\#3} & {\#4} & {Total} & {\begin{math}\overline{WER}\end{math}} \\ \midrule
Collider & 51 & 2 & - & - & 53 & 2\% \\
All Things Consd. & 4 & 2 & - & - & 6 & 26\% \\
Looper & 4 & 16 & 37 & - & 57 & 2\% \\
Bob Mondello & 20 & 15 & 5 & - & 40 & 10\% \\
Morning Ed. & 3 & 11 & - & - & 14 & 23\% \\
Screen Junkies & 20 & 30 & 3 & 11 & 64 & 2.5\% \\
Angry Joeshow & 18 & - & - & - & 18 & 4\% \\
Fresh Air & 14 & 12 & - & - & 26 & 25\% \\
SlashFilm & 38 & - & - & - & 38 & 38\%
\end{tabular}
\vspace{-1.0em}
\end{table}
Each podcast is subjected to the same set of pre-processing steps. First, raw podcast audio is transcribed using an open-source speech recognition framework from \cite{Zhang}. The output is a long sequence of lowercase words separated by whitespace; there are not any orthographic features or punctuation, which notably have been deemed critical in existing NER systems. Applying \cite{Zhang}'s model to audio from a National Public Radio podcast in which a human-curated transcript is available, we empirically observe a 23\% Word Error Rate (WER) on average. Given this relatively high WER, we reiterate the importance of an entity-detection system that is flexible enough to handle errors in the transcription procedure.
To aid in the inference of syntactic features, punctuation is inferred using a bidirectional recurrent neural network (with attention) that has been trained on European Parliment speech data \cite{tilk2016bidirectional}. To evaluate the quality of this inference, we apply it to the same podcast used to test WER. The punctuation model performs significantly better than chance and achieves a precision, recall, and F1 score of 0.78, 0.64, and 0.70, respectively.
After punctuation is complete, all numeric values are converted to text (e.g. 1984 to nineteen eighty four). Finally, the processed text string is tokenized using the sentence and word tokenizers from NLTK \cite{bird2009natural}.
\section{Methods}
\subsection{Identifying Candidates}
\subsubsection{Gazetteer}
We use a proprietary database of 9000 films produced between 2000 and 2016 as our gazetteer. Each film has the following metadata available: production budget, keywords, plot summary, and logline. Notably, 70\% of films are missing data from at least one of these fields.
\subsubsection{Entity Lookup}
\cite{Ashwini} performs exact string matching using suffix trees to identify words and phrases which may be entities (a.k.a. entity candidates). However, exact string matching tends to miss titles that have minor word errors (e.g. film "Coco" transcribed as "cocoa"). In our approach, we use a Levenshtein ratio \cite{fuad2012towards} with a threshold that varies with the number of tokens present in the phrase to determine whether a match exists. The threshold for each n-gram length \begin{math}n \in [1,6]\end{math} is determined using cross-validation within the training set. This similarity measure is implemented to account for the relatively high WER of the \cite{Zhang} audio transcription model. Accordingly, long movie titles such as "Three Billboards Outside Ebbing, Missouri" can have minor transcription errors without being ignored during the entity candidate identification stage.
\begin{figure*}[h]
\centering
\includegraphics[scale=0.45]{FigureCOM.png}
\caption{Left: Results from LOCO cross-validation. Variation due to the differences in WER and total number of true film title mentions. Right: Comparison of our NER system to 3 baselines across LOCO cross-validation. Substantial gain in precision achieved using metadata-based rules and features.
\end{figure*}
\subsubsection{Feature Extraction}
Using the tokenized text output from the preprocessing stage in Section 3, we infer part of speech (POS) tags for each token in the audio transcript using the Stanford POS Tagger \cite{bird2009natural}. Then, we add several features to each entity candidate based on metadata from their potential film match. All features can be found in Table 2.
\begin{table*}
\centering
\caption{Feature Properties Extracted from Transcription}
\label{tab:commands}
\def1.5{1.5}
\begin{tabular}{ccl}
\toprule
\makecell{Features} &\makecell{Properties} & \makecell{Definition}\\
\midrule
\texttt{Closeness} & \begin{math} Closeness(m,k) = 1 - \frac{|w_m-w_{k}|} {||transcript||} \end{math} & \makecell{Closeness value between movie \begin{math}m\end{math} and keyword \begin{math}k\end{math}} \\
\texttt{Levenshtein Ratio}& \begin{math} \frac{lev(m_{matched},m_{original})} {\max(||m_{matched}||,||m_{original}||)} \end{math} & \makecell{\begin{math}lev(a,b)\end{math} measures the minimum number of deletions,\\insertions and substitutions to transform \begin{math}a\end{math} into \begin{math}b\end{math}} \\
\texttt{Production Budget}& \begin{math} \frac{budget_m-budget_{min}} {budget_{max}-budget_{min}} \end{math} & \makecell{Normalized budget for film \begin{math}m\end{math} within the transcript}\\
\texttt{Title POS-tags}& \begin{math} [\textit{POS-tags}]_m \end{math} & \makecell{Bag of POS-tags in the title of movie \begin{math}m\end{math}}\\
\texttt{Pre/Post POS-tags}& \begin{math} [\textit{POS-tag}]_{m-1} \end{math}, \begin{math} [\textit{POS-tag}]_{m+1} \end{math} & \makecell{POS-tags of preceeding and succeeding word for candidate}\\
\texttt{N-gram Levels}& \begin{math} [1-6] \end{math} & \makecell{Number of tokens in entity candidate}\\
\bottomrule
\end{tabular}
\vspace{-1.5em}
\end{table*}
Notably, we design a metric to capture the thematic relevance the context around an entity candidate has to its associated film. Informally, we define \textit{closeness} to be a normalized value representing the number of words in the transcript which seperate the entity candidate, \begin{math}m\end{math}, from relevant keyword, \begin{math}k\end{math}, where \begin{math}k\in K\end{math} and \begin{math}K\end{math} denotes keywords of the movie \begin{math}m\end{math} in our database. Their word indices within the transcript are noted by \begin{math}w_m\end{math} and \begin{math}w_{k}\end{math} respectively. We define this metric mathematically in Table 1. After calculating each \begin{math}k\end{math}'s closeness value to corresponding entity candidate \begin{math}m\end{math}, we extract \begin{math}\frac{1}{||K||}\sum_{k\in K}closeness({w_m,k})\end{math}, \begin{math}\min_{k\in K}closeness({w_m,k})\end{math} and \begin{math}\max_{k\in K}closeness({w_m,k})\end{math} values as model features.
\noindent For example, if the entity candidate is "Godzilla", we may expect words like "monster", "large", or "japan" to appear in close proximity to the entity candidate's position within the transcript.
\subsection{Classifying Entities}
After identifying potential film mentions via our fuzzy matching algorithm, the entity candidates are subjected to binary classification via logistic regression. We evaluated baseline approaches and our system using 9-fold Leave One Channel Out (LOCO) cross-validation. Results from our approach are highlighted in the left-side of Figure 1. Hyperparameters (regularization and penalty) are selected to optimize F1 score within each training set.
Since the model is agnostic to specific words or phrases, this NER system does not require any retraining when the gazetteer is updated to reflect new movie releases.
\subsubsection{Feature Selection}
We find that the most predictive features of true entity mentions include the following: n-gram levels, POS-tags, closeness, and Levenshtein ratio (using null hypothesis testing p-values<0.05 for mentioned features). While n-gram levels and POS-tags provide adequate performance, the most significant performance gain comes from the addition of the metadata-based features: closeness and production budget.
\section{Results and Discussion}
We compare performance of our model-based entity recognition to three baseline approaches (right-side of Figure 1). Baseline 1 classifies all entity candidates as true film mentions. Baseline 2 is similar to Baseline 1, except that we limit candidates to those inferred to be a noun-phrase using the Stanford POS Tagger. For Baseline 3, we consider all candidates identified in the first stage of our process and then remove those whose closeness statistics are below thresholds determined via cross-validation over the training data.
Although \cite{ritter2011named} has proven useful in the social media domain, we find it does not serve as an adequate baseline for our task. The pre-trained model from \cite{ritter2011named} identifies zero film titles across our dataset of transcripts. To understand why, we applied this model to a human annotated transcription from National Public Radio (NPR) with and without capitalization. While the model identified 16 out of 33 true film title mentions in the capitalized transcript, it did not identify any within the uncapitalized version.
\noindent The lack of capitalization dramatically reduces the performance and highlights the value of a gazetteer-based approach.
The rule-based approach (Baseline 3) demonstrates that the metadata adds most of the predictive power. The linear model is better at taking into account multiple features as compared to the rule-based approach. As shown in the right-side of Figure 1, our model achieves an average F1 score, precision and recall of 0.61, 0.67 and 0.65, respectively.
To estimate the effect that the high WER from \cite{Zhang}'s speech transcription model has on our overall method, we also apply our two-stage NER system to the NPR podcast used to evaluate \cite{ritter2011named}'s NER model. We find that our system correctly identifies 27 true film mentions in the human-curated transcript as opposed to 22 true mentions in the computer-generated transcript. As such, we believe our system has room to improve given access to more accurate speech transcription models
Future research will explore two key directions. First, we plan to include additional film metadata fields such as release date, production studio, and cast members as features in the candidate classification model. We hypothesize that several of these fields can be represented within the model in a similar fashion to keyword mentions. Second, we plan to source a larger and more granularly labeled set of podcast transcript data to allow the use of data-greedy sequence learning models.
\bibliographystyle{ACM-Reference-Format}
\balance
|
\section{Introduction}\label{sectionIntro}
\hspace{12pt}We start with some background. In~\cite{ErdosPurdy}, Erd\H{o}s and Purdy posed the following problem. Given a set $P$ of $n$ points in the plane, not all on a line, how many new points (different from points in $P$) have to be chosen so that every line spanned by $P$ meets a new point? They conjectured that the answer is at least $(1 + o(1))n$ lines. We refer to the set $P$ as the \emph{initial points} and to the newly chosen points as the \emph{blocking points}. We also say that a blocking point $x$ \emph{blocks} a line $l$ if $x \in l$. The current record is due to Pinchasi~\cite{Pinch}.
\begin{theorem}[Pinchasi~\cite{Pinch}] Given a set $P$ of $n$ points in the plane, not all on a line, we need at least $\frac{n-1}{3}$ blocking points to block all lines spanned by $P$.\end{theorem}
It is worth mentioning that the first bounds of the form $\Omega(n)$ were due to Szemer\'edi and Trotter~\cite{SzTrotter} and Beck~\cite{Beck}. Those papers are in fact about Dirac's conjecture, which states that for a set $\mathcal{L}$ of $n$ non-concurrent lines in the plane, there is a line $l \in \mathcal{L}$ that forms at least $\frac{n}{2} - O(1)$ different intersection points with lines in $\mathcal{L}$. Szemer\'edi and Trotter, and Beck, prove that there is a line with $\Omega(n)$ different intersection points, which implies in particular that at least $\Omega(n)$ blocking points are necessary. Let us also remark that in~\cite{SzTrotter}, Szemer\'edi and Trotter prove their celebrated theorem on the number of incidence between lines and points, and they use this theorem to deduce the weak form of Dirac's conjecture. On the other hand, Pinchasi's proof is purely combinatorial and avoids any use of the incidence theorem.\\
In the same paper~\cite{ErdosPurdy}, Erd\H{o}s and Purdy propose a variant of the question where $P$ has to be in general position (i.e. no three points are collinear), but remark that Gr\"{u}nbaum pointed out to them that $2\lfloor\frac{n}{2}\rfloor$ new points suffice. It is easy to see that we need at least $n$ new points when $n$ is odd, and $n-1$ new points when $n$ is even, so this remark actually has a typo, and should probably be $2\lceil\frac{n}{2} \rceil -1$. However, with this obvious typo corrected, what Erd\H{o}s and Purdy actually imply is that there are examples where $n-1$ new points are sufficient, for infinitely many $n$. Therefore, the right question is: when do $n-1$ new points suffice? This was answered by Ackerman, Buchin, Knauer, Pinchasi and Rote in~\cite{magic}.
\begin{theorem}[Ackerman, Buchin, Knauer, Pinchasi and Rote~\cite{magic}]\label{blockingTheorem} Let $P$ be a set of $n \geq 5$ points in general position and let $B$ be a set of some other points in the plane, such that every line determined by two points in $P$ meets a point in $B$. Then $|B| \geq n$.\end{theorem}
Observe immediately that the theorem is trivial when $n$ is odd. Indeed, if $n$ is odd, each point in $B$ can meet at most $\lfloor \frac{n}{2} \rfloor = \frac{n-1}{2}$ lines spanned by $P$, so at least $n$ points are needed. (If $n$ is even, the same counting argument gives only the bound of $n-1$.) On the other hand, the fact that there is a non-trivial example for $n=4$ shows that we cannot hope for such a short argument in the general case. Similarly, the theorem may naturally be compared to the Sylvester-Gallai theorem (posed by Sylvester~\cite{SylvesterProblem} and solved by Gallai~\cite{GallaiSolution}), but once again, the $n=4$ case tells us that we can expect the proof to be more involved than, for example, looking at the minimal height of a triangular region formed in the dual (as in the usual proof of the Sylvester-Gallai theorem).\\
\indent Let us also remark that the regular $n$-gon in the projective plane with $n$ points on the line at infinity corresponding to directions of diagonals, show that for every $n$, $n$ blocking points suffice for certain configurations. Of course, using suitable transformations, we can make such examples affine.\\
Our proof of Theorem~\ref{blockingTheorem} is based on the following classification theorem (Theorem~\ref{informalClassification}), which is the main result of this paper. At this stage, we state this theorem informally, as otherwise the definitions would occupy significant part of the introduction. For an arbitrary convex polygon $R$, we write $\partial R$ for the \emph{boundary} of $R$, which is the union of its edges.\\
\indent Before we state the classification theorem, let us very briefly explain what the theorem is about. Namely, we consider a triangular region which has two sets of segments $\mathcal{S}$ and $\mathcal{B}$ inside it, with vertices on the boundary of the triangle. These have the property that whenever two segments in $\mathcal{S}$ intersect, then there is a unique segment in $\mathcal{B}$ passing through their intersection. Also, when a segment in $\mathcal{S}$ meets a segment in $\mathcal{B}$, then there is a unique other segment in $\mathcal{S}$ passing through their intersection. These two conditions come from considering the dual of the hypothetical extremal arrangement of points in general position and their blocking points, and are requirements \textbf{(i)} and \textbf{(ii)} in the classification theorem. Our aim is
to classify all such collections of segments that satisfy an additional condition. This is the condition \textbf{(iii)} in the statement. Let us remark here that although this condition looks somewhat artificial compared to the other two, it actually develops naturally in the proof of Theorem~\ref{blockingTheorem}.
\begin{theorem}\label{informalClassification} Let $T$ be a triangle, with edges $e_1, e_2, e_3$, and let $\mathcal{S}$ and $\mathcal{B}$ be collections of segments inside $T$, with endpoints on the edges of $T$, but no internal point of a segment meets $\partial T$. Write $\overline{\mathcal{S}} = \mathcal{S} \cup \{e_1, e_2, e_3\}$. Suppose that
\begin{itemize}
\item[\textbf{(i)}] No three segments of $\overline{\mathcal{S}}$ are concurrent, and for any two such segments that intersect at a point $p$, there is a unique segment $\beta(p) \in \mathcal{B}$ that passes through $p$, except possibly when the two segments are edges of $T$, in which case there might not be any such segment in $\mathcal{B}$.
\item[\textbf{(ii)}] For every intersection $p$ of a segment in $\overline{\mathcal{S}}$ and a segment in $\mathcal{B}$, there is a unique second segment in $\overline{\mathcal{S}}$ that passes through $p$.
\item[\textbf{(iii)}] In every minimal $\overline{\mathcal{S}}$-region $R$, for any consecutive vertices $v_1, v_2, v_3, v_4, v_5$ appearing in this order on $\partial R$, we have that, if $l(v_1v_2)$ and $\beta(v_3)$ intersect in $T$, and $\beta(v_3)$ crosses the interior of $R$, then $l(v_1v_2), \beta(v_3), l(v_4v_5)$ are concurrent.\end{itemize}
Then the configuration formed by $T, \mathcal{S}$ and $\mathcal{B}$ must have one of the structures shown in the Figure~\ref{structuresFigure}.
\end{theorem}
\begin{figure}
\includegraphics[width=\textwidth]{StructureTheoremQuotes.pdf}\\[24pt]
\includegraphics[width=\textwidth]{StructureTheoremQuotes2.pdf}
\caption{Types of triangle blocking arrangements}
\label{structuresFigure}
\end{figure}
We call an arrangement satisfying conditions \textbf{(i)} and \textbf{(ii)} a \emph{triangle blocking arrangement}. If a triangle blocking arrangement~additionally satisfies the condition \textbf{(iii)}, then we call it a \emph{strong triangle blocking arrangement}.
Given that there is such a strong structure theorem in this setting, it is plausible that an extremal result could hold. Recall that Theorem~\ref{blockingTheorem} resembles the Sylvester-Gallai theorem, which has its extremal version in the following theorem of Green and Tao. For a given set of points in the plane, we say that a line is \emph{ordinary} if it passes through exactly two points in the set.
\begin{theorem}[Green and Tao~\cite{GreenTao}] \label{GTord}There is an $n_0$ such that, whenever we have $n \geq n_0$ points in the plane that span at most $\frac{n}{2}$ ordinary lines, there is a cubic curve containing the given points.\end{theorem}
With this in mind, we formulate the following conjecture.
\begin{conjecture} Suppose that $P$ is a set of $n$ points in the plane in general position, and let $B$ be a set of blocking points for $P$. If $|B| = n$, then $P \cup B$ lie on a cubic curve.\end{conjecture}
We postpone the discussion of the connection between proof of Theorem~\ref{GTord} and the classification theorem (Theorem~\ref{informalClassification}) we prove here to the concluding remarks. There we also discuss why the condition \textbf{(iii)} is necessary in the classification theorem.
The plan of the paper is as follows. In the next section we describe the classification theorem. Then, in Section~\ref{sectionDeduct}, we see how to deduce Theorem~\ref{blockingTheorem} from the classification theorem. In Section~\ref{sectionClassificationProof}, we prove the classification theorem. This is actually the main part of the paper.
\section{Detailed description of the structural theorem}\label{sectionDesc}
\hspace{12pt}Before stating the classification result, we first need to introduce some terminology. Recall that \emph{triangle blocking arrangement}~is a triple $\Delta = (T, \mathcal{S}, \mathcal{B})$ consisting of a triangle $T$, with vertices $x_1, x_2$ and $x_3$, a two collections of segments $\mathcal{S}$ and $\mathcal{B}$ such that the endpoints of each segment lie on the boundary of $\partial T$ (possibly coinciding with some vertex $x_i$), and no interior point of a segment lies on the boundary $\partial T$, and the following intersection condition is satisfied: for every pair of segments in $\overline{\mathcal{S}} \colon= \mathcal{S} \cup \{x_1 x_2, x_2 x_3, x_3 x_1\}$, except possibly the pairs of sides of $T$, if they intersect, there is a unique segment in $\mathcal{B}$ that passes through their intersection, and for every intersecting pair of segments, where one segment is in $\overline{\mathcal{S}}$ and the other in $\mathcal{B}$, there is a unique second segment in $\overline{\mathcal{S}}$ that passes through their intersection. We call the elements of $\overline{\mathcal{S}}$ the \emph{initial segments}, the elements of $\mathcal{S}$ the \emph{proper initial segments} when we have to distinguish them from $\mathcal{S}$, and the elements of $\mathcal{B}$ the \emph{blocking segments}. Furthermore, if $x, y$ are two points on a segment $s \in \overline{\mathcal{S}} \cup \mathcal{B}$, we also say that $xy$ is \emph{initial segment} if $s \in \overline{\mathcal{S}}$, that $xy$ is \emph{proper initial segment} if $s \in \mathcal{S}$, and that $xy$ is \emph{blocking segment} if $s \in \mathcal{B}$. We refer to intersections of initial segments as \emph{vertices} in $T$. We will write $\beta(v)$ for the unique blocking segment through the vertex $v$, and in general, for any two points $x,y$ we write $l(xy)$ for the line through $x$ and $y$.\\
\indent Given a triangle $T'$, whose vertices lie in $T$ (possibly on the edges of $T$), and whose edges are subsets of segments in $\overline{\mathcal{S}}$, we define $\overline{\mathcal{S}_{T'}} \colon= \{T' \cap s\colon s \in \overline{\mathcal{S}}\}$, $\mathcal{B}_{T'} \colon= \{T' \cap s\colon s \in \mathcal{B}\}$ and $\mathcal{S}_{T'} \colon= \overline{\mathcal{S}_{T'}} \setminus \{\text{edges of }T'\}$. We say that $\Delta' = (T', \mathcal{S}_{T'}, \mathcal{B}_{T'})$ is a \emph{sub-triangle blocking arrangement}~of $\Delta$, \emph{induced by} $T'$. One can check that sub-triangle blocking arrangement~is itself a triangle blocking arrangement.\\
We now define some special types of triangle blocking arrangements. Let us stress that these do not include all possible triangle blocking arrangements. An example not included in definitions is shown in the Figure~\ref{TBAnotInDef}.\\[6pt]
\begin{figure}
\begin{center}
\includegraphics[width = 0.7\textwidth]{counterexample.pdf}
\caption{Example of a~triangle blocking arrangement~not among definitions}
\label{TBAnotInDef}
\end{center}
\end{figure}
\noindent\textbf{Basic types }$\tbty{B}{0}, \tbty{B}{1}, \tbty{B}{2}, \tbty{B}{3}$\textbf{.} We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\tbty{B}{0}$ if $\mathcal{S} = \mathcal{B} = \emptyset$.\\
We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\tbty{B}{1}$ if it has the following structure. There is an ordering $y_1, y_2, y_3$ of $x_1, x_2, x_3$ such that the vertices of $\mathcal{S} \cup \mathcal{B}$ are $v_1, v_2, \dots,$ $v_n \in y_1y_2$ and $u_1, u_2, \dots,$ $u_n \in y_1y_3$, with ordering $y_1, v_1, v_2, \dots,$ $v_n, y_2$ on $y_1y_2$ and ordering $y_1, u_1, u_2, \dots,$ $u_n, y_3$ on $y_1y_3$, such that $\mathcal{S} = \{v_1u_1, v_2u_2, \dots,$ $v_nu_n\}$ and $\mathcal{B} = \{v_1 u_2,$ $v_2 u_1,$ $v_3u_4,$ $v_4u_3, \dots,$ $v_{t-1} u_t,$ $v_t u_{t-1}\}$, where $t = n$ if $n$ is even, and $t = n+1$, if $n$ is odd, and in that case, we set $v_{n+1} = y_2, u_{n+1} = y_3$. We say that $y_1$ is the \emph{first vertex} of $\Delta$. Also, when $n$ is even, we say that $\Delta$ is \emph{even}, and if $n$ is odd, we say that $\Delta$ is \emph{odd}.\\
We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\tbty{B}{2}$ if it has the following structure. There is an ordering $y_1, y_2, y_3$ of $x_1, x_2, x_3$ such that the vertices of $\mathcal{S} \cup \mathcal{B}$ are $v_1, v_2, \dots, v_n \in y_1y_2$, then $u_1, u_2, \dots,$ $u_n,$ $u'_m, \dots,$ $u'_1 \in y_1y_3$ and $w_1, w_2, \dots,$ $w_m \in y_3y_2$, with ordering $y_1, v_1, v_2, \dots,$ $v_n, y_2$ on $y_1y_2$, ordering $y_1, u_1, u_2, \dots,$ $u_n, u'_m, \dots,$ $u'_1, y_3$ on $y_1y_3$, and ordering $y_3, w_1, \dots,$ $w_m, y_2$ on $y_3y_2$, and $m$ and $n$ are even. The segments are $\mathcal{S} = \{v_1u_1, v_2u_2, \dots,$ $v_nu_n,$ $u'_1w_1, \dots,$ $u'_mw_m\}$ and $\mathcal{B} = \{v_1 u_2, v_2 u_1,$ $v_3u_4, v_4u_3, \dots,$ $v_{n-1} u_n, v_n u_{n-1}\}$ $\cup$ $\{u'_1w_2, u'_2w_1, \dots,$ $u'_nw_{n-1}, u'_{n-1}w_n\}$.\\
We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\tbty{B}{3}$ if it has the following structure. There is an ordering $y_1, y_2, y_3$ of $x_1, x_2, x_3$ such that the vertices of $\mathcal{S} \cup \mathcal{B}$ are $v_1, \dots, v_k,$ $v'_l, \dots, v'_1 \in y_1y_2$, $w_1, \dots, w_l,$ $w'_m, \dots,$ $w'_1 \in y_2y_3$, $u_1, \dots, u_m,$ $u'_k, \dots,$ $u'_1 \in y_3y_1$, appearing in orders $y_1, v_1, \dots, v_k,$ $v'_l, \dots, v'_1, y_2$ on $y_1y_2$, $y_2, w_1, \dots, w_l,$ $w'_m, \dots, w'_1, y_3$ on $y_2y_3$, and $y_3, u_1, \dots, u_m,$ $u'_k, \dots, u'_1, y_1$ on $y_3y_1$, and $k,l,m$ are of the same parity.\\
\indent When $k,l,m$ are even, then the segments are $\mathcal{S} = \{v_1u'_1, \dots,$ $v_ku'_k\} \cup$ $\{w_1v'_1, \dots,$ $w_lv'_l\} \cup$ $\{u_1w'_1, \dots,$ $u_mw'_m\}$ and $\mathcal{B} = \{v_1u'_2,$ $v_2u'_1, \dots,$ $v_{k-1}u'_k, v_ku'_{k-1}\}$ $\cup$ $\{w_1v'_2, w_2v'_1, \dots,$ $w_{l-1}v'_l, w_l v'_{l-1}\}$ $\cup$ $\{u_1w'_2, u_2w'_1, \dots,$ $u_{m-1}w'_m, u_m w'_{m-1}\}$.\\
\indent When $k,l,m$ are odd, then the segments are $\mathcal{S} = \{v_1u'_1, \dots,$ $v_ku'_k\}$ $\cup$ $\{w_1v'_1, \dots,$ $w_lv'_l\}$ $\cup$ $\{u_1w'_1, \dots,$ $u_mw'_m\}$ and $\mathcal{B} = \{v_1u'_2,$ $v_2u'_1, \dots,$ $v_{k-1}u'_{k-2},$ $v_{k-2}u'_{k-1}\}$ $\cup$ $\{w_1v'_2, w_2v'_1, \dots,$ $w_{l-1}v'_{l-2},$ $w_{l-2} v'_{l-1}\}$ $\cup$ $\{u_1w'_2,$ $u_2w'_1, \dots,$ $u_{m-1}w'_{m-2},$ $u_{m-2} w'_{m-1}\}$ $\cup$ $\{v_kw'_m,$ $v'_lu_m, w_l u'_k\}$.\\
\noindent\textbf{Intersecting types }$\tbty{I}{1}, \tbty{I}{2}$\textbf{.} We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\tbty{I}{1}$ if it has the following structure. There is an ordering $y_1, y_2, y_3$ of $x_1, x_2, x_3$ and there are vertices $z_2 \in y_1y_2, z_3 \in y_1y_3$ such that
\begin{enumerate}
\item The segment $z_2z_3$ is in $\mathcal{S}$, and the sub-triangle blocking arrangement in the triangle $T' \colon= y_1z_2z_3$ is of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with $y_1$ as the first vertex and it is even.
\item There are vertices $v_1, v_2, \dots, v_{2k}$ on $y_2z_2$, appearing in that order from $y_2$ to $z_2$, and there are vertices $u_1, u_2, \dots, u_{2k}$ on $y_2y_3$, appearing in that order from $y_2$ to $y_3$, such that
\[\mathcal{S} \setminus \overline{\mathcal{S}_{T'}} = \{u_1v_2, u_2v_1, u_3v_4, u_4v_3, \dots, u_{2k - 1}v_{2k}, u_{2k}v_{2k-1}\}\]
and
\[\begin{split}\mathcal{B} \setminus \mathcal{B}_{T'} = &\{u_1 v_1, u_2v_3, u_3v_2, u_4v_5, u_5v_4, \dots, u_{2k-2}v_{2k-1}, u_{2k-1}v_{2k-2}\} \\\hspace{1cm}&\cup \{y_2z_3, y_3 v_{2k}, z_2u_{2k}\}.\end{split}\]
\item For $i=1,2, \dots, k$, the segments $u_{2i-1}v_{2i}, u_{2i}v_{2i-1}, y_2z_3$ are concurrent. Let $p_i$ be their intersection point. The intersections between $\overline{\mathcal{S}}$ and $\mathcal{B}$, and the intersections of pairs of segments in $\overline{\mathcal{S}}$ are either on $\partial T$, or the points $p_i$.
\end{enumerate}
We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\tbty{I}{2}$ if it has the following structure. There is an ordering $y_1, y_2, y_3$ of $x_1, x_2, x_3$ and there are vertices $z_2 \in y_1y_2, z_3 \in y_1y_3$ such that
\begin{enumerate}
\item The segment $z_2z_3$ is in $\mathcal{S}$, and the sub-triangle blocking arrangement~in the triangle $T' \colon= y_1z_2z_3$ is of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with $y_1$ as the first vertex and it is even.
\item The vertices from $y_2$ to $z_2$ are $u_1, u_2, \dots, u_{2k}$, in that order, from $y_2$ to $y_3$ are $v_1, v_2, \dots, v_{2k}, w_{2l}, w_{2l-1}, \dots, w_1$, in that order, and from $y_3$ to $z_3$ are $t_1, t_2, \dots, t_{2l}$, in that order. The segments of $\Delta$ are
\[\begin{split}\mathcal{S} \setminus \overline{\mathcal{S}_{T'}} = &\{u_1 v_2, u_2 v_1, \dots, u_{2k-1}v_{2k}, u_{2k}v_{2k-1}\} \\
&\hspace{1cm}\cup \{w_1t_2, w_2t_1, \dots, w_{2l-1}t_{2l}, w_{2l}t_{2l-1}\}\end{split}\]
and
\[\begin{split}\mathcal{B} \setminus \mathcal{B}_{T'} = &\{u_1 v_1, u_2v_3, u_3v_2,\dots, u_{2k-2}v_{2k-1}, u_{2k-1}v_{2k-2}\}\\
&\hspace{1cm}\cup\{w_1 t_1, w_2t_3, w_3t_2, \dots, w_{2k-2}t_{2k-1}, w_{2k-1}t_{2k-2}\}\\
&\hspace{1cm} \cup \{u_{2k}y_3, t_{2l}y_2\} \cup \mathcal{B}',\end{split}\]
where $\mathcal{B}' = \{v_{2k}z_3, w_{2l} z_2\}$ or $\mathcal{B}' = \{v_{2k}z_2, w_{2l} z_3\}$.
\item For $i=1,2, \dots, k$, the segments $u_{2i-1}v_{2i}, u_{2i}v_{2i-1}, y_2t_{2l}$ are concurrent, at point $p_i$, and for $i=1,2, \dots, l$, the segments $w_{2i-1}t_{2i}, w_{2i}t_{2i-1}, y_3u_{2k}$ are concurrent, at point $q_i$. The intersections between segments in $\overline{\mathcal{S}}$ and $\mathcal{B}$, and the intersections of pairs of segments in $\overline{\mathcal{S}}$ are either on $\partial T$, or the points $p_i$ and $q_i$.
\end{enumerate}
\noindent\textbf{Triangular type }$\mathbf{T}$\textbf{.} We say a triangle blocking arrangement~$\Delta = (T, \mathcal{S}, \mathcal{B})$ is of type $\mathbf{T}$ if it has the following structure. There is an integer $k \geq 2$ and there are vertices $u_1, u_2, \dots, u_{2k} \in x_1x_2$, appearing in that order from $x_1$ to $x_2$, vertices $v_1, v_2, \dots, v_{2k} \in x_2x_3$, appearing in that order from $x_2$ to $x_3$, and vertices $w_1, w_2, \dots, w_{2k} \in x_3x_1$, appearing in that order from $x_3$ to $x_1$.
\begin{enumerate}
\item Segments $u_1 w_{2k}, v_1 u_{2k}$ and $w_1 v_{2k}$ belong to $\mathcal{S}$, and writing $T_1 = x_1 u_1 w_{2k},$ $T_2 = x_2 v_1 u_{2k}$ and $T_3 = x_3 w_1 v_{2k}$, each sub-triangle blocking arrangement~induced by $T_i$ is of type $\tbty{B}{0}$ or $\tbty{B}{1}$ with $x_i$ as the first vertex, and it is even.
\item The segments are given by
\[\begin{split}\mathcal{S} \setminus \cup_{i=1}^3 \mathcal{S}_{T_i} = \{u_{2i - 1} w_{2k + 2 - 2i}\colon i \in [k]\} &\cup \{v_{2i-1} u_{2k + 2 - 2i}\colon i \in [k]\}\\
&\cup \{w_{2i-1} v_{2k + 2 - 2i}\colon i \in [k]\}\end{split}\]
and
\[\begin{split}\mathcal{B} \setminus \cup_{i=1}^3 \mathcal{B}_{T_i} = \{u_{2i} w_{2k + 1 - 2i}\colon i \in [k]\} &\cup \{v_{2i} u_{2k + 1 - 2i}\colon i \in [k]\} \\
&\cup \{w_{2i} v_{2k + 1 - 2i}\colon i \in [k]\}.\end{split}\]
\item For every triple $(a,b,c) \in [2k]^3$ such that not all of $a,b,c$ are even and $a+b+c = 4k + 2$, the triple of segments $u_a w_{2k+1-a}, v_b u_{2k+1 -b}, w_c v_{2k+1-c}$ is concurrent at the point $p_{a,b,c}$. The intersections between segments in $\overline{\mathcal{S}}$ and $\mathcal{B}$, and the intersections of pairs of segments in $\overline{\mathcal{S}}$ are either on $\partial T$, or the points $p_{a,b,c}$.
\end{enumerate}
We remark that allowing $k = 1$ in the definition of type $\mathbf{T}$ would actually give type $\tbty{B}{3}$. We keep $\tbty{B}{3}$ as a basic type, as in this case the intersections between initial segments lie on $\partial T$ only, while in the type $\mathbf{T}$ we insist on having at least one intersection of initial segments that is in the interior of $T$.\\
Classifying all triangle blocking arrangements~currently seems out of reach, however we are able to prove the following. (The assumption \textbf{(A)} below is exactly the assumption \textbf{(iii)} of Theorem~\ref{informalClassification}.)
\begin{theorem}[A classification theorem for triangle blocking arrangements.]\label{classificationTheorem} Suppose that $\Delta = (T, \mathcal{S}, \mathcal{B})$ is a strong triangle blocking arrangement, i.e. a triangle blocking arrangement~such that the following assumption \textbf{(A)} holds.\begin{itemize}
\item[\textbf{(A)}] In every minimal $\overline{\mathcal{S}}$-region $R$, for any consecutive vertices $v_1, v_2, v_3, v_4, v_5$ appearing in this order on $\partial R$, we have that, if $l(v_1v_2)$ and $\beta(v_3)$ intersect in $T$, and $\beta(v_3)$ meets the interior of $R$, then $l(v_1v_2), \beta(v_3), l(v_4v_5)$ are concurrent.\end{itemize}
Then $\Delta$ has one of the types among $\tbty{B}{0}, \tbty{B}{1}, \tbty{B}{2}, \tbty{B}{3}, \tbty{I}{1}, \tbty{I}{2}$ and $\mathbf{T}$.
\end{theorem}
\section{Deducing Theorem~\ref{blockingTheorem} from the Classification theorem}\label{sectionDeduct}
\hspace{12pt}In this section we prove Theorem~\ref{blockingTheorem}. Immediately, we move to the dual, where the theorem has the following formulation. For $n \geq 4$, we say that a pair of disjoint sets $(\mathcal{L}, \mathcal{B})$ of lines in $\mathbb{P}^2 = \mathbb{P}^2(\mathbb{R})$ is an \emph{$n$-blocking configuration} if $|\mathcal{L}| = |\mathcal{P}| + 1 = n$, no three lines in $\mathcal{L}$ are concurrent, and for every pair $l_1, l_2$ of lines in $\mathcal{L}$ there is a unique line in $\mathcal{B}$ that passes through $l_1 \cap l_2$. We refer to the lines in $\mathcal{L}$ as the \emph{initial lines}, and to the lines in $\mathcal{B}$ as the blocking lines.
\begin{theorem}\label{dualBlockingTheorem}Let $n \geq 4$ and let $(\mathcal{L}, \mathcal{B})$ be an $n$-blocking configuration. Then $n = 4$.\end{theorem}
We begin the proof by deducing some structural information about the configuration of lines in $\mathcal{L} \cup \mathcal{B}$, which will enable us to apply Theorem~\ref{classificationTheorem} and deduce Theorem~\ref{dualBlockingTheorem}.\\
By an \emph{$\mathcal{L}$-region}, we mean the closure of any connected subset of $\mathbb{P}^2 \setminus L$, where $L$ is a subset of some lines in $\mathcal{L}$. A \emph{minimal $\mathcal{L}$-region} is the closure of a connected subset of $\mathbb{P}^2 \setminus \cup \mathcal{L}$. For a $\mathcal{L}$-region $R$, we define its \emph{edges} to be the segments of lines in $\mathcal{L}$ that intersect $R$, and \emph{vertices} as the intersections of lines in $\mathcal{L}$ that lie in $R$. Finally, a vertex $v$ of $\mathcal{L}$-region is \emph{internally blocked} if the unique blocking line through $v$ meets the interior of $R$, otherwise, it is \emph{externally blocked}.
\begin{lemma}\label{evenInternalBlocks} Let $n \geq 4$ and let $(\mathcal{L}, \mathcal{B})$ be an $n$-blocking configuration. Let $R$ be any $\mathcal{L}$-region, not necessarily minimal. Then the number of internally blocked vertices of $R$ is even.\end{lemma}
\begin{proof} We proceed by a double-counting argument. Draw $l \cap R$ for all the initial lines $l$ that meet $R$, partitioning $R$ into minimal $\mathcal{L}$-regions $R_1, R_2, \dots, R_t$. Observe that the total number $N$ of the internally blocked vertices of the regions $R_1, R_2, \dots, R_t$ can be written as $N = N_1 + N_2 + N_3$, where $N_1$ is the number of internally blocked vertices of $R$, $N_2$ is the number of internally blocked vertices which lie on the interiors of edges of $R$ and $N_3$ is the number of internally blocked vertices that lie in the interior of $R$. Our goal is to show that $N_1$ is even.
Every blocking line $b$ that crosses $R$ intersects $\partial R$ at two points, and these contribute to $N$ by 2, and $b$ also passes through some vertices in $\operatorname{int} R$. But, each such vertex is blocked internally two times by $b$, for some minimal $\mathcal{L}$-regions. Hence, $N$ is even.
Observe that every initial line $l$ that meets $R$, but is not one of its edges, intersects $\partial R$ twice at interiors of edges of $R$. Thus, these two intersections contribute to $N_2$ by 2, and every such intersection is defined by a unique such initial line $l$. This shows that $N_2$ is even.
Finally, every vertex in $\operatorname{int} R$ is internally blocked twice, so $N_3$ is even, hence $N_1 = N - N_2 - N_3$ is also even, as desired.\qed\end{proof}
\begin{lemma} Let $n \geq 4$ and let $(\mathcal{L}, \mathcal{B})$ be an $n$-blocking configuration. Let $R$ be any minimal $\mathcal{L}$-region. Then either all vertices of $R$ are internally blocked or all vertices of $R$ are externally blocked.\end{lemma}
\begin{proof} Suppose contrary, $R$ has two consecutive vertices $u$ and $v$, such that $u$ is internally blocked, but $v$ is externally blocked. Since a blocking line meets $\operatorname{int} R$, $R$ cannot be a triangle. Let $u', v'$ be another two vertices of $R$, such that $u', u, v, v'$ are consecutive. Let $p = u'u \cap v'v$. Consider $\mathcal{L}$-region $S$ with vertices $u,v,p$. Inside $S$, among $u$ and $v$, exactly one is internally blocked vertex. Therefore, $p$ is an internally blocked vertex of $S$, with a blocking line $b \in \mathcal{B}$. But $b$ must cross the interior of $uv$, which is a contradiction with the fact that $R$ is a minimal $\mathcal{L}$-region.\qed\end{proof}
\begin{lemma} Let $n \geq 4$ and let $(\mathcal{L}, \mathcal{B})$ be an $n$-blocking configuration. Let $R$ be any minimal $\mathcal{L}$-region, with vertices $v_1, v_2, \dots, v_k$, sorted in the order as they appear on $\partial R$. Suppose that the vertices of $R$ are internally blocked. Then $k$ is even, and for every $i, j \in [k/2]$, the line $v_i v_{i + k/2} \in \mathcal{B}$, and the lines $v_{i - j - 1} v_{i-j}, v_i v_{i + k/2}, v_{i+j} v_{i + j + 1}$ are concurrent, (indices of vertices are taken modulo $k$). \end{lemma}
\begin{proof} Let $b_i$ be the blocking line at the vertex $v_i$, and let $l_i = v_i v_{i+1} \in \mathcal{L}$. We first prove that $l_{i - j - 1}, b_i, l_{i + j}$ are concurrent by induction on $j \in \{0, 1, 2, \dots, k/2\}$. Observe that when $j = k/2$, we have that $l_{i - k/2 - 1}, b_i, l_{i + k/2}$ are concurrent. But the lines $l_{i - k/2 - 1} = v_{i - k/2 - 1} v_{i - k/2}$ and $l_{i + k/2} = v_{i + k/2}v_{i + k/2 + 1}$ already meet at $v_{i + k/2}$, which is blocked by $b_{i + k/2}$, so by uniqueness of blocking lines $b_i = b_{i + k/2} = v_i v_{i + k/2}$.\\[3pt]
\indent For the base of induction, when $j = 0$, the lines $l_{i - 1}, b_i, l_{i}$ meet at $v_i$, so the claim holds.\\
\indent Suppose that the claim holds for some $0 \leq j < k/2$, and consider $l_{i - j - 2},$ $b_i,$ $l_{i + j + 1}$. Look at the $\mathcal{L}$-region $S$ formed by lines $l_{i - j - 2}, l_{i - j - 1}, l_{i + j}, l_{i + j + 1}$. By induction hypothesis, the triple of lines $l_{i - j - 1}, b_i, l_{i+j}$ is concurrent with the common point $p_1$. Let $p_2 \colon= l_{i-j-2} \cap l_{i+j}, p_3 \colon= l_{i-j-1} \cap l_{i+j+1}$ and $p_4 \colon= l_{i-j-2} \cap l_{i + j + 1}$, so the vertices of $S$ are precisely $p_1, p_2, p_3$ and $p_4$, and our goal is to show that $p_4 \in b_i$. Look at $\mathcal{L}$-region with vertices $v_{i-j-1}, p_1, p_2$, formed by initial lines $l_{i-j-2}, l_{i-j-1}$ and $l_{i+j}$. For this region, $v_{i-j-1}$ and $p_1$ are externally blocked, so by Lemma~\ref{evenInternalBlocks}, the vertex $p_2$ must be externally blocked as well, therefore $p_2$ is internally blocked in $S$. Likewise, the vertex $p_3$ is internally blocked in $S$, so, since $p_1$ is also internally blocked, by Lemma~\ref{evenInternalBlocks}, the remaining vertex $p_4$ is internally blocked, by some blocking line $b$. But $b$ meets $R$ between vertices $v_{i - j - 1}$ and $v_{i + j}$, so it must contain some vertex $v_l$, $l \in \{i - j, i - j + 1, \dots, i + j - 1\}$. By induction hypothesis, $b_l$ meets $l_{i-j-2}$ at a point other than $p_4$, when $l < i$, and $b_l$ meets $l_{i+j + 1}$ at point other than $p_4$, for $l > i$, proving that $l = i$, as desired.\qed\end{proof}
Having acquired enough structural information about blocking configurations, we are ready to prove Theorem~\ref{dualBlockingTheorem}.
\begin{proof}[Proof of Theorem~\ref{dualBlockingTheorem}.] Observe that since no three lines in $\mathcal{L}$ are concurrent, there are $m_1, m_2, m_3 \in \mathcal{L}$ that define a region which is a minimal $\mathcal{L}$-region (this follows from the fact that any line that crosses a triangle splits that triangle into two regions, one of which is also a triangle). Denote the other $\mathcal{L}$-regions formed by $m_1, m_2$ and $m_3$ by $S_1, S_2$ and $S_3$. Applying Theorem~\ref{classificationTheorem} to the triangle blocking arrangements~induced by $S_1, S_2$ and $S_3$, we see that each of them is of type $\tbty{B}{1}$. Let $x_1 \colon= m_2 \cap m_3, x_2 \colon= m_3 \cap m_1, x_3 \colon= m_1 \cap m_2$ and let $e_1 \colon= m_1 \cap S, e_2 \colon= m_2 \cap S, e_3 = \colon= m_3 \cap S$. Let $a_1, a_2, \dots, a_r$ be the vertices on $m_1 \setminus e_1$, listed from $x_2$ to $x_3$, let $b_1, b_2, \dots, b_s$ be the vertices on $m_2 \setminus e_2$, listed from $x_3$ to $x_1$, and finally let $c_1, c_2, \dots, c_t$ be the vertices of $m_3 \setminus e_3$, listed from $x_1$ to $x_2$. By the definition of type $\tbty{B}{1}$, we have $r = s = t$ and $a_i b_{r + 1-i}, b_i c_{r+1-i}, c_i a_{r+1- i} \in \mathcal{L}$, for $i \in [r]$. However, $a_1$ belongs to the initial lines $a_1 b_r, a_1c_r$ and $m_1$, and $m_1$ is different from $a_1b_r$ and $a_1 c_r$, therefore $a_1b_r = a_1c_r$, making $a_1, b_r, c_r$ collinear. Similarly, $b_r, a_1, c_1$ are collinear, so $c_1, c_r, b_r, a_1$ are collinear. If $r > 1$, then $c_1 \not= c_r$, but $c_1, c_r \in m_3$, making $a_1b_r = m_3$, which is a contradiction. Therefore $r = 1$, so $n=4$, as desired.\qed\end{proof}
\section{Proof of the Classification theorem}\label{sectionClassificationProof}
\hspace{12pt}As the title suggests, this section is devoted to the proof of the Classification theorem. For an integer $n \geq 0$, we say that $n$-Classification theorem holds, if the conclusion of Theorem~\ref{classificationTheorem} holds for all triangle blocking arrangements~$\Delta = (T, \mathcal{S}, \mathcal{B})$ with $|\mathcal{S}| + |\mathcal{B}| \leq n$. We denote the quantity $|\mathcal{S}| + |\mathcal{B}|$ by $|\Delta|$ and call it the \emph{size} of $\Delta$. The argument will be based on induction on $|\Delta|$. Note that 0-Classification theorem holds, as $|\Delta| = 0$ implies that $\mathcal{S} = \mathcal{B} = \emptyset$, so $\Delta$ is of type $\tbty{B}{0}$.\\
Before we proceed with the proof, we need a couple of pieces of notation. Firstly, a segment is \emph{minimal} if there are no other vertices in its interior. Also recall that for a segment $xy$, we write $l(xy)$ for the line that contains the segment, and given a point $p$, write $\beta(p)$ for the blocking segment through $p$. Further, also for a segment $xy$, we write $s(xy)$ for the unique element of $\overline{\mathcal{S}}$ which contains both $x$ and $y$, if it exists. Given a $\overline{\mathcal{S}}$-region $R$, and a vertex $v$ of $R$, we say that $v$ is \emph{internally blocked} in $R$ if $\beta(v)$ passes through the interior of $R$, and otherwise we say that $v$ is \emph{externally blocked} in $R$. In particular, if there are no blocking segments through $x_i$, we say that $x_i$ is externally blocked. (By a vertex of $R$ here, we mean an intersection of initial segments that are edges of $R$, so a vertex on boundary, but in interior of an edge of $R$ is not counted as a vertex when we talk about internally or externally blocked vertices.)\\
We restate Lemma~\ref{evenInternalBlocks} here, which will be crucial to our work. As we shall be using this lemma all the time, we will not refer to it explicitly.
\begin{lemma}Suppose that $\Delta = (T, \mathcal{S}, \mathcal{B})$ is a triangle blocking arrangement, and let $R$ be any, not necessarily minimal, $\overline{\mathcal{S}}$-region. Then, the number of internally blocked vertices of $R$ is even.\end{lemma}
\begin{proposition} \label{ForbStruct1}Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that no blocking segment passes through $x_1$. Suppose that $s_1, s_2 \in \mathcal{S}$ are two initial segments, each with the property that one of the endpoints is on the edge $x_1x_2$ and the other is on the edge $x_1x_3$. Then $s_1$ and $s_2$ are disjoint.\end{proposition}
\begin{proof}Suppose contrary, let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, with vertices of $T$ given by $x_1, x_2$ and $x_3$, and suppose that no blocking segment passes through $x_1$, but two initial segments $a_2 a_3$ and $b_2 b_3$ intersect at the point $c$ and $a_2,b_2 \in x_1 x_2$, $a_3,b_3 \in x_1x_3$. Without loss of generality, the vertices appear in order $x_1, b_2, a_2, x_2$ and $x_1, a_3, b_3, x_3$ on the segments $x_1x_2$ and $x_1x_3$. We consider the following cases on the positions of blocking segments.
\begin{itemize}
\item[\textbf{Case 1}] Inside the region $a_2b_2c$, the vertices $b_2$ and $c$ are both internally blocked.
\item[\textbf{Case 2}] Inside the region $a_2b_2c$, the vertex $c$ is internally blocked, but $b_2$ is externally blocked.
\item[\textbf{Case 3}] Inside the region $a_2b_2c$, the vertex $c$ is externally blocked. Thus $c$ is internally blocked in the region $x_1 a_3 c b_2$, and by the parity of the number of internally blocked vertices, exactly one of $b_2$ and $a_3$ is internally blocked in that region, without loss of generality, $b_2$. Thus, $b_2$ is externally blocked in the region $a_2b_2c$.
\end{itemize}
We treat each case separately and we depict the steps of the proof in Figures~\ref{prop8case1Fig},~\ref{prop8case2Fig} and~\ref{prop8case3Fig}.\\
\noindent\framebox{\textbf{Case 1.}} Consider the triangle blocking arrangement~$\Delta_1$ induced by the triangle $x_1 a_2 a_3$. Since $|\Delta_1| < |\Delta| = n$, we may apply the classification theorem~to $\Delta_1$. All of $x_1, a_2, a_3$ are externally blocked, so the type of $\Delta_1$ is either $\tbty{B}{i}$, for some i, or $\mathbf{T}$.
\begin{figure}
\includegraphics[width = \textwidth]{prop8case1.pdf}
\caption{Case 1 of proof of Proposition~\ref{ForbStruct1}}
\label{prop8case1Fig}
\end{figure}
If the type of $\Delta_1$ is $\mathbf{T}$, then there is an initial segment $s \in \mathcal{S}$, whose restriction to $\Delta_1$ is $y_1y_2$, with $y_1 \in x_1a_3$ and $y_2 \in ca_3$. However, that implies that $s$ intersects segment $cb_3$, at some point $b_4$, so the triangle blocking arrangement~$\Delta_2$, induced by the triangle $x_1b_2b_3$ has the vertex $b_3$ externally blocked, and initial segments $y_1b_4$ and $a_3c$ intersect. However, this cannot occur in any of the seven types we defined, and by the $(n-1)$-classification theorem,~$\Delta_2$ has one of these types, which is a contradiction.\\
Hence, the type of $\Delta_1$ is one of the four basic types, and analogously the triangle blocking arrangement~$\Delta_2$ (induced by the triangle $x_1b_2b_3$) has also a basic type. By the definition of the basic types, if we write $d', d'', c', c''$ for the first vertices next to $c$ on $a_2c$, next to $c$ on $b_3c$, next to $b_2$ on $a_2b_2$, next to $a_3$ on $b_3a_3$, respectively, then $c'd', b_2d', cc', c''d'', a_3d''$ and $cc''$ are minimal segments, and $c'd' \in \mathcal{S}_{\Delta_1}, b_2d', cc' \in \mathcal{B}_{\Delta_1}, c''d'' \in \mathcal{S}_{\Delta_2}, a_3d'', cc'' \in \mathcal{B}_{\Delta_2}$. Furthermore, we also have that the segments $b_2c, ca_3$ are minimal. Let $p,q$ be vertices on $\partial T$ such that $c'p \colon= s(c'd')$ and $c''q\colon= s(c''d'')$.\\
\noindent\textbf{Claim.} The vertices $p$ and $q$ lie on $x_2x_3$.
\begin{proof}[Proof of the claim.] We prove the claim for $p$, the proof for $q$ is similar. Suppose contrary, $p \in x_1x_3$. Looking at triangle blocking arrangement~$\Delta_p$ induced by $x_1 c' p$, we see that $c'$ is internally blocked, and $|\Delta_p| < n$, so the $(n-1)$-classification theorem~implies that $\Delta_p$ has one of types $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$. We can immediately discard type $\tbty{B}{1}$ as $cc'$ is blocking, but $c \notin x_1p$. On the other hand, intersecting types do not permit $\beta(c')$ crossing segment $d'a_3$ in its interior, as $a_3 \in px_1$, so we get a contradiction.\qed\end{proof}
Next, we consider the cases whether the segments $c'p$ and $c''q$ intersect or not.\\
\noindent\textbf{Case 1.1, $c'p$ and $c''q$ intersect.} Let $e \colon= c'p \cap c''q$. In the region $x_1 c'' e c'$, the vertices $c'$ and $c''$ are internally blocked, while $x_1$ is not, so $e$ is externally blocked.\\
\indent Look at the triangle blocking arrangement~$\Delta_3$ induced by the triangle $c'' q x_3$. Its size is smaller than $n$, so the $(n-1)$-classification theorem~applies. As we have initial segments $d''b_3$ and $ep$, the type of $\Delta_3$ is not $\tbty{B}{0}$, nor $\tbty{B}{1}$. Suppose for a moment that is one of the two remaining basic types. Then, $\beta(e)$ crosses $x_3p$, at a point $p'$, say, and $\beta(p)$ crosses $ec''$ at a point $e'$, say, and we know that $ee', pp', e'p'$ are minimal initial segments. If $e'$ is in the interior of $ed''$, then $\beta(p)$ crosses $d''c \cup cd'$, which is impossible, as $d''c, d'c$ are minimal and $c$ is externally blocked in $cd''ed'$. But, $c''d''$ is minimal, so we must have $e' = d''$, but then $d''$ has three initial segments passing through it, $d''p', d''e, d''b_3$, which is impossible as well. Hence, $\Delta_3$ is not one of basic types.\\
\indent If the type is $\tbty{I}{1}$ or $\tbty{I}{2}$, then $q$ is internally blocked in the region $c''qx_3$. Since $e$ is externally blocked in the region $pqe$, $p$ is internally blocked in that region. Let the blocking segment through $q$ intersect $ep$ at a point $r$, say. By the definition of types $\tbty{I}{1}$ and $\tbty{I}{2}$, there is another initial segment $r_1r_2$ through $r$, with $r_1 \in px_3, r_2 \in eq$. Let $r_3 \in \partial T$ be the vertex such that $r_1r_3 = l(r_1r_2) \cap T$. Since $r_1 \in x_2x_3$ and $r_1$ and $r_3$ are on different sides of $l(c'p)$, we have $r_3 \in x_1x_2$. However, $p$ is internally blocked in the region $pqe$, and thus in the region $x_2c'p$, which thus have one of the types $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$, and also intersecting segments $eq$ and $r r_3$, which is a contradiction with the definitions of all these three types. Hence, $c''qx_3$ has type $\mathbf{T}$.\\
\indent Since the type of $\Delta_3$ is $\mathbf{T}$ and $c''d''$ is a minimal segment, it follows that $c''b_3$ and $b_3 d''$ are also minimal, and that these three segments bound a minimal $\mathcal{S}_{\Delta_3}$-region. Let $R$ be the other minimal $\mathcal{S}_{\Delta_3}$-region that contains the segment $b_3d''$, which is a hexagon by definition of $\mathbf{T}$, as $\beta(d'')$ crosses $qx_3$. Let $u,v, w$ be the vertices of $R$ such that $u, b_3, d'', v, w$ are consecutive on $\partial R$. As $l(a_3d'') \cap T$ is the blocking segment through $d''$, by the assumption \textbf{(A)} $l(ub_3), l(a_3 d'')$ and $l(vw)$ are concurrent. But $l(ub_3) \cap l(a_3 d'') = a_3$, and $a_2a_3$ is another initial segment through $a_3$, which is a contradiction, as otherwise three initial segments would be concurrent.\\
\noindent\textbf{Case 1.2, $c'p$ and $c''q$ are disjoint.} Look at the triangle blocking arrangement~$\Delta_4$ induced by the triangle $x_2 c' p$. Let $f' \colon= l(b_2d') \cap x_2x_3$ and $f'' \colon= l(a_3 d'') \cap x_2x_3$. Applying the $(n-1)$-classification theorem, $\Delta_4$ has one of the seven defined types. However, it cannot have a basic type, as $d' f'$ gives a blocking segment with endpoints on $x_2p$ and $c'p$, while $a_2d'$ is an initial segment and $a_2 \in c'x_2$.\\
\indent Suppose for a moment that $\Delta_4$ has type $\tbty{I}{1}$ or $\tbty{I}{2}$. By definition of these types we see that $d'f'$ does not cross any initial segment, except at its endpoints. Thus, the segments $f'd', d'c, cd'', d''c''$ are minimal. Thus, there is no initial segment that crosses $qx_3$ and $qc''$ in their interiors. But looking at the triangle blocking arrangement~$\Delta_5$ induced by $c''qx_3$, and applying $(n-1)$-classification theorem, we see that $\Delta_5$ has to have one of the seven types defined, and no type satisfies the requirements that $d''b_3$ is an initial segment, $d''f''$ is a blocking segment, and there are no initial segments with a vertex on $qx_3$ and a vertex on $qc''$, which is a contradiction. Therefore, $\Delta_4$ must be of type $\mathbf{T}$, and similarly, $\Delta_5$ must be of the same type.\\
\indent By definition of the type $\mathbf{T}$, there are vertices $p' \in px_2, p'' \in p c'$, such that $pp',pp''$ are minimal segments and $p'p''$ is an initial segment. Similarly, there are vertices $q' \in qx_3, q'' \in q c''$ such that $qq', qq''$ are minimal segments and $q'q''$ is an initial segment. Also, recalling that $c'd'$ and $c''d''$ are minimal segments, and using the definition of $\mathbf{T}$, we have that $a_2d'$ and $b_3d''$ are also minimal segments. Thus $l(p'p'')$ and $l(q'q'')$ cannot cross $a_2c \cup cb_3$, so $l(p'p'')$ crosses $x_1 x_3$, at $r_3$, say, and $l(q'q'')$ crosses $x_1x_2$, at $r_2$, say. In particular, $r_2q'$ and $r_3 p'$ intersect, at some point $r$. Also, $r_2q'$ must intersect $c'p$, at some point $r'$. However, we then have the following structure. In the triangle blocking arrangement~induced by $x_2 r_2 q'$, the vertex $q'$ is externally blocked (by minimality of the region $qq'q''$), $p, p' \in x_2q', r, r' \in r_2q'$ and $pr', p'r$ are intersecting initial segments. Applying $(n-1)$-classification theorem, we obtain a contradiction, as none of the seven types has this structure. This concludes the proof in the \textbf{Case 1}.\\
\noindent\framebox{\textbf{Case 2.}} Consider the triangle blocking arrangement~$\Delta_1$ induced by the triangle $x_1a_2a_3$. By the classification theorem~$\Delta_1$ has type $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$. Immediately, we see that the basic type is not possible here, as initial segment $b_2c$ and $\beta(a_2)$ intersect. Further, the point $e_2 \colon= b_2c \cap \beta(a_2)$, has the property that $b_2e_2$ and $e_2 c$ are both minimal segments. Let $d_2 e_3$ be the other initial segment through $e_2$, with $d_2 \in x_1a_2, e_3 \in a_2a_3$. By the definition of types $\tbty{I}{1}$ and $\tbty{I}{2}$, we have $d_2 \in a_2b_2, e_3 \in ca_3$. Let $d_3 \colon= l(d_2e_2) \cap x_1x_3$. Arguing similarly for the triangle blocking arrangement~$\Delta_2$ induced by the triangle $x_1b_2b_3$, we obtain that the segments $b_2d_2, d_2e_2, e_2b_2, e_2c, ce_3, e_2e_3, e_3d_3, d_3a_3, a_3e_3$ are all minimal. From this, we also see that $\Delta_1$ and $\Delta_2$ must both have type $\tbty{I}{1}$. Writing $x_2' \colon= \beta(b_3) \cap x_1x_2, x'_3 \colon= \beta(a_2) \cap x_1x_3$, we also obtain that $x'_2x'_3$ is a minimal initial segment, that $x'_2b_2$ and $x'_3a_3$ are also minimal and $b_2a_3$ is a blocking segment.\\
\indent Observe $d_2c$ and $d_3c$ cannot both be blocking segments, as otherwise $d_2d_3$ would simultaneously be blocking and initial. Without loss of generality, $d_3c$ is not a blocking segment. By the definition of type $\tbty{I}{1}$, there are vertices $p,q \in d_3b_3$ and $p',q' \in cb_3$, such that $d_3p, pq, cp', p'q'$ are minimal and $p'q, pq'$ are initial segments and $p'q, pq', b_3e_3$ are concurrent. We now consider cases on the position of intersection $r$ of $l(p'q)$ and $\partial T$, other than $q$.\\[6pt]
\begin{figure}
\includegraphics[width = \textwidth]{prop8case2.pdf}
\caption{Case 2 of proof of Proposition~\ref{ForbStruct1}}
\label{prop8case2Fig}
\end{figure}
\noindent\textbf{Case 2.1, $r \in x_1x_2$.} The type of triangle blocking arrangement~$\Delta_3$ induced by $x_1rq$ must be $\mathbf{T}$, as $x_1$ and $q$ are externally blocked, and some of the initial segments intersect. As $d_2e_2, e_2c, cp'$ are minimal segments, we must have $r \in d_2x_2$. If $r \in a_2x_2$, then we obtain a contradiction from the type $\mathbf{T}$ for $\Delta_3$ and the fact that $a_2a_3$ and $d_2d_3$ intersect, while $a_2, d_2 \in x_1r, a_3d_3 \in x_1q$. Therefore $r \in d_2a_2$, and $rq$ intersects $a_2c$, at point $r'$, say. Note that the segments $q'p'$ and $p'r'$ are minimal. Returning to the triangle blocking arrangement~$\Delta_1$ induced by $x_1a_2a_3$, which is of type $\tbty{I}{1}$, as $r$ is internally blocked in $a_2rr'$, we have vertices $s \in d_2r, s' \in a_2r'$, such that $rs, r's'$ are minimal and $rr', ss'$ are initial segments that are concurrent with $a_2e_2$.\\
\indent Extend $pq'$ to the other intersection $t$ of $l(pq')$ and $\partial T$, other than $p$. If $t \in x_1x_2$, applying the $(n-1)$-classification theorem~for the triangle blocking arrangement~$\Delta_4$ induced by $x_1tp$, we have that $\Delta_4$ has type $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$, as $p$ is internally blocked. But, the fact that hexagonal region $x'_2 x'_3 a_3 e_3 e_2 b_2$ is minimal $\overline{\mathcal{S}_{\Delta_4}}$-region with edges on $x_1 t$ and $x_1p$ is in contradiction with the definitions of these two types. Therefore, $t \in x_2x_3$.\\
\indent Let $f \colon= pq' \cap p'q$. Applying $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_3pt$, and observing that $q'f, qf$ are minimal segments, $q'b_3$ is initial and $fb_3$ is blocking, we must have that $qq'$ is also a blocking segment. Consider the minimal $\overline{\mathcal{S}}$-region $R$, that contains vertices $s', r', p'$ and $q'$. Let $u,v$ be the two vertices of $R$, such that $p',q', u, v$ are consecutive. As $p'r'$ and $\beta(q') = q'q$ intersect at $q$, by the assumption \textbf{(A)}, the line $l(uv)$ must also pass through $q$. However, the only initial segment other than $rq$ through $q$ is $x_1x_3$, which is disjoint from $R$, and we reach a contradiction.\\
\noindent\textbf{Case 2.2, $r \in x_2x_3$.} Applying the classification theorem~to the triangle blocking arrangement~induced by $qrx_3$, we obtain a contradiction as $q$ is internally blocked in this region, while $y_1y_2 \colon= l(pq') \cap qrx_3$ and $p'b_3$ intersect, while $p', y_1 \in qr, y_2 \in rx_3, b_3 \in qx_3$, and none of the defined types has this substructure.\\
\noindent\framebox{\textbf{Case 3.}} Let $\Delta_1$ be the triangle blocking arrangement~induced by the triangle $x_1b_2b_3$. Applying the classification theorem~we see that $\Delta_1$ has type $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$. But, $\beta(b_3)$ crosses $ca_3$, discarding $\tbty{B}{1}$ as an option. Therefore, there are collinear vertices $d_3 \in x_1a_3, e_3 \in a_3c, f_3 \in cb_3$ such that $d_3a_3, e_3d_3, e_3f_3, e_3a_3, e_3c$ and $cf_3$ are minimal, $d_3f_3$ is initial and $e_3b_3$ is blocking. Let $p \not= d_3$ be the other intersection point of $l(d_3f_3)$ and $\partial T$. If $p \in x_1x_2$, then segments $b_2b_3$ and $d_3p$ satisfy conditions of the \textbf{Case 2}, which we have proved to be impossible. Hence, $p \in x_2x_3$.\\
\begin{figure}
\includegraphics[width = \textwidth]{prop8case3.pdf}
\caption{Case 3 of proof of Proposition~\ref{ForbStruct1}}
\label{prop8case3Fig}
\end{figure}
\indent Apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $pd_3x_3$. Since $e_3f_3$ and $e_3a_3$ are minimal, it follows that $f_3b_3$ and $a_3b_3$ are also minimal and $a_3f_3$ is blocking. But, look at the minimal $\overline{\mathcal{S}}$-region $R$, that has $f_3$ and $c$ as two vertices, but not $e_3$. Since $l(a_2c)$ and $\beta(f_3)$ meet at $a_3$, by the assumption \textbf{(A)}, we have $a_3 \in l(s)$, where $s$ is another segment of $R$. But, the only other initial segment through $a_3$ is $x_1x_3$, which is disjoint from $R$, thus we have a contradiction.\qed\end{proof}
\begin{proposition} \label{ForbStruct2} Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose there are blocking segments through $x_1$ and $x_2$. Suppose also that $a_1a_2, b_1b_2 \in \mathcal{S}$ are two initial segments, such that $a_1 \in x_1x_2, b_2 \in x_2x_3, a_2,b_1 \in x_1x_3$. Then $a_1a_2$ and $b_1b_2$ are disjoint.\end{proposition}
\begin{proof} Let $c$ be the intersection $a_1a_2 \cap b_1b_2$. Depending on the blocking segment through $c$ and $a_1$, we have the following four cases.
\begin{itemize}
\item[\textbf{Case 1}] In the region $x_2a_1cb_2$, both $a_1$ and $c$ are internally blocked.
\item[\textbf{Case 2}] In the region $x_2a_1cb_2$, $a_1$ is internally blocked, while $c$ is externally blocked.
\item[\textbf{Case 3}] In the region $x_2a_1cb_2$, $a_1$ is externally blocked, while $c$ is internally blocked.
\item[\textbf{Case 4}] In the region $x_2a_1cb_2$, both $a_1$ and $c$ are externally blocked.
\end{itemize}
We treat each case separately and we depict the steps of the proof in Figures~\ref{prop9case1Fig},~\ref{prop9case2Fig},~\ref{prop9case3Fig} and~\ref{prop9case4Fig}.\\
\noindent\framebox{\textbf{Case 1.}} We have $b_2$ externally blocked in the region $b_1b_2x_3$, and in the region $b_1ca_2$, the vertex $b_1$ is externally blocked, while $a_2$ is internally blocked. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $b_1b_2x_3$, it has a basic type or $\mathbf{T}$. In either of these cases, from the definition of the types, there are vertices $d \in b_1a_2, e \in b_1c$ such that $ce, ed, da_2$ are minimal initial segments, and $cd, a_2e$ are blocking. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by the triangle $x_1a_1a_2$, we see that it has type $\tbty{I}{1}$ or $\tbty{I}{2}$, so $l(de)$ must cross the segment $a_1c$, at some point $f$, and also $fc$ is minimal. Let $R$ be the minimal $\overline{\mathcal{S}}$-region with vertices $c,f$, but not $e$. Let $f', u, v$ be the vertices of $R$, such that $f', f,c,u,v$ are consecutive, appearing in this order of $\partial R$. As $l(ff') = l(de)$ and $\beta(c)$ intersect at $d$, by the assumption \textbf{(A)}, we must have $d \in l(uv)$. However, the other initial segment through $d$, apart from $de$, is $x_1x_3$, which is disjoint from $R$, and we have a contradiction in this case.\\[6pt]
\begin{figure}
\begin{center}
\includegraphics[width = 0.7\textwidth]{prop9case1.pdf}
\caption{Case 1 of proof of Proposition~\ref{ForbStruct2}}
\label{prop9case1Fig}
\end{center}\end{figure}
\noindent\framebox{\textbf{Case 2.}} We have $b_2$ internally blocked in the region $b_1b_2x_3$, and in the region $b_1ca_2$, the vertices $b_1$ and $a_2$ are internally blocked. Let $e \colon= \beta(a_2) \cap b_1c$. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_1a_1a_2$, we see that its type is either $\tbty{I}{1}$ or $\tbty{I}{2}$. In either case, there are vertices $d \in x_1b_1, f \in a_2c$ such that $b_1d, de, ef, fc$ are minimal initial segments. Note that $l(ef)$ intersects the segment $b_2x_3$, at some point $p$, say. However, applying $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $b_1b_2x_3$, we obtain a contradiction, as no type permits a configuration where $b_1, b_2$ are internally blocked, and $ca_2, ep$ cross.\\[6pt]
\begin{figure}\begin{center}
\includegraphics[width = 0.7\textwidth]{prop9case2.pdf}
\caption{Case 2 of proof of Proposition~\ref{ForbStruct2}}
\label{prop9case2Fig}
\end{center}\end{figure}
\noindent\framebox{\textbf{Case 3.}} We have $b_2$ internally blocked in the region $b_1b_2x_3$, and in the region $b_1ca_2$, the vertex $b_1$ is internally blocked, while $a_2$ is externally blocked. Apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_1a_1a_2$. Its type is one of $\tbty{B}{1}, \tbty{I}{1}$ and $\tbty{I}{2}$, but in any case, $b_1c$ is minimal, and there are vertices $d \in b_1a_2, e \in ca_2$ such that $ce,b_1d$ and $de$ are minimal initial segments, and $cd$ and $b_1e$ are blocking. Apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $b_1b_2x_3$, which must then have type $\tbty{I}{1}$ or $\tbty{I}{2}$, and in particular $l(de)$ crosses $cb_2$, at a point $f$, say.\\
\indent Let $p \not= d$ be the other point of intersection of $l(de)$ with $\partial T$. If $p \in x_2x_3$, then the segments $pd, b_1b_2$ and the vertex $x_3$ form a configuration that is impossible by Proposition~\ref{ForbStruct1}, hence $p \in x_1x_2$. As $pd$ crosses $a_1a_2$ at $e$, we actually have $p \in a_1x_2$. However, applying $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $px_1d$, we obtain a contradiction, as no type allows a subconfiguration, where $b_1f$ and $a_1e$ cross, $x_1$ and $d$ are internally blocked, and $e,f \in pd, b_1 \in x_1d, a_1 \in x_1p$.\\[6pt]
\begin{figure}\begin{center}
\includegraphics[width = 0.7\textwidth]{prop9case3.pdf}
\caption{Case 3 of proof of Proposition~\ref{ForbStruct2}}
\label{prop9case3Fig}
\end{center}\end{figure}
\noindent\framebox{\textbf{Case 4.}} We have $b_2$ externally blocked in the region $b_1b_2x_3$, and in the region $b_1ca_2$, the vertices $b_1$ and $a_2$ are externally blocked. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_1$ induced by $x_1a_1a_2$, we see that its type is $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$. In particular, the triangle blocking arrangement~induced by $b_1a_2c$ is of type $\tbty{B}{0}$ or $\tbty{B}{1}$ and $b_1c$ is minimal. On the other hand, applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_2$ induced by $b_1b_2x_3$, which must have a basic type or $\mathbf{T}$, we also see that as $b_1c$ is minimal, so are $b_1a_2$ and $a_2c$.
\begin{figure}\begin{center}
\includegraphics[width = 0.7\textwidth]{prop9case4.pdf}
\caption{Case 4 of proof of Proposition~\ref{ForbStruct2}}
\label{prop9case4Fig}
\end{center}\end{figure}
Assume now that there is no initial segment with one vertex on $cb_2$ and the other vertex on $a_2x_3$. Thus $\Delta_2$ has type $\mathbf{T}$, and there are vertices $d \in a_2x_3, f,f' \in x_3b_2, e \in b_2c$ such that $a_2d, df', f'f, fe, ec$ are minimal initial segments, that bound a minimal $\overline{\mathcal{S}}$-region $R$, and $a_2f, de, cf'$ are blocking. Observe that by the assumption \textbf{(A)} for $R$, as $l(a_2c)$ and $l(ff')$ are disjoint in $T$, we must have $l(de)$ disjoint from these two lines in $T$, as well. Hence $l(de)$ crosses segment $a_1x_2$. Thus, $\beta(x_2)$ does not pass through $e$ and $c$, and $fe, ec, cb_1$ are minimal, so $\beta(x_2)$ has to cross $a_1c$ and either crosses $x_1b_1$ or contains $b_1$.\\
\indent Next, we show that $l(cf')$ crosses $x_1x_3$. If $\Delta_1$ has type $\tbty{B}{1}$, then this is true. If $l(cf')$ does not cross $x_1x_3$, then, $\Delta_1$ has type $\tbty{I}{1}$ or $\tbty{I}{2}$, but in that case, the only blocking segments that could cross $l(cf')$ in $\Delta_1$ are $\beta(x_1), \beta(a_1)$ and $\beta(b_1)$. Thus $\beta(x_2)$ passes through $b_1$. But $\beta(b_1)$ crosses the interior of $a_1c$, but types $\tbty{I}{1}$ and $\tbty{I}{2}$ imply that $\beta(b_1)$ crosses $a_1x_1$, which is a contradiction.\\
\indent Hence, $l(cf')$ crosses $x_1x_3$. Then, by the assumption \textbf{(A)} for region $R$, $l(ef), l(cf'), l(a_2d)$ are concurrent at a point $q \not= x_2$. However, Proposition~\ref{ForbStruct1} gives a contradiction, when applied to the vertex $x_3$ and the segments $fq$ and $b_1b_2$.\\[3pt]
\indent Finally, we assume that there is an initial segment $de$ with $e \in cb_2$ and $d \in a_2x_3$. To obtain a contradiction, we consider the following three cases on the position of $p \not= d$, the other intersection of $l(de)$ with $\partial T$.\\[3pt]
\noindent\textbf{Case 4.1. Suppose that $p \in x_1a_1$.} Applying $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_1pd$, we see that $d$ must be internally blocked in this region. It follows that $e$ is internally blocked in the region $ca_2de$. However, $pd$ and $b_1b_2$ satisfy the conditions of the \textbf{Case 1}, which is impossible.\\[3pt]
\noindent\textbf{Case 4.2. Suppose that $p \in a_1x_2$.} Applying $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_1pd$ results in a contradiction as $a_1a_2$ and $b_1e$ cross, but $x_1$ is internally blocked.\\[3pt]
\noindent\textbf{Case 4.3. Suppose that $p \in x_2x_3$.} We actually have $p \in x_2b_2$ and Proposition~\ref{ForbStruct1} says that configuration where $x_3$ is externally blocked, and $b_1b_2$ and $dp$ cross is impossible.\qed\end{proof}
\begin{proposition} \label{ForbStruct3} Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose there are blocking segments through $x_1$ and $x_3$. Suppose also that $a_1a_2, b_1b_2 \in \mathcal{S}$ are two initial segments, such that $a_1 \in x_1x_2, b_2 \in x_2x_3, a_2,b_1 \in x_1x_3$. Then $a_1a_2$ and $b_1b_2$ are disjoint.\end{proposition}
\begin{proof} Suppose contrary, let $c\colon= a_1a_2 \cap b_1b_2$. Depending on the blocking segments through $c$ and $b_1$, up to symmetry, we have the following three cases.
\begin{itemize}
\item[\textbf{Case 1}] In the region $cb_1a_2$, $c$ is externally blocked, while $b_1$ is internally blocked.
\item[\textbf{Case 2}] In the region $cb_1a_2$, both $b_1$ and $c$ are externally blocked.
\item[\textbf{Case 3}] In the region $cb_1a_2$, both $b_1$ and $c$ are internally blocked.
\end{itemize}
As before, we treat each case separately and we depict the steps of the proof in Figures~\ref{prop10case1Fig},~\ref{prop10case2Fig} and~\ref{prop10case3Fig}.\\
\noindent\framebox{\textbf{Case 1.}} Note that the vertex $a_2$ is internally blocked in the region $cb_1a_2$, and that $a_1$ and $b_2$ are internally blocked in the region $x_2a_1cb_2$. Let $d\colon= \beta(a_2) \cap cb_1, e\colon= \beta(b_1) \cap ca_2$. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_1$ induced by $a_1a_2x_1$, we see that $\Delta_1$ has type among $\tbty{I}{1}$ and $\tbty{I}{2}$. In either case, $dc$ and $db_1$ are minimal segments, and no initial segment may cross the interior of the segment $b_1e$. Similarly, by looking at the triangle blocking arrangement~$\Delta_2$ induced by $b_1b_2x_3$, we have that $ce, ea_2$ are minimal, and no initial segment crosses the interior of $a_2d$. It follows that the initial segment through $d$, which is different from $cb_1$, must be $de$, so $de$ is an initial segment.\\
\indent However, the type of $\Delta_1$ implies that $l(de)$ crosses $x_1b_1$, while the type of $\Delta_2$ implies that $l(de)$ crosses $a_2x_3$, so $l(de)$ crosses segment $x_1x_3$ twice, which is impossible.\\[6pt]
\begin{figure}\begin{center}
\includegraphics[width = 0.7\textwidth]{prop10case1.pdf}
\caption{Case 1 of Proposition~\ref{ForbStruct3}}
\label{prop10case1Fig}
\end{center}
\end{figure}
\noindent\framebox{\textbf{Case 2.}}Note that the vertex $a_2$ is externally blocked in the region $cb_1a_2$, and that $a_1$ and $b_2$ are externally blocked in the region $x_2a_1cb_2$. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_1$ induced by $a_1a_2x_1$, the type of $\Delta_1$ is among $\tbty{B}{1}, \tbty{I}{1}$ or $\tbty{I}{2}$. In either case, the segment $b_1c$ is minimal. Similarly, looking at the triangle blocking arrangement~$\Delta_2$ induced by $b_1b_2x_3$, the segment $ca_2$ is also minimal.\\
\indent Suppose for a moment that $\beta(c)$ crosses $a_2x_3$, at a point $e$. Let $R$ be the minimal $\overline{\mathcal{S}}$-region with vertices $b_1$ and $c$, but not $a_2$. Let $u, v, w$ be the vertices of $R$, such that $u, b_1, c, v, w$ are consecutive and appear in that order on $\partial R$, thus $wv, vc$ are minimal. As $l(ub_1)$ and $\beta(c)$ intersect at $e$, by the assumption \textbf{(A)}, $d \in l(vw)$. Also $l(vw)$ crosses $cb_2$, let $d$ be their intersection point. From the $(n-1)$-classification theorem~applied to $\Delta_2$, we see that $cd, de, a_2e$ are minimal and $a_2d$ is blocking. As $cd, cv$ are minimal, so is $vd$. But, if we look at the minimal $\overline{\mathcal{S}}$-region $R'$, with vertices $v,d$, but not $c$, since $l(vc) = l(s_1)$ and $\beta(d)$ intersect at $a_2$ (where $s_1$ is the segment of $\partial R'$ through $v$, different from $vd$), it follows by the assumption \textbf{(A)}, that $a_2 \in l(s_2)$ for another segment $s_2$ of $\partial R'$. However, $a_2 \in x_1x_3$, and $x_1x_3$ is disjoint from $R'$, which is a contradiction. Therefore, $\beta(c)$ is disjoint from $a_2x_3$, and by symmetry $\beta(c)$ is also disjoint from $x_1b_1$.\\
\indent Thus $\beta(c)$ crosses $a_1x_1$, at some point $d$, and crosses $b_2x_3$, at some point $e$. Further, $\beta(x_1)$ crosses $a_1c$, at a point $f$, and $\beta(x_3)$ crosses $b_2c$, at a point $f'$. From the types of $\Delta_1$ and $\Delta_2$, we also have that $fc, f'c$ are minimal, and there are points $p \in a_1d$, and $p' \in b_2e$, such that $pf$ and $p'f'$ are initial segments. Further, also from the types of $\Delta_1$ and $\Delta_2$, we have that if an initial segment $s$ crosses $fx_1$, then $s$ must have one vertex on $a_1x_1$ and the other on $x_1b_1$, and we have that if an initial segment $s$ crosses $x_3f'$, then $s$ has one vertex on $b_2x_3$, and the other on $a_2x_3$. It follows that $l(pf) \cap \operatorname{int} f'x_3 = l(p'f') \cap \operatorname{int} fx_1 = \emptyset$.\\
\indent Observe that if it happens that $l(pf) = l(p'f')$, then Proposition~\ref{ForbStruct2}~applies to segments $pp', a_1a_2$ inside $x_1x_2x_3$ to give a contradiction. Therefore, $l(pf)$ crosses $x_2x_3$ at some point $q \not= p'$ (and $x_1x_2$ at $p$) and $l(p'f')$ crosses $x_1x_2$ at some point $q' \not= p$ (and $x_2x_3$ at $p'$). Finally, as $cf$ and $cf'$ are minimal, $pq$ and $p'q'$ must cross, but then the segments $pq, p'q'$ and the vertex $x_2$ are in a contradiction with Proposition~\ref{ForbStruct1}.\\[6pt]
\begin{figure}\begin{center}
\includegraphics[width = 0.7\textwidth]{prop10case2.pdf}
\caption{Case 2 of Proposition~\ref{ForbStruct3}}
\label{prop10case2Fig}
\end{center}
\end{figure}
\noindent\framebox{\textbf{Case 3.}} Note that the vertex $a_2$ is externally blocked in the region $cb_1a_2$, and that $a_1$ is externally and $b_2$ is internally blocked in the region $x_2a_1cb_2$. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_1$ induced by $a_1a_2x_1$, we have that the type of $\Delta_1$ is among $\tbty{B}{1}, \tbty{I}{1}$ and $\tbty{I}{2}$. In either case, $b_1c$ is a minimal segment. Next, applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_2$ induced by $b_1b_2x_3$, the type of $\Delta_2$ is either $\tbty{I}{1}$ or $\tbty{I}{2}$. Since $b_1c$ is minimal, setting $d \colon= \beta(c) \cap b_1a_2, e\colon= \beta(b_1) \cap ca_2$, we must have that $de$ is an initial segment, $l(de)$ crosses $cb_2$, at some point $f$, and $cf, ce, fe, de, ea_2,da_2$ are minimal. Let $R$ be a minimal $\overline{\mathcal{S}}$-region such that $c,f \in \partial R$, but $e \notin \partial R$. Let $u, v, w$ be the vertices of $R$, such that $u, f, c, v, w$ are consecutive and appear in this order on $\partial R$. Since $\beta(c)$ and $l(uf) = l(fe)$ meet at $d$, by the assumption \textbf{(A)}, it follows that $d \in l(vw)$. However, the other initial segment through $d$, apart from $de$, is $x_1x_3$, which is disjoint from $R$, which is a contradiction.
\begin{figure}\begin{center}
\includegraphics[width = 0.7\textwidth]{prop10case3.pdf}
\caption{Case 3 of Proposition~\ref{ForbStruct3}}
\label{prop10case3Fig}
\end{center}
\end{figure}\qed\end{proof}
\begin{lemma}\label{triangularBlocksLemma} Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that there are vertices $a_1 \in x_1x_2, a_2,b_1 \in x_1x_3, b_2 \in x_2x_3$ such that $a_1a_2$ and $b_1b_2$ are initial segments, intersecting at a point $c$. Then, in the region $a_1cb_2x_2$, the vertices $c, x_2$ are externally blocked and $a_1, b_2$ are internally blocked. Also, $x_1, x_3$ are externally blocked, and in the region $cb_1a_2$, the vertices $a_2,b_1$ are externally blocked.\end{lemma}
\begin{proof} By Propositions~\ref{ForbStruct2} and~\ref{ForbStruct3}, we have $x_1, x_2, x_3$ all externally blocked. Looking at the blocking segments $\beta(c)$ and $\beta(a_2)$, we have the following three cases, up to symmetry.
\begin{itemize}
\item[\textbf{Case 1}] In the region $a_1cb_2x_2$, the vertices $a_1, c, b_2, x_2$ are externally blocked. In the region $cb_1a_2$, the vertices $a_2,b_1$ are internally blocked.
\item[\textbf{Case 2}] In the region $a_1cb_2x_2$, the vertices $a_1, c$ are internally blocked and $b_2, x_2$ are externally blocked. In the region $cb_1a_2$, the vertex $b_1$ is internally blocked and $a_2$ is externally blocked.
\item[\textbf{Case 3}] In the region $a_1cb_2x_2$, the vertices $a_1, b_2$ are internally blocked and $c, x_2$ are externally blocked. In the region $cb_1a_2$, the vertices $a_2,b_1$ are externally blocked.
\end{itemize}
Observe that the \textbf{Case 3} is exactly the conclusion of the lemma, so we just need to discard the first two cases. As before, we depict the steps in the proof in Figure~\ref{triangleBlocksFig}.\\[6pt]
\begin{figure}
\begin{center}
\includegraphics[width = 0.7\textwidth]{lemma11.pdf}
\label{triangleBlocksFig}
\caption{Cases in the proof of Lemma~\ref{triangularBlocksLemma}}
\end{center}
\end{figure}
\noindent\framebox{\textbf{Case 1.}} Let $d \colon= \beta(a_2) \cap b_1c, e \colon= \beta(b_1) \cap a_2c$. By the $(n-1)$-classification theorem~applied to the triangle blocking arrangement~$\Delta_1$ induced by $a_1a_2x_1$, the segment $cd$ is minimal, and there is another initial segment through $d$ that crosses $ce$, possibly through $e$. Similarly, by applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_2$ induced by $b_1b_2x_3$, $ce$ is also minimal segment, so it follows that $de$ is itself a minimal initial segment. But, as $\Delta_1$ has type $\tbty{I}{1}$ or $\tbty{I}{2}$, it follows that $l(de)$ crosses $x_1b_1$. Similarly, from the type of $\Delta_2$, $l(de)$ crosses $a_2x_3$, hence $l(de)$ meets $x_1x_3$ twice, which is impossible.\\
\noindent\framebox{\textbf{Case 2.}} Let $e \colon= \beta(b_1) \cap ca_2$. By the $(n-1)$-classification theorem~applied to the triangle blocking arrangement~$\Delta_1$ induced by $b_1b_2x_3$, $\Delta_1$ has type $\tbty{I}{1}$ or $\tbty{I}{2}$, and in particular, there are vertices $d \in b_1a_2,f \in cb_2$, such that $d,e,f$ lie on the same initial segment, and $ce,ef,fc,ed,da_2,a_2e$ are minimal. Let $p \not= d$, be the other intersection of the line $l(de)$ with $\partial T$. If $p \in x_2x_3$, then we have a substructure where $b_1b_2, dp$ intersect, and $x_3$ is externally blocked, which is forbidden by Proposition~\ref{ForbStruct1}, yielding a contradiction. Therefore, $p \in x_1x_2$. Once again, Proposition~\ref{ForbStruct1} gives contradiction, as $a_1a_2, dp$ intersect and $x_1$ is externally blocked.\qed\end{proof}
\begin{proposition}\label{triangleCoversProp}Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that there are vertices $a_3, b_3 \in x_1x_2$, appearing in order $x_1, b_3, a_3, x_2$, $c_2, a_2 \in x_1x_3$, appearing in order $x_1, c_2, a_2, x_3$, and $b_1, c_1 \in x_3x_2$, appearing in order $x_3, b_1, c_1, x_2$, such that $a_2a_3, b_1b_3, c_1c_2 \in \mathcal{S}$. Let $d_1 = b_1b_3\cap c_1c_2, d_2 = a_2a_3 \cap c_1c_2, d_3 = a_2a_3\cap b_1b_3$. Suppose additionally that $d_1$ is in the interior of $x_1a_2a_3$. Then $\Delta$ has type $\mathbf{T}$.\end{proposition}
\begin{proof} Apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_a$ induced by $x_1a_3a_2$. As $b_1b_3 \cap \Delta_a$ and $c_1c_2 \cap \Delta_a$ intersect, it must have type $\mathbf{T}$. By definition, there are an integer $k \geq 2$ and vertices $u^{(a)}_1, u^{(a)}_2, \dots,$ $u^{(a)}_{2k}$ appearing on $x_1a_3$, in that order from $x_1$ to $a_3$, there are vertices $v^{(a)}_1, v^{(a)}_2, \dots,$ $v^{(a)}_{2k}$ appearing on $a_3a_2$, in that order from $a_3$ to $a_2$, and there are vertices $w^{(a)}_1, w^{(a)}_2, \dots,$ $w^{(a)}_{2k}$ appearing on $a_2x_1$, in that order from $a_2$ to $x_1$, such that segments $u^{(a)}_iu^{(a)}_{i+1}, v^{(a)}_iv^{(a)}_{i+1},$ $w^{(a)}_iw^{(a)}_{i+1}$ are minimal, and the triangle blocking arrangements~induced by $x_1u^{(a)}_1w^{(a)}_{2k}, a_3v^{(a)}_1u^{(a)}_{2k}$ and $a_2w^{(a)}_1v^{(a)}_{2k}$, are of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with minimal initial segments $u^{(a)}_1w^{(a)}_{2k}, v^{(a)}_1u^{(a)}_{2k}$ and $w^{(a)}_1v^{(a)}_{2k}$. The initial segments, outside the three small regions $x_1u^{(a)}_1w^{(a)}_{2k},$ $a_3v^{(a)}_1u^{(a)}_{2k},$ $a_2w^{(a)}_1v^{(a)}_{2k}$ are given by $u^{(a)}_{2i-1} w^{(a)}_{2k+2-2i}, v^{(a)}_{2i-1} u^{(a)}_{2k+2-2i},$ $w^{(a)}_{2i-1} v^{(a)}_{2k+2-2i}$, for $i = 1, 2, \dots, k$, and the blocking segments outside the three small regions are given by $w^{(a)}_{2i-1} u^{(a)}_{2k+2-2i},$ $u^{(a)}_{2i-1} v^{(a)}_{2k+2-2i},$ $v^{(a)}_{2i-1} w^{(a)}_{2k+2-2i}$, for $i = 1, 2, \dots, k$.\\
\indent Similarly, apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_b$ induced by $x_2b_1b_3$. It must have type $\mathbf{T}$. By definition, there are an integer $l \geq 2$ and vertices $u^{(b)}_1, u^{(b)}_2, \dots, u^{(b)}_{2l}$ appearing on $b_3x_2$, in that order from $b_3$ to $x_2$, there are vertices $v^{(b)}_1, v^{(b)}_2, \dots, v^{(b)}_{2l}$ appearing on $x_2b_1$, in that order from $x_2$ to $b_1$, and there are vertices $w^{(b)}_1, w^{(b)}_2, \dots, w^{(b)}_{2l}$ appearing on $b_1b_3$, in that order from $b_1$ to $b_3$, such that segments $u^{(b)}_iu^{(b)}_{i+1},$ $v^{(b)}_iv^{(b)}_{i+1},$ $w^{(b)}_iw^{(b)}_{i+1}$ are minimal, and the triangle blocking arrangements~induced by $b_3u^{(b)}_1w^{(b)}_{2l},$ $x_2v^{(b)}_1u^{(b)}_{2l}$ and $b_1w^{(b)}_1v^{(b)}_{2l}$, are of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with minimal initial segments $u^{(b)}_1w^{(b)}_{2l},$ $v^{(b)}_1u^{(b)}_{2l},$ $w^{(b)}_1v^{(b)}_{2l}$. The initial segments, outside the three small regions $b_3u^{(b)}_1w^{(b)}_{2l},$ $x_2v^{(b)}_1u^{(b)}_{2l},$ $b_1w^{(b)}_1v^{(b)}_{2l}$ are given by $u^{(b)}_{2i-1} w^{(b)}_{2l+2-2i},$ $v^{(b)}_{2i-1} u^{(b)}_{2l+2-2i},$ $w^{(b)}_{2i-1} v^{(b)}_{2l+2-2i}$, for $i = 1, 2, \dots, l$, and the blocking segments outside the three small regions are given by $w^{(b)}_{2i-1} u^{(b)}_{2l+2-2i},$ $u^{(b)}_{2i-1} v^{(b)}_{2l+2-2i},$ $v^{(b)}_{2i-1} w^{(b)}_{2l+2-2i}$, for $i = 1, 2, \dots, l$.\\
\indent Apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_c$ induced by $x_3c_1c_2$. It must have type $\mathbf{T}$, so by definition, there are and integer $m \geq 2$ and vertices $u^{(c)}_1, u^{(c)}_2, \dots, u^{(c)}_{2m}$ appearing on $c_2c_1$, in that order from $c_2$ to $c_1$, there are vertices $v^{(c)}_1, v^{(c)}_2, \dots, v^{(c)}_{2m}$ appearing on $c_1x_3$, in that order from $c_1$ to $x_3$, and there are vertices $w^{(c)}_1, w^{(c)}_2, \dots,$ $w^{(c)}_{2m}$ appearing on $x_3c_2$, in that order from $x_3$ to $c_2$, such that segments $u^{(c)}_iu^{(c)}_{i+1},$ $v^{(c)}_iv^{(c)}_{i+1},$ $w^{(c)}_iw^{(c)}_{i+1}$ are minimal, and the triangle blocking arrangements~induced by $c_2u^{(c)}_1w^{(c)}_{2m},$ $c_1v^{(c)}_1u^{(c)}_{2m}$ and $x_3w^{(c)}_1v^{(c)}_{2m}$, are of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with minimal initial segments $u^{(c)}_1w^{(c)}_{2m},$ $v^{(c)}_1u^{(c)}_{2m},$ $w^{(c)}_1v^{(c)}_{2m}$. The initial segments, outside the three small regions $c_2u^{(c)}_1w^{(c)}_{2m},$ $c_1v^{(c)}_1u^{(c)}_{2m},$ $x_3w^{(c)}_1v^{(c)}_{2m}$ are given by $u^{(c)}_{2i-1} w^{(c)}_{2m+2-2i},$ $v^{(c)}_{2i-1} u^{(c)}_{2m+2-2i},$ $w^{(c)}_{2i-1} v^{(c)}_{2m+2-2i}$, for $i = 1, 2, \dots, m$, and the blocking segments outside the three small regions are given by $w^{(c)}_{2i-1} u^{(c)}_{2l+2-2i},$ $u^{(c)}_{2i-1} v^{(c)}_{2m+2-2i},$ $v^{(c)}_{2i-1} w^{(c)}_{2m+2-2i}$, for $i = 1, 2, \dots, m$.\\
If we look at the triangle blocking arrangement~induced by $a_3u^{(a)}_{2k}v^{(a)}_1$, which we know has type $\tbty{B}{0}$ or $\tbty{B}{1}$ with $u^{(a)}_{2k}v^{(a)}_1$ minimal, since $\Delta_b$ has type $\mathbf{T}$, it follows that $a_3u^{(a)}_{2k}v^{(a)}_1$ actually is a minimal $\overline{\mathcal{S}}$-region, so $a_3u^{(a)}_{2k}$ is also minimal. Similarly, it follows that $c_1 v^{(c)}_1, b_1v^{(b)}_{2l}, a_2 w^{(a)}_1, c_2 w^{(c)}_{2m}$ and $b_3 u^{(b)}_1$ are minimal segments. Let $i_a, i_b, i_c$ be such that $a_3 = u^{(b)}_{i_a}, b_1 = v^{(c)}_{i_b}, c_2 = w^{(a)}_{i_c}$. Define
\begin{itemize}
\item[] $u_i = u^{(a)}_i$, for $1\leq i \leq 2k$, and $u_{2k + i} = u^{(b)}_{i_a + i - 1}$, for $1\leq i \leq 2l + 1 - i_a$,
\item[] $v_i = v^{(b)}_i$, for $1\leq i \leq 2l$, and $v_{2l + i} = v^{(c)}_{i_b + i - 1}$, for $1\leq i \leq 2m + 1 - i_b$,
\item[] $w_i = w^{(c)}_i$, for $1\leq i \leq 2m$, and $w_{2m + i} = w^{(a)}_{i_c + i - 1}$, for $1\leq i \leq 2k + 1 - i_c$.
\end{itemize}
Let $p_u = 2k + 2l + 1 - i_a, p_v = 2l + 2m + 1 - i_b, p_w = 2m + 2k + 1 - i_c$ (which are the lengths of these sequences). We now show that $p_u = p_v = p_w$. Let $p_{u,1}$ be the number of vertices $u_i$, whose initial segment $\not= x_1x_2$ crosses $x_2x_3$, let $p_{v,1}$ be the number of vertices $v_i$, whose initial segment $\not= x_2x_3$ crosses $x_3x_1$, let $p_{w,1}$ be the number of vertices $w_i$, whose initial segment $\not= x_3x_1$ crosses $x_1x_2$, and let $p_{u,2} = p_u - p_{u,1}, p_{v,2} = p_v - p_{v,1}, p_{w,2} = p_w - p_{w,1}$.\\
\indent Firstly, we show that $p_{u,1} = p_{u,2}$ (so by symmetry $p_{v,1}= p_{v,2}$ and $p_{w,1}=p_{w,2}$). Note immediately that the initial segment $u_1w_{p_w}$ crosses $x_1x_2$ and $x_1x_3$, and that the initial segment $u_{p_u} v_1$ crosses $x_1x_2$ and $x_2x_3$. Also, for any $i < p_u$, observe that, by the definition of type $\mathbf{T}$, the initial segments $\not=x_1x_2$ through $u_i$ and $u_{i+1}$ cannot both intersect $x_1x_3$, nor can both intersect $x_2x_3$. To spell out details, looking at $\Delta_a$, if $i < 2k$, then one of $u_i, u_{i+1}$ has the other initial segment with a vertex on $x_1a_2$, and the second has the other initial segment with a vertex on $a_3a_2$. Write temporarily $q$ for this second vertex among $u_i, u_{i+1}$ and $r\in a_3a_2$ for the vertex such that $qr$ is initial. But $l(qr)$ crosses $a_3a_2$, and we must have $l(qr)$ cross $x_2x_3$, as otherwise we obtain a contradiction by Proposition~\ref{ForbStruct1}. We argue similarly for $i \geq 2k$, by considering $\Delta_b$. The claim follows.\\
\indent Secondly, we show $p_{u,1} = p_{v,2}$ (and by symmetry $p_{v,1} = p_{w,2}$ and $p_{w,1} = p_{u,2}$). But, if a segment crosses $x_1x_2$ at some $u_i$, and crosses $x_2x_3$ at a point $q$, then by minimality of $u_{p_u} v_1$ and $v_{p_v} w_1$, $q = v_j$ for some $j$. However, this is injective map $i \mapsto j$, so $p_{u,1} \geq p_{v,2}$, and by symmetry $p_{u,1} = p_{v,2}$. From these observations, it follows that $p_u = p_v = p_w$, and we may write $p$ for this common value.\\
Next, we show that initial and blocking segments at $u_i, v_i, w_i$ satisfy the conditions of the type $\mathbf{T}$. As we have seen already, initial segments through $u_i$ are $u_1 w_{i_1}, u_3 w_{i_3}, \dots, u_{p-1} w_{i_{p-1}}$ and $u_2 v_{i_2}, u_4v_{i_4}, \dots, u_p v_{i_p}$, where $i_1, i_3, \dots,$ $i_{p-1} \in [p]$ are distinct and even, and $i_2, i_4, \dots, i_p \in[p]$ are distinct and odd. However, if $i_{j} > i_{j'}$ holds for some $j > j'$ of the same parity, then we obtain a contradiction by Proposition~\ref{ForbStruct1}. Hence, $i_{j} = p + 1 - j$, for all $j$, as desired, and a similar argument shows that all initial segments have desired structure. For blocking segments, observe that all initial and blocking segments between $x_1x_2$ and $x_2x_3$ through some $u_i$ are disjoint, so the blocking segments have the desired structure. The intersections structure follows from the structure of $\Delta_a, \Delta_b, \Delta_c$.\\
\indent Finally, we know from before that the triangle blocking arrangements~induced by $x_1u_1w_{p}, x_2v_1u_{p}, x_3 w_1 v_{p}$ are of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with minimal segments $u_1w_{p}, v_1u_{p}, w_1 v_{p}$. This completes the proof.\qed\end{proof}
\begin{corollary}\label{triangleS3}Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that there are vertices $a_3, b_3 \in x_1x_2$, appearing in order $x_1, b_3, a_3, x_2$, $c_2, a_2 \in x_1x_3$, appearing in order $x_1, c_2, a_2, x_3$, and $b_1, c_1 \in x_3x_2$, appearing in order $x_3, b_1, c_1, x_2$, such that $a_2a_3, b_1b_3, c_1c_2 \in \mathcal{S}$. Then $\Delta$ has type $\mathbf{T}$.\end{corollary}
\begin{proof}Let $d_1 = b_1b_3\cap c_1c_2, d_2 = a_2a_3 \cap c_1c_2, d_3 = a_2a_3\cap b_1b_3$. By previous proposition, we may suppose that $d_1$ is not in the interior of $x_1a_2a_3$, and moreover that there are no triples of segments where each pair intersects and form a small triangle that satisfies the conditions of Proposition~\ref{triangleCoversProp}.\\
\indent By Lemma~\ref{triangularBlocksLemma}, we have that $x_1, x_2, x_3$ are externally blocked, and in regions $a_3b_3d_3, a_2c_2d_2, b_1c_1d_1$, the vertices $a_3, b_3, d_3, a_2, c_2, d_2, b_1, c_1, d_1$ are externally blocked as well. Suppose for a moment that $a_3d_3$ is not a minimal segment. Let $q \in a_3d_3$ be another vertex. Let $pr$ be another initial segment through $q$, with $p,r \in \partial T$, with $p \in a_3x_1 \cup x_1a_2$ and $r \in a_3x_2 \cup x_2x_3 \cup x_3a_2$. If $pr$ crosses $b_3b_1$, then either Proposition~\ref{ForbStruct1} gives a contradiction, for $pr$ and $b_3b_1$ or $pr$ and $a_3a_2$, or the segments $pr, b_3b_1, a_3a_2$ form a triple we forbade at the beginning of the proof. Hence, $pr \cap b_3b_1 = \emptyset$, so $p \in a_3b_3, r \in x_2b_1$. However, if we apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_2b_3b_1$, it must have type $\mathbf{T}$, and it follows that there is an initial segment $s$ which crosses both $d_3a_3$ and $d_3b_3$. By Proposition~\ref{ForbStruct1}, it follows that $s$ crosses $x_1x_3$ and $x_2x_3$, thus $s, b_3b_1, a_3a_2$ forms a structure that we forbade at the beginning of the proof. This is a contradiction, and it follows similarly that $a_3d_3, b_3d_3, d_2a_2, d_2c_2, d_1b_1, d_1c_1$ are minimal, and further, $a_3b_3, a_2c_2, b_1c_1$ are minimal as well.\\
\indent Therefore, we have actually shown that any configuration of segments like $a_2a_3, b_1b_3, c_1c_2$ implies the minimality of the segments $a_3d_3, b_3d_3$, etc. Using this observation and Proposition~\ref{ForbStruct1}, it follows also that $d_1d_2, d_2d_3$ and $d_3d_1$ are minimal. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangements~induced by $x_1a_3a_2, x_2b_3b_1$ and $x_3c_2c_1$, all three have the type $\tbty{B}{3}$, and there are vertices $a'_2 \in x_1c_2, a'_3 \in b_3x_1, b'_3 \in a_3x_2, b'_1 \in x_2c_1, c'_1 \in b_1x_3, c'_2 \in a_2x_3$, such that $c_2a'_2, a'_2a'_3, a'_3b_3, a_3b'_3, b'_3b'_1, b'_1c_1, b_1c'_1, c'_1c'_2, c'_2a_2$ are minimal initial segments, and $\beta(d_1) = b'_3c'_2, \beta(d_2) = c'_1a'_3, \beta(d_3) = a'_2b'_1$. From the types of same triangle blocking arrangements, it follows that triangle blocking arrangements~induced by $x_1a'_2a'_3, x_2b'_3b'_1, x_3c'_1c'_2$ are of type $\tbty{B}{0}$ or $\tbty{B}{1}$, with $a'_2a'_3, b'_1b'_3, c'_1c'_2$ minimal. Thus, $\Delta$ has the type $\mathbf{T}$.\qed\end{proof}
\begin{proposition}\label{finalS3prop}Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that there are vertices $a_1 \in x_1x_2, a_2,b_1 \in x_1x_3, b_2 \in x_2x_3$ such that $a_1a_2$ and $b_1b_2$ are initial segments, intersecting at a point $c$. Then $\Delta$ has type $\mathbf{T}$.\end{proposition}
\begin{proof} By Lemma~\ref{triangularBlocksLemma}, the vertices $x_1,x_2,x_3$ are externally blocked. Let $c_1, c_2, \dots,$ $c_r$ be the vertices that lie on $a_1a_2$, in that order from $a_1$ to $a_2$. Thus, $r \geq 1$. Let the initial segment $\not=a_1a_2$ through $c_i$ be $p_iq_i$. Note that at least one of $p_i,q_i$ must be on $x_2x_3$, otherwise we obtain a contradiction using Proposition~\ref{ForbStruct1}, without loss of generality, $q_i \in x_2x_3$. Also, if $p_i \in x_1x_3, p_{i+1} \in x_1x_2$, then by Corollary~\ref{triangleS3}, $\Delta$ has type $\mathbf{T}$. Thus, assume that there is $i_0$ such that $p_i \in x_1x_2$ for $i \leq i_0$ and $p_i \in x_1x_3$ for $i > i_0$. Moreover, by Proposition~\ref{ForbStruct1}, on $a_1x_1$, the vertices $a_1, p_1, p_2, \dots, p_{i_0}, x_1$ appear in this order, and on $x_1a_2$, the vertices $x_1 p_{i_0+1} p_{i_0 + 2} \dots p_r a_2$ appear in this order. By Lemma~\ref{triangularBlocksLemma}, we also have $a_1, p_i$ and $c_i$ externally blocked in the region $a_1p_ic_i$, for $i \leq i_0$, and $a_2, p_i$ and $c_i$ externally blocked in the region $a_2p_ic_i$ for $i > i_0$. However, applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_a$ induced by $a_1a_2x_1$, the only type that can be satisfied by $\Delta_a$ is $\tbty{B}{3}$.\\
\indent From the definition of type $\tbty{B}{3}$, it follows that $p_1c_1, c_1c_2, c_2p_2$ are minimal, that there are vertices $r_1 \in x_1p_1, r_2 \in x_1p_2$ such that $r_1r_2$ is a minimal initial segment, and also that $r_1p_1, r_2p_2$ are minimal. Furthermore, $r_2c_1, r_1c_2, p_1p_2$ are all blocking segments.\\
\indent Observe that, by the assumption \textbf{(A)}, if $l(c_1r_2)$ crosses $x_1x_2$, then $l(p_2c_2)$ must pass through the same point, however, $l(p_2c_2)$ cuts $x_2x_3$, which is a contradiction. Hence, $\beta(c_1)$ crosses the segment $x_2q_1$, and similarly, $\beta(c_2)$ cuts $q_2x_3$. Therefore, $(n-1)$-classification theorem~applied to the triangle blocking arrangements~$\Delta_1, \Delta_2$ induced respectively by regions $p_1x_2q_1$ and $p_2x_3q_2$, both have type $\mathbf{T}$ or $\tbty{B}{3}$. In particular, looking at $\Delta_1$, there are vertices $u \in x_2q_1,v \in c_1q_1$, such that $uv$ is an initial segment. As $c_1c_2$ and $c_2a_2$ are minimal, $l(uv)$ is disjoint from $c_1a_2$, so $l(uv)$ must cross $a_2x_3$, at some point $w$. But, $uw$ and $p_2q_2$ must cross as well, and this is a final contradiction granted by Proposition~\ref{ForbStruct1}.\qed\end{proof}
\begin{lemma}\label{S12basicintersectionLemma}Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that there are vertices $a_1, b_1 \in x_1x_2, a_2,b_2 \in x_2x_3$ such that $x_1, b_1, a_1, x_2$ and $x_3, a_2, b_2, x_2$ appear in these orders, and $a_1a_2, b_1b_2$ are initial segments, with an intersection point $c$. Then, in the region $a_1x_2b_2c$, the vertices $x_2$ and $c$ are internally blocked.\end{lemma}
\begin{proof} By Proposition~\ref{ForbStruct1}, we have $x_2$ internally blocked. Thus, $\Delta$ cannot have type $\mathbf{T}$, so by Proposition~\ref{finalS3prop} it follows that any two initial segments that intersect have to have their endpoints on the same edges of $T$.\\
\indent Suppose contrary, vertex $c$ is externally blocked in $x_2a_1cb_2$. Thus, exactly one of $a_1, b_2$ is internally blocked in this region, by symmetry, we may assume that $a_1$ is internally blocked. So $b_1$ is internally blocked in $a_1b_1c$. Let $q = \beta(b_1) \cap a_1c$. Applying the $(n-1)$-classification theorem~to the triangle blocking arrangement~induced by $x_2b_1b_2$, implies that it has type $\tbty{I}{1}$ or $\tbty{I}{2}$. Thus, there are vertices $p \in a_1b_1,r \in cb_2$, such that $p,q,r$ are collinear, and $pq, qr$ are minimal initial segments. Let $s$ be the intersection $l(pr) \cap b_2a_2$. However, $cb_2, qs$ intersect, while $a_2$ is externally blocked in $a_1a_2x_2$, so application of Proposition~\ref{ForbStruct1} results in contradiction.\qed\end{proof}
\begin{lemma}Let $\Delta$ and vertices $x_1,x_2,x_3,a_1,a_2,b_1,b_2,c$ satisfy the assumptions of Lemma~\ref{S12basicintersectionLemma}. Then, in the regions $a_1b_1c$ and $a_2b_2c$, the vertices $a_1,a_2,b_1,b_2,c$ are externally blocked.\end{lemma}
\begin{proof} As in the proof of Lemma~\ref{S12basicintersectionLemma}, intersecting initial segments must have endpoints on the same edges of $T$. Also, by that lemma, $x_2$ is internally blocked and $c$ is internally blocked in the region $a_1cb_2x_2$.\\
\indent Consider the vertex $b_1$. If we prove that $b_1$ is externally blocked in the $a_1b_1c$, then it follows that so is $a_1$, and looking at regions $a_1a_2x_2$ and $b_1b_2x_2$, the conclusion follows. Therefore, assume contrary, that $b_1$ is internally blocked in $a_1b_1c$.\\
\indent Set $q = \beta(b_1) \cap a_1c$. By $(n-1)$-classification theorem~applied to the triangle blocking arrangement~$\Delta_b$ induced by $b_1b_2x_2$, the type of $\Delta_b$ is either $\tbty{I}{1}$ or $\tbty{I}{2}$, but in either case we have vertices $p \in a_1x_2, r \in b_1c$, such that $p,q,r$ are collinear, and $pq, qr$ are minimal initial segments. Looking at $l(pr)$ and $b_1b_2$, as these intersect, the line $l(pr)$ must cross $a_2x_3$, with intersection point $s$, say. Recalling that the type of $\Delta_b$ is either $\tbty{I}{1}$ or $\tbty{I}{2}$, we have that $p$ is externally blocked in $pqa_1$, hence in $x_2ps$, the vertices $x_2, p$ are internally blocked, and therefore $s$ is externally blocked. But $qa_2$ and $rb_2$ cross at $c$, which is a contradiction by Proposition~\ref{ForbStruct1} applied to $x_2ps$.\qed\end{proof}
\begin{lemma} Let $\Delta$ and vertices $x_1,x_2,x_3,a_1,a_2,b_1,b_2,c$ satisfy the assumptions of Lemma~\ref{S12basicintersectionLemma}. Then, $a_1c$ and $b_2c$ are minimal and $x_2c$ is blocking.\end{lemma}
\begin{proof} By previous Lemma, we have that $a_1,b_1, a_2, b_2, c$ are externally blocked in regions $a_1b_1c$ and $a_2b_2c$. Apply the $(n-1)$-classification theorem~to the triangle blocking arrangement~$\Delta_b$ induced by $b_1b_2x_2$, thus $\Delta_b$ has type $\tbty{B}{1}$, $\tbty{I}{1}$ or $\tbty{I}{2}$. But, in either case, $a_1c$ is minimal. Similarly, $b_2c$ is minimal, and $\beta(x_2)$ must pass through $c$.\qed\end{proof}
\begin{lemma}\label{S12finalintersectionLemma} Let $\Delta$ and vertices $x_1,x_2,x_3,a_1,a_2,b_1,b_2,c$ satisfy the assumptions of Lemma~\ref{S12basicintersectionLemma}. Then, $a_2c$ and $b_1c$ are minimal. \end{lemma}
\begin{proof} Suppose contrary, $b_1c$ is not minimal. Thus, there is a vertex $q \in b_1c$. Let $pr \not= b_1b_2$ be the initial segment through $q$, with $p, r \in \partial T$. As $a_1c$ is minimal, without loss of generality, $p \in a_1b_1$. As in the proof of Lemma~\ref{S12basicintersectionLemma}, since $pr$ and $b_1b_2$ intersect, $r \in x_2x_3$. However, we may apply the previous lemma to $b_1b_2$ and $pr$, to obtain that $b_2q$ is minimal, which is a contradiction as $c \in b_2q$.\qed\end{proof}
\begin{corollary}\label{segmentsCorollary} Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. If $\Delta$ is not of type $\mathbf{T}$, then for every $\{j_1, j_2, j_3\} = [3]$ there are vertices $p_1, p_2, \dots ,p_k \in x_{j_1} x_{j_2}$ and $q_1, q_2, \dots, q_k \in x_{j_1}x_{j_3}$ such that $x_{j_1} p_1, p_i p_{i+1}, x_{j_1}q_1, q_i q_{i+1}$ are minimal, for all $i < k$, and one of the following alternatives holds.
\begin{enumerate}
\item Vertex $x_{j_1}$ is externally blocked. Each $p_iq_i$ is a minimal initial segment, and every initial segment with one vertex on $x_{j_1}x_{j_2}$ and the other vertex on $x_{j_1}x_{j_3}$ is among $p_iq_i$. Furthermore, $k$ is odd, and $p_{2i-1}q_{2i}, p_{2i}q_{2i-1}$ are blocking segments, for $i \leq \frac{k-1}{2}$, and $p_k, q_k$ are externally blocked in $x_{j_1} p_k q_k$.
\item Vertex $x_{j_1}$ is externally blocked. Each $p_iq_i$ is a minimal initial segment, and every initial segment with one vertex on $x_{j_1}x_{j_2}$ and the other vertex on $x_{j_1}x_{j_3}$ is among $p_iq_i$. Furthermore, $k$ is even, and $p_{2i-1}q_{2i}, p_{2i}q_{2i-1}$ are blocking segments, for $i \leq \frac{k}{2}$.
\item Vertex $x_{j_1}$ is internally blocked and $k$ is even. For all $i \leq k/2$, the segments $p_{2i-1}q_{2i}$ and $p_{2i}q_{2i-1}$ are initial segments and intersect at point $r_i$. Every initial segment with one vertex on $x_{j_1}x_{j_2}$ and the other vertex on $x_{j_1}x_{j_3}$ is among these. The initial segments $r_ip_{2i-1}, r_ip_{2i}, r_iq_{2i-1}r_iq_{2i}$ are minimal. The vertices $r_1, \dots, r_{k/2}$ all lie on $\beta(x_{j_1})$. Also, $p_1q_1$ is blocking, and $p_{2i}q_{2i+1}, p_{2i+1}q_{2i}$ are blocking for $i < k/2$, and $p_k$ is externally blocked in $x_{j_1}p_kq_{k-1}$, and $q_k$ is externally blocked in $x_{j_1}q_kp_{k-1}$.
\end{enumerate}
\end{corollary}
\begin{proof} Without loss of generality, $j_1 = 1, j_2 = 2, j_3 = 3$. We split into two cases, depending on whether some initial segments between $x_1x_2$ and $x_1x_3$ intersect or not. The possible outcomes are shown in Figure~\ref{cor19Fig}.\\[6pt]
\begin{figure}
\begin{center}
\includegraphics[width = 0.8\textwidth]{cor19.pdf}
\caption{Possibilities in Corollary~\ref{segmentsCorollary}}
\label{cor19Fig}
\end{center}
\end{figure}
\noindent\textbf{Case 1: there is an intersecting pair.} By Lemma~\ref{S12finalintersectionLemma}, $x_1$ is internally blocked. Let $pq$ be any initial segment with $p \in x_1x_2, q \in x_1x_3$. Then $\beta(x_1)$ crosses $pq$, at a point $r$, say, and let $p'q'$ be another initial segment through $r$, with $p',q' \in \partial T$. However, $\Delta$ is not of type $\mathbf{T}$, and by Proposition~\ref{finalS3prop}, without loss of generality, $p' \in x_1x_2, q' \in x_1x_3$. Applying Lemma~\ref{S12finalintersectionLemma} to $pq$ and $p'q'$, it follows that $pr, p'r, qr, q'r, pp', qq'$ are minimal. Observe further that if an initial segment $s$ has an endpoint on $x_1p$, unless the second endpoint is on $x_1q$, $s$ crosses $pq$, and thus $s = p'q'$. Combining these observations, we conclude that there are points $p_1, p_2, \dots, p_k \in x_1x_2, q_1, q_2, \dots, q_k \in x_1x_3$, such that $x_1p_1, p_1p_2, \dots, p_{k-1}p_k, x_1q_1, q_1q_2, \dots, q_{k-1}q_k$ are minimal, $k$ is even, $p_{2i-1}q_{2i}, p_{2i}q_{2i-1}$ are initial segments, and every initial segment with a vertex on $x_1x_2$ and another vertex on $x_1x_3$ is one of $p_{2i-1}q_{2i}, p_{2i}q_{2i-1}$. Furthermore, $p_{2i-1}q_{2i}, p_{2i}q_{2i-1}$ intersect at a point $r_i$, and $r_ip_{2i-1}, r_ip_{2i}, r_iq_{2i-1}, r_iq_{2i}$ are minimal for all $i \leq k/2$.\\
\indent From the information about minimal segments, we are forced to have $r_1, r_2, \dots,$ $r_{k/2} \in \beta(x_1)$, $p_1q_1$ blocking, in the minimal $\overline{\mathcal{S}}$-region $p_{2i}r_iq_{2i} q_{2i+1}$ $r_{i+1} p_{2i+1}$, all six vertices are internally blocked, for every $1 \leq i < k/2$, and finally $p_k$ is externally blocked in $x_1 p_k q_{k-1}$ and $q_k$ is externally blocked in $x_1 q_k p_{k-1}$. It remains to prove that for $i < k/2$, $p_{2i}q_{2i+1}$ and $p_{2i+1}q_{2i}$ are blocking.\\
Suppose contrary, there is some $i < k/2$, such that, without loss of generality, $p_{2i}q_{2i+1}$ is not blocking. Looking at minimal $\overline{\mathcal{S}}$-region $p_{2i}r_iq_{2i}q_{2i+1} r_{i+1}p_{2i+1}$, it follows that $p_{2i} q_{2i}$ is blocking. However, looking at $\beta(p_{2i})$ and $l(r_i q_{2i})$, which meet at $q_{2i}$, by assumption \textbf{(A)}, it follows that $l(p_{2i+1} r_{i+1})$ also contains $q_{2i}$, which is a contradiction.\\[6pt]
\noindent\textbf{Case 2: there are no intersecting pairs.} Observe that if $p \in x_1x_2, q \in x_1x_3$ and $pq$ is an initial segment, then, if any other segment $s$ crosses $pq$, then, by the assumption of this case, $s$ must have at least one vertex on $x_2x_3$, which is impossible by Proposition~\ref{finalS3prop}. Thus, $pq$ is minimal, and as in the previous case, if a segment $s$ has an endpoint in $x_1p$, its other endpoint is bound to be in $x_1q$. From this, we conclude that there are points $p_1, p_2, \dots,$ $p_k \in x_1x_2,$ $q_1, q_2, \dots,$ $q_k \in x_1x_3$, such that $x_1p_1, p_1p_2, \dots,$ $p_{k-1}p_k, x_1q_1,$ $q_1q_2, \dots,$ $q_{k-1}q_k$ are minimal, and $p_iq_i$ is a minimal initial segment for every $i \leq k$, and if $s$ is an initial segment with endpoints on $x_1x_2$ and $x_1x_3$, then $s = p_iq_i$ for some $i$. As $x_1p_1, x_1q_1$ are minimal, $p_1, q_1$ are externally blocked in $x_1p_1q_1$. Thus, $p_1q_2$ and $p_2q_1$ are blocking segments. Hence, in the region $p_2q_2q_3p_3$, $p_2$ and $q_2$ externally blocked and so are $p_3, q_3$. Proceeding in this fashion, the conclusion of the corollary follows.\qed\end{proof}
\begin{proposition}\label{FinalInternalBlocks}Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that $x_1, x_2$ are internally blocked. Then, $\Delta$ has one the types $\tbty{B}{1},\tbty{I}{1}$ or $\tbty{I}{2}$.\end{proposition}
\begin{proof} We consider three cases, depending on the outcomes of Corollary~\ref{segmentsCorollary}. We say that $x_i$ \emph{has no segments} if there are no segments with one vertex on $x_ix_{i'}$ and the other on $x_ix_{i''}$, where $i \in \{1,2\}$ and $\{i, i', i''\} = \{1,2,3\}$. Otherwise, we say that $x_i$ \emph{has segments}.\\[6pt]
\noindent\textbf{Case 1: both $x_1, x_2$ have no segments.} Let $u = \beta(x_1) \cap x_2x_3, v = \beta(x_2) \cap x_1x_3$. By Corollary~\ref{segmentsCorollary}, any initial segment in $\mathcal{S}$ is of the form $pq$, where $p \in x_1x_3, q \in x_2x_3$ and all these are minimal (and hence disjoint). In particular, no initial segment can cross $x_1u, x_2v$, and also, $x_1v, x_2u$ are minimal segments. Thus, the other initial segment through $u$, must cross $x_3v$, and the other initial segment through $v$ must cross $x_3u$. However, all initial segments are disjoint, so actually $uv$ is an initial segment, and it is minimal. It follows from Corollary~\ref{segmentsCorollary} at vertex $x_3$ that the type of $\Delta$ is $\tbty{B}{1}$.\\[6pt]
\noindent\textbf{Case 2: $x_1$ has, but $x_2$ has no segments.} By Corollary~\ref{segmentsCorollary}, we have vertices $a_1, a_2, \dots,$ $a_k \in x_1x_3,$ $b_1, b_2, \dots,$ $b_k \in x_1x_2$ such that $k$ is even and $x_1a_1, a_1a_2, \dots,$ $a_{k-1}a_k, x_1b_1, b_1b_2, \dots,$ $b_{k-1}b_k, b_kx_2$ are minimal, $a_{2i-1}b_{2i}, a_{2i} b_{2i-1}$ are initial segments that intersect at a point $c_i$. Let $u = \beta(x_1) \cap x_2x_3$. By Corollary~\ref{segmentsCorollary} applied to vertex $x_3$, we see that every initial segment is either among $a_ib_j$, or has vertices on $x_1x_3$ and $x_2x_3$ and is minimal. Hence, $x_2u$ is minimal. Let $v \in x_1x_3$ be such that $uv$ is an initial segment, and thus minimal. Hence, $va_k$ is also minimal, as otherwise, an initial segment with a vertex on $va_k$ would have the second endpoint on $x_2x_3$, so it would have to be minimal, but would cross $\beta(x_1) = x_1u$ or $vu$, which is impossible. Finally, $a_kc_{k/2}b_kx_2uv$ is a minimal $\overline{\mathcal{S}}$-region. Using assumption \textbf{(A)} as before, we see that $x_2a_k, vb_k$ are blocking, and it follows that the type of $\Delta$ is $\tbty{I}{1}$.\\[6pt]
\noindent\textbf{Case 3: both $x_1, x_2$ have segments.} By Corollary~\ref{segmentsCorollary}, we have vertices $a_1, a_2, \dots,$ $a_{k} \in x_1x_3$, $b_1, b_2, \dots,$ $b_{k}, d_1, d_2, \dots,$ $d_l \in x_1x_2$, $e_1, e_2, \dots,$ $e_l \in x_2x_3$ such that $x_1a_1, a_1a_2, \dots,$ $a_{k-1}a_k$, $x_1b_1, b_1b_2, \dots,$ $b_{k-1}b_k$, $x_2d_1, d_1d_2, \dots,$ $d_{l-1}d_l$, $x_2e_1, e_1e_2, \dots,$ $e_{l-1}e_l$ are minimal, and $a_1b_2, a_2b_1, \dots, a_{k-1}b_k, a_kb_{k-1}$, $d_1e_2, d_2e_1, \dots,$ $d_{l-1}e_l, d_l e_{l-1}$ are initials segments, $k$, $l$ are even and these are all initial segments that have at least one vertex on $x_1x_2$. Furthermore, $a_{2i-1}b_{2i}, a_{2i}b_{2i-1}$ intersect at point $c_i$, $d_{2i-1}e_{2i}, d_{2i}e_{2i-1}$ intersect at a point $f_i$, such that $c_1, c_2, \dots,$ $c_{k/2} \in \beta(x_1),$ $f_1, f_2, \dots,$ $f_{l/2} \in \beta(x_2)$. We also have that $b_kd_l$ is minimal.\\
\indent Suppose for a moment that $a_kx_3$ is minimal. Then, as $e_lf_{l/2},$ $f_{l/2}d_l,$ $d_lb_k,$ $b_k c_{k/2},$ $c_{k/2}a_k$ are minimal, it follows that $e_l x_3$ is also minimal, and hence $a_kc_{k/2}b_kd_l f_{l/2}e_lx_3$ is a minimal $\mathcal{L}$-region. But inside this minimal region, $\beta(f_{l/2})$ can only pass through $a_k$, and $\beta(c_{k/2})$ can only pass through $e_l$. However, then we have $\beta(b_k) = \beta(d_l)$, which is a contradiction. Therefore, $a_kx_3$ is not minimal.\\
\indent Let $v$ be the vertex in $a_kx_3$ such that $a_kv$ is minimal. As in the previous case, any initial segment with vertices on $x_1x_3$ and $x_2x_3$ is minimal. It follows that the initial segment through $v$, not equal to $x_1x_3$, is $vu$ with $u \in x_3e_l$ and it is minimal. Since $uv, va_k, a_kc_{k/2}, c_{k/2}b_k, b_kd_l, d_lf_{l/2}, f_{l/2}e_l$ are minimal, it follows that $ue_l$ is minimal as well. Therefore, $R = a_kc_{k/2}b_kd_lf_{l/2}e_luv$ is a minimal $\overline{\mathcal{S}}$-region. Using Corollary~\ref{segmentsCorollary}, to prove that $\Delta$ has type $\tbty{I}{2}$, it suffices to show that $a_k \in \beta(x_2), e_l \in \beta(x_1)$.\\
\indent Suppose contrary, that $\beta(x_1) \cap x_2x_3 \not= e_l$. By minimality of $R$, we must have $\beta(x_1) \cap x_2x_3 = u$. But, we would then have $\beta(u) \cap l(va_k) = x_1$. By the assumption \textbf{(A)}, it follows that $l(e_lf_{l/2})$ also passes through $x_1$, which is impossible. Thus $e_l \in \beta(x_1)$, and similarly we obtain $a_k \in \beta(x_2)$. Thus, $\Delta$ has type $\tbty{I}{2}$.\qed\end{proof}
Finally, it remains to classify the triangle blocking arrangements~without intersecting initial segments.
\begin{proposition} \label{DisjointSegmentsProp} Let $n \geq 1$ be an integer and suppose that $(n-1)$-classification theorem~holds. Let $\Delta = (T, \mathcal{S}, \mathcal{B})$ be a triangle blocking arrangement~of size $n$, and let $x_1, x_2$ and $x_3$ be the vertices of the triangle $T$. Suppose that no two initial segments in $\mathcal{S}$ intersect. Then $\Delta$ has a basic type or $\mathbf{T}$ (with $k=1$ in the definition of $\mathbf{T}$). \end{proposition}
\begin{proof} We say that a vertex $x_i$ is empty, if for $\{i,j,j'\} = \{1,2,3\}$, there are no initial segments between $x_ix_j$ and $x_ix_{j'}$. We distinguish between four cases, depending on the number of empty vertices.\\[3pt]
\noindent\textbf{Case 0: All four vertices are empty.} Then $\mathcal{L} = \emptyset$, and $\Delta$ has type $\tbty{B}{0}$.\\[3pt]
\noindent\textbf{Case 1: Only $x_1$ is non-empty.} Applying Corollary~\ref{segmentsCorollary}, $\Delta$ has type $\tbty{B}{1}$.\\[3pt]
\noindent\textbf{Case 2: Vertices $x_1, x_2$ are non-empty.} By Corollary~\ref{segmentsCorollary}, there are vertices $a \in x_1x_3,b,c \in x_1x_2,d \in x_2x_3$ such that $ab, cd$ are initial segments and all initial segments are either in $x_1ab$ or in $x_2cd$, and are disjoint. Moreover, we obtain the desired structure of blocking lines in regions $x_1ab, x_2cd$. Moreover, $bc, dx_3, ax_3$ are minimal, so $R = x_3abcd$ is a minimal $\overline{\mathcal{S}}$-region, inside which $a$ is internally blocked iff $b$ is, and $c$ is internally blocked iff $d$ is. If none of these four vertices are internally blocked in $R$, $\Delta$ is of type $\tbty{B}{2}$. Assume for contradiction that some vertex among them is internally blocked in $R$. Without loss of generality, one of $a,b$ is internally blocked, so both of them must be internally blocked. However, $x_3$ is externally blocked in $R$, so $\beta(b)$ must pass through $d$. However, we have $l(cd) \cap \beta(b) = d$, so by the assumption \textbf{(A)} applied to $R$ and vertex $b$, we have that $l(ax_3)$ passes through $d$, which is a contradiction, as desired.\\[3pt]
\noindent\textbf{Case 3: All three vertices are non-empty.} Similarly to the previous case, there are vertices $a,f \in x_1x_3, b,c \in x_1x_2, d,e \in x_2x_3$ such that $ab, cd, ef$ are minimal initial segments, and all initial segments are in regions $x_1ab,$ $x_2cd,$ $x_3ef$, and are minimal. From this, $fa, bc, de$ are also minimal. Moreover, we know the structure of blocking lines in $x_1ab, x_2cd, x_3ef$, and it remains to determine the structure of blocking lines in the minimal $\overline{\mathcal{S}}$-region $R = abcdef$.\\
\indent We have that in $R$, $a$ is internally blocked iff $b$ is, $c$ is internally blocked iff $d$ is, and $e$ is internally blocked iff $f$. If all these are externally blocked, then $\Delta$ has $\tbty{B}{3}$, as desired. Now, assume that, without loss of generality, one of vertices $a,b$ is internally blocked in $R$. But then both $a$ and $b$ must be internally blocked. Suppose for a moment that there is a blocking segment in $R$, which is a small diagonal of hexagon $abcdef$. By symmetry, we may suppose it contains $a$, so it is $ac$ or $ae$. If $ac$ is blocking, however, $\beta(a) \cap l(bc) = c$, so by the assumption \textbf{(A)}, $l(ef)$ has to contain $c$, which is impossible. Similarly, if $ae$ is blocking, $\beta(a)\cap l(fe) = e$, so by the assumption \textbf{(A)}, $l(bc)$ has to contain $e$, which is also impossible. Hence, the only possible blocking segments in $R$ are the main diagonals $ad, be, cf$. As $a,b$ are internally blocked, we have that $ad, be$ are blocking segments. But, as $d$ is internally blocked in $R$, so is $c$, so $cf$ is also blocking, showing that $\Delta$ has type $\tbty{B}{3}$.\qed\end{proof}
Combining all ingredients, we are ready to prove the classification theorem.
\begin{proof}[Proof of the classification theorem.] We prove the theorem by induction on the size $n$ of triangle blocking arrangement. The base of induction is $n = 0$, when the triangle blocking arrangement~has type $\tbty{B}{0}$.\\
\indent Now, assume that $n\geq 1$ and $(n-1)$-classification theorem~holds and let $\Delta$ be a triangle blocking arrangement~of size $n$ with vertices $x_1,x_2,x_3$. If $\Delta$ has intersecting initial segments that satisfy the conditions of Proposition~\ref{finalS3prop}, we are done. Otherwise, if there are any intersecting initial segments in $\Delta$ at all, by Proposition~\ref{ForbStruct1}, we must have some of $x_1, x_2,x_3$ internally blocked. Then, we are done by Proposition~\ref{FinalInternalBlocks}. Finally, if there are no intersecting initial segments in $\Delta$, we may apply Proposition~\ref{DisjointSegmentsProp} to finish the proof.\qed\end{proof}
\section{Concluding remarks}
Our first remark is that it would also be very interesting to classify all triangle blocking arrangements, without the assumption \textbf{(A)}. However, this is probably much harder, as the following discussion suggests.\\
\noindent\textbf{A comment about the assumption \textbf{(A)}.} As we have seen before, the assumption \textbf{(A)} is necessary in the classification theorem. However, there could be hope that we are using this assumption only locally, and that the arrangement types are rigid enough so that after some point, large arrangements are forced to combine as in the classification theorem. However, Figure~\ref{TBAnotInDef} shows that we cannot localize the assumption \textbf{(A)}. Namely, a natural weaker assumption would be that for some fixed $K$, and for any minimal $\bar{\mathcal{S}}$-region $R$, for any consecutive vertices $v_1, v_2, v_3, v_4, v_5$ appearing in this order on $\partial R$, we have that, if $l(v_1v_2)$ and $\beta(v_3)$ intersect in $T$ as some point $p$, $\beta(v_3)$ meets the interior of $R$, and $v_2p$ or $v_3p$ has at most $K$ points, then $l(v_1v_2), \beta(v_3), l(v_4v_5)$ are concurrent. But this figure shows that we may have as many points between as we want; the only region where \textbf{(A)} fails is $abcdef$, namely $l(ab), \beta(c)$ meet at $x_1$, but $x_1 \notin l(de)$, and this region satisfies the weaker assumption.\\[6pt]
\noindent\textbf{Relationship with Green Tao theorem on ordinary lines.} We discuss very briefly the proof of the result about ordinary lines of Green and Tao~\cite{GreenTao}. It can be summarized as follows.
\begin{itemize}
\item[\textbf{Step 1.}] Move to the dual.
\item[\textbf{Step 2.}] Apply Melchior's inequality (which is a consequence of Euler's formula) to get some control over point-line incidences.
\item[\textbf{Step 3.}] Use the incidence information to find large pieces with `triangular structure'.
\item[\textbf{Step 4.}] Study `triangular structure' to show that it looks like a hexagonal lattice.
\item[\textbf{Step 5.}] Apply the dual version of Chasles' theorem to place the points on a cubic.
\end{itemize}
\textbf{Step 4} corresponds to our classification theorem, and to emphasize the similarity, we phrase it as the following Classification Lemma. The conclusion is written slightly informally.
\begin{lemma}[Classification of triangular arrangements, Green and Tao~\cite{GreenTao}.] Let $T = x_1x_2x_3$ be a triangle in the plane, and let $\mathcal{S}$ be a collection of segments with endpoints on $\partial T$ with the property that whenever two segments in $\bar{\mathcal{S}} = \mathcal{S} \cup \{x_1x_2, x_2x_3, x_3x_1\}$ intersect, there is a unique third segment in $\bar{\mathcal{S}}$ that contains the intersection point, except possibly if the intersection is one of $x_1, x_2, x_3$, in which case there might not be the third segment. Then, $\bar{\mathcal{S}}$ forms a hexagonal grid shown in Figure~\ref{hexagonalGridStructFig}.
\end{lemma}
\begin{figure}
\begin{center}
\includegraphics[width = 0.3\textwidth]{hexagonalgrid.pdf}
\caption{Hexagonal grid}
\label{hexagonalGridStructFig}
\end{center}
\end{figure}
\begin{proof} We prove the claim by induction on $|\mathcal{S}|$. If $\mathcal{S}$ is empty, we are done. Assume now that we are given $\mathcal{S}$ and the claim holds for all smaller arrangements.\\
\indent Observe immediately that if $v$ is an intersection point on some edge of $T$, but not among the vertices $x_1, x_2, x_3$, then, we have $u, w \in \partial T$ such that $uv, wv \in \mathcal{S}$. Without loss of generality, $v \in x_1x_2, u \in x_1x_3$. If $w \in x_1x_3$ also, then, without loss of generality, $u$ is between $x_1$ and $w$, so applying the induction hypothesis to $vx_1w$ gives a contradiction, as hexagonal structure does not allow three segments at $v$. Hence, we must always have the two segments that meet on $\partial T$ between different pairs of edges of $T$.\\
\indent Similarly, we show that if two segment intersect, then they are between different pairs of edges of $T$. Suppose for the sake of contradiction that $a, c \in x_1x_2$ and $b,d \in x_1x_3$ are such that $ab, cd$ intersect at $e$. Without loss of generality, $c$ is between $x_1$ and $a$. Applying the induction hypothesis to $x_1ab$, we obtain a segment $ef$ with $f \in x_1b$. But, applying the induction hypothesis to $x_1cd$, we obtain a contradiction.\\
\indent Without loss of generality, we have a segment between $x_1x_2$ and $x_1x_3$. Pick an endpoint $v \in x_1x_2$ of such a segment with the property that $v$ is closest to $x_2$ among all such points. Let $u \in x_1x_3$ be such that $uv \in \mathcal{S}$. By observations before, there are no other points in $vx_2$ and all the segments between $x_1x_2$ and $x_1x_3$ are in $ux_1v$. In particular, $ux_3$ also has no points in its interior. We may apply the induction hypothesis to $x_1vu$, to obtain hexagonal structure there, with points $w_1, w_2, \dots, w_k$ appearing from $v$ to $u$. Consider segment $w_1b_1$ with $b_1 \in x_1u$. Then $l(w_1b_1)$ must cross $x_2x_3$, at some point $t_1$. But, then at $t_1$ we also have a segment with other endpoint on $x_1x_2$. However, by the choice of $w_1$, this may only be $v$. Next, consider $w_2$, and apply the same argument. We obtain a point $t_2 \in x_2x_3$ such that $t_2w_2$ is a part of a segment with other endpoint on $x_1x_3$, so similarly we obtain $t_2w_1$ is a subset of a segment in $\mathcal{S}$. Proceeding further in this fashion, we eventually obtain the hexagonal grid.\qed\end{proof}
It is therefore plausible that an extremal result could be proved with a similar general strategy, but given the significant differences in the difficulty of the relevant Classification Theorem, we expect that the new interesting difficulties will arise, in particular because not all types we defined come from duals of points on cubic curves. Nevertheless, we will investigate this further.\\[6pt]
\noindent \textbf{Classification Theorem for curves in the plane.} Going back to the proof of~classification theorem, we made a heavy use of topological properties of the real plane. However, we mainly focused on order of points on a line, and did not rely too much on the fact that the lines are straight (except that at intersection points the lines change sides with respect to one another). Instead of asking what happens over a different field, it could be possible that a similar, if not the same theorem holds for curves instead lines. Here we need some conditions on the curves, e.g. that we have some family of curves $\mathcal{C}$ with the property that through any two distinct points, there is a unique line in $\mathcal{C}$ containing them. Then, we could consider configurations where segments are intersections of curves in $\mathcal{C}$ with $T$. Or, we might not need to go that far and maybe we could consider curves with endpoints on $\partial T$ which are not self-intersecting and any two intersect in at most one point. This is something we shall also study further.\\
\indent Returning to the possibility of using a different field, this is of course another interesting question. However, over $\mathbb{C}$ we have, for example, the Hesse configuration (which can be realized as inflection points of a cubic curve), which gives 9 points, without ordinary lines. In this setting the interesting phenomenon is actually of a different nature.
\begin{theorem}[Kelly,~\cite{Kelly}] Any finite set of points in a complex space without ordinary lines is coplanar.\end{theorem}
We also expect that a classification theorem over finite fields would be very different from the one proved here.
\subsection{Acknowledgements}
\hspace{12pt}I would like to thank Trinity College and the Department of Pure Mathematics and Mathematical Statistics of Cambridge University for their generous support while this work was carried out and Imre Leader for helpful discussions concerning this paper. I also acknowledge the support of the Ministry of Education, Science and Technological Development of the Republic of Serbia, Grant III044006.
|
\section{Introduction}
\label{int}
Flavor anomalies present in various $B$ meson decays compelled physicist to look for beyond the standard model~(SM) physics. The deviation
from the SM prediction observed in the measured ratio
of branching ratios $R_D$ and $R_{D^{\ast}}$~\cite{Lees:2013uzd,Huschle:2015rga,Sato:2016svk,Hirose:2016wfn,Aaij:2015yra} in
$B \to (D,\,D^{\ast})\,l\nu$ decays and very recently measured $R_{J/\Psi}$~\cite{Aaij:2017tyk} in
$B_c \to J/\Psi\,l\,\nu$ decays provide hints of lepton flavor universality violation. Currently, the discrepancy in $R_D$ and $R_{D^{\ast}}$
with the SM prediction~\cite{Lattice:2015rga,Na:2015kha,Aoki:2016frl,Bigi:2016mdz,Bernlochner:2017jka,Jaiswal:2017rve,Wang:2017jow,
Fajfer:2012vx,Bigi:2017jbd} stands at
around $3.78\sigma$~\cite{Amhis:2016xyh}. Similarly, a deviation of around $1.3\sigma$ is observed in
$R_{J/\Psi}$ from the SM prediction~\cite{Ivanov:2005fd,Wen-Fei:2013uea,Dutta:2017xmj,Cohen:2018dgz}. Furthermore, a deviation of around
$1.5\sigma$~\cite{Ciezarek:2017yzh} is observed
in $\mathcal B(B \to \tau\nu)$ decays mediated via $b \to u\,l\,\nu$ charged current interactions as well.
If this persists in future experiments, it would be a definite hint of existence of beyond the SM physics.
Study of semileptonic $B_c$ meson decays will play an important role in the future. Being consist of two heavy quarks $b$ and $c$, the weak
decays of $B_c$
meson can occur via $b \to (u,\,d,\,c,\,s)$ and $c \to (u,\,d,\,s)$ transition decays. Although, measurements on non-leptonic decays of
$B_c$ meson exist~\cite{Patrignani:2016xqp,Aaij:2016xas}, however, for semileptonic decays, so far, LHCb has measured only the ratio
of branching ratios $R_{J/\Psi}$
in $B_c \to J/\Psi\,l\,\nu$ decays. With more and more data accumulated by the LHCb on $B_c$ meson,
detection of semileptonic $B_c$ meson decays to various other mesons mediating via $b \to (c,\,u)\,l\,\nu$ charged current and
$b \to s\,l\,\bar{l}$ neutral current interactions will be feasible. It should be
noted that study of various $B_c$ meson decays will act as complimentary decay channels to similar decays in the $B$ meson sector.
Moreover, study of semileptonic $B_c$ meson decays both theoretically and experimentally are not only important but also necessary to
maximize future sensitivity to new physics~(NP) contributions. Again, a precise determination of $B_c$ semileptonic branching ratio will be
crucial to determine the not so well known CKM matrix elements as well.
The semileptonic $B_c \to (D,\,D^{\ast})\,l\,\nu$ decay amplitude can be factorized into leptonic and hadronic matrix elements, where, the
leptonic matrix element can be
easily evaluated. The hadronic matrix element, on the other hand, depends on various meson to meson transition form factors that involve
QCD in the non perturbative regime. Several theoretical approaches exist such as the QCD sum rules, the covariant light front quark model,
the relativistic constituent quark model, and the non relativistic QCD to evaluate $B_c \to M$ transition form factors~\cite{Du:1988ws,
Colangelo:1992cx,Nobes:2000pm,Ivanov:2000aj,Kiselev:2000pp,Kiselev:2002vz,Ebert:2003cn,Ebert:2003wc}. We follow the covariant
light front quark model adopted in Ref.~\cite{Wang:2008xt} for the $B_c \to D$ and $B_c \to D^{\ast}$ transition form factors. Our main aim
is to provide prediction on various observables related to $B_c \to (D,\,D^{\ast})\,l\,\nu$ decays in the SM. We give first prediction on
various observables such as the ratio of branching ratios, lepton side forward backward asymmetry, polarization fraction of the charged
lepton, convexity parameter, longitudinal and transverse polarization fraction of $D^{\ast}$ meson and the forward backward asymmetry of the
transversely polarized $D^{\ast}$ meson.
The paper is organized as follows. In section~\ref{efh}, we start with the effective weak Hamiltonian for $b \to u\,l\,\nu$ transition decays
and subsequently write down the general form of the differential decay distribution for the $B_c \to (D,\,D^{\ast})\,l\,\nu$ semileptonic
decays. We write down explicit expressions of various observables pertaining to $B_c \to D\,l\,\nu$ and $B_c \to D^{\ast}\,l\,\nu$
decays in section~\ref{bcdlnu} and section~\ref{bcdslnu}, respectively. We present our results and discussion in section~\ref{rd} with
a brief summary and conclusion in section~\ref{con}.
\section{Formalism}
\subsection{Effective weak Hamiltonian}
\label{efh}
Effective field theory approach is the natural way to separate the long distance and the short distance effects coming from different scales
involved in the electroweak interaction~\cite{Buchalla:1995vs,Buras:1998raa}. Another advantage of using effective field theory is that it can
be easily extended to include NP effects by introducing new operators~\cite{Cirigliano:2009wk,Bhattacharya:2011qm}.
At the quark level, the $B_c \to (D,\,D^{\ast})\,l\,\nu$ decays are governed by the effective weak Hamiltonian in which the heavy degrees of
freedom are integrated out. The effects of the heavy degrees of freedom are absorbed into the short distance coefficients. The effective weak
Hamiltonian for the $b \to u\,l\,\nu$ quark level transition, within the SM, can be written as
\begin{eqnarray}
\label{effh}
\mathcal H_{eff} = \frac{G_F}{\sqrt{2}}\,V_{ub}\,\bar{u}\,\gamma_{\mu}(1-\gamma_5)\,b\,\bar{l}\,\gamma^{\mu}\,(1-\gamma^5)\,\nu_l\,,
\end{eqnarray}
where $G_F$ is the Fermi coupling constant and $V_{ub}$ is the relevant Cabibbo-Kobayashi-Maskawa~(CKM) matrix element. Using the effective
Hamiltonian of Eq.~\ref{effh}, the three body differential decay distribution for the $B_c \to (D,\,D^{\ast})\,l\,\nu$ decays can be
expressed in terms of two kinematic variables $q^2$ and $\theta_l$ as follows:
\begin{eqnarray}
\frac{d\Gamma}{dq^2 d\cos\theta_l}= \frac{G_{F}^2 |V_{ub}|^2 |\vec{P}_{(D,D^{\ast})}|}{2^9\,\pi^3\,m_{B_c}^2} \left(1-\frac{m_{l}^2}{q^2}
\right)\,L_{\mu \nu} H^{\mu \nu}\,,
\end{eqnarray}
where $q^2$ denotes lepton invariant mass squared and $\theta_l$ denotes the the angle between the $\vec{P}_{D,\,D^{\ast}}$ and lepton three
momentum vector in the $l-\nu$ rest frame. Here
$|\vec{P}_{(D,D^{\ast})}|= \sqrt{\lambda (m_{B_c}^2, m_{(D,D^{\ast})}^2, q^2)}/2m_{B_c}$ denotes the three momentum vector of the outgoing
meson, where, $\lambda(a,b,c)=a^2+b^2+c^2-2(ab+bc+ca)$. The covariant contraction of $L_{\mu\nu}$~(lepton tensor) and
$H^{\mu\nu}$~(hadron tensor) can be obtained by using the helicity techniques. A detail discussion of the helicity techniques can be found in
Refs.~\cite{Korner:1989qb,Kadeer:2005aq}.
\subsection{$B_c \to D\,l\,\nu$ decay observables}
\label{bcdlnu}
The full angular distribution of three body $B_c \to D\,l\,\nu$ decays can be written as
\begin{eqnarray}
\label{dlnutheta}
\frac{d\Gamma^D}{dq^2\,d\cos\theta_l} = 2\,N\,|\vec{P}_{D}|\,\Bigg\{H_0^2\,\sin^2\theta_l\, + \frac{m_l^2}{q^2}\,
\Big[(H_0\,\cos\theta_l - H_{t})^2\Big]\Bigg\}\,,
\end{eqnarray}
with
\begin{eqnarray}
&&N = \frac{G_F^2\,|V_{u\, b}|^2\,q^2}{256\,\pi^3\,m_{B_c}^2}\,\Big(1 - \frac{m_l^2}{q^2}\Big)^2\,, \qquad\qquad
H_0 = \frac{2\,m_{B_c}\,|\vec{P}_{D}|}{\sqrt{q^2}}\,f_{+}(q^2)\,, \qquad\qquad
H_t = \frac{m_{B_c}^2 - m_{D}^2}{\sqrt{q^2}}\,f_0(q^2)\,,
\end{eqnarray}
where $f_0(q^2)$ and $f_+(q^2)$ are the relevant $B_c \to D$ transition form factors.
By performing the integration over $\cos\theta_l$, we obtain the differential decay rate as follows:
\begin{eqnarray}
\frac{d\Gamma^D}{dq^2} = \frac{8\,N\,|\vec{P}_{D}|\,}{3}\Bigg\{\,H_0^2\,\Big(1 + \frac{m_l^2}{2\,q^2}\Big) +
\frac{3\,m_l^2}{2\,q^2}\,H_t^2\Bigg\}\,.
\end{eqnarray}
We also investigate the $q^2$ dependent differential ratio defined by
\begin{eqnarray}
\label{rbcd}
&&R(q^2)=\frac{{\rm DBR}(B_c \to D\tau\nu)(q^2)}{{\rm DBR}(B_c \to D\,l\,\nu)(q^2)}\,,
\end{eqnarray}
where, $l$ is either an electron or muon and ${\rm DBR}(B_c \to D\tau\nu)(q^2) = \frac{d\Gamma^D}{dq^2}/\Gamma$. Note that
$\Gamma = 1/\tau_{B_c}$ denotes the total decay width of the $B_c$ meson. Furthermore, we also explore several other observables such as
the lepton side forward backward asymmetry $A_{FB}^l(q^2)$, the longitudinal polarization fraction of the charged lepton $P^l(q^2)$,
and the convexity parameter $C_F^l(q^2)$ for the $B_c \to D\,l\,\nu$ decays.
The forward backward asymmetry in the angular distribution is defined as
\begin{eqnarray}
\label{afbd}
A^l_{FB}(q^2) = \frac{\Big(\int_{-1}^{0}-\int_{0}^{1}\Big)d\cos\theta\frac{d\Gamma^D}{dq^2\,d\cos\theta}}{\frac{d\Gamma^D}{dq^2}} \, =
\frac{3m_{l}^{2}}{2q^2}\frac{H_0\,H_t}{H_0^2\,\Big(1+\frac{m_l^2}{2\,q^2}\Big)+\frac{3\,m_l^2}{2q^2}\,H_t^2}\,.
\end{eqnarray}
Similarly, the longitudinal polarization fraction of the charged lepton is defined as
\begin{eqnarray}
\label{pd}
P^l(q^2)=\frac{d\Gamma^D(+)/dq^2 - d\Gamma^D(-)/dq^2}{d\Gamma^D/dq^2} = \frac{\frac{m_l^2}{2\,q^2}\,(H_0^2+3\,H_t^2) - H_0^2}
{\frac{m_l^2}{2\,q^2}\,(H_0^2+3\,H_t^2) + H_0^2}\,,
\end{eqnarray}
where $d\Gamma^D(+)/dq^2$ and $d\Gamma^D(-)/dq^2$ define the differential decay rate of the positive and negative helicity leptons,
respectively.
Again, by taking second derivative of Eq.~\ref{dlnutheta} with respect to $\cos\theta_l$, we obtain the convexity parameter. That is
\begin{eqnarray}
\label{cfd}
C_F^l(q^2) = \frac{1}{\left(d\Gamma^D/dq^2\right)}\frac{d^2}{d(\cos\theta)^2}\left[\frac{d\Gamma^D}{dq^2\,d\cos\theta}\right]\, =
\frac{3}{2}\frac{H_0^2\Big(\frac{m_l^2}{q^2}-1\Big)}
{\Big[H_0^2\,\Big(1+\frac{m_l^2}{2\,q^2}\Big)+\frac{3\,m_l^2}{2\,q^2}\,H_t^2\Big]}\,.
\end{eqnarray}
We also give predictions on the average values of the ratio of branching ratios $R_{B_cD}$, forward backward asymmetry $A_{FB}^l$, the
longitudinal polarization fraction $P^l$, and the convexity parameter $C_F^l$ for the $B_c \to D\,l\,\nu$ decay mode by separately
integrating the numerator and the denominator of Eq.~\ref{rbcd}, Eq.~\ref{afbd}, Eq.~\ref{pd}, and Eq.~\ref{cfd} over $q^2$.
\subsection{$B_c \to D^{\ast}\,l\,\nu$ decay observables}
\label{bcdslnu}
We write the three body $B_c \to D^{\ast}\,l\,\nu$ differential~($q^2$, $\cos\theta_l$) decay distribution as
\begin{eqnarray}
\frac{d\Gamma^{D^{\ast}}}{dq^2\,d\cos\theta_l}= N\,|\vec{P}_{D^{\ast}}|\Bigg\{2\mathcal{A}_0^2\,\sin^2\theta_l
+ \Big[\Big(1+\cos^2\theta_l\Big) + \frac{m_l^2}{q^2}\sin^2\theta_l\Big]\Big[\mathcal{A}_{\|}^2 + \mathcal{A}_{\bot}^2\Big]
-4\,\mathcal{A}_{\|}\mathcal{A}_{\bot}\cos\theta_l + \frac{2\,m_l^2}{q^2}\Big[\mathcal{A}_0\cos\theta_l - \mathcal{A}_t\Big]^2\Bigg\}\,.
\end{eqnarray}
By performing the integration over $\cos\theta_l$, we obtain
\begin{eqnarray}
\frac{d\Gamma^{D^{\ast}}}{dq^2}=\frac{8\,N\,|\vec{P}_{D^{\ast}}|}{3}\Big[\Big(\mathcal{A}^2_0+ \mathcal{A}^2_{\|}+
\mathcal{A}^2_{\bot}\Big)\Big(1 + \frac{m_l^2}{2\,q^2}\Big)+\frac{3\,m_l^2}{2\,q^2}
\,\mathcal{A}^2_t\Big]\,
\end{eqnarray}
with
\begin{eqnarray}
&&\mathcal{A}_0=\frac{1}{2\,m_{D^{\ast}}\,\sqrt{q^2}}\Big[\Big(\,m_{B_c}^2-m_{D^{\ast}}^2-q^2\Big)(m_{B_c}+m_{D^{\ast}})A_1(q^2)\,-\,
\frac{4\,m_{B_c}^2|\vec{p}_{D^{\ast}}|^2}{m_{B_c}+m_{D^{\ast}}}A_2(q^2)\Big]\,,\nonumber\\
&&\mathcal{A}_{\|}=\frac{2\,(m_{B_c}+m_{D^{\ast}})\,A_1(q^2)}{\sqrt{2}}\,, \qquad
\mathcal{A}_{\bot}=-\frac{4\,m_{B_c}\,V(q^2)\,|\vec{P}_{D^{\ast}}|}{\sqrt{2}\,(m_{B_c}+m_{D^{\ast}})}\,, \qquad
\mathcal{A}_{t}=\frac{2\,m_{B_c}|\vec{P}_{D^{\ast}}|\,A_0(q^2)}{\sqrt{2}}\,,
\end{eqnarray}
where $V(q^2)$, $A_0(q^2)$, $A_1(q^2)$, and $A_2(q^2)$ are the relevant $B_c \to D^{\ast}$ transition form factors.
Similar to $B_c \to D\,l\,\nu$ decay observables, we investigate differential ratio $R(q^2)$, forward backward asymmetry parameter
$A_{FB}^l(q^2)$, longitudinal polarization fraction of the charged lepton $P^l(q^2)$, and the convexity
parameter $C_F^l(q^2)$ for the $B_c \to D^{\ast}\,l\,\nu$ decays. The explicit expressions are
\begin{eqnarray}
\label{q2_obs_ds}
&&R(q^2) = \frac{{\rm DBR}(B_c \to D^{\ast}\tau\nu)(q^2)}{{\rm DBR}(B_c \to D^{\ast}\,l\,\nu)(q^2)}\,,\qquad\qquad
l\in (e,\,\mu)\,,\nonumber \\
&&A_{FB}^{l}(q^2) = \frac{3}{2}\frac{\mathcal{A}_{\|}\mathcal{A}_{\bot} + \frac{m_{l}^{2}}{q^{2}}\mathcal{A}_{0} \mathcal{A}_{t}}
{\Big(\mathcal{A}_{0}^2 + \mathcal{A}_{\|}^2 + \mathcal{A}_{\bot}^2 \Big) \Big( 1+ \frac{m_{l}^{2}}{2q^{2}} \Big) +
\frac{3\,m_{l}^{2}}{2\,q^{2}}\mathcal{A}_{t}^2}\,, \nonumber \\
&&P^{l}= \frac{\Big(\mathcal{A}_{0}^2 + \mathcal{A}_{\|}^2 + \mathcal{A}_{\bot}^2 \Big)\Big(\frac{m_{l}^{2}}{2q^{2}}-1 \Big) -
\frac{3\,m_{l}^{2}}{2q^{2}}\mathcal{A}_{t}^2}{\Big(\mathcal{A}_{0}^2 + \mathcal{A}_{\|}^2 + \mathcal{A}_{\bot}^2 \Big)
\Big( 1+ \frac{m_{l}^{2}}{2q^{2}} \Big) + \frac{3\,m_{l}^{2}}{2q^{2}}\mathcal{A}_{t}^2}\,, \nonumber \\
&&C_F^l= \frac{3}{4}\frac{\Big(\frac{m_l^2}{q^2}-1 \Big) \Big(2\,\mathcal{A}_{0}^2 - \mathcal{A}_{\|}^2 -
\mathcal{A}_{\bot}^2 \Big)}{\Big(\mathcal{A}_{0}^2 + \mathcal{A}_{\|}^2 + \mathcal{A}_{\bot}^2 \Big) \Big( 1+ \frac{m_l^2}{2\,q^2}
\Big) + \frac{3\,m_l^2}{2\,q^2} \mathcal{A}_{t}^2}\,.
\end{eqnarray}
We also explore the forward backward asymmetry for the transversely polarized $D^{\ast}$ meson $A_{FB}^T(q^2)$ obtained by dropping the
$\mathcal A_0$ and $\mathcal A_t$ from $A_{FB}^{l}(q^2)$ of Eq.~\ref{obs_ds}. Similarly, we define the longitudinal and transverse
polarization fractions $F_L^{D^{\ast}}(q^2)$ and $F_T^{D^{\ast}}(q^2)$ of the $D^{\ast}$ meson. The explicit expressions are as follows:
\begin{eqnarray}
\label{q2_obs_ds1}
&&A_{FB}^T(q^2) = \frac{3}{2}\frac{\mathcal{A}_{\|}\mathcal{A}_{\bot}}
{\Big(\mathcal{A}_{\|}^2 + \mathcal{A}_{\bot}^2 \Big) \Big( 1+ \frac{m_{l}^{2}}{2\,q^{2}}\Big)}\,,\nonumber \\
&&F_L^{D^{\ast}}(q^2) = \frac{\mathcal A_0^2\Big(1 + \frac{m_l^2}{2\,q^2}\Big)+\frac{3\,m_l^2}{2\,q^2}\,A_t^2}{\Big(\mathcal{A}^2_0+
\mathcal{A}^2_{\|}+ \mathcal{A}^2_{\bot}\Big)\Big(1 + \frac{m_l^2}{2\,q^2}\Big)+\frac{3\,m_l^2}{2\,q^2}\,\mathcal{A}^2_t}\,, \nonumber \\
&&F_T^{D^{\ast}}(q^2) = \frac{\Big(\mathcal A_{\|}^2 + \mathcal A_{\bot}^2\Big)\Big(1 + \frac{m_l^2}{2\,q^2}\Big)}{\Big(\mathcal{A}^2_0+
\mathcal{A}^2_{\|}+ \mathcal{A}^2_{\bot}\Big)\Big(1 + \frac{m_l^2}{2\,q^2}\Big)+\frac{3\,m_l^2}{2\,q^2}\,\mathcal{A}^2_t}\,,
\end{eqnarray}
where $F_L^{D^{\ast}} + F_T^{D^{\ast}} = 1$. Similar to $B_c \to D\,l\,\nu$ decays in section~\ref{bcdlnu}, we also give perdictions on
the average values of all the observables defined in Eq.~\ref{q2_obs_ds} and Eq.~\ref{q2_obs_ds1}.
\section{Results}
\label{rd}
\subsection{Inputs}
\label{in}
For our numerical estimates, we use $m_D = 1.86486\,{\rm GeV}$, $m_{D^{\ast}}=2.00698\,{\rm GeV}$, and $m_{B_c} = 6.2751\,{\rm GeV}$.
Similarly for the lepton masses, we use $m_e=0.5109989461\times 10^{-3}\,{\rm GeV}$ and $m_{\tau}=1.77682\,{\rm GeV}$. The mass of $b$ quark
at the renormalization scale $\mu = m_b$ is taken to be $m_b(m_b) = 4.18\,{\rm GeV}$. For the
lifetime of $B_c$ meson, we use $\tau_{B_c} = 0.507d-12\,{\rm S}$. We use Ref.~\cite{Patrignani:2016xqp} for all the above mentioned inputs.
The CKM matrix element is taken to be $|V_{ub}| = (36.1\pm 3.2)\times 10^{-4}$~\cite{Patrignani:2016xqp}.
In order to make predictions on various observables, we need the information of
various form factors related to $B_c \to D$ and $B_c \to D^{\ast}$ transitions. We follow Ref.~\cite{Wang:2008xt} and report in
Table.~\ref{tab_ff} the relevant form factor inputs that are obtained in the covariant light-front quark model.
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$F$ & $F(0)$ &$c_1$ &$c_2$ \\
\hline
\hline
$f_+^{B_c\,D}$ &$0.16\pm 0.03$ &$3.46\pm 0.31$ & $0.90\pm 0.08$ \\
\hline
$f_0^{B_c\,D}$ &$0.16\pm 0.03$ &$2.41\pm 0.28$ &$0.47\pm 0.06$ \\
\hline
$V^{B_c\,D^{\ast}}$ &$0.13\pm 0.03$ &$4.21\pm 0.39$ &$1.09\pm 0.10$ \\
\hline
$A_0^{B_c\,D^{\ast}}$ & $0.09\pm 0.01$ & $4.18\pm 0.40$ & $0.96\pm 0.11$ \\
\hline
$A_1^{B_c\,D^{\ast}}$ & $0.08\pm 0.01$ & $3.18\pm 0.37$ & $0.65\pm 0.08$ \\
\hline
$A_2^{B_c\,D^{\ast}}$ & $0.07\pm 0.01$ & $3.78\pm 0.35$ & $0.80\pm 0.07$ \\
\hline
\hline
\end{tabular}
\caption{$B_c \to D,\,D^{\ast}$ form factor inputs from Ref.~\cite{Wang:2008xt}.}
\label{tab_ff}
\end{table}
The momentum dependence of the form factors is parametrized as follows:
\begin{eqnarray}
F(q^2) = F(0)\,\exp(c_1\,s + c_2\,s^2)\,,
\end{eqnarray}
where $s = q^2/m_{B_c}^2$ and $F$ represents the form factors $f_+$, $f_0$ and $V$, $A_0$, $A_1$, and $A_2$, respectively.
Uncertainties in the calculation of the $B_c \to (D,\,D^{\ast})\,l\,\nu$ decay amplitudes may come from two kinds of inputs.
First kind of inputs are very well known such as mass of mesons, mass of quarks, mass of leptons and lifetime of the $B_c$ meson.
Second kind of inputs are not very well known hadronic parameters such as the CKM matrix element $|V_{ub}|$ and $B_c \to (D,\,D^{\ast})$
transition form factors. To estimate
the uncertainties associated with each observable, we perform a $\chi^2$ analysis based on the uncertainties coming from the not so
well known hadronic inputs only. Uncertainties associated with the well known inputs are omitted in our $\chi^2$ analysis.
\subsection{SM prediction of $B_c \to D\,l\,\nu$ decay observables}
The SM prediction of $B_c \to D\,l\,\nu$ decay observables are reported in Table.~\ref{tab_d}. We use the central values of all the input
parameters to obtain the central values of each observables. To find the SM range, we define a $\chi^2$ as follows:
\begin{eqnarray}
\chi^2 = \sum_{i=1}^{7}\frac{(\mathcal O_i - \mathcal O_i^0)^2}{\sigma_{i}^2}\,,
\end{eqnarray}
where $\mathcal O_i = \Big(|V_{ub}|,\,f_0,\,f_+,\,c_1^{f_0},\,c_2^{f_0},\,c_1^{f_+},\,c_2^{f_+}\Big)$. Here $\mathcal O_i^0$ represents the
central value of each parameter and $\sigma_i$ represents the $1\sigma$ uncertainty associated with each parameter. The range of each
observable is obtained by demanding $\chi^2 \le 0.989$.
The branching ratio of $B_c \to D\,l\,\nu$ is found to be of order $10^{-4}$ for both $e$ and the $\tau$ modes which is in good agreement with
the values obtained in Ref.~\cite{Wang:2008xt}. The slight difference may come from different choices of the input parameters. The value of
each observables for the $e$ mode differs significantly from that of the $\tau$ mode. As expected, we obtain $A_{FB}^e=0$, $P^e = -1$, and
$C_F^e = -1.5$ in the SM.
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|c||c|c|c|}
\hline
Observables & Central value &Range &Observables &Central value &Range \\
\hline
\hline
$\mathcal B(B_c \to D\,e\nu)\times 10^4$ &$0.277$ &$[0.183, 0.406]$ &$\mathcal B(B_c \to D\tau\nu)\times 10^4$ &$0.195$ &$[0.138, 0.264]$\\
\hline
$P^e$ &$-1.000$ &$-1.000$&$P^{\tau}$ &$-0.047$ &$[-0.263, 0.183]$\\
\hline
$A^e_{FB}$ &$0.000$ &$0.000$&$A^{\tau}_{FB}$ &$0.294$ &$[0.269, 0.301]$\\
\hline
$C^e_F$ &$-1.500$ &$-1.500$&$C^{\tau}_F$ &$-0.556$ &$[-0.433, -0.672]$\\
\hline
$R_{B_c\,D}$ &$0.703$ &$[0.589, 0.891]$ & & &\\
\hline
\hline
\end{tabular}
\caption{SM prediction of various observables for the $B_c \to D\,l\,\nu$ semileptonic decays.}
\label{tab_d}
\end{table}
We show in Fig.~\ref{dbr_r_d} the differential branching ratio ${\rm DBR}(q^2)$ and differential ratio $R(q^2)$ as a function of $q^2$.
The solid lines correspond to the central values of the input parameters, whereas, the bands correspond to
uncertainties in the CKM matrix element $|V_{ub}|$ and various $B_c \to D$ transition form factor inputs. It is observed that the differential
decay distribution is zero at $q^2 = m_l^2$ and $q^2 = q^2_{\rm max} = (m_{B_c} - m_D)^2$. This is obvious because at $q^2 = m_l^2$, $N$ goes
to zero and at $q^2 = (m_{B_c} - m_D)^2$, $\vec{P}_D$ goes to zero. The peak of the differential decay distribution for the $e$
mode~($3.0\times 10^{-6}\,{\rm GeV}^{-2}$) and the $\tau$ mode~($2.5\times 10^{-6}\,{\rm GeV}^{-2}$) occurs at around $q^2 = 11\,{\rm GeV}^2$
and $q^2 = 12.5\,{\rm GeV}^2$, respectively. For the differential ratio $R(q^2)$, we obtain the maximum value of $R(q^2) = 4$ at
$q^2 = q^2_{\rm max}$. It is worth mentioning that the uncertainty is
less in $R(q^2)$ than ${\rm DBR}(q^2)$ due to cancellation of various uncertainties that are common to both the numerator and the denominator.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=5.8cm,height=5.0cm]{dbrtauq2_smcentral_d.png}
\includegraphics[width=5.8cm,height=5.0cm]{rq2_smcentral_d.png}
\end{center}
\caption{Differential branching ratio ${\rm DBR}(q^2)$ and differential ratio $R(q^2)$ for $B_c \to D\,l\,\nu$ decays in the SM. The pink
band~($e$ mode) and the blue band~($\tau$ mode)
correspond to uncertainties in $|V_{ub}|$ and $B_c \to D$ form factor inputs. The pink and blue solid lines represent the SM prediction
for the $e$ and the $\tau$ modes obtained using the central values of all the input parameters.}
\label{dbr_r_d}
\end{figure}
The $q^2$ dependence of the forward backward asymmetry $A_{FB}^l(q^2)$, longitudinal polarization fraction $P^l(q^2)$ and convexity
parameter $C_F^l(q^2)$ in the SM is shown in Fig.~\ref{obs_d}. We observe that $A_{FB}^l(q^2)$, $P^l(q^2)$, and $C_F^l(q^2)$ for the $e$
mode remain constant throughout the whole $q^2$ region. However, for the $\tau$ mode, $A_{FB}^{\tau}(q^2)$ peaks~($0.5$) at low $q^2$ and
gradually decreases as $q^2$ increases. It becomes zero at $q^2 = q^2_{\rm max}$. For the longitudinal polarization fraction of the $\tau$
lepton, $P^{\tau}(q^2)$ is around $0.5$ at low $q^2$, reaches its minimum value~($-0.47$) at $q^2 = 12.5\,{\rm GeV}^2$ and gradually increases
to $1$ at high $q^2$. It can assume
both positive and negative values depending on the value of $q^2$. Considering the central blue solid line, we find two zero crossings in the
$P^{\tau}(q^2)$ observable at $q^2 = 7.2\,{\rm GeV}^2$ and $q^2 = 16.5\,{\rm GeV}^2$, respectively. The convexity parameter $C_F^{\tau}(q^2)$
is zero at low and high $q^2$. It is found to be minimum~($-0.8$) at $q^2 = 15\,{\rm GeV}^2$. Although the forward backward asymmetry
$A_{FB}^{\tau}(q^2)$ is always positive, the convexity parameter $C_F^{\tau}(q^2)$ is always negative throughout the whole $q^2$ region.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=5.8cm,height=5.0cm]{afbtauq2_smcentral_d.png}
\includegraphics[width=5.8cm,height=5.0cm]{taptauq2_smcentral_d.png}
\includegraphics[width=5.8cm,height=5.0cm]{cftauq2_smcentral_d.png}
\end{center}
\caption{Forward backward asymmetry $A_{FB}^l(q^2)$, longitudinal polarization fraction $P^l(q^2)$ and convexity parameter $C_F^l(q^2)$ for
the $B_c \to D\,l\,\nu$ decays within the SM. Notations are same as in Fig.~\ref{dbr_r_d}.}
\label{obs_d}
\end{figure}
\subsection{SM prediction of $B_c \to D^{\ast}\,l\,\nu$ decay observables}
We first report in Table.~\ref{tab_ds} the central value and the allowed range of each observables for the $B_c \to D^{\ast}\,l\,\nu$
semileptonic decays, where, $l$ is either an electron or a tau lepton, respectively. The central value of each observable is obtained
by using the central value of each input parameters, whereas, the allowed range is obtained by effecting a $\chi^2$ defined as
\begin{eqnarray}
\chi^2 = \sum_{i=1}^{13}\frac{(\mathcal O_i - \mathcal O_i^0)^2}{\sigma_{i}^2}\,,
\end{eqnarray}
where $\mathcal O_i$ denotes the CKM matrix element $|V_{ub}|$ and the $B \to D^{\ast}$ form factor inputs such as $V$, $A_0$, $A_1$, $A_2$
and the corresponding values of $c_1$ and $c_2$, respectively. Here $\mathcal O_i^0$ represents the
central value of each parameter and $\sigma_i$ represents the $1\sigma$ uncertainty associated with each parameter. We obtain the SM range
by demanding $\chi^2 \le 3.565$. The branching ratio $\mathcal B(B_c \to D^{\ast}\,l\,\nu)$ is of the order of $10^{-4}$ for both $e$ and
the $\tau$ modes, respectively. This is in good agreement with the results obtained in Ref.~\cite{Wang:2008xt}. In the SM, we obtain the
ratio $F_L^{D^{\ast}}/F_T^{D^{\ast}}$ to be around $2/3$ for the $e$ as well as the $\tau$ mode.
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|c||c|c|c|}
\hline
Observables & Central value &Range &Observables &Central value &Range \\
\hline
\hline
$\mathcal B(B_c \to D^{\ast}\,e\nu)\times 10^4$ &$0.395$ &$[0.217,0.715]$ & $\mathcal B(B_c \to D^{\ast}\tau\nu)\times 10^4$ &$0.238$ &
$[0.135,0.424]$\\
\hline
$P^e$ &$-1.000$ &$-1.000$&$P^{\tau}$ &$-0.654$ & $[-0.544,-0.715]$\\
\hline
$A^e_{FB}$ &$-0.365$ &$[-0.234, -0.504]$&$A^{\tau}_{FB}$ &$-0.222$ &$[-0.113,-0.345]$\\
\hline
$C^e_F$ &$-0.162$ &$[0.079, -0.319]$&$C^{\tau}_F$ &$-0.039$ &$[0.088,-0.124]$\\
\hline
$(A_{FB}^T)^e$ & $-0.613$ & $[-0.435,-0.720]$ & $(A_{FB}^T)^{\tau}$ & $-0.527$ & $[-0.364, -0.633]$\\
\hline
$(F_L^{D^{\ast}})^e$ & $0.405$ & $[0.298,0.475]$ &$(F_L^{D^{\ast}})^{\tau}$ & $0.412$ & $[0.334,0.465]$\\
\hline
$(F_T^{D^{\ast}})^e$ & $0.595$ & $[0.525,0.702]$ &$(F_T^{D^{\ast}})^{\tau}$ & $0.588$ & $[0.535,0.666]$\\
\hline
$R_{B_c\,D^{\ast}}$ &$0.603$ &$[0.573, 0.651]$ & & &\\
\hline
\hline
\end{tabular}
\caption{SM prediction of various observables for the $B_c \to D^{\ast}\,l\,\nu$ semileptonic decays.}
\label{tab_ds}
\end{table}
In Fig.~\ref{dbr_r_ds}, we show the differential branching ratio ${\rm DBR}(q^2)$ and differential ratio $R(q^2)$ as a function of the
lepton invariant mass squared $q^2$. The solid lines correspond to the central values of the input parameters, whereas, the bands correspond
to the uncertainties associated with the CKM matrix element $|V_{ub}|$ and various $B_c \to D^{\ast}$ transition form factor inputs.
The pink band represents the $e$ mode and the blue band represents the $\tau$ mode, respectively.
We observe that the ${\rm DBR}(q^2)$ for $B_c \to D^{\ast}\,e\,\nu$ peaks~($8\times 10^{-6}\,{\rm GeV}^{-2}$) at
$q^2 = 15\,{\rm GeV}^2$, whereas, for the $B_c \to D^{\ast}\,\tau\,\nu$, it peaks~($5.5\times 10^{-6}\,{\rm GeV}^{-2}$) at
$q^2 = 15\,{\rm GeV}^2$. We observe that as $q^2$ increases, $R(q^2)$ gradually increases and reach its maximum value $0.75$ at
$q^2 = q^2_{\rm max}$. It is evident that the uncertainty in $R(q^2)$ is greatly reduced in comparison to the uncertainty in
${\rm DBR}(q^2)$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=5.8cm,height=5.0cm]{dbrtauq2_smcentral_ds.png}
\includegraphics[width=5.8cm,height=5.0cm]{rq2_smcentral_ds.png}
\end{center}
\caption{Differential branching ratio ${\rm DBR}(q^2)$ and differential ratio $R(q^2)$ for $B_c \to D^{\ast}\,l\,\nu$ decays in the SM.
The pink band~($e$ mode) and the blue band~($\tau$ mode) correspond to uncertainties in $|V_{ub}|$ and $B_c \to D^{\ast}$ form factor inputs.
The pink and blue solid lines represent
the SM prediction for the $e$ and the $\tau$ modes obtained using the central values of all the input parameters.}
\label{dbr_r_ds}
\end{figure}
We show in Fig.~\ref{obs_ds} various other $q^2$ dependent observables for the $B_c \to D^{\ast}\,l\,\nu$ decays. If we consider the central
values denoted by the solid lines, we find that there exists a zero crossing in the forward backward asymmetry parameter $A_{FB}^{\tau}(q^2)$
at $q^2 = 6.5\,{\rm GeV}^2$ above which it becomes negative. $A_{FB}^{\tau}(q^2)$ peaks~($0.3$) at low $q^2$ and reach its minimum of around
$-0.4$ at $q^2 = 15\,{\rm GeV}^2$. It becomes zero at high $q^2$. However, $A_{FB}^{e}(q^2)$ remains negative throughout the whole $q^2$
region. Similarly, we observe a zero crossing in the $\tau$ polarization fraction
$P^{\tau}(q^2)$ at $q^2 = 4.3\,{\rm GeV}^2$. $P^{\tau}(q^2)$ is around $0.4$ at low $q^2$ and gradually decreases to around $-0.8$ at high
$q^2$. We observe that $P^{e}(q^2)$ remains constant~($-1.0$) in the whole $q^2$ region. It is observed that, if we consider the central
solid lines, $C_F^l(q^2)$ for both $e$ and $\tau$ modes remains negative throughout the whole $q^2$ region. However considering the bands,
they can assume positive values depending on the value of $q^2$. The magnitude of $C_F^{e}(q^2)$ is much larger than the magnitude of
$C_F^{\tau}(q^2)$. Magnitude of $C_F^{e}(q^2)$ can be as large as $-1.2$ at low $q^2$, whereas, $C_F^{\tau}(q^2)$ remains very small in the
whole $q^2$ region. The asymmetry
$A_{FB}^T(q^2)$ for both $e$ and $\tau$ mode are always negative. For the $e$ mode, $A_{FB}^T(q^2)$ can be as large as $-0.75$ below
$q^2 = 15\,{\rm GeV}^2$, whereas, for the $\tau$ mode, it can assume large value of around $-0.65$ at $q^2 = 15\,{\rm GeV}^2$. In the SM,
the longitudinal polarization
fraction of the $D^{\ast}$ meson $F_L^{D^{\ast}}(q^2)$ for the $\tau$ mode can be as large as $0.85$ at low $q^2$ and gradually decreases to
about $0.3$ at high $q^2$. Similarly for the $e$ mode, it peaks~($1.0$) at low $q^2$ and decreases to about $0.2$ at high $q^2$. Again,
as expected, $F_T^{D^{\ast}}(q^2)$ for the $\tau$ mode is minimum~($0.15$) at low $q^2$ and it peaks~($0.7$) at high $q^2$. For the $e$ mode,
$F_T^{D^{\ast}}(q^2)$ is zero at low $q^2$ and peaks~($0.8$) at high $q^2$. $F_L^{D^{\ast}}(q^2)$ and $F_T^{D^{\ast}}(q^2)$ remain positive
in the entire $q^2$ region.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=5.8cm,height=5.0cm]{afbtauq2_smcentral_ds.png}
\includegraphics[width=5.8cm,height=5.0cm]{taptauq2_smcentral_ds.png}
\includegraphics[width=5.8cm,height=5.0cm]{cftauq2_smcentral_ds.png}
\includegraphics[width=5.8cm,height=5.0cm]{afbtq2_smcentral_ds.png}
\includegraphics[width=5.8cm,height=5.0cm]{flq2_smcentral_ds.png}
\includegraphics[width=5.8cm,height=5.0cm]{ftq2_smcentral_ds.png}
\end{center}
\caption{Forward backward asymmetry $A_{FB}^l(q^2)$, longitudinal polarization fraction $P^l(q^2)$, convexity parameter $C_F^l(q^2)$, forward
backward asymmetry for the transversely polarized $D^{\ast}$ meson $A_{FB}^T(q^2)$, longitudinal polarization fraction of $D^{\ast}$ meson
$F_L(q^2)$ and the transverse polarization of the $D^{\ast}$ meson $F_T(q^2)$ for the
$B_c \to D^{\ast}\,l\,\nu$ decays within the SM. Notations are same as in Fig.~\ref{dbr_r_ds}.}
\label{obs_ds}
\end{figure}
\section{Conclusion}
\label{con}
We have investigated semileptonic decays of $B_c$ meson to $D$ and $D^{\ast}$ meson with a charged lepton and a neutrino in the final state.
We use the effective Hamiltonian for $b \to u\,l\,\nu$ quark level transition decays and give predictions on various observables for
these decay modes within the SM. We give first prediction of various observables such as the ratio of branching ratios, lepton side forward
backward asymmetry, longitudinal polarization fraction of the charged lepton, convexity parameter, forward backward asymmetry of the
transversely polarized $D^{\ast}$ meson, longitudinal and transverse polarization fraction of the $D^{\ast}$ meson for these decays modes.
These results can be tested at future experiments and, in principle, can provide complimentary information regarding lepton flavor
universality violation observed in various $B$ meson decays.
\bigskip
|
\section{Introduction}
\label{sec:introduction}
Knowledge base curation, or the task of populating knowledge bases, is one of the main research challenges of crowdsourcing the Semantic Web~\cite{sarasua2015crowdsourcing}. Knowledge base curation can be done either manually, by asking annotators to populate the knowledge graph by manually extracting triples from unstructured data, or automatically by using information extraction methods that are trained and evaluated on ground truth collected from human annotators. In both cases, the process of gathering the human annotations is the a bottleneck in the entire knowledge base population process. The traditional approach to gathering human annotation is to employ experts to perform annotation tasks~\cite{welty2012query}, which is a costly and time consuming process. Additionally, in order to prevent high disagreement among expert annotators, strict annotation guidelines are designed for the experts to follow. On the one hand, creating such guidelines is a lengthy and tedious process, and on the other hand, the annotation task becomes rigid and not reproducible across domains. And, as a result, the entire process needs to be repeated over and over again in every domain and task. Finally, expert annotators are not always available for specific tasks such as open domain question-answering or news events, while many annotation tasks can require multiple interpretations that a single annotator cannot provide~\cite{aroyo2012harnessing}.
As a solution to those problems, crowdsourcing has become a mainstream approach. It has proved to provide good results in multiple domains: annotating cultural heritage prints~\cite{oosterman2014crowdsourcing}, medical relation annotation~\cite{aroyo2013measuring}, ontology evaluation~\cite{noy2013mechanical}. Following the central feature of volunteer-based crowdsourcing introduced by~\cite{von2009human} that majority voting and high inter-annotator agreement~\cite{Carletta1996} can ensure truthfulness of resulting annotations, most of those approaches are assessing the quality of their crowdsourced data based on the hypothesis~\cite{nowak2010reliable} that there is only one right answer to each question.
However, this assumption often creates issues in practice. Recent work in collecting annotations for text~\cite{poesio2005reliability,chang2016linguistic}, sounds~\cite{doi:10.1080/09298215.2016.1200631} and images~\cite{schaekermann2016,cheplygina2018crowd} found that disagreement between annotators is not just a result of poor quality work, and can actually be an indicator for other properties of the data, such as ambiguity and uncertainty~\cite{aroyo2018aimag}.
Previous experiments we performed~\cite{aroyo2013crowd} also identified issues with the assumption of the one truth: inter-annotator disagreement is usually never captured, either because the number of annotators is too small to capture the full diversity of opinion, or because the crowd data is aggregated with metrics that enforce consensus, such as majority vote. These practices create artificial data that is neither general nor reflects the ambiguity inherent in the data.
To address these issues, we proposed the {\bf CrowdTruth} methodology for crowdsourcing human annotation by harnessing inter-annotator disagreement, i.e representing the diversity of human interpretations in the ground truth. This is a novel approach for crowdsourcing human annotation that, instead of enforcing agreement between annotators, captures the ambiguity inherent in semantic annotation through the use of ambiguity-aware metrics for aggregating crowdsourcing responses. Based on this principle, we have implemented the CrowdTruth methodology as part of a framework~\cite{inel2014crowdtruth} for machine-human computation, that first introduced the ambiguity-aware metrics and built a pipeline to process crowdsourcing data with these metrics.
In this paper, we extend the definition of our ambiguity-aware methodology (CrowdTruth version 1.0~\cite{inel2014crowdtruth}) to work both with crowdsourcing tasks that are {\it closed}, i.e. the annotations that can occur in the data are already known, and the workers are asked to validate their existence (e.g. given a news event, decide whether it is expressed in a tweet), and tasks that are {\it open}, i.e. the annotation space is not known, and workers can freely select all the choices that apply (e.g. given a news piece, select all events that appear in the text). The code for the extended CrowdTruth version 1.1 methodology and metrics is available at: \url{https://git.io/fA3Mq}.
We investigate tasks of text and sound annotation, in both domains that typically require expertise from annotators (e.g. medical) and those that don't (open domain). In particular, we look at four crowdsourcing tasks: {\it Medical Relation Extraction}, {\it Twitter Event Identification}, {\it News Event Extraction} and {\it Sound Interpretation}. The aim is to investigate the role of inter-annotator disagreement as part of the crowdsourcing system by applying the CrowdTruth methodology to collect data over a set of diverse use cases.
Through the use of CrowdTruth aggregation metrics, the interpretations collected from the crowd are transformed into explicit semantics for the various tasks presented in this paper -- i.e. relations expressed in sentences, topics / events expressed in tweets and news articles, words describing sounds -- thus enabling knowledge base curation for these specific tasks. Furthermore, we prove that capturing disagreement is essential for acquiring high quality semantics. We achieve this by comparing the quality of the data aggregated with CrowdTruth metrics with majority vote, a method which enforces consensus among annotators. By applying our analysis over a set of diverse tasks we show that, even though ambiguity manifests differently depending on the task (e.g. each task has an optimal number of workers necessary to capture the full spectrum of opinions), our theory of inter-annotator disagreement as a property of ambiguity is generalizable for any semantic annotation crowdsourcing task.
The paper makes the following contributions:
\begin{enumerate}
\item {\bf comparative analysis of crowdsourcing aggregation methods:} we compare the performance of {\it ambiguity-aware metrics} and {\it consensus - enforcing metrics} over a diverse set of crowdsourcing tasks (Sections \ref{sec:results}, \ref{sec:discussion});
\item {\bf stability of crowd results:} we show in several crowdsourcing tasks that {\it an increased number of crowd workers leads to growth and stabilization in the quality of annotations}, going against the usual practice of employing a small number of annotators (Sections \ref{sec:results}, \ref{sec:discussion});
\item {\bf measuring quality in open-ended tasks:} we present an extension to the CrowdTruth methodology that allows the ambiguity-aware metrics to deal {\it both with open-ended and closed tasks} (Sections \ref{sec:methodology}, \ref{sec:experimental_setup}), as opposed to the initial version of the CrowdTruth metrics which only processed closed tasks;
\item {\bf semantics of ambiguity:} applying the CrowdTruth methodology we collect richer data that allows to reason about ambiguity of content (in all modality formats, e.g. images, videos and sounds), which is intrinsically relevant to the Semantic Web community.
\end{enumerate}
\section{CrowdTruth Methodology}
\label{sec:methodology}
In this section, we describe the CrowdTruth {\it methodology} version 1.1, for aggregating crowdsourcing data, which offers methods to aggregate both closed an open-ended tasks. Version 1.1 presented in this paper is a generalization of the initial version 1.0 of CrowdTruth~\cite{inel2014crowdtruth}.
In Section~\ref{sec:results} we use a number of annotation tasks in different domains to illustrate its use and gather experimental data to prove the main claim of this research - CrowdTruth methodology provides a viable alternative to traditional consensus-based majority vote crowdsourcing and expert-based ground truth collection. The elements of the CrowdTruth methodology are:
\begin{itemize}
\item annotation modeling with the \emph{triangle of disagreement};
\item quality \emph{metrics} for media units (input data), annotations and crowd workers;
\item identification of workers with low quality annotations.
\end{itemize}
Each of these elements is applicable across a variety of domains, content modalities, \emph{e.g.}, text, sounds, images and videos and annotation tasks, \emph{e.g.}, closed and open-ended annotations. The following sub-sections briefly introduce the overview of the methodology elements.
\subsection{CrowdTruth quality metrics}
\label{subsec:metrics}
Measuring quality in CrowdTruth is done with the triangle of disagreement model (based on the triangle reference \cite{knowlton1966definition}), which links together media units, workers, and annotations, as seen in Fig.\ref{fig:triangle_of_reference}. It allows us to assess the quality of each worker, the clarity of each media unit, and the ambiguity, similarity and frequency of each annotation. This model makes it possible to express how the ambiguity in any of the corners disseminates and influences the other components of the triangle. For example, an unclear sentence or an ambiguous annotation scheme would cause more disagreement between workers \cite{aroyo2014threesides}, and thus, both need to be accounted for when measuring the quality of the workers.
\begin{figure}[!hpt]
\centering
\includegraphics[width=0.8\linewidth]{img/triangle.png}
\caption{Triangle of Disagreement}
\label{fig:triangle_of_reference}
\end{figure}
The CrowdTruth quality metrics~\cite{aroyo2014threesides} are designed to capture inter-annotator disagreement in crowdsourcing. The metrics were introduced for {\it closed tasks}, i.e. multiple choice tasks, where the annotation set is known before running the crowdsourcing task. In this paper, we present an extended version of these metrics (version 1.1), that can be used for both {\it closed tasks} as well as {\it open-ended tasks} (i.e. the annotation set is not known beforehand, and the workers can freely select all the choices that apply). The code for the CrowdTruth version 1.1 metrics is available at: \url{https://git.io/fA3Mq}.
The quality of the crowdsourced data is measured using a \textbf{vector space representation} of the crowd annotations. For \emph{closed tasks}, the annotation vector contains the given answer options in the task template, which the crowd can choose from. For example, the template of a \emph{closed task} can be composed of a multiple choice question, which appears as a list checkboxes or radio buttons, thus, having a finite list of options to choose from.
While for closed tasks the number of elements in the annotation vector is known in advance, for \emph{open-ended tasks} the number of elements in the annotation vector can only be determined when all the judgments for a media unit have been gathered. An example of such a task can be highlighting words or word phrases in a sentence, or as an input text field where the workers can introduce keywords. In this case the answer space is composed of all the unique keywords from all the workers that solved that media unit. As a consequence, all the media units in a closed task have the same answer space, while for open-ended tasks the answer space is different across all the media units.
Although the answer space for open-ended tasks is not known from the beginning, it is still possible to deduce a finite answer space. To achieve this, we added an {\it answer space dimensionality reduction step} to the methodology for open-ended tasks. Additional goals of this step are to reduce redundancy in the answer space through similarity clustering (e.g. by making sure that synonymous words do not count as disagreement between annotators), and to keep the vector space representation small enough so that the CrowdTruth quality metrics still produce meaningful values. The method for performing dimensionality reduction is dependent on the annotation task itself.
In the annotation vector, each answer option is a boolean value, showing whether the worker annotated that answer or not. This allows the annotations of each worker on a given media unit to be aggregated, resulting in a \textbf{media unit vector} that represents for each option how often it was annotated.
Three core \textbf{worker metrics} are defined to differentiate between low-quality and high-quality workers. \emph{Worker-Worker Agreement} ($wwa$) measures the pairwise agreement between two workers across all media units they annotated in common - indicating how close a worker performs compared to workers solving the same task. \emph{Worker-Media Unit Agreement} ($wma$) measures the similarity between the annotations of a worker and the aggregated annotations of the rest of the workers. The average of this metric across all the media units solved gives a measure of how much a worker disagrees with the crowd in the context of all media units. \emph{Average annotations per media unit} ($na$) measures for each worker the total number of annotations they chose per media unit, averaged across all media units they annotated. Since in many tasks workers can choose all the possible annotations, a low quality worker can appear to agree more with the rest of the workers by repeatedly choosing multiple annotations, thus increasing the chance of overlap.
\begin{table*}
\caption {Consider an open-ended sound annotation task where 10 workers have to describe a given sound with keywords. The media unit for this task is a sound, the annotation set contains all the keywords workers provide for a sound. The table shows the media unit metrics, as well as the majority vote score for the media unit.}
\label{tab:example_annotation}
\begin{tabular}{ | r | c c c c c |}
\hline
{\bf worker annotations} & {\it dog barking} & {\it walking} & {\it animal} & {\it echo} & {\it loud} \\
{\bf media unit vector} & 3 & 2 & 5 & 1 & 1 \\
{\bf media unit -- annotation score} & 0.47 & 0.31 & 0.79 & 0.15 & 0.15 \\
{\bf majority vote} & 0 & 0 & 1 & 0 & 0 \\ \hline
\end{tabular}
\end{table*}
Two \textbf{media unit metrics} are defined to assess the quality of each unit. In this paper, we focus on the \emph{Media Unit-Annotation Score} -- the core CrowdTruth metric, used to measure the clarity with which the media unit expresses a given annotation. This metric is computed for each media unit and each possible annotation as the cosine between the media unit vector and the unit vector for each possible annotation. This metric is used in evaluating the quality of the CrowdTruth annotations.
\subsection{Spam Removal}
\label{subsec:spam_removal}
After collecting the crowd annotations, but before the evaluation of the data, we perform spam removal. The purpose of this step is to identify the adversarial and low quality workers -- e.g. those workers that always pick the same annotations, regardless of the unit. Once identified, the spam workers are removed from the dataset, and their annotations are not used in the evaluation. The methodology for spam removal is based on our previous work in~\cite{soberon2013}, extended in this paper to work also for open-ended tasks.
We identify the low quality workers by applying the core CrowdTruth worker metrics, the worker-worker agreement ($wwa$), worker-media unit agreement ($wma$) and the average number of annotations ($na$) submitted by a worker for one sentence. The first two metrics are used to model the extent to which a given worker agrees with the other annotators. The purpose is not to penalize disagreement with the majority, but rather to identify outliers, \emph{i.e.}, workers that are in constant disagreement. For \emph{closed tasks} where the semantics of the annotations in the answer space could rarely overlap, it is unlikely that a large number of possible annotations will occur for the same media unit. Therefore, the number of annotations per sentence can also indicate spam behavior.
In \emph{open-ended tasks} we apply the same approach. However, we need to acknowledge the fact that open-ended tasks are more prone to disagreement due to the large answer space and thus, the overall agreement between the workers can occur with lower values. Thus, we do not have predefined values for identifying the low-quality workers, but for every task or job we use the following main heuristic: given worker $w$, if the agreement $wwa(w)$, $wsa(w)$ and optionally, annotations per sentence $na(w)$, parameters do not fall within the standard deviation for the task, then worker $w$ is marked as a spammer. To confirm the validity of this metrics we also perform manual evaluation based on sampling of the results.
Based on the specificity of each task, closed or open-ended, the effort required to pick different annotations might vary. For instance, when no good annotation exists in the media unit, the time to complete the annotation is considerably reduced. This can bias the workers towards selecting the option that requires the least work. In order to prevent this, we introduce {\it in-task effort consistency checks}. Such annotations do not count towards building the ground truth, and are used to reduce the bias from picking the quickest option. For instance, when stating that no annotation is possible in the media unit, the workers also have to write an explanation in a text box for why no annotation were provided.
\begin{table*}[!htp]
\caption {Crowdsourcing Task Details}
\begin{tabular}{ | r | c c l |}
\hline
{\bf Task} & {\bf Type} & {\bf Media Unit} & {\bf Annotations} \\
\hline \hline
\multirow{3}{*}{Medical Relation Extraction} & \multirow{3}{*}{closed} & \multirow{3}{*}{sentence} & medical relations: {\it cause}, {\it treat}, {\it prevent}, {\it symptom}, {\it diagnose}, \\
& & & {\it side effect}, {\it location} {\it manifestation}, {\it contraindicate}, \\
& & & {\it is a}, {\it part of}, {\it associated with}, {\it other}, {\it none} \\ \hline
\multirow{5}{*}{Twitter Event Identification} & \multirow{5}{*}{closed} & \multirow{5}{*}{tweet} & tweet events: {\it Davos world economic forum 2014}, {\it FIFA World Cup 2014}, \\
& & & {\it Islands disputed between China and Japan}, {\it 2014 anti-China protests in Vietnam}, \\
& & & {\it Korean MV Sewol ferry ship sinking}, {\it Japan whaling and dolphin hunting}, \\
& & & {\it Disappearance of Malaysia Airlines flight 370}, {\it Ukraine crisis 2014}, \\
& & & {\it none of the above} \\ \hline
News Event Extraction & open-ended & sentence & words in the sentence \\ \hline
Sound Interpretation & open-ended & sound & tags describing sound \\ \hline
\end{tabular}
\label{tab:crowd_data}
\end{table*}
\begin{table*}[!htp]
\caption {Crowdsourcing Task Data}
\begin{tabular}{ | r | c c c c c | }
\hline
{\bf Task} & {\bf Source} & {\bf Expert annotation} & {\bf Media Units} & {\bf Workers / Unit} & {\bf Cost / Judgment} \\
\hline \hline
Medical Relation Extraction & PubMed article abstracts & yes & 975 & 15 & \$0.05 \\
Twitter Event Identification & Twitter (2014) & no & \np{3019} & 7 & \$0.02 \\
News Event Extraction & TimeBank & yes & \np{200} & 15 & \$0.02 \\
Sound Interpretation & Freesound.org & yes & \np{284} & 10 & \$0.01 \\ \hline
\end{tabular}
\label{tab:crowd_tasks}
\end{table*}
\section{Experimental Setup}
\label{sec:experimental_setup}
The aim of the crowdsourcing experiments described and analyzed in this paper is to show that the CrowdTruth ambiguity-aware crowdsourcing approach produces data with a higher quality than the traditional majority vote where consensus among annotators is enforced. In order to show this, we perform an experiment over a set of four diverse crowdsourcing tasks:
\begin{itemize}
\item two closed tasks, i.e. {\it Medical Relation Extraction}, {\it Twitter Event Identification},
\item two open-ended tasks, i.e. {\it News Event Extraction} and {\it Sound Interpretation}.
\end{itemize}
These tasks were picked from diverse domains (medical, sound, open), to aid in the generalization of our results. To evaluate the quality of the crowdsourcing data, we constructed a trusted judgments set by combining expert and crowd annotations. The rest of this section describes the details of the crowdsourcing tasks, trusted judgments acquisition process, as well as the evaluation methodology we employed.
\subsection{Crowdsourcing Overview}
\label{subsec:detasets}
\begin{figure*}[!tb]
\centering
\caption{Templates of the Crowdsourcing Tasks}
\label{fig:taskdesigns}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/relex}
\caption{Medical Relation Extraction}
\label{fig:screenshot_medical}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/tweets.png}
\caption{Twitter Event Identification}
\label{fig:screenshot_tweets}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/sound-task.png}
\caption{Sound Interpretation}
\label{fig:screenshot_sounds}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/news_event_extraction.png}
\caption{News Event Extraction}
\label{fig:screenshot_news}
\end{subfigure}
\end{figure*}
Tables \ref{tab:crowd_data} and \ref{tab:crowd_tasks} present an overview of the crowdsourcing tasks, as well as the datasets used. The results of the crowdsourcing tasks were processed with the use of CrowdTruth metrics (Sec.~\ref{subsec:metrics}), and we removed consistently low quality workers based on the spam removal procedure (Sec~\ref{subsec:spam_removal}). The tasks were implemented and ran on Figure Eight\footnote{\url{https://figure-eight.com/}} (formerly known as CrowdFlower). The templates are available on the CrowdTruth platform\footnote{tasks marked with $*$: \url{https://github.com/CrowdTruth/CrowdTruth/wiki/Templates}}.
The payment per judgment was determined through a series of pilot runs of the tasks where we started with a \$0.01 cost per judgment, and then gradually increased the payment until a majority of Figure Eight workers rated our tasks as having fair payments. As a result, we were able to get a constant stream of workers to participate in the tasks. The values shown in Table~\ref{tab:crowd_tasks} show the final cost per judgment we reached after the pilot runs. Since crowd pay has a complex effect on the quality of the annotation~\cite{mao2013volunteering}, and in order to remove confounding factors, judgments collected with costs lower than those in Table~\ref{tab:crowd_tasks} were left out of this evaluation. In total, it took two months to perform the pilot runs and then collect the judgments for all of the tasks.
The number of workers per media unit was determined experimentally with the goal of capturing all possible results from the crowd and stabilizing the quality of the annotations; this process is explained at length further on in Section~\ref{sec:results}, with the results of the experiment shown in Figure~\ref{fig:f1_workers}.
The {\bf Medical Relation Extraction dataset} consists of 975 sentences extracted from PubMed\footnote{\url{http://www.ncbi.nlm.nih.gov/pubmed}} article abstracts. The sentences were collected using distant supervision~\cite{mintz2009distant}, a method that picks positive sentences from a corpus based on whether known arguments of the seed relation appear together in the sentence (\emph{e.g.}, the $treat$ relation occurs between the terms $antibiotics$ and $typhus$, so find all sentences containing both and repeat this for all pairs of arguments that hold). The MetaMap parser~\cite{aronson2001effective} was used to extract medical terms from the corpus and the UMLS vocabulary~\cite{bodenreider2004unified} was used for mapping terms to categories, and relations to term types. The intuition of distant supervision is that since we know the terms are related, and they are in the same sentence, it is more likely that the sentence expresses a relation between them (than just any random sentence). We started with a set of 8 UMLS relations important for clinical decision making~\cite{P14-1078}, that became the seed in distant supervision, but this paper only discusses results for the relations $cause$ and $treat$, as these were the only relations for which we could also collect expert annotations. The expert judgment collection is detailed in Section~\ref{subsec:expert_data}.
The \emph{medical relation extraction task} (see Figure \ref{fig:screenshot_medical}) is a {\it closed task}. The crowd is given a medical sentence with the two highlighted terms collected with distant supervision, and is then asked to select from a list all relations that are expressed between the two terms in the sentence. The relation list contains eight UMLS\footnote{https://www.nlm.nih.gov/research/umls/} relations, as well as {\it is a}, {\it part of}, {\it associated with}, {\it other}, {\it none} relations, added to make the choice list complete. Multiple choices are allowed in this task. To reduce the bias of selecting $none$, we also added an in-task effort consistency check by asking workers to explain in a text box why no relation is possible between the terms. The task results are processed into an annotation vector containing a component for each of the relations. A detailed description of the crowdsourcing data collection is given in \cite{DBLP:journals/corr/DumitracheAW17}.
The {\bf Twitter Event Identification dataset} consists of 3,019 English tweets from 2014, crawled from Twitter. The tweets are selected as been relevant to eight events, such as, ``Japan whale hunt'', ``China Vietnam relation'' among other controversial events. The dataset was created by querying a Twitter dataset from 2014 with relevant phrases for each of the eight events, \emph{e.g.}, ``Whaling Hunting'', ``Anti-Chinese in Vietnam''. The \emph{Twitter event identification task} (see Figure \ref{fig:screenshot_tweets}) is a {\it closed task}. The crowd is asked to choose for each tweet the relevant events out of the list of eight, as well as to highlight for each of the relevant events the event mentions in the tweet. The crowd could also pick that none of the events was present in the tweet. Multiple choices of events were permitted. Since tweets and tweet annotations typically are not done by experts, we did not collect expert data for this task. To reduce the bias of selecting no event, we also added an in-task effort consistency check by asking workers to explain in a text box why none of the events is present in the tweet. The task results are processed into an annotation vector containing a component for each of the events.
The {\bf News Event Extraction dataset} consists of 200 randomly selected English sentences from the English TimeBank corpora~\cite{pustejovsky2003timebank}, which were also presented in \cite{CASELLI16.966}. The \emph{news event extraction} (see Figure \ref{fig:screenshot_news}) is an {\it open-ended task}. The crowd receives an English sentence, and is asked to highlight words or word phrases (multiple words) that describe an event or a time expression. For each sentence, the crowd is allowed to highlight a maximum of 30 event expressions or time expressions. For the purpose of this research we only focus on evaluating the extraction of event expressions. We define an \emph{event} as something that happened, is happening, will or happen. On this dataset we employed expert annotators as described in Section \ref{subsec:expert_data}. To reduce the bias of selecting fewer events than actually expressed in the task, we implemented an in-task effort consistency check by asking workers that annotated 3 events or less to explain in a text box why no other events are expressed in the sentence. As part of the {\it answer set dimensionality reduction step}, we removed the stop words from the sentence (we consider that the stop words are not meaningful for our analysis and they could add unsubstantial disagreement), and split the expressions collected from the crowd into words. The annotation vector is composed of the words in the sentence, where a word is selected in the worker vector if it appears in at least one of the expressions identified by the worker.
The {\bf Sound Interpretation dataset} consists of 284 unique sounds sampled from the Freesound\footnote{\url{https://www.freesound.org/}} online database. All these recordings and their metadata are freely accessible through the Freesound API\footnote{\url{https://www.freesound.org/docs/api/}}. We focused on SoundFX sounds, \emph{i.e.}, sound effects category, as classified by \cite{font2014audio}. The \emph{Sound interpretation task} (see Figure \ref{fig:screenshot_sounds}) is an {\it open-ended task}, where the crowd is asked to listen to three sounds and provide for each sound a comma separated list of keywords that best describe what they heard. For each sound, any number of answers is possible. In the {\it answer set dimensionality reduction step}, the annotated keywords were clustered syntacticly using spell checking and stemming, and semantically using a word2vec model~\cite{mikolov2013distributed} pre-trained on the Google News corpus. The annotation vector contains a component for each of the keywords used to describe the sound, after clustering. A detailed description of the crowdsourcing data collection and processing is given in \cite{miltenburg2016}. For this dataset we also collected expert annotations from the sound creators as described in Section \ref{subsec:expert_data}.
\subsection{Evaluation Methodology}
\label{subsec:evaluation}
The purpose of the evaluation is to determine the quality of the annotations generated with CrowdTruth ambiguity-aware aggregating metrics. To this end, we label each media unit and annotation pair with its media unit-annotation score (see Section \ref{subsec:metrics}), and compare it with three other methods for labeling the data, as described below:
\begin{itemize}
\item \textbf{Majority vote}: Each media unit-annotation pair receives either a positive or a negative label, according to the decision of the majority of crowd workers. For each annotation performed by a crowd worker over a given media unit, we calculate the ratio of workers that have selected this annotation over the total number of workers that have annotated the unit, and assess whether it is greater or equal to 0.5. This allows for multiple annotations to be picked for one media unit. For some units, however, none of the annotations were picked by half or more of the workers. This is especially the case for open-ended tasks, such as sound interpretation, where workers put in a large number of annotations, and agreement is seldom. In these situations, we picked the annotations that were selected by the most workers (even if they do not constitute more than half). Judgments from workers labeled as spammers were not employed in the aggregation. An example of the majority vote aggregation is shown in Table~\ref{tab:example_annotation}.
\item \textbf{Single}: Each media unit-annotation pair receives either a positive or a negative label, according to the decision of a single crowd worker. For every media unit, this score was randomly sampled
from the set of workers annotating it. Judgments from workers labeled as spammers were not employed. While a single annotator is not used as often as the majority vote in traditional crowdsourcing, we use this dataset as a baseline for the crowd, to show that having more annotators generates better quality data.
\item \textbf{Expert}: Each media unit-annotation pair receives either a positive or a negative label, according to the expert decision. The details of how expert data was collected for each tasks are discussed in Section \ref{subsec:expert_data}.
\end{itemize}
The \emph{evaluation of the quality of the CrowdTruth method} was done by computing the micro-F1 score over each task. The micro-F1 score was used in order to treat each case equally, without giving advantage to annotations that appear less frequently in our datasets. Using the trusted judgments collected according to Section~\ref{subsec:expert_data}, we evaluate each media unit -- annotation pair as either a true positive, false positive etc. We compute the value of the micro-F1 score using the following formulas for the micro precision (Equation \ref{eq:precision}) and micro recall (Equation \ref{eq:recall}):
\begin{equation}
P_{micro} = \frac{\sum_{i=1}^{n}{TP_i}}{\sum_{i=1}^{n}{TP_i} + \sum_{i=1}^{n}{FP_i}}
\label{eq:precision}
\end{equation}
\begin{equation}
R_{micro} = \frac{\sum_{i=1}^{n}{TP_i}}{\sum_{i=1}^{n}{TP_i} + \sum_{i=1}^{n}{FN_i}}
\label{eq:recall}
\end{equation}
where $TP_i$, $FP_i$, $FN_i$, with $i$ from 1 to $n$ (the number of media units in the dataset), represent the number of true positive, false positive and false negative annotations for media unit $i$. Finally, the micro-F1 score is computed as the harmonic mean of the micro-precision and micro-recall.
An important variable in the evaluation is the {\it media unit-annotation score threshold} for differentiating between a negative and a positive classification. Traditional crowdsourcing aims at reducing disagreement, and therefore corresponds to high values for this threshold. Lower values means accepting more disagreement in the classification of positive answers by the crowd. In our experiments, we tried a range of threshold values for each task, to investigate with which one we achieve the best results. The media unit-annotation score threshold was also used in gathering the set of trusted judgments for the evaluation (Section~\ref{subsec:expert_data}). All the data used in this paper can be found in our data repository\footnote{\url{https://github.com/CrowdTruth/Cross-Task-Majority-Vote-Eval}}.
\subsection{Trusted Judgments Collection}
\label{subsec:expert_data}
To perform the evaluation, a set of trusted judgments is necessary to assess the correctness of crowd annotations. For each dataset, we manually evaluated the correctness of all the media unit annotations that were generated by the crowd and the experts. Depending on the task, the number of media unit-annotation pairs can become quite high, so we explored methods to make the manual evaluation more efficient.
For the datasets that contain expert annotation, we calculated the thresholds which yielded the maximum agreement in number of annotations between the crowd and expert annotations. These annotations were then added to the trusted judgments collection, as the judgment in this case is unambiguous. The interesting cases appear when crowd and expert disagree. Previous work we performed in crowdsourcing {\it Medical Relation Extraction}~\cite{aroyo2015truth} has indicated that experts might not always provide better annotations than crowd workers. Additionally, for the {\it Sound Interpretation} task we noticed that experts provided considerably fewer tags than the crowd, and there was a large discrepancy between annotations of crowds and experts, with a very small overlap between their annotations. Therefore, instead of simply relying on expert judgment, the annotations where crowd and expert disagree were manually relabeled by exactly one of the authors, and then added to the trusted judgments set, which is also published in our data repository. In Appendix~\ref{sec:appendix} we present a selection of examples where the expert judgment is different from the trusted judgment. While these cases might call into question the level of expertise of the domain experts, inconsistencies and disagreement in expert annotation are regularly reported in various annotation tasks~\cite{cheatham2014conference,mcdonnell2016relevant,INEL16.635}. Furthermore, in Section~\ref{sec:results} we will show that using the trusted judgments for evaluation still results in the expert performing the best for 2 out of 3 tasks. The only task where the expert underperforms is {\it Sound Interpretation}, where the set of annotations provided by the expert is much smaller than the one provided by the crowd.
\begin{figure*}[!b]
\centering
\caption{CrowdTruth F1 scores for all crowdsourcing tasks.}
\label{fig:f1_mv}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/medical_mv.png}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/tweets_mv_2.png}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/events_mv.png}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/sounds_mv_2.png}
\end{subfigure}
\end{figure*}
We collected expert annotations for the {\it Medical Relation Extraction} data by employing medical students. Each sentence was annotated by exactly one person. The annotation task consisted of deciding whether or not the UMLS seed relation discovered by distant supervision is present in the sentence for the two selected terms.
For the {\it Sound Interpretation} task, each sound in the dataset contains a description and a set of keywords that were provided by the authors of the sounds. We consider the keywords provided by the sounds' authors as trusted judgments given by domain experts.
The \emph{news event extraction} data was annotated with events by various linguistic experts. In total, 5 people annotated each sentence but we only have access to the final annotations, a consensus among the annotators. In the annotation guidelines described in \cite{pustejovsky2003timebank}, events are defined as situations that happen or occur, but are not generic situations. In contrast to the crowdsourcing task, where the workers had very loose instructions, the experts had very strict rules for identifying events, strictly based on linguistic features: \emph{(i)} tensed verbs: has called, will leave, was captured, \emph{(ii)} stative adjectives: sunken, stalled, on board and \emph{(iii)} event nominals: merger, Military Operation, Gulf War.
The only task without expert annotation is {\it Twitter Event Identification} -- as it is in the open domain, no experts exist for this type of data.
\section{Results}
\label{sec:results}
We begin by evaluating {\bf how the majority vote method compares with CrowdTruth}, by calculating the precision/recall metrics using the gold standards we collected for each of the four crowdsourcing tasks. Figure~\ref{fig:f1_mv} shows the F1 score for CrowdTruth over the four tasks. The results are calculated for different media unit-annotation score thresholds for separating the data points into positive and negative classifications. Table~\ref{tab:f1_mv} shows the detailed scores for CrowdTruth, given the highest F1 media unit-annotation score threshold.
\begin{table*}[!tb]
\centering
\caption {CrowdTruth evaluation results, given the highest F1 media unit-annotation score threshold.}
\label{tab:f1_mv}
\begin{tabular}{|lr|ccccl|}
\hline
{\bf Task} & {\bf Dataset} & {\bf Precision} & {\bf Recall} & {\bf F1 score} & {\bf Accuracy} & {\bf media unit-annotation score threshold} \\ \hline \hline
\multirow{4}{*}{\parbox{1.8cm}{{\bf Medical \\ Relation \\ Extraction}}} & CrowdTruth & 0.86 & 0.962 & 0.908 & 0.932 & 0.6 \\
& expert & 0.899 & 0.89 & 0.895 & 0.927 & \\
& majority vote & 0.924 & 0.781 & 0.847 & 0.902 & \\
& single & 0.222 & 0.776 & 0.346 & 0.748 & \\ \hline
\multirow{3}{*}{\parbox{1.8cm}{{\bf Twitter \\ Event \\ Identification}}} & CrowdTruth & 0.965 & 0.945 & 0.955 & 0.995 & 0.4 \\
& majority vote & 0.984 & 0.885 & 0.932 & 0.984 & \\
& single & 0.959 & 0.819 & 0.884 & 0.972 & \\ \hline
\multirow{4}{*}{\parbox{1.8cm}{{\bf News \\ Event \\ Extraction}}} & CrowdTruth & 0.984 & 0.929 & 0.956 & 0.931 & 0.05 \\
& expert & 0.983 & 0.944 & 0.963 & 0.942 & \\
& majority vote & 0.985 & 0.375 & 0.544 & 0.492 & \\
& single & 0.99 & 0.384 & 0.554 & 0.501 & \\ \hline
\multirow{4}{*}{\parbox{1.8cm}{{\bf Sound \\ Interpretation}}} & CrowdTruth & 1 & 0.729 & 0.843 & 0.815 & 0.1 \\
& expert & 1 & 0.291 & 0.45 & 0.515 & \\
& majority vote & 1 & 0.148 & 0.258 & 0.418 & \\
& single & 1 & 0.098 & 0.178 & 0.383 & \\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{$p$-values for McNemar's test of statistical significance in the CrowdTruth classification, compared with the others.}
\label{tab:stat_sig}
\centering
\begin{tabular}{|r|ccc|}
\hline
{\bf Task} & {\bf Maj. Vote} & {\bf Expert} & {\bf Single} \\ \hline \hline
Medical Relation Extraction & $0.0001$ & $0.629$ & $< 2.2 \times 10^{-16}$ \\ \hline
Twitter Event Identification & $0.0001$ & N/A & $6.145 \times 10^{-15}$ \\ \hline
News Event Extraction & $< 2.2 \times 10^{-16}$ & $0.505$ & $< 2.2 \times 10^{-16}$ \\ \hline
Sound Interpretation & $< 2.2 \times 10^{-16}$ & $< 2.2 \times 10^{-16}$ & $< 2.2 \times 10^{-16}$ \\ \hline
\end{tabular}
\end{table*}
Across all four tasks, the CrowdTruth method performs better than both majority vote and the single annotator dataset. While majority vote unsurprisingly performs the best on precision, as a consequence of its lower rate of positive labels, CrowdTruth consistently scores the best for both recall, F1 score and accuracy. These differences in classification are statistically significant, as shown in Table~\ref{tab:stat_sig} -- this was calculated using McNemar's test~\cite{mcnemar1947note} over paired nominal data.
The evaluation of CrowdTruth compared with the expert is more nuanced. For the {\it Medical Relation Extraction} and {\it news event extraction tasks}, CrowdTruth performs as well as the expert annotators, with p-values indicating there is no statistically significant difference in the classifications. In contrast, for the task of {\it Sound Interpretation}, CrowdTruth performs better than the expert by a large margin.
\begin{figure*}[!tb]
\centering
\caption{The effect of the number of workers per unit on the F1 score, calculated at the best media unit-annotation score threshold (Table~\ref{tab:f1_mv}). For every point, the F1 is calculated with at most the given number of workers. The number of units used in the calculation of the F1 is shown in the y-axis on the right.}
\label{fig:f1_workers}
\centering
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/medical_workf1_2.png}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/tweets_workf1_2.png}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/events_workf1_2.png}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/sounds_workf1_2.png}
\end{subfigure}
\end{figure*}
The second evaluation shows the {\bf influence of the number of workers on the quality of the CrowdTruth data}. Figure~\ref{fig:f1_workers} shows the CrowdTruth F1 score in relation to the number of workers. Given one task, the number of workers per unit varies because of spam removal, so the F1 score was calculated using at most the number of workers at every point in the graph. The number of units annotated with the given number of workers is also shown in the graph.
The effects of the number of workers on the CrowdTruth F1 is clear -- more workers invariably leads to a higher F1 score. For the tasks of {\it Medical Relation Extraction}, {\it Twitter Event Identification} and {\it News Event Extraction}, the CrowdTruth F1 grows into a straight line, showing that the opinions of the crowd stabilize after enough workers. For the {\it Sound Interpretation} task, the CrowdTruth F1 score is still on an upwards trend after 10 workers, possibly indicating that more workers are necessary to get the full spectrum of annotations.
Figure~\ref{fig:f1_workers} also shows that CrowdTruth performs better than majority vote regardless of the number of workers per task. For closed tasks, increasing the number of workers has a positive impact on the majority vote F1 score. For open tasks, adding more workers has less of an effect -- more workers increase the size of the annotation set for a unit, which is typically larger than for closed tasks, but the agreement is low because opinions are split between possible annotations.
\begin{figure*}[!tbh]
\centering
\caption{CrowdTruth F1 score evaluation, using expert annotation as ground truth.}
\label{fig:f1_exp}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/exp_eval_relex.png}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/exp_eval_events.png}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\includegraphics[width=\linewidth]{img/exp_eval_sounds.png}
\end{subfigure}
\end{figure*}
Finally, Figure~\ref{fig:f1_exp} shows an evaluation of CrowdTruth using only the expert annotations as ground truth (the {\it Twitter Event Identification} task does not have experts, so it could not be evaluated). The F1 scores are lower than in the evaluation over the trusted judgments collection. For the {\it Medical Relation Extraction Task}, majority vote performs essentially the same as CrowdTruth, whereas for the open-ended tasks, CrowdTruth still performs better. However, as we have shown in Appendix~\ref{sec:appendix}, the expert annotations contain errors and are sometimes incomplete, particularly in the case of open-ended tasks. The evaluation using expert ground truth was done to show that the trusted judgments set is not biased in favor of CrowdTruth.
\section{Discussion}
\label{sec:discussion}
The first goal in this paper was to show that the {\bf ambiguity-aware CrowdTruth approach with multiple annotators and disagreement-based quality scores can perform better than majority vote}, a method that enforces consensus among annotators. Our results over several crowdsourcing tasks, as seen in Figure~\ref{fig:f1_mv}, show this clearly.
The gap in performance between CrowdTruth and majority vote is the most striking for open tasks ({\it News Event Extraction} and {\it Sound Interpretation}). These tasks also require the lowest agreement threshold for achieving the best performance with CrowdTruth. During the trusted judgments collection process, we observed how these tasks are prone to a wide range of opinions -- for instance, in the case of {\it Sound Interpretation}, there are frequent examples of labels that are semantically dissimilar, but could reasonably be applied to the same sound (e.g. the same sound was annotated with the tag {\tt balloon popping} by one worker, and with {\tt gunshot} by another worker). Because of this, enforcing consensus does not work for these tasks, and ambiguity-aware annotation aggregation appeared to be a viable solution.
Our evaluation also shows that processing crowd data with ambiguity-aware metrics performs at least as well as expert annotators, which is not the case for majority vote. Crowdsourcing annotation is significantly cheaper in cost than experts -- e.g. even with 15 workers per unit, crowdsourcing for the task of {\it Medical Relation Extraction} cost 2/3 of what the experts did. The crowd also has the advantage of being readily available on platforms such as Figure Eight, while the process of finding and hiring expert annotators can incur significant time costs. As our results showed, in order for the crowdsourcing to produce results comparable in quality to that of experts, appropriate processing with ambiguity-aware metrics is a necessity.
The variation in the optimal media unit-annotation score thresholds across the tasks shows that the level of ambiguity is dependent on the crowdsourcing task, thus supporting our triangle of disagreement model (Section~\ref{subsec:metrics}). It is not surprising that the task with the highest agreement threshold ({\it Medical Relation Extraction}) also has the most exact definition of a correct answer (i.e. whether a medical relation is expressed or not in a given sentence). The definition of a medical relation is fairly clear; in contrast, the definition of an event is more subjective, therefore workers were able to come up with a wider range of correct annotations.
The experimental setup provides an empirical method for selecting the optimal threshold for media unit-annotation score. However, if performing an evaluation with trusted judgments is not possible, selecting the optimal threshold becomes more difficult. For open-ended tasks, the experiments indicate that almost all opinions matter, and the agreement threshold should be as low as possible. In these cases, spam workers can be successfully eliminated by in-task effort consistency checks, and there is no need to enforce agreement beyond that. In contrast, the experiments for closed tasks show higher agreement thresholds tend to work better. The difficulty as well as the subjectivity of the domain also appear to have an impact. The threshold should grow together with the difficulty, and inversely with subjectivity. However, both difficulty and subjectivity might be difficult to measure in practice. In the end, the tuning of the threshold should be regarded similarly to a precision-recall trade-off analysis, where the optimal value depends on the requirements of the ground truth (high precision but many false negative crowd labels, or high recall but more false positives). The high variability for optimal threshold values also shows the limitations of traditional evaluation metrics like precision and recall that rely on discrete labels. CrowdTruth metrics were constructed to measure ambiguity on a continuous scale, but the use of standard metrics resulted in losing this information by forcing the conversion to either positive or negative. Ultimately, our goal is to move away from a binary ground truth that needs to be calculated using a fixed threshold, and instead to use the CrowdTruth metrics to express ambiguity on a continuous scale.
The second goal of the experiment was to show {\bf the effect of the number of workers on the quality of CrowdTruth annotations}. The results in Figure~\ref{fig:f1_workers} clearly show the increase in F1 score for CrowdTruth as more workers contribute to the tasks. This combined with the poor performance of the single annotator dataset proves the importance in considering a large enough pool of workers to be able to accurately capture the full spectrum of opinions.
The stabilization of the F1 score for {\it Medical Relation Extraction}, {\it Twitter Event Identification} and {\it News Event Extraction} is an indication that we have indeed managed to collect the entire set of opinions for these tasks. The fact that the scores all stabilize at different points in the graph (around 8 workers for {\it Medical Relation Extraction}, 5 for {\it Twitter Event Identification}, and 10 for {\it News Event Extraction}) indicates that the optimal number of workers is dependent on the task type, thus also confirming our hypothesis that more workers than what is typically being considered in crowdsourcing studies are necessary for acquiring a high quality ground truth.
There exists a trade-off between cost and quality of annotations that should also be considered when optimizing the number of workers. The higher cost was justified for these tasks, as the expert annotation was three times more expensive than the crowdsourced annotations at expert quality level.
An interesting observation is that the optimal number of workers per task does not seem to influence the optimal media unit-annotation score threshold for the task. The {\it News Event Extraction} requires a high number of workers, but the optimal media unit-annotation score threshold is low, while the {\it Twitter Event Identification} requires a low number of workers, and also a low media unit-annotation score threshold, at least compared to {\it Medical Relation Extraction}.
While four tasks is a small sample to draw conclusions from, our findings seem to indicate that ambiguity in the crowdsourcing system has an impact on both the optimal number of workers per task, as well as the clarity of the media units. These observations will form the basis for our future research in modeling crowd disagreement.
Finally, it is worth discussing the outlier characteristics of the {\it Sound Interpretation} task. It is the only task that does not achieve a stable F1 curve (Figure~\ref{fig:f1_workers}) possibly due to insufficient workers assigned to it. It is also unique in its lack of false positive examples -- precision is 1 for the optimal media unit-annotation score threshold (Table~\ref{tab:f1_mv}), meaning that all labels collected from the crowd were accepted as part of the trusted judgments, with the exception of the spam workers that were removed from the set. {\it Sound Interpretation} is also the only task for which the expert annotator performed comparatively poor, with a statistically significant difference from CrowdTruth. As mentioned in the beginning of this section, after collecting the trusted judgments for this task, it became clear that the main challenge for the {\it Sound Interpretation} task is not to achieve consensus between annotators, but to collect the entire spectrum of annotations that describe a sound, given that this spectrum is so large (e.g. the tags {\tt balloon popping} and {\tt gunshot} can both reasonably apply to the same sound). For this reason, it was difficult to label tags as false positives, and the annotations of the workers, experts included, were largely non-overlapping, as they tended to interpret the sounds quite differently. The {\it Sound Interpretation} task is therefore an extreme example of subjective ground truth.
\section{Related Work}
\label{sec:relatedwork}
\subsection{Crowdsourcing Ground Truth}
Crowdsourcing has grown into a viable alternative to expert ground truth collection, as crowdsourcing tends to be both cheaper and more readily available than domain experts. Experiments have been carried out in a variety of tasks and domains: medical entity extraction~\cite{zhai2013web,Finin2010,van2012eu}, medical relation extraction~\cite{kilicoglu2011constructing,van2012eu}, open-domain relation extraction~\cite{kondreddi2014combining}, clustering and disambiguation~\cite{Lee2013}, ontology evaluation~\cite{noy2013mechanical}, web resource classification~\cite{castano2016human} and taxonomy creation~\cite{bragg2013crowdsourcing}. \cite{Snow2008} have shown that aggregating the answers of an increasing number of unskilled crowd workers with majority vote can lead to high quality NLP training data. The typical approach in these works is to assume the existence of a universal ground truth. Therefore, disagreement between annotators is considered an undesirable feature, and is usually discarded by using either of the following methods: restricting annotator guidelines, picking one answer that reflects some consensus usually through majority voting, or using a small number of annotators.
\subsection{Disagreement and Ambiguity in Crowdsourcing}
Besides CrowdTruth, there exists some research on how disagreement in crowdsourcing should be interpreted and handled. In assessing the OAEI benchmark, \cite{cheatham2014conference} found that disagreement between annotators (both crowd and expert) is an indicator for inherent uncertainty in the domain knowledge, and that current benchmarks in ontology alignment and evaluation are not designed to model this uncertainty. \cite{plank-hovy-sogaard:2014:P14-2} found similar results for the task of crowdsourced part-of-speech tagging -- most inter-annotator disagreement was indicative of debatable cases in linguistic theory, rather than faulty annotation. \cite{Bayerl2011} also investigate the role of inter-annotator disagreement as a possible indicator of ambiguity inherent in natural language. \cite{lau2014measuring} propose a method for crowdsourcing ambiguity in the grammatical correctness of text by giving workers the possibility to pick various degrees of correctness, but inter-annotator disagreement is not discussed as a factor in measuring this ambiguity. \cite{schaekermann2016} propose a framework for dealing with uncertainty in ground truth that acknowledges the notion of ambiguity, and uses disagreement in crowdsourcing for modeling this ambiguity. For the task of word sense disambiguation, \cite{jurgens2013embracing} show that, in modeling ambiguity, the crowd was able to achieve expert-level quality of annotations. \cite{Chang:2017:Revolt} implemented a workflow of tasks for collecting and correcting labels for text and images, and found that ambiguous cases cannot simply be resolved by better annotation guidelines or through worker quality control. Finally, \cite{lin2014re} shows that often, machine learning classifiers can achieve a higher accuracy when trained with noisy crowdsourcing data. To our knowledge, our paper presents the first experiment across several tasks and domains that explores ambiguity as a property of crowdsourcing systems, and how it can be interpreted to improve the quality of ground truth data.
\subsection{Crowdsourcing Aggregation beyond Majority Vote}
The literature on alternative crowdsourcing aggregation metrics typically focuses on analyzing worker performance -- identifying spam workers~\cite{Bozzon:2013,Kittur2008,Ipeirotis:2010}, and analyzing workers' performance for quality control and optimization of the crowdsourcing processes~\cite{Singer:2013}. \cite{NIPS2009_3644} and \cite{welinder2010multidimensional} have used a latent variable model for task difficulty, as well as latent variables to measure the skill of each annotator, to optimize crowdsourcing for image labels. \cite{werling2015job} use on-the-job learning with Bayesian decision theory to assign the most appropriate workers for each task, for both text and image annotation. Finally, \cite{prelec2017solution} show that the surprisingly popular crowd choice (i.e. the answer that most workers thought would not be picked by other workers, even though it is correct) gave better results than the majority vote for a variety of tasks with unambiguous ground truths (state capitals, trivia questions and price of artworks).
All of these approaches show promising improvements over the use of majority vote as an aggregating method. These methods were developed only for closed tasks, primarily dealing with classification. However, the novel approach of CrowdTruth allows to explore both closed and open-ended tasks. Furthermore, our focus is on modeling ambiguity as a latent variable in the crowdsourcing system, as well as its role in generating inter-annotator disagreement, which these approaches currently do not take into account. We believe an optimal crowdsourcing approach would combine both ambiguity modeling, as well as specialized task assignment to workers. For instance, \cite{felt2015early} developed a generative model to aggregate crowd scores that incorporates features of the data (e.g. number of words), although they do not evaluate the performance of specific features. Ambiguity as measured with CrowdTruth, like the media unit-annotation score, could be used as a data feature in such a system.
\section{Conclusions}
\label{sec:conclusions}
Gathering human annotation is a major bottleneck in the process of knowledge base curation. Crowdsourcing-based approaches are gaining popularity in the attempt to solve the issues related to volume of data and lack of annotators. Typically these practices use inter-annotator agreement as a measure of quality. However, by ignoring inter-annotator disagreement, these practices tend to create artificial data that is neither general nor reflects the ambiguity inherent in the source.
In this paper we presented an empirically derived methodology for efficiently gathering of human annotation by aggregating crowdsourcing data with CrowdTruth metrics, which harness the inter-annotator disagreement. We applied this methodology over a set of diverse crowdsourcing tasks: closed tasks ({\it Medical Relation Extraction}, {\it Twitter Event Identification}), and open-ended tasks ({\it News Event Extraction} and {\it Sound Interpretation}). Our results showed that the ambiguity-aware CrowdTruth approach allows us to collect richer data, which enables reasoning about the ambiguity of the content being annotated. This is intrinsically relevant to the Semantic Web community, i.e. to identify the semantics of ambiguity across all modalities, e.g. text, images, videos and sounds. Our results also showed that, in all the tasks we considered, such ambiguity-aware quality scores provide better ground truth data than the traditional majority vote. Moreover, we have shown that CrowdTruth annotations have at least the same quality, even better in the case of {\it Sound Interpretation}, as expert annotations. Finally, we showed that, contrary to the common crowdsourcing practice of employing a small number of annotators, adding more crowd workers actually can lead to significantly better annotation quality.
In the future, we plan to expand our methodology to more complex annotation tasks, that require multiple or combined types of input beyond the closed/open-ended categorization we presented in this paper. We are also working on expanding the CrowdTruth metrics for ambiguity to incorporate the state-of-the art in modeling crowd worker and data features~\cite{felt2015early}. Finally, we want to use the CrowdTruth data in practice for training and evaluating information extraction models used to populate the Semantic Web.
\section*{Acknowledgements}
We would like to thank Emiel van Miltenburg for assisting with the exploration of feature analysis of sounds, Chang Wang and Anthony Levas for providing and assisting with the medical data, Zhaochun Ren for the help in gathering the Twitter dataset, Tommaso Caselli for providing the news dataset, and the anonymous crowd workers for their contributions to our crowdsourcing tasks.
|
\section{Introduction}
In recent years, robots have become part of human living space and have been expected to perform various tasks in complex environments. If robots could use tools as humans do, they could improve in versatility, overcome some physical limitations and adapt to the environment. Therefore, research on tool-use by robots has aimed at robots that are useful for daily life.\\
Piaget \cite{Piaget} proposed that infants learn tool-use as part of the cognitive developmental process as detailed in \cite{b1}, \cite{b2}. In the first to third substages of Piaget's sensorimotor developmental stage (0--8 months), infants come to understand motor behavior and learn to couple perception and action. Next, in the fourth substage (8--12 months), they learn the causal relationship between actions and effects. In other words, they become able to understand what results from their actions. Then, in the fifth substage (12--18 months), they acquire the features of tools and objects by trial-and-error of object manipulation with tools. The features are unique to each combination and can be inferred from the causality between the actions and the effects. Finally, in the sixth (and final) substage (18--24 months), they become able to make plans for tool-use, even without conducting actions. They can understand how they should act and what they should use, understanding this from the effects. In other words, they become able to detect proper features of tools, objects, and actions from observed effects.\\
Despite this development process being known, most conventional research on robot tool-use has focused on either tools or objects. Nishide et al.~\cite{b3} and Montesano et al.~\cite{b4} enabled robots to learn the relations between actions and object movements. Along another line, Stoytchev~\cite{b5}, Nishide et al.~\cite{Nishide}, Mar et al.~\cite{Mar}, and Takahashi et al.~\cite{b6} enabled robots to acquire tool function for a specific target object. In each of these studies, robots did not consider tools, objects, and actions simultaneously, the way humans do. A study by Goncalves et al.\cite{b7} focused on both tools and objects, but they aimed to determine features of tools and objects on the basis of categories, such as area and circularity, set in advance by the experimenters. Therefore, it was difficult for the robot to acquire the features autonomously and manipulate arbitrary objects with arbitrary tools without requiring human assistance. \\
In this study, we refer to Piaget's developmental process and use this analogy to enable a robot to detect the correct tools, objects, and actions from information on effects. To realize this, it is necessary to consider four factors: (1) tools, (2) target objects, (3) actions, and (4) effects. These must be considered simultaneously because the effects will change if one of the other three factors changes. For example, a pulling action targeting a ball will result in motion if the tool is a rake but not if the tool is a stick. Therefore, we set tasks that include all four factors and have the robot experience them. Then, we construct a model that can consider the four factors simultaneously. \\
This paper is organized as follows. Section I\hspace{-.1em}I describes the method to construct the tool-use model. Section I\hspace{-.1em}I\hspace{-.1em}I presents the experimental setup and explains how the four factors are considered simultaneously. Section I\hspace{-.1em}V presents the experimental results, and Section V discusses them. Finally, Section V\hspace{-.1em}I concludes.
\section{Tool-use Model}
In this section, we describe the tool-use model, which can consider the four factors simultaneously. Figure \ref{model} shows an overview of the model based on the study of Takahashi et al. \cite{b6}. The model contains two modules: an image feature extraction module and a tool-use module. The model is constructed and used as follows.
\begin{enumerate}
\item A robot experiences some object manipulation with tools, deriving motion and visual information from this.
\item The visual information is trained by applying the image-feature extraction module, and the visually useful elements are extracted as image features.
\item Time-series data of the image features and motion information are integrated. Then, the integrated data are fed to the tool-use module and learned as integrated information.
\end{enumerate}
\begin{figure}[tbp]
\centerline{\includegraphics[width=7.7cm]{model.eps}}
\caption{Proposed model, which can consider all four factors at the same time. This model is constructed and used in three steps. First, a robot experiences some tasks and records sensory-motor data during the experience. Second, image features are extracted by the image-feature extraction module. Third, the image features and motor information are integrated and used to train the tool-use module. MTRNN: multiple time-scale recurrent neural network}
\label{model}
\end{figure}
\subsection{Tool-use Experience by a Robot}
A robot experiences several tasks in combination of multiple tools, objects and actions. The joint angles of the robot arm are recorded as motion information and images taken by a camera mounted on the robot are recorded as visual information. It is important that the tasks display differences between effects, such as object falling, sliding, and failing to move. Choosing tasks like this allows the recorded sensory-motor data to include information about all four factors.
\subsection{Extraction of Image Features}
For the image-feature extraction module, we use Convolutional Auto-Encoder (CAE) \cite{CAE}. The Auto-Encoder (AE) \cite{Hinton} technique uses a structure that narrows in the middle, like an hourglass. The input data passes through this middle layer with a small number of nodes, and then the processed data are output with the original dimensionality. Since the output data are learned with the aim of reconstructing the input data, it is possible to extract features of images from only a few dimensions in the middle-layer nodes. CAE is AE with a convolution layer. The raw images taken by the robot are compressed by CAE, and the features of tools, objects, and robot-arm movement shown in the images are processed in low dimensions. Because of the good generalization offered by CAE, it is possible to represent unknown tools and objects as image features.
\subsection{Learning Sensory-motor data}
For the tool-use module, we use a multiple timescale recurrent neural network (MTRNN) \cite{MTRNN}, a type of recurrent neural network that can predict the next state from a current state. MTRNN uses three types of node, and these types different by time constant: input--output (IO) nodes, fast context (Cf) nodes, and slow context (Cs) nodes. The Cf nodes learn primitives of movement in the data, whereas the Cs nodes learn sequences in the data. By combining the three node types, the dynamics of time-series data can be learned. This module learns to integrate the time series of image features and motor angles. \\
In forward calculation, the output value is computed. First, the internal value of the \(i\)th neuron \(u_i\) at step \(t\) is calculated as
\begin{equation}
u_i(t)=\biggl(1-\frac{1}{\tau_i}\biggr)u_i(t-1)+\frac{1}{\tau_i}\left[\sum_{j\in N}w_{ij}x_j(t)\right],
\end{equation}
where \(N\) is the index sets of neural units, \(\tau_i\) is the time constant of the \(i\)th neuron, \(w_{ij}\) is the weight of the connection between the \(j\)th and \(i\)th neuron, and \(x_j(t)\) is the value input to the \(i\)th neuron by the \(j\)th neuron. Then, the output value is calculated by
\begin{equation}
y_i(t) = tanh\left(u_i(t)\right).
\end{equation}
The value of \(y_i(t)\) is used as the next input value:
\begin{equation}
x_i(t+1)=y_i(t).
\end{equation}
In backward calculation, we use the back-propagation through time (BPTT) algorithm \cite{BPTT} to minimize the training error in \eqref{y}, and update the weights by \eqref{weight}.
\begin{equation}
E=\sum_{i}\sum_{i\in{IO}}\left(y_i(t-1)-T_i(t)\right)^2\label{y}
\end{equation}
\begin{equation}
w^{n+1}_{ij}=w^{n}_{ij}-\alpha \frac{\partial E}{\partial w^{n}_{ij}}
\label{weight}
\end{equation}
Here, \(T_i(t)\) is the \(i\)th input given as teaching data, \(\alpha\) is the learning rate, and \(n\) is the number of iterations. At the same time, the initial value of the Cs layer (Cs(0)) is also updated to store the features of the dynamics information, as
\begin{equation}
Cs^{n+1}_i(0)=Cs^n_i(0)-\alpha \frac{\partial E}{\partial Cs^n_i(0)}.\label{Cs}
\end{equation}
Thanks to this, it is expected that each feature of the four factors will be self-organized in Cs(0) space. In other words, the features of the four factors are accumulated in Cs(0) space. \\
In addition, when dealing with unknown tools and objects, the Cs(0) value that best matches the task can be calculated from a trained network. This process is called recognition. The dynamics information detected by the robot is embedded in the recognized member of Cs(0). Therefore, by inputting the value of Cs(0) to the trained network with initial input data (IO(0)), it is possible to generate a sequence of predicted images for the behavior and motions corresponding to the value of Cs(0). We can evaluate whether the model has the ability to detect the features by examining the generated result.
\section{Experimental Setup}
\subsection{Task Design}
To guide a robot to experience information that includes all four factors, it is necessary to use experiments that have a variety of combinations of tool, object, and action. We used a humanoid robot, NEXTAGE developed by Kawada Robotics, and guided it to try some tasks with its right arm, which has 6 degrees of freedom. As shown in Fig. \ref{setup}, we prepared two kinds of tool (a stick shape and a rake shape) and three kinds of target object (a ball, a tall box, a short box, a tall cylinder, and a short cylinder). For actions, we used four kinds: pushing sideways and pulling toward the robot, with each action performed at either a high or low position. The robot was not instructed to pull tall objects at a low position because doing so might cause destruction of the tools or objects and there was a risk of imposing a high load on the robot. Therefore, we set a total of 36 tasks (=\(2\times5\times4-4\); tools \(\times\) objects \(\times\) actions $-$ forbidden cases). In all tasks, the robot began with the tool gripped, and we set the initial joint angles and objects in the same way each time.
\begin{figure}[tbp]
\centerline{\includegraphics[width=8cm]{setup.eps}}
\caption{Tools, objects, and actions for training. Tools are rake- or stick-shaped. Objects are a ball, a tall box, a short box, a tall cylinder, and a short cylinder. Actions are pull (high and low) and push (high and low).}
\label{setup}
\end{figure}
\subsection{Setting for Training}
The time length to perform the 36 tasks ranged from 7.5 to 14.4 s. By keeping the robot stationary at the final position after task completion, we recorded sensory-motor data for 14.4 s in each task, sampling each 0.1 s (144 frames for each action). Among the sensory-motor data, the visual information had 9216 dimensions (64 \(\times\) 48 \(\times\) 3; width \(\times\) height \(\times\) channel). The high-dimensional data were compressed to 20 dimensions by CAE. Figure \ref{CAE} shows the structure of CAE. The extracted 20-dimensional image features and 6-dimensional joint angle data were put into MTRNN. The values of the extracted image features were rescaled to [-1, 1], and joint angle data were rescaled to [-0.8, 0.8]. Table \ref{tab1} shows the structure of MTRNN.
\begin{figure*}[tbp]
\centerline{\includegraphics[width=12.cm]{CAE.eps}}
\caption{Structure of CAE. CAE consists of convolution layers and full connection layers. This module compressed 9216-dimensional raw images to 20-dimensional image features.}
\label{CAE}
\end{figure*}
\begin{table}[t]
\caption{Structure of MTRNN}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Node Name} & \textbf{ Number of Nodes}& \textbf{Time Constant} \\
\hline
\textbf{IO nodes} & 26& 1 \\
\hline
\textbf{Cf nodes} & 50& 5 \\
\hline
\textbf{Cs nodes} & 6& 40 \\
\hline
\end{tabular}
\label{tab1}
\end{center}
\end{table}
\subsection{Experimental Evaluation}
For model evaluation, we evaluated whether the model can well predict images and motions of tool-use tasks involving ``unknown'' tools and objects. If so, the trained model is regarded as detecting the features of tools, objects, and actions.\\
For the recognition task in Cs(0), we provided the initial and final states of unknown tasks to the model. At this time, the final state included only the image data. These data show the effects because behavior of the object can be understood by comparing the initial and goal images. Cs(0) was calculated as \eqref{Cs} such that the error between the target image and the image at the final step obtained by repeating next-state prediction from the initial state was progressively smaller. At this time, the error was calculated using the trained MTRNN as follows.
\begin{equation}
E=\left\{ \begin{array}{lll}
\sum_{i\in{IO}}\left(y_i(t-1)-T_i(t)\right)^2 & t=1\\
\sum_{i\in{img}}\left(y_i(t-1)-T_i(t)\right)^2 & t=144\\
0 & otherwise\\
\end{array}\right.
\end{equation}
Evaluation experiments were conducted using two unknown short boxes and an unknown rake, shown in Fig. \ref{unknown}. We prepared two experiments, A and B, and performed the experiments using the objects.
\begin{description}
\vspace{1pt}
\item[Experiment A]\mbox{}\\
\hspace{-10pt}The robot performs ``pull low'' with the tool, and the object ``slides'' to the front of the robot.
\item[Experiment B]\mbox{}\\
\hspace{-10pt}The robot performs ``pull low'' with the tool, but the object ``does not move.''
\vspace{1pt}
\end{description}
We put initial state and goal images of each experiment A and B to the model. Since the actual task execution of experiment B is impossible, the goal images of this experiment were artificially prepared. We set the same tool and object and showed the same arm position in the images, but the robot was expected to detect different features due to the difference of the effects. In experiment A, we expected the robot to detect the tool as ``rake,'' which can make marks, and the object as ``short object,'' which can slide without falling over. In experiment B, there were two patterns of proper detection. First, the robot could detect the action as ``low pull,'' the tool as ``stick'' (which cannot make marks), and the object as ``short object,'' which does not fall over even if it is hit by a tool. Second, the robot could detect ``high pull'' paired with ``short object'' because the object would not be moved by this combination of action and object. In this pattern, the detected tool is expected to be more likely to be ``rake'' because it has the least difference from the setting. In Table \ref{expect}, we summarize all four factors' relations suggested above. The detection of tool, object, and action is said to be correct if the detected combination satisfies this condition. By putting the recognized Cs(0) and initial state (IO(0)) into the trained model, the robot predicted images and motions. We evaluated the robot's detection from the generated results. We also checked whether the generated images and motions connect the first and the final state sequentially, considering the movement of robot arm (i.e., the action), the tool, and the object.
\begin{table}[t]
\caption{Expected Combinations of Four Factors}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Experiment} & \textbf{Tool}& \textbf{Object}& \textbf{Action}& \textbf{Effect} \\
\hline
A &Rake&Short object&Low pull&Slide \\
\hline
B & Stick& Short object&Low pull&No movement \\
\hline
B & Rake&Short object &High pull& No movement\\
\hline
\end{tabular}
\label{expect}
\end{center}
\end{table}
\begin{figure}[tbp]
\centerline{\includegraphics[width=5cm]{unknown.eps}}
\caption{Untrained rake-shaped tool and box-shaped objects. Evaluation experiments A and B were conducted using them.}
\label{unknown}
\end{figure}
\section{Experimental Results}
\begin{figure}[tbp]
\centerline{\includegraphics[width=7.5cm]{angle.eps}}
\caption{Trajectory of the robot arm six joint angles in experiment A using tool X. The data are rescaled to [-1, 1]. The dotted lines show the trained angle data for the ``low pull'' action, which is a target for generation. The solid lines show the motor angle data generated by using the recognized Cs(0).}
\label{angle}
\end{figure}
\begin{figure*}[tbp]
\centerline{\includegraphics[width=15.5cm]{img_gen.eps}}
\caption{Figures in blue squares show the given initial images and goal images. Figures in red square show the predicted image data generated by the robot. In order from the top row, we show the result of experiment A with object X, experiment B with object X, and experiment B with object Y. In experiment A, the generated images show the sequential task of drawing the object, and in experiment B, the generated images show the tasks that do not move the objects. In experiment A and experiment B with object Y, the generated figures show the tool as ``rake'', whereas in experiment B with object X, the generated figures show the tool as ``stick.''}
\label{imggen}
\end{figure*}
\begin{figure*}[tbp]
\centerline{\includegraphics[width=15cm]{5.eps}}
\caption{Transition of the robot's movement of experiment A with object X. The joint angles are generated with trained MTRNN using recognized Cs(0). The robot pulled the box with the rake in the low position and succeeded in drawing it in front.}
\label{gen}
\end{figure*}
\begin{figure*}[tbp]
\centerline{\includegraphics[width=15cm]{b5.eps}}
\caption{Transition of the robot's movement of experiment B with object Y. The robot conducted ``high pull'' action with the rake. Therefore, the box did not move. The result of the object behavior matched the goal image. }
\label{gen2}
\end{figure*}
We trained MTRNN with the data of 36 tasks, performing this 150,000 times. Then, we performed Cs(0) recognition 150,000 times for each of the two experiments using the initial and final states as in the blue squares in Fig. \ref{imggen}. Using the recognized Cs(0) value, the robot generated predicted images and motions. Fig. \ref{angle} shows, as an example of the training data and generated result, the trajectory of the robot arm's motor angles using object X in experiment A. The dotted lines show the training data and the solid lines show the data generated by using the recognized Cs(0). Although the $x$-axis (i.e., time) direction of the solid lines is shifted, the $y$-axis direction (i.e., angle) is close to the target dotted lines. In other words, it was possible to generate data to recreate the target operation, though with different timing.\\
The top row of Fig. \ref{imggen} shows the predicted images in experiment A using unknown object X. We can see that the robot properly drew a short box pulled forward with a rake. Therefore, we can say the robot detected the combination of the four factors as ``rake,'' ``short box,'' ``low pull,'' and ``slide,'' which matches the requirements shown at the top row of Table \ref{expect}. We also confirmed that the robot generated the motor angles to actually move according to the data. As a result, the robot was able to perform the task properly and succeeded in sliding the box with the rake, as shown in Fig. \ref{gen}. The result using object Y was similar.\\
The predicted images for experiment B using unknown object X are shown in the middle row of Fig. \ref{imggen}. As can be seen, the robot predicted a ``low pull'' action with a ``stick,'' and the tool passed over the ``short box.'' This combination matched the requirements shown for the case in Table \ref{expect}. Actual task execution by the robot is not shown because this experimental setting cannot be tested in the real environment. Similarly, the predicted images for experiment B using unknown object Y are shown in the bottom row of Fig. \ref{imggen}. These images show that the robot conducted a ``pull'' action with a rake, and the tool passed over the ``short box.'' The robot also generated the motor angles and to actually move according to the data, as shown in Fig. \ref{gen2}. As can be seen, the object did not move because the robot pulled at the high position. Therefore, the robot detected the combination of the four factors as ``rake,'' ``short box,'' ``high pull,'' and ``no movement'', which matched the requirements shown in the bottom row of Table \ref{expect}. \\
In addition, we carried out principal component analysis (PCA) of the Cs(0) value of the trained tasks to confirm whether the robot could self-organize features. We select three axes that are easy to see and plotted the result of 16 ``pull'' tasks in Fig. \ref{actool}. We indicate different position of the action by different colors and different tools by different markers. As can been seen, clustering and separation distinguish among different features. In other words, the robot acquired the features of actions and tools properly. We also analyzed the detection described above by PCA of Cs(0) from experiments A and B, then plotted the results on the trained tasks' Cs(0) space. As shown in Fig. \ref{actool}, each experiment's Cs(0) is located as follows. Two plots of A are located in the cluster of ``low pull'' and ``rake.'' The plot of B using object X is located in the cluster of ``low pull'' and ``stick.'' The plot of B using object Y is located in the cluster of ``high pull'' and ``rake.'' In all experiments, every plot is located exactly in the cluster that presents the detected feature's combination of actions and tools as mentioned above. This analysis confirms reliable recognition.
\begin{figure}[tbp]
\centerline{\includegraphics[width=7cm]{actool.eps}}
\caption{PCA of Cs(0) of trained tasks. PC1, PC4, PC6 are shown. We plot 16 ``pull'' tasks with a rake and a stick. Different features of pulling height and tools are separated and clustered. In addition, we plot PCA of Cs(0) of experiment A and B on the space.}
\label{actool}
\end{figure}
\section{Discussion}
In contrast with previous studies, we constructed a model that can simultaneously consider four factors of object manipulation (tool, target object, action, and effect) and thereby acquire combined features according to the trained experience. Human infants develop this process during the fifth substage of Piaget's sensorimotor developmental stage. As shown in Section I\hspace{-.1em}V, the detected combinations of the four factors matched the expectations shown in Table \ref{expect} for the experiments. Note that in experiments A and B with object X, the same tool was used, but the robot recognized different tool features from the effect of the operation. In addition, comparing experiments A and B using object Y, the robot conducted different actions in response to the different effects despite the robot arm having the same placement in the input images. This indicates that the robot could detect the proper features of tools, objects, and actions from the effects, even without concretely performing the task. In addition, we showed that the robot could generate sequences of predicted images. This is similar to the effect of infants mentally planning actions. We also showed that the model could generating motions, which enabled the robot to execute tasks correctly. In other words, the robot could reproduce the effect from the detected features. This is one of the abilities acquired by infants during the sixth substage of Piaget's sensorimotor developmental stage.
\section{Conclusion}
In this study, we proposed a tool-use model that can detect tools, objects, and actions. This model mimics infant developmental of similar ability. We set tasks that include information about four factors (tools, objects, actions, and effects), and the robot experienced them. We trained the sensory-motor data recorded during the experience with a tool-use model consisting of an image-feature extraction module and a tool-use module. The results showed that it was possible to detect the appropriate features from an initial and final state of a task using unknown tools and objects. The robot succeeded in executing actual tasks with generated joint angles and properly generated predictive images. In future work, we plan to construct a model that can select proper tools and conduct proper actions according to the state at the moment of choice and the final state. This model will enable a robot to behave according to the position of the object and its arm, which would improve versatility and ability to cope with more complicated tasks.
\section*{Acknowledgments}
This work was supported in part by a JSPS Grant-in-Aid for Scientific Research (S) (No. 25220005), a JST CREST Grant (No. JPMJCR15E3), and the ``Fundamental Study for Intelligent Machine to Coexist with Nature'' program of the Research Institute for Science and Engineering, Waseda University, Japan.
|
\section{Introduction.}
Associative separable algebras play an important role and have found
many-sided application (see, for example,
\cite{bourbalgch123b,flosaam11,gesagmj78,mapesaam09,montsmith,
saoystha,pierceb,rumsaal94} and references therein). Studies of
their structure are based on cohomology theory. On the other hand,
cohomology theory of associative algebras was investigated by
Hochschild and other authors
\cite{cartaneilenbb56,hochschild46,pommb}, but it is not applicable
to nonassociative algebras. Cohomology theory of group algebras is
an important and great part of algebraic topology. It is worth to
mention that nonassociative algebras with some identities in them
found many-sided applications in physics, noncommutative geometry,
quantum field theory, PDEs and other sciences (see
\cite{dickson,girardgb,guertzeb,kansol,krausshb,ludwrgrijmgta}-
\cite{lustsdoadaca11,schaeferb,serodaaca07} and references therein).
\par An extensive area of investigations of
PDE intersects with cohomologies and deformed cohomologies
\cite{pommb}. Therefore, it is important to develop this area over
octonions, Cayley-Dickson algebras and more general metagroup
algebras. Some results in this area are presented in
\cite{ludcohmalosal17}.
\par This article is devoted to a separability of nonassociative
metagroup algebras. Conditions are found under which they are
separable. Algebras satisfying these conditions are described.
\par A formula "$(m)$" within the same subsection "n" and section $"k"$
is referred as "$(m)$"; in another subsection within the same
section as "$n(m)$"; in another section as "$k.n(m)$". A theorem,
lemma, etc. number "n" within the same section "k" is referred as
"n", in another section as "k.n".
\par All main results of this paper are obtained for the first time.
They can be used for further studies of nonassociative algebras,
their cohomologies, algebraic geometry, PDEs, their applications in
the sciences, etc.
\section{Separable nonassociative algebras.}
To avoid misunderstandings we first give our definitions and
notations.
\par {\bf 1. Definition.} Let \par $(1)$ $A$ be a nonassociative metagroup algebra over a commutative
associative unital ring $\cal T$ with a metagroup $G$ such that
$G\cap {\cal T} = {\bf \Psi }$. \par A $G$-graded $A$-module $P$
(see subsection 2.7 in \cite{ludcohmalosal17} also) is called
projective if it is isomorphic with a direct additive of a free
$G$-graded $A$-module. The metagroup algebra $A$ is called separable
if it is a projective $G$-graded $A^e$-module.
\par One puts $\mu (z)=1_Az$ for each $z\in A^e$, where $A$ is
considered as the $G$-graded right $A^e$-module.
\par {\bf 2. Proposition.} {\it Suppose that $A$ is a nonassociative algebra
satisfying condition 1$(1)$. Then the following conditions are
equivalent:
\par $(1)$ $A$ is separable;
\par $(2)$ the exact sequence $0\rightarrow Ker ~ \mu \rightarrow A^e\mbox{ }_{\overrightarrow{\mu }} A\rightarrow 0$
splits;
\par $(3)$ an element $b\in A^e$ exists such that $\mu (b)=1_A$ and
$xb=bx$ and $b(xy)=(bx)y$ and $(xb)y=x(by)$ and $(xy)b=x(yb)$ for
all $x$ and $y$ in $A$, where $A^e$ is considered as the $G$-graded
two-sided $A$-module.}
\par {\bf Proof.} The implication $(1)\Rightarrow (2)$ is evident.
\par $(2)\Rightarrow (3)$. If the exact sequence $(2)$ splits, then $A^e$ as
the $A^e$-module is isomorphic with $A\oplus ker (\mu )$. Therefore,
$A$ is separable. The sequence $(2)$ splits if and only if there
exists $p\in Hom_{A^e}(A,A^e)$ such that $\mu p=id_A$. With this
homomorphism $p$ put $b=p(1_A)$. Then
$(xb)y=(xp(1_A))y=p(x1_A)y=p(x(1_Ay))=p((xy)1_A)=(xy)b$, hence $\mu
(b)=\mu p(1_A)=1_A$ and $xb=xp(1_A)=p(x1_A)=p(1_Ax)=p(1_A)x=bx$.
Thus $(3)$ is valid.
\par $(3)\Rightarrow (1)$. Suppose that condition $(3)$ is
fulfilled. Then a mapping $p : A\to A^e$ exists such that $p(x)=bx$.
The element $b$ has the decomposition $b=\sum_j b_j g_j$ with
$g_j=g_{j,1}\otimes g_{j,2}$, where $g_{j,1}\in G$ and $g_{j,2}\in
G^{op}$ and $b_j\in \cal T$ for each $j$. Therefore, using
conditions $(3)$ and 2.4$(1-4)$ \cite{ludcohmalosal17} we infer that
\par $p(xy)=\sum_j \sum_k \sum_l b_jg_j((c_kx_k)(d_ly_l)) = \sum_j
b_j(g_jx)y= (bx)y= p(x)y$ and
\par $p(yx)=(by)x=(yb)x=y(bx)=yp(x)$ for each $x$ and $y\in A$, where
$x=\sum_k c_kx_k$ and $y=\sum_ld_ly_l$ with $x_k$ and $y_l$ in $G$,
$c_k$ and $d_l$ in ${\cal T}$ for each $k$ and $l$. Thus $p\in
Hom_{A^e}(A,A^e)$. Moreover, $\mu (p(x))=\mu (bx)=\mu (b)x=1_Ax=x$
for each $x\in A$, consequently, the exact sequence $(2)$ splits.
\par {\bf 3. Definition.} An element $b\in A^e$ fulfilling condition
$(3)$ is called a separating idempotent of an algebra $A$.
\par {\bf 4. Lemma.} {\it Let $A$ be a nonassociative algebra
satisfying condition 1$(1)$. Let also $M$ be a two-sided $A$-module.
\par $(1)$. If $p\in Hom_{A^e}(ker (\mu ), M)$ and $\kappa : A\to A^e$ with
$\kappa (x)=x\otimes 1-1\otimes x$ for each $x\in A$, then $p\kappa
$ is a derivation of $A$ with values in $M$. \par $(2)$. A mapping
$\chi : p\mapsto p\kappa $ is an isomorphism of $Hom_{A^e}(ker (\mu
), M)$ onto $Z^1_{\cal T}(A,M)$.
\par $(3)$. $\chi ^{-1}(B^1_{\cal T}(A,M))= \{ \psi |_{\ker (\mu )}:
\psi \in Hom_{A^e}(A^e,M) \} $.}
\par {\bf Proof.} $(1)$. Since $\mu \kappa =0$, then $Im (\kappa )\subseteq
ker (\mu )$. By virtue of Proposition 2.12 \cite{ludcohmalosal17}
$\mu \kappa $ is the derivation having also properties $(2)$ and
$(3)$.
\par {\bf 5. Proposition.} {\it Suppose that $A$ is a nontrivial
nonassociative algebra satisfying condition 1$(1)$. Then $H^1_{\cal
T}(A,M)=0$ for each two-sided $A$-module $M$ if and only if $A$ is a
separable ${\cal T}$-algebra.}
\par {\bf Proof.} In view of Proposition 2 the algebra $A$ is
separable if and only if the exact sequence 2$(2)$ splits. That is a
homomorphism $h$ exists $h\in Hom _{A^e}(A, ker (\mu ))$ such that
its restriction $h|_{ker (\mu )}$ is the identity mapping.
Therefore, if $H^1_{\cal T}(A, ker (\mu ))=0$, then the algebra $A$
is separable due to Lemma 4. \par Vice versa if a homomorphism $h\in
Hom _{A^e}(A^e, ker (\mu ))$ exists with $h|_{ker (\mu )}=id$, then
each $p\in Hom _{A^e}(ker (\mu ), M)$ has the form $f|_{ker (\mu )}$
with $f=ph\in Hom _{A^e}(A^e, M)$. By virtue of Lemma 4 $Z^1_{\cal
T}(A,M)=B^1_{\cal T}(A,M)$ for each two-sided $A$-module $M$.
\par {\bf 6. Theorem.} {\it Let a noncommutative algebra $A$ fulfill
condition 1$(1)$ and
\par $(1)$ $Dim (A/J(A))\le 1$ and
\par $(2)$ $A/J(A)$ is projective as the $\cal T$-module and
\par $(3)$ $J(A)^k=0$ for some $k\ge 1$, where $J(A)$ denotes the radical of $A$.
\par Then a subalgebra $D$ in $A$ exists such that $A=D\oplus J(A)$
as $\cal T$-modules and $A/J(A)$ is isomorphic with $D$ as the
algebra.}
\par {\bf Proof.} For $k=1$ we get $A=D$. \par For $k=2$ a natural
projection $\pi : A\to A/J$ exists, where $J=J(A)=rad (A_A)$, since
$J^2=0$. The algebra $A$ is $G$-graded and ${\cal T}\subset Z(A)$,
hence $rad ((A_e)_{A_e}) \subseteq (rad (A_A))_e$, where $e$ is the
unit element of $G$. In view of conditions 2.7$(1-3)$ in
\cite{ludcohmalosal17} $J$ is the two-sided ideal in $A$ and
$J_r^m=J_l^m$ for each positive integer $m$, where $J_l^1=J$,
$J_r^1=J$, $J_l^{m+1}=JJ_l^m$ and $J_r^{m+1}=J_r^mJ$. Conditions
2.1$(9)$ and 2.7$(1-3)$ in \cite{ludcohmalosal17} imply that $A/J$
is also $G$-graded, since ${\cal T}\subset Z(A)$. \par By condition
$(2)$ the $\cal T$-module $A/J$ is projective, consequently, an
exact splitting sequence of $\cal T$-modules exists $0\to J\to A\to
A/J\to 0$. Thus a homomorphism $\kappa : A/J\to A$ of $\cal
T$-modules exists such that $\pi \kappa =id$ on $A/J$. For any two
elements $x$ and $y$ in $A/J$ put $\Phi (x,y)=\kappa (xy)-\kappa
(x)\kappa (y)$. Therefore, $\pi \Phi (x,y)=\pi \kappa (xy)-\pi
(\kappa (x)\kappa (y))=xy-xy=0$, since $\pi $ is the algebra
homomorphism and $\pi \kappa =id$. Thus $\Phi (x,y)\in ker (\pi
)=J$.
One has by the definition that \par $Dim (A/J)=\sup \{ n: \exists \mbox{ two-sided }
A/J \mbox{-module } M ~ ~ H^n_{\cal T}(A/J,M)\ne 0 \} $.
\par Then put $ux=u\kappa (x)$ and $xu=\kappa (x)u$ to be the right
and left actions of $A/J$ on $J$. Since $\kappa $ is the
homomorphism of $\cal T$-modules and ${\cal T}\subseteq Z(A)$, then
for each pure states $x$, $y$ and $u$: $(xy)u-{\sf t}_3x(yu)=\kappa
(xy)u-(\kappa (x)\kappa (y))u=\Phi (x,y)u\in J^2=0$, where ${\sf
t}_3={\sf t}_3(x,y,u)$. Then $u(xy)-{\sf t}_3^{-1}(ux)y=u\kappa
(xy)-u(\kappa (x)\kappa (y)) = u\Phi (x,y)\in J^2=0$, where ${\sf
t}_3={\sf t}_3(u,x,y)$. Thus $J$ has the structure of the two-sided
$A/J$-module.
\par Evidently, $\Phi $ is $\cal T$-bilinear. Then for every
pure states $x$, $y$ and $z$ in $A/J$: \par $(\delta ^2\Phi
)(x,y,z)={\sf t}_3x(\kappa (yz)-\kappa (y)\kappa (z))-(\kappa
((xy)z)-\kappa (xy)\kappa (z))+{\sf t}_3(\kappa (x(yz))-\kappa
(x)\kappa (yz))-(\kappa (xy)-\kappa (x)\kappa (y))z$\par $={\sf
t}_3\kappa (x)\kappa (yz)- {\sf t}_3\kappa (x)(\kappa (y)\kappa
(z))- \kappa ((xy)z)+\kappa (xy)\kappa (z) +{\sf t}_3\kappa
(x(yz))-{\sf t}_3\kappa (x)\kappa (yz)-\kappa (xy)\kappa (z)+(\kappa
(x)\kappa (y))\kappa (z)=0$, \\ consequently, $\Phi \in B^2_{\cal
T}(A/J,J)$, where ${\sf t}_3={\sf t}_3(x,y,z)$. Thus by the $\cal
T$-linearity a homomorphism $h$ in $Hom_{\cal T}(A/J,J)$ exists
possessing the property $\Phi (x,y)=xh(y)-h(xy)-h(x)y$ for each $x$
and $y$ in $A/J$.
\par Let now $p=\kappa +h\in Hom_{\cal T}(A/J,J)$, consequently,
$\pi p=\pi \kappa =id|_{A/J}$, since $\pi (J)=0$. This implies that
$p(xy)-p(x)p(y)=0$ for each $x$ and $y$ in $A/J$, since $\kappa
(xy)-\kappa (x)\kappa (y)=\Phi (x,y)=xh(y)-h(xy)+h(x)y$ and
$h(x)h(y)\in J^2=0$. Since $p(1_{A/J})-1_A\in J$, then
$(p(1)-1)^2=1-p(1)$. Therefore, $p$ is the algebra homomorphism.
This implies that $D=Im (p)$ is the subalgebra in $A$ such that
$A=D\oplus J$.
\par Let now $k>2$ and this theorem is proven for $1,...,k-1$.
Put $A_1=A/J^2$, then $J/J^2$ is the two-sided ideal in $A_1$ and
$A_1/(J/J^2)$ is isomorphic with $A/J$, also $(J/J^2)^2=0$. Thus
$J(A_1)=J/J^2$ and $A_1$ satisfies conditions $(1-3)$ of this
theorem and is $G$-graded, since $A$ and $J$ are $G$-graded and
${\cal T}\subset Z(G)$ due to conditions 2.1$(9)$ and 2.7$(1-3)$ in
\cite{ludcohmalosal17}.
\par From the proof for $k=2$ we get that a subalgebra $D_1$ in $A_1$
exists such that $A_1=D_1\oplus J/J^2$. Consider a subalgebra $E$ in
$D$ such that $E\cap J=J^2$ and $D_1=E/J^2$. Then $E/J$ is
isomorphic with $E/(E\cap J) \approx (E+J)/J=A/J$. Moreover,
$(J^2)^{k-1}=J^{k+k-2}\subseteq J^k=0$, hence $J(E)=J^2$. Thus the
algebra $E$ fulfills conditions $(1-3)$ of this theorem, is
$G$-graded and $J(E)^{k-1}=0$. \par By the induction supposition a
subalgebra $F$ in $E$ exists such that $E=F\oplus J^2$,
consequently, $F+J=E+J=A$ and $F\cap J=F\cap E\cap J=F\cap J^2=0$.
Thus $A=F\oplus J$.
\par {\bf 7. Proposition.} {\it Suppose that conditions of Theorem 6
are satisfied and $(1)$ takes the form $Dim (A/J(A))=0$. Then for
any two $G$-graded subalgebras $B$ and $C$ in $A$ such that
$A=B\oplus J(A)$ and $A=C\oplus J(A)$ an element $v\in J(A)$ exists
for which $(1-v)C=B(1-v)$ such that $(1-v)$ has a right inverse and
a left inverse.}
\par {\bf Proof.} Let $q: A\to B$ and $r: A\to C$ be the canonical projections
induced by the decompositions $A=B\oplus J$ and $A=C\oplus J$, where
$J=J(A)$. Then $p \pi =q$ and $s \pi =r$, where $\pi : A\to A/J$ is
the quotient homomorphism, $p: A/J\to A$ and $s: A/J\to C$ are
homomorphisms as in the proof of Theorem 6, since $q$ and $r$ are
homomorphisms of algebras. Put $w(x)=p(x)-s(x)$ for each $x\in A/J$,
$w: A/J\to J$. Then $\pi (w\pi )=\pi (p\pi )-\pi (s\pi )=\pi q-\pi
r=\pi (id_A-r)-\pi (id_A-q)=0$, since $Im (id_A-q)=Im (id_A-r)=J=ker
(\pi )$. Therefore, $Im (w)=Im (w\pi )\subseteq J$, hence $w\in
Hom_{\cal T} (A/J,J)$. Then
$w(xy)=p(xy)-s(xy)=p(x)(p(y)-s(y))+(p(x)-s(x))s(y)=xw(y)+w(x)y$,
consequently, $w$ is the derivation of the algebra $A/J$ with values
in the two-sided $A$-module $A/J$ (see also the proof of Theorem 6).
Since $Dim (A/J)=0$, then $w$ is the inner derivation by Proposition
2.12 in \cite{ludcohmalosal17}. Thus an element $v\in J$ exists for
which $w(x)=xv-vx$ for each $x\in A/J$. This implies that
$p(x)(1-v)=(1-v)s(x)$ for each $x\in A/J$. The element $(1-v)$ has a
right inverse and a left inverse, since $J^k=0$ implies $v_l^k=0$
and $v_r^k=0$, where $v_l^1=v$, $v_r^1=v$, $v_l^{m+1}=vv_l^m$ and
$v_r^{m+1}=v_r^mv$ for each positive integer $m$. Therefore,
$B(1-v)=p(A/J)(1-v)=(1-v)s(A/J)=(1-v)C$.
|
\section{INTRODUCTION}
In the near future, robots will interact, explore, and cooperate with human beings, and they will permeate into people's lives and our societies. This application calls for new social functions in robotic systems. Several studies have explored social navigation based on robots being aware of people \cite{sisbot2007human}\cite{nonaka2004evaluation} \cite{shi2008human}. Several studies introduced the concept of social distance, or proxemics, as a personal space that influences robot navigation performance \cite{mumm2011human} \cite{rios2015proxemics}. These kinds of studies highlighted the need to consider the social and cognitive side of people for effective navigation. However, they do not consider the people's awareness of robot which we propose as a new metric for robot navigation.
To insert human parameters (social parameters) into planning problems, a human-aware motion planner has been used to not only provide safe robot paths, but also to synthesize socially-acceptable and legible paths in the presence of humans \cite{sisbot2007human}. Another study characterizes the concept of comfort, naturalness, and sociability of navigation performance \cite{kruse2013human}. A popular method has considered the above social characteristics to build a cost map in combination with a conventional sampling-based planner \cite{mainprice2011planning}. On the other hand, there exist other approaches that utilize human teaching to navigate by demonstrating or providing feedback \cite{chernova2009interactive}. Although these kinds of researches showed us the possibility of social navigation, real-time motion planning in human-crowded environments still has
not reached a satisfactory level, especially when navigating in crowds.
\begin{figure}[t]
\centering
{\includegraphics[width=1.0\linewidth]{front.png}}
\caption{Ambiguous situation in navigation: Who has the lowest trust to the robot? A. A person who is riding a bike with a high velocity? B. A couple who are talking to each other? C. A person who is quite close to the robot? Through incorporating notions of awareness, a robot can navigate confidently for each situation while maintaining a level of comfort for the surrounding people.}
\label{Figure:Problem definition}
\end{figure}
For planning motions of mobile robots in dynamic situations, many studies regarding the prediction of pedestrians have been conducted, from a social force model \cite{helbing1995social} to machine learning-based estimations \cite{scovanner2009learning} \cite{keller2014will}. However, they did not consider an human's awareness state, which could be major factor for human robot interactions. Although the study presented in \cite{nonaka2004evaluation} uses mental states, it does not use them for navigation.
In fact, a human mental state is quite difficult to examine becauseit is fundamentally uncertain, both due to the unobservable and unpredictable nature of mental state; the same state could produce many different actions, while many different mental states could lead to the same action. Rather, this ambiguous problem can be simplified if it is confined to the problem of navigation with assumptions. We assume that a mental state can be only determined
by eye-contact between two agents. This assumption can be justified by studies from the social cognitive field \cite{macrae2002you} \cite{baker2014modeling} \cite{hollands2002look}. According to these studies, it can be confirmed that eye contact plays an important role in anticipating other's intentions for navigation. Focusing on this point, the proposed idea is quite simple: insert this eye contact model for robots to predict future movement of humans by defining
a mental state with the concept of awareness.
Many researchers have already pointed out that probabilistic representations of target states and reasoning over them are quite effective for navigation of mobile robots in dynamic environments. In situations of probabilistic decision making, Partially Observable Markov Decision Processes (POMDPs) have been widely used in robot navigation and human interactions \cite{foka2007real}. Since finding optimal control strategies in POMDP cases is computationally intractable due to the continuous and high dimensional belief space, POMDPs have usually been applied to topological navigation \cite{pineau2003point}.
In this paper, we propose a navigation planning framework of mobile robots based on human detection, POMDP and human awareness estimated by an eye contact (gaze) model.
More specifically, the proposed approach can be applied to human-crowded environments, including moving pedestrians and dynamic obstacles. The main contribution of this study
is to integrate human state estimation from real-time human detection and tracking and navigation planning to manage uncertainties, both position and awareness, from people. In
particular, the concept of human awareness of a robot is incorporated in the state model and the reward model of POMDP to improve social navigation performance in a way inspired by humans.
This paper is organized as follows. In section II, we provide background for the proposed methods containing Markov Decision Process (MDP) and POMDP models. Section III presents the framework of the proposed model, including detailed algorithms and the method for measuring human awareness based on gaze detection. In Section IV, the proposed methodology is validated in simulation and real hardware. We discuss the simulation results in terms of the effect of awareness in the navigation process.
\section{Background}
\subsection{MDP $\&$ POMDP}
The basic concept of MDP is that of a decision-making problem formulated as a set of states and actions given defined costs. A crucial assumption in MDP is that the transition of states is a Markov process, and that their future distribution is conditionally independent of the history of states and only affected by the current state. Highly probable states are determined by a reward (so-called "value") function. The goal of the agent is to select an action which will generate the maximum value for a predetermined time horizon.
Partially Observable MDP, or POMDP, is proposed to enable the MDP to be applied to the real world, which has a lot of uncertainties and disturbances. States usually cannot be measured directly, so we have to use incomplete sensors to perceive an environment. A POMDP model can be described by the tuple ($S, \pi, A, T, Z, E, R$), a finite set of states $S=\{s_1,\cdots s_{|S|}\} $, an initial probability distribution over these states $\pi$, a finite set of actions $A=\{a_1,\cdots a_{|A|}\} $, a finite set of observations $Z=\{z_1,\cdots z_{|Z|}\} $, and a transition function $T^{a,z}(s_i,s_j)=P(s_j|s_i,a,z)$ that maps $S\times A$ into discrete probability distributions over $S$.
The transition model $T(s',a,s)$ specifies the conditional probability distribution of shifting from state $s$ to $s'$ by applying action policy $a$. $O(s',a,z)$ is the observation mapping that computes the probability of observing $z$ in state $s'$ when executing action $a$. Usually, the transition model and observation model can be rewritten as $T(s',a,s)= p(s'|s,a)$ , $Z(s',a,z) = P(z|s',a)$
where $s \in S$ , $a \in A$ , $z \in Z$.
\subsection{DESPOT}
An on-line approach to solve POMDP is to combine planning and execution together through calculating and executing optimal action based on current belief state which is updated recursively over time. These on-line methods apply algorithmic techniques for computational efficiency. For example, heuristic search, branch-and-bound pruning, Monte Carlo sampling \cite{silver2010monte}, POMCP \cite{ross2008online}, and DESPOT \cite{somani2013despot} are among the fastest on-line POMDP methods recently.
We adopted DESPOT as an on-line POMDP solver. The key concept of DESPOT is to reduce all policies under $K$ sampled scenarios. Under each scenario, a policy traces out a path in the belief tree consisting of a particular sequence of actions and observations. DESPOT is defined by a tree, which keeps only the belief-tree nodes and edges that are generated from all policies under the sampled scenarios. Assuming that there is height, $H$, in the belief tree, DESPOT is more sparse, only including $O(|A|^HK)$ nodes, than the original belief tree, which contains $O(|A|^H|Z|^HK)$ nodes, leading to a dramatic improvement in computational efficiency for moderate $K$ values. Equally importantly, it can be proven that a small DESPOT tree is sufficient to generate suboptimal policy, which admits a compact representation, with bounded regret \cite{somani2013despot}.
\begin{figure}[b]
\centering
{\includegraphics[width=0.98\linewidth]{flowchart_new.png}}
\caption{Overall flowchart of proposed framework : a global planner calculates a global path while local planner seeks the optimal action policy. States for local planner (POMDP) are determined based on the combination of the human position and awareness from Human tracker (dotted block) and dynamic occupancy grid. Finally, a global path and an optimal action policy are delivered to the robot navigation controller.}
\label{Figure:OverallFramework}
\end{figure}
\section{Methods}
\subsection{Overall framework}
Our framework consists of double-layered planners: a global path planner with MDP and a local planner with POMDP, as shown in Fig.\ref{Figure:OverallFramework}. A global planner generates a collision-free path based on the environment, which can be pre-known or perceived in advance. This collision-free path can be described as a static occupancy grid map, which considers only static obstacles. On the other hand, a local planner acts as a reactive planner to cope with various possible situations that can occur during real-time navigation. Human tracking continuously detects and tracks people in a local window, a relatively small area around the robot, in order to react according to the movement of people. A POMDP-based local planner is designed to search for the optimal action policy that can obtain the maximum reward based on the belief states which can be estimated from observations. The proposed framework properly discretizes the global path given bound on the sub-path segments until arriving at the final goal position. Based on the action policy calculated from the local planner, the robot navigation controller finds the input command for the robot to reach the desired position.
\begin{algorithm}[t]
\caption{MDP Planner ($\mathcal{M}$)}
\begin{algorithmic}
\Procedure {MDPsolve}{}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\Require {$S$$\,$(state), $A$$\,$(Action), $R$$\,$(Reward), $T$$\,$(Transition)}
\Ensure{$\pi^*(s)$ : The Optimal action policy}
\State $V_0(s) \leftarrow 0$, $\,$ $k \leftarrow 0$
\Repeat
\For {$k \leftarrow k+1$}
\ForAll {$s \in S$} \\
$V_k(b) =\max_{a \in A} \{\sum_{s^{'}}T(s,a,s^{'})[R(s,a,s^{'})+\gamma V_{k-1}(s^{'})] $ \EndFor
\EndFor
\Until{$\forall$ $s\|V_k(s)-V_{k-1}(s) \| < \epsilon $}
\ForAll {$s \in S$} \\
$\pi^{*}(s)=argmax_{a} \{\sum_{s^{'}}T(s,a,s^{'})[R(s,a,s^{'})+\gamma V_{k}(s^{'})]$
\EndFor
\EndProcedure
\end{algorithmic}
\end{algorithm}
\subsection{Global Path Planning: MDP}
The 2D environment can be represented by an occupancy grid, which is described as $o(i,j)$ for $i$, $j$ are 2D coordinates of grid map respectively \cite{elfes1989using}. This occupancy value becomes 0 if it is free while it becomes 1 or 2 if it is occupied with obstacles or human. Assuming that there is no human and mapping and localization are done in advance, based on this grid map occupancy, MDP-based global planner finds a collision-free path as shown in \ref{Figure:MDP_path}. Start position is the current robot position and goal position can be set with Graphic User Interface. In fact, this MDP planner obtains mapping from occupancy grid to action policy ($\mathcal{M}:\Re^1 \rightarrow \Re^1$). Value iteration can be used to solve the MDP problem as shown in Algorithm 1. Since we have optimal solutions for all the lattice points (grid map), robot is able to flexibly cope with the environment. Then, every cell which is not occupied with obstacles can have a desired action policy:
\begin{equation}
\bar{S}(i,j)=
\Big\{ \begin{array}{cl} \pi^*(s) \in A & \textrm{if} \quad o(i,j) = 0 \quad (Free) \\ \phi & \textrm{otherwise} \quad (Occupied) \end{array}
\end{equation}
where $A = \{E,EN,N,NW,W,WS,S,SE\}$, or eight possible action sets, and $o(i,j)$ means occupancy grid. Particularly, starting from the robot's current position, a desired path can be generated by choosing consecutive grid cells until reaching goal position. Once the robot reaches the goal position, this path can be represented as
\begin{equation}
\bm{P} = \left\{ \left(p^1_x, p^1_y \right), \left(p^2_x, p^2_y \right)\cdots, \left(p^n_x, p^n_y \right) \right\}_{ \mathcal{D}_{grid}}
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width = 0.88\linewidth]{global_path.png}
\label{Figure:MDP_path}
\caption{ Global path (MDP solution) in a grid space, and the robot view.}
\end{figure}
where $\mathcal{D}_{grid}$ is the resolution of grid, which is set as 0.75 (m) in this study. We selected this value so that actual mobile robot hardware fits in one; 0.75 (m) $\times$ 0.75 (m) cell, which is applicable and reasonable for real hardware implementation. Based on this path information, we construct a smooth path with cubic spline interpolation methods. Fig. \ref{Figure:MDP_path} shows the collision-free path with static obstacles.
\subsection{People Detection \& Tracking}
In order to successfully navigate human-robot coexisting environment, real-time detection and tracking of people are essential. In practice, tracking a human is quite a challenging task due to the limitation of sensor visibility, noise in sensor readings, possible target occlusions, and confusions from multiple targets. (RGB-D camera and laser scan data) is used to detect and track humans around the robot. This sensor fusion technique can improve the accuracy of detection and tracking people.
\subsubsection{Vision-based Detection}
For the vision based detection algorithm, one of the state of the art, a deep learning based real-time object detection algorithm, or You Only Look Once (YOLO) \cite{redmon2016you} is implemented to detect the existence of people. This algorithm has the advantage of being able to accurately recognize human only with some part of the shape of a person. In addition, this method is known for its rapid detection capability, as time latency is less than 25 milliseconds with optimal settings, which is far faster than conventional vision-based approaches. We can obtain the number of existing humans in the camera region and approximate 3D position of findings. Using point cloud data from the RGB-D camera, 3D bounding boxes of the human can be obtained, which is described in Fig. \ref{Figure:human_recognition}.
\begin{figure}[t]
\centering
{\includegraphics[width=1.0\linewidth]{human_recognition.png}}
\caption{ (a) Robot in grid space with the global path and human in front of it. (b) Snapshot of real laboratory. (c) Image-based human detection with YOLO. (d) Cropped face from face recognition.}
\label{Figure:human_recognition}
\end{figure}
\subsubsection {Laser Sensor-based Detection}
Laser scan data also can be utilized to detect a human leg through pattern recognition. Human leg pattern can be detected as three categories - Leg Apart (LA), Forward Straddle (FS), and Single Leg (SL) with feature-based classifiers, as described in \cite{bellotto2009multisensor}. We adopted their work to find all possible human leg patterns based on laser scan information. An advantage of using laser sensors is that algorithm is able to rapidly detect human legs when people are moving at a high speed, while vision-based methods might not able to detect human movements due to their processing and computational costs.
\subsubsection{Measurement Fusion- Data processing}
The main reason for using sensor fusion method is to simultaneously utilize the advantages of the observations. To combine two different types of observations (vision-based and laser-based measurements), the measurement fusion method is known for its effectiveness to produce better estimation results. We assumed that YOLO detection is much more reliable because the laser detection algorithm only provides possible candidates of human leg, rather than exact human leg information. This means that laser-based observations need to be filtered based on YOLO detections. Consequently, once robot detects the number of humans around it, referring to $N$, data processing method is designed to extract the same number of human legs out of all candidates from laser sensors. This processing method can be used in combination with gaze control since two sensor's field of view are different. Outputs of data processing are delivered to the Kalman filter.
\subsubsection{Kalman filter}
The position of each detected humans is individually tracked over time using the Kalman filter technique \cite{julier1997new}. The Kalman filter algorithm consists of two steps: prediction and correction (update step). The first step predicts the current state from the previous states and the second step uses sensor measurement to update or correct the estimation from the previous step.
Each filter estimates each position of human candidates, or $x_k$, over time as one element of target clusters which is denoted as $X_k=\{x^1_k,x^2_k, \cdots x_k^N\}$, where N is the total target numbers at time step $k$. Here, a state includes 2D position and velocity of particle as $x_k^i=(x,y,\dot{x}, \dot{y})^T$, and the Kalman filter model uses the linear dynamic model and measurement model which are formulated as
\begin{eqnarray}
\dot{x}&=&Ax+Bu+w \\
z&=&Hx+v
\end{eqnarray}
where $A$ is a state transition matrix, $B$ is the input matrix, $u$ is input variable, and $w$ is a white Gaussian noise with covariance $Q$. The measurement variable, $z$, can be modeled with $H$, which is the observation matrix, and $v$ is observation noise variable, of which the covariance is $R$. Based on this formulation, the Kalman filter iteratively estimates a state variable, $x$, with consequent measurements, $z$. System model parameters used in this study can be referred to in the previous study \cite{leigh2015person}.
\subsubsection{Face Recognition}
\label{face_recognition}
The face recognition package from \cite{ageitgey2013} is also implemented in our system. It basically uses Histogram of Oriented Gradients (HOG) \cite{dalal2005histograms} to detect faces and face landmark estimation \cite{kazemi2014one} to extract face features. Then, extracted features are used to train Deep Convolutional Neural Network to recognize face. Therefore, human faces can be recognized from the video stream, which leads to capturing human gaze. This face recognition is only used for gaze detection, not for human detection and tracking.
\subsubsection{Gaze detection (Awareness)}
In order to observe $p_{gaze}$, given face image from the above algorithm, a gaze tracker using a simple image gradient-based eye center algorithm is applied to track the gaze of a human \cite{timm2011accurate}. This algorithm provides relatively high accuracy results with low computational cost.
The robot's success of eye contact recognition is proportional to the time a person stares at the robot. It can be detected based on the fact that the human eyes are looking at the center of the camera of the robot. Therefore, the proposed method calculates time spent with both eyes on the center rectangular region, which can be defined as $C(R)$ with width $w$, and height $h$, as shown in Fig.\ref{Figure:gaze detection model description}. If measured time duration exceeds the time threshold, or $\epsilon$, then, the awareness variable is activated. This time threshold can be determined as 5 seconds through a trial and error approach. Based on this detection model, we can judge whether a person is aware of a robot or not, which can be described as the awareness variable. Consequently, the following equation is used to define the awareness variable
\begin{equation}
G(t)=
\Big\{ \begin{array}{cc} 1 & \int_{t_0}^{t}{P_{gaze}}(\tau)d\tau> \epsilon \\ -1 & \textrm{otherwise} \end{array}
\end{equation}
\begin{figure}[t]
\centering
{\includegraphics[width=1.0\linewidth]{gaze_final.png}}
\caption{Gaze detection model}
\label{Figure:gaze detection model description}
\end{figure}
If a person is aware of a robot, when the awareness variable is active, the robot is allowed to approach closer than to those who did not make eye contact. This feature can be obtained by differentiating the reward model according to pedestrians. In our model, the collision reward is individually set for each observed pedestrian to change the permission range of distance between human and robot.
\subsection{Local planner: POMDP}
\subsubsection{States}
As a local planner, we utilize the on-line POMDP solver, DESPOT, to obtain the optimal action policy based on current belief. The state variable contains a robot state and people state. A robot state includes robot position, $R_{pos}$, which can be measured directly by sensors in a local window describing the surroundings in the form of an occupancy grid. Pedestrians state contains the current position, and awareness, which are represented as $p_{pos}$ and $p_{awareness}$ respectively. The dimension of states varies according to the number of pedestrians in the local planner, $N$. To sum up, the state of the POMDP model can be described as
\begin{equation}
\begin{split}
\bm{S} &=\{S_{\textrm{robot}}, \, S^1_{\textrm{ped}}, \, \cdots, \, S^N_{\textrm{ped}} \}. \\
S_{\textrm{robot}} &= [R_{pos}] \\
S_{\textrm{ped}} &=[p_{pos}, \, p_{awareness} ]
\end{split}
\end{equation}
\label{Eq : POMDP STATE MODEL}
\subsubsection{Observations}
The observation information can be measured by a scanning laser range finder and a RGB-D Camera. The field of view of laser sensor is 270˚ range. The distance range is assumed to be 5 meters ahead, to sense humans, static and dynamic obstacles. From this sensor model, our observation model is written as:
In a simulation environment, each sensor has its own noise model, which is based on its specification. Based on sensor information, we can update our belief over the state variable at each time step.
\subsubsection{Actions (Policy Set)}
Simplifying current mobile robot movements, we have 3 possible action policies: Go, Wait and Avoid. The "Go" policy make a robot go forward along the global path, while "Wait" means do not move for one time step. These two policies are included in the POMDP action set. Lastly, the "Avoid" action makes the robot move to the collision-free position when a robot can not move along the path due to pedestrians or dynamic obstacles. This action is only activated when the robot can not move for the predefined amount of time, and this command makes the global planner regenerate a global path, depending on dynamic occupancy grid. The default step size of this action policy is set to one grid resolution. Reducing the dimension of action policy sets can lessen the computational burden of the DESPOT algorithm.
\begin{equation}
\textrm{A} = \{\textrm{Go},\, \textrm{Wait}, \, \textrm{Avoid}\}
\end{equation}
\begin{algorithm}[t]
\caption{Proposed Planner (Main loop)}
\begin{algorithmic}[1]
\State SetStartGoal()
\State $P\leftarrow$ MDPSolve(), $k \leftarrow 0$, $N\leftarrow \text{size}(P)$
\While{Not Goal}
\State StateUpdate() \Comment{Static/Dynamic Obs}
\Repeat
\For {$k \leftarrow k+1$}
\State TrackerUpdate() \Comment{People}
\State BeliefUpdate() \Comment{Particle Filtering}
\State Action$^*$ $\leftarrow$ POMDPSolve() \Comment{DESPOT}
\If {Action$^*$ $=$'Avoid'}
\State Go To Line 2
\EndIf
\State RobotControl(Action$^*$)
\EndFor
\Until{$k<N$}
\EndWhile
\end{algorithmic}
\end{algorithm}
\subsubsection{Rewards}
Establishing a reward model is quite a sensitive problem because we can design the characteristics of the desired action. Our proposed reward basically consists of reward for goal, penalty for collision, and time. Reward for goal state is set to the highest value. If a robot collides with a human or wall, reward gives a penalty. Lastly, since a navigation time is also one indicator that can evaluate performance, it is regarded as a penalty. Thus, reward function is written as
\begin{equation}
R(s) = w_gR_{goal}(s)+w_cR_{col}(s)+w_tR_{time}(s)
\end{equation}
where $w_g$,$w_c$,and $w_g$ are weighting factors of each reward function. In particular, a human-collision reward model must be designed more carefully, depending on the degree of awareness that we defined above. A potential field approach \cite{ge2000new} is used to model the collision function based on the awareness variable and the distance between robot and human. This collision reward enables the planner to consider an awareness effect for each pedestrian. In other words, the reward collision model varies depending on awareness to change an acceptable permission range in equation (\ref{equation:R_col}). By differentiating this range, robot flexibly navigates with pedestrians who have different levels of awareness between robot and human. $R_{\textrm{Aware}}$ is set to smaller than $R_{\textrm{Non-Aware}}$ to allow the robot to have a high proximity to humans who are aware of it.
\begin{equation}
R_{col}=
\Big\{ \begin{array}{cl} R_{col} & dist(R_{pos},P_{pos}) \leq \rho_{Aware} \\ 0 & \textrm{otherwise} \end{array}
\label{equation:R_col}
\end{equation}
\begin{equation}
\rho_{Aware}=(\frac{1-G(k)}{2}){R_{\textrm{Aware}}}+(\frac{1+G(k)}{2}){R_{\textrm{Non-Aware}}} \\
\label{Equation: Reward+collision}
\end{equation}
\begin{figure*}[t]
\centering
{\includegraphics[width=1.0\linewidth]{subplot4-eps-converted-to.pdf}}
\caption{Time series of grid map of local window (10 $\times$ 10). The robot goes on its own pre-calculated MDP path because the robot knows pedestrians who are aware of it. Otherwise, a robot executes "Wait" action if a pedestrian who is not aware of robot is nearby. }
\label{Figure:Simulation result}
\end{figure*}
\subsection{Belief Update} \label{Section:Belief Update}
For every time step, POMDP maintains distribution over the states, or belief space. This belief space can be updated with the following equation:
\begin{equation}
b^{'}(s^{'}) = \eta Z(s^{'},a,o)\sum_{s \in S} T(s,a,s^{'})
\end{equation}
where $\eta$ is a normalizing factor. DESPOT does not obtain the exact belief space, but calculates an approximate belief by a set of $K$ particles. Each particle corresponds to a sampled POMDP state, which can be written as,
\begin{align}
B_t &:= (s^1_t,s^2_t, \cdots , s^K_t) \\
s^i_{t+1} &= p(s_{t+1}|s_t,a_{t+1},o_{t+1})
\end{align}
where each state vector, $s^i_t$ represents the state vector for the $i$th particle. A general particle filter is applied to update belief space with k, the number of particles, equal to 5000. This filtering is an approximation of Bayes filter approach.
\subsection{Navigation Control}
For basic navigation functions, ROS navigation stack is utilized. When the global path planner finds the desired trajectory, it can generate safety velocity commands to control the wheels of mobile base to follow a desired trajectory according to information from odometry, sensor streams, and a goal pose.
\section{Experimental Results}
\subsection{System Description}
A Toyota Human Support Robot (HSR) mobile robot is used as the hardware platform for both simulation and experiments. The mobile base consists of two drive omni-wheels and three casters, which are located to the front and rear of those. It can smoothly change the direction of navigation and has the capability to avoid obstacles. The maximum speed of HSR is approximately 0.22m/s, the maximum step size of the mobile base is 5mm, and the maximum incline is $5^{\circ}$. HSR has a variety of sensors. For vision information, two stereo cameras are mounted around the two eyes, a wide angle camera is on the forehead; a depth camera (Xtion, Asus) is placed on the top of the head to get RGB-D video stream. Furthermore, a laser range scanner (UST-20LX, Hokuyo) is equipped at the front side of the mobile base platform. The robot uses two different computers; one is for the main operating programs regarding basic functions and the other, which has a GPU (NVIDIA Jetson TK1), is only for running YOLO for object detection. All sub-programs to operate the robot are able to communicate, and to send and receive useful information to each other via the ROS interface.
\subsection{Pedestrian Model (Human)}
One important feature of simulation is how to make movements in a pedestrian model. In our simulation, some pedestrians have their own path, others move randomly to collision-free space. They can even go out of the grid map, and the robot does not take them in to consideration anymore. The key feature of a pedestrian, the awareness variable, $G$, is determined by the gaze-detection. Although there are the limitations of continuously tracking people's faces in real situation, it is assumed that a camera has been able to keep track of the people's faces so it can track the movement of the pupil. During simulation, the awareness variables are manually set to each pedestrian. In contrast, for the real experiment, we assumed that once $G$ is activated, it never turns off so that the robot regards the person as being aware of it.
\subsection{Scenario Analysis}
We conducted both simulation and real platform experiments. For an actual experiment, there was a limit to closely analyzing all the situations. The evaluation of the algorithm is focused with the analysis result of the simulation.
Each scenario has its own start and goal position. To test a robot's navigation performance according to the movements of pedestrians, we fixed the initial position and repeated the scenarios. The initial and goal positions of the robots are set to (0,1) and (7,7), respectively. For each scenario, 3 or 4 pedestrians are set to roam within the grid space. The initial positions of pedestrians are manually set to near the goal position. One pedestrian moves along the predefined path and others walk around with random direction in a local window. Each pedestrian has their mental state, $G$, the awareness variable, which is randomly chosen as a constant for each scenario.
Fig. \ref{Figure:Simulation result} shows the one case scenario which contains the desired navigation strategy. A Robot is supposed to follow a global path (\bm$P$) and the DESPOT solver generates the optimal action policy depending on the belief states of the robot. In this scenario, it is assumed that pedestrians 1 and 2 (represented within blue triangle and red circle in Fig. \ref{Figure:Simulation result}) are aware of the robot and that pedestrian 3 (green cross) has no awareness of robot. Looking closely, at $t=7$ and $t=10$, a robot executed the 'Go' action because pedestrians who were aware are in front of it. This is because the robot assumes that they will not move in the direction where they will collide with it. At $t=12$, robot chose the 'Wait' action when the pedestrian who was not aware of it is near its predefined path. At the next time step, $t=13$, we confirmed that the robot moved along the path when the location of the untrustworthy pedestrian is a certain distance away. This scenario is a good example, showing that the robot changed its navigation policy according to how people are aware of a robot.
\subsection{Performance Analysis}
The proposed algorithm was tested in simulation with twenty-five sampled scenarios in simulation that have the same initial condition to ensure reliability.
\subsubsection{Navigation time}
Due to a randomness of movement by pedestrians and the corresponding reactions of a robot, navigation time is different for each scenario. Since navigation time is also an important factor, we analyzed average navigation time to evaluate performance. If a robot never stops and keeps going along the path, a robot can reach the goal position point through 14 movements. As shown in Fig. \ref{Figure:Simulation result2}.(a), average navigation time (the number of steps to reach the goal position) is 20.54 and standard deviation equals to 3.97. When considering that there are 4 peoples, this time seems to be a reasonable navigation time.
\begin{figure}[h]
\centering
{\includegraphics[width=0.95\linewidth]{performance_edit-eps-converted-to.pdf}}
\caption{Navigation performance : (a) Navigation Time (b) The distance between the closest human and robot for each cases. In both cases, there is one pedestrian with awareness, pedestrian without awareness.}
\label{Figure:Simulation result2}
\end{figure}
\subsubsection{Proximity to pedestrians}
The evaluation of navigation performance is quite difficult to define. To measure the social efficiency of the proposed navigation methods, we analyzed an average distance to the pedestrian when the robot executes the "Wait" action. By calculating this distance, we can examine how robot reacts to close pedestrians based on awareness. In Fig.\ref{Figure:Simulation result2}.(b), we first measured the mean distance from the closest pedestrian. The distance data is divided into two categories based on the existence of awareness. The two average distances between the pedestrian and robot, for the aware and non-aware cases were equal to 1.64 and 2.25, respectively. We confirmed that the robot achieves low proximity to a person who is not aware of the robot while it obtain high proximity to those who are aware of it. In other words, if pedestrians make eye contact with the robot during navigation, the robot can trust and navigate closer to them. As a result, by acquiring this characteristic of a distinctive proximity to humans, the social navigation ability of the robot improved.
\section{CONCLUDING REMARKS}
The awareness-based navigation planning method is proposed for coexistence of humans and mobile robots. The proposed motion planner utilizes the concept of awareness as a simplified version of a human mental state. Both MDP and POMDP planners are integrated to increase social navigation performance while reducing real-time computational cost. To achieve the social interaction ability of a robot, a human detection and tracking system which includes a gaze detection model is implemented to obtain human positions and human awareness of the robot, which can be a key factor for socially-acceptable navigation. Adopting the concept of awareness, robots can react to or handle dynamic situations in a social manner, which is a key characteristic of human navigation. The simulation results and actual experiments with the HSR robot showed that the proposed planner makes it possible for a robot to handle ambiguous situations flexibly. If a person is aware of a robot, the robot is allowed to approach closer than to those who do not make eye contact, which indicates that they are not aware of the robot.
However, several future works still remain; currently, our proposed on-line POMDP planner currently can only select actions from discrete set. A next step is to extend this study to apply continuous-action sets. In addition, there exist hardware-implementation issues for mobile robots to detect awareness or other mental states. Human intention recognition technologies such as gaze-tracking or facial expression detection should be upgraded to estimate human mental state accurately. Lastly, an actual navigation of a robot should be evaluated properly from a human subject perspective. In other words, a metric to measure social-acceptableness for navigation need to be defined. Since performance of social navigation is quite difficult to measure, conducting human-related experiments and analyzing feedback from pedestrians can be a good reference for building desired strategies for a motion planner.
\section*{ACKNOWLEDGMENT}
This research was partially supported by a NASA Space Technology Research Fellowship (NSTRF), Grant number NNX15AQ42H.
\bibliographystyle{IEEEtran}
|
\section{I. Introduction}
Originating from quantum fluctuations, the Casimir effect describes the phenomenon where an attractive force emerges between two non-contacted, uncharged plates in vacuum \cite{Casimir}. The Casimir effect tells us that vacuum is not empty, but full of fluctuations with photons popping in and out. In fact, there are many other effects that can manifest the fluctuating nature of vacuum. For example, quantum fluctuations can exert a torque on bodies that lack rotational symmetry, called Casimir torque \cite{YBarash,measureCasimirtorque}. If some discrete symmetries are broken in materials, quantum fluctuation can transmit symmetry breaking effect to nearby atoms and perturbs the atom's spectra, namely the quantum atmosphere effect \cite{jiangWilczek}.
In recent years, another interesting phenomenon, called Casimir friction, was discovered. Here, objects moving relative to each other can feel a dissipative viscous force due to the exchange of Doppler-shifted photons \cite{JBPendry}. Perhaps counter-intuitively, a spinning object in vacuum will eventually slow down due to Casimir friction \cite{AManjavacas1}. In recent years, theorists have proposed many models that feature the Casimir friction \cite{AIVolokitin}, and some of them are closely related to experimental phenomena \cite{MOigrebinskii,AVolokitin2}.
However, to our best knowledge, all the proposed Casimir friction phenomena (motion-induced vacuum forces) are dissipative. A natural question then arises: is it possible to find a dissipationless motion-induced vacuum force? This question is partially motivated by the recent progress in quantum Hall physics, where dissipationless Hall viscosity emerges as a new topological signature \cite{JAvron}. We address this question in this paper by examining a rotating particle near a bi-isotropic material (BIM) plate. Existing commonly in nature, BIMs include materials that break time-reversal symmetry (TRS) or parity symmetry (PS) or both (PTS) \cite{ILindell}. In recent years, the widely studied Chern insulators \cite{FHaldane} and chiral metamaterials \cite{STretyakov} can be classified as bi-isotropic materials breaking TRS and PS, respectively.
We show that, in addition to the dissipative Casimir friction, a dissipationless force can emerge for a rotating particle near a PS or TS (or both) breaking BIMs. Since the dissipationless rotation-induced force is always parallel to the particle's rotation axis and changes sign when its spinning direction is reversed, we, therefore, call it the axial Casimir force (ACF). Two cases are of particular interest: (i) when the rotation axis is parallel to the BIM plate, the axial Casimir force is lateral (L-ACF); (ii) when the rotation axis is perpendicular to the BIM plate, the axial Casimir force is vertical (V-ACF) [Fig. 1]. We calculate ACF both numerically and analytically, and show that TS breaking is crucial for V-ACF, whereas, by contrast, PS breaking is important for L-ACF. Let us observe that very recent experiments have already achieved a superfast rotation of nanoparticles, making the ACF within the experimental reach \cite{RReimann}.
\begin{figure}[!htb]
\centering
\includegraphics[height=3.7cm, width=7cm, angle=0]{figure1.eps}
\caption{Schematic of the structure.
$d$ is the distance from the center of the rotating object to the BIM plane. $\Omega$ represents the rotating frequency of the object. $\bold{\hat n}$ is the unit vector in the rotating direction. $\theta$ is the angle between $\bold{\hat n}$ and the $x$ direction. (a) shows the general case, while (b) shows two special rotating directions, perpendicular to the BIM plane (top) and parallel with the BIM plane (bottom). \label{fig:1}}
\end{figure}
\section{II. Model}
We consider a spherical, isotropic particle rotating with frequency $\Omega$ located at the position $\bold r_0=(0,0,d)$ above a BIM plate at $z=0$ plane [Figure 1 (a)]. Without loss of generality, we assume that the rotating axis lies in the $x$-$z$ plane and forms a $\theta$ angle with $x$ axis. In this paper, we study the dissipationless ACF along the rotating axis $\bold{\hat n}$. Particularly, when $\theta=0$, the ACF lies in the $x$-direction, becoming a L-ACF; when $\theta=\pi/2$, the ACF is parallel with $z$-direction, leading to the V-ACF.
We assume that the particle is small enough so that it can be safely described by polarization function $\alpha(\omega)$, for instance a small metallic ball. (In other words, the size of the particle $\rm R$ is much smaller that the distance $d$.) The electromagnetic force that exerts on an electric dipole in the direction $\bold{\hat n}$ can be evaluated via the formula
$
F_{n}=p_i(t)\partial_{\hat n} E_i(\bold r_0, t)
$ ($i\in \{x,y,z\}$),
where $p_i(t)$ and $E_i(\bold r_0,t)$ are, respectively, the instantaneous electric dipole moment at time $t$ and the electric field at the particle. (Einstein summation rule is implied through out this paper.) Note that we have omitted any magnetic dipole contribution, which is much smaller than the electric dipole contribution \cite{AVolokitin2}. We will further elaborate this point later in the paper.
Although the average electric dipole and electric field are zero in vacuum, quantum fluctuation can still induce a instantaneous dipole, therefore exert a force on the particle. (This is also the mechanism of Wan der Vaals force.)
There are two kinds of fluctuations that contribute to the ACF: (i) fluctuations of the dipole moment of the particle, and (ii) fluctuations of the field caused by the electromagnetic response of the BIM plate.
Therefore, the total ACF includes two terms, i.e.,
\begin{eqnarray}\label{eq1}
F_{n}=\langle p_i^{fl}(t)\partial_{\hat n}E_i^{ind}(\bold r_0,t)\rangle+\langle p_i^{ind}(t)\partial_{\hat n}E_i^{fl}(\bold r_0,t)\rangle,
\end{eqnarray}
where $\langle \cdots \rangle$ stands for the average over fluctuations in vacuum. In this formula $p_i^{fl (ind)}$ and $E_i^{fl(ind)}$ are, respectively, the fluctuating (induced) electric dipole moment and electric field at the particle. (Note that the cross terms $\langle p^{fl}\partial_{\hat n} E^{fl}\rangle$ and $\langle p^{ind}\partial_{\hat n} E^{ind}\rangle$ vanish in vacuum because dipole moment and electric field arise from different sources.) When the particle is not rotating, the force in z-direction is the usual Casimir-Polder force. As soon as the particle rotates, ACF will emerge, have an additional component in z-direction.
Applying Fourier transformation, one can write down the induced field (dipole moment) in terms of the fluctuation of the dipole moment (electric field) in $\omega-$space, yielding
$E_i^{ind}(\bold r,\omega)=G_{ij}(\bold r,\bold r_0,\omega) p_j^{fl}(\omega)$ and
$p_i^{ind}(\omega)=\alpha_{ij}(\omega) E_j^{fl}(\bold r_0,\omega)$,
where $G_{ij}$ and $\alpha_{ij}$ represent Green's tensor and polarization tensor, respectively. Substitute the above equations into Eqn. \eqref{eq1}, and one can obtain
\begin{eqnarray}\label{eq4}
F_{\hat n}=\int_{-\infty}^{\infty} \frac{d\omega d\omega^{\prime}}{4\pi^2}\, e^{-i(\omega+\omega^{\prime})t}\left\{\langle p_i^{fl} p_j^{fl}\partial_{\hat n}G_{ij}(\bold r_0,\bold r_0,\omega^{\prime})\rangle\right.\nonumber\\+\left.\langle \alpha_{ij}(\omega)\partial_{\hat n}E_i^{fl}(\bold r_0,\omega^{\prime})E_j^{fl}(\bold r_0, \omega)\rangle\right\},\qquad
\end{eqnarray}
where one should notice that the derivative only acts on the first component of Green's tensor, i.e., $\partial_{\hat n}G_{ij}(\bold r_0,\bold r_0,\omega)\equiv \partial_{\hat n} G_{ij}(\bold r,\bold r_0,\omega)|_{\bold r=\bold r_0}$. We emphasize that, in Eqn. \eqref{eq4}, $p_i$ and $\alpha_{ij}$ are the effective electric dipole moment and electric polarizability in the laboratory frame, respectively. However, the electric dipole and polarizability are defined in the rotating frame of the particle. Therefore, one needs to identify the transformation from electric dipole or polarizability ($\tilde p_i$ or $\tilde \alpha_{ij}$) in the rotating frame to those in the laboratory frame \cite{AManjavacas1,AManjavacas}:
$p_i(\omega)=\Lambda_{ij}^+\tilde p_j(\omega_+)+\Lambda_{ij}^0\tilde p_j(\omega)+\Lambda_{ij}^{-} \tilde p_j(\omega_-)$ and
$\alpha_{ij}(\omega)=\Gamma_{ijkl}^{+}\tilde \alpha_{kl}(\omega_+)+\Gamma_{ijkl}^{0}\tilde \alpha_{kl}(\omega)+\Gamma_{ijkl}^{-}\tilde \alpha_{kl}(\omega_-)$,
where $\omega_{\pm}=\omega\pm\Omega$ is the Doppler-shifted frequency due to rotation. Here, $\Lambda^{0}$, $\Lambda^{\pm}$ and $\Gamma^{0}$, $\Gamma^{\pm}$ represent the transformation tensor for dipole moment and polarizability. (See Appendix A.)
By applying fluctuation-dissipation theorem (FDT) to Eqn. \eqref{eq4}, we obtain a compact expression of \textbf{axial Casimir force}:
\begin{eqnarray}\label{eq8}
F_{n}(\Omega)=F_x(\Omega)\cos^2\theta+F_z(\Omega)\sin^2\theta.
\end{eqnarray}
In this formula, $\theta$ denotes the rotating direction of the particle [Fig. 1(a)], $F_{x/z}$ denotes the ACF in x/z direction with expressions:
\begin{eqnarray}\label{eq9}
\begin{aligned}
F_{x/z}(\Omega)&=\frac{\hbar}{\pi}\int_0^{\infty} d\omega\, \mathrm{Im} \left\{\Sigma_{x/z} \right\}\times \\&\left[\mathrm{Im}\,\alpha (\omega_+)N(\omega_+)- \mathrm{Im}\,\alpha (\omega_-)N(\omega_-)\right].
\end{aligned}
\end{eqnarray}
Here, the differential Green's functions $\Sigma_{x/z}$ are determined by the surface Green's tensor $G_{ij}$ of the BIM plate via $\Sigma_x=\partial_xG_{yz}-\partial_xG_{zy}$ and $ \Sigma_z=\partial_zG_{xy}-\partial_zG_{yx}$; $N(\omega_{\pm})\equiv n(T_1,\omega_\pm)-n(T_2,\omega)$ is defined by the difference of Bose-Einstein distribution, where $T_1$ and $T_2$ are temperatures at the rotating particle and the BIM plate, respectively. Note that, in deriving the above formula, we have used the isotropic assumption of the electric polarizability of the particle, i.e., $\tilde \alpha_{ij}(\omega)=\tilde \alpha(\omega) \delta_{ij}$ ($i,j\in \left\{x, y, z\right\}$). Eqns.
\eqref{eq8} and \eqref{eq9} are the main results of this paper. We stress that the ACF is different from the usual Casimir-Polder force, because ACF exists only when the particle is rotating with a finite speed. We shall compare ACF with the usual Casimir-Polder force later in this paper and in the appendix.
\section{III. Criterion of ACF - TRS/PS breaking}
In this part, we demonstrate that the emergence of an ACF requires TRS/PS breaking of the underlying BIM plate. A BIM plate
can generally be described by the constitutive relations $\bold D=\epsilon \bold E+(\chi-i\kappa)\sqrt{\epsilon_0\mu_0}\bold H$ and $\bold B=\mu\bold H+(\chi+i\kappa)\sqrt{\epsilon_0\mu_0}\bold E$, where $\epsilon$ ($\epsilon_0$) and $\mu$ ($\mu_0$) are, respectively, the permittivity and permeability of the BIM plate (vacuum). The essence of BIMs is encoded in the magnetoelectric parameters $\chi$ and $\kappa$, which characterize the non-reciprocity and chirality of the system, respectively. BIM with $\chi\neq 0$ and $\kappa=0$ has been called Tellegen medium, where TRS is broken. By contrast, BIM with $\chi= 0$ and $\kappa\neq 0$ has been labeled Pasteur medium, where PS is violated. Materials with $\chi=0$ and $\kappa=0$ is usually called simple isotropic medium, whereas, by contrast, both $\chi\neq 0$ and $\kappa\neq 0$ represent more general BIMs. [See Table I]
\begin{table}
\centering
\caption{Classification of BIMs \cite{ILindell} with axial Casimir force.}
\begin{tabular}{cccc}
\hline\hline
non-reciprocity & chirality & classification & \color{blue}{axial Casimir force}\\
\hline
\rowcolor[gray]{0.88}
$\chi=0$ & $\kappa=0$ & simple BIM & $F_{n}=0$\\
\hline
\rowcolor[gray]{0.95}
$\chi=0$ & $\kappa\neq 0$ & Pasteur & $F_x \neq 0$; $F_z=0$\\
\hline
\rowcolor[gray]{0.88}
$\chi\neq 0$ & $\kappa=0$ & Tellegen & $F_x= 0$; $F_z\neq 0$\\
\hline
\rowcolor[gray]{0.95}
$\chi\neq 0$ & $\kappa\neq 0$ & general BIM & $F_{x}\neq 0$; $F_{z}\neq 0$\\
\hline
\end{tabular}
\end{table}
With the constitutive relations of BIMs, one can study the electromagnetic response of BIMs. To obtain the ACF, one needs the expression of Green's tensor of BIM plate. In general, the surface Green's tensor $\mathds G$ can be expressed in terms of Fresnel coefficients for reflection at the BIM plate \cite{JCrosse}, i.e.,
\begin{eqnarray}\label{eq10}
\mathds G(\bold r,\bold r^{\prime},\omega)&=&\frac{i}{2\pi}\int\,{d^2 k_\rho}\,\frac{e^{i\bold{k_\rho}\cdot (\bold r-\bold r^{\prime})+ik_z(z+z^{\prime})}}{k_z} r_{\mu\nu}\bold M_{\mu\nu},\nonumber\\
\end{eqnarray}
where $\bold{k_\rho}=(k_x, k_y)$ and $k_z=\sqrt{\omega^2-k_\rho^2}$ represent the wave vectors in x-y plane and z-direction, respectively; $r_{\mu\nu}=E_\mu^{ref}/E_\nu^{in}$ ($\mu,\nu \in \left\{ s, p\right\}$) stands for the reflection coefficient from $\nu$-polarized photons to $\mu$-polarized photons; the superscript \textit{in} (\textit{ref}) simply denotes incident (reflection) photons. The explicit expressions of the matrices $\bold M_{\mu\nu}$ are given in Appendix C. In contrast to common PTS materials, the cross reflection coefficients $r_{sp}$ and $r_{ps}$ are usually nonzero for BIMs due to the fact that magnetoelectric effect can mix s- and p- polarizations in general.
Based on the constitutive relations and boundary conditions, one can obtain the cross-reflection coefficients \cite{ILindell}
\begin{eqnarray}\label{rsp}
r_{sp} (r_{ps})=\frac{2\eta_0\eta c_0}{\Delta}\left[\pm i(c_+-c_-)\cos\beta- (c_++c_-)\sin\beta\right]. &{}&\nonumber\\
&{}&
\end{eqnarray}
Here, $\eta=\sqrt{\mu/\epsilon}$ ($\eta_0=\sqrt{\mu_0/\epsilon_0}$) represents the impedance of the BIM (vacuum); $c_0=\cos\theta_0$, where $\theta_0$ is the incident angle of an EM wave; $c_{\pm}=\cos\theta_{\pm}=\sqrt{k_{\pm}^2-k_\rho^2}/k_{\pm}$, where $\theta_{\pm}$ stand for refractive angles and $k_{\pm}=k(\cos\beta\pm \kappa_r)$; $\sin\beta=\chi_r=\chi (\sqrt{\epsilon_0\mu_0}/\sqrt{\epsilon\mu})$ and $\kappa_r=\kappa (\sqrt{\epsilon_0\mu_0}/\sqrt{\epsilon\mu})$ are the relative magnetoelectric parameters; $\Delta=(\eta_0^2+\eta^2)c_0(c_++c_-)+2\eta_0\eta(c_0^2+c_+c_-)\cos\beta$. For lossless media, $k_{\pm}\geq 0$ implies the relationship $\chi_r^2+\kappa_r^2\leq 1$ \cite{RZhao}.
The key element that induces the ACF is the differential Green's functions $\Sigma_{x/z}$, which can be expressed by the cross-reflection coefficients through
\begin{eqnarray}\label{Txz}
\Sigma_{x/z}(\omega)=\frac{\omega}{2\pi}\int_{-\infty}^{\infty}d^2 k_\rho \,e^{2 i \sqrt{w^2-k_\rho^2}d}g_{x/z}(r_{sp}\mp r_{ps}).\nonumber\\
\end{eqnarray}
Here, $g_x={k_x^2}/{\sqrt{\omega^2-k_\rho^2}}$ and $g_z=\sqrt{\omega^2-k_\rho^2}$.
Substituting Eqn. \eqref{Txz} into Eqn. \eqref{eq8} and \eqref{eq9}, one can immediately obtain the ACF $F_{n}$ in an arbitrary direction $\hat n$.
\begin{figure}[!htb]
\centering
\includegraphics[height=5.8cm, width=8.2cm, angle=0]{figure2.eps}
\caption{Numerical calculation of ACF vs. angle $\theta$, rotating frequency $\Omega$ and distance $d$.
Figure (a) shows the normalized ACF $|F_n|_{norm}=|F_n/(F_n)_{max}|$ at different angles. Red, blue, and black curves correspond to Pasteur BIM ($\chi=0$, $\kappa=1$), Tellegen BIM ($\chi=1$, $\kappa=0$), and general BIM ($\chi=0.1$, $\kappa=0.5$), respectively. Other parameters are set as: $d=12 R$, $T=0.01 R_0^{-1}$, $\Omega=10^{-9} R_0^{-1}$, $\omega_p=0.1 R_0^{-1}$, $\tau_0^{-1}=0.01 R_0^{-1}$. Figure (b) gives distance-dependence of ACF. Note that the force values are normalized by $F_0=\hbar c/R^2$. The inset of (c) gives the zoomed-in ACF at $\theta=\pi/2$. Parameters are set as: $\kappa=-0.5$, $\chi=0.1$, and other parameters are the same as those in (a). Figure (c) shows the frequency-dependence of ACF. Red, blue, black curves correspond to angle $\theta=0$, $\theta=\pi/2$, $\theta=\pi/4$, respectively. Parameters are set as: $d=12 R_0$, and other parameters are the same as those in (b). Figure (d) corresponds to the temperature-dependence of ACF at angles $\theta=0$, $\theta=\pi/2$, and $\theta=\pi/4$, respectively.
\label{fig:2}}
\end{figure}
Based on the above formulas, we give the criterion for the emergence of ACF in the following:\\
(\textit{i}) For simple isotropic materials (with PS and TRS), $\chi =\kappa =0$, and one can find $\Sigma_{x}=\Sigma_{z}=0$. Consequently, both $F_x$ and $F_z$ vanish, leading to the vanishing of ACF in any direction. \\
(\textit{ii}) For Pasteur materials (with TRS but without PS), $\chi =0$ and $\kappa \neq0$ lead to $\Sigma_{z}=0$ and $\Sigma_x\neq 0$. As a result, $F_x\neq 0$ and $F_z= 0$ indicate that the ACF only vanishes in the z-direction. \\
(\textit{iii}) For Tellegen materials (with PS but without TRS), $\kappa=0$ but $\chi\neq 0$, and one can show $\Sigma_x=0$ whereas $\Sigma_z\neq 0$, which results in $F_x=0$ and $F_z\neq0$. In this case, the ACF only vanishes in the x-direction. \\
(\textit{iv}) For more general cases (without PS and TRS) where $\chi\neq 0$ and $\kappa \neq 0$, $\Sigma_x\neq 0$ and $\Sigma_z \neq 0$, ACF can persist in any direction.
Based on the above analysis, ACF is a general phenomenon that exists in many materials, including topological materials and chiral materials. [See Table I.]
\section{IV. Numerical Calculation and analytical limit}
We calculate the ACF numerically by considering a rotating particle described by Drude model,
where the electric permittivity is modeled by $\epsilon=\epsilon_b+\frac{\omega_p^2}{\omega(\omega+ i\tau^{-1})}$. In the formula, $\epsilon_b$ is background static electric permittivity, $\tau$ is the scattering time of electrons, and $\omega_p$ is called as plasmonic frequency. The polarizability can be obtained from $\epsilon$, and reads as $\alpha(\omega)= 4\pi R_0^3 \epsilon_0 \frac{\epsilon-\epsilon_0}{\epsilon+2\epsilon_0}$, where $R_0$ is the radius of the particle, and $\epsilon_0$ is the vacuum permittivity. For simplicity, we set $\eta=\eta_0$ in the calculation. The numerical results are summarized in Figure 2. In Figure 2 (a), the angle-dependent ACF is shown, where L-ACF $F_x$ exists at $\kappa\neq 0$, whereas V-ACF $F_z$ exists at $\chi\neq 0$. In general case where $\chi\neq 0$ and $\kappa \neq 0$, both $F_x$ and $F_z$ exist. Figure 2 (b) shows the distance dependence of ACF at different angles. One can see that the decaying behavior of ACF dependents on the angle $\theta$. At $\theta=0$, the L-ACF is two-orders smaller than V-ACF (inset of figure 2(b)). Figure 2 (c) gives the rotating-frequency dependence of ACF, from which one can see the ACF increases linearly with $\Omega$. In figure 2 (d), the temperature-dependence of ACF is shown, where ACF increase with $T$ non-linearly at low temperature, but linearly at high temperature.
We analytically obtain ACF in the low-frequency limit $\omega d \rightarrow 0$, which can be fulfilled at low temperatures due to the differential distribution function $N(T,\omega)$. For Tellegen materials, one has $\chi=1$ and $\kappa=0$ and $r_{sp}=r_{ps}=-1$, which leads to $\Sigma_x=0$ and $\mathrm{Im}\left\{\Sigma_z\right\}=-\omega/2 d^3$. By contrast, for Pasteur materials, $\chi=0$ and $\kappa=1$ results in $r_{sp}=-r_{ps}=-2 i\sqrt{1-s^2}/(2\sqrt{1-s^2}+\sqrt{4-s^2})$, and leads to $\Sigma_z=0$ and $\mathrm{Im}\left\{\Sigma_x\right\}=-4\omega^4/3+3\pi\omega^5d/4$. Further, if $\omega \ll \omega_p$, one can have the imaginary polarizability $\rm Im\{\alpha\}\approx - {12\pi \omega}/{(\omega_p^2\tau)}$. Under realistic conditions, when the rotating frequency is much smaller than plasmonic frequency, one can approximately obtain $\rm Im\{\alpha(\omega_+)\}N(\omega_+)-Im\{\alpha(\omega_-)\}N(\omega_-)\approx 2\Omega \,Im\{\alpha\}\partial_{\omega}N(\omega,T)+\Omega^3\,Im\{\partial_{\omega}\alpha\} \left[\partial_{\omega}^2 N(\omega,T)\right]$ assuming $T_1=T_2=T$. By substituting these approximations into the ACF expression Eqn. \eqref{eq9}, one can readily obtain the analytical expressions of V-ACF and L-ACF:
\begin{eqnarray}
\label{acffz}F_z&=&-\frac{\hbar R_0^3}{\omega_p^2\tau d^3}\left[4\pi^2T^2\Omega+\Omega^3\right]\\
F_x&=&-\frac{16\hbar R_0^3}{\omega_p^2\tau}\left[240\zeta(5) T^5 \Omega+16\zeta(3)T^3\Omega^3\right.\nonumber
\\&&\quad\quad\quad\quad+\left.\frac{2 T}{5} \Omega^5+\frac{\mathrm{Sign}[\Omega]}{30}\Omega^6\right]
\end{eqnarray}
Note that we set parameters $\chi\rightarrow 1$ ($\chi\rightarrow 0$) and $\kappa\rightarrow 0$ ($\kappa\rightarrow 1$) in obtaining analytical result of laterally (vertically) ACF. In figure 3 we compare our analytical result with numerical results showing consistence at low temperature. From the analytical expressions, one can understand why ACF depends on rotating frequency linearly and non-linearly on temperature at low temperatures.
\begin{figure}[!htb]
\centering
\includegraphics[height=3.3cm, width=8.4cm, angle=0]{figure3.eps}
\caption{Comparison between Numerical calculation and analytical calculation of ACF. Figure (a) and (b) show the L-ACF and V-ACF, where analytical result (red curve) and numerical result (blue curve) are consistent at low temperature. The distance is set $d=5 R_0$, while other parameters are the same as that in Figure 2. \label{fig:3}}
\end{figure}
\section{V. Magnetic contribution}
Our previous calculation is based on the dipolar approximation, i.e., the size of the particle is much smaller than the cut-off wave length of photons $1/\omega_p$ and the distance $d$. For a large object, the dipolar approximation is not valid, and one should use the scattering-matrix method to calculate the Casimir force \cite{ACanaguier-Durand}. Notice that, the magnetic contribution is neglected in our previous calculation. The magnetic contribution can be easily included by replacing the electric polarization $\alpha$ with the magnetic polarization $\beta$. Also, the electric Green's tensors $\mathcal G_{ij}$ should be replaced by the magnetic Green's tensors $\mathcal H_{ij}$ which can be easily obtained from the electric counterparts by swapping polarization indices, i.e. $\mathcal H_{ij}=\mathcal G_{ij}(s\leftrightarrow p)$.
We calculate the magnetic polarizability, and show that, for small particles, the magnetic contribution is vanishingly small. The magnetic polarizability for a spherical particle with radius $R$ is given by
$\beta(\omega)=-R^3\left[\frac{1}{2}-\frac{3}{2(\epsilon-1)\omega^2 R^2}+\frac{3}{2(\epsilon-1)\omega R}cot( \sqrt{\epsilon-1} \omega R)\right]$ \cite{AVolokitin2}. With the magnetic polarizability and magnetic Green's tensor, one can calculate the magnetic contribution of the ACF. In figure 4, we numerically calculated the ratio between the electric contribution and the magnetic contribution. Our result shows that the magnetic contribution to the ACF is vanishingly small. In fact, the small magnetic contribution can be understood analytically. In the small particle limit ($\omega_p\,R \ll 1$), the magnetic polarizability becomes $\beta(\omega)\rightarrow -R^3(\omega_p R)^2/30$, and the magnetic and the electric polarization ratio $\beta(\omega)/\alpha(\omega) \ll 1$. This is quite different from the large metallic sphere case where magnetic fluctuations contribute the same order as the electric counterpart \cite{milton1,milton2}.
\begin{figure}[!htb]
\centering
\includegraphics[height=3.6cm, width=5.4cm, angle=0]{figure4.eps}
\caption{Magnetic contribution to the ACF. Red (blue) curve represents the force ratio between the electric L-ACF (V-ACF) and magnetic L-ACF (V-ACF). \label{fig:4}}
\end{figure}
\section{VI. Discussion and comments}
(1) The ACF is parallel with rotating axis, and it does not exert torque on the rotating particle. Consequently, the ACF can not induce the heat transfer between the particle and the BIM plate. Hence, we again illustrate the dissipationless nature of ACF \cite{AManjavacas1}.
(2) Let's compare the ACF with the usual Casimir-Polder force. The usual Casimir-Polder force exists only in z-direction, and depends on $r_{ss}$ and $r_{pp}$. In the Appendix C, we show that the usual Casimir-Polder force can be made as small as possible and even vanished in certain cases. Also, from equation \eqref{acffz}, one can see that V-ACF decays slower that the usual Casimir force ($\propto 1/d^5$).
(3) In this paper, the non-equilibrium effects and spatial dispersion effects are ignored in the calculation, which could become important in some circumstances \cite{DReiche}. The non-inertial effect is also neglected in our approach, which may be interesting for further study \cite{YZeldovich}.
(4) Quantum levitation may be possible by using V-ACF. For example, a particle with radius $R_0=10 \,\rm{nm}$ and density $\rho_0=0.1 \,\rm{g\, cm^{-3}}$ has gravitational force $F_G \approx 4\times 10^{-21}$ N. If it rotates above a BIM plate, the V-ACF can reach $F_n\approx 10 F_G$ for parameters $\Omega = 10^{-5} R_0^{-1}\approx100$ GHz, $d=3 R_0=30$ nm, $\omega_p=0.1 R_0^{-1}\approx 1$ eV, $\tau^{-1}=200$ {meV}, $T=0.03 R_0^{-1}\approx 500$ K, $\chi=1$, and $\kappa=0$. Note that the very recent experiments have already achieved a superfast rotation of nanoparticles, making the ACF within the experimental reach \cite{RReimann}.
\section{Summary}
We have identified the first dissipationless rotation-induced force in vacuum, named axial Casimir force. The axial Casimir force emerges when a particle rotating above a plate that has either time-reversal symmetry breaking or parity-symmetry breaking. Various topological materials and chiral materials are promising candidates to observe the axial Casimir force. Due to V-ACF, quantum levitation is also possible for a particle rotating nearby a BIM plate. Furthermore, the axial Casimir force has a slower scaling law with distance, and can dominate over the common Casimir-Polder force in certain cases.
\section{Acknowledgement}
We are grateful to T. H. Hansson for carefully reviewing our manuscript and many helpful discussions. This work was supported by the Swedish Research Council under Contract No. 335-2014-7424. In addition, FW's work is supported by the U.S. Department of Energy under grant Contract No. DE-SC0012567 and by the European Research Council under grant 742104.
\hspace{2mm}
\begin{center}
\textbf{Appendix A: Derivation of L-ACF}
\end{center}
In this section, we calculate the L-ACF in detail.
The L-ACF is contributed from two pieces including the dipole fluctuation and the field fluctuation. We calculate them separately.
\textit{Dipole fluctuation distribution---}We can compute the first term of Eqn. \eqref{eq4}:
\begin{eqnarray}\label{fxp2}
F_{x,p}=&&\langle p_i^{fl}(t)\partial_x E_i^{ind}(\bold r_0, t)\rangle\nonumber\\
=&&\int_{\infty}^{\infty} \frac{d\omega d\omega^{\prime}}{4\pi^2}\, e^{-i(\omega+\omega^{\prime})t} \langle p_i(\omega)\partial_x E_i(\bold r_0, \omega^{\prime})\rangle\nonumber\\
=&&\int_{\infty}^{\infty} \frac{d\omega d\omega^{\prime}}{4\pi^2}\, e^{-i(\omega+\omega^{\prime})t}\nonumber\\
&&\times \langle p_i(\omega)\partial_x G_{ij}(\bold r_0, \bold r_0, \omega^{\prime})p_j(\omega^{\prime})\rangle,
\end{eqnarray}
where the derivative only acts on the first component of Green function, i.e. $\partial_x G(\bold r_0,\bold r_0,\omega)=\partial_x G(\bold r,\bold r_0,\omega)|_{\bold r=\bold r_0}$. Now, comes to the important part. Based on the FDT, one can connect the quantity $\langle p_i p_j\rangle$ with atomic polarizability $\alpha_{ij}$. Since the particle is rotating, the atomic polarizability is only well defined in the rotating frame. In the expression of Eqn. \eqref{fxp2}, $p_i$ is the electric dipoles that are defined in the lab frame. In order to express electric polarizability in terms of electric dipoles $\tilde p_{i}$ in the rotating frame, one needs the coordinate transformation \cite{AManjavacas1}
\begin{eqnarray}\label{px}
&&p_x(\omega)=\tilde p_{x}(\omega)\nonumber\\&&
p_y(\omega)=\frac{1}{2}\left[ \tilde p_{y}(\omega_+)+i\tilde p_{z}(\omega_+)+\tilde p_{y}(\omega_-)-i\tilde p_{z}(\omega_-)\right]\nonumber\\&&
p_z(\omega)=\frac{1}{2}\left[-i \tilde p_{y}(\omega_+)+\tilde p_{z}(\omega_+)+i\tilde p_{y}(\omega_-)+\tilde p_{z}(\omega_-)\right]\nonumber\\
\end{eqnarray}
where $\omega_{\pm}=\omega\pm\Omega$ is the Doppler-shifted frequency with $\Omega$ denotes the rotation frequency of the particle. In the main text, we expressed the coordinate transformation of electric dipole by the form: $p_i(\omega)=\Lambda_{ij}^+\tilde p_j(\omega_+)+\Lambda_{ij}^0\tilde p_j(\omega)+\Lambda_{ij}^{-} \tilde p_j(\omega_-)$. Now, one can read out $\Lambda_{ij}^{\pm}$ and $\Lambda_{ij}^0$ from above equations. Due to the translational symmetry in x, y and z direction of the surface, the surface Green function satisfies $\partial_x G_{xx}=\partial_x G_{yy}=\partial_x G_{zz}=0$. One can find the explicit expression of surface Green's tensor in \textbf{Appendix C}, where we show that only the terms $\partial_x G_{yz}$, $\partial_x G_{zy}$, $\partial_xG_{xz}$, and $\partial_xG_{zx}$ need to be calculated. (Due to the isotropic assumption of the rotating particle ($\alpha_{xy}=\alpha_{xz}=0$), we do not need to calculate $\partial_x G_{xz}$ and $\partial_x G_{xy}$, which alway appear, respectively, with $\alpha_{xy}$ and $\alpha_{xz}$ at the same time.) The L-ACF induced from dipole fluctuation is
\begin{eqnarray}\label{fxp}
F_{x,p}
=&&\int_{-\infty}^{\infty} \frac{d\omega d\omega^{\prime}}{4\pi^2}\left\{\left[\langle p_y(\omega)\partial_x G_{yz}(\bold r_0, \bold r_0, \omega^{\prime})p_z(\omega^{\prime})\rangle\right]\right.\nonumber\\
&&\left.+\left[\langle p_z(\omega)\partial_x G_{zy}(\bold r_0, \bold r_0, \omega^{\prime})p_y(\omega^{\prime})\rangle\right]\right\}\,\delta(\omega+\omega^{\prime})\nonumber\\
=&&\int_{-\infty}^{\infty} \frac{d\omega }{4\pi^2}\, \left[\langle p_y(\omega)\partial_x G_{yz}(\bold r_0, \bold r_0, -\omega)p_z(-\omega)\rangle\right]\nonumber\\
&&+\left[\langle p_z(\omega)\partial_x G_{zy}(\bold r_0, \bold r_0, -\omega)p_y(-\omega)\rangle\right].
\end{eqnarray}
Substitute Eqn. \eqref{px} into Eqn. \eqref{fxp} , and one can obtain
\begin{eqnarray}
\langle p_y &&(\omega) \partial_x G_{yz}(\bold r_0, \bold r_0, -\omega)p_z(-\omega)\rangle\nonumber\\
=&&\frac{1}{4}\partial_x G_{yz}^*\langle\left[{{\tilde p_{y}(\omega_+)}}+{{i\tilde p_z(\omega_+)}}+{{\tilde p_y(\omega_-)}}-i\tilde p_z(\omega_-)\right]\nonumber\\
&&\times\left[{{-i \tilde p_y(\omega_+^{\prime})}}+\tilde p_z(\omega_+^{\prime})+{{i\tilde p_y(\omega_-^{\prime})}}+{{\tilde p_z(\omega_-^{\prime})}}\right]\rangle\nonumber\\
=&&i \pi\hbar\partial_x G_{yz}^*\left\{\mathrm{Im}\left\{{{\tilde \alpha_{yy}(\omega_+)}}+{{\tilde \alpha_{zz}(\omega_+)}}\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\right.\nonumber\\
&&\left.-\mathrm{Im}\left\{{{\tilde \alpha_{yy}(\omega_-)}}+\tilde \alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\},
\end{eqnarray}
where $\omega_\pm=\omega\pm\Omega$, $\omega_\pm^{\prime}=-\omega\pm\Omega$, and $T_1$ is the particle temperature.
In the same way, one can obtain
\begin{eqnarray}
\langle &&p_z(\omega)\partial_x G_{zy}(\bold r_0, \bold r_0, -\omega)p_y(-\omega)\rangle\nonumber\\
=&&-i \pi\hbar\partial_x G_{zy}^*
\left\{\mathrm{Im}\left\{\tilde \alpha_{yy}(\omega_+)+\tilde \alpha_{zz}(\omega_+)\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\right.\nonumber\\
&&\left.-\mathrm{Im}\left\{\tilde \alpha_{yy}(\omega_-)+\tilde \alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\}.
\end{eqnarray}
Several comments in order: (i) Since the electric field $\bold E(t)$ is real, the Green function $G_{ij}(t)=\langle E_i(t)E_j(0)\rangle$ is also real, i.e., $G_{ij}^*(t)=G_{ij}(t)$. Due to the expression of Green function in $\omega$-space, $G_{ij}(t)=\int \frac{d\omega}{2\pi} e^{-i\omega t}G_{ij}(\omega)$ we can obtain $G^*(\omega)=G(-\omega)$. The same reason also suggests $\alpha^*(\omega)=\alpha(-\omega)$.
(ii) When we make the simplification $\int_{-\infty}^{\infty}\mapsto\int_{0}^{\infty}$, it's not as simple as $\int_{-\infty}^{\infty}=2\int_{0}^{\infty}$. In fact, we should use the equality
\begin{eqnarray}
\int_{-\infty}^0 &&d\omega ~\partial_x G_{yz}^{*}(\omega){\rm Im}\{\alpha(\omega_+)\}\left(n(\omega_+)+\frac{1}{2}\right)\nonumber\\
=&&\int_{0}^{\infty} d\omega\partial_x G_{yz}^{*}(-\omega){\rm Im}\{\alpha(-\omega_-)\}\left(n(-\omega_-)+\frac{1}{2}\right)\nonumber\\
=&&\int_{0}^{\infty} d\omega\partial_x G_{yz}(\omega){\rm Im}\{\alpha(\omega_-)\}\left(n(\omega_-)+\frac{1}{2}\right).
\end{eqnarray}
Note that the FDT in the rotating frame leads to the relation $\langle \tilde p_i(\omega)\tilde p_j(\omega^{\prime})\rangle=2\pi\hbar \delta(\omega+\omega^{\prime})\mathrm{Im}\left\{\tilde\alpha_{ij}(\omega)\right\}coth(\frac{\beta \omega}{2})$.
Therefore, the Casimir force $F_{x,p}$ due to electric dipole fluctuation is
\begin{eqnarray}
F_{x,p}=&&\frac{i\hbar}{4\pi}\int_{-\infty}^{\infty} d\omega (\partial_x G_{yz}^*-\partial_x G_{zy}^*)\times\nonumber\\
&&\left\{\mathrm{Im}\left\{\tilde\alpha_{yy}(\omega_+)+\tilde\alpha_{zz}(\omega_+)\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\right.\nonumber\\
&&\left.-\mathrm{Im}\left\{\tilde\alpha_{yy}(\omega_-)+\tilde\alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\}\nonumber
\end{eqnarray}
One can express the above result only in the frequency region $\omega \geq 0$ by using ${\rm Im}\{\tilde\alpha(-\omega)\}=\rm{Im}\{\tilde\alpha^*(\omega)\}=-{\rm Im}\{\tilde\alpha(\omega)\}$ and $\left(n(T_1,-\omega)+\frac{1}{2}\right)=-\left(n(T_1,\omega)+\frac{1}{2}\right)$:
\begin{eqnarray}
F_{x,p}
=&&\frac{\hbar}{2\pi}\int_{0}^{\infty} d\omega \mathrm{Im}\left\{\partial_x G_{yz}-\partial_x G_{zy}\right\}\times\nonumber\\
&&\left\{\mathrm{Im}\left\{\tilde\alpha_{yy}(\omega_+)+\tilde\alpha_{zz}(\omega_+)\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\right.\nonumber\\
&&\left.-\mathrm{Im}\left\{\tilde\alpha_{yy}(\omega_-)+\tilde\alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\}\nonumber
\end{eqnarray}
\textit{Field fluctuation distribution---}One can compute the second term of Eqn. \eqref{eq4} induced by the electric field fluctuation.
\begin{eqnarray}\label{fxe}
F_{x,E}
=&&\langle {\alpha}_{ij} E_j^{fl}(\bold r_0, t) \partial_x E_i^{fl}(\bold r_0, t) \rangle\nonumber\\
=&&\int_{-\infty}^{\infty}\frac{d\omega d\omega^{\prime}}{(2\pi)^2} e^{-i(\omega+\omega^{\prime})t}\langle {\alpha}_{ij}(\omega)E_j^{fl}(\omega)\partial_x E_i^{fl}(\omega^{\prime})\rangle,\nonumber\\
\end{eqnarray}
where ${\alpha}_{ij}$ denotes the effective polarizability seen in the lab frame, corresponding to the polarizability $\tilde \alpha$ in its rotating frame via \cite{AManjavacas1}:
\begin{eqnarray}
{\alpha}_{xx}(\omega)=&& \tilde\alpha_{xx}(\omega)\nonumber\\
{\alpha}_{yy}(\omega)=&&\frac{1}{4}\left(\tilde\alpha_{yy}(\omega_+)+\tilde\alpha_{zz}(\omega_+)+\tilde\alpha_{yy}(\omega_-)+\tilde\alpha_{zz}(\omega_-)\right)\nonumber\\
=&&{\alpha}_{zz}(\omega)\nonumber\\
{\alpha}_{yz}(\omega)=&&\frac{i}{4}\left(\tilde\alpha_{yy}(\omega_+)+\tilde\alpha_{zz}(\omega_+)-\tilde\alpha_{yy}(\omega_-)-\tilde\alpha_{zz}(\omega_-)\right),\nonumber\\
=&&-{\alpha}_{zy}(\omega)
\end{eqnarray}
from which one can read out $\Gamma_{ijkl}^{\pm}$ and $\Gamma_{ijkl}^{0}$ in the main text. Substituting $\tilde{\alpha}_{yz}(\omega)$ and $\langle E_z^{fl}(\omega)\partial _x E_y^{fl}(\omega^{\prime})\rangle=\langle \partial _x E_y^{fl}(\omega^{\prime}, \bold r_0)E_z^{fl}(\omega,\bold r_0)\rangle=4\pi\delta(\omega+\omega^{\prime})\mathrm{Im}\{\partial_xG_{yz}(\bold r_0, \bold r_0,\omega^{\prime})\} \left(n(T_2,\omega^{\prime})+\frac{1}{2}\right)$ ( $T_2$ is the temperature of the surface.) into Eqn. \eqref{fxe}, one can get
\begin{eqnarray}
F_{x,E}=&&\langle p_i^{ind}(t)\partial_x E_i^{fl}(\bold r_0, t)\rangle\nonumber\\
=&&-\frac{\hbar}{2\pi}\int_{0}^{\infty} d\omega ~\mathrm{Im}\{\partial_x G_{yz}-\partial_xG_{zy}\}\nonumber\\
&&\times\mathrm{Im}\{\tilde\alpha_{yy}(\omega_+)+\tilde\alpha_{zz}(\omega_+)-\tilde\alpha_{yy}(\omega_-)-\tilde\alpha_{zz}(\omega_-)\}\nonumber\\
&&\times\left[n(T_2,\omega)+\frac{1}{2}\right].
\end{eqnarray}
Combine the force induced by dipole fluctuation and field fluctuation distribution, and we can obtain the final expression of ACF:
\begin{eqnarray}
F_x=&&\frac{\hbar}{2\pi}\int_{0}^{\infty} d\omega\, \mathrm{Im}\left\{\partial_x G_{yz}-\partial_x G_{zy}\right\}\nonumber\\
&&\left[\mathrm{Im}\left\{\tilde\alpha_{yy}(\omega_+)+\tilde\alpha_{zz}(\omega_+)\right\}N(\omega_+)\right.\nonumber\\
&&\left.- \mathrm{Im}\left\{\tilde\alpha_{yy}(\omega_-)+\tilde\alpha_{zz}(\omega_-)\right\}N(\omega_-)\right],
\end{eqnarray}
where $N(\omega_\pm)=n(T_1, \omega_\pm)-n(T_2, \omega)$. Even at zero temperature $T_1=T_2=0$, $N(\omega_\pm)\neq0$ which indicates that the ACF is totally contributed from quantum fluctuation. By assuming the particle is isotropic, i.e., $\alpha_{ij}=\alpha \delta_{ij}$, the formula in the main text is obtained.
\vspace{2mm}
\begin{center}
\textbf{Appendix B: Derivation of Axial Casimir force (ACF) in Arbitrary Direction.}
\end{center}
In this section, we show how to calculate the ACF of a particle rotating along an arbitrary axis $\hat n=(\cos\theta, 0,\sin\theta)$. We calculate the Casimir force in x, y, z directions, respectively, and project them along the rotating axis. $(p_x^{\theta}, p_y^{\theta}, p_z^{\theta})$ represents electric dipole of the particle in the lab frame, which can be obtained via coordinate rotation:
\begin{eqnarray}
\left(\begin{array}{cc}
p_x^{\theta}\\ p_z^{\theta}
\end{array}\right)=
\left(\begin{array}{cc}
\cos\theta & -\sin\theta\\
\sin\theta & \cos\theta
\end{array}\right)\left(\begin{array}{cc}
p_x\\ p_z
\end{array}\right),
\end{eqnarray}
where $p_x$ and $p_z$ are the electric dipoles that are obtained in last section. The polarization at $\hat n=(\cos\theta, 0, \sin\theta)$ corresponds to the polarization at $\hat n=(1,0,0)$ via
\begin{eqnarray}
\alpha_{xx}^{\theta}=&&\cos^2\theta {\alpha}_{xx}+\sin^2\theta {\alpha}_{zz}; \,
\alpha_{yy}^{\theta}=\langle p_y p_y\rangle= {\alpha}_{yy}; \nonumber\\
\alpha_{zz}^{\theta}=&&\cos^2\theta {\alpha}_{zz} +\sin^2\theta \alpha_{xx} ;\,
\alpha_{xy}^{\theta}=-\sin\theta {\alpha}_{zy}; \nonumber\\
\alpha_{yx}^{\theta}=&&-\sin\theta {\alpha}_{yz};
\alpha_{xz}^{\theta}=\sin\theta \cos\theta {\alpha}_{xx}-\sin\theta \cos\theta {\alpha}_{zz} \nonumber\\
\alpha_{zx}^{\theta}=&&\alpha_{xz}^{\theta}; \,
\alpha_{yz}^{\theta}=\cos\theta {\alpha}_{yz}; \,
\alpha_{zy}^{\theta}=\cos\theta {\alpha}_{zy},
\end{eqnarray}
In the following, we obtain the Casimir force (induced by rotation) in the x-direction $F_x$ and in the z-direction $F_z$, respectively. Then, the total ACF is \begin{eqnarray}
F_n(\theta)=F_x(\theta)\cos\theta+F_z(\theta)\sin\theta.
\end{eqnarray}
\textit{The axial Casimir force in the x-direction---} Notice that the L-ACF calculated here is different from that in \textbf{Appendix A}. Because the rotating axis is not parallel with the plate anymore, i.e., $F_x(\theta)\neq F_x$, thus one needs to re-calculate the L-ACF in this case. Again, the L-ACF is induced from two parts contributions: (i) the electric dipole fluctuation and (ii) the electric field fluctuation, i.e.,
$F_x(\theta)=F_{x,p}+F_{x,E},$
where
\begin{eqnarray}
F_{x,p}=&&\int_{-\infty}^{\infty}\frac{d\omega d\omega^{\prime}}{(2\pi)^2}e^{-i(\omega+\omega^{\prime})t}\nonumber\\
&&\times\langle p_i^{\theta}(\omega)\partial_x G_{ij}(\bold r_0,\bold r_0, \omega^{\prime}) p_j^{\theta}(\omega^{\prime})\rangle;\\
F_{x,E}=&&\int_{-\infty}^{\infty}\frac{d\omega d\omega^{\prime}}{(2\pi)^2} e^{-i(\omega+\omega^{\prime})t}\langle \alpha^{\theta}_{ij}(\omega)E_j^{fl}(\omega)\partial_x E_i^{fl}(\omega^{\prime})\rangle.\nonumber\\
\end{eqnarray}
Substituting the electric dipole into the expression of the dipole-induced Casimir force, and one can obtain
\begin{eqnarray}
F_{x,p}=\frac{\hbar}{2\pi}\int_0^{\infty}d\omega \left[\sin\theta \cos\theta f_{p1}+\cos\theta f_{p2}-\sin\theta f_{p3}\right],\nonumber\\
\end{eqnarray}
where
\begin{eqnarray}\label{fp1}
f_{p1}=&&{\rm Re}\,\left\{(\partial_xG_{xz}+\partial_xG_{zx})\right\}\nonumber\\
&&\times\left\{4\,{\rm Im}\{\tilde \alpha_{xx}(\omega) \}\left(n(T_1,\omega)+\frac{1}{2}\right)\right.\nonumber\\
&&-{\rm Im}\left\{\tilde \alpha_{yy}(\omega_{+})+\tilde \alpha_{zz}(\omega_+)\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\nonumber\\
&&\left.-{\rm Im}\left\{\tilde \alpha_{yy}(\omega_{-})+\tilde \alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\}.\nonumber\\
\end{eqnarray}
\begin{eqnarray}\label{fp2}
f_{p2}=&&{\rm Im}\left\{(\partial_xG_{yz}-\partial_xG_{zy})\right\}\nonumber\\
&&\times\left\{{\rm Im}\left\{\tilde \alpha_{yy}(\omega_{+})+\tilde \alpha_{zz}(\omega_+)\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\right.\nonumber\\
&&\left.-{\rm Im}\left\{\tilde \alpha_{yy}(\omega_{-})+\tilde \alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\};\nonumber\\
\end{eqnarray}
\begin{eqnarray}
\label{fp3}
f_{p3}=&&{\rm Im}\left\{(\partial_xG_{yx}-\partial_xG_{xy})\right\}\nonumber\\
&&\times\left\{{\rm Im}\left\{\tilde \alpha_{yy}(\omega_{+})+\tilde \alpha_{zz}(\omega_+)\right\}\left(n(T_1,\omega_+)+\frac{1}{2}\right)\right.\nonumber\\
&&\left.-{\rm Im}\left\{\tilde \alpha_{yy}(\omega_{-})+\tilde \alpha_{zz}(\omega_-)\right\}\left(n(T_1,\omega_-)+\frac{1}{2}\right)\right\}.\nonumber\\
\end{eqnarray}
The electric fluctuating field contribution to the lateral Casimir force is
\begin{eqnarray}\label{fye}
F_{x,E}=\frac{\hbar}{2\pi}\int_0^{\infty} d\omega \left[\sin\theta \cos\theta f_{E1}+\cos\theta f_{E2}- \sin\theta f_{E3}\right],\nonumber\\
\end{eqnarray}
where
\begin{eqnarray}\label{fE1}
f_{E1}
=&& \rm {Im}\left\{\partial_x G_{xz}+\partial_xG_{zx}\right\} {\rm Re}\left\{ 4\tilde \alpha_{xx}\right.-\left(\tilde \alpha_{yy}(\omega_+)
+\tilde \alpha_{zz}(\omega_+)\right.\nonumber\\
&&\left.\left.+\tilde \alpha_{yy}(\omega_-)+\tilde \alpha_{zz}(\omega_-)\right)\right\}\times\left(n(T_2,\omega)+\frac{1}{2}\right)\nonumber\\
f_{E2}=&&-{\rm Im}\{\partial_x G_{yz}-\partial_x G_{zy}\}\,{\rm Im}\left\{\tilde \alpha_{yy}(\omega_+)+\tilde \alpha_{zz}(\omega_+)\right.\nonumber\\
&&\left.-\tilde \alpha_{yy}(\omega_-)-\tilde \alpha_{zz}(\omega_-)\right\}
\times\left(n(T_2,\omega)+\frac{1}{2}\right)\nonumber\\
f_{E3}
=&&-{\rm Im}\{\partial_x G_{yx}-\partial_x G_{xy}\}\,{\rm Im}\left\{\tilde \alpha_{yy}(\omega_+)+\tilde \alpha_{zz}(\omega_+)\right.\nonumber\\
&&\left.-\tilde \alpha_{yy}(\omega_-)-\tilde \alpha_{zz}(\omega_-)\right\}\left(n(T_2,\omega)+\frac{1}{2}\right)
\end{eqnarray}
Therefore, the total L-ACF in x direction is
\begin{eqnarray}
F_x(\theta)=&&\frac{\hbar}{2\pi}\int_0^{\infty} d\omega \left[\sin\theta \cos\theta \left(f_{p1}+ f_{E1}\right)\right.\nonumber\\
&&\left.+\cos\theta \left(f_{p2}+f_{E2}\right)- \sin\theta \left(f_{p3}+f_{E3}\right)\right]
\end{eqnarray}
\textit{The Casimir force in the y-direction $F_{y}$---} The force in the y-direction has the same form as that in the x-direction. The only difference is that all derivatives on Green tensors changes from $\partial_x G_{ij}$ to $\partial_y G_{ij}$. If we let $\theta=0$, the expression coincides with the expressions in reference \cite{AManjavacas}.
\textit{The Casimir force in the z-direction $F_z$---} The form of $F_z$ is different from $F_x$ and $F_y$ due to the non-vanishing diagonal terms $\partial_zG_{ii}$. We can write the Casimir force in the form
$F_{z}^{tot}(\theta)=F_z^d(\theta)+F_z(\theta)$,
where $F_z^d(\theta)$ is the Casimir force due to the diagonal terms $\partial_zG_{ii}$, and $F_z(\theta)$ is the rotation-induced Casimir force. $F_z(\theta)$ has the similar form as $F_x(\theta)$ and $F_y(\theta)$, and one can obtain $F_z(\theta)$ by the substitution $\partial_{x/y} G_{ij}\rightarrow \partial_z G_{ij}$. The diagonal Casimir force $F_z^d(\theta)$ corresponds to the usually referred Casimir-Polder force, whereas the off-diagonal Casimir force is induced by rotation.
The total ACF along the rotating axis is
\begin{eqnarray}\label{appfn}
F_n=&&F_x(\theta) \cos\theta+F_z(\theta) \sin\theta\nonumber\\
\approx&& F_x\cos^2\theta +F_z\sin^2\theta,
\end{eqnarray}
where $F_x\equiv F_x(\theta=0)$ and $F_z\equiv F_z(\theta=\pi/2)$. In deriving Eqn. \eqref{appfn}, we have used the approximation that $\Omega/\omega \ll 1$. We also use the fact that $f_{p2}+f_{E2}$ and $f_{p3}+f_{E3}$ vanish for $F_z(\theta)$ and $F_x(\theta)$, respectively. (The reason relies on the Green's tensor form in Appendix C.) Note that we derived the formula Eqn. \eqref{eq8} announced in the main text.
\\
In the following, one can derive the diagonal term $F_z^d(\theta)$ by considering the dipole fluctuation and the field fluctuation, respectively.
The Casimir force in the z-direction induced by dipole fluctuation is
\begin{eqnarray}
F_{z,p}^d=&&\int_{-\infty}^{\infty}\frac{d\omega d\omega^{\prime}}{(2\pi)^2}e^{-i(\omega+\omega^{\prime})t}\left\{\langle p_x^{\theta}(\omega)\partial_z G_{xx}(\bold r_0,\bold r_0,\omega^{\prime}) p_x^{\theta}(\omega^{\prime})\rangle\right.\nonumber\\
&&\left.+\langle p_y^{\theta}\partial_z G_{yy} p_y^{\theta}\rangle+\langle p_z^{\theta}\partial_z G_{zz} p_z^{\theta}\rangle\right\}
\end{eqnarray}
The Casimir force in the z-direction induced by field fluctuation is
\begin{eqnarray}
F^d_{z,E}
=&&\int_{-\infty}^{\infty}\frac{d\omega d\omega^{\prime}}{(2\pi)^2} e^{-i(\omega+\omega^{\prime})t} (4\pi \hbar\delta(\omega+\omega^{\prime}))\left\{\alpha_{xx}^{\theta} {\rm Im}\{ \partial_z G_{xx}\}\right.\nonumber\\
&&\left.+ \alpha_{yy}^{\theta} {\rm Im}\{\partial_z G_{yy}\}+\alpha_{zz}^{\theta} {\rm Im}\{ \partial_z G_{zz}\}\right\}\left(n(T_2,\omega)+\frac{1}{2}\right).\nonumber\\
\end{eqnarray}
Unlike the the case rotation-induced Casimir force, the signs before the integral in the expressions of $F^d_{z,p}$ and $F^d_{z,E}$ are the same. This is definitely reasonable, meaning that, even without rotation, these two terms still exist.
Add up the dipole contribution and the field contribution, We get the total diagonal Casimir force in the z-direction:
\begin{eqnarray}
F_z^d=&&F_{z,p}^d+F_{z,E}^d\nonumber\\
\approx &&\frac{2\hbar}{\pi}\int_0^{\infty}d\omega\, \mathrm{Im}\left\{\partial_z(G_{xx}+G_{yy}+G_{zz})\times\alpha(\omega)\right\}\nonumber\\
&&\times \left[n(T,\omega)+\frac{1}{2}\right],
\end{eqnarray}
where following assumptions is implied: equal temperature $T_1=T_2$, isotropic polarizability $\alpha_{ij}=\alpha\,\delta_{ij}$, and $\left[\alpha(\omega_+)\left(n(T,\omega_+)+\frac{1}{2}\right)+\alpha(\omega_-)\left(n(T,\omega_-)+\frac{1}{2}\right)\right]\approx 2\alpha(\omega)\left[n(T,\omega)+\frac{1}{2}\right]$ for $\Omega \ll k_B T$. In the limit of $T\rightarrow 0$, the diagonal Casimir force $F_z^d=\frac{\hbar}{\pi} \int_0^{\infty} d\omega\,\, \mathrm{Tr}\left[\alpha_{ij}\partial_zG_{ij}\right]$, which agrees with the Casimir-Polder formula in Ref. \cite{FIntravaia}. According to \textbf{Appendix C}, the differential Green's function reads $\partial_z(G_{xx}+G_{yy}+G_{zz})=r_{ss}k^2+r_{pp}(k_\rho^2-k_z^2)$. For a metallic surface, $r_{ss}=-r_{pp}$, and $ \partial_z(G_{xx}+G_{yy}+G_{zz})=-\frac{1}{2\pi}\int dk_x dk_y\,e^{2i k_z z}\,r_{ss}(2k_z^2)$, consequently, the above formula become consistent with the Casimir-Polder formula in Refs. \cite{AVolokitin2} and \cite{FIntravaia}.
\begin{center}
\textbf{Appendix C: Surface Green's Tensor For BIM Plate.}
\end{center}
\textit{ General expression of surface Green's tensor---}The surface Green's tensor can be derived from the knowledge of Fresnel coefficients, i.e., $r_{ss}$ and $r_{pp}$ in usual cases, where $r_{ss}$ ($r_{pp}$) stands for the reflection coefficients from TE(TM) wave to TE(TM) wave \cite{lnovothy}. However, when there is a mix between TE wave and TM wave, the surface Green's tensor also depends on $r_{sp}$ and $r_{ps}$ \cite{JCrosse}, where $r_{ps}$ ($r_{sp}$) stands for the cross-reflection coefficients from TE (TM) wave to TM (TE) wave. We then give the general expression of the surface Green function by taking $r_{ps}$ ($r_{sp}$) into consideration
\begin{eqnarray}
\mathds G(\bold r,\bold r^{\prime},\omega)&&=\frac{i}{2\pi}\int\,\frac{d k_x dk_y}{k_z}\,e^{ik_x(x-x^{\prime})}e^{ik_y(y-y^{\prime})}e^{ik_z(z+z^{\prime})}\nonumber\\
&&\times\left[r_{ss}M_{ss}+r_{pp}M_{pp}+r_{sp}M_{sp}+r_{ps}M_{ps}\right],
\end{eqnarray}
where \cite{JCrosse}
\begin{eqnarray}
{M}_{ss}&=&
\left(\begin{array}{ccc}
\frac{k_y^2}{k_\rho^2}k^2&-\frac{k_xk_y}{k_\rho^2}k^2 & 0\\
-\frac{k_xk_y}{k_\rho^2}k^2 & \frac{k_x^2}{k_\rho^2}k^2 & 0\\
0 & 0 & 0
\end{array}\right),
\\
{M}_{pp}&=&
\left(\begin{array}{ccc}
-\frac{k_x^2k_z^2}{k_\rho^2}&-\frac{k_xk_yk_z^2}{k_\rho^2} & -k_xk_z\\
-\frac{k_xk_yk_z^2}{k_\rho^2} & -\frac{k_y^2k_z^2}{k_\rho^2} & -k_yk_z\\
k_xk_z & k_yk_z & k_\rho^2
\end{array}\right),
\\
{M}_{sp}&=&
\left(\begin{array}{ccc}
\frac{k_xk_yk_z}{k_\rho^2}k&\frac{k_y^2k_z}{k_\rho^2}k & k_yk\\
-\frac{k_x^2k_z}{k_\rho^2}k & -\frac{k_xk_yk_z}{k_\rho^2}k & -k_xk\\
0 & 0 & 0
\end{array}\right),
\\
{M}_{ps}&=&
\left(\begin{array}{ccc}
-\frac{k_xk_yk_z}{k_\rho^2}k&\frac{k_x^2k_z}{k_\rho^2}k & 0 \\
-\frac{k_y^2k_z}{k_\rho^2}k & \frac{k_xk_yk_z}{k_\rho^2}k & 0 \\
k_y k & -k_xk & 0
\end{array}\right).
\end{eqnarray}
And the Fresnel reflection coefficients are given by \cite{ILindell}
\begin{eqnarray}
r_{ss}=&&\frac{1}{\Delta}\left\{(\eta^2-\eta_0^2)c_0(c_++c_-)+2\eta_0\eta (c_0^2-c_{+}c_{-})\cos\beta\right\};\nonumber\\
r_{pp}=&&\frac{-1}{\Delta}\left\{(\eta^2-\eta_0^2)c_0(c_++c_-)-2\eta_0\eta (c_0^2-c_{+}c_{-})\cos\beta\right\};\nonumber\\
r_{sp}=&&\frac{2\eta_0\eta c_0}{\Delta}\left[i(c_{+}-c_{-})\cos\beta-(c_{+}+c_{-})\sin\beta\right];\nonumber\\
r_{ps}=&&\frac{-2\eta_0\eta c_0}{\Delta}\left[i(c_{+}-c_{-})\cos\beta+(c_{+}+c_{-})\sin\beta\right],
\end{eqnarray}
Corresponding definitions e.g., $k_\rho$, $\Delta$, etc. are the same as those in the main text.
Let's check the trivial case for an ideal metal plate, where $\chi=\kappa=0$, and $\eta=\sqrt{\mu/\epsilon}\rightarrow 0$. In this case, $r_{pp}=-r_{ss}=1$ and $r_{sp}=r_{ps}=0$.
In another interesting case, by assuming $\eta \mapsto \eta_0$ and $\beta \mapsto \pi/2$ (perfect Tellegen Materials), $r_{ss}=r_{pp}=0$ leads to the vanishing of the usual Casimir-Polder Force.
\textit{ Symmetry Analysis---}We analyze the TRS and PS of the system, and demonstrate why the L-ACF (V-ACF) vanishes for systems with PS (TRS).
Our demonstration is based on Figure \ref{figuresupp1} (a) and (b). We use four kinds of style lines (solid and dashed, blue and red) to represent the plate, in order to capture the chiral nature of the plate. With our definition of ACF (Eqn. (3) and (4) in the main text), it must satisfy
\begin{eqnarray}
F_{n}(\Omega)=-F_n(-\Omega).
\end{eqnarray}
When the rotating axis is parallel with the plate, the parity operation (with respect to $z=0$ plane) will transform left figure of (a) to the middle figure of (a), which is equivalent to right figure of (a). One can find that, if there is parity symmetry of the plate (solid lines = dashed lines), then one has $F_x(\Omega)=F_x(-\Omega)$. Therefore, the only solution of this equation is $F_x(\Omega)=0$, i.e., L-ACF vanishes. Analogously, if there is time-reversal symmetry of the plate, one can show that $F_z(\Omega)=F_z(-\Omega)$. And the only solution is $F_z(\Omega)=0$, i.e., V-ACF vanishes.
\begin{figure*}[!htb]
\centering
\includegraphics[height=4.8cm, width=10.6cm, angle=0]{figuresupp1.eps}
\caption{Symmetry analysis of L-ACF and V-ACF. Figure (a) and (b) show that the L-ACF and V-ACF must, respectively vanish for materials with PS and TRS. $\odot$ and $\otimes$ represent the rotating direction of the particle, perpendicular out of plane or into plane. $\uparrow$ and $\downarrow$ mean that the rotating direction of the particle is upward or downward. \label{figuresupp1}}
\end{figure*}
\vspace{2mm}
\begin{center}
\textbf{Appendix D: Derivation of analytical limit of axial Casimir force}
\end{center}
In this part, we give the detail derivation of the ACF for Tellegen and Pasteur materials in the low frequency limit $\omega d\rightarrow 0$.
For Tellegen materials, $\chi\rightarrow 1$ and $\kappa\rightarrow 0$. Thus, the reflection coefficients can be obtained
$r_{sp}=r_{ps}\rightarrow -1$ leading to $\Sigma_x=0$ and
\begin{eqnarray}
\Sigma_z
=-2\omega^4 \int_0^{\infty} ds (s\sqrt{1-s^2}\,\,e^{2 i \sqrt{1-s^2}\omega d})
\end{eqnarray}
In the limit $\omega d\rightarrow 0$, one can obtain $\mathrm{Im}\left\{\Sigma_z\right\}\approx-{\omega}/{2d^3}$.
For Pasteur materials, $\chi\rightarrow 0$ and $\kappa\rightarrow 1$. Thus, the reflection coefficients can be obtained
$r_{sp}=-r_{ps}\rightarrow \frac{- 2 i \sqrt{1-s^2}}{2\sqrt{1-s^2}+\sqrt{4-s^2}}$ leading to $\Sigma_z=0$ and
\begin{eqnarray}
\Sigma_x
=\int_{0}^{\infty} ds\,\,\frac{s^3~e^{2 i \sqrt{1-s^2}\omega d}}{\sqrt{1-s^2}}\left[\frac{{(-2 i)\omega^4}~\sqrt{1-s^2}}{2\sqrt{1-s^2}+\sqrt{4-s^2}}\right].
\end{eqnarray}
In the limit $\omega d\rightarrow 0$, one can obtain $\mathrm{Im}\left\{\Sigma_x\right\}\approx-4 \omega^4/3+3\pi \omega^5 d/4\approx -4 \omega^4/3$.
Under the assumption $\omega \ll \omega_p$, the imaginary part of the polarizability reads $\mathrm{Im} \left\{\alpha(\omega)\right\}\approx -12\pi R_0^3 \omega/\omega_p^2\tau$.
|
\section{Introduction}
\label{sec:intro}
The need for regrasp is ubiquitous in assembly operations. In-hand manipulation skills allow one to pick up a part, reorient it in the hand, and then use/insert/assemble it. Robots, especially those with parallel-jaw grippers, lack the dexterity to autonomously adjust the grasp on an object. They are often assisted by systems such as part feeders and fixtures. Part feeders take care of reorienting the parts, while customized fixtures allow robots to regrasp through place-and-pick maneuvers. This approach leads to highly-customized, non-flexible, and expensive automation solutions.
This paper presents an in-hand manipulation technique that allows a robot with a simple parallel-jaw gripper to regrasp an object without any customized hardware. The robot pushes the object against external contact/s in the environment such that the object sticks to the external contact while the fingers slide over the object. Since the object is held stationary in the environment without any fixtures, we call this approach \emph{fixtureless fixturing}.
Fixtureless fixturing is a variant of prehensile pushing that is robust against uncertainties in the system parameters. In our prior work on prehensile pushing, we present an idea of manipulating an object with external pushes~\cite{ChavanDafle2014,ChavanDafle2015a,ChavanDafle2017,ChavanDafle2018a}. However, there we assume precise knowledge of system parameters such as the object inertia, grasping force, and friction coefficients at the fingers and at the external contacts. In practice, there is always some uncertainty in the knowledge of these parameters. Based on the mechanics of pushing, we define in this paper a subspace of robust prehensile pushes and refer to it as a \emph{robust motion cone}. We show that the robust motion cone is invariant to all the system parameters listed above except friction at the external contact/s.
\begin{figure*}
\centering
\includegraphics[scale=0.98]{figures/tpush_exp_small.pdf}
\caption{An example of regrasping using fixtureless fixturing. A T-shaped object is regrasped in a gripper by pushing it against an edge contact in the environment. Note that during the regrasp process, the object sticks to the environment while the fingers slide on it.}
\label{fig:tpush_exp}
\end{figure*}
We present a sampling-based planning framework that uses robust motion cones to build long pushing strategies for regrasping. Exploiting the knowledge of the motion cones for dynamics propagation, the planner rapidly builds a tree of reachable grasps. A path is this tree is a pushing strategy, possibly involving pushes from different sides, to manipulate the object in the grasp. These pushing strategies force the object to the desired pose in the grasp irrespective of any uncertainty in the object inertia, grasping force, friction at the fingers. \figref{fig:tpush_exp} shows one such pushing strategy to manipulate a T-shaped object.
For experimental validation, we consider different examples of regrasping performed on a manipulation platform well-instrumented to capture the pose of the object and the robot. We show that the pushing strategies generated by our planner consistently achieve the desired regrasp and are robust against the uncertainties in the system parameters.
The efficient computation of the motion cones and their application to propagate the dynamics allows our planner to generate pushing strategies in less than a second. Moreover, the planner can use and reuse a simple feature in the environment to manipulate many different objects. Such a fast and robust framework for regrasping by fixtureless fixturing takes us a step closer towards a versatile and flexible method for in-hand manipulation for industrial automation.
\section{Related Work}
\label{sec:related}
Fixturing is one of the widely used practices in industry for applications such as machining, assembly, and inspection. Researchers have extensively studied the principles and designs for fixturing~\cite{american1962handbook, Asada85, Chou89}. However, the role of friction in fixturing is often neglected or considered secondary to that of the kinematics.
One of the very few works that describe the use of force analysis with friction in the context of fixture planning is by Hong and Cutkosky~\cite{Hong91}.
They point out that fixture planning, similar to grasp planning, is primarily considered from the kinematic perspective only.
They advocate the analysis of the mechanics of fixturing and use the limit surface model~\cite{GoyalPhD89} to capture the force-motion relationship at the contacts between the object and the fixture.
Their work focuses on the stability of a fixtured part against the forces exerted by a tool and provides estimates for the clamping force required to prevent the part from sliding. Using similar contact mechanics tools, the work presented in this paper studies the controlled sliding of the object in the grasp. With the fixtureless fixturing, we demonstrate the application of the mechanics of frictional contact for fixturing.
The work presented in this paper is motivated by our recent work on stable prehensile pushing~\cite{ChavanDafle2018a}. In that work, we present a planner to regrasp an object in a gripper using the external pushes for which the object sticks to the pusher. To find feasible stable prehensile pushes, the planner evaluates a dynamics check for multiple random push directions. This dynamics check depends on the knowledge of the object inertia, grasping force, and coefficient of friction at the fingers and at the features in the environment. Consequently, the stable prehnsile pushing is sensitive to these parameters. In contrast, the fixtureless fixturing presented in this paper uses only the stable prehensile pushes whose dynamics is invariant to these parameters except for the friction at the features in the environment.
Moreover, we demonstrate a method to compute bounds on the valid robust stable pushes in the form of robust motion cones. The knowledge of the motion cones avoids the need of dynamics check and speeds up the regrasp planning.
The idea of robust motion cones is inspired by the original concept of the motion cone presented by Mason~\cite{mason86}. Mason studies the mechanics of pushing a planar object on a horizontal plane with a point pusher. He proposes the \emph{motion cone} as a set of pushes that lead to sticking contact between the pusher and the object. Our recent work~\cite{ChavanDafle2018b} extends the idea of motion cones to prehensile tasks in a more general setting and demonstrates their application for fast in-hand manipulation planning. The application of robust motion cones for regrasp planning brings robustness to in-hand manipulation planning.
\section{Problem Formulation}
\label{sec:formulation}
We present a robust in-hand manipulation technique using a simple parallel-jaw gripper. An object in a gripper is forced to a desired pose in the grasp using prehensile pushes.
We implement the prehensile pushes by pushing the object against features in the environment. If provided with a dual-arm robot or a multi-finger gripper, the second arm or the extra fingers of the gripper can be used for the pushes.
We propose a manipulation planning framework that is compatible with several of the practical implementations for the prehensile pushes. We consider that an object is grasped in a gripper that is fixed in the world. A moving pusher pushes the object to a new pose in the grasp.
In our implementation, the pusher motion is a reflection of the robot motion against the fixed environment. Now, planning the robot motion to regrasp an object is equivalent to finding a pushing strategy that forces the object to the desired pose.
We assume the following knowledge about the physical properties of the manipulation system:
\begin{itemize}
\item[$\raisebox{-0.25ex}{\scalebox{1.75}{$\cdot$}}$] Object geometry.
\item[$\raisebox{-0.25ex}{\scalebox{1.75}{$\cdot$}}$] Initial and goal pose of an object in a grasp, specified by the locations and geometries of each fingers contacts.
\item[$\raisebox{-0.25ex}{\scalebox{1.75}{$\cdot$}}$] Discrete set of pusher contacts, specified by their locations and geometries.
\item[$\raisebox{-0.25ex}{\scalebox{1.75}{$\cdot$}}$] Coefficients of friction at the pusher contacts (an approximate value is sufficient).
\end{itemize}
Our proposed planning framework works at two levels.
At the high-level, a T-RRT$^{*}$ based planning architecture samples the configuration space of different grasps, similar to our work in \cite{ChavanDafle2018a}.
At the low level, the planning tree is grown in the direction of the sampled poses using the knowledge of local reachable poses in the form of robust motion cones.
Exploiting the efficient dynamics propagation via motions cones, the planner rapidly explores the configuration space and generates feasible pushing strategies to move the object in the grasp.
\section{Mechanics of Fixtureless Fixturing}
\label{sec:dynamics}
In this section, we start with a brief review on fundamental contact modelling techniques that allow us to model the force-motion interaction during prehensile pushing.
Based on the contact modelling in \secref{sec:contact} and dynamics of stable prehensile pushing in \secref{sec:pushing_dyn}, we discuss the mechanics of fixtureless fixturing and characterize the set of fixtureless fixturing pushes in \secref{sec:fixture_dyn} and \ref{sec:robust_motioncone}.
\subsection{Contact Modelling}
\label{sec:contact}
In this paper, we assume that all bodies are rigid and all contacts have dry isotropic friction.
\subsubsection{Limit Surface}
\label{sec:limit_surface}
Goyal~\cite{GoyalPhD89} presented the \emph{limit surface} as the boundary of the friction wrenches (force-torque pairs) that a contact with finite area can provide.
Howe and Cutkosky~\cite{howe96} and Xydas and Kao~\cite{xydas99} showed that the limit surface geometry can be approximated as an ellipsoid. We consider circular patch contacts at the fingers. To model the force-motion interaction at the finger contacts, we assume the ellipsoidal approximation to the limit surface, which has been shown to be computationally efficient for pushing motions~\cite{lynch96,Dogar2011,lynch15,Jiaji17b}
Let $\boldsymbol{w}_c=[f_\textnormal{x}, f_\textnormal{z}, m_\textnormal{y}]$ be a frictional wrench at a finger contact in the contact frame.
The ellipsoidal limit surface constraint can be written as: $\boldsymbol{w}_c^T A \boldsymbol{w}_c=1$, where $A=Diag(a_1^{-2}, a_2^{-2}, a_3^{-2})$.
For isotropic friction, the maximum friction force is $a_1= a_2=\mu_{c} N$, where $\mu_{c}$ is the friction coefficient, and $N$ is the the normal force at the contact. The maximum friction torque about the contact normal is $a_3=rc\mu_{c} N$, where $r$ is the radius of the finger contact and $c \in [0,1]$ is the constant from numerical integration. For a uniform pressure distribution, $c$ is about 0.6~\cite{xydas99,lynch15}.
When the object slides on the contact, the friction wrench lies on the limit surface such that the normal to the limit surface is in the direction of the object motion at the contact.
If the object twist (linear and angular velocity pair), $\boldsymbol{v_\textnormal{obj}}=[v_x, v_z, \omega_y]^T$, is known, we can find the friction wrench at the contact as:
\begin{equation}
\label{eq:vel2wrench}
\boldsymbol{w}_c= \frac{A^{-1}\boldsymbol{v_\textnormal{obj\_c}}}{\sqrt{\boldsymbol{v_\textnormal{obj\_c}}^T A^{-1} \boldsymbol{v_\textnormal{obj\_c}}}} = \mu_c N \boldsymbol{\overline{w}}_c
\end{equation}
where $\boldsymbol{v}_{obj\_c}$ is the twist of the object in the contact frame. Here, $\boldsymbol{v}_{obj\_c}=[v_{x\_c}, v_{z\_c}, \omega_{y\_c}]^T=\boldsymbol{{J}_\textnormal{c}} \cdot \boldsymbol{v}_{obj}$, where $\boldsymbol{{J}_\textnormal{c}}$ is the Jacobian that maps the velocity from the object frame to the contact frame.
$\boldsymbol{\overline{w}}_c=[\overline{f}_\textnormal{x}, \overline{f}_\textnormal{z}, \overline{m}_\textnormal{y}]^T$ is a unit wrench on the limit surface and is scaled by maximum linear friction at the contact ($\mu_c N$) to estimate the net frictional wrench.
For the ellipsoidal limit surface assumption, the linear velocity $[v_\textnormal{x\_c}, v_\textnormal{z\_c}]^T$ of the object in the contact frame is parallel and opposite to the linear frictional force $[\overline{f}_x, \overline{f}_z]^T$ applied by the contact in the contact frame~\cite{lynch92}. Moreover, the relationship between the friction wrench and the normal to the limit surface, which defines the motion direction, sets the following constraint between the angular velocity at the contact and the linear velocity:
%
\begin{equation*}
\frac{v_\textnormal{x\_c}}{{\omega}_\textnormal{y\_c}}={(rc)^2} \frac{\overline{f}_\textnormal{x}}{\overline{m}_\textnormal{y}} \ \ and \ \
\frac{v_\textnormal{z\_c}}{{\omega}_\textnormal{y\_c}}={(rc)^2} \frac{\overline{f}_\textnormal{z}}{\overline{m}_\textnormal{y}}
\end{equation*}
Given the friction wrench on the object from the contact ($\boldsymbol{\overline{w}}_c$), we can find the object velocity as:
\begin{equation}
\label{eq:wrench2vel}
\boldsymbol{v}_{obj} = k\boldsymbol{\Tilde{J}_\textnormal{c}} \boldsymbol{B} \cdot \boldsymbol{\overline{w}}_c \ , \ \boldsymbol{B}= Diag (1, 1, (rc)^{-2} ), \ \ k \in {\rm I\!R}^+
\end{equation}
where $\boldsymbol{\Tilde{J}_\textnormal{c}}$ maps the object velocity from the contact frame to the object frame.
\subsubsection{Generalized Friction Cone}
\label{sec:generalized_cone}
\begin{figure}
\centering
\includegraphics[scale=1.04]{figures/fric_cone.pdf}
\caption{The generalized friction cone for a line pusher modeled with point contacts at the ends. The convex hull of the set of wrenches imposed by the forces in the friction cones at the two points contacts is a polyhedral cone.}
\label{fig:fric_cone}
\vspace{-3mm}
\end{figure}
Erdmann~\cite{Erdmann94} introduced the concept of the \textit{generalized friction cone} ($\boldsymbol{W}$) as a representation of the local Coulomb friction cone at a contact in the wrench-space and in the object frame. The generalized friction cone for a patch contact modelled with multiple point contacts is the convex hull of the generalized friction cones for each constituent contact~\cite{Erdmann93}.
We model the friction at a pusher contact with its generalized friction cone:
\begin{equation}
\label{eq:wrench_cone}
\boldsymbol{W_\textnormal{pusher}} = \{ \boldsymbol{\overline{w}_\textnormal{pusher}} = \boldsymbol{J^\top_\textnormal{p}}\cdot \boldsymbol{\overline{f}_\textnormal{p}} \ | \ \boldsymbol{\overline{f}_\textnormal{p}} \in FC_\textnormal{pusher}\}
\end{equation}
Here, $\boldsymbol{{f}_\textnormal{p}}$ is the collection of forces at the constituent point contact/s at the pusher, and $FC_\textnormal{pusher}$ is the Coulomb friction cone/s at those constituent contact/s. The Jacobian $\boldsymbol{J^\top_\textnormal{p}}$ maps the local contact forces at the pusher ($\boldsymbol{{f}_\textnormal{p}}$) to the wrenches in the object frame; $\boldsymbol{\overline{w}_\textnormal{pusher}}$ is the unit wrench corresponding to unit force/s $\boldsymbol{\overline{f}_\textnormal{p}}$ inside the friction cone/s at the constituent contact/s of the pusher.
\figref{fig:fric_cone} shows the generalized friction cone for a line pusher modeled with point contacts at the ends. Note that the polyhedral cone form of the generalized friction cone is not an approximation. The convex hull of the wrenches exerted by the forces in the friction cones at the two points contacts in fact takes the form of a polyhedral cone.
\subsection{Dynamics of Stable Prehensile Pushing}
\label{sec:pushing_dyn}
An object in a grasp moves following the net wrench acting on the object. Under the quasi-static assumption, which is valid for slow pushing operations, the inertial forces on the object are negligible and there is a force balance:
\begin{equation}
\label{eq:force_balance}
\boldsymbol{w_\textnormal{grasp}} + \boldsymbol{w_\textnormal{pusher}} + m \boldsymbol{g}=0
\end{equation}
Equation~\eref{eq:force_balance} is written in the object frame located at the center of gravity of the object, where $\boldsymbol{w_\textnormal{grasp}}$ is the friction wrench provided by the grasp; $\boldsymbol{w_\textnormal{pusher}}$ is the wrench exerted by the pusher; $m$ is the mass of the object; and $\boldsymbol{g}$ is the gravitational wrench.
This paper focuses on manipulations in a parallel-jaw grasp, so the object motions are restricted in the plane of the grasp.
We assume uniform pressure distribution at the finger contacts.
The normal forces at the finger contacts balance each other. All forces involved in these manipulations (friction forces at the fingers, the pusher forces, and the gravitational force) are in the plane of the grasp.
Both the fingers contribute equally to constitute the grasp wrench, i.e., $\boldsymbol{w_\textnormal{grasp}}=2 (\boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{{w}_\textnormal{c}})$. Here, $\boldsymbol{J^\top_\textnormal{c}}$ maps the contact wrench from the finger contact frame to the object frame. We can rewrite \eref{eq:force_balance} as:
\begin{equation}
\label{eq:force_balance_prepush}
2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) + \boldsymbol{J^\top_\textnormal{p}}\cdot \boldsymbol{f_\textnormal{p}} + m\boldsymbol{g} = 0
\end{equation}
A prehensile push for which the pusher contact sticks to the object during the push is called a stable prehensile push~\cite{ChavanDafle2018a}. For a stable prehensile push, constraint \eref{eq:force_balance_prepush} is balanced by the force/s inside the friction cone/s at the pusher constituent contact/s:
\begin{equation*}
\begin{split}
2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) + \boldsymbol{J^\top_\textnormal{p}}\cdot \boldsymbol{f_\textnormal{p}} + m\boldsymbol{g} = 0 \ , \ \boldsymbol{{f}_\textnormal{p}} \in FC_\textnormal{pusher}
\end{split}
\end{equation*}
Using \eref{eq:wrench_cone}, we can rewrite the previous equation as:
\begin{equation}
\label{eq:stable_check_prepush1}
\begin{split}
-2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) - m\boldsymbol{g} = k_1 \boldsymbol{\overline{w}_\textnormal{pusher}} \\
\boldsymbol{\overline{w}_\textnormal{pusher}} \in \boldsymbol{W_\textnormal{pusher}} \ , \ k_1 \in {\rm I\!R}^+
\end{split}
\end{equation}
where $k_1$ is the magnitude of the pusher normal force.
To find if a particular object motion can be achieved with a stable push, we can check if the net required wrench falls inside the generalized friction cone of the pusher:
\begin{equation}
\label{eq:stable_check_prepush2}
\begin{split}
-2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) - m\boldsymbol{g} \in \boldsymbol{W_\textnormal{pusher}}
\end{split}
\end{equation}
\subsection{Dynamics of Fixtureless Fixturing}
\label{sec:fixture_dyn}
We define \emph{fixtureless fixturing} as a subset of stable prehensile pushes that are invariant to the object inertia, grasping force and friction parameters at the finger contacts.
The presence of gravity in \eref{eq:stable_check_prepush1} makes the dynamics of prehensile pushing sensitive not only to the object inertia ($m\boldsymbol{g}$), but also to the grasping force ($N$) and the coefficient of friction at the fingers ($\mu_\textnormal{c}$).
If the pusher is aligned such that the pusher contact normal is along the direction of gravity, the gravitational force on the object is entirely balanced by the part of the normal force at the pusher.
Then the dynamics check for stable prehensile pushing \eref{eq:stable_check_prepush1} becomes:
\begin{equation*}
\begin{split}
-2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) - k_2\boldsymbol{\overline{w}_\textnormal{pusher\_n}} = k_1 \boldsymbol{\overline{w}_\textnormal{pusher}} \\
\boldsymbol{\overline{w}_\textnormal{pusher}} \in \boldsymbol{W_\textnormal{pusher}} \ , \ \ \ k_1, k_2 \in {\rm I\!R}^+
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
\vspace{-1mm}
-2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) = k_1 \boldsymbol{\overline{w}_\textnormal{pusher}} + k_2\boldsymbol{\overline{w}_\textnormal{pusher\_n}} \\
\boldsymbol{\overline{w}_\textnormal{pusher}} \in \boldsymbol{W_\textnormal{pusher}} \ , \ \ \ k_1, k_2 \in {\rm I\!R}^+
\end{split}
\end{equation*}
where $\boldsymbol{\overline{w}_\textnormal{pusher\_n}}$ is the unit normal wrench at the pusher contact. We know that $\boldsymbol{\overline{w}_\textnormal{pusher\_n}} \in \boldsymbol{W_\textnormal{pusher}}$, so we can rewrite the above constraint as:
\begin{equation*}
\vspace{-1mm}
\begin{split}
-2(\mu_\textnormal{c} N \boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}}) = k_3 \boldsymbol{\overline{w}_\textnormal{p}} \\
\boldsymbol{\overline{w}_\textnormal{p}} \in \boldsymbol{W_\textnormal{pusher}} \ , \ \ \ k_3 \in {\rm I\!R}^+
\end{split}
\end{equation*}
\noindent As $\mu_\textnormal{c}$ and $N$ are scalar constants, we further simplify it as:
\begin{equation}
\label{eq:stable_check_prepush_gfree1}
\begin{split}
-\boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}} = k_4 \boldsymbol{\overline{w}_\textnormal{p}} \\
\boldsymbol{\overline{w}_\textnormal{p}} \in \boldsymbol{W_\textnormal{pusher}} \ , \ \ \ k_4 \in {\rm I\!R}^+
\end{split}
\end{equation}
Now, we can write the dynamics condition for stable prehensile pushing with gravity balancing pushers as:
\begin{equation}
\label{eq:stable_check_prepush_gfree2}
\begin{split}
-\boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{c}} \in \boldsymbol{W_\textnormal{pusher}}
\end{split}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=1]{figures/motion_cone.pdf}
\caption{An example of a robust motion cone. Note that the motion cone is a three-dimensional cone in the space of the linear velocity of the object in the $XZ$ plane and the rotational velocity of the object about $Y$ axis.}
\label{fig:motion_cone}
\vspace{-3mm}
\end{figure}
Note the invariance of \eref{eq:stable_check_prepush_gfree1} and \eref{eq:stable_check_prepush_gfree2} on the object inertia, grasping force and the friction coefficient at the fingers.
All the object motions that satisfy \eref{eq:stable_check_prepush_gfree2} can be achieved with fixtureless fixturing when pushing in the gravity-balancing orientation. For planning regrasps with fixtureless fixturing, rather than checking different instantaneous object motions for their validity for \eref{eq:stable_check_prepush_gfree2}, a direct bound on the feasible object motions is more efficient.
\subsection{Computation of Robust Motion Cone}
\label{sec:robust_motioncone}
A motion cone abstracts the dynamics of pushing and provides direct bounds on the object motions that can be achieved while keeping the pusher contact sticking to the object~\cite{mason86}.
In this section, we will extend the idea of the motion cone for fixtureless fixturing.
\textbf{\textit{Problem}}: Find a set of instantaneous object motions that can be achieved with fixtureless fixturing.
This is equivalent to finding a set of object motions that satisfy constraint \eref{eq:stable_check_prepush_gfree2}.
To find a motion cone, we find a set of support contact wrenches ($\boldsymbol{\overline{w}_\textnormal{c}}$) that satisfy \eref{eq:stable_check_prepush_gfree2} and then map this wrench-set ($\boldsymbol{\overline{W}_\textnormal{c}}$) to a set of object twists. We denote this object twist-set by $\boldsymbol{\overline{V}_\textnormal{obj}}$ and call it \emph{robust motion cone}.
From the constraint \eref{eq:stable_check_prepush_gfree2} we can observe that $-\boldsymbol{J^\top_\textnormal{c}} \cdot \boldsymbol{\overline{W}_\textnormal{c}}=\boldsymbol{W_\textnormal{pusher}}$ or $\boldsymbol{\overline{W}_\textnormal{c}}=-\boldsymbol{\Tilde{J}^\top_\textnormal{c}} \cdot \boldsymbol{W_\textnormal{pusher}}$ where $\boldsymbol{\Tilde{J}^\top_\textnormal{c}}$ maps wrenches from the object frame to the finger contact frame. From definition \eref{eq:wrench_cone}, $\boldsymbol{W_\textnormal{pusher}}$ is convex and polyhedral. Therefore, $\boldsymbol{\overline{W}_\textnormal{c}}$ and $\boldsymbol{\overline{V}_\textnormal{obj}}$ are also convex and polyhedral.
The generators of the robust motion cone $\boldsymbol{\overline{V}_\textnormal{obj}}$ can be computed by a linear mapping of the generators of the generalized friction cone of the pusher $\boldsymbol{W_\textnormal{pusher}}$. \figref{fig:motion_cone} shows a graphical illustration of the motion cone computed for a pushing scenario.
\myparagraph{Procedure to compute a robust motion cone}:
\begin{enumerate}
\item Solve $\boldsymbol{\overline{w}_\textnormal{c}}=-\boldsymbol{\Tilde{J}^\top_\textnormal{c}} \cdot \boldsymbol{\overline{w}_\textnormal{p}}$ for $\boldsymbol{\overline{w}_\textnormal{p}}$ equal to every edge of the generalized friction cone $\boldsymbol{W_\textnormal{pusher}}$.
\item Define the set of $\boldsymbol{\overline{w}_\textnormal{c}}$ computed in step 1 as the generators/edges of the wrench-set $\boldsymbol{\overline{W}_\textnormal{c}}$.
\item Map $\boldsymbol{\overline{W}_\textnormal{c}}$ to the object twist space using \eref{eq:wrench2vel} to obtain the robust motion cone $\boldsymbol{\overline{V}_\textnormal{obj}}$.
\end{enumerate}
\section{Application to Regrasp Planning}
\label{sec:planning}
In this section, we discuss the application of fixtureless fixturing to regrasp planning. We present a sampling-based planning framework that uses the knowledge of robust motion cones and generates a tree of grasps that can be reached with fixtureless fixturing. A path in this tree is a pushing strategy to regrasp an object in the gripper through a series of robust stable prehensile pushes.
The high-level planning framework is similar to that presented in our prior work~\cite{ChavanDafle2017,ChavanDafle2018a, ChavanDafle2018b}. It is based on T-RRT$^*$ -- an optimal sampling based method for planning on configuration space cost-maps~\citep{trrt_star,trrt}.
\myparagraph{For selective sampling}, the T-RRT* framework uses a transition test and filters the sampled configurations to prefer the exploration in low configuration-cost regions. We define the configuration cost as the distance from the goal. The transition test\footnote{The implementation of the transition test and its adaptive nature is discussed in detail in~\cite{ChavanDafle2017}} gives preference to the stochastic exploration towards the goal grasp, while allowing the flexibility to move the object away from the goal if that is necessary to move the object finally to the goal.
\myparagraph{For effective connections}, the T-RRT$^*$ exploits the underlying RRT$^*$~\citep{rrt_star} framework to make and rewire the connections in the tree such that the cost of the nodes is reduced.
We define the cost of a node as the sum of the cost of the parent node and the cost to push the object to the sampled node. The cost of a push is $0.1$ if the parent node uses the same pusher as the child and $1$ otherwise.
In practice, every time a pusher contact is changed, it introduces a possibility to add noise to the system. Reducing the pusher switch-overs can add to the robustness of the manipulations planned.
With our node cost definition, the planner tries to reduce the number of pusher switch-overs required to push the object to the desired pose.
\begin{algorithm}[t]
\caption{: In-Hand Manipulation Planner}\label{alg:full_planner}
$ \textbf{input}: q_{init}, q_{goal}$ \par
$ \textbf{output}:$ {tree} $\ \mathcal{T}$
\begin{algorithmic}
\State $\mathcal{T}\gets \textrm{initialize tree}(q_{init})$
\State $ \textrm{generate\_robustCones}(\mathcal{T},q_{init})$
\While{$q_{goal} \notin \mathcal{T}$ \textbf{or} cost($q_{goal}) > \textrm{cost threshold} $}
\State $q_{rand}\gets \textrm{sample random configuration}(\mathcal{C})$
\State $q_{parent}\gets \textrm{find nearest neighbor}(\mathcal{T},q_{rand})$
\State $q_{sample}\gets \textrm{take unit step}(q_{parent},q_{rand})$
\If{$q_{sample} \notin \mathcal{T}$}
\If{\textrm{transition test}$(q_{parent},q_{sample},\mathcal{T})$}
\State $q_{new} \gets \textrm{robust\_push}(q_{parent},q_{sample})$
\If{\textrm{transition test}$(q_{parent},q_{new},\mathcal{T})$ \textbf{and} \\ \hspace{19mm}\textrm{grasp maintained}$(q_{new})$}
\State $q\mbox{*}_{parent}\gets \textrm{optimEdge}(\mathcal{T},q_{new},q_{parent})$
\State $\textrm{add new node}(\mathcal{T},q_{new})$
\State $\textrm{add new edge}(q\mbox{*}_{parent},q_{new})$
\State $ \textrm{generate\_robustCones}(\mathcal{T},q_{new})$
\State $\textrm{rewire tree}(\mathcal{T},q_{new},q\mbox{*}_{parent})$
\EndIf
\EndIf
\EndIf
\EndWhile
\end{algorithmic}
\end{algorithm}
Algorithm \ref{alg:full_planner} shows our in-hand manipulation planner.
Let $q$ be a configuration of an object, i.e., the pose of the object in the gripper frame that is fixed in the world.
In this paper, we consider manipulations in a parallel-jaw grasp, so the configuration space $\mathcal{C}$ is $[X, Z, \theta_y] \in {\rm I\!R}^3$. The object can translate in the grasp plane ($XZ$ plane) and rotate around a perpendicular to the grasp plane ($Y$ axis).
Let $q_{init}$ and $q_{goal}$ be an initial and desired pose of the object in the grasp frame, respectively.
The planner initiates a tree $\mathcal{T}$ with $q_{init}$ and generates robust motion cones at it.
While the goal pose is not reached within some cost threshold, a random configuration ($q_{rand}$) is sampled. A nearest configuration ($q_{parent}$) to $q_{rand}$ in the tree $\mathcal{T}$ is found, and an unit-step object pose ($q_{sample}$) towards the $q_{rand}$ is computed.
Using the transition test, the planner evaluates whether a move from the parent node $q_{parent}$ to $q_{sample}$ is beneficial or not.
If it is beneficial, the \textit{robust\_push} routine computes an object configuration ($q_{new}$) closest to $q_{sample}$ that can be reached using the robust motion cones at $q_{parent}$. The node $q_{new}$ is added to the tree such that the node costs of $q_{new}$ and the nodes near $q_{new}$ are lowered if possible. The robust motion cones are generated for every new node added to the tree.
\textit{generate\_robustCones} and \textit{robust\_push} are two important routines for the regrasp planning using fixtureless fixturing:
\myparagraph{\textit{generate\_robustCones}} computes polyhedral robust motion cones for a given object configuration in the grasp using the procedure explained in \secref{sec:robust_motioncone}. At every node, the number of motion cones is the same as that of the pushers.
\myparagraph{\textit{robust\_push}} finds an object pose closest to the sampled pose ($q_{sample}$) that can be reached with fixtureless fixturing. This computation is done using the robust motion cones at the parent node ($q_{parent}$), which provide the knowledge of local reachability.
If the object twist needed from $q_{parent}$ to $q_{sample}$ is inside any of the robust motion cones, the sampled pose can be reached from $q_{parent}$.
If the required object twist is outside all the motion cones, a twist that is inside one of the motion cones and closest to the desired twist is selected.
The motion cone based unit-step propagation and T-RRT$^*$-based high-level framework work together to rapidly explore the configuration space of different grasps and generate pushing strategies for the desired in-hand manipulation. The manipulation plans generated by the planner are invariant to the object inertia, grasping force and friction at the fingers.
\section{Regrasp Examples and Experimental Validation}
\label{sec:Exp_examples}
\begin{figure}
\centering
\includegraphics[scale=0.7]{figures/linpush_small.pdf}
\caption{Simulated motion of the object and snapshots of the experimental run for a pushing strategy to offset the object in the grasp using low coefficient pushers. The circular finger contact is shown in green, while the line pusher contact is shown in magenta.}
\label{fig:linpush}
\end{figure}
\begin{table}[b]
\vspace{-3mm}
\caption{Physical properties of the objects used.}
\label{tab:objects}
\centering
\begin{tabular}{|l|l|c|r|}
\hline
\textbf{Shape} & \textbf{Material} & \textbf{Dim. [L, B, H]} (mm) & \textbf{Mass} (g) \\ \hline
square prism & Al 6061 & 100, 25, 25 & 202\\ \hline
rectangular prism & Delrin & 80, 25, 38 & 130\\ \hline
T-shaped & ABS & 70, 25, 50 & 62\\ \hline
\end{tabular}
\end{table}
We validate our manipulation planner with regrasping tasks in simulation and with experiments. The experiments were performed on a manipulation platform instrumented with an industrial robot arm, a parallel-jaw gripper with force control, and a Vicon system for object tracking.
\subsection{Simulation and Experimental Setup}
\label{sec:setup}
We consider different regrasp tasks for the objects listed in Table \ref{tab:objects}. We use a parallel-jaw gripper with flat circular finger contacts. The planner is initiated with pusher contacts on either sides of the object and under the object. In practice, all these contacts are implemented using a single feature in the environment.
The initial pose of the object is treated as $[0,0,0]$, and the goal poses for different regrasp examples are listed in \tabref{tab:timing}. The plan time for each regrasp, shown in \tabref{tab:timing}, is the median time in seconds over 10 planning trials. A computer with an Intel Core i7 2.8 GHz processor and MATLAB R2017a was used for all the computations and planning.
\subsection{Regrasp Examples}
\label{sec:examples}
\subsubsection{Regrasping an object offset to the center}
The goal in this example is to regrasp the square prism 20 mm offset from the center in the horizontal direction. The gripping force, and frictional parameters at the fingers and at the features in the environment, are chosen such that pushing the object horizontally from side will not be a valid solution. Due to the gravitational force, the object will also slide downwards by a few millimeters if pushed from the side, as observed in~\cite{Kolbert16}.
\begin{table}[b]
\vspace{-4mm}
\caption{Planning times (in seconds) with different methods for unit-step propagation}
\label{tab:timing}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Manipulation} & \textbf{Goal} & \textbf{Plan} & \textbf{Plan} & \textbf{Plan}\\
& [$X, Z, \theta_Y $] & \textbf{Time} & \textbf{Time}~\cite{ChavanDafle2018a} & \textbf{Time}~\cite{ChavanDafle2017} \\
\hline
Horz. offset (low $\mu$) & 20, 0, 0 & 0.39 & 2.83 & 592.8\\ \hline
[$X, Z, \theta_Y$] Regrasp & 15, -13, 45 & 0.65 & 2.54 & 17684\\ \hline
T-shaped & 25, 17.5, 0 & 0.60 & 0.82 & 32657\\ \hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[scale=0.7]{figures/prepush_small.pdf}
\caption{Simulated motion of the object and snapshots of the experimental run for a general regrasp in [$X, Z, \theta_y$]. In the simulation, the direction of gravity remains constant in the pusher frame because, in reality, the pusher is a fixed feature in the environment}
\label{fig:prepush}
\vspace{-5mm}
\end{figure}
The planner proposed in this paper rotates the gripper such that the object is pushed from a side face but using the feature located horizontally, as shown in \figref{fig:linpush}. The gravitational force on the object is along the pusher contact normal and the regrasp is achieved using fixtureless fixturing. This strategy is simpler and more robust than the solutions that the planners in~\cite{ChavanDafle2017,ChavanDafle2018a} generated for the same problem.
In~\cite{ChavanDafle2017} and~\cite{ChavanDafle2018a}, the planner uses multiple pushers, specifically one from side and one from bottom face, to achieve this regrasp.
With the planner proposed in this paper, no contact switch-over is needed for the desired regrap as shown in \figref{fig:linpush}.
Moreover, the time taken to plan such a strategy is 5 to 1500 times faster than our prior work as shown in Table \ref{tab:timing}.
\subsubsection{General manipulation in $[X, Z, \theta_y]$}
This example demonstrates large in-hand manipulation for the rectangular prism, involving both displacement and rotation of the fingers on the object. The planner quickly converges to a strategy which does not require any pusher switch-over. The knowledge of motion cones allows the planner to reason about all feasible motions that can be made instantaneously and to choose the motion that best moves the object towards the goal. \figref{fig:prepush} shows the pushing strategy our planner generated.
\subsubsection{Manipulation with a complex non-convex object}
Here we show manipulation of a T-shaped object. Manipulating such a non-convex object is particularly hard, because pushing the object directly towards the goal may lead to losing the grasp on the object. The long horizon planning feature of the planner and the capability to reason about discreetly changing the pushers are essential for regrasping this object.
As shown in \figref{fig:tpush_exp} and \figref{fig:tpush_sim}, our planner generates a strategy that respects the object shape and corresponding geometric constraints. The planner utilizes pushes from two different sides to finally move the object to the goal grasp.
\begin{figure}
\centering
\includegraphics[scale=0.85]{figures/tpush_sim.pdf}
\caption{Simulated motion of the object for a pushing strategy for manipulating a T-shaped object. \figref{fig:tpush_exp} shows the experimental run.}
\label{fig:tpush_sim}
\vspace{-3mm}
\end{figure}
\subsection{Experimental Observations}
\label{sec:observations}
In this section, we highlight some of the key characteristics of fixtureless fixturing that are supported experimentally. We carried out multiple runs of the pushing strategies in \secref{sec:examples} for different grasping forces and recorded the object motion in the grasp.
\subsubsection{Minimal displacement of the object}
The experimental data shows that the object moves a minimal amount with respect to the environment during the planned regrasps. From the mechanics of robust stable prehensile pushing, we expect the object to stick to the environment during the regrasps.
\figref{fig:linpush_error} shows the displacement of the object observed during the pushing strategy shown in \figref{fig:linpush}.
The displacement in the X direction is less than $\pm 0.2$~mm, whereas that in the Z direction is even smaller, less than $\pm 0.1$~mm.
These values are close to the precision level of the Vicon object tracking system, so some of the object displacement can be attributed to the noise from the camera system.
\figref{fig:prepush_error} shows the displacement of the object during the pushing strategy shown in \figref{fig:prepush}.
The object moves during the regrasp by a very small amount, less than $3\%$ of the object displacement required for this regrasp.
\subsubsection{Robustness against variation in the friction at the fingers}
\label{sec:robustness}
We observe a minimal effect on the outcome of the regrasp actions planned with fixtureless fixturing when the friction at the fingers is changed.
As the grasping force changes, the friction force at the fingers change. However, as seen in the plots in \figref{fig:linpush_error} and \figref{fig:prepush_error}, the variation in the outcome of the pushing strategy for three different grasping forces is negligible. This supports the invariance we expect theoretically from fixtureless fixturing.
The repeatability and robustness of the pushing actions allowed the robot to place the T-shaped object after every regrasp and restart the next run autonomously twenty times in a row. The T-shaped object does not have Vicon markers on it for tracking. Twenty runs for the manipulation shown in \figref{fig:tpush_exp} were performed in a sequence in an open-loop fashion with different grasping forces.
\subsubsection{Universal Fixture} For all the examples discussed in \secref{sec:examples}, a single feature in the environment was sufficient to manipulate different objects. Fixtureless fixturing allows robots to use a simple feature as a ``universal'' fixture for different objects and different regrasping tasks.
\begin{figure}[t]
\centering
\includegraphics[scale=0.98]{figures/linpush_error.pdf}
\caption{Displacement of the object with respect to the environment for the regrasp action shown in \figref{fig:linpush}. As expected, during fixtureless fixturing, the object sticks to the environment and moves by a negligible amount as the fingers slide on it.}
\label{fig:linpush_error}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.98]{figures/prepush_error.pdf}
\caption{Displacement of the object with respect to the environment for the regrasp action shown in \figref{fig:prepush}.
}
\label{fig:prepush_error}
\vspace{-3mm}
\end{figure}
\section{Discussion}
\label{sec:discussion}
We present a manipulation method that acts as a fixturing technique for regrasping without using a conventional fixturing hardware. The object in a gripper is manipulated with a series of pushes against simple features in the environment such that the environment keeps the object stationary while the fingers slide on the object to a desired pose. We refer to this approach as \emph{fixtureless fixturing}.
Based the mechanics of prehensile pushing, we construct convex polyhedral sets of object motions that can be produced with fixtureless fixturing. We call this subsets of pushes robust motion cones and show that they are invariant to the object inertia, gripping force, and friction at the fingers. The robust motion cones abstract the dynamics of prehensile pushing and provide direct bounds on the object motions that are feasible with fixtureless fixturing.
We demonstrate the application of the robust motion cones for planning in-hand manipulations with a parallel-jaw gripper. Our planner generates pushing strategies in a fraction of a second to force the object to the goal pose. The generated pushing plans are robust against the uncertainty in the object inertia, friction at the fingers, and gripping force.
We validate this robustness with controlled experiments on a platform instrumented with a Vicon object tracking system.
The experimental results confirm that the object remains fixtured to the environment during the manipulations. Moreover, minimal variation is observed in the outcome of the regrasp experiments performed with different grasping forces.
Using the fixtureless fixturing technique, robots can exploit simple features in the environment as versatile fixtures. The flexibility gained by using fixtureless fixturing facilitates dexterous manipulation even with simple grippers. With such an approach to in-hand manipulation, we move a step closer to the goal of versatile, yet simple, robust, and flexible robotic automation.
|
\section{Proofs}
\label{sec:proof}
First, we provide a helpful technical lemma.
\begin{lemma}\label{lem:piecewise-direction}
Let $z \in \{-1,1\}^T$, and suppose the functions $p_1(x), \ldots p_T(x)$ are analytic
functions from $\mathbb{R}^n\rightarrow \mathbb{R}$.
Let $S_z \subset
\mathbb{R}^n$ be a set defined as follows:
$$
S_z=\{x \, \vert \, \mathrm{sign}(p_1(x))=z_1, \ldots \mathrm{sign}(p_T(x))=z_T \} \, .
$$
Fix any $x\in \mathbb{R}^n$. For all $z \in \{-1,1\}^T$ and for almost all $u$, there
is an open neighborhood $ B(u,r)$ such that for all $u' \in B(u,r)$ ,
\[
\lim_{\delta \downarrow 0} \mathbb{I}_{S_z} (x+\delta u)=\lim_{\delta \downarrow 0} \mathbb{I}_{S_z} (x+\delta u') \, .
\]
\end{lemma}
\begin{proof}
We need only consider one such $z$ for the proof, as the set of
possible $z$ is finite. Without loss of generality, we can say $S_z=S $ is a set of the form:
$$
S=\{p_1(x)\ge 0,\ldots, p_k(x) \ge 0, p_{k+1}(x) <0,\ldots,p_T(x) <0\} \, .
$$
For the proof, it suffices to show that for all $j$ and almost all $u$, there
is an open neighborhood $ B(u,r)$ such that for all $u' \in B(u,r)$ ,
\[
\lim_{\delta \downarrow 0} \mathrm{sign}(p_j(x+\delta u)) =
\lim_{\delta \downarrow 0} \mathrm{sign}(p_j(x+\delta u'))
\, .
\]
Let us split the constraints into a set of active constraints $p_j(x)
=0$ for $j \in A$ and inactive constraints $p_j (x) \neq 0$ for $j \in I$.
For inactive constraints, the above holds due to continuity. It
remains to show the above for active constraints. Let us also assume that
the functions $p_j$ in the active set are not identically equal to $0$ (the claim
is true for any zero function).
For every active constraint $j \in A$ , testing $\lim_{\delta \downarrow 0}
\mathrm{sign}(p_j(x+\delta u) )= 1$ can be done by the Taylor expansion,
\begin{align*}
p_j(x+\delta u) = p_j(x)+ \sum_{i=1}^\infty\frac{\delta^i \norm{u}^i}{i!}
\nabla^i p_j(x) [u^{\otimes i}] \, ,
\end{align*}
which exists since $p_j$ is analytic.
For any $i$, note that $\nabla^i p_j(x) [u^{\otimes i}]$ is a
polynomial function in $u$. As it is a polynomial, this function will
either be equal to $0$ (for all $u$) or
it will be nonzero for almost all $u$\footnote{This can proven
by induction on the dimension of $u$. }.
Let $i_j$ be the first index that the function
$\nabla^i p_j(x) [u^{\otimes i}]$ is nonzero for almost all $u$ ($i_j$
is finite as $p_j$ is not the zero function). Then for almost every direction $u$,
\[
\lim_{\delta \downarrow 0} \mathrm{sign} (p_j (x+\delta u) ) =\mathrm{sign}\big(
\nabla^{i_j} p_j(x) [ u^{\otimes i_j} ]\big).
\]
For almost all $u$,
$\nabla ^{i_j} p_j(x) [u^{\otimes i_j}] \neq 0$. Since $\nabla ^{i_j}
p_j(x) [u^{\otimes i_j}]$ is strictly non-zero, then,
by continuity,
\[
\mathrm{sign} (\nabla ^{i_j} p_j(x) [u^{\otimes i_j}])
=
\mathrm{sign}(\nabla^{i_j} p_j (x) [ u'^{\otimes i_j}])
\]
for $\norm{u'-u}<r$ (for sufficiently small $r$). This implies that for $\norm{u'-u}<r$
\[
\lim_{\delta \downarrow 0} \mathrm{sign}(p_j(x+\delta u)) =
\lim_{\delta \downarrow 0} \mathrm{sign}(p_j(x+\delta u'))
\, ,
\]
which completes the proof.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:chain}}
\iffalse
\begin{reptheorem}{thm:chain}
Assume $h:\mathbb{R}^m \rightarrow \mathbb{R}$ and $g_1$, \ldots
$g_m$ (where $g_i: \mathbb{R}^d \rightarrow \mathbb{R}$) are locally
Lipschitz functions computable in our
computational model and that the function $h$ is overloaded with an ASD subroutine
as specified in Assumption~\ref{assumption:library}.
Define:
\[
f(x) := h(g_1(x), \ldots g_m(x)) =h(g(x))\, ,
\]
where $g(x)$ is the vector valued function $(g_1(x), \ldots
g_m(x))^\top$.
Denote the $m\times 1$ vector of (one-sided) directional
derivatives as $D[g;v](x)$. If it exists, let $\partial[g;v](x)$ denote $m
\times d$ limiting Jacobian
matrix (whose rows are given by the vectors $\partial[g_i;v](x)$'s). Set:
\[
[a, d, u] = \text{ASD}[h] (g(x); D[g;v](x))
\]
For all but a measure $0$ set of $v$, we have that $\partial[f;v](x)$
and $\partial[g;v](x)$ exist and that:
\[
\partial[f;v](x) = \partial[g;v](x)^\top u
\]
\end{reptheorem}
\fi
\begin{proof}
Suppose that a program for $h(y)=h(y_1,\ldots, y_m)$ has the representation:
\begin{eqnarray*}\label{eq:h}
h(y) = \sum_{z \in \{-1,1\}^T} \mathbb{I}_{S_z} (y) p_z (y),
\end{eqnarray*}
where $S_z$ is specified as in Equation~\ref{eq:piecewise-def-f}.
Since $f$ is the composition of two programs, it also has a
representation of the form:
\begin{eqnarray*}
f(x) = \sum_{z \in \{-1,1\}^{T'}} \mathbb{I}_{\tilde{S}_z} (x) \tilde{p}_z (x),
\end{eqnarray*}
for different polynomials $\tilde{p}_z$ and sets $\tilde{S}_z$.
By Rademacher's theorem, $\partial [f;v]$ exists for almost all
$v$. Since $f$ itself is a piecewise polynomial, $\partial[f;v](x) =
\nabla \tilde{p}_z(x)$ where $z$ is such that $\lim_{\delta \downarrow 0}
\mathbb{I}_{\tilde{S}_z} (x+\delta v)=1$.
Using Lemma \ref{lem:piecewise-direction}, for all $z' \in
\{-1,1\}^{T'}$ and almost all $v$,
there is a (full dimensional) neighborhood $\mathcal{B}$ around $v$,
in which $\lim_{\delta \downarrow 0} \mathbb{I}_{\tilde{S}_{z'}}
(x+\delta v')=1$, which implies $\partial[f;v'](x) = \nabla
\tilde{p}_{z}(x)$ (using our choice of $z$). Hence,
\[
\partial[f;v](x) = \partial[f;v'](x) \,
\]
for all $v' \in \mathcal{B}$.
Furthermore, using this and the definition of the directional
derivative, we have for all $v' \in \mathcal{B}$,
\begin{equation}\label{eq:dir_deriv_pbar}
D[f;v'](x)= \partial[f;v'](x) ^\top v' =\partial[f;v](x)^\top v'
\, .
\end{equation}
The remainder of the proof seeks to show this directional derivative can be
written as:
\[
D[f;v'](x)= u^\top \partial[g;v](x) \, v'
\]
(for $v'$ in a sufficiently
small full dimensional ball). Note this would imply that for all $v'$
in a full dimensional ball, $u^\top \partial[g;v](x) \, v'
=\partial[f;v](x)^\top v' $, which would complete the proof by choosing $d$
linearly independent $v'$'s.
Define $M=\partial[g;v](x)\in \mathbb{R}^{m \times d}$ be a matrix whose rows are $\partial[g_i;v](x)$. Observe $D[g;v](x) =M v$. Now let us
show that:
\[
D[g;v'](x) =M v'
\]
for all $v' \in \mathcal{B}_1$, where $\mathcal{B}_1 $ is a
sufficiently small (full dimensional) neighborhood
around $v$.
To see this, note the function $g(x)$ itself is a piecewise (vector) valued
function, and, by Lemma \ref{lem:piecewise-direction}, for almost all
$v$, there is a neighborhood
$\mathcal{B}_1$ around $v$, in which for all $v' \in
\mathcal{B}_1$, $v'$ selects the same piece as $v$, which implies
$\partial[g;v'](x)=M$ as claimed.
Now due to the assumed properties of $ASD$, we know
\[
u = \nabla p_{z}(y) \vert_{y=g(x)} \, .
\]
for some $z$ which satisfies:
\[
\lim_{\delta \downarrow 0} \mathbb{I}_{S_z} (g(x)+
\delta D[g;v](x) ) = 1 \, .
\]
Since $D[g;v](x) =M v$, we have:
\[
\lim_{\delta \downarrow 0} \mathbb{I}_{S_z} (g(x)+
\delta Mv ) = 1 \, .
\]
Also, note
\[
D[h;Mv](y) =\nabla p_z(y)^\top Mv
\]
by the definition of the directional derivative and where
$y=g(x)$.
Using Lemma
\ref{lem:piecewise-direction} again, we will show that
\begin{equation} \label{eq:claim}
D[h;Mv'](y) =\nabla p_z(y)^\top Mv'
\end{equation}
for all $v' \in \mathcal{B}_2$ (for a sufficiently small ball
$\mathcal{B}_2$). Note that the above holds if:
\begin{align}
\lim_{\delta \downarrow 0} \mathbb{I}_{S_z} (g(x)+
\delta Mv' ) = 1 \, , \label{eq:Mv'-set}
\end{align}
for all $v' \in \mathcal{B}_2$.
Now consider the set
$\bar{S}_z$ defined by
\[
\bar{S}_z=\{\tilde v \, \vert \, \mathrm{sign}(p_1(g(x)+M\tilde
v))=z_1, \ldots \mathrm{sign}(p_T(g(x)+M\tilde v))=z_T \}.
\]
Lemma \ref{lem:piecewise-direction} guarantees
that there exists a ball $\mathcal{B}_2$ (centered
on $v$) such that:
\[
\lim_{\delta \downarrow 0} \mathbb{I}_{\bar{S}_{z}} (
\delta v' ) = 1 \, ,
\]
for all $v' \in \mathcal{B}_2$, which is equivalent to Equation
\eqref{eq:Mv'-set}. Thus we have proven the Equation \eqref{eq:claim},
for all $v' \in \mathcal{B}_2$.
As the chain rule holds for directional derivatives (of locally Lipschitz piecewise polynomial/semialgebraic functions)~\citep{Shapiro1990}, we have that:
\[
D[f;v'](x) =D[h;Mv'](y) =\nabla p_z(y)^\top Mv' =u^\top \partial[g;v](x) v'
\]
for $v'$ that is in some sufficiently small full dimensional neighborhood
$\mathcal{B}_3$ (that is contained in $\mathcal{B}_2$ and $\mathcal{B}_1$).
Now define the full dimensional neighborhood $\mathcal{B}_4$ so that it is contained in
both $\mathcal{B}_3$ and $\mathcal{B}$. We have shown, for $v' \in \mathcal{B}_4$,
\[
u^\top \partial[g;v](x) \, v' =\partial[f;v](x)^\top v' \, ,
\]
using Equation~\ref{eq:dir_deriv_pbar}. The proof is completed by choosing $d$ linearly independent $v'$'s, which implies that $u^\top \partial[g;v](x) =\partial[f;v](x)^\top$.
\end{proof}
\subsection{Proof of Lemma~\ref{lemma:black_box}}
\iffalse
\jnote{Uncommented for now, I wanted to read what these were.}
\begin{lemma}[Property of piecewise functions \eqref{eq:piecewise-def-f}]\label{lemma:semialgebraic-full-dim}
Fix $x$ and $v \in \mathbb{S}^{n-1}$.
There exists a $\delta_0 =\delta_0 (v)$ such that for all $0<\delta < \delta_0$, that $x +\delta v \in S_{z(v)}$.
\end{lemma}
\begin{lemma}[Piecewise polynomial fact]
Let $ f : \mathbb{R}^n \to \mathbb{R} $ be a piecewise polynomial
$$
f(u) = \sum_{z} \mathbb{I}_{S_z} (u) p_z (u),
$$
where $S_z \subset \mathbb{R}^n$.
Let $S_{z(u)} $ be the set that contains $u$. For almost all $u$, there is an open neighborhood $ B(u,r)$ such that for all $u' \in B(u,r)$, $ u' \in S_{z(u)}$.
\end{lemma}
\fi
\iffalse
\begin{replemma}{lemma:black_box}
Assume $g$
satisfies the constraint qualification
conditions in Assumption~\ref{assumption:cq}.
Suppose the corresponding representation is given by $g(x)=\sum_{z \in \{-1,1\}^T} \mathbb{I}_{S_z} (x) p_z (x)$,
On an arbitrary input $x$ and $v$, Algorithm~\ref{alg:ASD} returns $g(x)$, $D[g;v](x)$,
and an element $u =\nabla p_z(x)$
where the set index $z$ is such that
$
\lim_{\delta \downarrow 0} (\mathbb{I}_{S_z}(x+\delta v))=1\, .
$
\end{replemma}
\fi
\begin{proof}[Proof of Lemma~\ref{lemma:black_box}]
Let $h_{t,z} (x) $ be polynomials that represent $g_{t,z} $ by expressing in terms of the input variables $x$ only. By the Constraint Qualification Assumption \ref{assumption:cq}, the branch taken by Algorithm \ref{alg:asd-subroutine} is the same branch that would be taken by running on the perturbed input $x+\delta v$ for all $\delta< \delta_0 (v)$.
To finish the proof, we simply note that the reverse mode is performed on a straight-line program that computes $\nabla p_z (x) $, where $z$ is such that $\lim_{\delta \downarrow 0} \mathbb{I}_{S_z} (x+\delta v) =1$.
\end{proof}
\subsection{Completing the proof of Theorem~\ref{thm:main}}
Lemma \ref{lemma:black_box} shows that every library function $g$ that
satisfies Assumption \ref{assumption:cq} can be implemented via
Algorithm \ref{alg:asd-subroutine} to satisfy Assumption
\ref{assumption:library}. Then Lemma \ref{lemma:library} shows that we compute an element $\partial[f;v](x) \in \partial f(x)$.
The second part of the proof counts the number of unit operations. See~\citet{Morgenstern1985HowTC,Griewank89onautomatic} for the counting
argument for a factor of $5$, where the reverse
mode (after we have computed the function and stored the intermediate
variables) is a factor of $4$ more than computing the function
itself.
To obtain a factor of $6$ in the existence claim in
Theorem~\ref{thm:main}, the construction uses a minor modification of
the
algorithm presented, where the algorithm is modified to only run one (global)
reverse mode (as opposed to using a reverse mode computation
associated with each library function). Algorithm~\ref{alg:asd-with-library},
the one presented, has a runtime that is within a factor of $10$ of the cost of
computing just $f(x)$, due to that it runs the reverse mode when
calling every library function (which costs a factor of $4$ more than just
computing $g(x)$). This algorithm presented is what we actually suggest
to use, as the algorithm that achieves a factor of
$6$ may require more memory in order to build a larger computational
graph (and the factor of $10$ is rather pessimistic). To see the factor of $10$, the reverse mode
(associated with each library function) gives an additional factor of $4$.
The directional derivative
costs one additional factor (this is referred to as the forward mode of AD~\citep{Griewank89onautomatic}). After all $T$ (overloaded) library
functions are completed, one additional reverse mode is called at the
end of Algorithm~\ref{alg:asd-with-library}, which incurs (at most) an additional factor of $4$.
To obtain the factor of $6$, it is straightforward to modify the
algorithm, as the following proof shows.
\begin{proof}[Proof of Theorem~\ref{thm:main}]
First, observe that the algorithm correctly computes the function
value on some (limiting branch), so the time it takes to compute
$f(x)$ on this branch is some $T$ where $T \leq
\mathrm{Runtime}^*(f;x)$. Now the directional derivative computation can
be computed when doing the forward pass of the algorithm, which incurs
at most one additional factor of $T$ (this is also referred to as the
forward mode of AD). Instead of computing the intermediate
derivatives $\frac{\partial x_k}{\partial x_j}$ (as in
Algorithm~\ref{alg:ASD}), we will unravel the entire computational
graph of $f(x)$, where we replace each function $g_k(\cdot)$, by
inserting the corresponding computational
graph of $g_k(\cdot)$ into a global computational graph. We then run the reverse mode
of AD on this global computational graph, giving only one additional factor
of $4$.
\end{proof}
\section{Introduction}
The widespread implementation of Automatic Differentiation (AD)
methods \citep{Baydin2015AutomaticDI} has had a transformative effect
on applied machine learning; these methods have eased the difficulty for
practitioners, across a range of disciplines, to learn sophisticated
machine learning models (including deep neural architectures
and richer inferential models). The paradigm is: one simply writes a
program to compute the function of interest, say a scalar (loss) function
$f(x):\mathbb{R}^d \rightarrow \mathbb{R}$, and then a correctly implemented AD method
will return both $f(x)$ and all $d$ of its partial derivatives when
provided with $x$ as an input. These partial derivatives are often
used in conjunction with some (stochastic) gradient-based optimization
approach.
Underlying the effectiveness of this
general black-box approach
is the \emph{Cheap Gradient Principle} \citep{Griewank:2008:EDP:1455489}: the computational cost of computing the
vector of partial derivatives $(\partial f
/ \partial x_1, \partial f
/ \partial x_2, \ldots \partial f
/ \partial x_d)$ is often nearly the same as
that of simply computing the scalar function $f(x)$ itself. In fact,
for all rational functions, the striking Baur-Strassen
theorem~\citep{Baur1983TCo,Griewank89onautomatic} shows
that this increase in computational complexity is a (dimension free) factor of $5$.
In many settings, our underlying function $f(x)$ is a nonsmooth function,
and we resort to subgradient methods.
This work considers the question:
is there a \emph{Cheap Subgradient Principle}? Specifically, given a program that computes a (locally Lipschitz) function $f$ and given a
point $x$, can we automatically compute an element of the (Clarke)
subdifferential $\partial f(x)$ \citep{origin}, and can we do this at
a cost which is comparable to computing the function $f(x)$ itself?
Informally, the set $\partial f(x)$ is the convex hull of limits of
gradients at nearby differentiable points. It can be thought of as generalizing the gradient (for smooth functions) and the subgradient
(for convex functions).
\iffalse
Subgradient methods are a widespread tool in machine learning and
optimization, where automatic differentiation (AD) procedures play a
central role in their implementation. The AD problem for
subdifferentiation is as follows: given a program which computes a
(locally Lipschitz) function $f$ and given a point $x$, compute an
element of the Clarke subdifferential $\partial f(x)$
\citep{origin}. Informally, the set $\partial f(x)$ is the convex hull
of limits of gradients at nearby differentiable
points. It can be thought of as generalizing the gradient (for smooth
functions) and the subgradient (for convex functions).
\fi
\iffalse Commonly used automatic differentiation (AD) libraries for machine learning and optimization
--- including TensorFlow, PyTorch, MxNet, Theano, AutoGrad --- incorrectly
compute (generalized) subderivatives, which can be observed even on
simple examples.
\fi
Let us briefly consider how current approaches handle nonsmooth functions, which are available to the user as functions in some library.
Consider the following three equivalent ways to write the identity
function, where $x\in \mathbb{R}$,
\[
f_1(x) = x, \quad f_2(x) = \relu{x}-\relu{-x}, \quad
f_3(x) = 10 f_1(x)-9f_2(x) \, ,
\]
where $\relu{x} = \max\{x,0\}$, and so
$f_1(x)=f_2(x)=f_3(x)$. As these functions are differentiable at $0$,
the unique derivative is $f'_1(0)=f'_2(0)=f'_3(0)=1$. However, both TensorFlow~\citep{tensorflow2015-whitepaper} and
PyTorch~\citep{paszke2017automatic},
claim that $f'_1(0)=1$, $f'_2(0)=0$, $f'_3(0)=10$. This particular
answer is due to using a subgradient of $0$ at $x=0$. One may ask if a more
judicious choice fixes such issues; unfortunately, it is not difficult to
see that no such universal choice exists\footnote{By defining
$\textrm{ReLU}^{\prime} (0) =1/2$, the reader may note we
obtain the correct derivative on $f_2, f_3$; however, consider $f_4
(x) = \relu{\relu{x}} - \relu{-x}$, which also equals
$f_1(x)$. Here, we would need $\textrm{ReLU}^{\prime} (0) = \frac{\sqrt{5}-1}{2}$ to
obtain the correct answer.}.
This example should be concerning for a number of reasons. The use of
nonsmooth functions in AD go well beyond simple one dimensional
nonsmooth functions (such as $\relu{\cdot}$ or the
$|\cdot |$); current methods permit utilizing eigenvalues, SVDs, QR
decompositions (there are AD procedures on these nonsmooth linear
algebra functions~\citep{autograd,DBLP:journals/corr/abs-1710-08717}).
\iffalse
The difficulty in generalized subdifferentiation is that there is no chain rule for nonsmooth
functions. Furthermore, even if our program is coding a smooth
function itself, where the derivative unambiguously exists, applying
AD to programs that allow for branching may result in incorrect
answers, as is the case here.
\fi
\iffalse
One may immediately ask if there are simple fixes to these issues.
Note that restricting $f(x)$ to be a smooth function does not fix this
issue, as the previous example shows; the ability to utilize
nonsmooth functions when coding $f(x)$ actually leads to incorrect
answers for differentiable $f(x)$.
We may hope that a reasonable approach would be to have a warning message printed if we our AD
method might be returning an incorrect answer (while such an approach
sidesteps the issues of correctness, it could provide more safety).
\fi
\paragraph{Is correctness important?}
One option is to disregard these
issues --- which is the current
state of affairs --- based on the observation that in most cases
these issues are unlikely to harm our optimization method. In numerical linear algebra, one could make the same argument:
we never truly encounter degenerate linear systems (or degenerate
eigenspaces); nonetheless, in retrospect,
numerical issues have made evident the importance of carefully addressing these ``corner cases''. The situation may be analogous here:
numerical issues in these approaches can easily lead to
unstable outputs. Note that some numerical instability is certainly to be expected due to
nonsmoothness (a point we return to in the Discussion under the
notion of \emph{mixed stability});
yet we would still hope to have
nontrivial stability guarantees in our widely used AD libraries,
much in the manner we have for our established numerical linear algebra libraries~\citep{tref-nla-1997,demmel}.
Ultimately, the importance of correctness in these methods is a
decision that must be made by the broader ML community. Here, it is
worthwhile to consider that AD software has a range of
applications: from physical simulators to health care/social science
applications to deployed online learning systems to differentiable
programming. For example, when using physical simulators (say in robotics or in
the sciences), a strong notion of stability may be critical when doing
AD through nonsmooth system dynamics. In safety-critical settings, we
may seek to have deployed online learning methods which are not
susceptible to errors due to misspecified input-output behavior in our
programs. Perhaps the most compelling reason for provably correct
software implementations is to avoid costly failure modes due to the
utilization of the methods in novel and unforeseen manners.
\paragraph{Related Work:}
These issues are in fact known in the mathematical AD literature (see
\citet[Chapter 14]{Griewank:2008:EDP:1455489}). Once we include
either nonsmooth primitive functions or permit branching in a program, the usual chain rule fails to hold and incorrect input-out
behavior is easy to observe. Due to that established calculus
properties of nonsmooth functions~\citep{kk_book,Mordukhovich} do not
seem amenable to AD approaches, the current provable methods
do not have general purpose, computationally
efficient AD methods for subdifferentials.
One influential and powerful idea is that of lexicographic
differentiation~\citep{DBLP:journals/mp/Nesterov05a}; it is a property
of a subclass of nonsmooth functions which allow these function to
inherit a generalized notion of a chain rule. This idea has been
utilized for obtaining correct generalized derivatives in
~\citet{khan_piecewise,Griewank_linearization}. The difficulty is that
lexicographic approach often is expensive in that it involves a
dimensional factor in the computational cost increase.
The other relatively few works that do focus on automatic generalized
differentiation go through some notion of algorithmic
linearization~\citep{dijon,DBLP:journals/mp/Nesterov05a,khan_piecewise,khan_vector,griewank_study},
where often piecewise smooth functions are considered, and the
approach attempts at correct AD through probing the pieces through
some linearization (see~\citet{GRIEWANK2014} for review). The
difficulties are due to understanding what information we can extract
through linear ``probes'' into the function.
One of the first ideas along this line of thought is due
to~\citep{dijon}, which shows how to compute directional
derivatives of nonsmooth functions through following a ``branch'' the
program would take on an input (where the branch corresponds to the
approach direction in the directional derivative). In fact, our work
uses this basic idea, as does the ``branch locking'' approach in
\citet{doi:10.1080/10556788.2017.1341506,Griewank_linearization}. The
difficulty in these approaches is in finding a means to relate this
linearization to properties of the (nonsmooth) functions, which will allow
the algorithm to succeed; naively, we can tell when a method might
have failed though it is difficult to guarantee if it will succeed.
As such, the extant body of work does not contain methods which
contain only a constant factor blow up in the computational cost.
Notable differences in this work is that our assumptions make
strong connections to nonlinear
programming~\citep{abadie,doi:10.1137/1015075,article}, which help in characterizing when the linearization approach is informative, and we
provide a key technical result showing a certain chain rule holds for
randomized algorithms. Furthermore, our focus is on generalizing the
reverse mode for scalar functions (as opposed to focusing on
multivariate functions where there is no known Cheap Gradient Principle).
\paragraph{Our contributions:}
Our main result provides --- under a natural set of assumptions widely
used in nonlinear programming --- a provably correct
Automatic Subdifferentiation procedure, which given some $x$, computes \emph{both} the functional value $f(x)$ and a $d$ dimensional
subdifferential $(u_1,\ldots u_d) \in \partial f(x)$, with a computational cost that is a factor of at most $6$ times that of
computing the scalar function $f(x)$ itself. Our assumption
is that our library of functions be implemented in a manner consistent
with the standard \emph{constraint qualification} assumptions
in nonlinear programming~\citep{abadie}.
In short, this work shows that in fact there is a
\emph{Cheap Subgradient Principle}.
\begin{algorithm}[t]
\caption{Straight Line Program for $f(x)$}
\label{alg:straight_program}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=(x_1, \ldots x_d)$
\FOR{$k=d+1,d+2,\ldots T$}
\STATE Compute:
\[
x_k = g_{k} (x_{\text{parents}(k)})
\]
where
$\text{parents}(k)$ is the index set of the ``parent'' variables of $k$.
\ENDFOR
\item[\textbf{Return:}] $x_T$.
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{The Reverse Mode of AD}
\label{alg:ReverseMode}
\begin{algorithmic}[1]
\item[\textbf{Input:}] variables $(x_1, \ldots x_T)$; a
computational graph $\{\text{children}(t)\}_{t\in \{1,\ldots T\}}$; the
associated derivatives
\STATE Initialize: $
\frac{\partial x_T}{\partial x_T} =1$
\FOR{$t= T,T-1, \ldots 1$}
\STATE Compute:
\[
\frac{\partial x_T}{\partial x_t} = \sum_{i \in \text{children}(t)} \frac{\partial x_T }{\partial x_i} \frac{\partial x_i}{\partial x_t}
\]
\ENDFOR
\item[\textbf{Return:}]
$\frac{\partial x_T}{\partial x}=\left(\frac{\partial x_T}{\partial x_1},\frac{\partial x_T}{\partial x_2},\ldots \frac{\partial x_T}{\partial x_d}\right) $.
\iffalse \jnote{Think of $x_k = g_{k,z}(x_{1:k-1})= p_{k,z}
(x_{1:d})$. For $t\ge d$, $\frac{\partial x_k}{\partial x_t}$ means we
perturb $x_{t:k-1}$, and $\frac{\partial x_k}{\partial x_t}$
means we only perturb $x_t$.} \sk{I HAVE USED PARTIALS}
\fi
\end{algorithmic}
\end{algorithm}
\section{Preliminaries}
\iffalse
In the current work, we focus on the optimization problem
$$\min_{x\in\mathbb{R}^d}~f(x),$$
where $f\colon\mathbb{R}^d\to\mathbb{R}$ is a locally Lipschitz function. The algorithms we consider are based on generalized derivatives, which we now review.
Recall that by Rademacher's theorem, $f$ is differentiable almost everywhere. The {\em Clarke subdifferential} of $f$ at any point $x$ is the set \citep[Theorem 8.1]{clarke2008nonsmooth}
\fi
Assume $f\colon\mathbb{R}^d\to\mathbb{R}$ is a locally Lipschitz
function, and recall, that by Rademacher's theorem, this implies that $f$ is
differentiable almost everywhere. The {\em Clarke subdifferential}
of $f$ at any point $x$ is the set \citep[Theorem
8.1]{clarke2008nonsmooth}
\begin{align}
\partial f(x):=\mathrm{conv}\left\{\lim_{i\to\infty} \nabla f(x_i):
x_i\xrightarrow[]{\Omega} x\right\} \label{eq:clarke-def},
\end{align}
where $\Omega$ is any full-measure subset of $\mathbb{R}^d$ such that $f$ is
differentiable at each of its points. Here, the limit is taken to be the set of all limit points.
In classical circumstances, the subdifferential reduces to more
familiar objects. Namely, when $f$ is $C^1$-smooth at $x$, the
subdifferential $\partial f(x)$ consists only of the gradient
$\nabla f(x)$, while for convex functions, it reduces to the
subdifferential in the sense of convex analysis.
\subsection{AD Review and The Baur-Strassen Theorem}
A \emph{straight line program} for computing $f(x): \mathbb{R}^d \rightarrow \mathbb{R}$ is specified
by a program of the form shown in
Algorithm~\ref{alg:straight_program}. Here the functions $g_1,g_2,\ldots$ are
assumed to be some function from a library of functions. In the
algebraic circuit complexity model, these functions are either
monomials or affine functions of its inputs.
More generally, we will be interested in utilizing a richer class of
functions where $g \in \mathcal{L}$, a library of functions, e.g. we may desire functions like
the $|\cdot|$, $\relu{x}$, or ever richer nonsmooth functions
like eigenvalues.
Define $\mathrm{Runtime}(f;x)$ to be the time it takes to compute $f(x)$ under a given
program for $f$.
\begin{theorem}
\citep{Baur1983TCo,Griewank89onautomatic} Assume all multiplications and additions have unit runtime cost.
If we restrict to the algebraic circuit
complexity model (where the functions $g_k$ are either monomials or
affine functions), then it is possible to compute
both $f(x)$ and all its partial derivatives $\nabla f(x)$ in time that is at most
$5*\mathrm{Runtime}(f;x)$.
\end{theorem}
An algorithm achieving
this guarantee is to first compute $f(x)$ and then use the \emph{reverse
mode} of AD, in Algorithm~\ref{alg:ReverseMode}.
To see the specific counting argument, see
\citep{Morgenstern1985HowTC}. This theorem is often more general: the
reverse mode also correctly returns the derivatives even with a richer
family of smooth
functions in our library $\mathcal{L}$, often with a constant factor cost increase as
well~\citep{Griewank89onautomatic}. The reverse mode
itself has been rediscovered many times
\citep{griewank_invtented}; the well known back-propagation
algorithm~\citep{Rumelhart:1986:LIR:104279.104293} is one
example of the reverse mode of AD. The reverse mode
(and the back-propagation algorithm) is not a
direct application of the chain rule; the direct application of the
chain rule is referred to as the \emph{forward mode} of AD (see
~\citet{Griewank:2008:EDP:1455489}), which is
$d$ times more expensive procedure to compute the gradient. The reverse mode can
be viewed as a form of dynamic programming. To compare the two, in
the reverse mode of AD, we compute the derivatives $\frac{\partial
x_T}{\partial x_t}$, referred to as the adjoints\footnote{For a variable $x_T = g(x_{\text{parents}} )$, the notation $\frac{\partial x_{T}}{\partial x_t}$ refers to the derivative with respect to $x_t$, but holding all parent variables of $x_t$ as fixed. If $x_t$ is an input variable, then this is the usual partial derivative.}, while in the
forward mode of AD we would compute ($d$-dimensional) derivatives of the form $\frac{\partial
x_t}{\partial x}$ (referred to as dual numbers).
\iffalse
One practical concern is that while the reverse mode of AD
is remarkable from a computational perspective, the memory used is often significantly more than computing the function itself, in that it is often difficult to free up various intermediate variables (see \citep{Baydin2015AutomaticDI}). \jnote{Is this relevant to us?}
\fi
\subsection{Nonsmooth functions and our computational model}
To specify how our nonsmooth functions are implemented, we
extend the computational model to allow for branching, using (a
restricted version\footnote{We avoid halting
concerns by assuming our programs halt in a bounded amount of time. We also explicitly avoid discussing tapes and registers in
our computational cost model.} of) the Blum-Shub-Smale model of
computation~\citep{DBLP:conf/focs/BlumSS88}.
\begin{definition}[Computation Model]
The computational model for computing any
$g(x): \mathbb{R}^d \rightarrow \mathbb{R}$ in our library ($d$
may be different for each function) is specified
by a program of the form shown in Algorithm~\ref{alg:branch_program}. We
assume that the function $g_{k,z}$ is either a monomial or an affine function of its inputs.
Furthermore, for every $g$, we assume that there exists
a time $T$, where the program terminates in at most this amount
of time.
\end{definition}
\begin{algorithm}[t]
\caption{Program for a Nonsmooth function $g(x)$ }
\label{alg:branch_program}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=(x_1, \ldots x_d)$
\STATE Initialize a vector $z$ to be all $-1$'s. $z$ is for
notational convenience to keep track of the branch.
\FOR{$k=d+1,d+2,\ldots T$}
\STATE Compute:
\[
x_k = g_{k,z} (x_{\text{parents}(k,z)})
\]
\STATE If the program branches at $(k,z)$, then\\
\begin{itemize}
\item If: $x_k \ge 0$, $z_k=1$.
\item Else: $z_k=-1$.
\end{itemize}
\STATE If the program halts at $(k,z)$, then terminate the \textbf{for} loop.
\ENDFOR
\item[\textbf{Return:}] $x_k$.
\end{algorithmic}
\end{algorithm}
Throughout, we make the following assumption:
\begin{assumption}\label{assumption:cost}
(Computational Cost) Assume all multiplications and additions have unit runtime cost
and that an execution of an ``If''
statement is also unit cost. For example, the cost of computing a
monomial is the number of multiplications.
\end{assumption}
The program implicitly encodes a
function that has the following representation:
\begin{align}
f(x) = \sum_{z \in \{-1,1\}^T} \mathbb{I}_{S_z} (x) p_z (x),
\label{eq:piecewise-def-f}
\end{align}
where each $p_z$ is a polynomial; $\mathbb{I}_{S_z}$ is
the indicator function on the set $S_z$; and $S_z$ consists of
all $x$ where the program executes branch $z$ when given $x$ as
input. The set $S_z$ can be explicitly defined as follows: for
steps $k$ where the programs branches on $z$, define
$h_{k,z}(x)=x_k$; on non-branching $k$, define $h_{k,z}(x)=-1$; define
the vector valued function $h_z(x)=(h_1(z),\ldots h_T(x))$;
\begin{equation} \label{eq:Sz-def}
S_z := \{ x | \, \mathbb{I} (\mathrm{sign}(h_z(x))=z) \}
\end{equation}
where the $\mathrm{sign}(\cdot)$ is the usual sign function (applied componentwise) taking values in
$\{-1,1\}$ (where we take $\mathrm{sign}(0)=1$). Note that $S_z$ is
specified by a set of polynomial inequalities as defined by the
functions $h_{k,z}(x)$.
\iffalse
\begin{remark}
(Relation to the algebraic circuit complexity model and the
Blum-Shub-Smale model of computation) If we restrict to a straight
line program, then this is the standard algebraic circuit complexity model.
With regards to the Blum-Shub-Smale (BSS) model of
computation\citep{bss,braverman},
the functions that halt in
finite time in the BSS model is equal to the set of
functions computable in our computational model.
\end{remark}
\fi
\section{Provable Automatic Subdifferentiation}
\begin{figure}[t]
\begin{minipage}{.45\textwidth}
\begin{algorithm}[H]
\caption{$\relu{x}$}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=x_1$
\STATE Branch:
\begin{itemize}
\item If: $x_1 \ge 0$, set $x_2=x_1$.
\item Else: set $x_2=0$.
\end{itemize}
\item[\textbf{Return:}] $x_2$.
\end{algorithmic}
\end{algorithm}
\end{minipage}
\begin{minipage}{.45\textwidth}
\begin{algorithm}[H]
\caption{$\relu{x}$}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=x_1$
\STATE Branch:
\begin{itemize}
\item If: $x_1^3 \ge 0$, set $x_2=x_1$.
\item Else: set $x_2=0$.
\end{itemize}
\item[\textbf{Return:}] $x_2$.
\end{algorithmic}
\end{algorithm}
\end{minipage}
\caption{\label{fig:relu} Two programs that implement $\relu{x}$:
Both programs are correct and return
the same value. However, the program on the right violates
Assumption \ref{assumption:cq} since the gradient of the constraint
function at $x=0$, $\nabla (x_1 ^3) = 3x_1^2 =0$. }
\end{figure}
In the algebraic circuit complexity model, where AD is provably
correct, branching is not permitted. The inclusion of branching into
our program leads to a number of subtle issues.
Branching allows us to implement the same nonsmooth
function in different manners, which have important
consequences in linearization approaches. Consider two different
programs (with the same input-output behavior) for the $\relu{x}$
function in Figure~\ref{fig:relu}. The left program returns $x$ on the
constraint set that is encoded as $S_1=\{x| x\geq 0\}$, while the right
program returns $x$ on the constraint set that is encoded as $S_1=\{x|
x^3\geq 0\}$. In nonlinear programming, the importance of avoiding
encoding constraints in the latter manner is well-known~\citep{abadie,doi:10.1137/1015075,article}.
This example motivates our
restriction to only consider library
functions that are encoded like the former set.
We will make the
standard constraint
qualification assumption\footnote{The
standard constraint
qualification assumption on a constraint set is that the tangent
cone of the constraint set equals the linearized cone (of the
functions which define the constraints).}. Roughly
speaking, the assumption states that first order information
characterizes the set of feasible perturbations. We state this
assumption in a manner more directly applicable to our setting (see
\citep{abadie,doi:10.1137/1015075,article}).
\begin{assumption}
\label{assumption:cq}
(Constraint Qualification on our Library)
Assume for all $g\in
\mathcal{L}$ that $g$ is locally Lipschitz and our program for $g$ (in our computational model) satisfies the constraint
qualification condition on all sets $S_z$ in the following sense:
suppose $\{h_z\}$
(for binary $z$) are the
corresponding constraint functions in our program. For
any $x,v$ (of the same input dimensionality of $g$), assume that for
all $z$:
\[
\lim_{\delta \downarrow 0} (\mathrm{sign} (h_z(x+\delta v)))
=
\lim_{\delta \downarrow 0}
(\mathrm{sign} (h_z(x)+\delta\nabla h_z (x)\cdot v))
\, .
\]
Roughly, this states that the set approached along the limiting direction $x + \delta
v$, when $\delta \downarrow 0$, can be determined with first order information.
\end{assumption}
Before we state our main theorem, one more definition is in order, due to that
$\mathrm{Runtime}(f;x)$ may not be continuous. Define the
limiting runtime $\mathrm{Runtime}^*(f;x)$ of $f$ at $x$ as the (supremum) runtime
to compute $f(x)$, as $x$ is approached
from nearby points. Precisely,
\[
\mathrm{Runtime}^*(f;x):=\sup \left\{\lim_{i\to\infty} \mathrm{Runtime}(f;x_i):
x_i\rightarrow x\right\},
\]
(where the limit is taken to be the set of all limit points).
\begin{theorem}\label{thm:main}
(A Cheap Subgradient
Principle) Assume that our
program for
$f(x)$, in Algorithm~\ref{alg:straight_program}, is allowed to use nonsmooth functions from our library
$\mathcal{L}$ (in addition to affine functions and monomials).
Suppose assumptions~\ref{assumption:cost} and ~\ref{assumption:cq} hold. There
exists a (randomized) algorithm, which upon input $x$,
terminates in time that is at most
$6*\mathrm{Runtime}^*(f;x)$, and, almost surely, returns both $f(x)$
and an element $u \in \partial f(x)$.
\end{theorem}
The following example shows one subtle issue with regards to
constraint qualification.
\begin{example}\label{example:relu_x_squared}
(Constraint qualification on programs do not compose) Consider the function
$f(x)=\relu{x^2}$ (which is equivalent to the smooth
function $x^2$). It is straight forward to see that the induced program for
$f(x)=\relu{x^2}$ (when we unravel it) does not satisfy the
constraint qualification assumption, even if we do use an
implementation of $\relu{\cdot}$ that does satisfy this
assumption. Regardless, in Example~\ref{example:return}, we show
that our algorithm does indeed correctly compute the gradient on
this (continuous) function.
\end{example}
Before we present the construction, we first provide a chain rule for
nonsmooth functions
\subsection{A Chain Rule for Nonsmooth Functions}
Let $D[g;v](x)$ denote the one-sided (Dini) directional derivative:
\[
D[g;v](x) := \lim_{\delta \downarrow 0} \frac{g(x+ \delta v) -g(x)}{\delta}.
\]
(note that we are not assuming that $v$ is a unit vector). This
derivative exists for all piecewise polynomials and semialgebraic functions~\citep[Lemma 6.2]{coste2000introduction}.
\begin{assumption}
\label{assumption:library}
(Overloading the library with ASD subroutines)
Assume we have a library of (locally Lipschitz) functions
$\mathcal{L}$ computable in our computational model.
For any $g \in \mathcal{L}$, with the
representation $g(x)=\sum_{z \in \{-1,1\}^T} \mathbb{I}_{S_z} (x) p_z (x)$,
assume we have the following associated automatic
subdifferentiation subroutine $\text{ASD}[g]$ with the
following behavior: upon input $(x;v)$, the output $[a, d, u] =
\text{ASD}[g](x;v)$ satisfies
\[
a =g(x), \,
d = D[g;v](x), \,
u = \nabla p_z(x)
\]
where $z$ is such that:
\[
\lim_{\delta \downarrow 0} (\mathbb{I}_{S_z}(x+\delta v))=1\, .
\]
Roughly speaking, $u$ is the derivative determined by the set $S_z$ which is
approached along the limiting direction $x + \delta v$, when $\delta \downarrow 0$.
\end{assumption}
For any locally Lipschitz function $h$, define the limiting total derivate as:
\[
\partial[h;v](x) := \lim_{\delta \downarrow 0} \nabla h(x+ \delta v)
\]
if the limit exists. For almost all $v$, the limit exists, and $\partial[h;v](x)$ is a
subdifferential of $h$.
\begin{theorem}\label{thm:chain}
(A Chain Rule for Nonsmooth Functions)
Assume $h:\mathbb{R}^m \rightarrow \mathbb{R}$ and $g_1$, \ldots
$g_m$ (where $g_i: \mathbb{R}^d \rightarrow \mathbb{R}$) are locally Lipschitz functions computable in our
computational model and that the function $h$ is overloaded with an ASD subroutine
as specified in Assumption~\ref{assumption:library}.
Define:
\[
f(x) := h(g_1(x), \ldots g_m(x)) =h(g(x))\, ,
\]
where $g(x)$ is the vector valued function $(g_1(x), \ldots
g_m(x))^\top$.
Denote the $m\times 1$ vector of (one-sided) directional
derivatives as $D[g;v](x)$. If it exists, let $\partial[g;v](x)$ denote $m
\times d$ limiting Jacobian
matrix (whose rows are given by the vectors $\partial[g_i;v](x)$'s). Set:
\[
[a, d, u] = \text{ASD}[h] (g(x); D[g;v](x))
\]
For all but a measure $0$ set of $v$, we have that $\partial[f;v](x)$
and $\partial[g;v](x)$ exist and that:
\begin{align}
\partial[f;v](x) = \partial[g;v](x)^\top u \, .
\label{eq:chain-rule}
\end{align}
\end{theorem}
\begin{example}
Consider the example $x= f_2 (x) = \relu{x} - \relu{-x}$. We define $h
(y_1,y_2) =y_1- y_2$, $g_1 (x) = \relu{x}$, and $g_2(x) = \relu{-x}$,
so that $f_2 = h(g_1(x),g_2 (x))$. By applying the \text{ASD} subroutine
to $h$, starting at $x=0$ with $v=1$ which leads to running
$\text{ASD}[h] ((0,0);(1,0))=[a,d,u]$ (where it is straightforward to
verify that $u=[1, -1]^\top$), we obtain
\begin{align*}
\partial[f_2;v] (0) &= \partial[g; v](0) ^T u\\
&= \begin{bmatrix}
1 \\ 0
\end{bmatrix} ^\top
\begin{bmatrix} 1 \\ -1 \end{bmatrix}\\
&=1,
\end{align*}
which is correct. Furthermore, note a correct answer is obtained for
any $v \neq 0$.
\end{example}
\begin{example}
We return to $f(x)=\relu{x^2}$ from
Example~\ref{example:relu_x_squared}. Define $h(y)=\relu{y}$, $g(x) = x^2$, and so
$f(x) = h(g(x))$. By applying the chain rule lemma at $x=0$ with $v=1$,
\[
\partial[f;v] (0) = \partial[g; v](0) u = 0\cdot u = 0
\]
Subtly, note that $[a, d, u] = \text{ASD}[h] (0; 0)$, so we are
feeding a degenerate direction $d=0$ into our subroutine. Regardless, the
chain rule lemma still applies (for any $v$ in this case).
\end{example}
\subsection{The algorithm}
We first present the algorithm that utilizes an overloaded library. We then provide a provably correct
construction of this overloaded library. All proofs are
provided in the appendix.
\subsubsection*{Subdifferentiation with the overloaded library}
Algorithm~\ref{alg:asd-with-library} is the Automatic Subdifferentiation
procedure. Correctness follows from Lemma \ref{lemma:library}.
\begin{algorithm}[t]
\caption{Automatic Subdifferentiation}
\label{alg:asd-with-library}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=(x_1, \ldots x_d)$, $v\in R^d$. \\
\item[\textbf{Initialize:}]
Set $\dot{x}_1=v_1,\, \dot{x}_2=v_2, \ldots \dot{x}_d=v_d$.
\FOR{$k=d+1,d+2,\ldots T$}
\STATE Compute $[a, d, u] = \text{ASD}[g_k]
(x_{\text{parents}(k)};\dot{x}_{\text{parents}(k)})$ and set: \label{state:AD-state}
\[
x_k = a, \,
\dot{x}_k = d, \quad
\frac{\partial x_k}{\partial x_{\text{parents}(k)}} = u
\]
\ENDFOR
\STATE Compute $\frac{\partial x_T}{\partial x}$ using the Reverse Mode on these precomputed variables.
\item[\textbf{Return:}] $x_T$,
and $\frac{\partial x_T}{\partial x}$.
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{Overloading the function $g(x)$ }
\label{alg:ASD}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=(x_1, \ldots x_d)$, $v\in R^d$. \\
\item[\textbf{Initialize:}] Set $\dot{x}_1=v_1,\, \dot{x}_2=v_2, \ldots \dot{x}_d=v_d$.
\FOR{$k=d+1,d+2,\ldots T$}
\STATE Compute $x_k$, its partial derivatives, and the directional derivative:
\begin{align*}
x_k &= g_{k,z} (x_{\text{parents}(k,z)})\, , \quad
\left\{ \frac{\partial x_k}{\partial x_j} \middle \vert j \in \text{parents}(k,z) \right\}, \,
\\
\dot{x}_k &= \sum_{j \in \text{parents}(k,z)} \frac{\partial x_k}{\partial x_j} \dot{x}_j
\end{align*}
\STATE If the program branches at $(k,z)$, then:
\begin{itemize}
\item If: $x_k > 0$, then $z_k=1$.
\item Elseif: $x_k = 0$ \emph{and} $\dot{x}_k \geq 0$, then $z_k=1$.
\item Else: $z_k=-1$
\end{itemize}
\STATE If the program halts at $(k,z)$, then terminate the \textbf{for} loop.
\ENDFOR
\STATE Compute $\frac{\partial x_k}{\partial x}$ using the Reverse Mode on these pre-computed variables.
\item[\textbf{Return:}] $[a,d,u] = [x_k, \dot{x_k},\frac{\partial x_k}{\partial x}]$.
\end{algorithmic}
\label{alg:asd-subroutine}
\end{algorithm}
\begin{lemma}\label{lemma:library}
Suppose
Assumptions \ref{assumption:cost} and \ref{assumption:library} hold.
Upon input of an arbitrary $x$, and if $v$ is sampled uniformly at random from
the unit sphere, then, almost surely, Algorithm~\ref{alg:asd-with-library}
returns both $f(x)$ and an element $u \in \partial f(x)$.
\end{lemma}
\begin{proof}
Fix $k \in[d+1,\ldots, T]$. Every parent variable
$j \in \text{parent}(k)$ can be expressed as
$x_j = \tilde{g}_j (x) $, where $g_j$ is a piecewise polynomial on
the $d$ dimensional input $x$. Thus
\begin{align*}
x_k = g_k ( \tilde{g}_1 (x),\ldots, \tilde{g}_{k-1} (x)).
\end{align*}
Now the usual chain rule holds for directional
derivatives~\citep{Shapiro1990}. As the forward mode of AD implements
the usual chain rule of directional derivatives, then we have
$\dot{x}_{j}= D[\tilde g_j;v]$.
By Assumption \ref{assumption:library} and Theorem \ref{thm:chain}, $\text{ASD}[g_k](x_{\text{parents}(k)}, \dot{x}_{\text{parents}(k)})$ returns $u= \frac{\partial x_k}{\partial x_{\text{parents}(k)}}= \partial [g_k ; \dot{x}_{\text{parents}(k)}]$ and this limiting total derivate satisfies the chain rule $\partial[x_k; v] (x) = \partial[\tilde g;v](x) ^\top u$. Since the limiting total derivates satisfies the chain rule and the validity of reverse mode AD algorithm relies only on the chain rule, Algorithm \ref{alg:asd-with-library} correctly computes $\partial [f(x); v]$.
By Rademacher's theorem and the definition of Clarke subgradient in Equation \eqref{eq:clarke-def}, $\partial [f(x);v] \in \partial f(x)$, for almost all $v$.
\end{proof}
\subsubsection*{Overloading the Library Functions}
The following lemma shows that we can provide a method to correctly
overload the library, which we use in Algorithm~\ref{alg:asd-with-library}.
\iffalse
\begin{example}
Figure 1 presents two different possible programs that implement the ReLU activation. Both programs are correct, and return the same value. However, the program on the right violates Assumption \ref{assumption:cq} since the gradient of the constraint function at $x=0$, $\nabla (x_1 ^3) = 3x^2 =0$.
\end{example}
\fi
\begin{lemma}\label{lemma:black_box}
(Correct Library Overloading) Assume $g$
satisfies the constraint qualification
conditions in Assumption~\ref{assumption:cq},
Suppose the corresponding representation is $g(x)=\sum_{z \in \{-1,1\}^T} \mathbb{I}_{S_z} (x) p_z (x)$,
On an arbitrary input $x$ and $v$, Algorithm~\ref{alg:ASD} returns $g(x)$, $D[g;v](x)$,
and an element $u =\nabla p_z(x)$
where $z$ is such that:
$
\lim_{\delta \downarrow 0} (\mathbb{I}_{S_z}(x+\delta v))=1\, .
$
\end{lemma}
\begin{example}\label{example:return}
We again return to $\relu{x^2}$ from Example~\ref{example:relu_x_squared}. Here
we examine how $h$ is overloaded based on the
implementation in Algorithm \ref{alg:ASD}. When $(x,v)=(0,1)$, we are
running $\text{ASD}[h] (0; 0)$ and this may not follow the same
branch had we run on the (infinitesimal) input $x = \epsilon
v$ which leads to running $h(\epsilon^2 v^2)$. However, the gradient is correctly computed, $\partial \relu{x^2} = 0$,
regardless of the branch taken.
\end{example}
\section{Discussion and Open Questions}
{\bf Overloading the Library Functions:} It is not difficult to see that piecewise univariate functions can be
implemented in our library.
\begin{example}{Univariate Piecewise Polynomial (Algorithm~\ref{alg:univariate-poly}).}
Let $\sigma: \mathbb{R} \to \mathbb{R}$ be a univariate piecewise polynomial, meaning that the domain $\mathbb{R}$ is partitioned into a set of $k$ intervals $(-\infty, b_1),(b_1,b_2),\ldots, (b_{k-1},\infty)$. On each interval, the function is equal to a polynomial $p_1,\ldots, p_k$.
Algorithm \ref{alg:univariate-poly} provides a constraint qualified program for the function $\sigma(\cdot)$, which can be used as a library function.
\end{example}
\begin{figure}[t]
\begin{algorithm}[H]
\label{alg:univariate-poly}
\caption{$\sigma(x)$}
\begin{algorithmic}[1]
\item[\textbf{Input:}] $x=x_1$
\STATE Branch:
\begin{itemize}
\item If: $x_1 \le b_1$, set $x_2=p_1 (x)$.
\item Elseif: $x_1 \le b_2$, set $x_2 =p_2 (x)$.\\
~ \vdots\\
\item Elseif: $x_1 \le b_{k-1}$, set $x_2 =p_{k-1} (x)$.\\
\item Else: set $x_2 = p_k (x)$.
\end{itemize}
\item[\textbf{Return:}] $x_2$.
\end{algorithmic}
\end{algorithm}
\end{figure}
An important step would be in extending our computational model to
allow the incorporation of provably correct automatic
subdifferentiation libraries for linear
algebra libraries. AutoGrad~\citep{autograd} does do AD through linear
algebra methods though it can not be used to obtain correct
subdifferentials in programs (at
nondifferentiable points);
obtaining correct generalized derivatives may be particularly important in cases where we deal with low
rank methods. We conjecture our results can be extended, by extending
the computational model, to handle these cases
(there is already much known about the first order structure of
these methods~\citep{DBLP:journals/corr/abs-1710-08717}); technically,
SVDs are not exactly computable in either the algebraic circuit
complexity model or the Blum-Shub-Smale model.
{\bf Numerical Analysis:} The most important open question is how to
obtain numerically stable and accurate
solutions~\citep{tref-nla-1997,demmel}. We conjecture the techniques
developed here will help in characterizing these issues. In
particular, the most natural question is how to develop algorithms that
satisfy the \emph{mixed stability} criterion: the algorithm should give
``nearly the right answer to nearly the right problem'' (as in
~\citep{tref-nla-1997}). For example, for the $\text{abs}(\cdot)$ function, it
should be acceptable for an AD method to provide a subgradient near to
$-1$ for a small input $\epsilon>0$ due to roundoff error; however, it would undesirable for
numerical error to lead vastly different gradients than those that
arise from any nearby
problem. This may be particularly important when doing AD in
physical simulators.
\paragraph{Acknowledgments:} We thank Dima Drusvyatskiy for many
helpful discussions. Sham Kakade acknowledges funding from
Washington Research Foundation Fund for Innovation in Data-Intensive
Discovery, the NSF through award CCF-1740551, and ONR award N00014-18-1-2247. Jason D. Lee acknowledges support of the ARO under MURI Award W911NF-11-1-0303. This is part of the collaboration between US DOD, UK MOD and UK Engineering and Physical Research Council (EPSRC) under the Multidisciplinary University Research Initiative.
|
\section*{}
\vspace{-1cm}
\footnotetext{\textit{$^{a}$~\'{E}cole Normale Sup\'{e}rieure de Lyon, 65 All\'ee d'Italie, 69007 Lyon, France.}}
\footnotetext{\textit{$^{b}$~H.H. Wills Physics Laboratory, Tyndall Avenue, Bristol, United Kingdom. E-mail: <EMAIL>}}
\footnotetext{\dag~Electronic Supplementary Information (ESI) available: [details of any supplementary information available should be included here]. See DOI: 10.1039/cXsm00000x/}
\section{Introduction}
It may be surprising, but crystalline solids can always flow. This happens not only in defective, i.e. finite temperature, solids via the motion of dislocations \cite{seeger1955cxxxii,rice1974ductile}, but also in perfect crystals, where a finite load induces distortions to relax the stresses \cite{ruelle1999,lebowitz1968statistical}. However, the solid's viscosity diverges as the applied deformation vanishes \cite{sausset2010solids}, which is consistent with our everyday experience.
Hence ,ideal crystals can undergo relaxation and observable flow if we are patient enough: yet, it is not clear how ordered solids enter the flowing state and how this process depends on the specific symmetries and composition of the crystal. Theoretical approaches based on nucleation theory \cite{sausset2010solids,nath2018} provide a promising route to account for the emergence of flow. Such approaches are however valid at vanishingly small stresses (or strains): a different scenario is expected for finite, large loads.
In an inspiring work by Sausset \textit{et al.} \cite{sausset2010solids} it was conjectured that (in analogy with ferromagnetic systems \cite{pyragas2000}) the application of oscillatory shear to crystalline solids would reveal a nonequilibrium, dynamical phase transition between a non-flowing regime (elastic, reversible and ordered) and a flowing regime (plastic, irreversible and amorphous) triggered by nucleation at a particular shear amplitude or frequency.
Here we precisely tackle such conjecture, and study a reversible-irreversible dynamical phase transition in several models under oscillatory shear. Depending on the dimensionality of the model and the nature of the ordered state, different order-disorder and reversible-irreversible transitions are observed as a function of the external driving amplitude. The dynamical transition that we observe is distinct from previous disorder-disorder transitions \cite{corte2008} and shear-melting studies \cite{stevens1991shear}. In particular, the transition we are ineterested in is an intrinsically nonequilibrium transition between different periodically driven steady states, as opposed to shear-melting transitions, where crystalline solids are linearly strained until their mechanical failure \cite{stevens1991shear,stevens1993simulations}.
In our analysis, we focus on the generic phenomenology of the transition, exploring how it depends on the dimensionality and the composition of the ordered phase. The transition is explored as a function of the amplitude of the periodic shear and the temperature of a stochastic thermostat coupled to the system. Given the generality of our investigation, we do not explore the effects of the particular orientation of the ordered state with respect to the shear direction or of the particular implementation of the thermostat. These details are known to affect the mechanical response of crystals \cite{stevens1993simulations}, but do not influence the general features of the transition that we discuss.
In our analysis we find that in two dimensions, for one-component systems whose ordered state is hexagonal, a continuous transition takes place. In three dimensions, a discontinuous transition from a flowing to a non-flowing steady state is observed, with a particular scaling for the relaxation time to steady state and the distribution of the amorphous region. suggesting the that an instability analogous to spindoal decomposition dominates the amorphization instead of nucleation only.
The article is organised as follows: we first introduce the three studied models (2D single component Lennard Jones (LJ) system, 3D single component LJ and 3D two-component LJ), and specify the preparation and external driving protocols; we then sequentially discuss the phenomenology of the dynamical in transition for the three models, as identified by their specific order parameters; we finally focus on the 3D two-component LJ case, as it shows the clearest evidence for a dynamical transition driven by a spinodal instability; we summarise and discuss the implications of our observations in the conclusions.
\section{Models and methods}
\label{sec:models}
We perform nonequilibrium molecular dynamics simulations of particle systems in two (2D) and three dimensions (3D) with Lennard Jones interactions. Each system is prepared in an ordered state at different temperatures and deformed according to a sinusoidal uniaxial shearing protocol of period $\tau_0$ and maximal strain amplitude $\gamma_0$. In two dimensions, we consider the LJ hexagonal lattice at fixed reduced density $\rho=0.86$, which is approximately $8\%$ larger than the solid density at the triple point \cite{barker1981}. In three dimensions, we consider two initial solid states: Firstly, the face centered cubic (fcc) LJ crystal, which is the stable phase for this model, at reduced density $\rho=1.0$, with melting temperature $T_{\rm melt}^{LJ3d}=0.96$ \cite{mastny2007}; Secondly, the $\rm Al_2 Cu$ crystal structure corresponding to the low energy crystalline state of a popular LJ nonadditive binary mixture of large (A) and small (B) particles \cite{Kob1994}, prepared at pressure $P=0$ with estimated melting temperature $T_{m}^{BLJ}=0.447(2)$ \cite{crowther2015}.
We study the amorphization of ordered phases in two (2D) and three (3D) dimensions through oscillatory shear. For the sake of simplicity, we focus on systems with spherically symmetric interactions which are initially prepared in their respective thermodynamically stable ordered phases.
The preparation protocol of the several systems is detailed in the following sub-sections.
\subsection{Two dimensional single component case} We consider a single component system of particles interacting via the Lennard-Jones potential
\begin{equation}
V(r)=4\varepsilon \left[ (\sigma/r)^{12}-(\sigma/r)^6\right].
\label{seq:lj}
\end{equation} The parameters $\varepsilon$ and $\sigma$ provide the units for energy and length scales, the particle mass $m$ is set to unity as well as the Boltzmann constant $k_B$, defining a set of reduced units. The potential is truncated and shifted at the conventional cutoff distance $r_{\rm cut}=2.5\sigma$ to speed-up the calculations.
For sufficiently high densities and low temperatures, the system attains an ordered hexagonal phase \cite{gribova2011}. This is a phase with only quasi-long range order, i.e. where spatial correlations decay slowly (algebraically as opposed to exponentially). We directly arrange the particles on an hexagonal lattice at number density $\rho = 0.86$, randomly initialize the velocities and equilibrate the ordered phase in the isothermal-isochoric (NVT) ensemble employing a deterministic Nose-Hoover thermostat with damping time $t_{\rm damp}=0.1$ reduced time units in a range of temperatures $T\in[0.4, 0.6]$. We consider systems of sizes $N=2800,5000$ particles.
In order to characterise the degree of local order, we measure the global hexagonal order parameter $\Psi=|1/N \sum_k \psi_k|$, with
\begin{equation}
\psi_k = N_{\rm neig}^{-1}\sum_{l \in \{ \rm neig\}_k }\exp[6i\phi_{kl}]
\end{equation}
where $\{\rm neig\}_k$ is the list of the $N_{\rm neig}$ neighbours of particle $k$, $\phi_{kl}$ is the orientation of a vector connecting the centres of particle $k$ and the neighbour $l$ with respect to the a reference axis, and $i$ is the imaginary unit.
\subsection{Three dimensions} For the three-dimensional case we consider two different ordered states: a single-component LJ crystal and two-component LJ crystal.
\subsubsection{Single-component case}
We employ precisely the same interaction potential as in Eq.\ \ref{seq:lj} which in 3D leads to the formation of a genuine crystalline face-centered-cubic (fcc) solid phase at sufficiently high densities and low temperature \cite{mastny2007melting}. Here we report results for a particular reduced density $\rho=1.0$ and a range of temperatures $T\in[0.6,1.4]$. Notice that so-called quasi-universality in the Lennard-Jones systems ensures that these specific choices are not particularly important for the overall phenomenology of the dynamical transition under study \cite{bacher2014explaining}.
Similar to the 2D case, the system is prepared in an ordered state through equilibration in the NVT ensemble via a Nos\'e-Hoover thermostat with damping time $t_{\rm damp}=0.1$ reduced units. We consider systems of $N=4000$ particles.
We track structural changes using the ring-statistics \cite{stukowski2012} and Voronoi-based metrics \cite{Malins2013a,lazar2017} to measure face centered cubic (fcc), body centered cubic (bcc) and hexagonal close packed (hcp) arrangements.
\subsubsection{Two-component case}
We also consider a binary crystal with interaction parameters originally fixed by Kob and Andersen to model the behaviour of a metallic glassformer \cite{Kob1994}. Here we have large (A) and small (B) particles interacting via Lennard-Jones potentials tuned to to favour the A-B interaction, so that $\varepsilon_{BB}=0.5\varepsilon_{AA}$, $\sigma_{BB}=0.88\sigma_{AA}$ and the cross interaction parameters are $\varepsilon_{AB}=1.5\varepsilon_{AA}$, $\sigma_{AB}=0.8\sigma_{AA}$. All quantities are expressed in terms of the large particles size $\sigma_{AA}=\sigma_A=\sigma$ and energy scale $\varepsilon_{AA}=\varepsilon$ and mass $m_{AA}=m_{BB}=m$. In the 2A:1B composition the system has been shown to grow at zero reduced pressure $p=0$ the $\rm Al_{2}Cu$ crystalline lattice \cite{crowther2015}. For consistency, we prepare our ordered samples at zero pressure and consider systems of several sizes $N=768,1500,12000, 96000, 768000$. To identify the corresponding density, we construct the equation of state following Vinet's law for Lennard-Jones crystals
\begin{equation}
p = 2 B_0 \, x^3 (x^2 - 1),
\label{eq:vinet}
\end{equation}
where $x=\rho/\rho_0$ and $B_0$ and $\rho_0$ are the bulk modulus and density at zero pressure.
As shown in the Supplementary Information$^\dag$ this expression provides a very good fit to the molecular dynamics results. It also allows us to estimate the temperature dependence of the bulk modulus $B_0$ and of the density at zero pressure $\rho_0$. A linear fit to low temperatures provides an estimate for the density at zero pressure and temperature $\rho_0(T=0) \approx 1.48$ in agreement with previous estimates \cite{Fernandez:2003he}.
For the binary crystal $\rm Al_2 Cu$, we use ring-statistics \cite{stukowski2012} and Voronoi-based metrics \cite{Malins2013a,lazar2017} to measure the number of B particles in the center of a bicapped square antiprism (simply \textit{antiprism} in the following), which is the distinctive motif of the crystal unit cell and a recurrent local motif (or \textit{locally favourite structure}) in the melt \cite{coslovich2007,malins2013faraday}.
\subsection{Oscillatory shear protocol}
For our nonequilibrium protocol, we shear the several systems according to a sinusoidal deformation as implemented in the molecular dynamics package LAMMPS \cite{plimpton1995fast}. The deformation is uniaxial along the x direction so that the strain as a function of time oscillates as in
\begin{equation}
\gamma(t) = \gamma_0 \sin(2\pi t/\tau_0),
\label{eq:protocol}
\end{equation}
where $\gamma_0$ is the maximum shear amplitude and $\tau_0$ is the period of the oscillations. The obtained triclinic box is periodic in all dimensions (effectively implementing Lees-Edwards's boundary conditions).
During the deformation, the temperature of the system is kept constant by a Langevin thermostat whose characteristic damping time is $t_{\rm damp}=100dt$, where $dt$ is the time step for the integration of the equations of motion, set to $dt=0.001$. For the different Lennard-Jones models considered here, we keep the period of the oscillations constant to $\tau_0=10$ in reduced Lennard Jones units and vary the other physical parameters, such as the amplitude of the oscillations or the temperature.
We typically consider 10 distinct realizations of the dynamics for a given set of parameters, starting from different initial conditions for both the prepared ordered state and the particle velocities, initialized from a pseudo-random sample of the Maxwell-Boltzmann distribution.
\section{Results on the order-disorder dynamical transition} We are interested in the transition between two distinct nonequilibrium steady states (NESS): an ordered NESS, where the system remains in the prepared ordered state, and an amorphous (or only semi-crystalline) NESS, attained while the system is continuously subject to oscillatory shear and to modest thermal fluctuations.
We start with the analysis of the 2D LJ system; continue with the 3D LJ single component system and finally focus on the 3D two component (binary) LJ system.
\subsection{Single-component 2D Lennard-Jones hexagonal solid}
Two-dimensional LJ disks, in absence of external driving, undergo a liquid-hexagonal transition which proceeds via a Kosterlitz-Thouless-Halperin-Nelson-Young mechanism, i.e. through a so-called hexatic phase \cite{Wierschem2011,gribova2011} in a weakly first-order phase transition. The equilibrium ordered phase is known to be characterised by a mosaic of locally ordered (hexagonal) patches and amorphous structures \cite{patashinski2010}. Here we promote the amoprhization of the hexagonal ordered state through sinusoidal shear of variable strain amplitude at several temperatures, $T\in[0.4,0.6]$. Starting from a thermalised hexagonal state, we monitor the time evolution of the global orientation order parameter $\Psi$ for different values of the maximum strain amplitude $\gamma_0$, see Fig.~\ref{figLJ2d}(a). For small values of $\gamma_0$, the system responds elastically, and at every cycle it reversibly attains the initial ordered state characterised by large values of $\Psi$. However, as we increase the maximum amplitude, we observe that the orientational order parameter gradually relaxes to lower and lower values. We fit the relaxation curves with a stretched exponential $\Psi(t)=\Psi_{\infty}+(\Psi_0-\Psi_\infty)\exp[-(t/\tau)^b]$, from which we derive the relaxation time $\tau$, the long and short time value of the order parameter $\Psi_\infty$ and $\Psi_0$ and the stretching exponent $b$ which tends to unity from below as we approach larger strain amplitudes.
\begin{figure}[t]
\centering
\includegraphics{figLJ2d}
\caption{\textbf{Dynamical order-disorder crossover in the 2D hexagonal lattice.} (a) Evolution of the single-run global hexagonal order parameter $\Psi$ for $\gamma_0\in[0.5,0.26]$. Dotted black lines are stretched exponential fits. (b) Steady state $\Psi_{\infty}$ for different temperatures, averaged over several trajectories. (c) Relaxation time to the steady state $\tau$ in units of the oscillation period $\tau_0$ as a function of the distance from the crossover strain amplitude $\tilde{\gamma}$. Full relaxation within the time of the simulations is for $\gamma_0>\tilde{\gamma}$ (filled symbols).}
\label{figLJ2d}
\end{figure}
As shown in Fig.~\ref{figLJ2d}(b), the transition from an ordered to a disordered steady state occurs over a broad range of strain amplitudes $\gamma_0$ at all the considered temperatures, so that we can successfully fit it with a hyperbolic tangent function $\Psi_{\infty}(\gamma_0)=a+b\tanh[(\gamma_0-\tilde{\gamma})/c]$, where $\tilde{\gamma}$ is the fitting parameter corresponding to the inflection point of the curve.
Plotting the resulting relaxation times $\tau(\gamma_0)$ as a function of the distance from the inflection point $\tilde{\gamma}$ we manage to collapse the data from different temperatures on a semilogarithmic plot, Fig.~\ref{figLJ2d}(c), suggesting that the temperature dependence is mainly encoded in the crossover value $\tilde{\gamma}$. Only for $\gamma<\tilde{\gamma}$ the collapse fails, but this is because reliable values of the relaxation time cannot be obtained through the stretched exponential fit when the system relaxes on time scales larger than the simulation time scale. For $\gamma>\tilde{\gamma}$, we observe that in the vicinity of $\tilde{\gamma}$, the relaxation time increases by less than two orders of magnitude with respect to the large $\gamma_0$ regime.
The transition can be analysed in terms of its effect on local orientational order, in terms of the the complex number $\psi_i$ associated to every particle $i$. The spatial distributions of its modulus and argument allow us to keep track of the formation of boundaries between differently oriented regions.
In Fig.\ \ref{fig:2dorder} we present configurations color-coded by the values of the argument and the modulus of $\psi_i$ respectively, for a particular temperature $T=0.4$ and strain amplitude $\gamma_0=0.22$ over several periods of oscillatory shear.
Both the modulus and the argument of $\psi_i$ identify domains of coherence and local order. However, grain boundaries lacking 6-fold coordination are more immediately identified in the modulus, as low values of $|\psi_i|$ typically correspond to four-fold coordinated particles.
The reduction of quasi-long range order over time takes place through the formation of extended defect lines which break the lattice into sub-domains of smaller sizes. This demonstrates that orientational correlations span over smaller scales as the system reaches the disordered steady state. The size of the orientational coherence can be estimated from the decay of the conventional two-point correlation function
\begin{equation}
g_6(r)=\langle \psi^*({\bf r}) \psi({\bf r})\rangle,
\end{equation}
where the average here is only performed on all the particles at a given time $t$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{local2d}
\caption{\textbf{Dissolution of the local hexagonal order for two-dimensional system.} Color-coded configurations at temperature $T=0.40$ and strain amplitude $\gamma_0=0.22$. Top row: local order as detected by the angular part of the local orientational order parameter $\psi_i$. Bottom row: the same, as detected by the modulus of $\psi_i$. Columns (a,d), (b,e) and (c,f) correspond to identical time frames $t=0, 50,100\tau_0$.}
\label{fig:2dorder}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics{g6}
\caption{\textbf{Orientational pair correlation function in the 2D LJ system}. At temperature $T=0.40$ for three distinct values of the strain amplitude $\gamma_0=0.15,0.17,0.22$ across the order-disorder crossover, the orientational correlations as measured by $ g_6(r)=\langle \psi^*({\bf r}) \psi({\bf r})\rangle$ decay very mildly for small strain amplitudes, and on much shorter scale for large amplitudes. In every, correlations are plotted after $0,50,100$ shear cycles. Dashed lines are exponential fits to extract the correlation length $\xi$. }
\label{fig:g6}
\end{figure*}
In Figure \ref{fig:g6} we plot $g_6(r)$ across the order-disorder crossover at a relatively cold temperature $T=0.4$ for after 0, 50 and 100 shearing cycles of duration $\tau_0$. For $\gamma_0=0.15$ correlations are strong well beyond ten particle diameters. In fact, the decay of the correlations can be fitted equally weel by exponential or power law decays, indications that correlations are quasi-long range, as in the hexatic phase of equilibrium disks. However, across the transition, orientational correlations decay more steeply and undergo a time evolution which is more rapid as the strain amplitude increases. At large strain amplitudes, $\gamma_0=0.22$, orientational correlations decay over a length $\xi(\gamma_0=0.22)\approx5\sigma$.
Hence, in two dimensions, we observe a dynamical \textit{continuous} crossover taking place through the formation of large regions of different orientations, which become progressively smaller as the strain amplitude is increased. Notice that at the chosen temperature and densities the quiescent (i.e. not driven) system would simply form an ordered hexagonal phase. The driven system, instead, present a wide spectrum of possible steady states of increasing orientational order as the amplitude of the strain oscillations is reduced.
\subsection{Single-component 3D Lennard-Jones crystal}
A very different scenario emerges from the three dimensional simulations. We first consider the single-component fcc crystal of LJ particles sheared along the $\langle100\rangle$ orientation. We follow a protocol analogous to the two-dimensional case, but now we monitor (via Common Neighbour Analysis \cite{stukowski2012,stukowski2009} and Voronoi cell analysis as implemented in VoroTop \cite{lazar2017}) a variety of multiple crystalline motifs: fcc, hcp, bcc. We prepare an equilibrated fcc crystal at reduced density $\rho=1.0$ at temperatures $T\in [0.6,1.4]$, from around the triple point to well above the melting temperature. While at high temperature bulk melting dominates, interesting transitions emerge at low temperatures. In Fig.~\ref{figLJ3d} we present the case of $T=0.6$.
As we increase the shear amplitude $\gamma_0$ the crystalline cells become highly distorted, so that a perfect fcc crystal is transformed into a defective fcc/hcp crystal with bcc inclusions, see Fig.~\ref{figLJ3d}(a) and \dag. However, for $\gamma_0\sim0.26$ (corresponding to the constrained tension ideal strain of the fcc LJ crystal of $\gamma_{\rm ideal}=0.26$ \cite{macmillan1972}) an abrupt transition to a non-crystalline steady state is observed. Interestingly, the structure of the crystal for $\gamma_0\rightarrow \gamma_{\rm ideal}$ is dominated by the bcc order instead of the fcc order. This is somehow surprising: for the quiescent bulk, the free energies of the hcp and fcc structures are close (with the fcc crystal being marginally lower in free energy) while the bcc crystal is unstable \cite{travesset2014phase}. However, it has been suggested that precritical nuclei are formed in the bcc arrangement \cite{rein1996numerical}: in our case, the oscillatory protocol appears to allow such nuclei to grow and become stable at steady state.
The transformation from fcc to bcc for $\gamma_0<\gamma_{\rm ideal}$ occurs with very limited amorphization: the rearrangement of the crystalline defects is sufficient to cause the change in crystalline symmetry. However, for larger $\gamma_0$ and in particular for $\gamma_0=0.28$ we observe partial re-crystallization of the system: as depicted in Fig.\ref{figLJ3d}(b) the initial fcc crystal is initially turned into a largely amorphous packing in a few cycles. However, bcc crystal nucleation and growth occurs in the following cycles so that the system attains a largely re-crystallized NESS. Hence, amorphization and bcc-nucleation are competing processes for $\gamma_0\sim \gamma_{\rm ideal}$. In such a regime, the time to attain an amorphous steady state $\tau$ is longer than the nucleation time for the bcc crystal $\tau_{\rm bcc}$: this means that the fcc-order to disorder transition is mediated by an intermediate family of nonequilibrium bcc steady states.
\begin{figure}[tp]
\centering
\includegraphics{figLJ3d}
\caption{\textbf{Dynamical order-disorder transition in the 3d single-component fcc crystal.} (a) Fraction of particles in crystalline arrangements (fcc, bcc, hcp and the sum crys=fcc+hcp+bcc) as a function of $\gamma_0$. Inset: bcc nucleation time $\tau_{bcc}$ with respect to $\gamma_0$. (b) Time evolution of the structure of the sheared crystal at $\gamma_0=0.28$ at time $t=0,5, 200 \tau_0$, color-coded as in (a).}
\label{figLJ3d}
\end{figure}
\subsection{Two-component Lennard-Jones crystal}
The results on the single-component Lennard-Jones crystal motivate us to consider a system less prone to re-crystallization. Previous studies on glass-forming systems have shown that a non-additive 2:1 mixture of large and small Lennard-Jones particles is highly stable to crystallization \cite{crowther2015}, probably due, firstly, to the vicinity of its stoichiometry to the eutectic point of the mixture \cite{pedersen2018} and secondly to the fact that binary crystals fail differently under shear compared to one-component crystals \cite{horn2014does}. The mixture crystallizes into the $\rm Al_2 Cu$ arrangement, whose unit cell is characterized by spindles of small particles surrounded by square rings of large particles, forming a \textit{bicapped square antiprism} \cite{fernandez2003}. In fact, this is the same symmetry of recurrent geometrical motifs (also called \textit{locally favoured structures}) in the liquid and supercooled liquid states of the same model, which play an important role in the emergence of slow relaxation within the glassy state \cite{coslovich2007,malins2013faraday, turci2017}. It has been shown that the number of particles involved in such motifs can be large in the liquid at low temperature without any sign of spontaneous crystal formation \cite{crowther2015}. We therefore expect that a large enough deformation of an equilibrated configuration in the $\rm Al_2 Cu$ crystalline arrangement of the model would lead to a substantial amorphization of the material which would be irreversible, with no secondary nucleation processes.
\begin{figure*}[ht]
\centering
\includegraphics{figBLJ.pdf}
\caption{\textbf{Dynamical order-disorder transition in the 3d binary crystal.} (a) Fraction of particles $n(t)$ forming bicapped square antiprisms for $\gamma_0\in[0.19, 0.5] $ at $T=0.336$. Dashed lines are stretched exponential fits, defining the relaxation time $\tau$. (b) Steady-state values of $n$ for different temperatures: as a function of $\gamma_0$ we observe a sharp transition around $\tilde{\gamma}\approx0.19$ between an ordered and disordered dynamical phase. In (c) and (d) we present two empirical scalings of the relaxation time $\tau$: (c) a collapse with respect to $\log(x)/x$ with $x=\gamma_0/\tilde{\gamma}$ and a critical-like scaling $\tau\propto (\gamma_0-\tilde{\gamma})^{-\nu}$, from which $\nu\approx3.0$. }
\label{figBLJ}
\end{figure*}
We monitor for this purpose the specific crystalline order of the binary crystal employing Voronoi face analysis (in which the antiprism can be detected with the [0,2,8] signature) as well as the Topological Cluster Classification method \cite{Malins2013a}. We have seen that in the two-dimensional case the increase of disorder is gradual in time, while in the three-dimensional one-component LJ this is mediated by the nucleation of a bcc phase within the original fcc phase. In the case of the BLJ crystal (for a choice of period $\tau_0=10$ in Lennard-Jones time units with respect to the large particles) we directly observe the formation of disordered pockets, without the interference of intermediate crystal formation. The time evolution of the fraction of particles in antiprisms, Fig.~\ref{figBLJ}(a), shows that when the shear amplitude is small ($\gamma_0\lesssim 0.19$), the system remains ordered for several hundreds of cycles before rapidly and irreversibly falling into the amorphous state.
In Fig.~\ref{figBLJ}(b) we plot the steady state value of the fraction of particles in an antiprism for several values of the temperature. As expected, the transition from a non-flowing to a flowing regime is very sharp with no further crystallisation beyond a characteristic value of $\tilde{\gamma}$. We also emphasize that higher temperatures correspond to slightly smaller steady state populations of antiprisms, both in the non-flowing and in the flowing dynamical phase.
As suggested by Fig.~\ref{figBLJ}(a) the relaxation time $\tau$ at which the irreversible plastic event occurs increases as $\gamma_0\rightarrow\tilde{\gamma}^+$. In Fig.~\ref{figBLJ}(c) and (d) we show two different scalings of this relaxation times: inspired by the estimate $\log \tau\propto \log(x)^\delta/x$ with $x=\gamma_0/c$ proposed in \cite{sausset2010solids} for the nucleation of amorphous patches in the limit of small stresses (where $c$ is a material specific constant and $\delta=3$ in the original calculations in three dimensions) we explored a $\log(x)/x$ scaling finding that only $\delta=1$ would account for our data, Fig.~\ref{figBLJ}(c). However, if for $\gamma<\tilde{\gamma}$ the steady states are truly irreversible, then $\lim_{\gamma_0\rightarrow\tilde{\gamma}^+}\tau=+\infty$, so we also model the singularity at $\tilde{\gamma}$ with a critical power law scaling, obtaining an equally good collapse of the data, see Fig.~\ref{figBLJ}(d), with $\tau\propto(\gamma_0-\tilde{\gamma})^{-\nu}$ with $\nu\approx3.0$.
\section{Analogy to spinodal decomposition in the 3D binary LJ crystal}
The 3D binary crystal presents a clear transition at a specific value of the strain amplitude and the timescales appear to diverge in a fashion compatible with scale-free, power law scaling. This suggests the possibility that the transition is dominated by a barrier-less process akin to spinodal decomposition, with the timescales for the transition from the reversible solid NESS to the irreversible amorphous NESS being controlled by the coarseing of the amorphous regions.
To test this, we further investigate the scaling behaviour measuring the growth of amorphous patches within the crystal in large simulations of up to $N=768000$ particles. We identify connected regions without antiprisms and measure the number of particles $K$ in each cluster. In Fig.~\ref{figlengths} (a) we show how the largest value $K^{\rm m}$ increase as time evolves for different $\gamma_0$ observing that $K^{m}\propto t^{\mu}$ before reaching a plateau value once the system turns amorphous, with the exponent reaching $\mu\approx 2$ for the smallest $\gamma_0$ for which we still observe amorphization. If we assume that the cluster length is approximately $\ell=\sqrt[3]{K}$ we have $\ell\propto t^{1/z}$ with the dynamical exponent $z=3/\mu\approx3/2$ to be compared with $z=2$ and $z=3$ of Model A (e.g. Ising) and model B (e.g. Cahn-Hilliard model for spinodal decomposition) respectively \cite{hohenberg1977theory}. At the same time, the probability distribution of the cluster sizes $P(K)$ appears to be inherently scale-free, as shown in Fig.~\ref{figlengths} (b) for the case of $\gamma_0=0.21$ at $T=0.336$, where, for the largest accessible system, we measure $P(K)\propto1/K$ until one full relaxation time. The spatial distribution of clusters corresponds to a branched network of amorphous regions across the system, as pictorially illustrated in the snapshots in Fig.~\ref{figlengths} (c), where the amorphous domains appear to span across the length of the system. The growth pattern of the amorphous network is reminiscent of coarsening or spinodal decomposition, occurring in other protocols of fast energy injection, such as superheating phenomena \cite{belonoshko2007}.
\begin{figure}[t]
\centering
\includegraphics{figlengths}
\caption{\textbf{Growth of amorphous regions in the 3D binary crystal.} (a) Number of particles in the largest amorphous cluster $K^{\rm m}$ for a BLJ crystal of $N=12000$ particles at $T=0.336$ for several $\gamma_0$. Time $t_0$ is chosen so that $K_0=K^m(t_0)=100$ particles. Dashed lines are power-law fits $K^{\rm m}(t)\propto t^{\mu(\gamma_0)}$ and in the inset we report $\mu(\gamma_0)$. (b-c) Time evolution of a system of $N=768000$ particles at $\gamma_0=0.215$ and $T=0.336$: (b) probability distribution of the sizes of amorphous patches, with the $1/K$ slope (dashed line); (c) successive snapshots of the system, with crystal/amorphous interfaces (green surfaces, grey the periodic faces) dotted by antiprism particles (blue and red).}
\label{figlengths}
\end{figure}
As a second test for a spinodal-like scenario, we consider the seeding of amorphous regions within the solid. In analogy to studies of crystal nucleation in metastable liquids \cite{espinosa2016seeding}, we probe whether the sharp transition in the binary Lennard-Jones mixture follows a nucleation/growth scenario through the seeding of disordered regions within the crystalline state.
We consider a system of $N=12000$ particles at low temperature $T=0.336$ and a strain amplitude of $\gamma_0=0.21$, 10\% larger than the estimated critical strain $\tilde{\gamma}\approx0.19$. In such conditions, we know that the relaxation from the crystal to the liquid occurs on timescales of the order of $10^3 \tau_0$ (see Fig.\ 3(d) in the main text).
\begin{figure*}[bt]
\centering
\includegraphics{nuclei}
\caption{\textbf{Growth of seeded amorphous regions in the BLJ crystal.} (a) Time evolution of the fraction of particles in crystalline regions for different values of the radius $R$ of the seeded amorphous nucleus, at temperature $T=0.336$ and strain amplitude $\gamma_0=0.21$. Dashed lines are linear fits to the initial decay, defining the growth speed $v=dn/dt|_{t=0}$ (b) Average growth speed $\langle v(r)\rangle$ as a function of the radius of the seeded amorphous nucleus: a linear fit points to zero growth speed only for nuclei of size $R\approx1\sigma_A$. See video in the Electronic Supplementary Material$^\dag$.}
\label{fig:seed}
\end{figure*}
We first prepare a crystalline arrangement of particles through the equilibration procedure described in Sec.\ \ref{sec:models}. We then identify particles outside a given radius $R$ from the origin of the reference frame and freeze their positions. For the particles within the radius $R$ we reassign their velocities, randomly selected them from a Maxwell-Boltzmann distribution, so that their kinetic temperature is $T_{\rm seed}=10$ and solve the equation fo motion (with frozen external particles and with no shear) for a time of approximately 100$\tau_0$. At such high temperatures the particles can cross each other more easily and the spherical region becomes disordered. Following this route we prepare 10-30 initial configurations from which we start oscillatory shear simulations. We track the number of particles in crystalline motifs $N_x$ and the relative fraction $n=N_x/N$ over time and obtain single-run curves as in Fig.\ \ref{fig:seed}(a). We define the growth speed
\begin{equation}
v = \left.\frac{dn}{dt}\right|_{t=0}
\end{equation}
and numerically evaluate it through a linear fit at early times. Averaging over several realizations, we obtain $\langle v(R)\rangle$ as a function of the radius of the disordered seed, see Fig.\ \ref{fig:seed}(b). We observe that we only measure positive growth speeds and that a linear fit through the data points $\langle v(R)\rangle=0$ only for $R\approx1$. Such a small size for the initial seed can hardly be called a nucleus. See video in the Supplementary Information $\dag$ for an example trajectory.
From the seeded nucleation analysis we infer that amorphous nuclei of very small sizes appear to lead to the irreversible transition from the ordered to the amorphous oscillatory steady state. We interpret this as a further indication in support of a spinodal-like instability scenario.
\section{Conclusions}
Under the action of an oscillatory external driving, atomistic models of ordered states in two and three dimensions undergo a novel nonequilibrium dynamical transition between a non-flowing, ordered phase and a flowing amorphous phase. This -- originally only conjectured -- dynamical transition is revealed to have a very rich phenomenology: in 2D, it appears merely as a broad crossover between a disordered state with residual orientational order and an ordered hexagonal phase; in 3D, it corresponds to a sharp transition that can be masked by the nucleation of ordered phases with symmetries different from the ones of the unperturbed solid (i.e. bcc instead of fcc); moreover, with a suitable choice of the ordered phase, as in the case of the binary Lennard-Jones crystal, the 3D transition occurs abruptly in a narrow range of shear amplitudes with a scaling of the time required to relax to the amorphous state which is compatible with power-law, critical scaling.
The sharpest transition is therefore observed in the case of the two-component system with non-additive Lennard-Jones interactions, which prevent recrystallization across the dynamical transition. In the case of this particular system, the numerical results suggest that -- even for small systems -- one can define a critical strain amplitude $\gamma_0$ at which amorphous pockets are formed and start growing irreversibly, in a process akin to coarsening. The relatively high density of amorphous pockets, their structure and growth are reminiscent of spinodal decomposition as opposed to a nucleation scenario, which is confirmed by the study of the spontaneous growth of very small seeded amorphous regions within the ordered phase. Coarsening of the emergent disordered phase appears to play a role as important as the nucleation-based description described in Ref.\cite{sausset2010solids}, which dominates the onset of flow in the small strain limit.
The phenomenology presented here can be tested in experiments, for example with colloidal crystals in shear cells, both in two and three dimensions. Further theoretical study will focus on the possible implications of the present dynamical transitions for amorphous solids such as glasses and gels, where critical behaviour is believed to play an important role in the emergence of rigidity \cite{jaiswal2016mechanical,parisi2017shear,ozawa2018random}.
\\
The authors would like to thank T.\ Speck, J.\ Lam, C.\ Valeriani, N.\ Wilding and R.\ L.\ Jack for insightful discussions. This work was carried out using the computational facilities of the Advanced Computing Research Centre, University of Bristol. FT acknowledges the support of the European Research Council under the FP7/ERC Grant agreement n$^{\circ}$617266 NANOPRS.
\balance
\bibliographystyle{rsc}
\providecommand*{\mcitethebibliography}{\thebibliography}
\csname @ifundefined\endcsname{endmcitethebibliography}
{\let\endmcitethebibliography\endthebibliography}{}
\begin{mcitethebibliography}{40}
\providecommand*{\natexlab}[1]{#1}
\providecommand*{\mciteSetBstSublistMode}[1]{}
\providecommand*{\mciteSetBstMaxWidthForm}[2]{}
\providecommand*{\mciteBstWouldAddEndPuncttrue}
{\def\unskip.}{\unskip.}}
\providecommand*{\mciteBstWouldAddEndPunctfalse}
{\let\unskip.}\relax}
\providecommand*{\mciteSetBstMidEndSepPunct}[3]{}
\providecommand*{\mciteSetBstSublistLabelBeginEnd}[3]{}
\providecommand*{\unskip.}}{}
\mciteSetBstSublistMode{f}
\mciteSetBstMaxWidthForm{subitem}
{(\emph{\alph{mcitesubitemcount}})}
\mciteSetBstSublistLabelBeginEnd{\mcitemaxwidthsubitemform\space}
{\relax}{\relax}
\bibitem[Seeger(1955)]{seeger1955cxxxii}
A.~Seeger, \emph{The London, Edinburgh, and Dublin Philosophical Magazine and
Journal of Science}, 1955, \textbf{46}, 1194--1217\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Rice and Thomson(1974)]{rice1974ductile}
J.~R. Rice and R.~Thomson, \emph{Philosophical magazine}, 1974, \textbf{29},
73--97\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Ruelle(1999)]{ruelle1999}
D.~Ruelle, \emph{Statistical mechanics: Rigorous results}, World Scientific,
1999\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Lebowitz(1968)]{lebowitz1968statistical}
J.~L. Lebowitz, \emph{Annual Review of Physical Chemistry}, 1968, \textbf{19},
389--418\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Sausset \emph{et~al.}(2010)Sausset, Biroli, and
Kurchan]{sausset2010solids}
F.~Sausset, G.~Biroli and J.~Kurchan, \emph{Journal of Statistical Physics},
2010, \textbf{140}, 718--727\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Nath \emph{et~al.}(2018)Nath, Ganguly, Horbach, Sollich, Karmakar, and
Sengupta]{nath2018}
P.~Nath, S.~Ganguly, J.~Horbach, P.~Sollich, S.~Karmakar and S.~Sengupta,
\emph{Proceedings of the National Academy of Sciences}, 2018, \textbf{115},
E4322--E4329\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Pyragas \emph{et~al.}(2000)Pyragas, Lange, Letz, Parisi, and
Kittel]{pyragas2000}
K.~Pyragas, F.~Lange, T.~Letz, J.~Parisi and A.~Kittel, \emph{Physical Review
E}, 2000, \textbf{63}, 016204\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Cort{\'{e}} \emph{et~al.}(2008)Cort{\'{e}}, Chaikin, Gollub, and
Pine]{corte2008}
L.~Cort{\'{e}}, P.~M. Chaikin, J.~P. Gollub and D.~J. Pine, \emph{Nat Phys},
2008, \textbf{4}, 420--424\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Stevens \emph{et~al.}(1991)Stevens, Robbins, and
Belak]{stevens1991shear}
M.~J. Stevens, M.~O. Robbins and J.~F. Belak, \emph{Physical review letters},
1991, \textbf{66}, 3004\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Stevens and Robbins(1993)]{stevens1993simulations}
M.~J. Stevens and M.~O. Robbins, \emph{Physical Review E}, 1993, \textbf{48},
3778\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Barker \emph{et~al.}(1981)Barker, Henderson, and Abraham]{barker1981}
J.~Barker, D.~Henderson and F.~F. Abraham, \emph{Physica A: Statistical
Mechanics and its Applications}, 1981, \textbf{106}, 226--238\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Mastny and de~Pablo(2007)]{mastny2007}
E.~A. Mastny and J.~J. de~Pablo, \emph{The Journal of chemical physics}, 2007,
\textbf{127}, 104504\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Kob and Andersen(1994)]{Kob1994}
W.~Kob and H.~C. Andersen, \emph{Phys. Rev. Lett.}, 1994, \textbf{73},
1376--1379\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Crowther \emph{et~al.}(2015)Crowther, Turci, and Royall]{crowther2015}
P.~Crowther, F.~Turci and C.~P. Royall, \emph{The Journal of Chemical Physics},
2015, \textbf{143}, 044503\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Gribova \emph{et~al.}(2011)Gribova, Arnold, Schilling, and
Holm]{gribova2011}
N.~Gribova, A.~Arnold, T.~Schilling and C.~Holm, \emph{The Journal of chemical
physics}, 2011, \textbf{135}, 054514\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Mastny and de~Pablo(2007)]{mastny2007melting}
E.~A. Mastny and J.~J. de~Pablo, \emph{The Journal of Chemical Physics}, 2007,
\textbf{127}, 104504\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Bacher \emph{et~al.}(2014)Bacher, Schr{\o}der, and
Dyre]{bacher2014explaining}
A.~K. Bacher, T.~B. Schr{\o}der and J.~C. Dyre, \emph{Nature communications},
2014, \textbf{5}, 5424\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Stukowski(2012)]{stukowski2012}
A.~Stukowski, \emph{Modelling and Simulation in Materials Science and
Engineering}, 2012, \textbf{20}, 045021\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Malins \emph{et~al.}(2013)Malins, Williams, Eggers, and
Royall]{Malins2013a}
A.~Malins, S.~R. Williams, J.~Eggers and C.~P. Royall, \emph{The Journal of
Chemical Physics}, 2013, \textbf{139}, 234506\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Lazar(2017)]{lazar2017}
E.~A. Lazar, \emph{Modelling and Simulation in Materials Science and
Engineering}, 2017, \textbf{26}, 015011\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Fern{\'a}ndez and Harrowell(2003)]{Fernandez:2003he}
J.~Fern{\'a}ndez and P.~Harrowell, \emph{Phys. Rev. E}, 2003, \textbf{67},
011403\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Coslovich and Pastore(2007)]{coslovich2007}
D.~Coslovich and G.~Pastore, \emph{The Journal of chemical physics}, 2007,
\textbf{127}, 124504\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Malins \emph{et~al.}(2013)Malins, Eggers, Tanaka, and
Royall]{malins2013faraday}
A.~Malins, J.~Eggers, H.~Tanaka and C.~P. Royall, \emph{Faraday discussions},
2013, \textbf{167}, 405--423\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Plimpton(1995)]{plimpton1995fast}
S.~Plimpton, \emph{Journal of computational physics}, 1995, \textbf{117},
1--19\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Wierschem and Manousakis(2011)]{Wierschem2011}
K.~Wierschem and E.~Manousakis, \emph{Physical Review B}, 2011, \textbf{83},
214108\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Patashinski \emph{et~al.}(2010)Patashinski, Orlik, Mitus, Grzybowski,
and Ratner]{patashinski2010}
A.~Z. Patashinski, R.~Orlik, A.~C. Mitus, B.~A. Grzybowski and M.~A. Ratner,
\emph{The Journal of Physical Chemistry C}, 2010, \textbf{114},
20749--20755\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Stukowski(2009)]{stukowski2009}
A.~Stukowski, \emph{Modelling and Simulation in Materials Science and
Engineering}, 2009, \textbf{18}, 015012\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Macmillan and Kelly(1972)]{macmillan1972}
N.~Macmillan and A.~Kelly, \emph{Proc. R. Soc. Lond. A}, 1972, \textbf{330},
291--308\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Travesset(2014)]{travesset2014phase}
A.~Travesset, \emph{The Journal of chemical physics}, 2014, \textbf{141},
164501\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Rein~ten Wolde \emph{et~al.}(1996)Rein~ten Wolde, Ruiz-Montero, and
Frenkel]{rein1996numerical}
P.~Rein~ten Wolde, M.~J. Ruiz-Montero and D.~Frenkel, \emph{The Journal of
chemical physics}, 1996, \textbf{104}, 9932--9947\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Pedersen \emph{et~al.}(2018)Pedersen, Schr{\o}der, and
Dyre]{pedersen2018}
U.~R. Pedersen, T.~B. Schr{\o}der and J.~C. Dyre, \emph{Physical review
letters}, 2018, \textbf{120}, 165501\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Horn and L{\"o}wen(2014)]{horn2014does}
T.~Horn and H.~L{\"o}wen, \emph{The Journal of chemical physics}, 2014,
\textbf{141}, 224505\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Fern{\'a}ndez and Harrowell(2003)]{fernandez2003}
J.~R. Fern{\'a}ndez and P.~Harrowell, \emph{Physical Review E}, 2003,
\textbf{67}, 011403\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Turci \emph{et~al.}(2017)Turci, Royall, and Speck]{turci2017}
F.~Turci, C.~P. Royall and T.~Speck, \emph{Physical Review X}, 2017,
\textbf{7}, 031028\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Hohenberg and Halperin(1977)]{hohenberg1977theory}
P.~C. Hohenberg and B.~I. Halperin, \emph{Reviews of Modern Physics}, 1977,
\textbf{49}, 435\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Belonoshko \emph{et~al.}(2007)Belonoshko, Davis, Skorodumova, Lundow,
Rosengren, and Johansson]{belonoshko2007}
A.~Belonoshko, S.~Davis, N.~V. Skorodumova, P.-H. Lundow, A.~Rosengren and
B.~Johansson, \emph{Physical Review B}, 2007, \textbf{76}, 064121\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Espinosa \emph{et~al.}(2016)Espinosa, Vega, Valeriani, and
Sanz]{espinosa2016seeding}
J.~R. Espinosa, C.~Vega, C.~Valeriani and E.~Sanz, \emph{The Journal of
Chemical Physics}, 2016, \textbf{144}, 034501\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Jaiswal \emph{et~al.}(2016)Jaiswal, Procaccia, Rainone, and
Singh]{jaiswal2016mechanical}
P.~K. Jaiswal, I.~Procaccia, C.~Rainone and M.~Singh, \emph{Physical review
letters}, 2016, \textbf{116}, 085501\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Parisi \emph{et~al.}(2017)Parisi, Procaccia, Rainone, and
Singh]{parisi2017shear}
G.~Parisi, I.~Procaccia, C.~Rainone and M.~Singh, \emph{Proceedings of the
National Academy of Sciences}, 2017, \textbf{114}, 5577--5582\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Ozawa \emph{et~al.}(2018)Ozawa, Berthier, Biroli, Rosso, and
Tarjus]{ozawa2018random}
M.~Ozawa, L.~Berthier, G.~Biroli, A.~Rosso and G.~Tarjus, \emph{Proceedings of
the National Academy of Sciences}, 2018, 201806156\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\end{mcitethebibliography}
\end{document}
|
\section{Conclusions and future work}
In this demo, we present an interactive prototype system, \textsc{Extra}, as the first step towards explaining team recommendation through the lens of the underlying network where teams embed. \hide{The key algorithmic idea is to identify the most influential network substructures and/or attributes that account for the team recommendation results.} The system can provide intuitive visual explanations from different perspectives for various team recommendation scenarios.
Future work includes applying the proposed algorithm to other networked recommendation scenarios, e.g., user-product networks to explain why a certain product is recommended to a given user.
\section{Introduction}
The emergence of network science has been significantly changing the landscape of team-based research in recent years.
A cornerstone behind team recommendation algorithms for various scenarios (e.g., {\em team member replacement}, {\em team expansion} and {\em team shrinkage}) is random walk graph kernel~\cite{borgwardt2007fast}. For instance, in team member replacement, by applying random walk graph kernel to the team networks before and after replacement, it encodes both the skill match, structure match as well as the interaction between the two during the replacement process~\cite{li2015replacing}. Team member replacement further enables other team recommendation scenarios~\cite{li2017enhancing}. \hide{Although being effective in answering questions like {\em who} is the best replacement, {\em what} is the best team expansion strategy,} However, these existing methods lack intuitive ways to explain {\em why} the underlying algorithm gives the specific recommendation for a given team optimization scenario.
\noindent\textbf{Purpose and Novelty:} In this paper, we present an interactive prototype system,~\textsc{Extra}, a paradigm shift from {\em what} to {\em why}, for the purpose of explaining the networked team recommendation results. The novelties of \textsc{Extra}{} are (1) on the algorithmic side, we propose an effective and efficient algorithm to explain random walk graph kernel, where the key idea is to identify the most influential network substructures and/or attributes whose removal or perturbation will impact the graph kernel/similarity the most, and (2) on the system side, \textsc{Extra}{} is able to provide informative visual analysis for explaining various team recommendation scenarios from a variety of perspectives (e.g., edges, nodes and attributes). For example, given a candidate for team member replacement, we are able to tell (a) what the key connections are between the candidate and the existing team members, and (b) what the key skills the candidate has that might make him/her a potentially good replacement.
\hide{
\noindent\textbf{Engaging the Audience:} We expect that our demo will primarily attract two kinds of audiences: (1) practitioners who are interested in the applications of network science of teams and more broadly network-based recommender systems, and (2) researchers who are interested in developing new algorithms and tools for explainable recommender systems, information retrieval, and data mining.
}
\section{0pt}{12pt plus 4pt minus 2pt}{0pt plus 2pt minus 2pt}
\newcounter{para}
\newcommand\mypara{\par\refstepcounter{para}\textbf{\thepara}\space}
\newcommand{\hide}[1]{}
\newcolumntype{C}[1]{>{\centering\let\newline\\\arraybackslash\hspace{0pt}}m{#1}}
\newcommand\newdagger{\mathrel{\ooalign{$\dagger$\cr\hidewidth$\ddagger$\hidewidth\cr}}}
\newcommand{\textsc{Extra}}{\textsc{Extra}}
\newcommand{\hh}[1]{{\small\color{red}{\bf hh: #1}}}
\newcommand{\lli}[1]{{\small\color{blue}{\bf LL: #1}}}
\newcommand{\qh}[1]{{\small\color{brown}{\bf QH: #1}}}
\newcommand{\mkclean}{
\renewcommand{\hh}[1]{}
\renewcommand{\lli}[1]{}
\renewcommand{\qh}[1]{}
}
\begin{document}
\setlength{\textfloatsep}{2pt}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\parindent}{1em}
\title{\textsc{Extra}: Explaining Team Recommendation in Networks}
\author{Qinghai Zhou$^\dagger$,~~~Liangyue Li$^\dagger$,~~~Nan Cao$^\ddagger$,~~~Norbou Buchler$^\newdagger{}$ and Hanghang Tong$^\dagger$}
\affiliation{\vspace{-4mm}
\institution{$^\dagger$Arizona State University, $^\ddagger$Tongji University, $^\newdagger{}$ARL\\
\vspace{-4mm}
$^\dagger$\{qinghai.zhou, liangyue, hanghang.tong\}@asu.edu, $^\ddagger$nan.<EMAIL>, $^\newdagger{}$norbou.<EMAIL>}}
\hide{\author{Qinghai Zhou$^\dagger$,~~~Liangyue Li$^\dagger$,~~~Nan Cao$^\ddagger$,~~~Norbou Buchler$^\newdagger{}$ and Hanghang Tong$^\dagger$}
\affiliation{
\institution{$^\dagger$Arizona State University, $^\ddagger$Tongji University, $^\newdagger{}$ARL\\\vspace{-4mm}
$^\dagger$\{qinghai.zhou, liangyue, hanghang.tong\}@asu.edu, $^\ddagger$nan.<EMAIL>, $^\newdagger{}$norbou.<EMAIL>}}}
%
\renewcommand{\shortauthors}{Q. Zhou et al.}
\begin{abstract}
\input{abstract}
\end{abstract}
\hide{\begin{CCSXML}
<ccs2012>
<concept>
<concept_id>10010520.10010553.10010562</concept_id>
<concept_desc>Computer systems organization~Embedded systems</concept_desc>
<concept_significance>500</concept_significance>
</concept>
<concept>
<concept_id>10010520.10010575.10010755</concept_id>
<concept_desc>Computer systems organization~Redundancy</concept_desc>
<concept_significance>300</concept_significance>
</concept>
<concept>
<concept_id>10010520.10010553.10010554</concept_id>
<concept_desc>Computer systems organization~Robotics</concept_desc>
<concept_significance>100</concept_significance>
</concept>
<concept>
<concept_id>10003033.10003083.10003095</concept_id>
<concept_desc>Networks~Network reliability</concept_desc>
<concept_significance>100</concept_significance>
</concept>
</ccs2012>
\end{CCSXML}
\ccsdesc[500]{Computer systems organization~Embedded systems}
\ccsdesc[300]{Computer systems organization~Redundancy}
\ccsdesc{Computer systems organization~Robotics}
\ccsdesc[100]{Networks~Network reliability}
}
\vspace{-6mm}
\keywords{Team Recommendation Explanation; Visualization; Random Walk Graph Kernel}
\copyrightyear{2018}
\acmYear{2018}
\acmConference[RecSys '18]{Twelfth ACM Conference on Recommender Systems}{October 2--7, 2018}{Vancouver, BC, Canada}
\acmBooktitle{Twelfth ACM Conference on Recommender Systems (RecSys '18), October 2--7, 2018, Vancouver, BC, Canada}
\acmDOI{10.1145/3240323.3241610}
\acmISBN{978-1-4503-5901-6/18/10}
\maketitle
\input{introduction}
\vspace{-2mm}
\input{system_overview.tex}
\input{system_demonstration.tex}
\input{conclusion}
\vspace{-2mm}
\begin{acks}
This work is supported by NSF (IIS-1651203, IIS-1715385, CNS-1629888 and IIS-1743040), DTRA (HDTRA1-16-0017), ARO (W911NF-16-1-0168), DHS (2017-ST-061-QA0001) and NSFC (61602306).
\end{acks}
\vspace{-3mm}
\bibliographystyle{ACM-Reference-Format}
\section{System Demonstration}
\begin{figure}[ht!]\label{fig:example}
\centering
\includegraphics[width=0.48\textwidth]{figures/system_ui.pdf}
\vspace{-5mm}
\caption{An illustrative example of influence analysis.}\label{experiment}
\end{figure}
\hide{\lli{can we mark the nodes w/ their last names in the figure instead of A,B,C..}}
Figure~\ref{experiment} presents the user interface of \textsc{Extra}\ and an example of visualization of influence analysis for team member replacement on a co-authorship network\footnote{A demo video of the system is available at \href{https://youtu.be/D4gcI-QHtps}{https://youtu.be/D4gcI-QHtps}. \\ System website: \href{http://144.202.123.224/system.html}{http://144.202.123.224/system.html }.}, which suggests that Dr. \emph{Jiawei Han} is the best replacement. The influence scores of all edges for both graphs are calculated by our proposed algorithm in Section~\ref{algorithms} and the width of edge is proportional to the influence score. The top-{\em 4} most influential edges are those that connect Dr. {\em Yu} with the 4 existing members, \emph{Lin, Zhao, Chen and Sun}, all of whom are considered as the key members and they also have strong collaborations with Dr. {\em Han}. After replacement, the top-{\em 5} most influential edges with Dr. {\em Han} overlap with those of Dr. {\em Yu} in the same ranking order. The only exception is that no edge exists between Dr. {\em Yu} and Dr. {\em Zhu} because there is no prior collaboration between them. In addition, the system can provide explanations from the attribute perspective. The key skills (represented in pie chart) shared by Dr. {\em Han} and Dr. {\em Yu} are {\it databases} and {\it data mining} in this case, which makes the team replacement recommendation more understandable.
\hide{
Figure~\ref{performance} compares the runtime of the proposed method (\textsc{Extra}) \hh{yep, qinghai: change the label in figure 2 to Extra}\lli{I thought our method is called "extra"...} with `Direct Computation'. We can see that the proposed method is much faster, with \hide{\hh{fill in}}$17.7\times$ speedup on average.
\begin{figure}\label{fig:time}
\centering
\captionsetup{justification=centering,margin=1.2cm}
\includegraphics[width=0.48\textwidth]{figure2.pdf}
\vspace{-8mm}
\caption{Algorithm Performance Comparison.}\label{performance}
\end{figure}
}
\section{Overview of the System} \label{section2}
\begin{table*}[ht]\label{tab:summary}
\centering
\small
\captionsetup{justification=centering}
\begin{tabular}{|C{1.5cm}|C{4.9cm}|C{5.3cm}|C{4.7cm}|}
\hline
& {\bf Team Replacement} & {\bf Team Expansion} & {\bf Team Shrinkage} \\
\hline
{\bf Edges} & important common collaborations shared by the candidate and the departure member & new collaborations that the new member might establish & the most important but lacking collaborations the candidate should have\\
\hline
{\bf Nodes} & key existing team members both candidate and departure member collaborate with & key existing team members the new member will work with & key existing team members that the candidate should have collaborated with\hide{\qh{consider overall collaboration is not good}}\\
\hline
{\bf Attributes} & common and important skills shared by the candidate and the departure member & the unique skills the new team member brings that are critical to the team's new need & the most important skills that the candidate lacks \\
\hline
\end{tabular}
\centering
\vspace{1mm}
\caption{Summary of system functionalities. }\label{table}
\vspace{-6mm}
\end{table*}
In this section, we present (1) the main functionalities of \textsc{Extra}\ to explain three different team recommendation scenarios
and (2) key algorithms that support the main functionalities.
\vspace{-3mm}
\subsection{System Functionality}\label{functionality_section}
The system provides intuitive visual explanations for the team recommendation results through the lens of the underlying network. Table~\ref{table} summarizes the main functionalities, where the rows represent explanations from three different perspectives (e.g., edges, nodes and attributes) for three team recommendation scenarios.
\noindent\textbf{A -- Explaining Team Replacement:} \textsc{Extra}\ identifies a few key (1) edges (the relationship between the candidate and other team members), (2) nodes (other existing team members) and (3) attributes (the skills of the candidate) that make the candidate and the departure member most similar. In this way, it could help the end-user (e.g., the team leader) understand why the underlying replacement algorithm thinks the given candidate is potentially a good replacement. \\
\textbf{B -- Explaining Team Expansion:} it provides the explanations from the following aspects, including (1) what are the unique new skills s/he brings to the team (i.e., attribute); and (2) what are the key collaborations the new team member might establish (i.e., edges) and with whom (i.e., nodes). \\
\textbf{C -- Explaining Team Shrinkage:} the prototype system flags the {\em absent} skills and connections with existing team members that makes the candidate most insignificant to the current team. In other words, we want to understand why the dismissal of this particular candidate would impose the least negative impact on the team.
\vspace{-3mm}
\subsection{Overview of Algorithm}\label{algorithms}
\noindent\textbf{A -- Graph Kernel for Team Recommendation: } random walk graph kernel provides a natural way to measure the similarity between two graphs
~\cite{borgwardt2005protein} (e.g., the two team networks before and after a replacement) and turns out to be the core building block behind a variety of team recommendation scenarios. For example, in \emph{team replacement}, the objective is to find a \emph{similar} person $m$ to replace an existing team member $r$ in the context of the team itself. A good replacement $m$ should have a similar skill set as well as a similar collaboration structure with the existing team member. With graph kernel, we aim to find a candidate that makes the new team most similar to the old team.
For details regarding other scenarios (i.e., team expansion and shrinkage), please refer to \cite{borgwardt2007fast,li2017enhancing}.
\noindent\textbf{B -- Explaining Team Recommendation:} we seek to understand the team recommendation results via the influence of various graph elements to the corresponding graph kernels. We define the influence score of one specific element as the {\em rate} of the change in graph kernel $k(\mathbf{G}, \mathbf{G'})$. Given random walk graph kernel for labelled graph~\cite{li2015replacing}, the influence of a graph edge $W_{ij}$ can be computed as
\begin{equation}\label{influence}
\mathcal{I}\left( W_{ij} \right) = \frac{\partial k\left( \mathbf{G},\mathbf{G'} \right)}{\partial W_{ij}} = c\mathbf{q_{\times}}^{T}\mathbf{R}\mathbf{L_{\times}}\left( (\mathbf{J}^{i,j} + \mathbf{J}^{j,i})\otimes\mathbf{W'}\right)\mathbf{R}\mathbf{L_{\times}}\mathbf{p} \nonumber
\end{equation}
\noindent where $\mathbf{R} = \left( \mathbf{I} - c~\mathbf{W_{\times}} \right)^{-1}$ \hide{, whose computation can be accelerated using power method,} and $\mathbf{J}^{i,j}$ is a single-entry matrix
with one at its $(i,j)$-th entry and zeros everywhere else.
In the scenario of team member replacement, an ideal candidate should have the key collaborations (with high influence score) as the departure member does, indicating their similarity in terms of how they collaborate with existing members.
The node influence of the $i^{\textrm{th}}$ member is defined as the aggregation of the influence of all the edges incident to this node, i.e., $\mathcal{I}\left(i\right) = \sum_{j | (i,j)\in \mathbf{E}}\mathcal{I}\left(W_{ij}\right)$.
Existing team members with the highest node influence scores are expected to be key members in the team.
\hide{Therefore a good candidate should also collaborate with these key members as the departure member does after replacement.}
\hide{In team expansion, these are the key team members the new member will work with. In team shrinkage, these are the key members that the candidate should have collaborated with.}
Likewise, to compute the influence of a team member's attributes on the graph kernel, we take the derivative of the graph kernel w.r.t. the member's skill. Denote the $k^{\textrm{th}}$ skill of the $i^{\textrm{th}}$ team member in graph $\mathbf{G}$ as $L_{k}\left( i \right)$, the attribute influence can be calculated as
\begin{equation}\label{attribute}
\mathcal{I}\left(L_{k}\left( i \right)\right) = \frac{\partial k\left( \mathbf{G},\mathbf{G'} \right)}{\partial L_{k}\left( i \right)} = \mathbf{q_{x}}^{T}\mathbf{R}\left( \frac{\partial \mathbf{L_{\times}}}{\partial L_{k}\left( i \right)}\right)\left( \mathbf{I} + c(\mathbf{W}\otimes \mathbf{W}')\mathbf{R}\mathbf{L_{\times}} \right)\mathbf{p_{\times}} \nonumber
\end{equation}
\noindent where $\frac{\partial \mathbf{L_{\times}}}{\partial L_{k}\left( i \right)} = \textrm{diag}\left( \mathbf{e}_i \right) \otimes \textrm{diag}\left(\mathbf{L'}\left(:, k\right)\right)$ and $\mathbf{e}_i$ is an $n \times 1$ vector with one at the $i^{\textrm{th}}$ entry and zeros everywhere else.
Skills with the highest influence scores are the key skills that (1) the candidate and the departure member have in common in team replacement, (2) the new team member might bring in team expansion, and (3) the dismissal member might lack in team shrinkage.
|
\section{Introduction}
Classically intractable and unsolvable problems are the driving force behind
the ever increasing search for new computational models. These new models can
be used to provide faster solutions to classically intractable
problems~\cite{dna} and, in extreme cases, they can be {\em used} to
{\em solve} classically unsolvable problems~\cite{quantum}. A major drawback
of such new models is that they seem to be useful only for the solution of a
particular class of problems. For example, DNA computing has proved
efficient for the solution of combinatorial problems. Thus, a new computational
model is really useful once we have the techniques that allow us to compute
{\em ordinary} things.
P systems are a new model of computation inspired by the way cells live and
function (see~\cite{membrane} for an overview of the theory). Although, in
general, it has been proved that one can compute with P~systems whatever
(universal) Turing machines are capable to compute; until recently it was not clear
how one can compute primitive recursive functions with P~systems~\cite{pcomp2}.
In this note we provide an alternative encoding of the basic functions as well as
three processes, which can be applied to the basic functions to yield
any $\mu$-recursive function. In order to encode these functions and processes
we had to introduce new rewriting rules and to make use of a more ``liberal''
membrane structures. Thus, our encoding has stretched to the limit the
capabilities of P~systems as far it regards numerical computation.
In what follows, we present a very brief overview of the theory of P~systems.
Next, we present our encoding and we conclude with some remarks concerning
our work and possible applications of it.
\section{Brief Overview of P Systems}
In what follows we will present an informal definition P~systems.
\begin{definition}
A P~system is an n-tuple
\begin{displaymath}
\Pi=(O, \mu, w_1,\ldots, w_n, R_1,\ldots, R_m, i_0)
\end{displaymath}
where:
\begin{enumerate}
\item $O$ is an alphabet (i.e., a set of distinct entities) whose elements
are called \textit{objects}.
\item $\mu$ is the membrane structure of the particular P~system; membranes
are injectivelly labeled with succeeding natural numbers starting with
one.
\item $w_i$, $1\le i\le m$, are strings that represent multisets over $O$
associated with each region $i$.
\item $R_i$, $1\le i\le m$, are finite sets of rewriting rules (called
\textit{evolution rules}) over $O$. An evolution rule is of the form
$u\rightarrow v$, $u\in O^{+}$ and $v\in O^{+}_{\mathrm{tar}}$,
where $O_{\mathrm{tar}} = O\times\mathrm{TAR}$,
$\mathrm{TAR}=\{\mathrm{here},\mathrm{out}\}\cup\{\mathrm{in}_{j} |
1\le j\le m\}$.
\item $i_0\in\{1,2,\ldots,m\}$ is the label of an elementary membrane (i.e.,
a membrane that does not contain any other membrane), called the
\textit{target} compartment.
\end{enumerate}
\end{definition}
Note that the keywords ``here,'' ``out,'' and ``$\mathrm{in}_j$'' are called
\textit{target} commands. Given a rule $u\rightarrow v$, the length of $u$
(denoted $|u|$) is called the radius of the rule. A P~system that contains
rules of radius greater than one is a system with cooperation.
The rules that belong to some $R_i$ are applied to the objects of the
compartment $i$ synchronously, in a non-deterministic maximally parallel way.
Given a rule $u\rightarrow v$, the effect of applying this rule to a
compartment $i$ is to remove remove the multiset of objects specified
by $u$ and to insert the objects specified by $v$ in the regions designated
by the target commands associated with the objects from $v$. In particular,
\begin{enumerate}
\item if $(a,\mathrm{here})\in v$, the object $a$ will be placed in
compartment $i$ (i.e., the compartment where the action takes
place)
\item if $(a,\mathrm{out})\in v$, the object $a$ will be placed in
the compartment that surrounds $i$
\item if $(a,\mathrm{in}_{j})\in v$, the object $a$ will be placed in
compartment $j$, provided that $j$ is immediately inside $i$, or else
the rule is not applicable
\end{enumerate}
\section{Encoding the Basic Functions as P Systems}
Our exposition on recursion theory is based on standard references~\cite{boolos,davis}.
Here are the three basic functions:
\begin{enumerate}
\item the {\em successor} function $S(x)=x+1$,
\item the {\em zero} function $z(x_1,\ldots,x_n)=0$, and
\item the {\em identity} function $U^{n}_{i}(x_1,\ldots,x_n)=x_i$,
$1\le i\le n$.
\end{enumerate}
If we want to define $A$-recursive functions, we also need to use the characteristic
function. However, $A$-recursive sets are not important for our discussion, so we
are going to restrict ourselves to general recursive functions only.
Two nested compartments (i.e., one inside another) make up the simplest
non-trivial membrane structure. For this reason, we have opted to encode
the basic functions as P~systems that have this particular
membrane structure. In what follows, the inner compartment will be the
target compartment. Let us now elaborate on the encoding of each function.
\paragraph{The zero function} This function simply discards its arguments
and returns the number zero. For each argument $x_i$ of the function we
pick up an object $\alpha_i$ and place $x_i$ copies of it in the outer
compartment. Next, we associate to each object $\alpha_i$ a multiset
rewrite rule of the form
\begin{displaymath}
\alpha_i\rightarrow\varepsilon
\end{displaymath}
This is a new kind of rule that simply annihilates all objects as it replaces
each object with the empty string. We can imagine that there is a pipe that
is used to throw the $\alpha_i$'s into the environment. This rule
can be considered to implement a form of ``catharsis'' of a compartment. Thus,
the outcome of the computation of this P~system is the number zero as
all objects are practically thrown to the environment. Obviously, this system
implements the zero function.
\paragraph{The successor function} Suppose we have a P~system with no data
that, at the same time, encodes the successor function, then this system must
place an object to the target compartment and stop. Obviously, the problem is
how this can happen, as there is no rule that can be applied to an empty
compartment. For this reason, we introduce the new {\empty} rule
\begin{displaymath}
\varepsilon\rightarrow(\alpha, \mathrm{in}_{i})
\end{displaymath}
This rule can be applied only if there are no objects in a given compartment.
And its effect is to place an object $\alpha$ to compartment $i$. After this,
the rule is ``disassociated'' from the current compartment. In other words,
this rule can be used only once. The rationale behind the introduction of
this rule is that at any moment a cell can absorb matter from its
environment that, in turn, will be consumed. Having this new rule, it is
now trivial to encode the successor function.
\paragraph{The identity function} This function is similar to the zero
function with one difference: it discards all of its arguments but one.
So, for each argument $x_i$ we place $x_i$ copies of an object $\alpha_i$.
Note that all objects $\alpha_i$ are distinct. Assume that the particular
instance of the identity function returns its $j$th argument, then we
associate with $\alpha_j$ the following rule:
\begin{displaymath}
\alpha_{j}\rightarrow(\alpha_{j},\mathrm{in}_{2})
\end{displaymath}
Similarly to the zero function case, all other objects are associated
with a catharsis rule:
\begin{displaymath}
\alpha_{i}\rightarrow\varepsilon,\; i\not=j
\end{displaymath}
\section{Encoding the Processes as P Systems}
Recursive functions can be defined by applying function builders,
or processes, to the basic functions. For example, the addition function
is build by applying primitive recursion to the identity and successor
functions. Thus, it is necessary to provide an encoding of the three
processes in order to be able to compute any recursive function using
P~systems. The three processes are described below:
\begin{description}
\item[Composition] Suppose that $f$ is a function of $m$ arguments and
each of $g_1,\ldots,g_m$ is a function of $n$ arguments, then the function
obtained by composition from $f, g1, \ldots, g_m$ is the function $h$
defined as follows:
\begin{displaymath}
h(x_1,\ldots,x_n)=f(g_1(x_1,\ldots,x_n),\ldots,g_m(x_1,\ldots,x_n))
\end{displaymath}
\item[Primitive Recursion] A function $h$ of $k+1$ arguments is said to be
definable by (primitive) recursion from the functions $f$ and $g$, having
$k$ and $k+2$ arguments, respectively, if it is defined as follows:
\begin{eqnarray*}
h(x_1,\ldots,x_k,0)&=&f(x_1,\ldots,x_k)\\
h(x_1,\ldots,x_k,S(m))&=&g(x_1,\ldots,x_k,m,h(x_1,\ldots,x_k,m))
\end{eqnarray*}
\item[Minimalization] The operation of minimalization associates with each
total function $f$ of $k+1$ arguments the function $h$ of $k$ arguments.
Given a tuple $(x_1,\ldots,x_k)$, the value of $h(x_1,\ldots,x_k)$
is the least value of $x_{k+1}$, if one such exists,
for which $f(x_1,\ldots,x_k,x_{k+1})=0$.
If no such $x_{k+1}$ exists, then its value is undefined.
\end{description}
In the rest of this section we will not provide encodings for the general
case. Instead, we will concentrate on simple cases that can be easily
generalized. To put it in another way, if the P~systems that we present
faithfully encode the basic processes, then one can easily generalize
the encoding scheme.
\paragraph{Encoding the composition process}
Assume that $f$ is function that has two arguments
and that $g_1$ and $g_2$ are two functions with three arguments each, then
we want to design a P~system that will encode a function $h$ defined as
follows
\begin{equation}\label{eq:1}
h(x_1,x_2,x_3)=f(g_1(x_1,x_2,x_3),g_2(x_1,x_2,x_3))
\end{equation}
\begin{figure}
\begin{center}
\begin{picture}(200,200)
\put(0,0){\framebox(180,180)[cc]{}}
\put(5,170){$\alpha^{x_1}$}
\put(20,170){$\beta^{x_2}$}
\put(35,170){$\gamma^{x_3}$}
\put(50,130){\oval(70,40)}
\put(65,130){\oval(20,10)}
\put(50,155){$g_1$}
\put(130,130){\oval(70,40)}
\put(145,130){\oval(20,10)}
\put(130,155){$g_2$}
\put(90,50){\oval(70,40)}
\put(90,75){$f$}
\put(65,130){\line(1,-4){17}}
\put(67,130){\line(1,-4){17}}
\put(145,130){\line(-1,-2){34}}
\put(147,130){\line(-1,-2){34}}
\end{picture}
\end{center}
\caption{A P system encoding the composition process.}
\end{figure}
Given three integers $x_1$, $x_2$ and $x_3$, we want to design a P~system
capable to compute $h(x_1,x_2,x_3)$. From equation~\ref{eq:1} it is clear
that we have first to compute $u_1=g_1(x_1,x_2,x_3)$ and
$u_2=g_2(x_1,x_2,x_3)$. This suggests that the P~system will contain two
nested P~systems that will compute the numbers $u_1$ and $u_2$, respectively.
In addition, these numbers (i.e., a multiset of objects representing them)
have to be directly supplied to a third nested P~system, which will compute
the value $u=f(u_1,u_2)$. This means that we need two pipelines that will
connect the output compartment of the first two nested P~systems with the
P~system that will compute the number $u$. Obviously, P~systems with
tree-like structure are not adequate to solve this problem. This means
that we need a way to more generally specify the communication channels
between the various compartments. An obvious solution, is to use P~systems
that have a (limited) graph-like structure. Indeed, in~\cite{pi-dist} the first
author of this paper provides such an extension.
\begin{definition}
A graph-structured symbol-object membrane system is a set of
(possibly nested) compartments that can be specified as follows:
\begin{displaymath}
\Pi=(O,g_m,w_1,\ldots,w_m,R_1,\ldots,R_m,i_0)
\end{displaymath}
Here:
\begin{enumerate}
\item $O$ is a set of objects that denote elements of information;
\item $g_m$ is a relation in $\{1,\ldots,m\}\times\{1,\ldots,m\}$,
describing the network structure of the system;
\item $w_i$, $1\le i\le m$, are strings representing multisets of objects
that are elements of $O$ that denote conglomerations of information;
\item $R_i$, $1\le i\le m$, are finite sets of \textbf{evolution rules} over
$O$; $R_i$ is associated with compartment $i$; an evolution rule is
of the form $u\rightarrow v$, where $u$ is a string over $O$ and $v$
is a string over $O_{\mathrm{NTAR}}$, where $O_{\mathrm{NTAR}}=O\times
\mathrm{NTAR}$, for $\mathrm{NTAR}=\{\mathrm{here}, \mathrm{out}\}
\cup\{\mathrm{to}_{j}|1\le j\le m\}$;
\item $i_0 \in \{1,2,\ldots,m\}$ is the label of the \textbf{output
compartment}.
\end{enumerate}
\end{definition}
The symbol ``$\mathrm{to}_j$'' should be used to directly place an object from
the host compartment to compartment $j$. The rule is applicable only if
$(i,j)\in g_m$ or $(j,i)\in g_m$.
\paragraph{Encoding primitive recursion}
In primitive recursion we a priori know the number of iterations. Thus, if we want to
compute $h(x,y+1)$ we actually have to compute the following:
\begin{displaymath}
g(x,y,g(x,y-1,g(x,y-2,\ldots g(x,0,f(x))\ldots)))
\end{displaymath}
Thus, if we want to compute $h(x,y)$, we need $n$ compartments that will
compute the successive values of $g$ and one compartment that will compute the value of
$f(x)$. In addition, all these compartments must form a pipeline where the $i$th
compartment will get its data from compartment $i-1$ and it will feed its output to
the compartment $i+1$. Naturally, the last compartment will populate with data the
output compartment. Initially, all compartments will get the same copies of a
designating object, say \#, which will denote the first argument of function $h$. Also,
the compartments that compute the successive values of $g$ will get the respective number
of copies of another designated object, say @. Figure~\ref{fig:2} depicts the initial
setting of the system.
\begin{figure}
\begin{center}
\begin{picture}(300,200)
\put(0,0){\framebox(270,190)[cc]{}}
\put(5,180){$\alpha^{x}$}
\put(20,180){$\beta^{y}$}
\put(20,100){\framebox(50,30){}}
\put(40,110){\framebox(20,10){}}
\put(30,140){$f(x,y)$}
\put(58,117){\line(1,0){40}}
\put(58,114){\line(1,0){40}}
\put(90,100){\framebox(50,30){}}
\put(110,110){\framebox(20,10){}}
\put(90,140){$g(x,0,f(x))$}
\put(128,117){\line(1,0){40}}
\put(128,114){\line(1,0){40}}
\put(172,115){\ldots}
\put(190,117){\line(1,0){20}}
\put(190,114){\line(1,0){20}}
\put(200,100){\framebox(50,30){}}
\put(220,110){\framebox(20,10){}}
\put(210,140){$g(\ldots)$}
\put(230,115){\line(0,-1){60}}
\put(233,115){\line(0,-1){60}}
\put(200,30){\framebox(50,30){}}
\end{picture}
\end{center}
\caption{A P system encoding primitive recursion.}\label{fig:2}
\end{figure}
\paragraph{\textbf{Encoding the minimalization function builder}}
As for the previous cases, we present a simplified system that computes function
$h(x)$. Assume that $f$ is a function of two arguments $x$ and
to compute $f(x,y)$ with a fixed $x$, which is equal to the argument of the original
function $h$, and a variable $y$, which is increased by one in each iteration, that halts
the computation the very moment the outcome of the function computation is zero.
Figure~\ref{fig:3} shows the general setup of a P~system that computes function $h(x)$.
Compartment~3 is directly connected to the output compartment of compartment~2, where
actual the computation is taking place. The output of this compartment is pipelined to
compartment~3 and it is transformed to the distinguished object \#. This symbol is
used as counter. The connection from the compartment
number~3 to compartment number~4 is used to move the object \# to it. In addition, at
each iteration, the $a$'s and $b$'s are also placed in compartment~3 in the form of
of two distinguished elements $\oplus$ and $\otimes$. More specifically, we move the
$a$'s and the $b$'s in compartment~1 and at the same time we create a number of
$\oplus$'s and $\otimes$'s that is to number of $a$'s and $b$'s respectively. Clearly,
if the result of the computation at compartment~2 is the number zero, no action should
take place at compartment~3. Otherwise, we first move \# to compartment~4 and place
a copy of $b$ at the compartment~1, then we
place the $\oplus$'s and $\otimes$ to compartment~1 in the form of $a$'s and $b$'s.
This way, we actually increase the $b$'s by one, which is what we actually need to
do to get the correct result.
\begin{figure}
\begin{center}
\begin{picture}(200,200)
\put(0,0){\framebox(200,170)[cc]{}}
\put(5,160){$\alpha^{x}$}
\put(20,160){$\beta^{y}$}
\put(5,5){$1$}
\put(30,100){\framebox(50,30){}}
\put(50,110){\framebox(20,10){}}
\put(40,140){$f(x,y)$}
\put(66,117){\line(1,0){40}}
\put(66,114){\line(1,0){40}}
\put(33,103){$2$}
\put(100,100){\framebox(50,30){}}
\put(103,103){$3$}
\put(130,110){\line(0,-1){60}}
\put(133,110){\line(0,-1){60}}
\put(100,30){\framebox(50,30){}}
\put(103,33){$4$}
\end{picture}
\end{center}
\caption{A P system encoding minimalization function.}\label{fig:3}
\end{figure}
It is important to note that in all cases our description is a little bit vague,
as we do not give the full details of the encoding. However, our intension was
to give the general idea of how things can be done. After all, we believe it is not
difficult to work out these details so to provide a complete description of the
encoding of each function builder.
\section{Conclusions}
We have presented an alternative encoding of the basic functions and the three
function builders of recursion theory. Since any functional programming language
can be used to compute any recursive function, one can build a front-end capable
of compiling (simple) functional programs into P~systems. Indeed, in~\cite{pi-dist},
the first author of this paper presents ideas related to this particular implementation
procedure. We believe that more work is needed to be done in order to provide
a complete and viable solution to the problem of encoding recursive functions as
P~systems. We just hope that our work is step towards this direction.
|
\section{Proof of Collapsing Conjecture for Strings of Length~3}
\label{subsec:scs3}
In this section, we show that the Collapsing Conjecture holds for
the special case
when input strings have length at most three.
Remarkably, this follows from a~more general theorem stated below.
\begin{theorem}\label{thm:cc3}
Let $\mathcal{S}$ contain strings of length at most~3 and let $L$~be an Eulerian solution that for each $s \in \mathcal{S}$ contains at least two copies of arcs $(\pref(s), s)$ and $(s, \suff(s))$. Then $CA(\mathcal{S}, L)=GHA(\mathcal{S})$.
\end{theorem}
It is not difficult to see that the theorem indeed implies the Collapsing Conjecture: clearly, $L=D \sqcup D$, where $D$ is an Eulaerian solution, satisfies the condition. Moreover, this also works for $L=D_1 \sqcup D_2$, where $D_1, D_2$ are arbitrary Eulerian solutions, and for $L=D \sqcup CC$, where $CC$ is a~cycle cover, i.e., a~set of cycles that go through all input strings. The main difference between an Eulerian solution and a~cycle cover is that the later is not required to be connected. For this reason, any Eulerian solution is also a~cycle cover (but not vice versa) and hence an~{\em optimal} cycle cover is definitely not longer than an optimal Eulerian solution: $OPT \le OPTCC$. Hence, Theorem~\ref{thm:cc3} says that the result of GHA is not just no longer than $2\cdot OPT$, but even no longer than $OPT+OPTCC$.
Before proving Theorem~\ref{thm:cc3}, we introduce some notation and
prove two auxiliary results. Recall that the Collapsing Algorithm processes the nodes level by level. Denote by $L_i$ an intermediate Eulerian solution right before is starts collapsing the nodes at level~$i$ (that is, in $L_i$ all the nodes at levels $>i$ are already collapsed).
For an arbitrary Eulerian solution~$U$, by $\operatorname{above}(U,i)$ denote the part of~$U$ that lies above the level~$i$:
$\operatorname{above}(U,i)=\{(u, v) \in U \colon |u|, |v| \ge i\}$.
We show that $\operatorname{above}(D,i)=\operatorname{above}(L_i, i)$ for every~$i$. This is enough since then \[CA(\mathcal{S}, L)=\operatorname{above}(L_0,0)=\operatorname{above}(D,0)=GHA(\mathcal{S}) \, .\]
\begin{lemma}\label{lem:path}
Let $w$~be a~walk from~$u$ to~$v$ in an Eulerian solution with all its nodes at levels~$\le k$. Consider a~single collapsing step for a~node~$t$ that is either an~{\em intermediate} node of~$p$ at level~$k$ or is a~node at level $<k$ that do not belong to~$p$. Then $w$~is still a~walk from~$u$ to~$v$ in the resulting solution.
\end{lemma}
\begin{proof}
Indeed, if $t$~does not belong to~$w$, then collapsing it does not change~$w$ at all. Otherwise $t$~is an intermediate node of~$w$ at level~$k$. Since $w$~does not have any node above level~$k$, $w$~goes through
$(\pref(t), t), (t, \suff(t))$. Clearly, collapsing~$t$ keeps $w$~a~walk.
\end{proof}
\begin{lemma}\label{lemma:toepsilon}
Let $v$~be a~node in~$L_2$ at level~$1 \le l \le 2$ (i.e., $l=|v|$). Then there is a~walk from~$\varepsilon$ to~$v$ and a~walk from~$\varepsilon$ to~$v$ in~$L_l$ that does not contain nodes at level~3.
\end{lemma}
\begin{proof}
We start by proving that there is a~walk from~$v$ to~$\varepsilon$
for $|v|=2$ (the existence of a~walk from~$\varepsilon$ to~$v$ is proved in a~similar fashion).
Consider a~walk~$w$ from~$v$ to~$\varepsilon$ in~$L$ (there is such a~walk as $L$~is an~Eulerian solution). All repeated nodes in~$w$ may be removed,
so one may assume that $w$~passes through its nodes at level~3 exactly once. Then, it is sufficient to show that each such node is collapsed.
Consider a~node~$s$ of~$w$ at level~3 and a~pair of arcs $(\pref(s), s), (s, \suff(s)) \in w$ going through it. If $s$~is not at input string (i.e., $s \not \in \mathcal{S}$), then CA collapses this pair of arcs and this does not disconnect~$w$. On the other hand, if $s$~is an input string ($s \in \mathcal{S}$), then there are two copies of $(\pref(s), s), (s, \suff(s))$ in~$L$. At least one copy of this pair is collapsed in~$L$ and therefore belongs to~$L_2$.
The statement for~$v$ with $|v|=1$ follows from Lemma~\ref{lem:path}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:cc3}]
As discussed above, it suffices to prove that $\operatorname{above}(D,i)=\operatorname{above}(L_i, i)$ for every $i=2,1,0$.
\begin{description}
\item {\em Level $i=2$.} The base case $i=2$ is straightforward: clearly, the Collapsing Algorithm leaves exactly one copy of arcs $(\pref(s), s)$ and $(s, \suff(s))$ for every $s \in \mathcal{S}$ and fully collapses all other nodes at level~$3$. Then, $\operatorname{above}(L_2,2)=\operatorname{above}(D,2)$ as $(\pref(s), s), (s, \suff(s))$ for $s \in \mathcal{S}$ are the only edges between levels~2 and~3 in $D$.
\item {\em Level $i=1$.} Note that $\operatorname{above}(L_2,2) \subseteq L_2$ and $L_2$ is an Eulerian cycle. Hence, $\operatorname{above}(L_2,2)$ is a~collection of walks. Consider such a~walk~$w$ and consider two cases.
\begin{itemize}
\item $w$ is a~closed walk. Let~$v$ be the lexicographically largest node of~$w$ at level~2. What we want to show is that in $L_1$ this closed walk~$w$ is connected to the rest of $L_1$ through a~pair of arcs $(\pref(v), v)$ $(v, \suff(v))$ only.
By Lemma~\ref{lemma:toepsilon}, there is a~path from $v$ to $\varepsilon$ in $L_2$ and hence $(v, \suff(v)) \in L_2$; similarly, $(\pref(v), v) \in L_2$. Since $v$~is lexicographically largest at level~$2$ in~$w$, when CA starts processing the node~$v$, all other nodes at level~2 in~$w$ are fully collapsed, i.e., for any such node~$u$, $(\pref(u), u) \not \in L_1$ and $(u, \suff(u)) \not \in L_1$. Moreover, CA does not collapse the pair of arcs
$(\pref(v), v), (v, \suff(v))$ as this would disconnect~$w$ from the rest of the solution.
\item $w$ is not closed. Denote by~$v_1$ and~$v_k$ its first and last nodes. All other nodes of~$w$ in $\operatorname{above}(L_2,2)$ are balanced. What we want to show is that in $L_1$ the only edges between levels~1 and~2 that connect~$w$ to the rest of the solution are $(\pref(v_1), v_1)$ and $(v_k, \suff(v_k))$.
We prove this for $v_k$ (for $v_1$ is it shown similarly). By Lemma~\ref{lemma:toepsilon}, there is a~path from $v_k$ to $\varepsilon$ in $L_2$ and hence $(v_k, \suff(v_k)) \in L_2$. The algorithm CA always works with an Eulerian solution and hence every node is balanced at every stage (i.e., its in-degree is equal to its out-degree). This means that $(v_k, \suff(v_k)) \in L_1$ and that all intermediate nodes of~$w$ are not connected to level~1 nodes in~$L_1$.
\end{itemize}
\item {\em Level $i=0$.} Note that $\operatorname{above}(L_1,1)$ is a~collection of walks. The case of a~non-closed walk in this case is easy as it must be connected to~$\varepsilon$ directly. For this reason, we focus on a~closed walk~$w$ in $\operatorname{above}(L_1,1)$.
We show that for every node~$v$ of~$w$ with $|v|=1$, $L_1$ contains arcs $(\varepsilon, v)$ and $(v, \varepsilon)$ (recall that for $|v|=1$, $\pref(v)=\suff(v)=\varepsilon$). This suffices as then CA (when processing level one nodes) collapses all nodes of~$w$ at level~1 except for the lexicographically largest one, and this is exactly how~$w$ is connected to~$\varepsilon$ in~$D_0$. Below, we show that $(\varepsilon, v) \in L_1$. It then follows that $(v, \varepsilon) \in L_1$ (as $L_1$ must be Eulerian).
Lemma~\ref{lemma:toepsilon} guarantees that $L_1$~contains a~path from~$v$ to~$\varepsilon$ that does not contain nodes at level~3. If the first arc of this path goes down to~$\varepsilon$, then there is nothing to prove. Hence, consider a~case when the first arc goes up to a~node~$u$ (and hence $v=\pref(u)$). The next arc then must go down to~$\suff(u)$.
Hence, $(\pref(u), u), (u, \suff(u)) \in D_1$. This may happen in two cases only: either $u$~is an~input string (i.e., $u \in \mathcal{S}$) or $u$~is the last chance of its component to be connected to the rest of the solution (i.e., exactly for this reason GHA added these two edges to the solution). The former case is straightforward: then there were at least two copies of the arcs $(\pref(u), u), (u, \suff(u))$ and CA collapsed at least one copy. Let us then focus on the latter case.
Let $x,y \in \mathcal{S}$ be such that $u=\suff(y)$ and $c := \suff(x)=\pref(y)$, see the picture below (solid arcs belong to~$L$, dashed arc belong to $L_2$).
\begin{mypic}
\node[vertex] (u) at (3,3) {$u$};
\node[vertex] (v) at (2,2) {$v$};
\node[vertex] (eps) at (1,1) {$\varepsilon$};
\node[vertex] (vv) at (0,2) {};
\node[inputvertex] (y) at (2,4) {$y$};
\node[inputvertex] (x) at (0,4) {$x$};
\node[vertex] (xl) at (-1,3) {};
\node[inputvertex] (z) at (4,4) {};
\node[vertex] (xy) at (1,3) {$c$};
\foreach \s/\t in {y/u, x/xy, xy/y, u/z, xl/x}
\draw[->] (\s) -- (\t);
\foreach \s/\t in {xl/vv, vv/xy, xy/v, v/u}
\draw[->,dashed] (\s) -- (\t);
\end{mypic}
Note that
\[v = \pref(u)=\pref(\suff(y))=\suff(\pref(y))=\suff(c) \, .\]
Hence, $(c,v), (v,u) \in L_2$ (resulting from collapsing at least one pair of arcs $(c,y), (y,u) \in L$). $L_2$ also contains a~pair of arcs $(\pref(x), \suff(\pref(x))), (\suff(\pref(x)), c)$. When processing the node~$c$, CA collapses the pair of arcs $(\pref(c), c), (c, v)$ as there is an arc~$(v,u)$.
Hence, $(\varepsilon, v) \in L_1$, as required. (It may be the case that $x=y$. Then $x={\tt aaa}$, $v=\{a\}$. Then the first pair of arcs of the considered path is ${\tt a} \to {\tt aa} \to {\tt a}$ and one may just drop them.)
\end{description}
As a~final remark, note that if a~walk~$w \in \operatorname{above}(L_i,i)$ is connected to the rest of a~solution through some a~of arcs $(\pref(v), v), (u, \suff(u))$ ($v$~and~$u$ may coincide), then any other balanced node in~$w$ at level~$i$ can be fully collapsed, as every such collapse, thanks to Lemma~\ref{lem:path}, does not disconnect~$w$ or any other walk from~$\operatorname{above}(L_i,i)$ from the rest of the solution.
\end{proof}
\section{Collapsing Conjecture}
\label{sec:collapsing}
We are now ready to conjecture an astonishing structural property of the hierarchical graph:
\begin{quote}
Take any Eulerian solution, double every arc of it, and normalize the resulting solution; the result is the same for all initial solutions!
\end{quote}
For the formal statement of the conjecture we use the following notation: If $U$ and $V$ are two multisets, then $U\sqcup V$ is the multiset~$W$ such that each $w\in W$ has multiplicity equal to the sum of multiplicities it has in sets $U$ and $V$.
Formally, the conjecture is stated as follows.
\newtheorem*{scs}{Collapsing Conjecture}
\begin{scs}
For any set of strings ${\cal S}$ and any two Eulerian solutions~$D_1, D_2$ of~${\cal S}$,
\begin{align*}
CA({\cal S}, D_1 \sqcup D_1) = CA({\cal S}, D_2 \sqcup D_2) \, .
\end{align*}
\end{scs}
Figures~\ref{fig:coll} and~\ref{fig:collnaive} illustrate the action of
the Collapsing Algorithm for optimal and naive solutions, respectively. Note that the resulting solutions are equal. When processing level $l>1$ nodes,
the collapsing procedure does not change the total length of the solution. What one normally sees at the beginning of the
$l=1$ iteration is an~Eulerian solution with many
redundant pairs of arcs of the form $({\tt a}, \varepsilon)$, $(\varepsilon, {\tt a})$. It is exactly this stage of the algorithm where the total length of a~solution is decreased by the Collapsing Algorithm.
\begin{figure}[!ht]
\begin{mypic}
\begin{scope}[scale=0.99,transform shape]
\we{0}{0}{a}{
\foreach \f/\t/\a in {eps/a/10, eps/a/20,
a/aa/10, a/aa/20,
aa/aaa/10, aa/aaa/20,
aaa/aa/10, aaa/aa/20,
aa/a/10, aa/a/20,
a/ae/0, a/ae/10,
ae/aec/0, ae/aec/10,
aec/ec/0, aec/ec/10,
ec/c/0, ec/c/10,
c/ca/0, c/ca/10,
ca/cae/0, ca/cae/10,
cae/ae/0, cae/ae/10,
ae/e/0, ae/e/10,
e/ee/10, e/ee/20,
ee/eee/10, ee/eee/20,
eee/ee/10, eee/ee/20,
ee/e/10, ee/e/20,
e/eps/10, e/eps/20}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{57}{0}{b}{
\foreach \f/\t/\a in {eps/a/10, eps/a/20,
a/aa/10, a/aa/20, a/aa/30, aa/a/30,
aa/aaa/10,
aaa/aa/10,
aa/a/10, aa/a/20,
a/ae/0, a/ae/10,
ae/aec/0, ae/e/20,
aec/ec/0, e/ec/0,
ec/c/0, ec/c/10,
c/ca/0, c/ca/10,
ca/cae/0, ca/a/0, a/ae/20,
cae/ae/0,
ae/e/0, ae/e/10,
e/ee/10, e/ee/20, e/ee/30,
ee/eee/10,
eee/ee/10,
ee/e/10, ee/e/20, ee/e/30,
e/ee/30,
e/eps/10, e/eps/20}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{114}{0}{c}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20, eps/a/30, eps/a/40, a/eps/10, a/eps/20,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
a/ae/0, a/ae/10,
ae/aec/0, ae/e/20,
aec/ec/0, e/ec/0,
ec/c/0, ec/c/10,
c/ca/0, c/ca/10,
ca/cae/0, ca/a/0, a/ae/20,
cae/ae/0,
ae/e/0, ae/e/10,
e/ee/10, e/ee/20, e/ee/30,
ee/eee/10,
eee/ee/10,
ee/e/10,
ee/e/20,
ee/e/30,
e/ee/30,
e/eps/10, e/eps/20}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{0}{-55}{d}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20, eps/a/30, eps/a/40,
a/eps/10, a/eps/20, a/eps/30, a/eps/40, a/eps/50,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
ae/aec/0,
aec/ec/0, e/ec/0,
ec/c/0, ec/c/10,
c/ca/0, c/ca/10,
ca/cae/0, ca/a/0,
cae/ae/0,
e/ee/10, e/ee/20, e/ee/30,
ee/eee/10,
eee/ee/10,
ee/e/10, ee/e/20, ee/e/30,
e/ee/30,
e/eps/10, e/eps/20, e/eps/30,
eps/e/10, eps/e/20, eps/e/30}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{57}{-55}{e}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20, eps/a/30, eps/a/40, eps/a/50,
a/eps/10, a/eps/20, a/eps/30, a/eps/40, a/eps/50,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
ae/aec/0,
aec/ec/0, e/ec/0,
ec/c/0, ec/c/10,
c/ca/0,
ca/cae/0,
cae/ae/0,
e/ee/10, e/ee/20, e/ee/30,
ee/eee/10,
eee/ee/10,
ee/e/10, ee/e/20, ee/e/30,
e/ee/30,
e/eps/10, e/eps/20, e/eps/30,
eps/e/10, eps/e/20, eps/e/30,
c/eps/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{114}{-55}{f}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20, eps/a/30, eps/a/40, eps/a/50,
a/eps/10, a/eps/20, a/eps/30, a/eps/40, a/eps/50,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
ae/aec/0,
aec/ec/0,
ec/c/0,
c/ca/0,
ca/cae/0,
cae/ae/0,
e/ee/10, e/ee/20, e/ee/30,
ee/eee/10,
eee/ee/10,
ee/e/10, ee/e/20, ee/e/30,
e/ee/30,
e/eps/10, e/eps/20, e/eps/30, e/eps/40,
eps/e/10, eps/e/20, eps/e/30,
c/eps/10, eps/c/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{0}{-110}{g}{
\foreach \f/\t/\a in {aa/aaa/10, aaa/aa/10, ca/cae/0, cae/ae/0, ae/aec/0, aec/ec/0, ee/eee/10, eee/ee/10, aa/a/10, a/aa/10, c/ca/0, ec/c/0, ee/e/10, e/ee/10, a/aa/10, aa/a/10, eps/c/10, c/eps/10,
e/eps/10, eps/e/10,
e/eps/20, eps/e/20,
e/eps/30, eps/e/30,
e/eps/40, eps/e/40,
e/eps/50, eps/e/50,
a/eps/10, eps/a/10,
a/eps/20, eps/a/20,
a/eps/30, eps/a/30,
a/eps/40, eps/a/40,
a/eps/50, eps/a/50}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{57}{-110}{h}{
\foreach \f/\t/\a in {aa/aaa/10, aaa/aa/10, ca/cae/0, cae/ae/0, ae/aec/0, aec/ec/0, ee/eee/10, eee/ee/10, aa/a/10, a/aa/10, c/ca/0, ec/c/0, ee/e/10, e/ee/10, a/aa/10, aa/a/10, eps/c/10, c/eps/10,
e/eps/10, eps/e/10,
a/eps/10, eps/a/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\end{scope}
\end{mypic}
\caption{Stages of applying the Collapsing Algorithm to the dataset $\{{\tt aaa}, {\tt cae}, {\tt aec}, {\tt eee}\}$ and its \textbf{optimal} solution. (a)~We start by doubling every arc of the optimal solution from Figure~\ref{fig:hgex}(c).
(b)~After collapsing all nodes at level $l=3$.
(c)~After processing the node {\tt aa} at level $l=2$. Note that the algorithm leaves a~pair of arcs $({\tt a}, {\tt aa}), ({\tt aa}, {\tt a})$ as they are needed to connect the component $\{{\tt aa}, {\tt aaa}\}$ to the rest of the solution. (d)~After processing the {\tt ae} node. The algorithm collapses all pairs of arcs for this node as it lies in the same component as the~node~{\tt c}.
(e)~After processing the {\tt ca} node.
(f)~After processing the {\tt ec} node.
(g)~After processing the {\tt ee} node. Note that at this point the solution has exactly the same length as at the very beginning (at stage~(a)).
(h)~Finally, after collapsing all the unnecessary pairs of arcs from the level~$l=1$.
}
\label{fig:coll}
\end{figure}
\begin{figure}[!ht]
\begin{mypic}
\we{0}{0}{a}{
\foreach \f/\t/\a in {
eps/a/10,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
a/eps/10,
eps/c/10,
c/ca/0,
ca/cae/0,
cae/ae/0,
ae/aec/0,
aec/ec/0,
ec/c/0,
c/eps/10,
eps/e/10,
e/ee/10,
ee/eee/10,
eee/ee/10,
ee/e/10,
e/eps/10
}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{57}{0}{b}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20,
a/aa/10, a/aa/20,
aa/aaa/10, aa/aaa/20,
aaa/aa/10, aaa/aa/20,
aa/a/10, aa/a/20,
a/eps/10, a/eps/20,
eps/c/10, eps/c/20,
c/ca/0, c/ca/10,
ca/cae/0, ca/cae/10,
cae/ae/0, cae/ae/10,
ae/aec/0, ae/aec/10,
aec/ec/0, aec/ec/10,
ec/c/0, ec/c/10,
c/eps/10, c/eps/20,
eps/e/10, eps/e/20,
e/ee/10, e/ee/20,
ee/eee/10, ee/eee/20,
eee/ee/10, eee/ee/20,
ee/e/10, ee/e/20,
e/eps/10, e/eps/20
}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{114}{0}{c}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20,
a/aa/10, a/aa/20, a/aa/30,
aa/aaa/10,
aaa/aa/10,
aa/a/10, aa/a/20, aa/a/30,
a/eps/10, a/eps/20,
eps/c/10, eps/c/20,
c/ca/0, c/ca/10,
ca/cae/0, ca/a/0,
cae/ae/0, a/ae/0,
ae/aec/0, ae/e/0,
aec/ec/0, e/ec/0,
ec/c/0, ec/c/10,
c/eps/10, c/eps/20,
eps/e/10, eps/e/20,
e/ee/10, e/ee/20,
ee/eee/10, ee/e/30,
eee/ee/10, e/ee/30,
ee/e/10, ee/e/20,
e/eps/10, e/eps/20
}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{0}{-55}{d}{
\foreach \f/\t/\a in {
eps/a/10, eps/a/20, eps/a/30, eps/a/40, eps/a/50,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
a/eps/10, a/eps/20, a/eps/30, a/eps/40, a/eps/50,
eps/c/10, eps/c/20, eps/c/30,
c/ca/0,
ca/cae/0,
cae/ae/0,
ae/aec/0,
aec/ec/0,
ec/c/0,
c/eps/10, c/eps/20, c/eps/30,
eps/e/10, eps/e/20, eps/e/30, eps/e/40, eps/e/50,
e/ee/10,
ee/eee/10,
eee/ee/10,
ee/e/10,
e/eps/10, e/eps/20, e/eps/30, e/eps/40, e/eps/50
}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{57}{-55}{e}{
\foreach \f/\t/\a in {
eps/a/10,
a/aa/10,
aa/aaa/10,
aaa/aa/10,
aa/a/10,
a/eps/10,
eps/c/10,
c/ca/0,
ca/cae/0,
cae/ae/0,
ae/aec/0,
aec/ec/0,
ec/c/0,
c/eps/10,
eps/e/10,
e/ee/10,
ee/eee/10,
eee/ee/10,
ee/e/10,
e/eps/10
}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\end{mypic}
\caption{Stages of applying the Collapsing Algorithm to the dataset $\{{\tt aaa}, {\tt cae}, {\tt aec}, {\tt eee}\}$ and its \textbf{naive} solution resulting from overlapping the input strings in the same order as they are given. (a)~The solution of length 10 corresponding to the superstring {\tt aaacaeceee}. (b)~The doubled solution. (c)~After collapsing the $l=3$ level. (d)~After collapsing the $l=2$ level. (e)~After collapsing the $l=1$ level.
}
\label{fig:collnaive}
\end{figure}
We have verified the conjecture on millions of datasets (both handcrafted and randomly generated), and we invite the reader to see its visualizations and to check the conjecture on arbitrary datasets at the webpage~\cite{webpage}.
Moreover,
we support the conjecture by proving that it holds in the (NP-hard) special case where the input strings have length at most~3 in Section~\ref{subsec:scs3}.
If the Collapsing Conjecture is true, then there is a~simple and natural 2-approximate algorithm for SCS: take {\em any} Eulerian solution (e.g., merge the input strings in arbitrary order), double it, and apply the Collapsing Algorithm. Under the conjecture, this results in the same Eulerian solution as for doubled optimal solution and hence the length of the result is at most twice the optimal length.
\section{Further Directions and Open Problems}
The most immediate open problems are to prove the Collapsing Conjecture or the Weak Greedy Hierarchical Conjecture.
\subsection{Applications of Hierarchical Graphs}
It would also be interesting to find other applications of the
hierarchical graphs. We list two such potential applications below.
\begin{description}
\item[Exact algorithms.] Can one use hierarchical graphs to solve SCS exactly in time $(2-\varepsilon)^n$?
It was shown in Section~\ref{sec:intro} that the SCS problem is a special case of the Traveling Salesman Problem. The best known exact algorithms for Traveling Salesman run in time $2^n \poly(|\inp|)$~\cite{B1962, HK1971, KGK1977, K1982, BF1996}. These algorithms stay the best known for the SCS problem as well. The hierarchical graphs were introduced~\cite{scs_exact} for an algorithm solving SCS on strings of length at most $r$ in time $(2-\varepsilon)^n$ (where $\varepsilon$ depends only on $r$). Can one use the hierarchical graph to solve exactly the general case of SCS in time $(2-\varepsilon)^n$ for a constant $\varepsilon$?
\item[Genome assembly.] The hierarchical graph in a~sense
generalizes de Bruijn graph. The latter one is heavily used
in genome assembly~\cite{pevzner2001eulerian}.
Can one adopt the hierarchical graph for this task? For this, one
would need to come up with a~compact representation of the graph
(as datasets in genome assembly are massive) as well as with a~way of
handling errors in the input data. Cazaux and Rivals~\cite{cazaux2018hierarchical} propose a linear-space counterpart of the hierarchical graph.
\end{description}
\subsection{Optimal Cycle Covers}
A superstring corresponds to a Hamiltonian path in the overlap graph, thus, a~minimum-weight cycle cover gives a~natural lower bound on its length.
The Greedy Conjecture claims that a greedy solution never exceeds twice the length of an optimal solution. It is also believed (see, e.g., \cite{weinard2006greedy,laube2005conditional}) that the greedy solution does not exceed the length of an optimal solution plus the length of an optimal cycle cover.
This has interesting counterparts in the hierarchical graphs.
\begin{itemize}
\item Note that an optimal cycle cover in the overlap graph can be constructed by a~straightforward greedy algorithm: keep taking heavy edges till the cycle cover is constructed. The proof of correctness of this algorithm relies on the Monge inequality. Interestingly, to construct an optimal cycle cover in the hierarchical graph, it suffices to invoke the Greedy Hierarchical Algorithm with lines 7--11 commented out! In a~sense, the Monge inequality is satisfied in the hierarchical graph automatically as it contains more information about input strings than just its pairwise overlaps. \item As discussed in Section~\ref{subsec:scs3}, for strings of length~3 even a~more general fact than Collapsing Conjecture holds: it suffices to have double edges adjacent to input strings. One simple way to force a~particular solution to satisfy this property is to double every edge of it. At the same time, adding a~shortest cycle cover to it is guaranteed to be as good.
\item Hence, the more general version of the Collapsing Conjecture is the following: take any solution, add any cycle cover to it, and collapse; the result is always the same. We tested this stronger conjecture and did not find any counter-examples.
\end{itemize}
\section{Definitions}
\subsection{Shortest Common Superstring Problem}
\label{sec:def_scs}
For a string $s$, by $|s|$ we denote the length of $s$.
For strings~$s$ and~$t$, by $\overlap(s,t)$
we denote the longest suffix of~$s$ that is also
a~prefix of~$t$. By $\pref(s,t)$
we denote the first $|s|-|\overlap(s,t)|$ symbols of $s$.
Similarly, $\suff(s,t)$ is the last
$|t|-|\overlap(s,t)|$ symbols of~$t$.
By $\pref(s)$ and $\suff(s)$ we denote, respectively,
the first and the last $|s|-1$ symbols of~$s$. See Figure~\ref{fig:overlap} for a~visual explanation. We denote the empty string by $\varepsilon$.
\begin{figure}[ht]
\begin{mypic}
\begin{scope}
\draw (0,2) rectangle (6,2.5);
\draw[step=5mm] (0,2) grid (6,2.5);
\node[left] at (0,2.25) {$s$};
\draw (3,0.5) rectangle (10,1);
\draw[step=5mm] (3,0.5) grid (10,1);
\node[right] at (10,0.75) {$t$};
\foreach \x in {3.25, 3.75, ..., 5.75}
\draw[gray,thick] (\x,2) -- (\x,1);
\foreach \f/\t/\y/\lab in {0/3/1.5/{\pref(s,t)},
3/6/0/{\overlap(s,t)}, 6/10/0/{\suff(s,t)}, 0/5.5/3/\pref(s),
0.5/6/3.5/\suff(s)}
\path (\f,\y) edge[<->] node[rectangle,inner sep=0.5mm,fill=white] {\strut $\lab$} (\t,\y);
\foreach \x/\a in {0/b, 0.5/a, 1/a, 1.5/c, 2/a, 2.5/b, 3/b, 3.5/c, 4/a, 4.5/a, 5/c, 5.5/b}
\node at (\x+0.25,2.25) {\tt \a};
\foreach \x/\a in {3/b, 3.5/c, 4/a, 4.5/a, 5/c, 5.5/b, 6/a, 6.5/c, 7/a, 7.5/a, 8/a, 8.5/b, 9/c, 9.5/a}
\node at (\x+0.25,0.75) {\tt \a};
\end{scope}
\end{mypic}
\caption{Pictorial explanations of $\pref$, $\suff$, and $\overlap$ functions.}
\label{fig:overlap}
\end{figure}
Throughout the paper by ${\cal S}=\{s_1, \dots, s_n\}$ we denote
the set of~$n$ input strings. We assume that no input string is a~substring of another (such a~substring can be removed from $\mathcal{S}$ in the preprocessing stage). Note that SCS is a~{\em permutation problem}: to find a~shortest string containing all $s_i$'s in a~{\em given order} one just
overlaps the strings in this order, see Figure~\ref{fig:permutation}. (This simple observation relates SCS to other permutation problems, including various versions of the Traveling Salesman Problem.) It will prove convenient to view the SCS problem as a~problem of finding an optimum permutation.
It should be noted at the same time that the correspondence between permutations and superstrings is not one-to-one: there are superstrings that do not correspond to any permutation. For example, the concatenation of input strings is clearly a~superstring, but it ignores the fact that neighbor strings may have non-trivial overlaps and for this reason may fail to correspond to a~permutation. Still, clearly, any {\em shortest} superstring corresponds to a~permutation of the input strtings.
\begin{figure}[ht]
\begin{mypic}
\foreach \x in {0, 2.5, 5.5, 10, 12.5, 16}
\draw[dashed,gray,thin] (\x,-0.5) -- (\x,4);
\foreach \x/\y/\len/\label in {0/3/5/s_{i_1}, 1/2.5/9/s_{i_2}, 2/2/16/s_{i_3}, 10.5/1/4/s_{i_{n-1}}, 12/0.5/8/s_{i_n}} {
\draw (\x,\y) rectangle (\x+0.5*\len,\y+0.5);
\node at (\x+0.25*\len,\y+0.25) {$\label$};
}
\foreach \f/\t/\label in {0/2.5/{s_{i_1}}, 2.5/5.5/{\suff(s_{i_1}, s_{i_2})}, 5.5/10/{\suff(s_{i_2}, s_{i_3})}, 12.5/16/{\suff(s_{i_{n-1}}, s_{i_n})}}
\path (\f,0) edge[<->] node[rectangle,inner sep=0.5mm,fill=white] {\strut $\label$} (\t,0);
\node at (11,0) {$\dotsb$};
\node at (10.5,1.75) {$\dotsb$};
\end{mypic}
\caption{SCS is a~permutation problem. The length of a~superstring corresponding to a~permutation $(s_{i_1}, \dotsc, s_{i_n})$ is $|s_{i_1}|$ plus the sum of the lengths of suffixes of consecutive pairs of strings. It is also equal to $\sum_{i=1}^n |s_i|-\sum_{j=1}^{n-1}|\overlap(s_{i_j}, s_{i_{j+1}})|$.}
\label{fig:permutation}
\end{figure}
\subsection{Hierarchical Graph}
\label{sec:def_hier}
For a~set of strings~${\cal S}$, the~\emph{hierarchical graph} $HG=(V,E)$ is a~weighted directed graph with $V=\{v \colon \text{$v$ is a~substring of some $s \in {\cal S}$}\}$. For every $v \in V,\, v \neq \varepsilon$, the set of arcs~$E$ contains an {\em up-arc} $(\pref(v), v)$ of weight~1 and a {\em down-arc} $(v, \suff(v))$ of weight~0. The meaning of an up-arc is appending one symbol to the end of the current string (and that is why it has weight~1), whereas the meaning of a down-arc is cutting down one symbol from the beginning of the current string.
Figure~\ref{fig:hgex}(a) gives an example of the hierarchical graph and shows that the terminology of up- and down-arcs comes from placing all the strings of the same length at the same level, where the $i$-th level contains strings of length~$i$. In all the figures in this paper, the input strings are shown in rectangles, while all other vertices are ellipses.
\newcommand{\we}[4]{
\begin{scope}[xshift=#1mm,yshift=#2mm]
\foreach \n/\x/\y in {aaa/0/3, cae/1/3, aec/3/3, eee/4/3}
\node[inputvertex] (\n) at (\x,\y) {\tt \n};
\foreach \n/\x/\y in {aa/0/2, ca/1/2, ae/2/2, ec/3/2, ee/4/2, a/1/1, c/2/1, e/3/1}
\node[vertex] (\n) at (\x,\y) {\tt \n};
\node[vertex] (eps) at (2,0) {$\varepsilon$};
\foreach \f/\t/\a in {eps/e/10, e/eps/10, eps/c/10, c/eps/10, eps/a/10, a/eps/10, a/aa/10, aa/a/10, aa/aaa/10, aaa/aa/10, c/ca/0, ca/cae/0, cae/ae/0, ae/aec/0, aec/ec/0, ee/eee/10, eee/ee/10, e/ee/10, ee/e/10, ca/a/0, a/ae/0, ae/e/0, e/ec/0, ec/c/0}
\path (\f) edge[hgedge,bend left=\a] (\t);
\node at (2,-1) {(#3)};
#4
\end{scope}
}
\begin{figure}[!ht]
\begin{mypic}
\we{0}{0}{a}{}
\we{57}{0}{b}{
\foreach \f/\t/\a in {eps/a/10, a/ae/0, ae/aec/0, aec/ec/0, ec/c/0, c/ca/0, ca/a/0, a/eps/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{114}{0}{c}{
\foreach \f/\t/\a in {eps/a/10, a/aa/10, aa/aaa/10, aaa/aa/10, aa/a/10, a/ae/0, ae/aec/0, aec/ec/0, ec/c/0, c/ca/0, ca/cae/0, cae/ae/0, ae/e/0, e/ee/10, ee/eee/10, eee/ee/10, ee/e/10, e/eps/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\end{mypic}
\caption{(a)~Hierarchical graph for the~dataset $\mathcal{S}=\{{\tt aaa}, {\tt cae}, {\tt aec}, {\tt eee}\}$. (b)~The~walk $\varepsilon \to {\tt a} \to {\tt ae} \to {\tt aec} \to {\tt ec} \to {\tt c} \to {\tt ca} \to {\tt a} \to \varepsilon$ has length (or weight)~4 and spells the~string {\tt aeca} of length~4. (c)~An~optimal superstring for~$\mathcal{S}$ is {\tt aaaecaeee}. It has length~9, corresponds to the~permutation $({\tt aaa}, {\tt aec}, {\tt cae}, {\tt eee})$, and defines the~walk of length~9 shown in black.}
\label{fig:hgex}
\end{figure}
What we are looking for in this graph is a shortest
walk from $\varepsilon$ to $\varepsilon$ going
through all the nodes from~$\mathcal{S}$.
It is not difficult to see that the length of a~walk
from $\varepsilon$ to $\varepsilon$ equals the
length of the string spelled by this walk.
This is just because each up-arc has
weight~$1$ and adds one symbol to
the current string. See Figure~\ref{fig:hgex}(b) for an~example.
Hence, the SCS problem is equivalent to finding a~shortest closed walk from $\varepsilon$ to $\varepsilon$ that visits all nodes from~${\cal S}$. Note that a walk may contain repeated nodes and arcs. The multiset of arcs of such a~walk must be Eulerian (each vertex must have the same in- and out-degree, and the set of arcs must be connected). It will prove convenient to define an~{\em Eulerian solution} in a~hierarchical graph as an~Eulerian multiset of arcs~$D$ that goes through $\varepsilon$ and all nodes from~${\cal S}$. Given such a~solution~$D$, one can easily recover an Eulerian cycle (that might not be unique). This cycle spells a~superstring of~${\cal S}$ of the same length as~$D$. Figure~\ref{fig:hgex}(c) shows an~optimal Eulerian solution.
A solution to SCS defines a permutation $(s_{i_1}, \dotsc, s_{i_n})$ of the input strings, and this permutation naturally gives a~``zig-zag'' Eulerian solution in the hierarchical graph:
\begin{align}
\label{eq:zigzag}
\varepsilon \to s_{i_1} \to \overlap(s_{i_1}, s_{i_2}) \to s_{i_2} \to
\overlap(s_{i_2}, s_{i_3}) \to \dotsb
\to s_{i_n} \to \varepsilon \, .
\end{align}
This Eulerian solution is shown schematically in Figure~\ref{fig:hgperm}(a). This schematic illustration is over simplified as the shown path usually has many self-intersections. Still, this point of view is helpful in understanding the algorithms presented later in the text.
Figure~\ref{fig:hgperm}(b) shows an ``untangled'' optimal Eulerain solution from Figure~\ref{fig:hgex}(c): by contracting nodes with equal labels into the same node, one gets exactly the solution from Figure~\ref{fig:hgex}(c).
\begin{figure}[ht]
\begin{mypic}
\begin{scope}[yscale=0.8]
\foreach \n/\x/\y in {aaa/0/3, cae/4/3, aec/2/3, eee/6/3}
\node[inputvertex] (\n) at (\x,\y) {\tt \n};
\foreach \n/\t/\x/\y in {aa1/aa/-0.5/2, aa2/aa/0.5/2, ae1/ae/1.5/2, ec/ec/2.5/2, ca/ca/3.5/2, ae2/ae/4.5/2, ee1/ee/5.5/2, ee2/ee/6.5/2, a1/a/0/1, a2/a/1/1, c/c/3/1, e1/e/5/1, e2/e/6/1}
\node[vertex] (\n) at (\x,\y) {\tt \t};
\node[vertex] (eps) at (3,0) {$\varepsilon$};
\foreach \f/\t in {aa1/aaa, aaa/aa2, ae1/aec, aec/ec, ca/cae, cae/ae2, ee1/eee, eee/ee2, a1/aa1, aa2/a2, a2/ae1, ec/c, c/ca, ae2/e1, e1/ee1, ee2/e2}
\draw[->] (\f) -- (\t);
\path (eps) edge[->,out=180,in=-45] (a1);
\path (e2) edge[->,out=-135,in=0] (eps);
\foreach \n/\x in {1/-8, 2/-7, 3/-6, n/-3}
\node[inputvertex] (\n) at (\x,3) {$s_{i_{\n}}$};
\node at (-4.5,3) {$\dotsb$};
\foreach \n/\x/\y in {12/-7.5/2, 23/-6.5/1, 34/-5.5/1.5, 67/-3.5/2}
\node[vertex] (\n) at (\x,\y) {};
\foreach \f/\t in {1/12, 12/2, 2/23, 23/3, 3/34, 67/n}
\draw[->,anypath] (\f) -- (\t);
\node[vertex] (eps) at (-5.5,0) {$\varepsilon$};
\path (eps) edge[anypath,->,out=180,in=-90] (1);
\path (n) edge[anypath,->,out=-90,in=0] (eps);
\node at (-5.5,-1) {(a)};
\node at (3,-1) {(b)};
\end{scope}
\end{mypic}
\caption{(a)~A~schematic illustration of a~normalized Eulerian solution.
(b)~Untangled optimal Eulerian solution from Figure~\ref{fig:hgex}(c).}
\label{fig:hgperm}
\end{figure}
Not every Eulerian solution in the hierarchical graph has a~nice zig-zag
structure described above. In the next section, we introduce a~normalization procedure (that we call collapsing)
that allows us to focus on nice Eulerian solutions only.
\subsection{Normalizing a~Solution}
\label{sec:def_normal}
In this section, we describe a~natural way of normalizing an~Eulerian solution~$D$. Informally, it can be viewed as follows.
Imagine that all arcs of~$D$ form one~circular thread, and that there is a~nail in every node~$s \in {\cal S}$ corresponding to an input string. We apply {\em ``gravitation''} to the thread, i.e., we replace every pair of arcs $(\pref(v), v)$, $(v, \suff(v))$
with a~pair $(\pref(v), \pref(\suff(v)))$, $(\pref(\suff(v)), \suff(v))$, if there is no nail in~$v$ and if this does not disconnect~$D$. We call this {\em collapsing}, see Figure~\ref{fig:collapsing}.
\begin{figure}[ht]
\begin{mypic}
\begin{scope}[minimum size=6mm]
\node[vertex] (a) at (0, 1.5) {$\pref(v)$};
\node[vertex,inner sep=1mm] (b) at (1.5, 2.5) {$v$};
\node[vertex] (c) at (3, 1.5) {$\suff(v)$};
\node[vertex] (d) at (1.5, 0.5) {$ \pref(\suff(v))$};
\draw[->,dashed] (a) -- (b);
\draw[->,dashed] (b) -- (c);
\draw[->] (a) -- (d);
\draw[->] (d) -- (c);
\begin{scope}[xshift=80mm]
\node[vertex] (a) at (0, 1.5) {\tt aba};
\node[vertex,inner sep=1mm] (b) at (1.5, 2.5) {\tt abac};
\node[vertex] (c) at (3, 1.5) {\tt bac};
\node[vertex] (d) at (1.5, 0.5) {\tt ba};
\draw[->,dashed] (a) -- (b);
\draw[->,dashed] (b) -- (c);
\draw[->] (a) -- (d);
\draw[->] (d) -- (c);
\end{scope}
\end{scope}
\end{mypic}
\caption{Collapsing a~pair of arcs is replacing a~pair of dashed arcs with a~pair of solid arcs: general case (left) and example (right). The~``physical meaning'' of this transformation is that to get {\tt bac} from {\tt aba} one needs to cut~{\tt a} from the beginning and append~{\tt c} to the end and these two operations commute.}
\label{fig:collapsing}
\end{figure}
A~formal pseudocode of the collapsing procedure is given in Algorithm~\ref{alg:collapse}. The pseudocode, in particular, reveals an important exception (not covered in Figure~\ref{fig:collapsing}):
if $|v|=1$, then $\pref(\suff(v))$ is undefined and we just remove the pair of arcs $(\pref(v), v)$ and $(v, \suff(v))$.
\begin{algorithm}[!ht]
\caption{Collapse}\label{alg:collapse}
\hspace*{\algorithmicindent} \textbf{Input:} hierarchical graph $HG(V,E)$, Eulerian solution~$D$, node~$v \in V$.
\begin{algorithmic}[1]
\If{$(\pref(v), v), (v, \suff(v)) \in D$}
\State\label{alg:col} $D \gets D \setminus \{(\pref(v), v), (v, \suff(v))\}$
\If{$|v| > 1$}
\State $D \gets D \cup \{(\pref(v), \pref(\suff(v))), (\pref(\suff(v)), \suff(v))\}$
\EndIf
\EndIf
\end{algorithmic}
\end{algorithm}
Algorithm~\ref{alg:ca}, that we call Collapsing Algorithm (CA), uses the property described above to normalize any solution.
It drops down all pairs of arcs that are not needed for connectivity.
\ab{(Recall that a~set of edges is called an Eulerian solution if it is connected and goes through all initial nodes and $\varepsilon$.)}
\begin{algorithm}[!ht]
\caption{Collapsing Algorithm (CA)}\label{alg:ca}
\hspace*{\algorithmicindent} \textbf{Input:} set of strings~$\mathcal{S}$, Eulerian solution~$D$ in~$HG$.\\
\hspace*{\algorithmicindent} \textbf{Output:} Eulerian solution $D'\colon |D'|\leq|D|$
\begin{algorithmic}[1]
\For{level~$l$ in $HG$ in descending order}\label{alg:ca_for}
\For{all $v \in V$ s.t. $|v|=l$ in lexicographic order:}
\While{$(\pref(v), v), (v, \suff(v)) \in D$ and collapsing it keeps~$D$ an Eulerian solution}
\State $\text{\sc Collapse}(HG, D, v)$
\EndWhile
\EndFor
\EndFor
\State return $D$
\end{algorithmic}
\end{algorithm}
It is easy to show (we prove this formally in Claim~\ref{claim:zigzag} on page~\pageref{claim:zigzag}) that any normalized solution is of the form~\eqref{eq:zigzag}. But it is not true that every zig-zag solution of the form~\eqref{eq:zigzag} is a normalized solution: see Figure~\ref{fig:abnormalzigzag} for an example. The normalization procedure does not just turn a~solution into some standard form, but it may also decrease its length.
\begin{figure}[ht]
\begin{mypic}
\foreach \n/\x/\y in {ae/1/2, aa/2/2, ca/3/2}
\node[inputvertex] (\n) at (\x,\y) {\tt \n};
\foreach \n/\x/\y in {a/2/1, e/1/1, c/3/1}
\node[vertex] (\n) at (\x,\y) {\tt \n};
\node[vertex] (eps) at (2,0) {$\varepsilon$};
\node at (2,-1) {(a)};
\path (aa) edge[->,bend left] (a);
\path (a) edge[->,bend left] (aa);
\path (eps) edge[->,bend left] (a);
\path (a) edge[->,bend left] (eps);
\path (eps) edge[->,bend left=10] (a);
\path (a) edge[->,bend left=10] (eps);
\foreach \f/\t in {a/ae, ae/e, e/eps, eps/c, c/ca, ca/a}
\draw[->] (\f) -- (\t);
\begin{scope}[xshift=50mm]
\foreach \n/\x/\y in {ae/1/2, aa/2/2, ca/3/2}
\node[inputvertex] (\n) at (\x,\y) {\tt \n};
\foreach \n/\x/\y in {a/2/1, e/1/1, c/3/1}
\node[vertex] (\n) at (\x,\y) {\tt \n};
\node[vertex] (eps) at (2,0) {$\varepsilon$};
\node at (2,-1) {(b)};
\path (aa) edge[->,bend left] (a);
\path (a) edge[->,bend left] (aa);
\foreach \f/\t in {a/ae, ae/e, e/eps, eps/c, c/ca, ca/a}
\draw[->] (\f) -- (\t);
\end{scope}
\end{mypic}
\caption{(a)~An~Eulerian solution corresponding to the~permutation $({\tt ae}, {\tt aa}, {\tt ca})$. (b)~The solution from~(a) after normalization results in a shorter solution corresponding to the~permutation~$({\tt ca}, {\tt aa}, {\tt ae})$. \ab{This example also shows that although collapsing a~pair of edges is a~local change in the graph, it may drastically change the resulting superstring. In this case, it replaces a~superstring {\tt aeaaca} with a~shorter superstring {\tt caae}.}}
\label{fig:abnormalzigzag}
\end{figure}
\section{Greedy Hierarchical Conjecture}
\label{sec:greedy_hier}
In this section, we present one more curious property of the Collapsing Algorithm that reveals its intricate connection to greedy algorithms. For this, we introduce the so called Greedy Hierarchical Algorithm (GHA) that constructs an Eulerian solution in a~stingy fashion, i.e., tries to add as few arcs as possible:
\begin{quote}
Proceed nodes from top to bottom. For each node, ensure that it is balanced and connected to the rest of the solution.
\end{quote}
This is best illustrated with an example, see Figure~\ref{fig:hgexa}. We start constructing an~Eulerian solution~$D$ by processing the nodes at level~$3$. The solution~$D$ must visit all these four nodes, so we add all incoming and outgoing arcs to~$D$, see Figure~\ref{fig:hgexa}(a). We then process the level~2. The node~{\tt aa} is balanced, but if we skip it, it will not be connected to the rest of the solution, so we add to~$D$ the arcs $({\tt a}, {\tt aa})$ and $({\tt aa}, {\tt a})$. The node {\tt ae} is balanced, we do nothing for it. The node {\tt ca} is imbalanced, so we add an arc $({\tt c}, {\tt ca})$ to~$D$. We balance the node {\tt ec} similarly. The node {\tt ee} is processed similarly to the node {\tt aa}. The result of processing the second level is shown in Figure~\ref{fig:hgexa}(b). On the last stage we connect the nodes {\tt a}, {\tt b}, and {\tt c} to $\varepsilon$ to ensure connectivity, see Figure~\ref{fig:hgexa}(c). Hence, when processing level~$l$, we only add arcs between levels~$l$ and~$l-1$.
\begin{figure}[!ht]
\begin{mypic}
\we{0}{0}{a}{
\foreach \f/\t/\a in {aa/aaa/10, aaa/aa/10, ca/cae/0, cae/ae/0, ae/aec/0, aec/ec/0, ee/eee/10, eee/ee/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{57}{0}{b}{
\foreach \f/\t/\a in {aa/aaa/10, aaa/aa/10, ca/cae/0, cae/ae/0, ae/aec/0, aec/ec/0, ee/eee/10, eee/ee/10, aa/a/10, a/aa/10, c/ca/0, ec/c/0, ee/e/10, e/ee/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\we{114}{0}{c}{
\foreach \f/\t/\a in {aa/aaa/10, aaa/aa/10, ca/cae/0, cae/ae/0, ae/aec/0, aec/ec/0, ee/eee/10, eee/ee/10, aa/a/10, a/aa/10, c/ca/0, ec/c/0, ee/e/10, e/ee/10, a/aa/10, aa/a/10, eps/c/10, c/eps/10, e/eps/10, eps/e/10, a/eps/10, eps/a/10}
\path (\f) edge[hgedge,bend left=\a,draw=black,thick] (\t);
}
\end{mypic}
\caption{(a)~After processing the $l=3$ level. (b)~After processing the $l=2$ level. Note that for the node {\tt aa} we add two lower arcs ($({\tt a}, {\tt aa})$ and $({\tt aa}, {\tt a})$) since otherwise the corresponding weakly connected component ($\{{\tt aa}, {\tt aaa}\}$) will not be connected to the rest of the solution. At the same time, when processing the node {\tt ae} we observe that it lies in a~weakly connected component that contains imbalanced nodes ({\tt ca} and {\tt ec}), hence there is no need to add two lower arcs to {\tt ae}. (c)~After processing the $l=1$ level. The resulting solution has length~10 and is, therefore, suboptimal (compare it with the optimal solution shown in Figure~\ref{fig:hgex}(c)).}
\label{fig:hgexa}
\end{figure}
More formally, GHA first considers the input
strings~${\cal S}$. Since we assume that
no $s \in {\cal S}$ is a~substring of another
$t \in {\cal S}$, there is no down-path from~$t$ to~$s$ in $HG$.
This means that any walk through $\varepsilon$ and ${\cal S}$ goes through the arcs $\{(\operatorname{pref}(s), s), (s, \operatorname{suff}(s)) \colon s \in {\cal S}\}$. The algorithm adds all of them to the constructed Eulerian solution~$D$ and starts processing all the nodes level by level, from top to bottom. At each level, we process the nodes in the lexicographic order. If the degree of the current node~$v$ is imbalanced, we balance it by adding an appropriate number of incoming (i.e., $(\pref(v),v)$) or outgoing (i.e., $(v, \suff(v))$) arcs from the previous (i.e., lower) level. In the case when $v$~is balanced, we just skip it. The only exception when we cannot skip it is when {\em $v$~lies in an Eulerian component and $v$ is the last chance of this component to be connected to the rest of the arcs in~$D$}. (See, for example, the vertex $ {\tt aa}$ in Figure~\ref{fig:hgexa}(a)). The pseudocode is given in~Algorithm~\ref{algo:gha}.
\begin{algorithm}[!ht]
\caption{Greedy Hierarchical Algorithm (GHA)}\label{algo:gha}
\hspace*{\algorithmicindent} \textbf{Input:} set of strings~${\cal S}$.\\
\hspace*{\algorithmicindent} \textbf{Output:} Eulerian solution~$D$.
\begin{algorithmic}[1]
\State $HG(V,E) \gets \text{hierarchical graph of ${\cal S}$}$
\State\label{alg:gha_init}$D \gets \{(\operatorname{pref}(s), s), (s, \operatorname{suff}(s)) \colon s \in {\cal S}\}$
\For{level $l$ from $\max\{|s| \colon s \in {\cal S}\}$ downto 1}\label{alg:for}
\For{node $v \in V$ with $|v|=l$ in the lexicographic order}
\If{$|\{(u,v) \in D \colon |u|=|v|+1\}| \neq |\{(v,w) \in D \colon |w|=|v|+1\}| $}
\State\label{alg:step6} balance the degree of $v$ in~$D$ by adding an appropriate number of lower arcs
\Else
\State\label{alg:else} ${\cal C} \gets \text{weakly connected component of $v$ in $D$}$
\State $u \gets \text{the lexicographically largest string among shortest strings in ${\cal C}$}$
\If{${\cal C}$ is Eulerian, $\varepsilon \not \in {\cal C}$, and $v = u$}
\State\label{alg:last} $D \gets D \cup \{(\pref(v), v), (v, \suff(v))\}$
\EndIf
\EndIf
\EndFor
\EndFor
\State return $D$
\end{algorithmic}
\end{algorithm}
While GHA is almost as simple as the standard Greedy Algorithm (GA), GHA has several provable advantages over GA:
\begin{description}
\item One advantage of GHA over GA is that GHA is more flexible in the following sense. On every step, GA selects two strings and fixes tightly their order. GHA instead works to ensure connectivity. When the resulting set~$D$ is connected, an actual order of input strings is given by the corresponding Eulerian cycle through~$D$. This is best illustrated on the following~toy example. For the~dataset $\mathcal{S}=\{{\tt ae}, {\tt ea}, {\tt ee}\}$, GA might produce a~suboptimal solution {\tt aeaee} if it merges the strings {\tt ae} and {\tt ea} at the first step. At the same time, it is not difficult to see that GHA finds an optimal solution for~$\mathcal{S}$.
\item Another advantage of GHA is that, in contrast to GA, it solves {\em exactly} two well known polynomially solvable special cases of SCS: when the input strings have length at most two and when the input strings form a~$k$-spectrum of an unknown string (that is, the input strings constitute all
$k$-substrings of a~string). We prove this formally in Sections~\ref{sec:ghatwo} and~\ref{sec:ghaspectrum}. Informally, this happens because for such datasets there are no connectivity issues for GHA: for $k$-spectrum, after processing the highest level GHA gets a~weakly connected component; for 2-SCS, after processing the level~2, GHA gets several weekly connected components such that different components do not share common letters and therefore are completely independent. Figure~\ref{fig:abnormalzigzag}(b) illustrates this: while GA may produce a~permutation $({\tt ca}, {\tt ae}, {\tt aa})$, GHA constructs an optimal
permutation $({\tt ca}, {\tt aa}, {\tt ae})$.
In Section~\ref{sec:ghatough}, we also show a~dataset where GHA produces a~solution that is almost two times longer than the optimal one.
\item In Section~\ref{sec:gr_im_wghc}, \ab{we show that the approximation guarantee of GHA is no worse than that of GA.
Combining with the result of Kaplan and Shafrir~\cite{KS2005},
this implies immediately that GHA is 3.5-approximate.}
Moreover, we prove that the standard Greedy Conjecture implies $2$-approximation of GHA, which makes it natural to study the approximation ratio of GHA.
\end{description}
We are now ready to state our second conjecture: the results of the Collapsing Algorithm and Greedy Hierarchical Algorithm coincide!
\newtheorem*{ghcc}{Greedy Hierarchical Conjecture}
\begin{ghcc}
For any set of strings~$\mathcal{S}$ and any Eulerian solution~$D$,
\[CA(\mathcal{S}, D \sqcup D) = GHA(\mathcal{S}) \, .\]
\end{ghcc}
While the Greedy Hierarchical Conjecture implies that GHA finds a $2$-approximate solution, we separately state this weak version of the conjecture.
\newtheorem*{wghcc}{Weak Greedy Hierarchical Conjecture}
\begin{wghcc}
GHA is a factor $2$ approximation algorithm for the Shortest Common Superstring problem.
\end{wghcc}
\section{Greedy Hierarchical Algorithm and Special Cases of SCS}
\subsection{Strings of Length~2}\label{sec:ghatwo}
Gallant et al.~\cite{GMS1980} show that SCS on strings of length $3$ is $\mathbf{NP}$-hard, but SCS on strings of length at most $2$ is solvable in polynomial time. In this section we show that GHA finds an optimal solution in this case as well. We note that the standard Greedy Algorithm does not necessarily find an optimal solution in this case. For example, if ${\cal S}=\{{\tt ab}, {\tt ba}, {\tt bb}\}$, the Greedy Algorithm may first merge {\tt ab} and {\tt ba}, which would lead to a~suboptimal solution {\tt ababb}
(recall also Figure~\ref{fig:abnormalzigzag}).
First, we can assume that all input strings from~${\cal S}$ have length exactly~$2$. Indeed, since we assume that no input string is a substring of another input string, all strings of length $1$ are unique symbols which do not appear in other strings. Take any such $s_i$ of length $1$. The optimal superstring length for ${\cal S}$ is $k$ if and only if the optimal superstring length for ${\cal S}\setminus\{s_i\}$ is $k-1$. The Greedy Hierarchical Algorithm has the same behavior: In Step~\ref{alg:gha_init}, GHA will include the arcs $(\varepsilon, s_i), (s_i, \varepsilon)$ in the solution, and it will never touch the vertex $s_i$ again (because it is balanced and connected to $\varepsilon$). Thus, $s_i$ adds $1$ to the length of the Greedy Hierarchical Superstring as well. By the same reasoning, we can assume that each string of length two is primitive, i.e., contains two distinct symbols.
When considering primitive strings ${\cal S}=\{s_1,\ldots,s_n\}$ of length exactly $2$, it is convenient to introduce the following directed graph $G=(V, E)$, where $V$ contains a vertex for every symbol which appears in strings from ${\cal S}$. The graph has $|E|=n$ arcs corresponding to $n$ input strings: for every string $s_i=ab$, there is an arc from $a$ to $b$. It is known~\cite{GMS1980} that the length of an optimal superstring in this case is $n+k$ where $k$ is the minimum number such that $E$ can be decomposed into $k$ directed paths, or, equivalently:
\begin{proposition}[\cite{GMS1980}]
Let $G$ be the graph defined above, and let $G_1=(V_1,E_1),\ldots,G_c=(V_c,E_c)$ be its weakly connected components. Then the length of an optimal superstring is
\begin{align}
\label{eq:gms}
n + \sum_{i=1}^{c} {\max\left( 1, \sum_{v \in V_i}{ \frac{ |\indegree(v) - \outdegree(v)|}{2} }\right)} \; .
\end{align}
\end{proposition}
We will now show that in this case, GHA finds an optimal solution.
\begin{lemma}
Let ${\cal S}=\{s_1,\ldots,s_n\}$ be a set of strings of length at most $2$, and let $s$ be an optimal superstring for ${\cal S}$. Then $GHA({\cal S})$ returns a~superstring of length~$|s|$.
\end{lemma}
\begin{proof}
We showed above that it suffices to consider the case of $n$ primitive strings $\{s_1,\ldots,s_n\}$ of length exactly $2$. For $1\leq i\leq n$, let $s_i=a_i b_i$, where $a_i\neq b_i$. Consider the partial greedy hierarchical solution $D$ after the Step~\ref{alg:gha_init} of the GHA algorithm: $D=\{(a_i, a_ib_i), (a_ib_i, b_i): 1\leq i\leq n \}$. (We abuse notation by identifying the set of arcs $D$ with the graph induced by $D$.) This partial solution has $n$ up-arcs, so its current weight is $n$.
Note that by the definition of the graph $G$ above, $G$ contains an arc $(a, b)$ if and only if $D$ has the arcs $(a, ab)$ and $(ab, b)$ of the graph HG. Thus, the indegree (outdegree) of a vertex $a$ in $G$ equals the indegree (outdegree) of the vertex $a$ in the partial solution $D$. Also, two vertices $a$ and $b$ of $G$ belong to one weakly connected component in $G$ if and only if they belong to one weakly connected component in $D$. Therefore, the expression~\eqref{eq:gms} in $G$ has the same value in the partial solution graph $D$. (Indeed, the vertices of $D$ corresponding to strings of length $2$ are balanced and do not form weakly connected components.)
Now we proceed to Steps~\ref{alg:for}--\ref{alg:last} of GHA. GHA will go through all strings of length $1$, and add $|\indegree(v) - \outdegree(v)|$ arcs for each unbalanced vertex $v$. The Steps~\ref{alg:else}--\ref{alg:last} ensure that each weakly connected component adds at least a pair of arcs. Since exactly a half of added arcs are up-arcs, we have increased the weight of the partial solution $D$ by
\begin{align*}
\sum_{i=1}^{c} {\max\left( 1, \sum_{v \in V_i}{ \frac{ |\indegree(v) - \outdegree(v)|}{2} }\right)} \; .
\end{align*}
\end{proof}
\subsection{Spectrum of a~String}\label{sec:ghaspectrum}
By a~$k$-spectrum of a~string $s$
(of length at least~$k$)
we mean a~set of all substrings of~$s$ of length~$k$.
Pevzner et al.~\cite{pevzner2001eulerian} give a polynomial time exact algorithm for the case when the input strings form a spectrum of an unknown string. We show that GHA also finds an optimal solution in this case.
\begin{lemma}
Let ${\cal S}=\{s_1,\ldots,s_n\}$ be a~$k$-spectrum of an unknown string~$s$. Then $GHA({\cal S})$ returns a~superstring of length at most~$|s|$.
\end{lemma}
\begin{proof}
Since $s$ has $n$ distinct substrings of length $k, \, |s|\geq n+k-1$. We will show that GHA finds a superstring of length $n+k-1$. After Step~\ref{alg:gha_init} of GHA, the partial solution $D = \{(\operatorname{pref}(s), s), (s, \operatorname{suff}(s)) \colon s \in {\cal S}\}$. In particular, $D$ is of weight $n$. For $1\leq i\leq k-1$, let $u_{i}$ be the first $i$ symbols of $s$, and let $v_{i}$ be the last $i$ symbols of $s$. Note that $u_{k-1}$ and $v_{k-1}$ are the only unbalanced vertices of the partial solution $D$ after Step~\ref{alg:gha_init}: all other strings of length $k-1$ appear equal number of times as prefixes and suffixes of strings from ${\cal S}$. Therefore, while processing the level $\ell=k-1$, GHA will add one arc to each of the vertices $u_{k-1}$ and $v_{k-1}$, and will not add arcs to other strings of length $k-1$.
In general, while processing the level $\ell=i$, GHA adds one up-arc to $u_i$ and one down-arc to $v_i$. In order to show this, we consider two cases. If $u_i \neq v_i$, then $u_i$ has an incoming arc from the previous step and does not have outgoing arcs, therefore GHA adds an up-arc to $u_i$ in Step~\ref{alg:step6}. Similarly, GHA adds a down-arc from $v_i$. Note that there are no other strings of length $i<k-1$ in the partial solution, so the algorithm moves to the next level. In the case when $u_i=v_i$, we have that all vertices are balanced, but the string $u_i$ is now the shortest string in this only connected component ${\cal C}$ of the graph. Therefore, for $i>0$ we have $\varepsilon\not\in{\cal C}$, and GHA adds an up- and down-arc to $u_i$ in Step~\ref{alg:last}.
We just showed that GHA solution for a $k$-spectrum of a string has the initial set of arcs $D = \{(\operatorname{pref}(s), s), (s, \operatorname{suff}(s)) \colon s \in {\cal S}\}$, and also the arcs $\{ (u_{i-1}, u_{i}), (v_i, v_{i-1})\colon 1\leq i\leq k-1 \}$. Thus, the total number of up-arcs (and the weight of the solution) is $n+k-1$.
\end{proof}
\subsection{Tough Dataset}\label{sec:ghatough}
There is a~well-known dataset consisting of just three strings where the classical greedy algorithm produces a~superstring that is almost twice longer than an optimal one: $s_1={\tt cc}({\tt ae})^n$, $s_2=({\tt ea})^{n+1}$, $s_3=({\tt ae})^n{\tt cc}$. Since $\overlap(s_1, s_3)=2n$,
while $\overlap(s_1,s_2)=\overlap(s_2,s_3)=2n-1$, the greedy algorithm produces a~permutation $(s_1, s_3, s_2)$ (or $(s_2,s_1,s_3)$). I.e., by greedily taking the massive overlap of length $2n$ it loses the possibility to insert $s_2$ between $s_1$ and $s_3$ and to get two overlaps of size $2n-1$. The resulting superstring has length $4n+6$. At the same time, the optimal superstring corresponds to the permutation $(s_1,s_2,s_3)$ and has length $2n+8$.
The algorithm GHA makes a~similar mistake on this dataset, see Figure~\ref{fig:tough}. When processing the node $({\tt ea})^n$, GHA does not add two lower arcs to it and misses a~chance to connect two components. It is then forced to connect these two components through~$\varepsilon$. This example shows that GHA also does not give a better than $2$-approximation for SCS.
\newcommand{\ged}[2]{
\tikzstyle{t}=[vertex,draw=black]
\tikzstyle{p}=[->,decorate,decoration=snake,bend left=10]
\begin{scope}[yshift=#1mm]
\node[inputvertex] (ccaen) at (0,5) {{\tt cc}({\tt ae})$^n$};
\node[inputvertex] (eann) at (10,5) {({\tt ea})$^{n+1}$};
\node[inputvertex] (aencc) at (4.5,5) {({\tt ae})$^n${\tt cc}};
\node[t] (ccaena) at (-1,4) {\tt cc(ae)$^{n-1}$a};
\node[t] (caen) at (1.5,4) {\tt c(ae)$^{n}$};
\node[t] (eane) at (9,4) {\tt (ea)$^{n}$e};
\node[t] (aean) at (11,4) {\tt a(ea)$^{n}$};
\node[t] (aenc) at (3.5,4) {\tt (ae)$^{n}$c};
\node[t] (eaencc) at (6,4) {\tt e(ae)$^{n-1}$cc};
\node[t] (ccaenn) at (1,3) {\tt cc(ae)$^{n-1}$};
\node[t] (aen) at (3.5,3) {\tt (ae)$^{n}$};
\node[t] (ean) at (10,3) {\tt (ea)$^{n}$};
\node[t] (aenncc) at (6,3) {\tt (ae)$^{n-1}$cc};
\node[t] (eaenn) at (3.5,2) {\tt e(ae)$^{n-1}$};
\node[t] (aeann) at (7,2) {\tt a(ea)$^{n-1}$};
\node[t] (eps) at (5,0) {$\varepsilon$};
\foreach \f/\t in {ccaena/ccaen, ccaen/caen, eane/eann, eann/aean, aenc/aencc, aencc/eaencc, ccaenn/ccaena, caen/aen, ean/eane, aean/ean, aen/aenc, eaencc/aenncc, ean/aeann, eaenn/ean}
\draw[->] (\f) -- (\t);
\path (eps) edge[p] (ccaenn);
\path (aenncc) edge[p] (eps);
#2
\end{scope}
}
\begin{figure}[!ht]
\begin{mypic}
\ged{0}{
\foreach \f/\t in {aeann/aen, aen/eaenn}
\draw[->] (\f) -- (\t);
}
\ged{-60}{
\path (eps) edge[p] (eaenn);
\path (aeann) edge[p] (eps);
}
\end{mypic}
\caption{Top: optimal solution for the dataset $\{ {\tt cc}({\tt ae})^n, ({\tt ea})^{n+1}, ({\tt ae})^n{\tt cc} \}$. Bottom: solution constructed by GHA.}
\label{fig:tough}
\end{figure}
\section{Introduction}
\label{sec:intro}
The {\em shortest common superstring problem} (abbreviated as SCS) is:
given a~set of strings, find a~shortest string that contains all of them as
substrings. This problem finds applications in genome assembly~\cite{waterman1995introduction, pevzner2001eulerian}, and data compression~\cite{GMS1980, phdthesis, storer1987data}. We refer the reader to the excellent surveys~\cite{gevezes2014shortest, mucha2007tutorial} for an overview of SCS, its applications and algorithms. SCS is known to be $\mathbf{NP}$-hard~\cite{GMS1980} and even $\mathbf{MAX}$-$\mathbf{SNP}$-hard~\cite{BJLTY1991}, but it admits constant-factor approximation in polynomial time.
The best known approximation ratios are $2\frac{11}{23}$ due to Mucha~\cite{M2013}
and $2\frac{11}{30}$ due to Paluch~\cite{P14}
(see \cite[Section~2.1]{GKM13} for an overview of the
previous approximation algorithms
and inapproximability results). While these approximation algorithms use an algorithm for Maximum Weight Perfect Matching as a~subroutine, the $30$ years old \emph{Greedy Conjecture}~\cite{storer1987data, TU1988, T1989, BJLTY1991} claims that the trivial \emph{Greedy Algorithm}, whose pseudocode is given in Algorithm~\ref{algo:ga}, is 2-approximate. Ukkonen~\cite{ukkonen1990linear} shows that for a fixed alphabet, the Greedy Algorithm can be implemented in linear time. It should be noted that GA is not deterministic as we do not specify how to break ties in case when there are many pairs of strings with maximum overlap. For this reason, GA may produce different superstrings for the same input.
\begin{algorithm}[ht]
\caption{Greedy Algorithm (GA)}\label{algo:ga}
\hspace*{\algorithmicindent} \textbf{Input:} set of strings~${\cal S}$.\\
\hspace*{\algorithmicindent} \textbf{Output:} a~superstring~for $\mathcal{S}$.
\begin{algorithmic}[1]
\While{$\mathcal{S}$ contains at least two strings}
\State extract from $\mathcal{S}$ two strings with the maximum overlap
\State add to $\mathcal{S}$ the shortest superstring of these two strings
\EndWhile
\State return the only string from $\mathcal{S}$
\end{algorithmic}
\end{algorithm}
\newtheorem*{gc}{Greedy Conjecture}
\begin{gc}
For any set of strings~$\mathcal{S}$, $\operatorname{GA}(\mathcal{S})$ constructs a~superstring that is at most twice longer than an optimal one.
\end{gc}
Blum et al.~\cite{BJLTY1991} prove that the Greedy Algorithm returns a $4$-approximation of SCS, and Kaplan and Shafrir~\cite{KS2005} improve this bound to~$3.5$. A~slight modification of the Greedy Algorithm gives a $3$-approximation of SCS~\cite{BJLTY1991}, and other greedy algorithms are studied from theoretical~\cite{BJLTY1991,rivals2018superstrings} and practical perspectives~\cite{romero2004experimental, cazaux2018practical}.
It is known that the Greedy Conjecture holds for the case when all input strings have length at most $3$~\cite{TU1988, cazaux20143}, and it was recently shown to hold in the case of strings of length $4$~\cite{kulikov2015greedy}. Also, the Greedy Conjecture holds if the Greedy Algorithm happens to merge strings in a particular order~\cite{weinard2006greedy, laube2005conditional}. The Greedy Algorithm gives a $2$-approximation of a~different metric called compression~\cite{TU1988}. The compression is defined as the sum of the lengths of all input strings minus the length of a superstring
(hence, it is the number of symbols saved with respect to a~naive superstring resulting from concatenating the input strings).
Most of the approaches for approximating SCS are based on the
{\em overlap graph} or the equivalent \emph{suffix graph}. The suffix graph has input strings as nodes, and a pair of nodes
is joined by an~arc of weight equal to their suffix (see Section~\ref{sec:def_scs} for formal definitions of overlap and suffix).
SCS is equivalent to (the asymmetric version of) the Traveling Salesman Problem (TSP) in the suffix graph. While TSP cannot be approximated within any polynomial time computable function unless $\mathbf{P}=\mathbf{NP}$~\cite{SG1976}, its special case corresponding to SCS can be approximated within a constant factor.\footnote{We remark that SCS is also a special case of TSP for costs satisfying the triangle inequaliy. This case of TSP can be approximated within a constant factor~\cite{svensson2018constant}, but this factor is currently much worse than that for SCS.} We do not know the full characterization of the graphs in this special case, but we know some of their properties: Monge inequality~\cite{monge} and Triple inequality~\cite{weinard2006greedy}. These properties are provably not sufficient for proving Greedy Conjecture~\cite{weinard2006greedy, laube2005conditional}.
While the overlap and suffix graphs give a~convenient graph structure, our current knowledge of their properties is provably not sufficient for showing strong approximation factors. Thus, the known approximation algorithms (including the Greedy Algorithm) estimate the approximation ratio via the overlap graph, and also separately take into account some string properties not represented by the overlap graph. The goal of this work is to develop a simple combinatorial framework which captures all features of the input strings needed for proving approximation ratios of algorithms.
\subsection{Our contributions}
We continue the study of the so-called {\em hierarchical graph}
introduced by Golovnev et al.~\cite{scs_exact}. (See also~\cite{cr16} for a related notion of the superstring graph.) This graph is designed specifically
for the SCS problem, in some sense it generalizes de~Bruijn graph, and it contains more information about the input strings
than just all pairwise overlaps. Given an instance of SCS, the vertex set of the corresponding hierarchical graph is just the set of substrings of all the input strings. For a string $s$ and two symbols $\alpha, \beta$, the graph contains the arcs: $(s,s\alpha)$ and $(\beta s, s)$. Now, every superstring of the given set of string corresponds to an Eulerian walk in the hierarchical graph (which passes through the vertices corresponding to the input strings), and vice versa. (See Section~\ref{sec:def_hier} for formal definition and statements.)
\subsubsection{Collapsing Conjecture}
We define a simple normalization procedure of a walk in the hierarchical graph: replace the pair of arcs $(\alpha s, \alpha s \beta), (\alpha s \beta, s\beta)$ with the pair $(\alpha s, s ), (s , s\beta)$ as long as it does not violate connectivity of the walk. It is easy to see that such a normalization never increases the length of the corresponding solution of SCS.
First, we observe a surprising property of this normalization procedure: if one takes \emph{any} solution, doubles all of its arcs in the hierarchical graph, and then applies the normalization procedure, then the resulting set of arcs is always the same (i.e., it does not depend on the initial solution). Collapsing Conjecture makes this observation formal (see Section~\ref{sec:collapsing}). Note that this conjecture implies an extremely simple 2-approximate algorithm for Shortest Common Superstring: take any solution (for example, write down all input strings one after another), then double each arc in the hierarchical graph, and apply the simple normalization procedure. This procedure will result in some superstring $S$. On the other hand, if one started with an optimal solution, doubled each of its arcs, and normalized the result, then the resulting solution would have length at most twice the length of the optimal solution. By Collapsing Conjecture, this resulting superstring would also be $S$, which implies that $S$ is a $2$-approximation.
\subsubsection{Greedy Hierarchical Conjecture}
We also propose a simple and natural greedy algorithm for SCS in the hierarchical graph: start from the nodes corresponding to the input strings, and greedily build an Eulerian walk passing through all of them.
While this Greedy Hierarchical Algorithm (GHA) is as simple as the Greedy Algorithm (GA), it provably performs better in some cases. For example, there are two well-known polynomially solvable special cases of SCS: strings of length~$2$ and
a~spectrum of a~string. While GA does not always find optimal solutions in these cases, GHA solves them exactly (see Sections~\ref{sec:ghatwo} and~\ref{sec:ghaspectrum}).
Greedy Hierarchical Conjecture (see Section~\ref{sec:greedy_hier}) claims that the set of arcs produced by GHA exactly matches the set of arcs from the Collapsing Conjecture: whichever initial solution one takes, after doubling its arcs and normalization, the resulting set of arcs is exactly the solution found by GHA. Clearly, this conjecture implies Collapsing Conjecture. Perhaps surprisingly, we prove that the two conjectures are equivalent (see Section~\ref{sec:equiv}): if all doubled solutions after normalization result in the same set of arcs, then this set of arcs is the GHA solution.
The weak form of Greedy Hierarchical Conjecture claims that GHA is a 2-approximate algorithm for SCS.
We prove (see Section~\ref{sec:gr_im_wghc}) that GHA is an instantiation of GA with some tie-breaking rule. That is, there is an algorithm which always merges some pair of strings with the longest overlap and outputs the same solution as GHA. This result has two consequences. First, by the known results for GA, we immediately have that GHA finds a $3.5$-approximation for SCS. Second, this gives us that Greedy Conjecture implies Weak Greedy Hierarchical Conjecture.
\subsubsection{Evidence for the Conjectures}
We support the Collapsing Conjecture (and the equivalent Greedy Hierarchical Conjecture) by proving its special case and verifying it empirically. We prove the conjecture for the special case where all input strings have length at most $3$, which until recently had been the only case where the Greedy Conjecture was proven (see Section~\ref{subsec:scs3}). Despite tesing the conjecture on millions of datasets (both hand-crafted and generated randomly according to various distributions), we have not found a counter-example. Note that even the Weak Greedy Hierarchical Conjecture suffices for getting a $2$-approximation for SCS, and this conjecture is not harder to prove than the standard Greedy Conjecture.
We implemented the Greedy Hierarchical Algorithm~\cite{github}, and we invite the reader to its~web interface~\cite{webpage} to see step by step executions of the described algorithms and to verify the conjectures on
custom datasets.
\section{Relations between the Conjectures}
In this section we prove some relations between the Collapsing and Greedy conjectures. Namely, in Section~\ref{sec:equiv} we prove the equivalence of Collapsing and Greedy Hierarchical conjectures. In Section~\ref{sec:gr_im_wghc} we prove that the standard Greedy Conjecture implies Weak Hierarchical Greedy Conjecutre (which is sufficient for a simple 2-approximate greedy algorithm for SCS). Finally, it is easy to see that Greedy Hierarchical Conjecture implies its weak version: indeed, if every doubled solution results in the solution obtained by GHA, then GHA does not exceed twice the optimal superstring length. In Figure~\ref{fig:relations}, we show the proven relations between the conjectures, together with 2-approximate algorithm which follow from each of the conjecture.
\begin{figure}
\begin{mypic}
\tikzstyle{r}=[rectangle,inner sep=2mm,draw,text centered, text width=40mm]
\node[r] (w) at (5.5,3) {Weak Greedy Hierarchical Conjecture};
\node[r] (cc) at (0,5) {Collapsing Conjecture};
\node[r] (gc) at (0,1) {Greedy Conjecture};
\node[r] (ghc) at (0,3) {Greedy Hierarchical Conjecture};
\node[r] (ca) at (12,5) {CA is 2-approximate};
\node[r] (gha) at (12,3) {GHA is 2-approximate};
\node[r] (ga) at (12,1) {GA is 2-approximate};
\foreach \f/\t in {cc/ghc, cc.east/w, ghc/w, gc.east/w, ghc/cc}
\draw[->] (\f) -- (\t);
\foreach \f/\t in {cc/ca, w/gha, gc/ga}
\draw[dashed,->] (\f) -- (\t);
\end{mypic}
\caption{Relations between the conjectures (left), and the 2-approximate algorithms they imply (right).
Collapsing and Greedy Hierarchical Conjectures are equivalent. They imply the weak version of the Greedy Hierarchical Conjecture, which also follows from the standard Greedy Conjecture. Each conjecture implies that the corresponding algorithm finds a $2$-approximate solution for SCS.}
\label{fig:relations}
\end{figure}
\subsection{Equivalence of Collapsing and Greedy Hierarchical Conjectures}
\label{sec:equiv}
In this section we prove the equivalence of Collapsing Conjecture and Greedy Hierarchical Conjecture. Recall that Collapsing Conjecture claims that for any pair of Eulerian solutions $D_1$ and $D_2$ for the input strings ${\cal S}$, we have
\[CA({\cal S}, D_1 \sqcup D_1) = CA({\cal S}, D_2 \sqcup D_2)\, .\]
The Greedy Hierarchical Solution extends this statement to
\[CA({\cal S}, D_1 \sqcup D_1) = GHA(\mathcal{S}) \, .\] Greedy Hierarchical Conjecture trivially implies Collapsing conjecture, and in order to prove their equivalence, it suffices to show that the collapsing procedure applied to the doubled GHA solution results in the GHA solution:
\[CA({\cal S}, GHA(\mathcal{S}) \sqcup GHA(\mathcal{S})) = GHA(\mathcal{S}) \, .\]
\begin{theorem}
For any set of strings $\mathcal{S}$,
\[
CA(\mathcal{S}, GHA(\mathcal{S}) \sqcup GHA(\mathcal{S})) = GHA(\mathcal{S}) \, .
\]
\end{theorem}
\begin{proof}
Let us denote two copies of the $GHA(\mathcal{S})$ solution by $B$ and $R$, which stand for a blue-copy and a red-copy. We will prove the theorem statement by showing that $CA(R \sqcup B)$ collapses all arcs of $B$ and keeps untouched the arcs of $R$, as this implies that $CA(R \sqcup B)=R=GHA(\mathcal{S})$. For this, without loss of generality assume that the Collapsing Algorithm collapses blue arcs first, that is, if for a vertex $v$, $CA$ can collapse a blue pair of arcs $(\pref(v), v), (v, \suff(v))$, it does so. Recall that $CA$ processes vertices in the descending order of levels (Algorithm~\ref{alg:ca}, line~\ref{alg:ca_for}). We will prove that before processing level $l$, all the arcs above it (i.e., the arcs with at least one vertex at a level $>l$) do satisfy the desired property: all blue arcs are collapsed, and all red arcs are untouched.
The base case trivially holds for $l := \max\{|s| \, : \, s\in\mathcal{S}\}$, since the set of arcs above the level $l$ is empty. Assume the claim is true for the level $k > 0$, and let us prove the claim for the level $k - 1$. Note that regardless of the number of collapse operations applied to $B$, $B$ remains a set of walks: indeed, the collapse procedure keeps the balance of incoming and outgoing arcs for each vertex. By the induction hypothesis all the blue arcs above the level $k$ are collapsed, so we have that if for a vertex $v$ at level $k$ there is an arc $(\pref(v), v)$ in $B$, then there is also an arc $(v, \suff(v))$ in $B$, and vice versa. Recall that $CA$ collapses blue arcs when possible, and since every vertex has the same number of blue incoming and outgoing arcs, all pairs collapsed at the level $k$ are monotone.
Now let us show that no red pair can be collapsed. Indeed, if for some vertex $v$ at level $k$ there is a red pair $(\pref(v), v)$, $(v, \suff(v))$, then by construction of $GHA$ $v$ is either in $\mathcal{S}$ or is the last chance of the corresponding component $\mathcal{C} \ni v$ to be connected to the remaining arcs in $R$ (note that the first case is a subcase of the second one, as then $\mathcal{C}$ contains only one vertex). It follows that if $CA$ collapses such a pair of arcs, then $v$ has no blue arcs (as they have been collapsed before the red arcs), and all other vertices in the component $\mathcal{C}$ at the level $k$ collapsed all arcs (since $v$ is the last vertex in $\mathcal{C}$ in lexicographic order). Therefore, this pair is also the last chance of $\mathcal{C}$ to be connected to the rest of the arcs in $R$, thus, $CA$ cannot collapse it.
It remains to show that all blue pairs at level $k$ are collapsed. This trivially holds because no red pair is collapsed, and, thus, the connectivity of $R \sqcup B$ is maintained by $R$. This finishes the proof.
\end{proof}
\subsection{Greedy Implies Greedy Hierarchical}
\label{sec:gr_im_wghc}
Consider a~permutation of the input strings. We say that it is a~{\em valid greedy permutation} if it can be constructed by the Greedy Algorithm: there exist $n-1$ merges of the $n$~input strings that lead to this permutation such that at every step the two merged strings have the largest overlap. We will prove that GHA always returns a solution which corresponds to a greedy permutation of the input string. That is, while the standard Greedy Algorithm does not determine how to break ties, the Greedy Hierarchical Algorithm is a specific instantiation of the Greedy Algorithm with some tie-breaking rule.
We will use the following simple property of solutions constructed by the GHA algorithm.
\begin{claim}
\label{claim:zigzag}
Let $D$ be an Eulerian solution constrtucted by GHA. Then $D$ has a ``zig-zag'' form as in~\eqref{eq:zigzag}.
\end{claim}
\begin{proof}
First we prove that $D$ is normalized, that is, any application of the collapsing procedure of Algorithm~\ref{alg:collapse} to $D$ will violate the property of Eulerian solution. Indeed, Algorithm~\ref{alg:collapse} can only collapse pairs of arcs of the form $(\pref(s), s), (s, \suff(s))$. The Greedy Hierarchical Algorithm adds such pairs to its solution in two cases: (i) $s$ is an input string (line~\ref{alg:gha_init} of Algorithm~\ref{algo:gha}); (ii) $s$ is the the lexicographically largest among the shortest strings in its Eulerian component (line~\ref{alg:last} of Algorithm~\ref{algo:gha}). Now note that in the former case, the collapsing procedure applied to $s$ would violate the property that $D$ must contain all input strings, and in the latter case, the collapsing procedure would violate the connectivity property of $D$.
We finish the proof by showing that every normalized solution is of the form~\eqref{eq:zigzag}. Let $\pi=(s_1, \dots, s_n)$ be the permutation of the input strings corresponding to a normalized Eulerian solution~$D$. Let us follow the arcs of~$D$ in the order of the permutation $\pi$, and let $P$ be the set of arcs between the input strings $s_i$ and $s_{i+1}$. We will prove that $P$ is the union of the sets of arcs of the paths $s_i \to \overlap(s_i,s_{i+1})$ and $\overlap(s_i,s_{i+1})\to s_{i+1}$. If $P$ contains a pair of consecutive up- and down-arcs, that is, there exists a pair of arcs $(\pref(s), s), (s, \suff(s))$ in $P$, then this pair would have been collapsed by Algorithm~\ref{alg:collapse}, line~\ref{alg:col}. Therefore, the path $P$ consists of a number of down-arcs followed by a number of up-arcs. It remains to show that the number of down-arcs in $P$ is $d=|s_i|-|\overlap(s_i,s_{i+1})|$. Note that by the definition of $\overlap(\cdot,\cdot)$, the number of down-arcs in~$P$ is at least~$d$. On the other hand, if the number of down-arcs in~$P$ is strictly greater than~$d$, then both the down-path and up-path in~$P$ contain the vertex $\overlap(s_i, s_{i+1})$. This implies that the pair of arcs $(\pref(s), s), (s, \suff(s))$ for $s=\overlap(s_i, s_{i+1})$ would have been collapsed by Algorithm~\ref{alg:collapse}, line~\ref{alg:col}, as it does not violate the connectivity of the solution. Therefore, the number of down-arcs in $P$ is exactly $d$, which implies that $P$ is the path $s_i \to \overlap(s_i,s_{i+1})$ followed by the path $\overlap(s_i,s_{i+1})\to s_{i+1}$.
\end{proof}
\begin{theorem}
\label{thm:gr_im_wghc}
Every permutation $\pi=(s_1, \dots, s_n)$ of the input strings constructed by GHA is a~valid greedy permutation.
\end{theorem}
\begin{proof}
Consider the following algorithm~$A$: it starts with the sequence $(s_1, \dots, s_n)$ obtained by GHA, and at every step it merges two neighboring strings in this sequence that have the largest overlap. It is a~greedy algorithm, but instead of considering all pairwise overlaps, it only considers overlaps of neighboring strings in the sequence. Of course, in the end, this algorithm constructs exactly the permutation~$\pi$. To show that $\pi$ is a~valid greedy permutation, we show that at every iteration of~$A$ no two strings have longer overlap than the two strings merged by~$A$.
Consider, for the sake of contradiction, the first iteration when the algorithm $A$~merges some pair of neighboring strings with overlap of length~$k$ whereas there are non-neighboring strings~$p$ and~$q$ with $v=\overlap(p,q)$, $|v|>k$.
At this point, $p$~is a~merger of input strings $s_a, s_{a+1}, \dotsc, s_b$
and $q$~is a~merger of input strings $s_c, s_{c+1}, \dotsc, s_d$.
Then, from the assumption that no input string contains another input string, we have that $v=\overlap(p,q)=\overlap(s_b,s_c)$. Since the algorithm~$A$
merges neighboring strings in the decreasing order of overlap lengths, we have that $|\overlap(s_b,s_{b+1})| \le k <|v|$ and $|\overlap(s_{c-1},c_c)| \le k < |v|$.\footnote{In the case when $s_b$ is the last string in the solution (or $s_c$ is the first string in the solution) we think of it being followed by $\varepsilon$, and $|\overlap(s_b,\varepsilon)|=0<|v|$ still holds.}
Now we consider the Eulerian solution $D$ constructed by GHA in the hierarchical graph.
By Claim~\ref{claim:zigzag}, $D$ has a ``zig-zag'' form, thus, it contains all arcs from the path $s_b\to\overlap(s_{b}, s_{b+1}) \to s_{b+1}$, and all arcs from the path $s_{c-1}\to\overlap(s_{c-1}, s_{c}) \to s_{c}$. Recall that $v=\overlap(s_b,s_c)$, and that $|\overlap(s_b,s_{b+1})| <|v|$ and $|\overlap(s_{c-1},s_c)|< |v|$. In particular, the paths $s_b\to\overlap(s_{b}, s_{b+1})$ and $\overlap(s_{c-1}, s_{c}) \to s_{c}$ pass through the vertex $v$, which implies that the vertex $v$ in the solution $D$ has at least one incoming arc from the previous level and at least one outgoing arc to the previous level (see~Figure~\ref{fig:gagha}(a)). Such a~pair of arcs in the Eulerian solution $D$ constructed by GHA may only occur when~$v$ is the last chance of
its connected component to be connected to the rest of the solution (see line~\ref{alg:last} of Algorithm~\ref{algo:gha}). This, in turn, implies that right before the pair of arcs $(\pref(v), v)$ and $(v, \suff(v))$ was added to the Eulerian solution, there was an~Eulerian component where $v$~was the lexicographically largest among all shortest nodes. This component is shown schematically in~Figure~\ref{fig:gagha}(b). All overlap-nodes (the nodes which are equal to $\overlap(s_i,s_{i+1})$) of this component lie on levels~$\geq k$. Note that the pair $(\pref(v), v)$ and $(v, \suff(v))$ is added to the solution by GHA exactly once (line~\ref{alg:last} of Algorithm~\ref{algo:gha}). Therefore, any path following the arcs of $D$, after going through the arc $(\pref(v), v)$ must traverse the overlying component containing $s_b$ and $s_c$ (as otherwise the path could not reach the overlying component). In turn, this implies that after considering all overlaps of length $|v|>k$, $s_b$ and $s_c$ are already merged into one string, so they cannot be merged at this stage.
\begin{figure}
\begin{mypic}
\node[inputvertex] (b) at (0,6) {$s_b$};
\node[inputvertex] (c) at (4,6) {$s_c$};
\node[vertex] (v) at (2,3) {$v$};
\node[vertex] (pv) at (1,2) {};
\node[vertex] (sv) at (3,2) {};
\draw[->] (pv) -- (v);
\draw[->] (v) -- (sv);
\draw[->,anypath] (b) -- (v);
\draw[->,anypath] (v) -- (c);
\node at (2,1) {(a)};
\begin{scope}[xshift=100mm]
\node[inputvertex] (b) at (0,6) {$s_b$};
\node[inputvertex] (c) at (4,6) {$s_c$};
\node[inputvertex] (d) at (2,6) {};
\node[inputvertex] (e) at (-2,6) {};
\node[vertex] (f) at (4,4) {};
\node[vertex] (g) at (0,3) {};
\node[vertex] (h) at (-0.5,5) {};
\node[vertex] (v) at (2,3) {$v$};
\node[vertex] (pv) at (1,2) {};
\node[vertex] (sv) at (3,2) {};
\draw[->] (pv) -- (v);
\draw[->] (v) -- (sv);
\foreach \f/\t in {b/v, v/c, c/f, f/d, d/g, g/e, e/h, h/b}
\draw[->,anypath] (\f) -- (\t);
\node at (2,1) {(b)};
\end{scope}
\end{mypic}
\caption{(a)~In the Eulerian solution the node $v=\overlap(s_b,s_c)$ has a~pair of lower arcs. (b)~For this reason, above~$v$, there is an~Eulerian component.}
\label{fig:gagha}
\end{figure}
\end{proof}
Theorem~\ref{thm:gr_im_wghc} has two immediate corollaries.
\begin{corollary}
The Greedy Conjecture implies the Weak Greedy Hierarchical Conjecture: if the Greedy Algorithm is 2-approximate, then so is the Greedy Hierarchical Algorithm.
\end{corollary}
Since every valid greedy permutation is a $3.5$-approximation to the Shortest Common Superstring problem~\cite{KS2005}, we have the following corollary.
\begin{corollary}
GHA is a factor $3.5$ approximation algorithm for the Shortest Common Superstring problem.
\end{corollary} |
\section{Introduction}
\label{Introd}
Cataclysmic variables (CVs) are close interacting binaries generally
consisting of a low-mass main-sequence (MS) star transferring matter
to the white dwarf (WD) primary via Roche lobe overflow
\citep{Warner1995}. In non-magnetic or weakly magnetic systems, which
make up the majority of observed CVs, the hydrogen-rich material
leaving the secondary star from the inner Lagrangian point forms an
accretion disc around the white dwarf. It is estimated that some
$20-25$\,per cent of all CVs host a magnetic white dwarf
\citep[MWDs,][]{WickFer2000,fer2015a}. These systems are the magnetic
cataclysmic variables (MCVs). Among MCVs we have the strongly magnetic
AM Herculis variables or polars. In polars the high magnetic field of
the white dwarf can thread and channel the material from the secondary
star directly from the ballistic stream to form magnetically confined
accretion funnels, so preventing the formation of an accretion
disc. In these systems the two stars are locked in synchronous
rotation at the orbital period. The radiation from the accretion
funnels \citep[e.g.][]{Ferrario1999} and the cyclotron radiation from
the shocks located at the funnels' footpoints of closed magnetic field
lines dominate the emission of these systems from the X-rays to the
infrared bands \citep[e.g.][]{MeggittWick1982,WickFer1988}. Cyclotron
and Zeeman spectroscopy and spectropolarimetry have revealed the
presence of strong fields in the range of a few $10^7-10^8$\,G
\citep[see, e.g.,][]{Ferrario1992,Ferrario1993,Ferrario1996}. Weaker fields of about $10^6$ to
$3\times 10^7$\,G are found in the DQ\,Herculis variables or
Intermediate Polars (IPs) where the white dwarf's magnetic field
cannot totally prevent the formation of an accretion disc
\citep[e.g. see][]{FWK1993}. In these systems the material is
magnetically threaded from the inner regions of a truncated accretion
disc and channelled on to the primary forming magnetically confined
accretion curtains \citep{FerrarioWick1993}. In the IPs the white
dwarf is not synchronously locked with the orbital period but is spun
up to a spin period shorter than the orbital period of the system.
\citet{Liebert2005} noticed the enigmatic lack of MWDs from the 501
detached binaries consisting of a white dwarf with a non-degenerate
companion found in the DR1 of the Sloan Digital Sky Survey
\citep[SDSS,][]{York2000}. They also noticed that among the 169 MWDs
known at the time, none had a non-degenerate detached companion. This
was puzzling because such a pairing is very common among non-magnetic
white dwarfs \citep[see, e.g.][]{Hurley2002,Ferrario2012}. A similar study
conducted on the much larger DR7 sample of SDSS detached binaries
consisting of a white dwarf with a non-degenerate companion
\citep{Kleinman2013} led to the same conclusion
\citep{Liebert2015}. Over the years, not a single survey conducted to
ascertain the incidence of magnetism among white dwarfs has yielded a
system consisting of a magnetic white dwarf with a non-degenerate
companion \citep[e.g.,][]{Schmidt2001,Kawka2007}. It is this curious
lack of pairing that led \citet{Tout2008} to propose that the
existence of magnetic fields in white dwarfs is intimately connected
to the duplicity of their progenitors and that they are the result of
stellar interaction during common envelope evolution. In this
picture, as the cores of the two stars approach each other, their
orbital period decreases and the differential rotation that takes
place in the convective common envelope generates a dynamo mechanism
driven by various instabilities. \citet{Regos1995} argued that it is
this dynamo mechanism that is responsible for the transfer of energy
and angular momentum from the orbit to the envelope which is
eventually, all or in part, ejected.
\citet{WTF2014} have shown that strong magnetic fields in white dwarfs
can be generated through an $\alpha-\Omega$ dynamo during common
envelope evolution where a weak seed poloidal field is wound up by
differential rotation to create a strong toroidal field. However
toroidal and poloidal fields are unstable on their own
\citep{Braithwaite2009}. Once the toroidal field reaches its maximum
strength and differential rotation subsides the decay of toroidal
field leads to the generation of a poloidal field with the two
components stabilising each other and limiting field growth until they
reach a final stable configuration. Thus, a poloidal seed field can be
magnified during common envelope evolution by an amount that depends
on the initial differential rotation but is independent of its initial
strength. According to this scenario the closer the cores of the two
stars are dragged at the end of common envelope evolution, before the
envelope is ejected, the greater the differential rotation and thus
the stronger the expected frozen-in magnetic field. If common envelope
evolution leads to the merging of the cores the result is an isolated highly
magnetic white dwarf. If the two stars do not coalesce they
emerge from the common envelope as a close binary that evolves into a
MCV. The viability of such model, in terms of incidence of magnetism
among single white dwarfs and their mass and magnetic field
distribution, have been shown by \citet{Briggs2015} and
\citet{Briggs2018}, henceforth referred to as paper\,I and paper\,II
respectively.
In this paper we continue our studies of the origin of fields in MWDs
to explain the parentage of MCVs. To this end we carry out a
comprehensive population synthesis study of binaries for different
common envelope efficiencies $\alpha$. We examine all paths that lead
to a system consisting of a white dwarf with a low-mass companion
star. We show that the observed properties of the MCVs are generally
consistent with their fields originating through common envelope
evolution for $\alpha<0.4$.
\section{Evolution and space density of MCVs}
Observed MCVs consist of a white dwarf that accretes matter from a
secondary star that has not gone through any significant nuclear
evolution when the transfer of mass begins. The mass ratio of an MCV
is given by $q=M_{\rm sec}/M_{\rm WD}<1$ where $M_{\rm WD}$ is the
mass of the white dwarf primary and $M_{\rm sec}$ is the mass of the
companion star. The mass accretion process in MCVs is relatively
stable over long periods of time, although the polars suffer from high
and low states of accretion. It is not known what sparks the change in
accretion mode but, because polars do not have a reservoir of matter
in an accretion disc, they can switch very quickly from high to low
states. IPs have never been observed in low states of accretion.
Stable mass transfer can be driven by nuclear-timescale expansion of
the secondary (not generally applicable to MCVs) and/or by loss of
angular momentum, driven by magnetic braking above the CV period gap
\citep[caused by the disrupted magnetic braking mechanism,
see][]{Spruit1983, Rappaport1983, Verbunt1984} and gravitational
radiation below the gap. Loss of angular momentum shrinks the orbit
keeping the companion star filling its Roche lobe and so drives mass
transfer. Therefore, as MCVs evolve, the orbital period diminishes
until it reaches a minimum when the secondary star becomes a
substellar-type object whose radius increases as further mass is
lost. Systems that have reached the minimum period and have turned
back to evolve toward longer periods are often called period bouncers
\citep[e.g.][]{Patterson1998}
The evolution of MCVs is expected to be similar to that of
non-magnetic CVs. However, \citet{LiWuWick1994} have shown that
angular momentum loss may not be as efficient in polars as it is in
non-magnetic or weakly magnetic CVs in bringing the two stars together
because the wind from the secondary star is trapped within the
magnetosphere of the white dwarf. This phenomenon slows down
the loss of angular momentum, reduces the mass transfer rate and leads to
longer evolutionary timescales. This mechanism provides a simple
explanation for the observed high incidence of magnetic systems among
CVs \citep{Araujo2005}.
\begin{figure}
\includegraphics[width=1.05\columnwidth]{CVperiods1.eps}
\caption{The orbital period distribution of MCVs (top) and CVs
(bottom). The MCVs are subdivided into Polars (solid black line histogram) and
IPs (shaded histogram). We have used the latest version (v7.20) of
Ritter \& Kolb's (2003) CV catalogue to create this figure.}
\label{fig:orbdistr}
\end{figure}
We show in Fig.\,\ref{fig:orbdistr} the period distribution of CVs and
MCVs where the MCVs have been subdivided into polars and intermediate
polars. The space density of CVs is not well known and, over the
years, there has been some considerable disagreement between
observations and theoretical predictions. The recent study of
\emph{Swift}\,X-ray spectra of an optically selected sample of nearby
CVs conducted by \citet{Reis2013} has unveiled a number of very low
emission X-ray systems. Hard X-ray surveys of the Galactic ridge have
shown that a substantial amount of diffuse emission can be
resolved into discrete low-luminosity sources. Because the MCVs are
generally strong X-ray emitters, \citet{Muno2004} and \citet{Hong2012}
have propounded that IPs could be the main components of these
low-luminosity hard X-ray sources.
\citet{Pretorius2013} have conducted a study of the X-ray flux-limited
\emph{ROSAT} Bright Survey (RBS) to determine the space density of
MCVs. They assume that the 30 MCVs in the RBS are representative of
the intrinsic population. They also allow for a 50 per cent high-state
duty cycle for polars under the assumption that polars are below the
RBS detection threshold while they are in low states of
accretion. They find that the total space density of MCVs is
$1.3^{+0.6}_{−0.4}\times 10^{−6}$ pc$^{-3}$ with about one IP per
200\,000 stars in the solar neighbourhood. They conclude that IPs are
indeed a possible explanation for most of the X-ray sources in the
Galactic Centre. These new findings seem to suggest that the space
density of CVs is somewhat larger than initially forecast and thus in
closer agreement with theoretical expectations.
\subsection{Where are the progenitors of the MCVs?}\label{WhereProg}
\citet{Liebert2005, Liebert2015} asked, ``Where are the magnetic white
dwarfs with detached, non-degenerate companions?'' This question is
awaiting an answer and thus the progenitors of the MCVs still need to
be identified. The proposal by \citet{Tout2008} that the existence of
high magnetic fields in isolated and binary white dwarfs is related to
their duplicity prior to common envelope evolution is gaining
momentum. Observational support for the binary origin of magnetic
fields in MCVs is also strengthening. \citet{Zorotovic2010} listed
about $60$ post common envelope binaries (PCEBs) from the SDSS and
other surveys consisting of a white dwarf with an M-dwarf
companion. The periods of these PCEBs range from about $0.08$ to
$20$\,d and the white dwarf effective temperatures in the range 7\,500
to 60\,000\,K. According to current binary evolution theory, one third
of these systems should lose angular momentum from their orbits by
magnetic braking and gravitational radiation and are expected to come
into contact within a Hubble time. However none of these 60 binaries
contains a MWD, even if observations indicate that 20 to 25\,per cent
of all CVs harbour one. This finding suggests that magnetic white
dwarf primaries are only present in those binaries that are already
interacting or are close to interaction. The magnetic systems
originally known as Low-Accretion Polars
\citep[LARPS,][]{Schwope2002,FWS2005} have been proposed to be the
progenitors of the polars. The first LARPS were discovered in the
Hamburg/ESO Quasar Survey \citep[HQS,][]{Wisotzki1991} and then by the
SDSS by virtue of their unusual colours arising from the presence of
strong cyclotron harmonic features in the optical band together with a
red excess owing to the presence of a low-mass red companion star. The
MWDs in LARPS are generally quite cool
($T_{\rm eff}\lsimeq10\,000$\,K) and have low-mass MS companions which
underfill their Roche lobes \citep[e.g.][]{Reimers1999, Schwope2002,
FWS2005, Parsons2013}. The MWDs in these systems accrete mass from
the wind of their companion at a rate substantially larger than in
detached non-magnetic PCEBs \citep{Parsons2013}. \citet{Tout2008}
suggested that LARPS could be pre-polars waiting for gravitational
radiation to bring the stars close enough to each other to allow Roche
lobe overflow to commence. These systems were renamed pre-polars
(PREPs) by \citet{Schwope2009} to avoid confusion with polars in a low
state of accretion. PREPs have orbital periods which are, on average,
only marginally longer than those of polars. The ages of the white
dwarfs in PREPs, as indicated by their effective temperatures, are
generally above a billion years. The absence of PREPs with hot white
dwarfs is puzzling but maybe still not alarming, if one considers the
small number of PREPs currently known and the initial rapid cooling of
white dwarfs. Thus, the hypothesis that the progenitors of MCVs are
expected to emerge from common envelope when they are close to
transferring mass via Roche Lobe overflow is well warranted. We show
in Fig.\,\ref{preps_fig} the period distribution of PCEBs and PREPs.
\begin{figure}
\includegraphics[width=1.03\columnwidth]{PCEB_obs_periods1.eps}
\caption{The orbital period distribution of PCEBs
\citep[solid black line histogram,][]{Nebot2011} and PREPs
\citep[shaded histogram,][]{fer2015a}.}
\label{preps_fig}
\end{figure}
\section{Population synthesis calculations}\label{popsynt}
\label{sec:calculations}
\begin{table*}
\caption{We have indicated with $N$ (second column) the fraction of PREPs for different efficiency parameters $\alpha$ (first column) in a single generation of binaries. The other columns give the smallest and the largest progenitor masses and initial orbital periods.}
\centering
\begin{tabular}{ c c c c r c c r}
\hline
\noalign{\smallskip}
$\alpha$ & $N$\,(per cent) & $M_{1_{\rm min}}/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ & $M_{2_{\rm
min}}/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ &$M_{1_{\rm max}}/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi
$&$M_{2_{\rm max}}/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi $ & $P_{0_{\rm min}}/$d & $P_{0_{\rm max}}/$d \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.10 & 1.518 & 1.08 & 0.10 & 8.16 & 1.42 & 369.7 & 3144.0 \\
0.15 & 1.672 & 1.08 & 0.10 & 8.16 & 1.42 & 293.3 & 2800.5 \\
0.20 & 1.663 & 1.08 & 0.10 & 8.16 & 1.42 & 246.6 & 2354.3 \\
0.25 & 1.213 & 1.08 & 0.10 & 8.16 & 1.36 & 207.3 & 2097.0 \\
0.30 & 1.163 & 1.08 & 0.10 & 8.16 & 1.14 & 184.6 & 2221.9 \\
0.50 & 0.808 & 1.08 & 0.10 & 8.16 & 0.58 & 123.2 & 2221.9 \\
0.70 & 0.804 & 1.08 & 0.10 & 8.16 & 0.19 & 87.0 & 1867.9 \\
0.80 & 0.859 & 1.08 & 0.10 & 8.16 & 0.13 & 69.1 & 1762.9 \\
\noalign{\smallskip}
\hline
\end{tabular}
\label{tab:statsCE}
\end{table*}
Each binary is assigned three initial parameters. These are the mass
$1.0\le M_1/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi\le 10.0$ of the primary star, the mass
$0.1\le M_2/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi\le 2.0$ of the secondary star, and the orbital
period $1\le P_0/{\rm d}\le 10\,000$ at the zero-age main sequence
(ZAMS). We set the eccentricity to zero. We sampled each parameter
uniformly on a logarithmic scale with 200\,divisions. This sampling
gives a synthetic population of about 70 million binary systems. The
actual number of binary systems is then calculated on the premise that
$M_1$ follows Salpeter's mass function distribution
\citep{Salpeter1955} and $M_2$ is according to a flat mass ratio
distribution with $q\le 1$. The initial period distribution is
assumed to be uniform in the logarithm.
We have used the rapid binary star evolution algorithm, {\sc bse},
developed by \citet{Hurley2002}, to evolve each binary system from the
ZAMS to $9.5\,$Gyr \citep[age for the Galactic
Disc,][]{Kilic2017}. {\sc bse} is an extension of the single star
evolution code written by \citet{Hurley2000}. It allows for stellar
mass loss and interaction between the two stars such as mass transfer, Roche
lobe overflow, common envelope evolution \citep{Paczynski1976}, tidal
interaction, supernova kicks, and angular momentum loss caused by
gravitational radiation and magnetic braking.
In our modelling we have used the $\alpha$ formalism for common envelope
evolution. If $\Delta E_{\rm orb}$ is the change in orbital energy during
the in-spiral phase and $E_{\rm bind}$ is the energy required to eject the
envelope then
\[
\Delta E_{\rm orb}=\alpha E_{\rm bind}.
\]
\noindent Here $\alpha$ is the common envelope efficiency parameter,
ranging between 0.1 and 0.9, that takes into consideration the fact
that the removal of the envelope is not totally efficient (see
papers\,I and II for more details). The expression for the envelope
binding energy contains another parameter $\lambda$ which denotes the
structure, or central condensation, of the envelope of the donor. This
parameter is very uncertain and could vary by up to an order of
magnitude \citep{TaurisDewi2001, Ivanova2011}. It depends on both the
envelope structure and on the mass above which the envelope is
expelled. In this work we follow \citet{Hurley2002} and set
$\lambda=0.5$. A review on common envelope evolution and on the
parameters that are used in its modelling can be found in
\citet{Izzard2012}.
Single star mass loss rates are described by \citet{Hurley2000}. In
our calculations we have adopted $\eta=1.0$ for the Reimers' mass-loss
parameter, as outlined in paper\,I, and a stellar metallicity
$Z = 0.02$.
Our theoretical sample of PCEBs consists of systems that (i)\,have
undergone common envelope evolution, (ii)\,have a primary that evolves
into a white dwarf, (iii)\,have a companion that remains largely
unevolved and (iv)\, have a mass ratio $q\le 1$. A subset of these
systems come into contact over the age of the Galactic Disc and become
classical CVs. Those systems with a white dwarf that develops a strong
magnetic field become MCVs.
Of our sample of PCEBs, we then select the subset consisting of the
PREPs (the progenitors of the MCVs). PREPs must fulfil two additional
criteria: (i) the primary star must have a degenerate core before
entering the last common envelope phase and (ii) no further core
burning occurs. The reason for the first criterion is that a
degenerate core is essential for a stellar magnetic field to persist,
in a frozen-in state, after its formation. The reason for the second
is that nuclear burning in the core would ignite convection that would
destroy any frozen-in magnetic field. Systems that violate either
criterion but come into contact over the age of the Galactic Disc are
expected to evolve into classical non-magnetic CVs. We show in
Table\,\ref{tab:statsCE} the limits of the parameter space within
which PREPs are formed. The minimum ZAMS masses of the systems that
give rise to PREPs are listed in the columns with headings
$M_{1_{\rm min}}$ and $M_{2_{\rm min}}$ and the maximum masses are
under the headings $M_{1_{\rm max}}$, $M_{2_{\rm max}}$. Minimum and
maximum initial periods are in the columns under $P_{0_{\rm min}}$ and
$P_{0_{\rm max}}$ respectively.
Once we have obtained our theoretical PREP sample, we assign a
magnetic field $B$ to each of their white dwarf primaries following
the prescription described in paper\,II to model the field
distribution of high field magnetic white dwarfs (HFMWDs). That is
\begin{equation}\label{EqBfield}
B = B_0\left(\frac{\Omega}{\Omega_{\rm crit}}\right)\, \mbox{G}.
\end{equation}
where $\Omega$ is the orbital angular velocity and
$\Omega_{\rm crit}=\displaystyle{\sqrt{GM_{\rm WD}/R_{\rm WD}^3}}$ is
the break-up angular velocity of a white dwarf of mass $M_{\rm WD}$
and radius $R_{\rm WD}$. The parameter $B_0$ is a free parameter that
was determined empirically in paper\,II, that is,
$B_0=1.35\times 10^{10}$\,G. The parameter $B_0$ does not influence
the shape of the field distribution which is only determined by
$\alpha$. Lower (or higher) $B_0$ shift the field distribution to
lower (or higher) field strengths. Unlike HFMWDs, both stars emerge
from common envelope evolution but on a much tighter orbit that
allows them to come into contact within a Hubble time and appear as MCVs.
Thus, white dwarfs in interacting binaries can only attain a fraction
of the upper field strength of single white dwarfs and this is the
reason why $B_0$ must be determined through the modelling of HFMWDs
(see paper\,II). Field strengths of MCVs are scaled down from the
maximum by equation (\ref{EqBfield}).
\section{Synthetic population statistics}\label{SyntPop}
We have time integrated each population, characterised by
$\alpha$, to the Galactic Disc age under the assumption that
the star formation rate is constant. We have listed in
Table\,\ref{tab:PCEBS_PREPS_Percent} the percentage by type of all
binaries that emerge from common envelope over the age of the Galactic
Disc.
\begin{table}
\caption{The number of PCEBs born, the fraction of PREPs from PCEBs
and of MCVs (magnetic systems already exchanging mass) from PREP as a
function of the common envelope efficiency parameter $\alpha$ over the
age of the Galactic Disc. The number of PREPs is maximum close to
$\alpha = 0.15$ while the number of MCVs is maximum at $\alpha= 0.10$.}
\centering
\begin{tabular} {l l l l}
\hline
\noalign{\smallskip}
$\alpha$ & Number of & \underline{PREPs} $\times100$& \underline{MCV~~} $\times100$ \\
& PCEBs~~~ &PCEBs~~ & PREPS~ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.10 & 30517472 &20.9 & 61.0 \\
0.15 & 36099023 &18.9 & 56.4 \\
0.20 & 38666876 &15.3 & 49.9 \\
0.30 & 41197674 & 8.7 & 45.0 \\
0.40 & 43654871 & 5.6 & 48.0 \\
0.50 & 46289395 & 4.5 & 51.0 \\
0.60 & 49010809 & 4.1 & 52.0 \\
0.70 & 51888317 & 3.8 & 52.4 \\
0.80 & 54664759 & 3.3 & 52.4 \\
\hline
\end{tabular}
\label{tab:PCEBS_PREPS_Percent}
\end{table}
Column 2 in Table\,\ref{tab:PCEBS_PREPS_Percent} shows that while the
number of PCEBs increases when $\alpha$ increases, the percentage of
PREPs (progenitors of the MCVs) decreases. This is because as $\alpha$
increases the envelope's clearance efficiency increases causing the
two stars to emerge from common envelope at wider separations and thus
less likely to become PREPs and thence MCVs. On the other hand, the
overall number of PCEBs increases because stellar merging events
become rarer at high $\alpha$, as shown in paper\,I. Fewer merging
events are also responsible for the high incidence of systems with
low mass He\,white dwarfs (He\,WDs) whose ZAMS progenitors were born at short orbital
periods and entered common envelope evolution when the primary star
became a Hertzsprung gap or a red giant branch (RGB) star. At larger
initial orbital periods common envelope evolution may occur on the
asymptotic giant branch (AGB). However as $\alpha$ increases only
stars in those systems that harbour massive enough white dwarfs can
come sufficiently close to each other to allow stable mass transfer to
occur within the age of the Galactic Disc (see section
\ref{mass_distribution}). In contrast, at low $\alpha$ the
clearance efficiency is low and so there is a longer time for the
envelope to exert a drag force on the orbit. This results in (i)
more merging events, (ii) tighter final orbits for all white dwarf
masses and (iii) a larger number of systems coming into contact over
the age of the Galactic Disc. Point (i) reduces the overall number of
PCEBs while (ii) and (iii) increase the number of PREPs.
\begin{table*}
\caption{Evolutionary history of an example binary system that
becomes a MCV after common envelope evolution with $\alpha=0.1$.
Here CE\,=\,Common Envelope, RLO\,=\,Roche Lobe Overflow. }
\centering
\begin{tabular}{ r r r r r r c l }
\hline
\noalign{\smallskip}
Stage & Time/Myr & $M_1/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ & $M_2/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ & $P/$\,d & $a/{\rm R}_\odot$ &$B/{\rm G}$& Remarks \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
1 & 0.000 & 4.577 & 0.230 & 2244.627 & 1218.030 & 0.000E+00 & ZAMS \\
2 & 128.515 & 4.577 & 0.230 & 2244.627 & 1218.030 & 0.000E+00 & S1 is a Hertzsprung gap star \\
3 & 129.078 & 4.577 & 0.230 & 2245.210 & 1218.188 & 0.000E+00 & S1 is a RGB star. Separation increases slightly. \\
4 & 129.445 & 4.574 & 0.230 & 2247.427 & 1218.790 & 0.000E+00 & S1 starts core He burning. Some mass loss occurs.\\
5 & 149.930 & 4.466 & 0.230 & 2352.896 & 1247.059 & 0.000E+00 & S1 is an AGB star. Further mass loss occurs. \\
6 & 150.947 & 4.390 & 0.230 & 2173.184 & 1176.321 & 0.000E+00 & S1 is a late AGB star. Separation decreases significantly \\
7 & 150.989 & 4.364 & 0.230 & 861.296 & 633.510 & 0.000E+00 & RLO \& CE start. Separation decreases dramatically. \\
8 & 150.989 & 0.918 & 0.230 & 0.117 & 1.053 & 1.218E+07 & S1 emerges from CE as a CO\,MWD and RLO ceases.\\
9 & 326.073 & 0.918 & 0.230 & 0.099 & 0.945 & 1.218E+07 & Separation decreases and MCV phase starts \\
10 & 9\,500.000 & 0.918 & 0.037 & 0.139 & 1.112 & 1.218E+07 & Separation reaches a minimum between stages 9 and 10 \\
& & & & & & & and increases again. S2 is a brown dwarf. \\
\hline
\end{tabular}
\label{tab:evolMCV1}
\end{table*}
\begin{table*}
\caption{Evolutionary history of a second example binary system that
becomes a MCV after common envelope with $\alpha=0.4$.}
\centering
\begin{tabular}{ r r r r r r c l }
\hline
\noalign{\smallskip}
Stage & Time/Myr & $M_1/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ & $M_2/\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ & $P$\,(d) & $a/{\rm R}_\odot$ &$B/{\rm G}$& Remarks \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
1 & 0.000 & 1.612 & 0.257 & 190.661 & 171.774 & 0.000E+00 & ZAMS \\
2 & 2197.329 & 1.612 & 0.257 & 190.661 & 171.774 & 0.000E+00 & S1 is a Hertzsprung gap star \\
3 & 2239.430 & 1.611 & 0.257 & 190.743 & 171.811 & 0.000E+00 & S1 is a RGB star, loses mass. Separation increases slightly. \\
4 & 2343.048 & 1.580 & 0.257 & 110.351 & 118.629 & 0.000E+00 & S1 loses more mass, separation decreases. \\
5 & 2343.048 & 0.386 & 0.257 & 0.149 & 1.020 & 3.577E+07 & RLO \& CE start. Separation decreases dramatically. \\
6 & 2343.048 & 0.386 & 0.257 & 0.149 & 1.020 & 3.577E+07 & S1 emerges from CE as a He\,MWD and RLO ceases.\\
7 & 3389.278 & 0.386 & 0.257 & 0.102 & 0.792 & 3.577E+07 & Separation decreases and MCV phase starts \\
8 & 9\,500.000 & 0.386 & 0.052 & 0.100 & 0.687 & 3.577E+07 & Separation reaches a minimum between stages 7 and 8 \\
& & & & & & & and increases again. S2 is a brown dwarf.\\
\noalign{\smallskip}
\hline
\end{tabular}
\label{tab:evolMCV2}
\end{table*}
\subsection{Magnetic CV evolution examples}
The evolutionary history of a binary system depends on the parameters
that characterise it. The number of common envelope events can vary from one to
several \citep{Hurley2002}. Whether a classical CV becomes magnetic
or not depends on the evolution before and after the common
envelope. Here we give two typical examples of systems that evolve
into a MCV. In the first the initially rather massive primary
star evolves into a CO\,white dwarf (CO\,WD) after common envelope evolution as
a late AGB star. In the second example the primary evolves into a
He\,white dwarf after common envelope evolution while ascending the RGB.
\emph{Example 1:} Table~\ref{tab:evolMCV1} illustrates the evolution
of a system that becomes a close binary after common envelope with
$\alpha=0.1$. The progenitors are a primary star (S1) of
$4.58\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ and a secondary star (S2) of sub-solar mass
$0.230\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$. At ZAMS the initial period is $2\,240\,$d with a
separation of $1\,220\,R _\odot$.
S1 evolves off the ZAMS and reaches the early AGB stage at
$149$\,Myr having lost $0.111$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ on the way. After a
further $1.02$\,Myr S1 has become a late AGB star. Further evolution
brings the stars closer together at a separation of
$634$\,R$_\odot$. Soon after dynamically unstable Roche lobe
overflow from S1 to S2 takes place and common envelope begins. At the
end of the short period of common envelope evolution the two stars
emerge with a separation of only $1.05$\,R$_\odot$ because of the
large orbital angular momentum loss during this stage. The ejection
of the envelope exposes the core of S1 that has now become a magnetic
$0.918$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ CO\,WD. After a further $175$\,Myr the
separation has further contracted to $0.945$\,R$_\odot$ via magnetic
braking and gravitational radiation. Roche lobe overflow begins and
the system becomes a bona fide mass-exchange MCV. During the MCV
evolutionary phase the mass of the donor star, separation and orbital
period steadily decrease until the mass of the companion star becomes
too low to maintain hydrogen burning and S2 becomes a degenerate
object. At this point separation and orbital period reach a
minimum. Further evolution sees these two quantities increase again
over time. At an age of $9\,500$\,Myr S2 has lost most of its mass and
has become a $0.037$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ brown dwarf with the separation from its white dwarf
primary increased to $0.112$\,R$_\odot$.
\emph{Example 2:} Table~\ref{tab:evolMCV2} shows the evolution of a
second system that becomes a close binary after common envelope. This
time we have $\alpha=0.4$. The progenitors are a MS
primary star (S1) of $1.61$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ and a secondary star (S2) of mass
$0.257$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$. At ZAMS the initial period is $191$\,d and the
separation $172$\,R$_\odot$.
S1 evolves off the ZAMS through the Hertzsprung gap to reach the RGB
after $2\,240$\,Myr having lost $0.001$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ on the way. Still on
the RGB at $2\,340$\,Myr S1 has lost $0.031$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ and the
separation has decreased to $119$\,R$_\odot$. Roche lobe overflow
from S1 to S2 and common envelope evolution begin. S1 emerges from
common envelope as a magnetic He\,WD with a mass of
$0.386$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$. The orbital separation has drastically decreased to
$1.02$\,R$_\odot$. S2 maintains its mass and remains an M-dwarf
star. From this time onwards magnetic braking and gravitational radiation
cause the orbit to shrink further until at $3\,390$\,Myr the
separation is $0.792$\,R$_\odot$ and Roche lobe overflow
commences. The system is now a MCV. Further evolution leads S2 to
lose mass, owing to accretion on to S1, until, at $9\,500$\,Myr, S2 has
become a brown dwarf of mass $0.052$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ and the separation is
$0.687$\,R$_\odot$.
\subsection{Property distributions of the synthetic population}
We create our population of putative PREPs by integration over time
from $t=0$ to $t=9.5\,$Gyr. The star formation rate is taken to be
constant over the age of the Galactic Disc. Whereas
Table\,\ref{tab:statsCE} shows the relative numbers of PREPs obtained
from a single generation of binaries, continuous star formation
over the age of the Galactic Disc builds up a population of PCEBs,
PREPs, CVs, and MCVs that, as birth time increases, favours systems with
progressively higher mass primaries because lower mass primaries,
especially in later generations, do not have enough time to evolve to
the white dwarf stage.
\begin{figure*}
\begin{center}
\includegraphics[width=1.01\textwidth]{RLOF-PeriodP-NEW.eps}
\caption{Theoretical period distribution of magnetic systems just
before they start RLOF for various $\alpha$'s.
The period distribution of the primary white dwarf types is shown as the
superimposed coloured categories. The total of the distribution is
shown as the pink background histogram peaking around 2.8
to 3.0\,hrs. This is to be compared with the observed distribution
for PREPs in Fig.\,\ref{preps_fig}}
\label{fig:RLOFPeriodP}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=1.00\textwidth]{RLOF-PeriodS-NEW.eps}
\caption{Same as Fig.\ref{fig:RLOFPeriodP} but with the secondary star
types shown as the superimposed coloured categories. Both secondary
star types are MS stars. The CS type is a deeply or fully
convective MS star with $M<0.7$\,M$_\odot$.}
\label{fig:RLOFPeriodS}
\end{center}
\end{figure*}
\subsubsection{Period distribution}\label{PerDist}
Figs\,\ref{fig:RLOFPeriodP} and \ref{fig:RLOFPeriodS} show the
theoretical period distribution of the PREPs just before the beginning
of Roche lobe overflow (RLOF) in a present day population formed over
the age of the Galactic Disc for various $\alpha$. The contributions
to the period distribution by white dwarf primaries of a certain
type are depicted in Fig.\,\ref{fig:RLOFPeriodP} while the
contributions to the period distribution by the secondaries of a given
type are displayed in Fig.\,\ref{fig:RLOFPeriodS}.
The period distribution peaks around $3\,$hr with a long tail
extending to about $10\,$hr for low $\alpha$. We note that at low
$\alpha$ our synthetic population tends to have orbital periods
clustering around the 1 to 4\,hr region while at higher $\alpha$ they
are confined to the 1 to 3\,hr region.
Fig.\,\ref{fig:RLOFPeriodP} shows that when $\alpha=0.1$ the main
contributors to the whole range of periods are systems with CO\,WD
primaries characterised by orbital periods from about 1 to 7\,hr and a
peak near 3\,hr. Systems with He\,WDs are also present but are fewer
and their periods are below 3\,hr. Massive Oxygen-Neon white dwarf
(ONe\,WD) primaries form a much smaller fraction of the distribution,
as expected from Salpeter's initial mass function, but make some
contribution to the full range of periods when $\alpha< 0.4$.
As $\alpha$ increases the fraction of CO\,WD systems decreases until
these all but disappear for $\alpha>0.5$ while the percentage of
He\,WDs increases dramatically. For $\alpha\ge 0.4$, the orbital
periods are always below 3\,hr and He\,WD systems well and truly
dominate the period distribution. For $\alpha>0.5$ the only systems
that are predicted to exist are those with He\,WDs. The fraction of
ONe\,WD systems reaches a maximum near $\alpha=0.2$ and then
decreases.
We note that systems with He\,WDs tend to populate the lowest period
range at all $\alpha$. These systems are generally characterised by
initially lower-mass primaries, and thus lower-mass companions
because $q\le 1$, and shorter orbital periods and initiate common
envelope evolution before helium ignition. Usually systems
characterised by short initial periods are unlikely to survive at low
$\alpha$ because the stronger drag force exerted on the two stars
during common envelope evolution causes them to merge.
Fig.\,\ref{fig:RLOFPeriodS} shows that most companions, particularly
at shorter orbital periods, are low-mass deeply convective stars.
More massive secondaries are generally found at longer periods for
three reasons. First, longer orbital periods require high-mass white
dwarfs to initiate stable mass transfer over the age of the Galactic
Disc and these massive white dwarfs can have secondaries with
masses all the way up to $1.44$\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi, provided $q\le 1$. Second,
during common envelope evolution for a fixed primary initial mass and
orbital period, systems with more massive secondaries have more
orbital energy and so a smaller portion of this energy is necessary to
eject the envelope. This leads to longer orbital periods. Third, for
a fixed white dwarf mass, more massive secondaries fill their Roche
lobes at longer orbital periods and so systems with more massive
companions are more likely to evolve into PREPS.
\subsubsection{Stellar pair distribution}\label{stellarpairs}
Table~\,\ref{tab:Contribs} lists fractions of the various combinations
of types of white dwarf primaries and secondary types just before RLOF
commences. At low $\alpha$ the predominant combination is a CO\,WD
primary with a low-mass M-dwarf secondary. Second in abundance are
systems comprised of a He\,WD with a low-mass M-dwarf secondary.
Other combinations are also found but in much smaller numbers. At
high $\alpha$ the two major categories are swapped and those systems
with He\,WD primaries become the predominant type. The observed
fraction of He\,WDs ($f_{\rm He}$) is generally low among classical
CVs ($f_{\rm He}\mbox{$\, \stackrel{\scriptstyle <}{\scriptstyle\sim}\,$} 10$\,per cent) and pre-CVs
($f_{\rm He}\mbox{$\, \stackrel{\scriptstyle <}{\scriptstyle\sim}\,$} 17\pm8$\,per\,cent as shown by
\citet{Zorotovic2011}. The results in Table\,\ref{tab:Contribs}
indicate that in order to reproduce the observed low fraction of
He\,WDs our models need to be restricted to $\alpha<0.3$.
\begin{table*}
\caption{The fraction of the
combinations of types of white dwarf primaries and
secondary types just before RLOF commences for various $\alpha$.
The stellar type CS is a deeply or fully convective low-mass MS star with $M<0.7$\,M$_\odot$.}
\centering
\label{tab:Contribs}
\begin{tabular}{c c c c c c c }
\hline
\noalign{\smallskip}
\multicolumn{7}{c}{MCV progenitor pairs, fraction per cent} \\
$\alpha$ & He\,WD/CS & CO\,WD/CS & ONe\,WD/CS & He\,WD/MS & CO\,WD/MS & ONe\,WD/MS \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.10 &14.86 &69.63 & 5.72 &0.00 &3.77 &6.03 \\
0.20 &30.27 &52.27 &12.99 &0.00 &0.38 &4.10 \\
0.30 &61.36 &25.69 &12.49 &0.00 &0.00 &0.46 \\
0.40 &96.44 & 7.78 & 5.78 &0.00 &0.00 &0.00 \\
0.50 &95.85 & 1.72 & 2.44 &0.00 &0.00 &0.00 \\
0.60 &98.75 & 0.28 & 0.98 &0.00 &0.00 &0.00 \\
0.70 &99.67 & 0.01 & 0.32 &0.00 &0.00 &0.00 \\
0.80 &99.92 & 0.00 & 0.00 &0.00 &0.00 &0.00 \\
\hline
\end{tabular}
\end{table*}
\subsubsection{Mass distribution}\label{mass_distribution}
Fig.\,{\ref{fig:RLOFMassP}} shows that all our models predict that, on
average, longer orbital period systems contain CO\,WDs while
shorter-period systems tend to have He\,WD primaries. At low $\alpha$
the primaries are predominantly CO\,WDs with masses in the range 0.5
to 1.1\,M$_\odot$ followed in lesser numbers by He\,WDs with masses in
the range 0.4 to 0.5\,M$_\odot$ while ONe\,WDs, with masses in the
range 1.2 to 1.4\,M$_\odot$, are rarer with their incidence reaching a
maximum near $\alpha=0.2$.
We note that there is a curious dip in the white dwarf mass
distribution near $M_{\rm WD}=0.8$\,M$_\odot$ which widens as
$\alpha$ increases until all CO and ONe\,WDs disappear for
$\alpha> 0.5$. This is because as $\alpha$ increases,
systems emerge from common envelope at progressively longer periods,
because large $\alpha$ means a high envelope clearance efficiency
which leads to larger stellar separation at the end of the common
envelope stage. However the longer the orbital period, the higher the
white dwarf mass needs to be for stable mass transfer to
commence. Thus the gap in the white dwarf mass distribution is caused
by those systems that emerge from common envelope at large separations
but with white dwarf primaries that are not massive enough to allow
RLOF to take place. Another, albeit much narrower, gap occurs near
0.5\,M$_\odot$ for all $\alpha$ but becomes wider for
$\alpha\ge0.2$. This gap also persists until all CO and ONe\,WDs
disappear at $\alpha > 0.5$. It divides systems with He\,WDs primaries
from those with CO\,WDs and is linked to whether the stars enter
common envelope evolution on the RGB, and so produce a He\,WD primary
with $M_{\rm WD}\mbox{$\, \stackrel{\scriptstyle <}{\scriptstyle\sim}\,$} 0.5$\,M$_\odot$, or on the AGB, and so produce
a CO\,WD primary with $M_{\rm WD}> 0.5$\,M$_\odot$.
Fig.\,\ref{fig:RLOFMassS} shows again that the secondaries are
predominantly low-mass deeply or fully convective M-dwarf stars. The
distribution has a broad peak around 0.1 to 0.3\,M$_\odot$ at
$\alpha=0.1$ to 0.2 with a long tail extending to $1.2$\,M$_\odot$. As
$\alpha$ increases, the peak in the secondary mass distribution shifts
to slightly lower masses (around 0.1 to 0.25\,M$_\odot$) but the
high-mass tail shrinks quite dramatically. At $\alpha\ge0.4$ the
distribution is confined to secondary masses of less than about
$0.3$\,M$_\odot$. As already noted in section \ref{PerDist}, the
majority of these very low-mass donor stars belong to systems that
underwent common envelope evolution during the Hertzsprung gap or RGB
phases and thus have He\,WD primaries with
$M_{\rm WD}\mbox{$\, \stackrel{\scriptstyle <}{\scriptstyle\sim}\,$} 0.5$\,M$_\odot$. We also note that systems with
low-mass secondaries ($M_{\rm sec}\mbox{$\, \stackrel{\scriptstyle <}{\scriptstyle\sim}\,$} 0.35$\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi) remain detached
for longer because magnetic braking is inefficient in these stars and
gravitational radiation is the main source of loss of angular
momentum.
\begin{figure*}
\begin{center}
\includegraphics[width=1.01\textwidth]{RLOF-MassP-Hi-NEW.eps}
\caption{Theoretical mass distribution of the white dwarf primary star
of magnetic systems just before they start RLOF for various
$\alpha$. The distributions of the three white dwarf types are shown
as three superimposed coloured categories.}
\label{fig:RLOFMassP}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=1.00\textwidth]{RLOF-MassS-Hi-NEW.eps}
\caption{As in Fig\,\ref{fig:RLOFMassP} but for the secondary star
types shown as the coloured categories. Both secondary star types
are MS stars. The CS type is a deeply or fully
convective MS star with $M<0.7$\,M$_\odot$.}
\label{fig:RLOFMassS}
\end{center}
\end{figure*}
\subsubsection{Magnetic field distribution}\label{MagFieldDistr}
Fig. \ref{fig:RLOFOvTBmag} shows the theoretical magnetic field
distribution and the breakdown of the primary white dwarf types for
our range of $\alpha$. The maximum field strength is a few $10^8$\,G
and is found mostly in systems whose primary is a He\,WD. The reason
for this is that systems that undergo common envelope evolution during
the RGB evolution have shorter initial orbital periods and create very
short period binaries with a highly magnetic white dwarf, as expected
from equation (\ref{EqBfield}).
The magnetic field distribution is dominated by systems with CO\,WD
primaries when $\alpha\le 0.2$. When $\alpha\ge 0.4$ the field
distribution becomes narrower and its peak shifts to higher field
strengths. For $\alpha\ge0.5$ the field distribution only contains
very highly magnetic He\,WD primaries with a peak near
$3.2\times 10^8$\,G. This shift to high fields is because those
systems that go through common envelope evolution while their
primaries are on the RGB merge for low $\alpha$ but can survive for
high $\alpha$ giving rise to very short orbital period systems with
strongly magnetic, low-mass white dwarfs.
We note that the magnetic field distribution has a dip near
$8\times 10^6$\,G appearing at $\alpha\ge 0.2$ and persisting until
all CO and ONe\,WDs disappear from the distribution. This is
reminiscent of the dip that we encountered in the white dwarf mass
distribution (see \ref{mass_distribution}) and has the same
explanation. The similar behaviour is because the magnetic field
strength is a function of white dwarf mass (by virtue of equations
\ref{EqBfield}). The field dip is thus
caused by the dearth of systems with white dwarf masses around
$0.8\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ (see Fig.\,\ref{fig:RLOFMassP}).
\section{Comparison to observations}\label{Comparison}
The optimal observational sample with which to compare our results
would be that formed by the known magnetic PREPs. However, this sample
is exceedingly small and observationally biased. To make things worse,
not all PREPS have well determined parameters, such as masses and
magnetic field strengths. Hence, for some of these studies we use the
observed sample of MCVs, noting the following important points (i) the
MCV sample is magnitude-limited, (ii) MCVs suffer from prolonged high
and low states of accretion and (iii) MCVs include systems at all
phases of evolution. Some of them began Roche lobe overflow billions
of years ago while others have only recently begun mass
exchange. Therefore, one should take such a comparison with some
degree of caution particularly when we compare quantities
that change over time, such as orbital periods and masses. When
comparing masses we will also use the observed sample of non-magnetic
Pre-CVs \citep{Zorotovic2011}.
The tables of \citet{fer2015a} show that the observed orbital periods
of MCVs are in the range 1 to 10\,hr, masses are between about $0.4$
and $1.1$\,$\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$ and that the magnetic field distribution is relatively
broad with a peak near $3.2\times10^7$\,G. A quick glance at
Figs\,\ref{fig:RLOFPeriodP}, \ref{fig:RLOFMassP} and
\ref{fig:RLOFOvTBmag} immediately reveals that models with
$\alpha>0.3$ are all unable to reproduce the general
characteristics expected from the progenitors of the observed
population of MCVs and we elaborate on this in more details
below. Generally, we see that models with $\alpha>0.3$ are
not realistic and evolutionary effects cannot account for the large
degree of discrepancy between theory and observations.
\begin{figure*}
\begin{center}
\includegraphics[width=1.0\textwidth]{RLOF-Bmag-NEW.eps}
\caption{Pink shaded histogram: Total theoretical magnetic field
distribution of the white dwarf primary stars in magnetic systems just before they
start RLOF for the indicated $\alpha$. The histograms of the three
types of white dwarfs making up the total theoretical magnetic field distribution are
shown as the foreground coloured histograms. These three are made partially
transparent so that details of the other histograms can be seen through them.}
\label{fig:RLOFOvTBmag}
\end{center}
\end{figure*}
We begin our analysis with the magnetic field distribution. There is
no evidence for field decay among MCVs \citep{fer2015a,Zhang2009} so we can
assume that the magnetic field strength remains unchanged over the
entire life of the magnetic binary.
We have used a Kolmogorov--Smirnov (K--S) test \citep{Press1992} to
compare the magnetic field distribution of the observed population
with the theoretical results. This test establishes the likelihood
that two samples are drawn from the same population by comparison of
the cumulative distribution functions (CDFs) of the two data
samples. The CDFs of the two distributions vary between zero and one
and the test is on the maximum of the absolute difference $D$ between
the two CDFs. It gives the probability $P$ that a random selection
would produce a larger $D$. Five model CDFs for five different
$\alpha$'s and the CDF for the known observed magnetic fields of 81 MCV
systems are compared in Fig.\,\ref{fig:MCVWDBMagCDF}.
\begin{figure}
\begin{center}
\includegraphics[width=1.00\columnwidth]{MCV-BMag-CDF.eps}
\caption{Theoretical cumulative distribution functions for the
magnetic fields of MCV white dwarfs at RLOF for
$\alpha=0.10, 0.15, 0.20, 0.30$ and $0.40$ and the CDF of the
observed magnetic field of 81 systems taken from \citet{fer2015a} }
\label{fig:MCVWDBMagCDF}
\end{center}
\end{figure}
\begin{table}
\caption{Kolmogorov–-Smirnov $D$ statistic and probability $P$ of the magnetic field
distributions of the observed and synthetic populations of MCVs for a range of $\alpha$.}
\centering
\begin{tabular} { c c c}
\hline
\noalign{\smallskip}
$\alpha$ & $D$ & $P$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.10 & 0.17476 & 0.36069 \\
0.15 & 0.19349 & 0.24632 \\
0.20 & 0.25141 & 0.05845 \\
0.30 & 0.22962 & 0.10500 \\
0.40 & 0.26939 & 0.04298 \\
0.50 & 0.35186 & 0.00429 \\
0.60 & 0.38035 & 0.00006 \\
0.70 & 0.61987 & 0.00000 \\
0.80 & 0.94366 & 0.00000 \\
\noalign{\smallskip}
\hline
\end{tabular}
\label{tab:KS_Bfield}
\end{table}
The observed samples of MCVs and magnetic PREPs are very biased,
particularly at the low and high ends of the magnetic field
distribution. At low fields ($B\mbox{$\, \stackrel{\scriptstyle <}{\scriptstyle\sim}\,$} 10$\,MG) the observed radiation
is dominated by the truncated accretion disc. In these low-field
systems the photosphere of the white dwarf is never visible and Zeeman
splitting cannot be used to determine field strengths. Nor can
cyclotron lines be used to measure fields because they are too weak
and invisible in the observed spectra. In the high field regime
($B\mbox{$\, \stackrel{\scriptstyle >}{\scriptstyle\sim}\,$} 100$\,MG) mass accretion from the companion star is
inhibited \citep{Ferrario1989, LiWuWick1994} and so high field MCVs
are very dim wind accretors often below the detection limits of most
surveys \citep[AR\,UMa][]{Hoard2004}. Because of these biases the
observed samples in these regimes are far from complete and
theoretical fits are unreliable. We therefore restrict our comparison
between theory and observations to field strengths in the range 10 to
70\,MG.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\columnwidth]{Obs-Bmag.eps}
\caption{Comparison of the theoretical magnetic field strength for
$\alpha= 0.1$ and the observed magnetic field strength of the 81
MCVs taken from \citet{fer2015a}}
\label{fig:CompareBmag}
\end{center}
\end{figure}
The results of the K--S test for our range of $\alpha$ are displayed
in Table \ref{tab:KS_Bfield} and show that the field distribution is a
better match to the observations at low $\alpha$. The comparison of
the magnetic field distribution between theory and observations is
shown in Fig.\,\ref{fig:CompareBmag} for $\alpha=0.1$.
We stress that for $\alpha>0.3$ all the theoretical magnetic field
distributions shown in Fig.\,\ref{fig:RLOFOvTBmag} are very
unrealistic because only very high field ($B>60$\,MG) He\,WDs
($M\lsimeq0.5$\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi) are predicted to exist. This is contrary to
observations that show that fields cover a much wider range of
strengths ($\mbox{a few}10^6$ to a few $\,10^8$\,G) and white dwarf
masses (0.4 to 1.2\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi) as seen in Tables\,2 and 3 of
\citet{fer2015a}. We note here that a larger $B_0$ with
$\alpha>0.3$ is not a good fit because high $\alpha$ models have a
very narrow distribution of field strengths skewed to high fields. A
larger $B_0$ would only exacerbate this by pushing the distribution
to even higher fields. If $B_0$ were chosen at the lower end of the
acceptable range found in paper\,II the field distribution would
shift to lower fields and yield an average field strength closer to
observations. However, because $B_0$ can only shift the field
distribution to lower (or higher) fields, the very narrow width of
the distribution that characterises all our high $\alpha$ models
would not be corrected by readjusting $B_0$. We can therefore
conclude that the $B_0$ calibration carried out in paper\,II holds
for the modelling of magnetic white dwarfs in binaries and supports
our common envelope dynamo theory, according to which the fields of
MCVs are scaled down, through equation\,(\ref{EqBfield}),
from the maxima obtained during merging events that generate
isolated HFMWDs.
We have performed a K--S study between the synthetic white dwarf mass
distribution and that of white dwarf masses in MCVs taken from
\citet{fer2015a}. In principle such a comparison can be
justified if we make the usual assumption that the mass of the white
dwarf does not grow in CVs because nova eruptions tend to expel all
material that is accreted. However, we have found that the
K--S test applied to the white dwarf masses of the theoretical and
observed population of MCVs yields poor results, as shown in the
second and third columns of Table \ref{tab:KS_WDmass}. Such
a conflict is not surprising because our assumption that the mass of
the white dwarf does not grow because of nova eruptions may not be
correct.
In this context, we note that \citet{Zorotovic2011} noticed a curious
discrepancy in their observational data of CVs and Pre-CVs. That is,
they found that the mean white dwarf mass in CVs
($0.83\pm 0.23$\,M$_\odot$) significantly exceeds that of pre-CVs
($0.67\pm0.21$\,\,M$_\odot$) and they excluded that this difference
could be caused by selection effects. The two possible solutions
advanced by \citet{Zorotovic2011} were that either the mass of the
white dwarf increases during CV evolution, or a short phase of thermal
time-scale mass transfer comes before the formation of CVs during
which the white dwarf acquires a substantial amount of mass via stable
hydrogen burning on the surface of the white dwarf \citep[as first
suggested by][]{Schenker2002}. During this phase the
system may appear as a super-soft X-ray source
\citep[][]{Kahabka1997}. Using this assumption \citet{Wijnen2015}
could build a large number of massive white dwarfs. However their
model still created too many low-mass He\,WDs and too many evolved
companion stars contrary to observations. Another possibility has
recently been advanced by \citet{Zorotovic2017}. In order to achieve
a better agreement between their binary population synthesis models
and observations of CVs they adopted the ad-hoc mechanism proposed by
\citet{Schreiber2016} which surmises the existence of additional
angular momentum losses generated by mass transfer during the CV
phase. Such losses are assumed to increase with decreasing white dwarf
mass and would cause CVs with low-mass white dwarfs to merge and
create an isolated white dwarf. By removing these merged systems from
the synthetic CV sample the average white dwarf mass
increases. Furthermore such a mechanism would explain the existence of
isolated low-mass white dwarfs ($M_{\rm WD}<0.5\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$) that
constitute around 10\,per cent of all single white dwarfs observed in
the solar neighbourhood \citep[e.g.][]{Kepler2007}.
Going back to our studies, if a comparison between white dwarf masses
in MCVs and our synthetic population may not be meaningful, the next
best sample to use for our K--S test is the observed white dwarf
masses of pre-CVs \citep{Zorotovic2011}. We show in
Fig.\,\ref{fig:MCV-WD-Mass-CDF} the CDFs of the mass distributions
for the observed sample and the theoretical distribution of the white
dwarf masses at the beginning of RLOF for various $\alpha$. The results are
reported in the fourth and fifth columns of Table \ref{tab:KS_WDmass}
and show that the agreement between observations and theory is greatly
improved. The comparison of the synthetic and observed Pre-CV\, white
dwarf mass distribution is shown in Fig.~\ref{fig:CompareMasses} for
the largest K--S probability at $\alpha = 0.10$. This suggests that
the white dwarf mass distribution of the binary population in which
primaries develop a magnetic field during common envelope evolution
and later evolve to MCVs does not differ from that of the binary
precursors of classical non-magnetic CVs.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\columnwidth]{MCV-WD-Masses-CDFs.eps}
\caption{Cumulative Distribution Functions of the mass distributions
for the observed pre-CV white dwarf masses taken from
\citet{Zorotovic2011} and the theoretical distribution of the
white dwarfs as the systems start RLOF for $\alpha=0.10,
0.15, 0.20, 0.30$ and $0.40$. The K-S statistics for this plot are
shown in the fourth and fifth columns of table~\ref{tab:KS_WDmass}}
\label{fig:MCV-WD-Mass-CDF}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1.0\columnwidth]{RLOF-TvO-MassP.eps}
\caption{Comparison of the mass distributions for the observed pre-CV white
dwarf masses taken from \citet{Zorotovic2011} and the theoretical mass
distribution of the white dwarfs as the systems start RLOF for
$\alpha= 0.10$.}
\label{fig:CompareMasses}
\end{center}
\end{figure}
\begin{table}
\caption{K--S $D$ statistic and probability $P$ of the
white dwarf mass distributions of the observed MCVs listed by
Ferrario et al. (2015a, second and third
columns) and our synthetic populations
for $\alpha$ given in the first column. In the fourth and fifth
columns we show the K--S results of the observed
Pre-CV masses of \citet{Zorotovic2011} and our synthetic
populations at the start of RLOF (fourth and fifth columns).}
\centering
\begin{tabular} {c c c c c}
\hline
\noalign{\smallskip}
$\alpha$&$D$&$P$&$D$&$P$\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.10 & 0.37687 & 0.02023088 & 0.12954 & 0.95281557 \\
0.15 & 0.49861 & 0.00064407 & 0.23478 & 0.34844783 \\
0.20 & 0.56677 & 0.00006150 & 0.26010 & 0.23507547 \\
0.30 & 0.62615 & 0.00000622 & 0.48014 & 0.18713800 \\
0.40 & 0.69590 & 0.00000032 & 0.66148 & 0.00106500 \\
\noalign{\smallskip}
\hline
\end{tabular}
\label{tab:KS_WDmass}
\end{table}
We note that the Pre-CV observational sample shows a dearth of systems
in the white dwarf mass distribution centred around $0.8\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$. This
mass gap was already noted in the theoretical {\sc{bse}} models and the
reasons for its existence were explained in section \ref{mass_distribution}.
The smaller size of this gap for models with $\alpha\le0.2$ explains why
we achieve a better fit with observations for $\alpha=0.1$, as indicated
by the K--S test.
If $\alpha>0.3$ the theoretical white dwarf mass distribution shown in
Fig.\,\ref{fig:RLOFMassP} is very unrealistic because only He\,WDs
($M<0.5$\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi) are predicted to exist by these models. This is
contrary to observations that show that masses cover the much wider
range 0.4 to 1.2\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi \citep[see Tables\,2 and 3 in][]{fer2015a}.
Next, we look at the secondary mass distribution, keeping in mind that
a comparison between our synthetic {\sc{bse}} mass sample and the
observed secondary masses in MCVs is definitely not appropriate
because secondary masses decrease over time as mass is transferred to
the white dwarf. Nonetheless it may still be pertinent to use the
observed Pre-CV sample to study and compare the overall
characteristics of these samples so that we can, at the very least,
discard some of the most extreme theoretical models.
Fig. \ref{fig:RLOFMassS} shows that if $\alpha>0.3$ then
$M_{\rm sec}< 0.3\,\ifmmode{{\rm M}_\odot}\else${\rm M}_\odot$\fi$, which is inconsistent with observations of
pre-CVs \citep[see][]{Zorotovic2011}. Furthermore, we can see that
when $\alpha>0.2$, the decline towards higher masses becomes far too
steep. This straightforward comparison seems again to indicate that
models with $\alpha>0.3$ are very unrealistic and therefore low $\alpha$
yields a better fit to observations.
\section{Discussion and Conclusions}\label{DandC}
The origin of large-scale magnetic fields in stars is still a puzzling
question \citep[see][]{fer2015b}. However, the results from
recent surveys such as the SDSS \citep{Kepler2013}, BinaMIcS
\citep{ale2015} and MiMes \citep{Wade2016} have provided us with a much
enlarged sample of magnetic stars that have allowed investigators to
conduct studies like this one. There are two main competing scenarios
to explain the existence of magnetic fields in white dwarfs. In 1981,
\citeauthor{ang1981} proposed that the magnetic Ap and Bp stars
are the most likely progenitors of the highly MWDs under the
assumption of magnetic flux conservation \citep[see
also][]{tout2004,wickramasinghe2005}.
The best clue so far on the origin of fields in white dwarfs (isolated
and in binaries) has come from the study of their binary properties
\citep{Liebert2005,Liebert2015}, as outlined in section
\ref{WhereProg}. This is why the proposal by \citet{Tout2008}, that
the origin of magnetic fields in white dwarfs is related to their
duplicity and stellar interaction during common envelope evolution, is
becoming more and more appealing.
We have extended our population synthesis study of binary systems
carried out for papers\,I and\,II for the HFMWDs to explain the origin
of fields in the accreting white dwarfs in MCVs. Similarly to the
investigations conducted in paper\,I and\,II, we have varied the
common envelope efficiency parameter $\alpha$ to investigate its
effects on the resulting synthetic population of MCVs. We have shown
that models with $\alpha\ge 0.4$ are not able to reproduce the large
range of white dwarf masses, field strengths, and secondary types and
masses that are observed in MCVs and therefore models with
$\alpha< 0.4$ best represent the observed data. K--S tests conducted
to compare our synthetic white dwarf mass and magnetic field
distributions with the observed populations have given us some
quantitative support in favour of models with $\alpha<0.4$.
However, we need to stress again some of the shortcomings of our work
and in particular those that arise from our comparison to
observations. Many of the parameters (e.g., white dwarf mass, magnetic
field, secondary star mass and type, orbital period) that characterise
the Galactic populations of MCVs and PREPs and are needed for
comparison studies are often hard to ascertain owing to evolutionary
effects and observational biases that are difficult to
disentangle. For instance, we mentioned in section \ref{Comparison}
that magnetic white dwarfs in PREPs would be the best objects with
which to compare our theoretical results and in particular the mass
distribution, because mass is not contaminated by accretion
processes. On the other hand there are far too few members of this
population. The white dwarf mass distribution provided by the much
larger sample of MCVs cannot be used either for comparison purposes
because masses vary over time, owing to accretion and nova
explosions. So we have used the sample provided by the
non-magnetic Pre-CVs of \citet{Zorotovic2011}.
The situation is somewhat ameliorated when we consider the magnetic
field distribution because fields are not expected to change over time
\citep[see][]{fer2015a}. However, the true magnetic field distribution
of MCVs is not well known because it is plagued by observational
biases. For example, at field strengths below a few\,$10^7$\,G most
systems (the intermediate polars) have an accretion disc from which
continuum emission and broad emission lines swamp the Zeeman and
cyclotron features arising from the white dwarf surface
and so hide those spectral signatures that are
crucial to determine their field strengths. Very high field polars are
also likely to be under-represented in the observational set because
mass accretion from the companion star is impeded by the presence of
strong fields \citep{Ferrario1989, Li1998} making these systems
very dim wind accretors.
Despite the limitations highlighted above, we have shown that the
characteristics of the MCVs are generally consistent with those of a
population of binaries that is born already in contact (exchanging
mass) or close to contact, as first proposed by \citet{Tout2008}. This
finding is also in general agreement with the hypothesis that the
binaries known as PREPs, where a MWD accretes matter from the wind of
a low-mass companion, are the progenitors of the MCVs.
\section*{Acknowledgements}
GPB gratefully acknowledges receipt of an Australian Postgraduate
Award. CAT thanks the Australian National University for supporting a
visit as a Research Visitor of its Mathematical Sciences Institute,
Monash University for support as a Kevin Westfold distinguished visitor
and Churchill College for his fellowship.\\
|
\section{Introduction and preliminaries}
Nowadays several parameters related to distances in graphs are highly attracting the attention of several researchers. One of them, namely, the metric dimension, has specifically centered several investigations. To see the richness of this topic, among several possible references, we would suggest for instance the three Ph. D. dissertations \cite{ale}, \cite{dk} and \cite{yunior}, and references cited therein. In this concern, a vertex $v$ of a connected graph $G$ \emph{distinguishes} (\emph{determines} or \emph{recognizes}) two vertices $u,w$ if $d(u,v)\ne d(w,v)$, where $d(x,y)$ represents the length of a shortest $x-y$ path in $G$. A subset of vertices $S$ of $G$ is a \emph{metric generator} for $G$, if any pair of vertices of $G$ is distinguished by at least one vertex of $S$. A metric generator of minimum cardinality is called a \emph{metric basis} for $G$ and its cardinality is the \emph{metric dimension} of $G$, which is denoted by $dim(G)$. These concepts were introduced by Slater in \cite{leaves-trees} in connection with some location problems in graphs. On the other hand, the concept of metric dimension was independently introduced by Harary and Melter in \cite{harary}.
A standard metric generator, as defined above, uniquely recognizes all the vertices of a graph in order to look out how they do ``behave'' in the graph. However, this ``surveillance'' does not succeed if an anomalous situations occurs in some edge between two vertices instead of in a vertex. It is possible that a metric generators properly identifies the edges in order to also see their behaving, but in general this is not possible. In this sense, in the way of correctly recognize the edges of a graph, a new parameter was recently introduced in \cite{KeTrYe}. Another variant on such direction was also presented in \cite{mix-dim} where not only edges are recognized between them, but also there is a recognition scheme between any two elements (vertices or edges) of a graph. In this work we only center our attention into recognizing edges.
Given a connected graph $G=(V,E)$, a vertex $v\in V$ and an edge $e=uw\in E$, the distance between the vertex $v$ and the edge $e$ is defined as $d_G(e,v)=\min\{d_G(u,v),d_G(w,v)\}$. A vertex $w\in V$ \emph{distinguishes} two edges $e_1,e_2\in E$ if $d_G(w,e_1)\ne d_G(w,e_2)$. A nonempty set $S\subset V$ is an \emph{edge metric generator} for $G$ if any two edges of $G$ are distinguished by some vertex of $S$. An edge metric generator with the smallest possible cardinality is called an \emph{edge metric basis} for $G$, and its cardinality is the \emph{edge metric dimension}, which is denoted by $\edim(G)$. In \cite{KeTrYe}, the concepts above were defined only for the case of connected graphs. However, if we consider non-connected graphs, then the parameter could be easily adapted by considering the distance between two vertices belonging to two different components as infinite. Nevertheless, such adapting make not much sense, since then we can readily seen the following.
\begin{remark}
If $G$ is a non-connected graph with components $G_1,\dots, G_r$, then
$$\edim(G)=\sum_{i=1}^r \edim(G_i).$$
\end{remark}
In connection with this remark above, we can consider every component of a graph separately. Moreover, any necessary concept, terminology and notation required in the work will be introduced throughout the exposition, right before it is firstly used, and for any remaining basic graph theory terminology we follow the book \cite{west}.
Studies of graph products have been intensively made in the last few decades and by now, a rich theory involving the structure and recognition of classes of these
graphs has emerged, cf. \cite{ImKl}. The most interesting and studied graph products are the Cartesian product, the strong product, the direct product and the
lexicographic product which are also called \emph{standard products}. There are several other operations made with graphs (from which some of them are also called as product graphs in the literature) that have also attracted the attention of several researchers. Some of them are for example the corona product, the join graphs, the rooted product and the hierarchical product. There are different styles of studying the graph products (or graph operations). One of them involves the analysis of the properties of the structure itself and a second one standard approach to graph products is to deduce properties of a product with respect to (usually the same) properties of its factors. This latter situation is the center of our work, in connection with the edge metric dimension. Some primary studies on the edge metric dimension of Cartesian product graphs were presented in \cite{KeTrYe}, where the value of the edge metric dimension was computed for the grid graphs (Cartesian product of two paths), and for some cases of torus graphs (Cartesian product of cycles). Moreover, some other results on this topic can be found in \cite{zubri}, where the edge metric dimension of the join graph $G\vee K_1$\footnote{The graph $G\vee K_1$ is obtained from a graph $G$ and a vertex $v$, by joining with and edge every vertex of $G$ with the vertex $v$.}, and of the Cartesian product of a path with any graph $G$ was studied. To the best of our knowledge there are no more results concerning the edge metric dimension of product graphs. In contrast with this fact, other variants of the standard metric dimension have been deeply studied in the last recent years. Some examples are for instance \cite{Barragan,Estrada,Jannesari} to just name those ones concerning the lexicographic product of graphs, which is one of the studied product of our work. In this sense, it is now our goal to make several contribution to this topic of edge metric dimension, and we precisely begin with studying the lexicographic product, the join and the corona product graphs.
\section{The join of graphs}
Given two graphs $G$ and $H$, the join graph $G\vee H$ is obtained from $G$ and $H$ by adding an edge between any vertex of $G$ and any vertex of $H$. In this sense, it is clearly observed that the join graph $G\vee H$ is always a connected independently of the connectivity of the graphs $G$ and $H$. We next study the edge metric dimension of join graphs.
To this end, we need the following terminology and notation. A set of vertices $D$ of a graph $G$ is a \emph{total dominating set} of $G$ if every vertex of $G$ is adjacent to a vertex of $D$. The minimum cardinality of any total dominating set of $G$ is the \emph{total domination number} of $G$ and is denoted by $\gamma_t(G)$. A set of vertices of cardinality $\gamma_t(G)$ is called a $\gamma_t(G)$-\emph{set}.
A graph $G$ is in a class of graphs $\mathcal G$ if for any vertex $u\in V(G)$ there is an edge $uv$ incident with $u$ ($v$ is a neighbor of $u$), such that $\{u,v\}$ is a $\gamma_t(G)$-set. It is easy to see that complete and complete bipartite graphs on at least two vertices are in $\mathcal G$. Let us mention that if $\gamma_t(G)$ does not exists, then $G$ contains an isolated vertex and is not in class $\mathcal G$, in particular $K_1$ is such.
\begin{theorem}
\label{join} For any non trivial graphs $G$ and $H$,
\begin{equation*}
\edim(G\vee H)=\left\{
\begin{array}{ll}
|V(G)|+|V(H)|-1, & \mbox{if $G\in\mathcal G$ or $H\in\mathcal G$}, \\[0.2cm]
|V(G)|+|V(H)|-2, & \mbox{if $G,H\notin\mathcal G$}. \\
\end{array
\right.
\end{equation*}
\end{theorem}
\begin{proof}
Let $G$ and $H$ be any graphs. The trivial upper bound for any graph is $\edim(G)\leq |V(G)|-1$ and therefore $\edim(G\vee H)\leq |V(G)|+|V(H)|-1$. Let $M$ be an edge metric basis for $G\vee H$. The distance between any edge $e=gh$ from $G\vee H$, where $g\in V(G), h\in V(H)$, to any vertex different from $g$ and $h$ equals one. Let $g$ be a fixed vertex from $G$. For $h\neq h'$, the edges $gh$ and $gh'$ have distance one to all vertices different from $g,h,h'$. Moreover, both edges have $g$ in common, so that the distance between them and $g$ equals zero in both cases. Therefore, at least one of $h$ and $h'$ must be in $M$. This argument can be repeated for any pair of edges of the form $gh$ and $gh'$ for any $h'\in V(H)-\{h\}$, and we see that at least $|V(H)|-1$ vertices from $H$ must be in $M$. Symmetrically we can see that at least $|V(G)|-1$ vertices from $G$ must be in $M$ and so, the lower bound $\edim(G\vee H)\geq |V(G)|+|V(H)|-2$ follows.
Assume now that one of the graphs, say $G$, belongs to $\mathcal G$. Suppose, with a purpose of contradiction, that $\edim(G\vee H)=|V(G)|+|V(H)|-2$ and let $S$ be an edge metric basis for $G\vee H$. In concordance with the above, it follows $|S\cap V(G)|=|V(G)|-1$ and $|S\cap V(H)|=|V(H)|-1$. Let $g\in V(G)$ and $h\in V(H)$ be outside of $S$. Since $G\in \mathcal G$, $G$ has no isolated vertices, and there exists $g'\in V(G)$ which is adjacent to $g$ and $\{g,g'\}$ is a $\gamma_t(G)$-set. This implies that $gg'$ is at distance one to every vertex from $V(G)-\{g,g'\}$. Moreover, the edge $gg'$ is also at distance one to every vertex from $H$. As already mentioned the edge $g'h$ has distance one to every vertex from $V(G\vee H)-\{g',h\}$. Thus, the only vertices that distinguish the edges $gg'$ and $g'h$ are $g$ and $h$, which are not in $S$, and this is a contradiction with $S$ being an edge metric basis for $G\vee H$. Therefore, $\edim(G\vee H)>|V(G)|+|V(H)|-2$ and we have the equality $\edim(G\vee H)= |V(G)|+|V(H)|-1$.
To finish the proof let now $G,H\notin\mathcal G$. In this sense, there exists a vertex $g'\in V(G)$ such that $\{g,g'\}$ is not a $\gamma_t(G)$-set for every vertex $g$ adjacent to $g'$. In other words, there exists a vertex $x_g\in V(G)$ such that $d_G(gg',x_g)\geq 2$. Similarly, there exists a vertex $h'\in V(H)$ such that for every neighbor $h$ of $h'$ there exists a vertex $y_h\in V(H)$ such that $d_H(hh',y_h)\geq 2$. We will show that the set $S=V(G\vee H)-\{g',h'\}$ is an edge metric generator for $G\vee H$. Clearly, any two edges with both end-vertices in $S$ are distinguished by at least one vertex from $S$ (one of the end-vertices will do so). Similarly, edges $g'u$ and $g'v$, $u\neq v$, are also identified, because at least one of $u$ and $v$ is in $S$. The same happens with two edges $h'w$ and $h'z$, $w\neq z$. Thus, let $g'u$ and $h'v$ be two edges. If $u\neq v$, then without loss of generality, $u\in S$ distinguish $g'u$ and $h'v$. Hence, let $u=v$ and, by the symmetry of $G$ and $H$, we may assume that $u=g\in V(G)$. Because $G\notin\mathcal G$ for every edge, also for $gg'$, there exists a vertex $x_g\in V(G)$, at distance at least two to $gg'$ in $G$. Clearly, $x_g\in S$, $d_{G\vee H}(gh',x_g)=1$, $d_{G\vee H}(gg',x_g)=2$ and so, $x_g$ distinguishes the edges $gg'$ and $gh'$. Therefore, $S$ is an edge metric generator for $G\vee H$, and the equality $\edim(G\vee H)=|V(G)|+|V(H)|-2$ follows.
\end{proof}
It is easy to see that Theorem \ref{join} covers the results from \cite{KeTrYe} for a wheel $K_1\vee C_n$, for a fan $K_1\vee P_n$ and for a complete bipartite graph $K_{p,q}=N_p\vee N_q$. This latter result concerning bipartite graphs can be generalized for complete multipartite graphs as next shown.
\begin{corollary}\label{completebipartite}
For any complete multipartite graph $K_{r_1,\ldots,r_t}$,
\begin{equation*}
\edim(K_{r_1,\ldots,r_t})=\left\{
\begin{array}{ll}
r_1+r_2-2, & \mbox{if $t=2$,} \\
\sum_{i=1}^tr_i-1, & \mbox{if $t>2$.} \\
\end{array
\right.
\end{equation*}
\end{corollary}
\begin{proof}
For $t=2$ we have $\edim(K_{r_1,r_2})=\edim(N_{r_1}\vee N_{r_2})=r_1+r_2-2$ by Theorem \ref{join} because both $N_{r_1},N_{r_2}\notin\mathcal G$. For $t>2$ we can consider $K_{r_1,\ldots,r_t}$ as the join of $N_{r_1}\vee K_{r_2,\ldots,r_t}$. Since $K_{r_2,\ldots,r_t}\in\mathcal G$, by Theorem \ref{join}, we have that $\edim(K_{r_1,\ldots,r_t})=\sum_{i=1}^tr_i-1$.
\end{proof}
\section{The lexicographic product}
For two graphs $G$ and $H$ the vertex set of the \emph{lexicographic product graph} $G[H]$ is $V(G)\times V(H)$. Two vertices $(g,h)$ and $(g',h')
$ are adjacent in the lexicographic product $G[H]$ if either $gg'\in
E(G)$ or ($g=g'$ and $hh'\in E(H)$). For $h\in V(H)$, we call the set $G^{h}=\{(g,h)\in V(G[H]):g\in V(G)\}$, a $G$ \emph{layer} throughout $h$ in
$G[H]$, and $^{g}H=\{(g,h)\in V(G[H]):h\in V(H)\}$ is an $H$ \emph{layer} throughout $g$ in $G[H]$. Note that the subgraph of $G[H]$ induced by $G^{h}$ is isomorphic to $G$, as well as the subgraph of $G[H]$ induced by $^{g}H$ is isomorphic to $H$. Note also that the lexicographic product is associative but not commutative, cf.~\cite{ImKl}. The map $p_{G}:V(G[H])\rightarrow V(G)$ defined with $p_{G}((g,h))=g$ is called a \emph{projection map onto} $G$. Similarly, we can define the \emph{projection map onto} $H$. We must remark also that the lexicographic product graph $G[H]$ is a connected if and only if $G$ is connected. It is well known that the distance between any two vertices $(g,h),(g',h')\in V(G[H])$ is given by
\begin{equation}\label{dist-lexico}
d_{G\circ H}((g,h),(g',h'))=\left\{\begin{array}{ll}
\min\{2,d_H(h,h')\}, & \mbox{if $g=g'$}, \\
d_G(g,g'), & \mbox{if $g\ne g'$}.
\end{array}
\right.
\end{equation}
Before we state the general result for the edge metric dimension of lexicographic product graphs we need to recall some well known concepts. Let $G$ be a graph. The vertex $u\in V(G)$ is called a \textit{false twin} of $v\in V(G)$ if $N_G(u)=N_G(v)$. Similarly, the vertices $x\in V(G)$ and $y\in V(G)$ are called \emph{true twins} if $N_G[x]=N_G[y]$. Clearly, each vertex is its own true twin and also its own false twin. Otherwise, different false twins $u$ and $v$ are not adjacent and if $x$ and $y$ are different true twins, then $x$ and $y$ are adjacent. It is easy to see that different false twins $u$ and $v$ are true twins only to themselves and vice versa, different true twins $x$ and $y$ are false twins only to themselves.
Both relations, being true or being false twins, are clearly equivalence relations. We are interested in all nontrivial equivalence classes of both relations, that is equivalence classes with at least two elements. Let $F=\{F_1,\ldots,F_k\}$ and $T=\{T_1,\ldots,T_{\ell}\}$ be the sets of all nontrivial equivalence classes of the false twin and true twin equivalence relations, respectively. Further, we need the number of elements in these classes, and therefore, we use notation $f(G)=\sum_{i=1}^k|F_i|$, $f'(G)=f(G)-k$, $t(G)=\sum_{i=1}^{\ell}|T_i|$ and $t'(G)=t(G)-\ell$.
The edges $e=uv$ and $e'=xy$ of a graph $G$ are \emph{twin edges} if $N_G[u]\cup N_G[v]=N_G[x]\cup N_G[y]$. If the twin edges $e=uv$ and $e'=xy$ are not incident with each other, then one can observe that vertices $u,v,x,y$ induce a subgraph that contains a four-cycle. Otherwise, if they are incident, say that $v=y$, then the condition $N_G[v]\subseteq N_G[u]\cup N_G[x]$ plays an important role. In particular, if there exists a neighbor $v$ of the false twins $u$ and $x$, such that $N_G[v]\subseteq N_G[u]\cup N_G[x]$, then the edges $uv$ and $xv$ are twin edges. Similarly, if $N_G[v]\subseteq N_G[u]\cup N_G[x]$ hold for a neighbor $v$ of the true twins $u$ and $x$, then again the edges $uv$ and $xv$ represent two twin edges.
As false twins, true twins, and twin edges play an important role while studying the edge metric dimension of lexicographic product graphs, we need to be careful to not count twice some of the vertices involved in the process. Let $G'$ be a graph obtained from a graph $G$ where we delete all vertices but one in every equivalence class of $F$ and of $T$. We call this operation \emph{twin deletion}. Let $Q=\{Q_1,\ldots,Q_m\}$ be nontrivial equivalence classes of the edge-twin relation of $G'$. We denote by $q(G')=\sum_{i=1}^m|Q_i|$ and $q'(G')=q(G')-m$ the number of nontrivial equivalence classes of the edge-twin relation.
It is easy to see that $t(K_n)=n, t'(K_n)=n-1$ and $f(K_n)=f'(K_n)=q(K'_n)=q'(K'_n)=0$. Similarly, for $m,n\geq 2$ we have that $f(K_{m,n})=m+n, f'(K_{m,n})=m+n-2$ and $t(K_{m,n})=t'(K_{m,n})=q(K'_{m,n})=q'(K'_{m,n})=0$. On the other hand, a graph $G$ obtained from a six-cycle $u_1u_2u_3u_4u_5u_6$ with two additional edges $u_2u_6$ and $u_3u_5$ has neither true nor false twins and we have $G'=G$. It is easy to see that $u_2u_3$ and $u_5u_6$ are twin edges and we have $q(G')=2$ and $q'(G')=1$. Another example is a graph $H$ that is obtained from vertices in $Q\cup N \cup Q'$, where the vertices of $Q$ and $Q'$ induce cliques while the set $N$ is an independent set of vertices and the cardinality of $Q,Q'$ and $N$ is at least two. In addition, we add all possible edges between $Q\cup Q'$ and $N$. The sets $Q$ and $Q'$ form two nontrivial equivalence classes of the true twin relation in $H$ and $N$ is the only nontrivial equivalence class of the false twin relation. Clearly, $H'$ is a path on three vertices $QNQ'$ and edges $QN$ and $NQ'$ are twins in $H'$.
We start this part of our exposition with a technical lemma that shows the independence of the values $q(G')$ and $q'(G')$ with respect to the deletion of true or false twins.
\begin{lemma}\label{technical}
Let $G$ be a graph. If the graphs $G'$ and $G''$ are obtained from $G$ by a twin deletion, then $q(G')=q(G'')$ and $q'(G')=q'(G'')$.
\end{lemma}
\begin{proof}
Let $uv$ and $xy$ be twin edges of a graph $G$. If none of the vertices $u,v,x,y$ is a true or a false twin, then $uv$ and $xy$ remain twin edges in $G'$ and in $G''$. If exactly one of $u,v,x,y$, say $u$, is a true or a false twin in $G$, then there exist $u'$ and $u''$ from the same equivalence class as $u$ which remain in $G'$ and in $G''$, respectively, after twin deletion process. Clearly, $u'v$ and $u''v$ are twin edges with $xy$ in $G'$ and $G''$, respectively. Similarly, if more than one vertex from $u,v,x,y$ are true or false twins, then we can always find their representatives in $G'$ and in $G''$ with the same properties as $u,v,x,y$. These representatives have the same properties in $G'$ and in $G''$ and the number of twin edges remains the same in both $G'$ and in $G''$. Therefore, $q(G')=q(G'')$ and $q'(G')=q'(G'')$ follows immediately.
\end{proof}
\begin{theorem}\label{lex1}
Let $G$ be any graph with at least three vertices in every component and let $H\ncong K_1$ be a graph. Then
$$\edim(G[H])\ge |V(G)|(|V(H)|-1)+f'(G)+t'(G)+q'(G').$$
Moreover, if $H\notin\mathcal G$, then
$$\edim(G[H])= |V(G)|(|V(H)|-1)+f'(G)+t'(G)+q'(G').$$
\end{theorem}
\begin{proof}
Let $G$ be a graph where every component contains at least three vertices and let $H$ be any graph on at least two vertices. Let $S$ be an edge metric basis for $G[H]$.
For an edge $gg'\in E(G)$ and vertices $h,h'\in V(H)$, $h\neq h'$, by $(\ref{dist-lexico})$, we can see that the edges $(g,h)(g',h)$ and $(g,h)(g',h')$ have the same distance to every vertex $(g_0,h_0)$ different from $(g',h)$ and $(g',h')$. Therefore, at least one vertex from $(g',h)$ and $(g',h')$ must be in $S$. Because $h,h'$ and $g'$ are arbitrarily taken, and $G$ contains no isolated vertices, we see that $S$ contains at least $|V(H)|-1$ vertices in each layer $^{g'}\!H$, for every $g'\in V(G)$.
Let now $g$ and $g'$ be different true or false twins of $G$ and let $g_0$ be any neighbor of $g$ and $g'$ different from $g$ and $g'$. Notice that $g_0$ exists because every component of $G$ contains at least three vertices. If both layers $^g\!H$ and $^{g'}\!H$ have exactly $|V(H)|-1$ vertices in $S$, then let $(g,h)$ and $(g',h')$ be such vertices outside of $S$. In such a case, the edges $(g,h)(g_0,h)$ and $(g',h')(g_0,h)$ have different distances only to the vertices $(g,h)$ and $(g',h')$, because of $(\ref{dist-lexico})$, and since $g$ and $g'$ are true or false twins. This is a contradiction with $S$ being an edge metric basis. Therefore at least one layer from $^g\!H$ and $^{g'}\!H$ must be entirely contained in $S$ for any pair of true or false twins. So, if $T_1,\ldots,T_{\ell}$ are nontrivial equivalence classes of the true twin relation, then at most one layer $^g\!H$, $g\in T_i$, is not entirely contained in $S$ for every $i\in\{1,\ldots,\ell\}$. Similarly, if $F_1,\ldots,F_k$ are the nontrivial equivalence classes of the false twin relation, then at most one layer $^g\!H$, $g\in F_i$, is not entirely contained in $S$ for every $i\in\{1,\ldots,k\}$. With this comments we arrive to the lower bound $\edim(G[H])\geq |V(G)|(|V(H)|-1)+f'(G)+t'(G)$.
We will now increase this bound in the case when $q'(G')>0$ ($G'$ is obtained from $G$ by the twin deletion previously described). We consider an edge metric generator $S$ of cardinality $|V(G)|(|V(H)|-1)+f'(G)+t'(G)$ as described in the previous paragraph. By Lemma \ref{technical}, for any nontrivial equivalence class of true or false twins, we can choose any vertex $g$ as the representative vertex for which $^g\!H$ is not entirely contained in $S$. Suppose now that $uv$ and $xy$ are twin edges of $G'$. By our choice of $G'$, there exist vertices $(u,h_1),(v,h_2),(x,h_3)$ and $(y,h_4)$ from $G[H]$ which are not in $S$. Note that the condition $N_{G'}[u]\cup N_{G'}[v]=N_{G'}[x]\cup N_{G'}[y]$ implies that $N_G[u]\cup N_G[v]=N_G[x]\cup N_G[y]$ holds as well because any true or false twin $z$ of $G$ is in $N_G[w]$ if and only if the representative of $z$ in $G'$ is in $N_{G'}[w]$. But then no vertex from $S$ distinguishes the edges $(u,h_1)(v,h_2)$ and $(x,h_3)(y,h_4)$, which is a contradiction. Therefore, at least one layer from $^u\!H,^v\!H,^x\!H$ and $^{y}\!H$ must be entirely contained in $S$ for any pair of twins $uv$ and $xy$ from $G'$. Moreover, if $uv$ and $xy$ are incident, say that $v=y$, then at least one layer from $^u\!H$ and $^x\!H$ must be entirely contained in $S$. In consequence, if $Q_1,\ldots,Q_m$ are nontrivial equivalence classes of the edge-twin relation of $G'$, then for at most one edge $wz\in Q_i$ both $^w\!H$ and $^z\!H$ are not entirely contained in $S$ for every $i\in\{1,\ldots,m\}$. From this facts, the lower bound $\edim(G[H])\geq |V(G)|(|V(H)|-1)+f'(G)+t'(G)+q'(G')$ follows.
We next show that this lower bound is exact when $H\notin\mathcal G$. So, suppose that $H\notin\mathcal G$ and let $h'$ be a vertex from $H$ such that for every edge $hh'$ there exists a vertex $x_h\in V(H)$ where $x_h$ is neither adjacent to $h$ nor to $h'$. Notice that $h'$ can also be an isolated vertex of $H$, but then we can take for $h$ any vertex different from $h'$ which exists by the assumption. For every nontrivial equivalence class $T_i$ of the true twin relation, we fix one vertex $t_i\in T_i$, $i\in\{1,\ldots,\ell\}$, and for every nontrivial equivalence class $F_j$ of the false twin relation, we also fix one vertex $f_i\in F_i$, $i\in\{1,\ldots,k\}$. Finally, let $Q_1,\ldots,Q_m$ be the nontrivial equivalence classes of the edge-twin relation of $G'$. We fix one edge $w_iz_i\in Q_i$ for every $i\in\{1,\ldots,m\}$ and for every other edge $u^j_iv^j_i\in Q_i$, $j\in\{1,\ldots,|Q_i|-1\}$, we fix a vertex which must be different from $w_i$ and from $z_i$. If the notation is chosen so that $u_i^j$ is this vertex, then we denote by $Q'_i$ the set $\{u_i^1,\ldots,u_i^{|Q_i|-1}\}$. We define the new sets $\mathcal T=\bigcup_{i=1}^{\ell}(T_i-\{t_i\})$, $\mathcal F=\bigcup_{i=1}^k(F_i-\{f_i\})$ and $\mathcal Q=\bigcup_{i=1}^m Q'_i$. We will show that the set
$$S=\left(V(G)\times (V(H)-\{h'\})\right)\cup\left((\mathcal T\cup\mathcal F\cup\mathcal Q)\times\{h'\}\right)$$
is an edge metric generator of cardinality $|V(G)|(|V(H)|-1)+f'(G)+t'(G)+q'(G')$. The equality $|S|=|V(G)|(|V(H)|-1)+f'(G)+t'(G)+q'(G')$ follows directly from the definitions of $\mathcal T, \mathcal F$ and $\mathcal Q$. To observe that $S$ is an edge metric generator, we need to check only pairs of edges $e_1$ and $e_2$ that have both end-vertices outside of $S$, or if $e_1$ and $e_2$ are incident, then the common end-vertex can be in $S$. If $e_1$ and $e_2$ (incident or not) have both end-vertices outside of $S$, then they must be lying over the layer $^{h'}\!H$. Let $e_1=(u,h')(v,h')$ and $e_2=(x,h')(y,h')$. By the definition of $S$ we have that $N_G[u]\cup N_G[v]\neq N_G[x]\cup N_G[y]$. Suppose without loss of generality that there exists $g\in (N_G[u]\cup N_G[v])-(N_G[x]\cup N_G[y])$. Clearly, $(g,h)\in S$ distinguishes $e_1$ and $e_2$ because $d_{G[H]}((g,h),e_1)=1$ and $d_{G[H]}((g,h),e_2)>1$ (recall that $h$ is different from $h'$ as they are adjacent or $h'$ is an isolated vertex). It remains that $e_1$ and $e_2$ are incident and that the common vertex is from $S$. Let now $e_1=(u,h')(v,h'')$ and $e_2=(x,h')(v,h'')$. If $u\neq v$ and $x\neq v$, then we can use the same argument that $N_G[u]\cup N_G[v]\neq N_G[x]\cup N_G[y]$ and we conclude as before. Thus we may assume that either $u=v$ or $x=v$ and in this case $h''$ must be adjacent to $h'$. By the symmetry we can assume that $u=v$. Recall that, since $H\notin\mathcal G$, there exists a vertex $x_{h''}$ that is nonadjacent to $h'$ and nonadjacent to $h''$. Clearly, $(u,x_{h''})\in S$ distinguishes $e_1$ and $e_2$ because $d_{G[H]}((u,x_{h''}),e_1)=2$ and $d_{G[H]}((u,x_{h''}),e_2)=1$. Because every pair of edges from $G[H]$ is distinguished by a vertex from $S$, we obtain that $\edim(G[H])\leq |V(G)|(|V(H)|-1)+f'(G)+t'(G)+q'(G')$ when $H\notin\mathcal G$ and the equality follows for this case.
\end{proof}
One could think that the bound given above is indeed an equality for any graphs $G$ and $H$ satisfying the statements of the theorem. However, this is not true, since other extra situations are also influencing the value of $\edim(G[H])$. We next comment some facts on this regard. To this end, we need the following terminology.
A vertex $v\in V(G)$ is called a \emph{satellite} of a vertex $u\in V(G)$ if $N_G[v]\subsetneq N_G[u]$. We also say that in such a case $u$ has a satellite $v$. If $v$ is a satellite of $u$, then $v$ and $u$ are adjacent. If $u$ has a false twin $v$, $u\neq v$, then $u$ cannot have a satellite, as any vertex $w$ adjacent to $u$ has $v$ in its closed neighborhood. On the other hand $u$, can be a satellite if it has a false twin $v$. If $x$ has a different true twin $y$, then $x$ can be a satellite of some vertex and can also have satellites. Similar as true twins, false twins and twin edges, vertices that have satellites are important for the edge metric dimension of the lexicographic product $G[H]$ when $H\in \mathcal G$.
\begin{lemma}\label{satellite}
Let $G$ and $H$ be any graphs, where every component of $G$ contains at least three vertices and $H\in\mathcal G$, and let $S$ be an edge metric basis for $G[H]$. If $g\in V(G)$ has a satellite $g'\in V(G)$, then $^{g}\!H$ or $^{g'}\!H$ is entirely contained in $S$.
\end{lemma}
\begin{proof}
By the proof of Theorem \ref{lex1}, there are at least $|V(H)|-1$ vertices of $^{g}\!H$ and of $^{g'}\!H$ in $S$. Suppose that, on the way to a contradiction, $(g,h)$ and $(g',h')$ do not belong to $S$. Because $H\in\mathcal G$, there exists a vertex $h_0\in V(H)$ that is adjacent to $h$ and $\{h,h_0\}$ is a $\gamma_t(H)$-set. We will see that edges $e_1=(g,h)(g,h_0)$ and $e_2=(g',h')(g,h_0)$ are not distinguished by any vertex form $S$. First, every vertex from $^{g}\!H$, with the exception of $(g,h)$ and $(g,h_0)$, is at distance 1 to both $e_1$ and $e_2$, because $\{h,h_0\}$ is a $\gamma_t(H)$-set and $gg'\in E(G)$. Second, the vertices from $^{g_0}\!H$ for every $g_0\in N_G[g']$, with the exception of $(g',h')$, are at distance 1 to $e_1$ and to $e_2$, because $g'$ is a satellite of $g$. Third, and finally, the other vertices from $G[H]$ are at the same distance to vertices of $^{g}\!H$, and so, also to $e_1$ and to $e_2$. Thus, the only vertices that distinguish $e_1$ and $e_2$ are $(g,h)$ and $(g',h')$ which are not in $S$, and this is a contradiction. Therefore, $^{g}\!H$ or $^{g'}\!H$ is entirely contained in $S$.
\end{proof}
This last lemma is one of the reasons causing that the bound from Theorem \ref{lex1} does not in general hold as equality, when $H\in\mathcal G$. We can expect that, if there are some satellite vertices, then one needs to add some additional vertices to a given set to become an edge metric generator. Again we observe that we need to be careful not to count twice some of them. For example, observe $K_p\vee N_r$. Every vertex from $N_r$ is a satellite from every vertex from $K_p$ and one can expect that $\min\{p,r\}$ of vertices need to be added to a given set to get an edge metric generator for some $H\in\mathcal G$, which yields a kind of minimization problem. However, this is not the right approach, as we have already seen in the second paragraph of the proof of Theorem \ref{lex1}, where all $H$-layers initiated by true twins (with one possible exception in every equivalence class), and by false twins (with one possible exception in every equivalence class) must belong to a given edge metric basis.
\section{The corona product}
Let $G$ and $H$ be two graphs of order $n_1$ and $n_2$, respectively. The corona product graph $G\odot H$ is defined as the graph obtained from $G$ and $H$, by taking one copy of $G$ and $n_1$ copies of $H$ and joining by an edge every vertex from the $i^{th}$-copy of $H$ with the $i^{th}$-vertex of $G$. Given a vertex $g\in V(G)$, the copy of $H$ whose vertices are adjacent to $g$ is denoted by $H_g$. We will first analyze the situation in which the second factor of this product is not isomorphic to the singleton graph $K_1$.
\begin{theorem}
\label{corona} For any graphs $G$ and $H$ where $G$ is connected and $|V(H)|\geq 2$,
$$\edim(G\odot H)=|V(G)|\cdot(|V(H)|-1)$$
\end{theorem}
\begin{proof}
Let $G$ and $H$ be any graphs and let $n=|V(H)|\geq 2$. Let $g\in V(G)$ and let $\{g_1,\ldots,g_n\}$ be the set of vertices of the copy $H_g$ of $H$. Any two edges $gg_i$ and $gg_j$, $i\neq j$, have the same distance to all vertices form $V(G\odot H)-\{g_i,g_j\}$. Therefore, at least one of them must be in any edge metric basis of $G\odot H$. Because $i,j$ and $g$, with $1\leq i<j\leq n$, are arbitrary, we see that every edge metric basis of $G\odot H$ must contain at least $n-1$ vertices from every copy $H_g$ of $H$ in $G\odot H$. This yields the lower bound $\edim(G\odot H)\geq |V(G)|\cdot(|V(H)|-1)$.
On the other hand, let $H_g$ be a copy of $H$ corresponding to a vertex $g\in V(G)$. We will show that the set $S=\cup_{g\in V(G)} (V(H_g)-\{h\})$ is an edge metric generator for $G\odot H$, where $h$ is an arbitrary vertex of $V(H_g)$. First notice that $S$ is nonempty because $n\geq 2$, and by the same reason in every copy of $H$ exists at least one vertex from $S$. Any two different edges from one copy of $H$ are distinguished by some vertex in $S$, because at least two end vertices of these two edges are in $S$. The same argument also holds if both edges are in two different copies of $H$. Similarly, an edge from $G$ and one from any copy of $H$ are distinguished by at least one end-vertex of the edge lying in the copy of $H$ which is in $S$. Two different edges from $G$ are distinguished by at least one vertex $g\in V(G)$ and so, any vertex in $S\cap V(H_g)$ distinguishes these two edges. So, let now consider one edge $gh'$ with $g\in V(G)$ and $h'\in V(H_g)$. If the second edge is $gh_1$ for some $h_1\in V(H_g)$ and $h_1\neq h'$, then these two edges are distinguish by $h'$ or by $h_1$. If the second edge $g'g''$ belongs to $G$, then at least one end-vertex, say $g'$, is different from $g$ and they are distinguished by any vertex from $S\cap V(H_{g'})$. Any edge $gh'$ is also clearly distinguished from any edge with at least one end-vertex in other copy of $H$ by any vertex from $S$ in that copy. So, let finally the second edge $h_1h_2$ be from $H_g$. Any vertex $x\in S\cap H_{g'}$, where $g'\ne g$, distinguishes $gh'$ and $h_1h_2$ because $d_{G\odot H}(h_1h_2,x)=d_{G\odot H}(hg,x)+1$. Therefore, we have $\edim(G\odot H)\le |V(G)|\cdot(|V(H)|-1)$ whenever $n\geq 2$, and the equality follows.
\end{proof}
In contrast with the case above, the corona product graph $G\odot K_1$ is in general complicated to deal with. In order to observe this, the following terminology and notation will be required. A vertex of degree at least $3$ in a tree $T$ will be called a \emph{major vertex} of $T$.
Any leaf $u$ of $T$ is said to be a \emph{terminal vertex} of a major vertex $v$ of $T$ if
$d(u, v)<d(u,w)$ for every other major vertex $w$ of $T$. The \emph{terminal degree} of
a major vertex $v$ is the number of terminal vertices of $v$. A major vertex $v$ of $T$ is an
\emph{exterior major vertex} of $T$ if it has positive terminal degree. Let $n_1(T)$ denote the number of leaves of $T$, and let $ex(T)$ denote the number of exterior major vertices of $T$. We can now state the formula for the edge metric dimension of a tree given in \cite{KeTrYe}. If $T$ is a tree that is not a path, then
\begin{equation}\label{formula-trees}\edim(T) = n_1(T) - ex(T).\end{equation}
Some situations can be easily deduced for $G\odot K_1$. For instance, if $G$ is the path $P_2$, then clearly $G\odot K_1\cong P_4$ and so, $\edim(P_2\odot K_1)=\edim(P_4)=1$. Also, if $G$ is a path of order $n\ge 3$, then $G\odot K_1$ is a tree such that $n_1(G\odot K_1)=n$ and $ex(G\odot K_1)=n-2$. Thus, from (\ref{formula-trees}) we get $\edim(P_n\odot K_1)=2$.
\begin{theorem}
For any graph $G$, $\edim(G\odot K_1)\ge \edim(G)$, and this bound is sharp.
\end{theorem}
\begin{proof}
Let $V(G)=\{g_1,\dots,g_n\}$ and for every $g_i\in V(G)$, let $u_i$ be the vertex adjacent to $g_i$ corresponding to the copy of $K_1$ used in the corona product. Assume $S$ is an edge metric basis for $G\odot K_1$, and consider the set of vertices $S'=\{g_i\,:\,\{g_i,u_i\}\cap S\ne \emptyset\}$. We will prove that $S'$ is an edge metric generator for $G$. Let $e,f\in E(G)$ be any two edges and let $x\in S$ such that $d_G(e,x)\ne d_G(f,x)$. If $x\in V(G)$, then $x\in S'$ and so, $x$ determines $e$ and $f$. If $x\notin V(G)$, then $x$ is a vertex corresponding to a copy of $K_1$ and is adjacent to a vertex $x'\in V(G)$ which is also in $S'$. Thus, $d_G(e,x')=d_{G\odot K_1}(e,x)-1\ne d_G(f,x)-1=d_{G\odot K_1}(f,x')$ and so, $x'$ distinguish $e$ and $f$. As a consequence, $S'$ is an edge metric generator for $G$ and the bound follows.
To see the sharpness of the bound we consider the graph $K_n\odot K_1$ where $n\ge 3$. Let $V(K_n)=\{g_1,\dots,g_n\}$ and, as above, for every $g_i\in V(K_n)$, let $u_i$ be the vertex adjacent to $g_i$ corresponding to the copy of $K_1$ used in the corona product. Assume $S$ is an edge metric basis for $K_n$ (note that an edge metric basis of $K_n$ is form by any set of $n-1$ vertices of $K_n$), and consider the set of vertices $S'=\{u_i\,:\,g_i\in S\}$. We shall prove $S'$ is a edge metric generator for $K_n\odot K_1$. Let $e,f\in E(K_n\odot K_1)$ be any two edges and consider the following situations.
\begin{itemize}[leftmargin=*]
\item $e,f\in E(K_n)$. Since there exists a vertex $g_j\in S$ such that $d_{K_n}(e,g_j)\ne d_{K_n}(f,g_j)$, we deduce that
$d_{K_n\odot K_1}(e,u_j)=d_{K_n}(e,g_j)+1\ne d_{K_n}(f,g_j)+1=d_{K_n\odot K_1}(f,u_j)$. As $u_j\in S'$, we have that $u_j$ recognizes $e,f$.
\item $e\notin E(K_n)$ and $f\in E(K_n)$. Let $e=g_iu_i$ and $f=g_jg_k$. If $u_i\in S'$, then clearly $e,f$ are identified by $u_i$. If $u_i\notin S'$ and ($u_j\in S'$ or $u_k\in S'$), say $u_j\in S'$, then $e,f$ are distinguished by $u_j$. Since the case whether $u_i,u_j,u_k\notin S'$ is not possible, because there is only one vertex of $V(K_n)$ which is not in $S$, we are done with this case.
\item $e,f\notin E(K_n)$. Let $e=g_iu_i$ and $f=g_ju_j$. Since there is only one vertex of $V(K_n)$ which is not in $S$, it must happen $u_i\in S'$ or $u_j\in S'$. Thus, $e,f$ are distinguished by $u_i$ or by $u_j$.
\end{itemize}
As a consequence of the cases above we obtain that $S'$ is an edge metric generator for $K_n\odot K_1$ and so, $\edim(K_n\odot K_1)\le \edim(K_n)$. Since $\edim(G\odot K_1)\ge \edim(G)$ for any graph $G$, we obtain the equality $\edim(K_n\odot K_1)=\edim(K_n)$ and the sharpness of the bound is completed.
\end{proof}
Although the bound above is tight, it is possible to observe that the difference between $\edim(G\odot K_1)$ and $\edim(G)$ can be arbitrarily large. To observe this, we consider a tree $T$ of order $n\ge 3$ which is not a path. Clearly, $n_1(T\odot K_1)=n$ and $ex(T\odot K_1)=n-n_1(T)$. As a consequence of (\ref{formula-trees}), $\edim(T\odot K_1)=n_1(T\odot K_1)-ex(T\odot K_1)=n-(n-n_1(T))=n_1(T)$ and so, $\edim(T\odot K_1)-\edim(T) = n_1(T)-(n_1(T)-ex(T))=ex(T)$, which can be as large as we would require.
|
\section{Introduction}
\label{sec:intro}
Objects in a triangulated category, such as equivariant KK-categories,
may be classified up to isomorphism using a primary homological
invariant and a secondary ``obstruction class'' provided they have a
projective resolution of length~\(2\)
in a suitable sense. This method was applied
in~\cite{Bentmann-Meyer:More_general} to objects in the
circle-equivariant KK-category \(\KKcat^\mathbb T\),
the equivariant KK-category \(\KKcat^X\)
for \(\textup C^*\)\nb-algebras
over a finite, unique path space~\(X\),
and graph \(\textup C^*\)\nb-algebras
with finite ideal lattice. Here we compute the obstruction classes
that occur in this classification. This makes the classification for
objects of \(\KKcat^\mathbb T\) and~\(\KKcat^X\) more explicit.
The result suggests, in fact, a different invariant for objects in
these categories that is fine enough to admit a Universal Coefficient
Theorem. The price to pay is that the invariant uses bivariant
\(\textup{K}\)\nb-theory
instead of ordinary \(\textup{K}\)\nb-theory.
We explain this for the equivariant bootstrap class
in the category~\(\KKcat^\mathbb Z\);
the latter is equivalent to~\(\KKcat^\mathbb T\) by Baaj--Skandalis duality.
An object of~\(\KKcat^\mathbb Z\)
is a \(\textup C^*\)\nb-algebra~\(A\)
with a \(\mathbb Z\)\nb-action,
which is generated by a single automorphism~\(\alpha\).
The most obvious homological invariant on~\(\KKcat^\mathbb Z\)
maps this to the \(\textup{K}\)\nb-theory
\(\textup{K}_*(A)\)
with the module structure over the group ring \(\mathbb Z[x,x^{-1}]\)
of~\(\mathbb Z\)
that is induced by~\(\alpha\).
Let~\(\Abel\)
be the category of countable, \(\mathbb Z/2\)\nb-graded
modules over \(\mathbb Z[x,x^{-1}]\).
This is a stable Abelian category. The \(\textup{K}\)\nb-theory
described above defines a stable homological functor
\(F^\mathbb Z_\textup{K}\colon \KKcat^\mathbb Z \to \Abel\).
The category~\(\Abel\)
has cohomological dimension~\(2\),
that is, any object has a projective resolution of length~\(2\).
Let~\(\Abel\delta\)
be the additive category of pairs~\((A,\delta)\)
with \(A\mathrel{\in\in}\Abel\)
and \(\delta\in\Ext^2_{\Abel}(\Sigma A, A)\);
morphisms from \((A,\delta)\)
to \((A',\delta')\)
in~\(\Abel\delta\)
are morphisms~\(f\)
from~\(A\)
to~\(A'\)
in~\(\Abel\)
with \(\delta' f=f\delta\).
It is shown in~\cite{Bentmann-Meyer:More_general} that isomorphism
classes of objects in the bootstrap class in~\(\KKcat^\mathbb Z\)
are in bijection with isomorphism classes of objects
in~\(\Abel\delta\).
In particular, any object~\(A\)
of~\(\Abel\)
lifts to an object in the bootstrap class in~\(\KKcat^\mathbb Z\).
The lifting is, however, not unique: different liftings of~\(A\)
are in bijection with \(\Ext^2_{\Abel}(\Sigma A, A)\);
here two liftings \(B_1,B_2\)
are identified if there is an isomorphism \(B_1\cong B_2\)
that induces the identity map on \(A=F^\mathbb Z_\textup{K}(B_1)=F^\mathbb Z_\textup{K}(B_2)\).
In this article, we compute the obstruction class of an object
of~\(\KKcat^\mathbb Z\)
explicitly. This question remained open
in~\cite{Bentmann-Meyer:More_general}. Let \((A,\alpha)\)
be a \(\textup C^*\)\nb-algebra
with an automorphism~\(\alpha\)
as above. The exact sequence in the Universal Coefficient Theorem
splits, so that
\[
\textup{KK}_0(A,A) \cong
\Hom\bigl(\textup{K}_*(A),\textup{K}_*(A)\bigr)
\oplus \Ext^1\bigl(\textup{K}_{*+1}(A),\textup{K}_*(A)\bigr).
\]
While the splitting above is not natural, it is canonical enough to
associate well defined elements
\(\alpha_0\in\Hom\bigl(\textup{K}_*(A),\textup{K}_*(A)\bigr)\)
and \(\alpha_1\in \Ext^1\bigl(\textup{K}_{*+1}(A),\textup{K}_*(A)\bigr)\)
to~\(\alpha\).
Here~\(\alpha_0\)
is the action on \(\textup{K}_*(A)\)
induced by~\(\alpha\),
which is part of the \(\mathbb Z[x,x^{-1}]\)-module
structure on~\(F^\mathbb Z_\textup{K}(A)\).
We show how to compute the obstruction class from~\(\alpha_1\).
Namely, it is~\(\mu^*(\alpha_1)\) for a canonical map
\[
\mu^*\colon \Ext^1\bigl(\textup{K}_{*+1}(A),\textup{K}_*(A)\bigr) \to
\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\textup{K}_{*+1}(A),\textup{K}_*(A)\bigr).
\]
Thus an object \((A,\alpha)\)
of \(\KKcat^\mathbb Z\)
is determined uniquely up to isomorphism by~\(A\)
as an object of~\(\KKcat\)
together with the class~\([\alpha]\)
of~\(\alpha\)
in \(\textup{KK}_0(A,A)\).
We treat the pair \((A,[\alpha])\)
as an object of a certain exact category~\(\KKcat[\mathbb Z]\)
with a suspension automorphism. The forgetful functor
\(F^\mathbb Z\colon \KKcat^\mathbb Z \to \KKcat[\mathbb Z]\)
is a stable homological functor. The fact that it gives a complete
invariant for objects of~\(\KKcat^\mathbb Z\)
follows from a Universal Coefficient Theorem computing
\(\textup{KK}^\mathbb Z_*(A,B)\).
This Universal Coefficient Theorem is already proven
in~\cite{Meyer-Nest:Homology_in_KK}, and it is shown there to be
equivalent to a Pimsner--Voiculescu like exact sequence for
\(\textup{KK}^\mathbb Z_*(A,B)\).
The Universal Coefficient Theorem for the invariant
\(F^\mathbb Z\colon \KKcat^\mathbb Z \to \KKcat[\mathbb Z]\)
is based on projective resolutions of length~\(1\)
for the stable homological ideal \(\ker F^\mathbb Z\).
We use these resolutions to compute the obstruction class related to
the homological invariant \(F^\mathbb Z_\textup{K}(A) \mathrel{\vcentcolon=} \textup{K}_*(A)\).
Namely, let \(0\to B_1 \to B_0\to A \to 0\)
be a projective resolution of length~\(1\)
for the ideal \(\ker F^\mathbb Z\)
in \(\KKcat^\mathbb Z\).
This remains exact with respect to the larger stable homological ideal
\(\ker F^\mathbb Z_\textup{K}\).
The objects \(B_1\)
and~\(B_0\)
are no longer projective for \(\ker F^\mathbb Z_\textup{K}\),
but they have projective resolutions of length~\(1\).
Thus \(\textup{KK}^\mathbb Z_0(B_1,B_0)\)
is computed by a Universal Coefficient Theorem, which splits
non-naturally. So the arrow \(B_1\to B_0\)
gives a class in \(\Ext^1_{\mathbb Z[x,x^{-1}]}\bigl(\textup{K}_{*+1}(B_1),\textup{K}_*(B_0)\bigr)\).
It must determine the obstruction class of~\(A\)
because~\(A\)
is \(\textup{KK}^\mathbb Z\)-equivalent
to the cone of the map \(B_1\to B_0\).
The result is the formula for the obstruction class of~\(A\)
asserted above. This computation of obstruction classes is carried
out in Section~\ref{sec:triangulated} in the same generality as
in~\cite{Bentmann-Meyer:More_general}, using a triangulated category
with a stable homological ideal with enough projectives.
The Universal Coefficient Theorem for \(\KKcat^\mathbb Z\)
is developed in Section~\ref{sec:UCT_Z}. This is followed by an
analogous treatment for the category of \(\textup C^*\)\nb-algebras
over a unique path space~\(X\)
in Section~\ref{sec:UCT_X}. In
Section~\ref{sec:compare_classification}, we compare the
classification theorems that our two invariants give for objects in
the bootstrap classes in \(\KKcat^\mathbb Z\)
and \(\KKcat^X\).
In Section~\ref{sec:filtrated_K}, we study \(\KKcat^X\)
in the special case where~\(X\)
is totally ordered. Then filtrated \(\textup{K}\)\nb-theory
is a third complete invariant for the same objects
(see~\cite{Meyer-Nest:Filtrated_K}), and we compare this invariant
with the new ones. This is a rather long homological computation. In
Section~\ref{sec:extensions}, the result of
Section~\ref{sec:filtrated_K} is formulated more nicely if~\(X\)
has only two points, so that \(\textup C^*\)\nb-algebras
over~\(X\) are extensions of \(\textup C^*\)\nb-algebras.
\section{Universal Coefficient Theorems with KK-valued functors}
\label{sec:KK-valued_UCT}
This section develops Universal Coefficient Theorems for the
equivariant Kasparov categories \(\KKcat^\mathbb Z\)
and \(\KKcat^X\)
for a unique path space~\(X\),
which are based on forgetful functors to~\(\KKcat\).
We shall use the general framework developed
in~\cite{Meyer-Nest:Homology_in_KK} for doing relative homological
algebra in triangulated categories. The starting point for this is a
stable homological ideal~\(\Ideal\)
in a triangulated category~\(\Tri\).
Being stable and homological means that there is a stable homological
functor \(F\colon \Tri \to \Abel\)
to some stable Abelian category~\(\Abel\) with
\[
\Ideal(A,B) = \ker F(A,B)
\mathrel{\vcentcolon=} \setgiven{g\in\Tri(A,B)}{F(g)=0}.
\]
Here stability of \(\Abel\) and~\(F\)
means that~\(\Abel\)
is equipped with a suspension automorphism and that
\(F\circ \Sigma_{\Tri} = \Sigma_{\Abel} \circ F\).
A stable homological ideal~\(\Ideal\)
allows to carry over various notions from homological algebra
to~\(\Tri\).
In particular, there are \(\Ideal\)\nb-exact
chain complexes, \(\Ideal\)\nb-projective
objects and \(\Ideal\)\nb-projective
resolutions in~\(\Tri\),
which then allow to define \(\Ideal\)\nb-derived
functors. We shall only be interested in cases where there are enough
\(\Ideal\)\nb-projective
objects. The thick subcategory~\(\gen{\Proj_\Ideal}\)
generated by the \(\Ideal\)\nb-projective
objects is an analogue of the ``bootstrap class'' in Kasparov theory.
If \(A\mathrel{\in\in} \gen{\Proj_\Ideal}\)
has an \(\Ideal\)\nb-projective
resolution of length~\(1\),
then the graded group \(\Tri_*(A,B)\)
for \(B\mathrel{\in\in}\Tri\)
may be computed by a Universal Coefficient Theorem (UCT). The Hom and
Ext groups in this UCT are those in a certain Abelian category,
namely, the target category of the universal \(\Ideal\)\nb-exact
stable homological functor~\(F^u\).
The functor~\(F^u\)
strongly classifies objects of~\(\gen{\Proj_\Ideal}\)
with an \(\Ideal\)\nb-projective
resolution of length~\(1\).
That is, any isomorphism \(F^u(A) \cong F^u(B)\)
between such objects is induced by an isomorphism \(A\cong B\)
in~\(\Tri\).
In many cases of interest, the universal functor~\(F^u\)
is quite explicit. The following theorem allows to recognise it:
\begin{theorem}[\cite{Meyer-Nest:Homology_in_KK}*{Theorem~57}]
\label{the:characterise_universal_homological}
Let~\(\Tri\)
be a triangulated category, let \(\Ideal \subseteq \Tri\)
be a stable homological ideal, and let \(F\colon \Tri\to\Abel\)
be an \(\Ideal\)\nb-exact
stable homological functor into a stable Abelian category~\(\Abel\).
Let~\(\Proj\Abel\)
be the class of projective objects in~\(\Abel\).
Suppose that idempotent morphisms in \(\Tri\)
split. The functor~\(F\)
is the universal \(\Ideal\)\nb-exact
stable homological functor to a stable Abelian category and there
are enough \(\Ideal\)\nb-projective
objects in~\(\Tri\) if and only if
\begin{enumerate}
\item \(\Abel\) has enough projective objects;
\item the adjoint functor~\(F^\lad\)
of~\(F\) is defined on~\(\Proj\Abel\);
\item \(F\circ F^\lad(A) \cong A\) for all \(A\mathrel{\in\in}\Proj\Abel\).
\end{enumerate}
\end{theorem}
The ideals to be treated in this section are defined as the kernel on
morphisms of a triangulated functor \(F\colon \Tri\to\Tri[S]\)
to another triangulated category~\(\Tri[S]\).
Ideals of this form are already treated
in~\cite{Meyer-Nest:Homology_in_KK}, using a construction of Freyd to
embed~\(\Tri[S]\)
into an Abelian category. Here we unify these two cases by allowing
the functor~\(F\)
to take values in an \emph{exact category} (see \cites{Buehler:Exact,
Keller:Handbook}). An exact category is an additive category with
a chosen class of ``admissible'' extensions. A typical example is a
full, additive subcategory of an Abelian category that is closed
under extensions. An Abelian category~\(\Abel\)
is exact where all extensions are admissible. And a triangulated
category~\(\Tri[S]\)
is exact where only split extensions are admissible. So exact
categories contain the two cases treated
in~\cite{Meyer-Nest:Homology_in_KK}. In many examples of
classification results using a UCT, the range of the universal
\(\Ideal\)\nb-exact
functor~\(F^u_\mathrm{old}\)
is a certain exact subcategory of a module category, and only objects
in this exact subcategory have projective resolutions of length~\(1\).
This holds for the UCTs in \cites{Bentmann:Thesis,
Bentmann-Koehler:UCT, Koehler:Thesis, Meyer-Nest:Filtrated_K} (see
Remark~\ref{rem:exact_NT} for one of these cases). The phenomenon is
understood better in~\cite{Ambrogio-Stevenson-Stovicek:Gorenstein}.
Quillen's Embedding Theorem allows to embed any exact category into an
Abelian one in a fully faithful, fully exact way. So all results in
the general theory carry over to the case where~\(F\)
takes values in an exact category. In particular, we may modify the
definition of the universal \(\Ideal\)\nb-exact
stable homological functor by allowing exact categories as target.
Let \(F^u_\mathrm{old}\colon \Tri\to\Abel\)
be the universal \(\Ideal\)\nb-exact
functor into an Abelian category as in
Theorem~\ref{the:characterise_universal_homological}. Let
\(\Abel[E]\subseteq\Abel\)
be the exact subcategory of~\(\Abel\)
generated by the range of~\(F^u_\mathrm{old}\).
Let \(F^u\colon \Tri\to\Abel[E]\)
be~\(F^u_\mathrm{old}\)
viewed as a functor to~\(\Abel[E]\).
This is the universal \(\Ideal\)\nb-exact
functor in the new sense. Its universal property follows from the one
of \(F^u_\mathrm{old}\)
and Quillen's Embedding Theorem. If~\(\Ideal\)
is defined by a forgetful functor \(F\colon \Tri\to\Tri[S]\)
to another triangulated category, then the universal homological
functor to an Abelian category typically uses Freyd's embedding
of~\(\Tri[S]\)
into an Abelian category. In contrast, we shall see that in many
examples the universal \(\ker F\)\nb-exact
functor to an exact category gives an exact category that is very
closely related to~\(\Tri[S]\).
\begin{remark}
\label{rem:universal_to_exact}
Assume that all objects in~\(\Tri\)
have an \(\Ideal\)\nb-projective
resolution of finite length. Then the image of~\(F^u_\mathrm{old}\)
is contained in the subcategory of objects in~\(\Abel\)
with a finite-length projective resolution. This subcategory is
exact. And it is already generated as an exact category by the
projective objects in~\(\Abel\):
this is proved by induction on the length of a resolution. Hence
the image of~\(F^u_\mathrm{new}\)
must be equal to the subcategory of objects with a finite-length
projective resolution. So \(F^u_\mathrm{new} = F^u_\mathrm{old}\)
if and only if \emph{all} objects of~\(\Abel\)
have a finite-length projective resolution.
\end{remark}
\subsection{Actions of the group of integers}
\label{sec:UCT_Z}
We now treat a concrete example, namely, the case
\(\Tri\mathrel{\vcentcolon=} \KKcat^\mathbb Z\).
This is a triangulated category. Its objects are pairs \((A,\alpha)\)
with a separable \(\textup C^*\)\nb-algebra~\(A\)
and \(\alpha\in\Aut(A)\);
the latter generates an action of~\(\mathbb Z\)
on~\(A\)
by automorphisms. The arrows in \(\KKcat^\mathbb Z\)
are the Kasparov groups \(\textup{KK}^\mathbb Z_0(A,B)\),
and the composition is the Kasparov product; we have dropped the
automorphisms from our notation as usual to avoid clutter. The
triangulated category structure on~\(\KKcat^\mathbb Z\)
is described in \cite{Meyer-Nest:BC}*{Appendix}. The relative
homological algebra in~\(\KKcat^\mathbb Z\)
is already studied in~\cite{Meyer-Nest:Homology_in_KK}. The main
result is the following variant of the Pimsner--Voiculescu sequence
for \(\textup{KK}^\mathbb Z_*(A,B)\).
Let \(A\)
and~\(B\)
be \(\textup C^*\)\nb-algebras
with automorphisms \(\alpha\)
and~\(\beta\), respectively. Then there is an exact sequence
\begin{equation}
\label{eq:PV-sequence}
\begin{tikzcd}[column sep=huge]
\textup{KK}_1(A,B) \arrow[r] &
\textup{KK}_0^\mathbb Z(A,B) \arrow[r, "\mathrm{forget}"] &
\textup{KK}_0(A,B) \arrow[d, "\alpha^*\beta_*^{-1}-1"] \\
\textup{KK}_1(A,B) \arrow[u, "\alpha^*\beta_*^{-1}-1"] &
\textup{KK}_1^\mathbb Z(A,B) \arrow[l, "\mathrm{forget}"] &
\textup{KK}_0(A,B) \arrow[l]
\end{tikzcd}
\end{equation}
(see \cite{Meyer-Nest:Homology_in_KK}*{Section~5.1}); here
\((\alpha^*\beta_*^{-1}-1)(x) = \beta^{-1}\circ x\circ\alpha - x\)
for all \(x\in \textup{KK}_*(A,B)\).
We may rewrite this as a pair of short exact sequences
\begin{multline*}
\coker \bigl(\alpha^*\beta_*^{-1}-1\colon
\textup{KK}_{1+i}(A,B) \to \textup{KK}_{1+i}(A,B)\bigr)
\rightarrowtail \textup{KK}_i^\mathbb Z(A,B)
\\ \twoheadrightarrow
\ker \bigl(\alpha^*\beta_*^{-1}-1\colon \textup{KK}_i(A,B) \to \textup{KK}_i(A,B)\bigr)
\end{multline*}
for \(i=0,1\).
We shall explain that the long exact sequence~\eqref{eq:PV-sequence}
is an instance of a Universal Coefficient Theorem as in
\cite{Meyer-Nest:Homology_in_KK}*{Theorem~66}. In particular, the
cokernel and kernel in it are the Ext and Hom groups in a certain
exact category. This is already proven
in~\cite{Meyer-Nest:Homology_in_KK}, but the consequences for
classification are not explored there. We shall also treat other
examples by the same method later.
The forgetful functor
\[
R^\mathbb Z\colon \KKcat^\mathbb Z\to\KKcat,\qquad
(A,\alpha)\mapsto A,
\]
is triangulated. So its kernel on morphisms
\[
\Ideal^\mathbb Z(A,B) \mathrel{\vcentcolon=} \setgiven{f\in\textup{KK}_0^\mathbb Z(A,B)}{R^\mathbb Z(f)=0
\text{ in }\textup{KK}_0(A,B)}
\]
is a stable homological ideal in~\(\KKcat^\mathbb Z\).
We now describe the universal \(\Ideal^\mathbb Z\)\nb-exact
functor. The main point here is to describe its target category. Let
\(\KKcat[\mathbb Z]\)
be the additive category of functors \(\mathbb Z\to\KKcat\)
with natural transformations as arrows. Equivalently, an object
of~\(\KKcat[\mathbb Z]\)
is an object \(A\mathrel{\in\in}\KKcat\)
with a group homomorphism \(a\colon \mathbb Z\to\textup{KK}_0(A,A)^\times\),
\(n\mapsto a_n\),
where \(\textup{KK}_0(A,A)^\times\)
denotes the multiplicative group of invertible elements in the ring
\(\textup{KK}_0(A,A)\).
And an arrow \((A,a) \to (B,b)\)
in \(\KKcat[\mathbb Z]\)
is an arrow \(f\in \textup{KK}_0(A,B)\)
that satisfies \(b_n\circ f = f\circ a_n\)
for all \(n\in\mathbb Z\).
The homomorphism~\(a\)
is determined by its value at \(1\in\mathbb Z\),
which may be any element \(a_1\in \textup{KK}_0(A,A)^\times\).
So we also denote objects of~\(\KKcat[\mathbb Z]\)
as \((A,a_1)\).
An element \(f\in \textup{KK}_0(A,B)\)
is an arrow \((A,a) \to (B,b)\)
in \(\KKcat[\mathbb Z]\)
if and only if \(b_1\circ f = f\circ a_1\)
or, equivalently, \(b_1^{-1}\circ f \circ a_1 - f=0\). Thus
\[
\ker \bigl(\alpha^*\beta_*^{-1}-1\colon \textup{KK}_0(A,B) \to \textup{KK}_0(A,B)\bigr)
\cong \Hom_{\KKcat[\mathbb Z]}\bigl((A,[\alpha]),(B,[\beta])\bigr).
\]
Here we have implicitly used the functor
\[
F^\mathbb Z\colon \KKcat^\mathbb Z\to \KKcat[\mathbb Z],\qquad
(A,\alpha) \mapsto (A,[\alpha]),
\]
where \([\alpha]\in \textup{KK}_0(A,A)^\times\)
is the \(\KKcat\)\nb-class
of the automorphism~\(\alpha\).
It has the same kernel on morphisms~\(\Ideal^\mathbb Z\)
as the forgetful functor~\(R^\mathbb Z\).
We call a kernel--cokernel pair \(K \rightarrowtail E \twoheadrightarrow Q\)
in \(\KKcat[\mathbb Z]\)
\emph{admissible} if it splits in~\(\KKcat\),
that is, \(E\cong K\oplus Q\)
as objects of~\(\KKcat\).
This turns \(\KKcat[\mathbb Z]\)
into an exact category (see \cite{Buehler:Exact}*{Exercise~5.3}). We
use this exact structure to define projective objects and projective
resolutions in \(\KKcat[\mathbb Z]\).
Given \(A\mathrel{\in\in}\KKcat\)
and a free Abelian group \(G= \mathbb Z[I]\)
on a countable set~\(I\), we define \(A\otimes G\mathrel{\in\in}\KKcat\) by
\begin{equation}
\label{def:tensor_with_Cstar}
A \otimes G \mathrel{\vcentcolon=} \bigoplus_{i\in I} A.
\end{equation}
In particular, \(A\otimes \mathbb Z \mathrel{\vcentcolon=} A\).
This construction is an additive functor in~\(G\).
Namely, let \(G\)
and~\(H\)
be free Abelian groups and let \(f\colon G\to H\)
be a group homomorphism. The Universal Coefficient Theorem implies
\(\textup{KK}_0(\mathbb C\otimes G,\mathbb C\otimes H) \cong \Hom(G,H)\).
So we get a functorial \(f_\mathbb C\in \textup{KK}_0(\mathbb C\otimes G,\mathbb C\otimes H)\).
Identifying \(A\otimes G = A \otimes (\mathbb C\otimes G)\)
(with the minimal tensor product of \(\textup C^*\)\nb-algebras),
we get a functorial
\[
f_A\mathrel{\vcentcolon=} \textup{id}_A \otimes f_\mathbb C\in\textup{KK}_0(A\otimes G,A\otimes H).
\]
Let \(A\mathrel{\in\in}\KKcat\).
Then \(A\otimes \mathbb Z[x,x^{-1}]\) with the invertible element
\[
x_A = \textup{id}_A \otimes x
\in \textup{KK}_0\bigl(A\otimes \mathbb Z[x,x^{-1}],A\otimes \mathbb Z[x,x^{-1}]\bigr)
\]
induced by multiplication with the invertible element \(x\in\mathbb Z[x,x^{-1}]\)
is an object of \(\KKcat[\mathbb Z]\).
It behaves like a free module over~\(A\) because
\[
\Hom_{\KKcat[\mathbb Z]}\bigl((A\otimes \mathbb Z[x,x^{-1}],x_A), (B,b) \bigr)
\cong \textup{KK}_0(A,B).
\]
In particular, \((A\otimes \mathbb Z[x,x^{-1}],x_A)\)
is projective in the exact category \(\KKcat[\mathbb Z]\).
The invertible element~\(x_A\)
in the \(\textup{KK}\)\nb-endomorphism
ring of \((A\otimes \mathbb Z[x,x^{-1}],x_A)\)
lifts to the shift automorphism
\[
\tau\in\Aut(\Cont_0(\mathbb Z,A)),\qquad
(\tau f)(n) \mathrel{\vcentcolon=} f(n-1).
\]
That is,
\(F^\mathbb Z(\Cont_0(\mathbb Z,A),\tau) \cong (A\otimes \mathbb Z[x,x^{-1}],x_A)\). And
\begin{equation}
\label{eq:FZ_adjoint}
\textup{KK}^\mathbb Z_0(\Cont_0(\mathbb Z,A), B)
\cong \textup{KK}_0(A,B)
\cong \Hom_{\KKcat[\mathbb Z]}\bigl((A\otimes \mathbb Z[x,x^{-1}],x_A),B\bigr).
\end{equation}
The first isomorphism here is \cite{Meyer-Nest:BC}*{Equation~(20)}.
It applies the forgetful functor
\[
\textup{KK}^\mathbb Z_0\bigl((\Cont_0(\mathbb Z,A),\tau), (B,\beta) \bigr) \to
\textup{KK}_0(\Cont_0(\mathbb Z,A),B)
\]
and then composes with the inclusion \texorpdfstring{$^*$\nb-}{*-}{}homomorphism
\(A\to\Cont_0(\mathbb Z,A)\)
that maps~\(A\)
identically onto the summand at \(0\in\mathbb Z\).
Equation~\eqref{eq:FZ_adjoint} says that the partial adjoint of the
functor~\(F^\mathbb Z\)
is defined on \((A\otimes \mathbb Z[x,x^{-1}],x_A)\)
and maps it to \((\Cont_0(\mathbb Z,A),\tau)\).
\begin{lemma}
\label{lem:PV-coker_is_Ext}
Any object of~\(\KKcat[\mathbb Z]\) has a free resolution of length~\(1\).
\end{lemma}
\begin{proof}
The trivial representation of the group~\(\mathbb Z\)
on~\(\mathbb Z\)
corresponds to~\(\mathbb Z\)
made a module over~\(\mathbb Z[x,x^{-1}]\)
by letting \(x\in \mathbb Z[x,x^{-1}]\)
act by~\(1\).
This module has the following free resolution of length~\(1\):
\begin{equation}
\label{eq:ZZ_resolution}
0\to \mathbb Z[x,x^{-1}] \xrightarrow{\mathrm{mult}(x-1)} \mathbb Z[x,x^{-1}]
\xrightarrow{\textup{ev}_1} \mathbb Z \to 0,
\end{equation}
where \(\textup{ev}_1\colon \mathbb Z[x,x^{-1}] \to \mathbb Z\)
is the homomorpism of evaluation at~\(1\),
so \(\textup{ev}_1(x^n)=1\)
for all \(n\in\mathbb Z\).
The extension in~\eqref{eq:ZZ_resolution} splits as an extension
of Abelian groups because~\(\mathbb Z\)
is free as an Abelian group.
The resolution~\eqref{eq:ZZ_resolution} induces a chain complex
\begin{equation}
\label{eq:AZ_resolution}
0 \to (A \otimes \mathbb Z[x,x^{-1}],[\alpha]\otimes x)
\to (A \otimes \mathbb Z[x,x^{-1}],[\alpha]\otimes x)
\to (A,[\alpha])
\to 0
\end{equation}
in the additive category~\(\KKcat[\mathbb Z]\).
That is, the maps are arrows in~\(\KKcat[\mathbb Z]\).
The extension~\eqref{eq:ZZ_resolution} splits by a group
homomorphism, and the tensor product construction is additive.
Hence~\eqref{eq:AZ_resolution} is exact, that is, it splits
in~\(\KKcat\). The arrow
\[
A \otimes \mathbb Z[x,x^{-1}]
= \bigoplus_{n\in\mathbb Z} A
\xrightarrow{\bigoplus_{n\in\mathbb Z} [\alpha^n]}
\bigoplus_{n\in\mathbb Z} A
= A \otimes \mathbb Z[x,x^{-1}]
\]
in~\(\KKcat\) is an isomorphism
\[
(A \otimes \mathbb Z[x,x^{-1}],\textup{id}_A\otimes x) \cong
(A \otimes \mathbb Z[x,x^{-1}],[\alpha]\otimes x).
\]
Thus~\eqref{eq:AZ_resolution} is isomorphic to a free resolution
\begin{equation}
\label{eq:AZ_resolution_free}
0 \to (A \otimes \mathbb Z[x,x^{-1}],1\otimes x)
\to (A \otimes \mathbb Z[x,x^{-1}],1\otimes x)
\to (A,[\alpha])
\to 0,
\end{equation}
in~\(\KKcat[\mathbb Z]\),
where the boundary map on \((A \otimes \mathbb Z[x,x^{-1}],1\otimes x)\)
has changed to
\([\alpha]^{-1} \otimes \mathrm{mult}(x) - \textup{id}_{A \otimes
\mathbb Z[x,x^{-1}]}\).
\end{proof}
Lemma~\ref{lem:PV-coker_is_Ext} implies that the exact category
\(\KKcat[\mathbb Z]\)
has enough projective objects and that an object of~\(\KKcat[\mathbb Z]\)
is projective if and only if it is a direct summand of a free object.
Since the partial adjoint of the functor~\(F^\mathbb Z\)
is defined on all free objects, it is defined also on all projective
objects of \(\KKcat[\mathbb Z]\),
and it is inverse to~\(F^\mathbb Z\)
on this subcategory. Idempotents in the category \(\KKcat^\mathbb Z\)
split because it is triangulated and has countable direct sums.
\begin{proposition}
\label{pro:FZ_universal_exact}
The functor \(F^\mathbb Z\colon \KKcat^\mathbb Z \to \KKcat[\mathbb Z]\)
is the universal \(\Ideal^\mathbb Z\)\nb-exact
stable homological functor from~\(\KKcat^\mathbb Z\) to an exact category.
\end{proposition}
\begin{proof}
To prove this, we first embed \(\KKcat[\mathbb Z]\)
into an Abelian category by Quillen's Embedding Theorem. One way to
do this is to embed~\(\KKcat\)
into an Abelian category~\(\Abel\)
by Freyd's Embedding Theorem and then form \(\Abel{}[\mathbb Z]\).
The Abelian category \(\Abel{}[\mathbb Z]\)
has enough projective objects, and all its projective objects
already belong to \(\KKcat[\mathbb Z]\).
So Theorem~\ref{the:characterise_universal_homological} applies and
shows that the functor to~\(\Abel{}[\mathbb Z]\)
is the universal \(\Ideal\)\nb-exact
functor to an Abelian category. The subcategory
\(\KKcat[\mathbb Z] \subseteq \Abel{}[\mathbb Z]\)
is exact and contains the range of the functor. In fact, we shall
show soon that the functor \(\KKcat^\mathbb Z \to \KKcat[\mathbb Z]\)
is surjective on objects (see Theorem~\ref{the:Z_classify}). Taking
this for granted, Remark~\ref{rem:universal_to_exact} shows that the
functor \(\KKcat^\mathbb Z \to \KKcat[\mathbb Z]\)
is the universal \(\Ideal\)\nb-exact functor to an exact category.
\end{proof}
Equip \(\Cont_0(\mathbb Z,A)\)
with the \(\mathbb Z\)\nb-action
generated by the shift automorphism~\(\tau\).
We use~\eqref{eq:ZZ_resolution} to lift the
resolution~\eqref{eq:AZ_resolution_free} in \(\KKcat[\mathbb Z]\)
to the following \(\Ideal^\mathbb Z\)\nb-projective
resolution of length~\(1\) in~\(\KKcat^\mathbb Z\):
\begin{equation}
\label{eq:lift_AZ_resolution}
0 \to \Cont_0(\mathbb Z) \otimes A
\xrightarrow{\varphi} \Cont_0(\mathbb Z) \otimes A
\xrightarrow{p} (A,[\alpha])
\to 0;
\end{equation}
here \(\varphi = [\tau]\otimes [\alpha]^{-1} - 1\),
and~\(p\)
is the counit of the adjunction~\eqref{eq:FZ_adjoint}; that is, the
first isomorphism in~\eqref{eq:FZ_adjoint} maps~\(p\)
to the identity element in \(\textup{KK}_0(A,A)\).
In other words, when we forget the \(\mathbb Z\)\nb-actions,
then~\(p\)
restricts to the identity map at the \(0\)th
summand in \(\Cont_0(\mathbb Z) \otimes A = \bigoplus_{n\in\mathbb Z} A\).
When we use the resolution~\eqref{eq:lift_AZ_resolution} to compute
derived functors, we get
\[
\Ext^1_{\KKcat[\mathbb Z],{\Ideal^\mathbb Z}}\bigl((A,[\alpha]),(B,[\beta])\bigr)
\cong
\coker \bigl((\alpha^*)^{-1}\beta_*-1\colon \textup{KK}_0(A,B) \to \textup{KK}_0(A,B)\bigr).
\]
The map \((\alpha^*)^{-1}\beta_*-1\),
has the same cokernel as \(\alpha^*\beta_*^{-1}-1\).
A rather deep result says that the \(\Ideal^\mathbb Z\)\nb-projective
objects generate \(\KKcat^\mathbb Z\).
Equivalently, if \(R^\mathbb Z(A) \cong 0\)
in \(\KKcat\),
then already \(A\cong 0\)
in \(\KKcat^\mathbb Z\).
This is related to the proof of the Baum--Connes conjecture for the
group~\(\mathbb Z\).
It follows immediately from the Pimsner--Voiculescu
sequence~\eqref{eq:PV-sequence}.
\begin{theorem}
\label{the:Z_classify}
Let \((A,\alpha)\)
and \((B,\beta)\)
be objects of \(\KKcat^\mathbb Z\).
Any isomorphism \(F^\mathbb Z(A) \cong F^\mathbb Z(B)\)
in \(\KKcat[\mathbb Z]\)
lifts to an isomorphism \(A\cong B\)
in \(\KKcat^\mathbb Z\).
And any object of \(\KKcat[\mathbb Z]\)
is isomorphic to \(F^\mathbb Z(A)\)
for some \(A\mathrel{\in\in}\KKcat^\mathbb Z\), which is unique up to isomorphism.
If \(t\in \textup{KK}_0(A,B)^\times\)
is a \(\textup{KK}\)\nb-equivalence
with \(t\circ [\alpha] = [\beta]\circ t\),
then there is an isomorphism in \(\textup{KK}^\mathbb Z_0(A,B)\)
that is mapped to~\(t\)
by the forgetful functor \(\KKcat^\mathbb Z\to\KKcat\).
\end{theorem}
\begin{proof}
The Universal Coefficient Theorem for \(\textup{KK}^\mathbb Z_0(A,B)\)
and the ideal~\(\Ideal^\mathbb Z\)
allows to lift any map \(t\colon (A,[\alpha])\to (B,[\beta])\)
in~\(\KKcat[\mathbb Z]\)
to an element
\(\hat{t}\in \textup{KK}^\mathbb Z_0\bigl((A,\alpha), (B,\beta)\bigr)\).
The naturality of the Universal Coefficient Theorem implies that the
composition vanishes for two elements of the \(\Ext^1\)-part
of \(\textup{KK}^\mathbb Z_0(A,B)\).
Hence~\(\hat{t}\)
is invertible if~\(t\)
is invertible. Any object of \(\KKcat[\mathbb Z]\)
has a projective resolution of length~\(1\).
This allows to lift it to an object of~\(\KKcat^\mathbb Z\)
(see \cite{Bentmann-Meyer:More_general}*{Proposition~2.3}). This
proves both assertions in the first paragraph. The second paragraph
only describes isomorphisms in \(\KKcat[\mathbb Z]\) more concretely.
\end{proof}
We make the isomorphism criterion in Theorem~\ref{the:Z_classify} more
explicit under the assumption that the \(\textup C^*\)\nb-algebras
\(A\)
and~\(B\)
belong to the bootstrap class. Let~\(A^\pm\)
be \(\textup C^*\)\nb-algebras in the bootstrap class with
\[
\textup{K}_0(A^+) = \textup{K}_0(A),\quad
\textup{K}_1(A^+) = 0,\qquad
\textup{K}_0(A^-) = 0,\quad
\textup{K}_1(A^-) = \textup{K}_1(A).
\]
Then \(\textup{K}_*(A^+ \oplus A^-) \cong \textup{K}_*(A)\).
So \(A\cong A^+ \oplus A^-\)
by the Universal Coefficient Theorem for \(\textup{KK}\).
We use such a \(\textup{KK}\)\nb-equivalence
to map~\(\alpha\)
to an element of \(\textup{KK}_0(A^+\oplus A^-, A^+\oplus A^-)\).
We rewrite this as a \(2\times2\)-matrix
\[
\begin{pmatrix}
\alpha^{++}& \alpha^{+-}\\
\alpha^{-+}& \alpha^{--}
\end{pmatrix},\qquad
\begin{aligned}
\alpha^{++} &\in\textup{KK}_0(A^+,A^+) \cong \Hom\bigl(\textup{K}_0(A),\textup{K}_0(A)\bigr),\\
\alpha^{+-} &\in\textup{KK}_0(A^-,A^+) \cong \Ext\bigl(\textup{K}_1(A),\textup{K}_0(A)\bigr),\\
\alpha^{-+} &\in\textup{KK}_0(A^+,A^-) \cong \Ext\bigl(\textup{K}_0(A),\textup{K}_1(A)\bigr),\\
\alpha^{--} &\in\textup{KK}_0(A^-,A^-) \cong \Hom\bigl(\textup{K}_1(A),\textup{K}_1(A)\bigr).
\end{aligned}
\]
Here we have used the Universal Coefficient Theorem for the \(\textup C^*\)\nb-algebras
\(A^\pm\).
A similar decomposition \(B\cong B^+ \oplus B^-\)
allows us to map~\(\beta\)
to an element of \(\textup{KK}_0(B^+\oplus B^-, B^+\oplus B^-)\),
which we then rewrite as a \(2\times2\)-matrix
\[
\begin{pmatrix}
\beta^{++}& \beta^{+-}\\
\beta^{-+}& \beta^{--}
\end{pmatrix},\qquad
\begin{aligned}
\beta^{++} &\in\textup{KK}_0(B^+,B^+) \cong \Hom\bigl(\textup{K}_0(B),\textup{K}_0(B)\bigr),\\
\beta^{+-} &\in\textup{KK}_0(B^-,B^+) \cong \Ext\bigl(\textup{K}_1(B),\textup{K}_0(B)\bigr),\\
\beta^{-+} &\in\textup{KK}_0(B^+,B^-) \cong \Ext\bigl(\textup{K}_0(B),\textup{K}_1(B)\bigr),\\
\beta^{--} &\in\textup{KK}_0(B^-,B^-) \cong \Hom\bigl(\textup{K}_1(B),\textup{K}_1(B)\bigr).
\end{aligned}
\]
And we may also transfer an element \(t\in \textup{KK}_0(A,B)\)
to such a \(2\times2\)-matrix
\[
\begin{pmatrix}
t^{++}& t^{+-}\\
t^{-+}& t^{--}
\end{pmatrix},\qquad
\begin{aligned}
t^{++} &\in\textup{KK}_0(A^+,B^+) \cong \Hom\bigl(\textup{K}_0(A),\textup{K}_0(B)\bigr),\\
t^{+-} &\in\textup{KK}_0(A^-,B^+) \cong \Ext\bigl(\textup{K}_1(A),\textup{K}_0(B)\bigr),\\
t^{-+} &\in\textup{KK}_0(A^+,B^-) \cong \Ext\bigl(\textup{K}_0(A),\textup{K}_1(B)\bigr),\\
t^{--} &\in\textup{KK}_0(A^-,B^-) \cong \Hom\bigl(\textup{K}_1(A),\textup{K}_1(B)\bigr).
\end{aligned}
\]
The naturality of the exact sequence in the Universal Coefficient Theorem implies that the
Kasparov product of an element of \(\Ext\bigl(\textup{K}_{1+*}(A),\textup{K}_*(B)\bigr)\)
with an element of \(\textup{KK}_0(B,C)\)
depends only on the image of the latter in \(\Hom\bigl(\textup{K}_*(B),\textup{K}_*(C)\bigr)\).
Thus the product of two \(\Ext\)-terms
always vanishes. Hence the condition \(t [\alpha]= [\beta] t\)
in Theorem~\ref{the:Z_classify} is equivalent to four equations
\begin{alignat*}{2}
t^{++} \alpha^{++} &= \beta^{++} t^{++},&\qquad
t^{+-} \alpha^{--} + t^{++} \alpha^{+-}
&= \beta^{+-} t^{--} + \beta^{++} t^{+-}, \\
t^{--} \alpha^{--} &= \beta^{--} t^{--},&\qquad
t^{-+} \alpha^{++} + t^{--} \alpha^{-+}
&= \beta^{-+} t^{++} + \beta^{--} t^{-+}.
\end{alignat*}
The equations on the left assert that the two diagonal entries of
\(t [\alpha]\)
and~\([\beta] t\)
are equal; those on the right assert equality of the two off-diagonal
entries. Theorem~\ref{the:Z_classify} says that \((A,\alpha)\)
and~\((B,\beta)\)
are \(\textup{KK}^\mathbb Z\)\nb-equivalent
if and only if there is an invertible element~\(t\)
for which these equations hold. The equations on the left mean that
the isomorphism \(t_* = (t^{++}, t^{--}) \in \Hom\bigl(\textup{K}_*(A),\textup{K}_*(B)\bigr)\)
induced by~\(t\)
intertwines the automorphisms \(\alpha_*\in \Aut(\textup{K}_*(A))\)
and \(\beta_*\in \Aut(\textup{K}_*(B))\).
Equivalently, \(t_*\)
is an isomorphism of \(\mathbb Z/2\)\nb-graded
\(\mathbb Z[x,x^{-1}]\)-modules.
The equations on the right mean that there is
\(t^1 \mathrel{\vcentcolon=} (t^{+-},t^{-+}) \in \Ext\bigl(\textup{K}_{1+*}(A),\textup{K}_*(B)\bigr)\) with
\[
t^1 \alpha_* - \beta_* t^1 = \beta^1 t_* - t_* \alpha^1,
\]
where we abbreviate \(\alpha_* = (\alpha^{++},\alpha^{--})\)
and \(\alpha^1 = (\alpha^{+-},\alpha^{-+})\),
and similarly for~\(\beta\).
The choice of~\(t^1\)
has no effect on the invertibility of~\(t\).
So the criterion in Theorem~\ref{the:Z_classify} is whether the image
of \(\beta^1 t_* - t_* \alpha^1\) vanishes in the cokernel of
\begin{equation}
\label{eq:cokernel_Ext_Z}
\Ext\bigl(\textup{K}_{1+*}(A),\textup{K}_*(B)\bigr) \to \Ext\bigl(\textup{K}_{1+*}(A),\textup{K}_*(B)\bigr),\qquad
t^1 \mapsto t^1 \alpha_* - \beta_* t^1.
\end{equation}
We shall later identify this cokernel with the group
\(\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\textup{K}_{1+*}(A),\textup{K}_*(B)\bigr)\)
and show that the image of \(\beta^1 t_* - t_* \alpha^1\)
in this cokernel is the relative obstruction class for \((A,\alpha)\)
and \((B,\beta)\)
and a \(\mathbb Z[x,x^{-1}]\)-module
isomorphisms \(t_*\colon \textup{K}_*(A) \cong \textup{K}_*(B)\).
This gives the rule to translate between the classifying invariants in
Theorem~\ref{the:Z_classify} and
in~\cite{Bentmann-Meyer:More_general}.
\begin{remark}
If \(A=B\)
is a unital Kirchberg algebra (separable, nuclear, unital, purely infinite
and simple), then a much finer classification theorem for
automorphisms is proved by Nakamura~\cite{Nakamura:Aperiodic}.
\end{remark}
\subsection{\texorpdfstring{$\textup C^*$}{C*}-Algebras over unique path spaces}
\label{sec:UCT_X}
Now we prove a Universal Coefficient Theorem for \(\KKcat^X\)
for a unique path space~\(X\).
Let~\(X\)
be a countable set and let~\(\to\)
be a relation on~\(X\),
which says for which points \(x,y\in X\)
there is an edge \(x\to y\).
Equip~\(X\)
with the partial order generated by~\(\leftarrow\),
that is, \(x \preceq y\)
if and only if there is a chain of edges
\(x=x_0 \leftarrow x_1 \leftarrow \dotsb \leftarrow x_\ell = y\)
with some \(\ell\ge0\)
and \(x_1,\dotsc,x_{\ell-1}\in X\).
Equip~\(X\)
with the Alexandrov topology generated by this partial order. We
assume~\((X,\to)\)
to be a \emph{unique path space}, that is, there is at most one chain
of edges
\(x=x_0 \leftarrow x_1 \leftarrow \dotsb \leftarrow x_\ell = y\)
between any two points \(x,y\in X\).
For \(x\in X\),
the subset \(U_x\mathrel{\vcentcolon=} \setgiven{y\in X}{x\preceq y}\)
is the minimal open subset containing~\(x\).
A \emph{\(\textup C^*\)\nb-algebra
over~\(X\)}
is equivalent to a \(\textup C^*\)\nb-algebra~\(A\)
with fixed ideals \(A(U_x)\idealin A\)
for all \(x\in X\),
such that \(A(U_x) \subseteq A(U_y)\)
for all \(x,y\in X\)
with \(x\to y\)
or, equivalently, with \(U_x \subseteq U_y\).
The equivariant Kasparov category~\(\KKcat^X\)
for \(\textup C^*\)\nb-algebras
over~\(X\)
has separable \(\textup C^*\)\nb-algebras
over~\(X\)
as objects and the KK-groups \(\textup{KK}_0^X(A,B)\)
as arrows (see also~\cite{Meyer-Nest:Bootstrap}, where this category
is denoted \(\mathfrak{KK}(X)\)). The forgetful functor
\[
R^X\colon \KKcat^X \to \prod_{x\in X} \KKcat,\qquad
A\mapsto (A(U_x))_{x\in X},
\]
is a triangulated functor between triangulated categories. Its kernel on
morphisms
\[
\Ideal^X(A,B) \mathrel{\vcentcolon=} \setgiven{f\in \textup{KK}^X_0(A,B)}
{f(U_x)=0 \text{ in } \textup{KK}_0(A(U_x),B(U_x)) \text{ for all } x\in X}
\]
is a stable homological ideal. We now describe the universal
\(\Ideal^X\)\nb-exact
stable homological functor as in Section~\ref{sec:UCT_Z}.
Let \(\KKcat[X]\)
be the category of functors \((X,\succeq) \to \KKcat\)
with natural transformations as arrows. Since~\(X\) is a unique
path space, the category associated to the partially ordered set
\((X,\succeq)\)
is the path category of the directed graph \((X,\to)\).
By the universal property of the path category, an object of
\(\KKcat[X]\)
is given by \(A_x\mathrel{\in\in}\KKcat\)
for \(x\in X\)
and \(\alpha_{y,x}\in \textup{KK}_0(A_x,A_y)\)
for \(x,y\in X\)
with \(x\to y\),
without any relations on the~\(\alpha_{y,x}\).
This uniquely determines \(\textup{KK}\)-classes
\(\alpha_{y,x}\in \textup{KK}_0(A_x,A_y)\)
for \(x,y\in X\)
with \(x\succeq y\)
such that \(\alpha_{x,x} = \textup{id}_{A_x}\)
and \(\alpha_{z,y}\circ\alpha_{y,x}=\alpha_{z,x}\)
for all \(x,y,z\in X\)
with \(x\succeq y\succeq z\).
An arrow \((A_x,\alpha_{y,x})\to (B_x,\beta_{y,x})\)
is a family of arrows \(f_x\in \textup{KK}_0(A_x,B_x)\)
for \(x\in X\)
with \(f_y \alpha_{y,x} = \beta_{y,x} f_x\)
in \(\textup{KK}_0(A_x,B_y)\)
for all \(x,y\in X\)
with \(x\to y\);
then \(f_y \alpha_{y,x} = \beta_{y,x} f_x\)
holds for all \(x,y\in X\) with \(x\succeq y\). Define
\[
F^X\colon \KKcat^X\to \KKcat[X]
\]
by mapping a \(\textup C^*\)\nb-algebra~\(A\)
over~\(X\)
to the object of \(\KKcat[X]\)
where \(A_x \mathrel{\vcentcolon=} A(U_x)\)
and where \(\alpha_{y,x}\in\textup{KK}_0(A_x,A_y)\)
for \(x,y\in X\)
with \(x\to y\)
is the \(\textup{KK}\)-class
of the inclusion map \(A(U_x) \hookrightarrow A(U_y)\).
Then \(\alpha_{y,x}\in\textup{KK}_0(A_x,A_y)\)
for \(x,y\in X\)
with \(x\succeq y\)
is the \(\textup{KK}\)-class
of the inclusion map as well. A kernel--cokernel pair \(K\to E\to Q\)
in \(\KKcat[X]\)
is called \emph{admissible} if it splits pointwise, that is,
\(K_x \to E_x \to Q_x\)
is a split extension in \(\KKcat\)
for all \(x\in X\);
so \(E_x \cong K_x \oplus Q_x\)
in \(\KKcat\)
for all \(x\in X\),
but the sections \(Q_x \to E_x\)
are not compatible with the structure maps \(E_x \to E_y\)
and \(Q_x \to Q_y\)
for \(x\to y\).
This turns~\(\KKcat[X]\)
into an exact category (see \cite{Buehler:Exact}*{Exercise~5.3}).
Let \(z\in X\)
and \(A\mathrel{\in\in}\KKcat\).
As in~\cite{Meyer-Nest:Bootstrap}, let \(i_z(A)\mathrel{\in\in}\KKcat^X\)
be the \(\textup C^*\)\nb-algebra~\(A\)
with
\[
i_z(A)(U_x) \mathrel{\vcentcolon=}
\begin{cases}
A&\text{if }z\in U_x\text{, that is, }x\preceq z,\\
0&\text{otherwise.}
\end{cases}
\]
Then
\begin{equation}
\label{eq:KK_out_ix}
\textup{KK}^X_0(i_z(A),B) \cong \textup{KK}_0\bigl(A,B(U_z)\bigr)
\end{equation}
for all \(B\mathrel{\in\in}\KKcat^X\)
by \cite{Meyer-Nest:Bootstrap}*{Proposition 3.13}. This isomorphism
applies the restriction map
\(\textup{KK}^X_0(i_z(A),B) \to \textup{KK}_0(i_z(A)(U_z),B(U_z))\)
and then identifies \(i_z(A)(U_z) = A\). The object
\[
j_z(A) \mathrel{\vcentcolon=} F^X(i_z(A)) \mathrel{\in\in}\KKcat[X]
\]
has \(j_z(A)_x = A\)
for \(x\preceq z\)
and \(j_z(A)_x = 0\)
otherwise,
and the map \(j_z(A)_x \to j_z(A)_y\)
for \(x\succeq y\)
is the identity map in \(\textup{KK}_0(A,A)\)
if \(z\succeq x\)
and the zero map in \(\textup{KK}_0(0,j_z(A)_y)\) otherwise. We compute
\begin{equation}
\label{eq:KKX_out_Fix}
\Hom_{\KKcat[X]}\bigl(j_z(A),(B_x,\beta_{y,x})\bigr) \cong \textup{KK}_0(A,B_z).
\end{equation}
Equations \eqref{eq:KK_out_ix} and~\eqref{eq:KKX_out_Fix} imply
\begin{equation}
\label{eq:KKX_out_Fix2}
\textup{KK}^X_0(i_z(A),B)
\cong \textup{KK}_0(A,B(U_z))
\cong \Hom_{\KKcat[X]}\bigl(j_z(A),F^X(B)\bigr)
\end{equation}
for all \(B\mathrel{\in\in}\KKcat^X\).
\begin{theorem}
\label{the:projectives_ups}
Let \((X,\to)\)
be a countable unique path space. The objects of the exact
category~\(\KKcat[X]\)
of the form \(\bigoplus_{z\in X} j_z(A_z)\)
for \(A_z\mathrel{\in\in}\KKcat\)
for \(z\in X\)
are projective. For any object \(A = (A_x,\alpha_{x,y})\)
of \(\KKcat[X]\),
there is an admissible extension
\begin{equation}
\label{eq:projective_resolution_ups}
\bigoplus_{x\to y} j_y(A_x)
\overset{\iota}\rightarrowtail \bigoplus_{x\in X} j_x(A_x)
\overset{\pi}\twoheadrightarrow A,
\end{equation}
which is a projective resolution of length~\(1\).
The exact category~\(\KKcat[X]\)
has enough projective objects, and the partial adjoint of~\(F^X\)
is defined on all projective objects of~\(\KKcat[X]\)
and is a section for~\(F^X\)
there. The functor \(F^X\colon \KKcat^X \to \KKcat[X]\)
is the universal \(\Ideal^X\)\nb-exact
stable homological functor to an exact category.
\end{theorem}
\begin{proof}
The objects~\(j_z(A_z)\)
are projective by~\eqref{eq:KKX_out_Fix}. This is inherited by the
direct sum \(\bigoplus_{z\in X} j_z(A_z)\).
The identity map on~\(A_x\)
has an adjunct \(a_x\colon j_x(A_x) \to A\)
by~\eqref{eq:KKX_out_Fix}. These maps induce the map
\(\pi= (a_x)_{x\in X}\colon \bigoplus_{x\in X} j_x(A_x) \to A\).
For each edge \(x\to y\)
in the directed graph \((X,\to)\),
the map \((\textup{id}_{A_x},-\alpha_{y,x})\colon A_x \to A_x \oplus A_y\)
is adjunct to a map
\[
a_{x\to y}\colon j_y(A_x) \to j_x(A_x) \oplus j_y(A_y)
\subseteq \bigoplus_{x\in X} j_x(A_x)
\]
by~\eqref{eq:KKX_out_Fix}. It satisfies
\(\pi\circ a_{x\to y} = 0\).
The maps~\(a_{x\to y}\) combine to a map
\[
\iota\colon \bigoplus_{x\to y} j_y(A_x) \to \bigoplus_{x} j_x(A_x)
\]
with \(\pi\circ\iota=0\).
Now we prove that the maps \(\pi\)
and~\(\iota\)
mapped to \(\prod_{x\in X} \KKcat\)
by the forgetful functor form a split exact sequence. We consider
the entries at a fixed \(z\in X\).
The entry of~\(A\)
at \(z\in X\)
is simply~\(A_z\).
The entry of \(\bigoplus_{x\in X} j_x(A_x)\)
at \(z\in X\)
is the direct sum of~\(A_x\)
over all \(x\in X\)
with \(x\succeq z\).
The entry of \(\bigoplus_{x\to y} j_y(A_x)\)
at \(z\in X\)
is the direct sum of~\(A_x\)
for all edges \(x\to y\)
in~\(X\) with \(y\succeq z\). The entry of~\(\pi\) at~\(z\) is
\[
(\alpha_{z,x})_{x\succeq z} \colon
\bigoplus_{x\succeq z} A_x \to A_z.
\]
This is split surjective with the canonical section that
maps~\(A_z\)
identically onto the summand~\(A_z\)
for \(x=z\).
The entry of~\(\iota\)
at~\(z\)
maps the summand~\(A_x\)
for \(x\to y\succeq z\)
to \(A_x \oplus A_y\subseteq \bigoplus_{t\succeq z} A_t\)
using \((\textup{id}_{A_x},-\alpha_{y,x})\). We are going to define a map
\[
s_1\colon \bigoplus_{t\succeq z} A_t \to
\bigoplus_{x\to y \succeq z} A_x
\]
which together with~\(s_0\)
forms a contracting homotopy for the short chain complex formed by
\(\iota|_z\)
and~\(\pi|_z\).
By assumption, if \(t\succeq z\)
then there is a unique chain
\(t=x_0 \to x_1 \to \dotsb \to x_\ell = z\).
We let \((s_1)_z|_{A_t}\)
for \(t\succeq z\)
map the summand \(A_t\)
in \(\bigoplus_{t\succeq z} A_t\)
to the direct sum of~\(A_{x_i}\)
for the edges \(x_i \to x_{i+1}\)
for \(i=0,\dotsc,\ell-1\),
where we use~\(\alpha_{x_i,t}\)
to map \(A_t\)
to~\(A_{x_i}\).
If \(x\to y \succeq z\),
then \(y=x_1\)
in the above chain. Therefore, \(s_1\circ\iota|_{A_x}\)
is a map to \(\bigoplus_{j=0}^\ell A_{x_j}\),
where the entry at~\(A_{x_j}\)
is \(\alpha_{x_j,x} - \alpha_{x_j,y}\circ \alpha_{y,x} = 0\)
for \(j=1,\dotsc,\ell\),
and the identity map for \(j=0\).
So \(s_1\circ \iota\)
is the identity map. Finally, we claim that
\(s_0\circ \pi + \iota\circ s_1\)
is the identity map on \(\bigoplus_{t\succeq z} A_t\).
This is checked on each summand~\(A_t\)
separately. Let \(t=x_0 \to x_1 \to \dotsb \to x_\ell = z\)
be the unique chain as above. Then \(\iota\circ s_1\)
is a telescoping sum of \(\pm\alpha_{x_j,t}\)
for \(j=0,\dotsc,\ell\),
where \(\alpha_{t,t}\)
occurs only with sign~\(+\)
and~\(\alpha_{z,t}\)
only with sign~\(-\).
And \(s_0\circ\pi\)
is the map~\(\alpha_{z,t}\).
Thus \(s_0\circ \pi + \iota\circ s_1\)
is the identity map on~\(A_t\).
This finishes the proof that~\eqref{eq:projective_resolution_ups} is
\(\Ideal^X\)\nb-exact.
Its entries are \(\Ideal^X\)\nb-projective.
So it is an \(\Ideal^X\)\nb-projective resolution.
The projective resolution~\eqref{eq:projective_resolution_ups}
implies that~\(\KKcat[X]\)
has enough projective objects and that an object is projective if
and only if it is a direct summand of an object of the form
\(\bigoplus_{z\in X} j_z(B_z)\)
for some separable \(\textup C^*\)\nb-algebras~\(B_z\)
for \(z\in X\).
Equation~\eqref{eq:KKX_out_Fix2} says that the adjoint functor to
\(F^X\colon \KKcat^X\to\KKcat[X]\)
is defined on \(j_z(A)\)
and maps it to~\(i_z(A)\).
Since the partial adjoint commutes with direct sums, it is also
defined on \(\bigoplus_{z\in X} j_z(B_z)\)
for any \(B_z\mathrel{\in\in}\KKcat\)
and maps it to \(\bigoplus_{z\in X} i_z(B_z)\).
Idempotents in the category~\(\KKcat^X\)
split because it is triangulated and has countable direct sums.
Therefore, the partial adjoint of~\(F^X\)
is defined on all projective objects of~\(\KKcat[X]\)
and is a section for~\(F^X\)
there. An argument as in the proof of
Proposition~\ref{pro:FZ_universal_exact} shows that~\(F^X\)
is the universal \(\Ideal^X\)\nb-exact
functor to an exact category. The target category is not smaller
because the functor \(\KKcat^X \to \KKcat[X]\)
is surjective on objects by Corollary~\ref{cor:classify_KKX}.
\end{proof}
Unlike in Section~\ref{sec:UCT_Z}, the \(\Ideal^X\)\nb-projective
objects in~\(\KKcat^X\)
do not generate~\(\KKcat^X\).
The localising subcategory \(\gen{\Proj_\Ideal}\)
generated by them is equal to the localising subcategory generated by
objects of the form \(i_x(A)\)
for \(x\in X\),
\(A\mathrel{\in\in}\KKcat\).
If~\(X\)
is finite, then it is described in several equivalent ways in
\cite{Meyer-Nest:Bootstrap}*{Definition~4.7}. It contains all
nuclear \(\textup C^*\)\nb-algebras over~\(X\). This is the subcategory
on which Theorem~\ref{the:projectives_ups} implies a Universal
Coefficient Theorem, using \(\Hom\)
and \(\Ext\)
groups in the category \(\KKcat[X]\).
This implies the following classification theorem:
\begin{corollary}
\label{cor:classify_KKX}
Let \((X,\to)\)
be a countable unique path space and let \(A\)
and~\(B\)
be \(\textup C^*\)\nb-algebras
over~\(X\)
that belong to the localising subcategory \(\gen{\Proj_\Ideal}\)
generated by objects of the form~\(i_x(B)\)
for \(x\in X\),
\(B\mathrel{\in\in}\KKcat\).
Any isomorphism between \(F^X(A)\)
and~\(F^X(B)\)
in~\(\KKcat[X]\)
lifts to an isomorphism in~\(\KKcat^X\).
And any object of~\(\KKcat[X]\)
lifts to an object of~\(\gen{\Proj_\Ideal}\).
An isomorphism \(F^X(A) \xrightarrow\sim F^X(B)\)
is a family of invertible elements \(t_x\in \textup{KK}_0(A(U_x),B(U_x))\)
for \(x\in X\) for which the diagrams
\[
\begin{tikzcd}
A(U_x) \arrow[d, "t_x"] \arrow[r, hookrightarrow] &
A(U_y) \arrow[d, "t_y"] \\
B(U_x) \arrow[r, hookrightarrow] &
B(U_y)
\end{tikzcd}
\]
commute in~\(\KKcat\) for all \(x,y\in X\) with \(x\to y\).
\end{corollary}
The criterion above may be made more explicit if, in addition,
\(A(U_x)\)
and \(B(U_x)\)
belong to the bootstrap class for all \(x\in X\).
As in Section~\ref{sec:UCT_Z}, we identify
\(A(U_x) \cong A(U_x)^+ \oplus A(U_x)^-\)
and rewrite the classes of the inclusion maps
\(\alpha_{y,x}\in \textup{KK}_0(A(U_x),A(U_y))\)
and \(\beta_{y,x}\in \textup{KK}_0(B(U_x),B(U_y))\)
for \(x\to y\)
as \(2\times2\)\nb-matrices,
and similarly for the arrows \(t_x \in \textup{KK}_0(A(U_x),B(U_x))\).
As in the case of~\(\KKcat^\mathbb Z\),
the equality \(\beta_{y,x} t_x = t_y \alpha_{y,x}\)
for \(x\to y\)
in Corollary~\ref{cor:classify_KKX} may be rewritten as four
equalities of matrix coefficients. The equality of the diagonal terms
says that the diagrams \(\bigl(\textup{K}_*(A(U_x)),\textup{K}_*(\alpha_{y,x})\bigr)\)
and \(\bigl(\textup{K}_*(B(U_x)),\textup{K}_*(\beta_{y,x})\bigr)\)
of countable \(\mathbb Z/2\)\nb-graded
Abelian groups are isomorphic. That is, for all \(x,y\in X\)
with \(x\to y\), the following diagram commutes:
\[
\begin{tikzcd}[column sep=large]
\textup{K}_*(A(U_x)) \arrow[d, "\textup{K}_*(t_x)"] \arrow[r, "\textup{K}_*(\alpha_{y,x})"] &
\textup{K}_*(A(U_y)) \arrow[d, "\textup{K}_*(t_y)"] \\
\textup{K}_*(B(U_x)) \arrow[r, "\textup{K}_*(\beta_{y,x})"] &
\textup{K}_*(B(U_y))
\end{tikzcd}
\]
The equality of the off-diagonal terms will be studied in
Section~\ref{sec:compare_ups}.
\section{Computation of the obstruction class}
\label{sec:triangulated}
We recall the setup of~\cite{Bentmann-Meyer:More_general}.
Let~\(\Tri\)
be a triangulated category with countable direct sums. Let~\(\Ideal\)
be a stable homological ideal in~\(\Tri\)
with enough projective objects. Let \(F\colon \Tri\to\Abel\)
be the universal \(\Ideal\)\nb-exact
stable homological functor (in this article, we allow~\(\Abel\)
to be exact). We assume that~\(\Abel\)
is paired as in \cite{Bentmann-Meyer:More_general}*{Definition~2.14},
that is, \(\Abel = \Abel_+ \times \Abel_-\)
with \(\Sigma \Abel_+ = \Abel_-\)
and \(\Sigma \Abel_- = \Abel_+\).
For instance, \(\Abel\)
could be the category of \(\mathbb Z\)\nb-graded
or \(\mathbb Z/2\)\nb-graded
modules over some ring, with the suspension automorphism shifting the
grading, and \(\Abel_+\)
and~\(\Abel_-\)
may be taken to be the graded modules concentrated in even or odd
degrees, respectively. We want to compute the obstruction class of an
object \(A\mathrel{\in\in}\Tri\).
This is only meaningful if~\(A\)
is constructed from simpler ingredients.
We assume that there is an exact, \(\Ideal\)\nb-exact triangle
\begin{equation}
\label{eq:triangle_B1B0A}
B_1 \xrightarrow{\varphi} B_0 \xrightarrow{p} A \xrightarrow{i}
\Sigma B_1,
\end{equation}
where \(B_0\)
and~\(B_1\)
are objects of~\(\gen{\Proj_\Ideal}\)
with projective resolutions of length~\(1\).
Then \(A\mathrel{\in\in}\gen{\Proj_\Ideal}\)
as well. The \(\Ideal\)\nb-exactness
assumption says that \(i\in\Ideal\).
Equivalently, \(F(i)=0\),
\(F(\varphi)\)
is monic, and \(F(p)\)
is epic. The objects \(B_0\)
and~\(B_1\)
are uniquely determined up to isomorphism by \(F(B_0)\)
and~\(F(B_1)\)
because they have projective resolutions of length~\(1\)
(compare \cite{Bentmann-Meyer:More_general}*{Proposition~2.3}). We
may split \(B_i \cong B_i^+ \oplus B_i^-\)
with \(F(B_i^\pm)\in \Abel_\pm\)
for \(i=0,1\). Then \(\varphi\) becomes a \(2\times2\)-matrix
\[
\varphi =
\begin{pmatrix}
\varphi_{++}& \varphi_{+-}\\
\varphi_{-+}& \varphi_{--}
\end{pmatrix}
\]
with \(\varphi_{-+}\colon B_1^+ \to B_0^-\),
and so on. The two diagonal entries give an element of
\[
\Tri(B_1^+,B_0^+) \oplus \Tri(B_1^-,B_0^-) \cong
\Hom_{\Abel}\bigl(F(B_1),F(B_0)\bigr),
\]
and the two off-diagonal entries give an element of
\[
\Tri(B_1^+,B_0^-) \oplus \Tri(B_1^-,B_0^+) \cong
\Ext^1_{\Abel}\bigl(\Sigma F(B_1),F(B_0)\bigr);
\]
here we have used the Universal Coefficient Theorem to compute
\(\Tri(B_1^\pm,B_0^\pm)\).
The results for these four groups only have a single Hom or a single
Ext group because of the parity assumptions. Thus the splitting of
\(\varphi\)
into \(\varphi^0\mathrel{\vcentcolon=}\varphi_{++}+\varphi_{--}\)
and \(\varphi^1\mathrel{\vcentcolon=}\varphi_{+-}+\varphi_{-+}\)
splits the exact sequence
\[
\Ext^1_{\Abel}\bigl(\Sigma F(B_1),F(B_0)\bigr) \rightarrowtail
\Tri_*(B_1,B_0) \twoheadrightarrow
\Hom_{\Abel}\bigl(F(B_1),F(B_0)\bigr).
\]
We shall compute the obstruction class of~\(A\)
in terms of~\(\varphi^1\).
By assumption, there is a short exact sequence
\[
\begin{tikzcd}
F(B_1) \arrow[r, rightarrowtail, "F(\varphi)"] &
F(B_0) \arrow[r, twoheadrightarrow, "F(p)"] &
F(A).
\end{tikzcd}
\]
This induces a long exact sequence
\begin{multline*}
0 \leftarrow \Ext^2_{\Abel}\bigl(\Sigma F(A),F(A)\bigr)
\xleftarrow{\partial} \Ext^1_{\Abel}\bigl(\Sigma F(B_1),F(A)\bigr)
\xleftarrow{F(\varphi)^*}\\
\Ext^1_{\Abel}\bigl(\Sigma F(B_0),F(A)\bigr)
\xleftarrow{F(p)^*} \Ext^1_{\Abel}\bigl(\Sigma F(A),F(A)\bigr)
\leftarrow \dotsb
\end{multline*}
because \(\Ext^k_{\Abel}\bigl(F(B_i),F(A)\bigr)=0\) for \(i=0,1\), \(k\ge2\).
\begin{theorem}
\label{the:compute_obstruction}
The obstruction class of~\(A\) is
\[
\partial(p_*\varphi^1)
= \partial(F(p)\circ \varphi^1)\in
\Ext^2_{\Abel}\bigl(\Sigma F(A),F(A)\bigr),
\]
where \(F(p)\in \Hom_{\Abel}\bigl(\Sigma F(B_0),\Sigma F(A)\bigr)\)
and \(\varphi^1\in \Ext^1_{\Abel}\bigl(\Sigma F(B_1),F(B_0)\bigr)\).
\end{theorem}
\begin{proof}
We shall recall the construction of obstruction classes
in~\cite{Bentmann-Meyer:More_general} along the way. It starts with
an \(\Ideal\)\nb-projective
resolution of~\(A\)
of length~\(2\).
So first we have to construct this. We use the projective
resolutions of~\(F(B_i)\)
of length~\(1\),
which exist by assumption. They lift canonically to
\(\Ideal\)\nb-projective resolutions
\begin{equation}
\label{eq:projective_resolutions_Bi}
0 \to P_{i1} \xrightarrow{d_{i1}} P_{i0} \xrightarrow{d_{i0}} B_i
\end{equation}
in~\(\Tri\)
for \(i=0,1\)
(see \cite{Meyer-Nest:Homology_in_KK}*{Theorem~59}).
Since~\eqref{eq:projective_resolutions_Bi} is a resolution,
\(d_{i1}\)
is \(\Ideal\)\nb-monic
and \(d_{i0}\)
is \(\Ideal\)\nb-epic.
The arrow \(\varphi\in\Tri(B_1,B_0)\) lifts to a chain map
\begin{equation}
\label{eq:lift_varphi_resolutions}
\begin{tikzcd}[baseline=(current bounding box.west)]
P_{11} \arrow[r, rightarrowtail, "d_{11}"]
\arrow[d, "\Phi_1"] &
P_{10} \arrow[r, twoheadrightarrow, "d_{10}"]
\arrow[d, "\Phi_0"] & B_1
\arrow[d, rightarrowtail, "\varphi"] \\
P_{01} \arrow[r, rightarrowtail, "d_{01}"] &
P_{00} \arrow[r, twoheadrightarrow, "d_{00}"] & B_0
\end{tikzcd}
\end{equation}
between the \(\Ideal\)\nb-projective
resolutions~\eqref{eq:projective_resolutions_Bi} (see
\cite{Meyer-Nest:Homology_in_KK}*{Proposition~44}). We
write~\(\rightarrowtail\)
for \(\Ideal\)\nb-monic
and~\(\twoheadrightarrow\) for \(\Ideal\)\nb-epic maps. We claim that
\begin{equation}
\label{eq:resolution_A}
0 \to
P_{11} \xrightarrow{(-d_{11},\Phi_1)} P_{10} \oplus P_{01}
\xrightarrow{(\Phi_0,d_{01})} P_{00}
\xrightarrow{p\circ d_{00}} A
\to 0
\end{equation}
is an \(\Ideal\)\nb-projective
resolution of~\(A\)
of length~\(2\).
The entries are \(\Ideal\)\nb-projective
by construction. Next we prove that~\eqref{eq:resolution_A} is a
resolution, that is, it becomes an exact chain complex when we
apply~\(F\) to it.
When we apply~\(F\)
to the diagram~\eqref{eq:lift_varphi_resolutions}, the two rows
become short exact sequences in~\(\Abel\),
and the vertical maps become a chain map between them. The mapping
cone of this chain map is again an exact chain complex in~\(\Abel\).
It has the form
\[
0 \to F(P_{11})
\to F(P_{10}) \oplus F(P_{01})
\to F(P_{00}) \oplus F(B_1)
\to F(B_0) \to 0.
\]
The map \(F(\varphi)\colon F(B_1) \to F(B_0)\)
is monic with cokernel~\(F(A)\).
So the direct summand \(F(B_1)\)
and its image in \(F(B_0)\)
together form a contractible subcomplex. The quotient by it is
again an exact chain complex in~\(\Abel\).
This is what we get by applying~\(F\)
to~\eqref{eq:resolution_A}. So this is a resolution as asserted.
An axiom for triangulated categories provides an exact triangle
\[
\begin{tikzcd}[column sep=large]
P_{11} \arrow[r, rightarrowtail, "{(-d_{11},\Phi_1)}"] &
P_{10} \oplus P_{01} \arrow[r, twoheadrightarrow] &
D \arrow[r] & \Sigma P_{11}
\end{tikzcd}
\]
containing \((-d_{11},\Phi_1)\).
Similarly, the map \(p\circ d_{00}\colon P_{00}\to A\)
in~\eqref{eq:resolution_A} is part of an exact triangle
\begin{equation}
\label{eq:DP00A}
\begin{tikzcd}[column sep=large]
D' \arrow[r, rightarrowtail, "\gamma"] &
P_{00} \arrow[r, twoheadrightarrow, "p\circ d_{00}"] &
A \arrow[r] & \Sigma D'.
\end{tikzcd}
\end{equation}
The long exact sequences for~\(F\)
applied to these two exact triangles show that \(F(D)\)
is the cokernel of the monomorphism \(F(-d_{11},\Phi_1)\),
and that \(F(D')\)
is the kernel of the epimorphism \(F(p\circ d_{00})\).
The exactness of~\eqref{eq:resolution_A} implies
\(F(D)\cong F(D')\).
Since \(B_0\)
and~\(B_1\)
belong to \(\gen{\Proj_\Ideal}\),
so do \(D\)
and~\(D'\).
And \(F(D)\)
has a projective resolution of length~\(1\)
by construction. Hence the Universal Coefficient Theorem applies to \(D\)
and~\(D'\).
Thus the isomorphism \(F(D)\cong F(D')\)
lifts to an isomorphism \(D\cong D'\).
We shall identify \(D=D'\).
The Universal Coefficient Theorem for~\(D\) gives a short exact sequence
\begin{equation}
\label{eq:UCT_D_P00}
\begin{tikzcd}[column sep=scriptsize]
\Ext^1_{\Abel}\bigl(\Sigma F(D), F(P_{00})\bigr) \arrow[r, rightarrowtail] &
\Tri(D,P_{00}) \arrow[r, twoheadrightarrow, "F"] &
\Hom_{\Abel}\bigl(F(D),F(P_{00})\bigr).
\end{tikzcd}
\end{equation}
We split \(D=D^+\oplus D^-\)
and \(P_{00} = P_{00}^+ \oplus P_{00}^-\)
into objects of even and odd parity as in the construction
of~\(\varphi^1\)
above the theorem. Then \(\Tri(D,P_{00})\)
splits accordingly as a \(2\times2\)-matrix.
The sum of the diagonal terms
\(\Tri(D^+,P_{00}^+) \oplus\Tri(D^-,P_{00}^-)\)
is isomorphic to \(\Hom_{\Abel}\bigl(F(D),F(P_{00})\bigr)\),
whereas the sum of the off-diagonal terms
\(\Tri(D^-,P_{00}^+) \oplus\Tri(D^+,P_{00}^-)\)
is isomorphic to \(\Ext^1_{\Abel}\bigl(\Sigma F(D),F(P_{00})\bigr)\).
This is the (unnatural) splitting of the Universal Coefficient Theorem exact
sequence~\eqref{eq:UCT_D_P00} that follows because~\(\Abel\)
is paired. In particular, we decompose
\(\gamma = \gamma^0 + \gamma^1\)
into its parity-preserving and parity-reversing parts.
Let~\(A'\)
be another object of~\(\gen{\Proj_\Ideal}\)
with an isomorphism \(F(A') \cong F(A)\).
The argument above shows that both \(A\)
and~\(A'\)
are cones of some \(\gamma,\gamma'\in\Tri(D, P_{00})\)
as in~\eqref{eq:DP00A}, which lift the inclusion map
\(F(D) \rightarrowtail F(P_{00})\)
that we get from the resolution~\eqref{eq:resolution_A}. The
\emph{relative obstruction class} is defined as follows: compose
\[
\gamma-\gamma'\in \Ext^1_{\Abel}\bigl(\Sigma F(D), F(P_{00})\bigr) \subseteq
\Tri(D,P_{00})
\]
with the map \(F(p\circ d_{00})\colon F(P_{00}) \to F(A)\)
and apply the boundary map for the extension
\(F(D) \rightarrowtail F(P_{00}) \twoheadrightarrow F(A)\).
That is, plug \(\gamma-\gamma'\) into the map
\begin{multline}
\label{eq:transfer_to_Ext2}
\Ext^1_{\Abel}\bigl(\Sigma F(D),F(P_{00})\bigr) \xrightarrow{F(p\circ d_{00})_*}
\Ext^1_{\Abel}\bigl(\Sigma F(D),F(A)\bigr)
\\ \xrightarrow{\partial_{DP_{00}A}} \Ext^2_{\Abel}\bigl(\Sigma F(A),F(A)\bigr).
\end{multline}
Let \(\gamma^0\colon D\to P_{00}\)
be the unique parity-preserving arrow that lifts the inclusion map
\(F(D) \rightarrowtail F(P_{00})\)
and let~\(A^0\)
be its cone. The \emph{obstruction class} of~\(A\)
is the relative obstruction class for \(A\)
and~\(A^0\).
That is, we plug \(\gamma^1 \mathrel{\vcentcolon=} \gamma- \gamma^0\)
into~\eqref{eq:transfer_to_Ext2}.
Finally, we relate the obstruction class defined above
to~\(\varphi^1\).
The solid square in the following diagram commutes:
\begin{equation}
\label{eq:compare_triangles_BBA_DPA}
\begin{tikzcd}[baseline=(current bounding box.west)]
B_1 \arrow[r, rightarrowtail, "\varphi"] &
B_0 \arrow[r, twoheadrightarrow, "p"] &
A \arrow[r] &
\Sigma B_1\\
D \arrow[r, rightarrowtail, "\gamma"]
\arrow[u, dotted, "\varepsilon"] &
P_{00} \arrow[r, twoheadrightarrow, "p\circ d_{00}"]
\arrow[u, twoheadrightarrow, "d_{00}"]&
A \arrow[r]
\arrow[u, equal] &
\Sigma D \arrow[u, dotted, "\Sigma \varepsilon"]
\end{tikzcd}
\end{equation}
By the third axiom of triangulated categories, there is an
arrow~\(\varepsilon\)
making all three squares commute. We shall only use the left
square. Since \(D\),
\(B_1\),
\(B_0\)
and~\(P_{00}\)
have projective resolutions of length~\(1\)
and~\(\Abel\)
is paired, the Universal Coefficient Theorem allows us to split them
into even and odd parts. Hence each of the arrows \(\gamma\),
\(\varphi\),
\(\varepsilon\)
and~\(d_{00}\)
splits into a parity-preserving and a parity-reversing part. We
have already used the splittings \(\varphi=\varphi^0+\varphi^1\)
and \(\gamma=\gamma^0+\gamma^1\),
and now we also split \(\varepsilon=\varepsilon^0+\varepsilon^1\).
The arrow \(d_{00}\)
is parity-preserving because
\(\Tri(P_{00},B_0) \cong \Hom_{\Abel}\bigl(F(P_{00}),F(B_0)\bigr)\)
has no parity-reversing part. The left commuting square
in~\eqref{eq:compare_triangles_BBA_DPA} implies
\begin{equation}
\label{eq:odd_part_relation}
d_{00}\circ \gamma^1
= \varphi^0\circ \varepsilon^1 + \varphi^1\circ\varepsilon^0
= \varphi\circ \varepsilon^1 + \varphi^1\circ\varepsilon.
\end{equation}
The second step uses that the composite of two odd terms always
vanishes, that is, the \(\Ext^1\)-term in the Universal Coefficient Theorem is nilpotent.
Composing \(\gamma^1\in\Ext^1_{\Abel}\bigl(\Sigma F(D),F(P_{00})\bigr)\)
with \(F(p\circ d_{00})\)
as in~\eqref{eq:transfer_to_Ext2} has the same effect as composing
with \(p\circ d_{00}\)
because the exact sequence in the Universal Coefficient Theorem is natural. Thus the
obstruction class of~\(A\)
is the image of
\(p\circ d_{00}\circ \gamma^1 \in\Ext^1_{\Abel}\bigl(\Sigma F(D),F(A)\bigr)\)
under the boundary map
\[
\partial_{DP_{00}A}\colon
\Ext^1_{\Abel}\bigl(\Sigma F(D),F(A)\bigr) \to \Ext^2_{\Abel}\bigl(\Sigma F(A),F(A)\bigr)
\]
Equation~\eqref{eq:odd_part_relation} and \(p\circ\varphi=0\) imply
\[
p\circ d_{00}\circ\gamma^1
= p\circ \varphi\circ \varepsilon^1 + p\circ \varphi^1\circ\varepsilon
= p\circ \varphi^1\circ\varepsilon.
\]
The composite
\(p\circ \varphi^1 = p_*(\varphi^1) \in
\Ext^1_{\Abel}\bigl(\Sigma F(B_1),F(A)\bigr)\)
also appears in the statement of the theorem. Composing
with~\(\varepsilon\)
in~\(\Tri\)
has the same effect as composing with \(F(\varepsilon)\)
in the graded category \(\Ext^*_{\Abel}\).
When we apply~\(F\)
to the morphism of exact
triangles~\eqref{eq:compare_triangles_BBA_DPA}, we get the following
morphism of extensions in~\(\Abel\):
\[
\begin{tikzcd}[column sep=huge]
F(B_1) \arrow[r, rightarrowtail, "F(\varphi)"] &
F(B_0) \arrow[r, twoheadrightarrow, "F(p)"] &
F(A)\\
F(D) \arrow[r, rightarrowtail, "F(\gamma)"]
\arrow[u, dotted, "F(\varepsilon)"] &
F(P_{00}) \arrow[r, twoheadrightarrow, "F(p)\circ F(d_{00})"]
\arrow[u, twoheadrightarrow, "F(d_{00})"]&
F(A) \arrow[u, equal]
\end{tikzcd}
\]
Since boundary maps in Ext-theory are natural, the boundary map
\[
\partial\colon \Ext^1_{\Abel}\bigl(\Sigma F(B_1),F(A)\bigr) \to
\Ext^2_{\Abel}\bigl(\Sigma F(A),F(A)\bigr)
\]
for the top row is equal to the composite of \(F(\varepsilon)\)
and the boundary map~\(\partial_{DP_{00}A}\).
Thus the obstruction class of~\(A\)
is \(\partial(F(p)\circ\varphi^1)\) as asserted.
\end{proof}
\section{Comparison of classification theorems}
\label{sec:compare_classification}
In this section, we apply Theorem~\ref{the:compute_obstruction} in
several cases to relate the classification theorem involving the
obstruction class to other classification theorems. We first compare
the classification for \(\mathbb Z\)\nb-actions
in Theorem~\ref{the:Z_classify} with the one obtained
in~\cite{Bentmann-Meyer:More_general}. Then we compare the
classification for \(\textup C^*\)\nb-algebras
over a unique path space~\((X,\to)\)
in Corollary~\ref{cor:classify_KKX} with the one
in~\cite{Bentmann-Meyer:More_general}. In both cases, the invariants
can be translated into each other rather directly.
\subsection{Actions of the integers}
\label{sec:Z-action}
The Universal Coefficient Theorem for \(\mathbb Z\)\nb-actions
in Section~\ref{sec:UCT_Z} is based on the stable homological
ideal~\(\Ideal^\mathbb Z\)
defined by the forgetful functor \(\KKcat^\mathbb Z\to\KKcat\).
Now we use another ideal~\(\Ideal^\mathbb Z_\textup{K}\).
Let~\(\Abel^\mathbb Z\)
be the category of countable \(\mathbb Z/2\)-graded
\(\mathbb Z[x,x^{-1}]\)\nb-modules. Let
\[
F^\mathbb Z_\textup{K}\colon \KKcat^\mathbb Z\to\Abel^\mathbb Z,\qquad
(A,\alpha) \mapsto \bigl(\textup{K}_*(A),\textup{K}_*(\alpha)\bigr),
\]
that is, we map \((A,\alpha)\mathrel{\in\in} \KKcat^\mathbb Z\)
to the \(\mathbb Z/2\)\nb-graded
Abelian group \(\textup{K}_*(A)\)
with the \(\mathbb Z[x,x^{-1}]\)\nb-module
structure given by the automorphism \(\textup{K}_*(\alpha)\)
of \(\textup{K}_*(A)\). Let
\[
\Ideal^\mathbb Z_\textup{K}(A,B) \mathrel{\vcentcolon=}
\setgiven{\varphi\in \textup{KK}^\mathbb Z_0(A,B)}{F^\mathbb Z_\textup{K}(\varphi)=0}
\]
be the kernel of~\(F^\mathbb Z_\textup{K}\)
on morphisms. This example is treated
in~\cite{Bentmann-Meyer:More_general}, but there \(\KKcat^\mathbb Z\)
is disguised as \(\KKcat^\mathbb T\)
for the circle group~\(\mathbb T\).
These two categories are equivalent by Baaj--Skandalis duality
(see~\cite{Baaj-Skandalis:Hopf_KK}). The functor~\(F^\mathbb Z_\textup{K}\)
corresponds to the functor on \(\KKcat^\mathbb T\)
that maps a \(\textup C^*\)\nb-algebra
with a continuous \(\mathbb T\)\nb-action
to \(\textup{K}_*^\mathbb T(A)\)
with the canonical module structure over the representation ring
\(\mathbb Z[x,x^{-1}]\)
of~\(\mathbb T\), which is used in~\cite{Bentmann-Meyer:More_general}.
If \(B\mathrel{\in\in}\textup{KK}^\mathbb Z\), then~\eqref{eq:FZ_adjoint} implies
\[
\textup{KK}^\mathbb Z_0(\Cont_0(\mathbb Z), B)
\cong \textup{KK}_0(\mathbb C,B)
\cong \textup{K}_0(B)
\cong \Hom_{\Abel^\mathbb Z}\bigl(\mathbb Z[x,x^{-1}], F^\mathbb Z_\textup{K}(B)\bigr).
\]
Thus the partial adjoint of~\(F^\mathbb Z_\textup{K}\)
is defined on the rank-\(1\)
free module \(\mathbb Z[x,x^{-1}]\)
and maps it to \(\Cont_0(\mathbb Z)\).
Then it is defined on all free modules, also in odd parity.
Since idempotents in~\(\KKcat^\mathbb Z\)
split, the partial adjoint~\((F^\mathbb Z_\textup{K})^\lad\)
of~\(F^\mathbb Z_\textup{K}\)
is defined on all projective objects of~\(\Abel^\mathbb Z\).
Since \(F^\mathbb Z_\textup{K}(\Cont_0(\mathbb Z))\)
is the rank-\(1\)
free module again, we get
\((F^\mathbb Z_\textup{K}) \circ (F^\mathbb Z_\textup{K})^\lad(P) \cong P\)
for all projective objects of~\(\Abel^\mathbb Z\).
Any object in~\(\Abel^\mathbb Z\)
has a projective resolution of length~\(2\).
Hence it belongs to the image of~\(F^\mathbb Z_\textup{K}\)
by \cite{Bentmann-Meyer:More_general}*{Lemma~2.4}.
Remark~\ref{rem:universal_to_exact} and
Theorem~\ref{the:characterise_universal_homological} show
that~\(F^\mathbb Z_\textup{K}\)
is the universal \(\Ideal^\mathbb Z_\textup{K}\)\nb-exact
stable homological functor both to an exact and to an Abelian
category.
The category~\(\Abel^\mathbb Z\)
is paired in an obvious way, using the subcategories~\(\Abel^\mathbb Z_\pm\)
of countable \(\mathbb Z[x,x^{-1}]\)\nb-modules
concentrated in degree \(0\)
and~\(1\),
respectively. So the obstruction class is defined for any object
of~\(\KKcat^\mathbb Z\).
Assume that \((A,\alpha)\mathrel{\in\in} \KKcat^\mathbb Z\)
is such that~\(A\)
belongs to the bootstrap class in~\(\KKcat\).
Equivalently, \((A,\alpha)\)
belongs to the localising subcategory of~\(\KKcat^\mathbb Z\)
generated by the \(\Ideal^\mathbb Z_\textup{K}\)\nb-projective
object \(\Cont_0(\mathbb Z)\).
Then the main result of~\cite{Bentmann-Meyer:More_general} shows
that~\(A\)
is determined uniquely up to isomorphism by
\(\bigl(\textup{K}_*(A),\textup{K}_*(\alpha)\bigr)\mathrel{\in\in} \Abel^\mathbb Z\)
and the obstruction class.
A chain complex in~\(\KKcat^\mathbb Z\)
that is \(\Ideal^\mathbb Z\)\nb-exact
is also \(\Ideal^\mathbb Z_\textup{K}\)\nb-exact
because \(\Ideal^\mathbb Z \subseteq \Ideal^\mathbb Z_\textup{K}\).
So the \(\Ideal^\mathbb Z\)\nb-projective
resolution in~\eqref{eq:lift_AZ_resolution} is
\(\Ideal^\mathbb Z_\textup{K}\)\nb-exact.
Its entries \(\Cont_0(\mathbb Z,A)\)
are no longer \(\Ideal^\mathbb Z_\textup{K}\)\nb-projective.
We claim, however, that they have \(\Ideal^\mathbb Z_\textup{K}\)\nb-projective
resolutions of length~\(1\).
Let \(P_1 \rightarrowtail P_0\twoheadrightarrow \textup{K}_*(A)\)
be a resolution of the \(\mathbb Z/2\)\nb-graded
Abelian group~\(\textup{K}_*(A)\) by countable free Abelian groups. Then
\[
\mathbb Z[x,x^{-1}] \otimes P_1 \rightarrowtail \mathbb Z[x,x^{-1}] \otimes P_0
\twoheadrightarrow \mathbb Z[x,x^{-1}] \otimes \textup{K}_*(A)
\]
is a projective resolution of
\[
F^\mathbb Z_\textup{K}(\Cont_0(\mathbb Z,A)) \cong \mathbb Z[x,x^{-1}] \otimes \textup{K}_*(A).
\]
Here the tensor products with \(\mathbb Z[x,x^{-1}]\)
carry the module structure defined by multiplication with~\(x\)
in the tensor factor~\(\mathbb Z[x,x^{-1}]\).
So we are in the situation of~\eqref{eq:triangle_B1B0A}.
Theorem~\ref{the:compute_obstruction} implies the following formula
for the obstruction class of~\(A\):
\begin{theorem}
\label{the:obstruction_class_Z}
Let~\(A\)
be a \(\textup C^*\)\nb-algebra
in the bootstrap class in~\(\KKcat\)
and \(\alpha\in\Aut(A)\).
Split \([\alpha]\in \textup{KK}_0(A,A)\)
into parity-preserving and parity-reversing parts
\(\alpha^0\in \Hom\bigl(\textup{K}_*(A),\textup{K}_*(A)\bigr)\)
and \(\alpha^1\in \Ext\bigl(\textup{K}_{1+*}(A),\textup{K}_*(A)\bigr)\). Define
\[
\gamma\colon \Ext^1_\mathbb Z\bigl(\textup{K}_{1+*}(A),\textup{K}_*(A)\bigr) \to
\Ext^1_\mathbb Z\bigl(\textup{K}_{1+*}(A),\textup{K}_*(A)\bigr),\qquad
x\mapsto \alpha\circ x - x\circ\alpha.
\]
Then
\(\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\Sigma F^\mathbb Z_\textup{K}(A),F^\mathbb Z_\textup{K}(A)\bigr)
\cong \coker \gamma\)
and the obstruction class of~\(A\)
is the image of \(-\alpha^1 (\alpha^0)^{-1}\) in this cokernel.
\end{theorem}
\begin{proof}
In our case, the map~\(\varphi\)
in~\eqref{eq:triangle_B1B0A} is the map
\(\Cont_0(\mathbb Z) \otimes A \to \Cont_0(\mathbb Z) \otimes A\)
in~\eqref{eq:lift_AZ_resolution}. Split \(A=A^+ \oplus A^-\)
into its even and odd parts as before. Then
\[
\Cont_0(\mathbb Z) \otimes A \cong \Cont_0(\mathbb Z) \otimes A^+ \oplus \Cont_0(\mathbb Z)\otimes A^-
\]
is the parity decomposition of~\(\Cont_0(\mathbb Z) \otimes A\).
The translation~\(\tau\)
and the identity are parity-preserving. Decompose \([\alpha]\)
and~\([\alpha^{-1}]\)
into their even and odd parts. The parity-reversing part of the map
on \(\Cont_0(\mathbb Z) \otimes A\)
is \(\varphi^1 = [\tau]\otimes [\alpha^{-1}]^1\).
The map \(p\in \textup{KK}^\mathbb Z_0(\Cont_0(\mathbb Z) \otimes A,A)\)
satisfies \(p (\tau\otimes \textup{id}_A) = [\alpha] p\)
because it is \(\mathbb Z\)\nb-equivariant,
and it restricts to the identity map on the \(0\)th
summand of \(\Cont_0(\mathbb Z) \otimes A\).
The isomorphism
\(\textup{KK}^\mathbb Z_0\bigl((\Cont_0(\mathbb Z,A),\tau), (B,\beta) \bigr) \cong
\textup{KK}_0(A,B)\)
in~\eqref{eq:FZ_adjoint} forgets the \(\mathbb Z\)\nb-action
and then evaluates at~\(0\).
Therefore, the image of~\(\varphi^1\)
in \(\textup{KK}_0(A,A)\)
is the restriction of \(p\circ \varphi^1\)
to the \(0th\)
summand. And this is \([\alpha] [\alpha^{-1}]^1 \in \textup{KK}_0(A,A)\).
Since \(\alpha \alpha^{-1} = \textup{id}_A\)
and products of two parity-reversing \(\textup{KK}\)\nb-classes
always vanish,
\([\alpha] [\alpha^{-1}]^1 + [\alpha]^1 [\alpha^{-1}] = [\textup{id}_A]^1 =
0\).
Theorem~\ref{the:compute_obstruction} now says that the obstruction
class for~\(A\)
is the image of
\[
[\alpha] [\alpha^{-1}]^1 = - [\alpha]^1 [\alpha^{-1}] = -\alpha^1 (\alpha^0)^{-1}
\in \Ext_\mathbb Z\bigl(\textup{K}_{1+*}(A),\textup{K}_*(A))\bigr)
\]
under the boundary map to
\(\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\Sigma F^\mathbb Z_\textup{K}(A), F^\mathbb Z_\textup{K}(A)\bigr)\).
The description of the latter Ext group in the theorem follows when
we compute it with the projective resolution
in~\eqref{eq:resolution_A} defined by the length-\(1\)
resolutions of \(\Cont_0(\mathbb Z) \otimes A\)
above. This computation has already been done in
\cite{Bentmann-Meyer:More_general}*{Section~3.2}. When the Ext
groups are described in this way, the boundary map in
Theorem~\ref{the:compute_obstruction} becomes a trivial map, mapping
an element of \(\Ext_\mathbb Z\bigl(\textup{K}_{1+*}(A),\textup{K}_*(A)\bigr)\)
to its image in \(\coker \gamma\).
\end{proof}
The formula for the obstruction class depends, of course, on the
isomorphism
\(\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\Sigma F^\mathbb Z_\textup{K}(A),F^\mathbb Z_\textup{K}(A)\bigr) \cong
\coker \gamma\),
and this depends on the \(\Ideal^\mathbb Z\)\nb-projective
resolution from which it is obtained. Our theorem uses the most
obvious resolution. One may apply the automorphism of
\(\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\Sigma F^\mathbb Z_\textup{K}(A),F^\mathbb Z_\textup{K}(A)\bigr)\)
that composes with the automorphism~\(-\Sigma\alpha^0\)
on \(\Sigma F^\mathbb Z_\textup{K}(A)\).
This replaces the obstruction class~\(-\alpha^1 (\alpha^0)^{-1}\)
by~\(\alpha^1\).
So the images of \(-\alpha^1 (\alpha^0)^{-1}\)
and~\(\alpha^1\)
in \(\Ext^2_{\mathbb Z[x,x^{-1}]}\bigl(\Sigma F^\mathbb Z_\textup{K}(A),F^\mathbb Z_\textup{K}(A)\bigr)\)
contain the same information.
Theorem~\ref{the:obstruction_class_Z} and the computations after
Theorem~\ref{the:Z_classify} allow to deduce the classification
by~\(F^\mathbb Z_\textup{K}\)
and the obstruction class from the classification by the
invariant~\(F^\mathbb Z\)
in Section~\ref{sec:UCT_Z}. Let \((A,\alpha)\)
and \((B,\beta)\)
belong to the bootstrap class in~\(\KKcat^\mathbb Z\).
Assume that there is an isomorphism
\(t^0\colon F^\mathbb Z_\textup{K}(A) \xrightarrow\sim F^\mathbb Z_\textup{K}(B)\)
in~\(\Abel^\mathbb Z\),
that is, a grading-preserving \(\mathbb Z[x,x^{-1}]\)-module
isomorphism \(\textup{K}_*(A) \cong \textup{K}_*(B)\).
We use this isomorphism to identify \(\textup{K}_*(A) = \textup{K}_*(B)\).
By Theorem~\ref{the:obstruction_class_Z}, the relative obstruction
class vanishes if and only if
\(\alpha^1 - \beta^1 \in \Ext\bigl(\textup{K}_{*+1}(A),\textup{K}_*(A)\bigr)\)
vanishes in \(\coker(\gamma)\).
Equivalently, there is \(t^1\in\Ext^1\bigl(\textup{K}_{1+*}(A),\textup{K}_*(B)\bigr)\)
for which \(t=t^0+t^1\in \textup{KK}_0(A,B)\)
satisfies \([\beta]\circ t = t\circ [\alpha]\).
Then~\(t\)
is an isomorphism \(F^\mathbb Z(A) \cong F^\mathbb Z(B)\)
in~\(\KKcat[\mathbb Z]\),
and Theorem~\ref{the:Z_classify} shows that such an isomorphism lifts
to an isomorphism in \(\textup{KK}^\mathbb Z_0(A,B)\).
Recall that Baaj--Skandalis duality is an equivalence of triangulated
categories \(\KKcat^\mathbb T \cong \KKcat^\mathbb Z\).
Hence everything said above about \(\mathbb Z\)\nb-actions
carries over to \(\mathbb T\)\nb-actions.
The functor \(A\mapsto \textup{K}_*(A)\)
becomes \(B\mapsto \textup{K}_*^\mathbb T(B)\)
on \(\KKcat^\mathbb T\),
equipped with the natural module structure over the representation
ring~\(\mathbb Z[x,x^{-1}]\)
of~\(\mathbb T\).
The automorphism~\(\alpha\)
is replaced by an automorphism of \(B\rtimes \mathbb T\),
namely, the generator~\(\beta\)
of the dual action of~\(\mathbb Z\).
Since \(\textup{K}_*(B\rtimes\mathbb T) \cong \textup{K}_*^\mathbb T(B)\),
the Universal Coefficient Theorem splits
\(\textup{KK}_0(B\rtimes\mathbb T,B\rtimes\mathbb T)\)
into the parity-preserving part
\(\Hom\bigl(\textup{K}_*^\mathbb T(B),\textup{K}_*^\mathbb T(B)\bigr)\)
and the parity-reversing part
\(\Ext\bigl(\textup{K}_{1+*}^\mathbb T(B),\textup{K}_*^\mathbb T(B)\bigr)\).
The obstruction class of \(B\rtimes\mathbb T\)
is the class of
\(-\beta^1 (\beta^0)^{-1} \in
\Ext\bigl(\textup{K}_{1+*}^\mathbb T(B),\textup{K}_*^\mathbb T(B)\bigr)\) in the cokernel of the map
\[
\gamma\colon \Ext\bigl(\textup{K}_{1+*}^\mathbb T(B),\textup{K}_*^\mathbb T(B)\bigr) \to
\Ext\bigl(\textup{K}_{1+*}^\mathbb T(B),\textup{K}_*^\mathbb T(B)\bigr),\qquad
x\mapsto \beta^0 x - x\beta^0.
\]
As above, a grading-preserving \(\mathbb Z[x,x^{-1}]\)-module
isomorphism \(t\colon \textup{K}_*^\mathbb T(A) \to \textup{K}_*^\mathbb T(B)\)
is compatible with the obstruction classes if and only if it lifts to
an isomorphism between \(A\rtimes\mathbb T\)
and~\(B\rtimes\mathbb T\)
in \(\KKcat[\mathbb Z]\),
and such an isomorphism lifts further to an isomorphism between
\(A\rtimes\mathbb T\)
and~\(B\rtimes\mathbb T\)
in \(\KKcat^\mathbb Z\).
By Baaj--Skandalis duality, the latter is equivalent to an invertible
element in \(\textup{KK}^\mathbb T_0(A,B)\).
\subsection{\texorpdfstring{$\textup C^*$}{C*}-Algebras over unique path spaces}
\label{sec:compare_ups}
Now we return to the setup of Section~\ref{sec:UCT_X}. So \((X,\to)\)
is a countable directed graph with the unique path property. Let~\(\preceq\)
be the partial order generated by~\(\leftarrow\)
and equip~\(X\)
with the Alexandrov topology defined by~\(\preceq\).
Let \(\KKcat^X\)
be the category of \(\textup C^*\)\nb-algebras
over~\(X\).
Objects in the appropriate bootstrap class in \(\KKcat^X\)
are classified in~\cite{Bentmann-Meyer:More_general} under the extra
assumption that~\(X\)
be finite. Actually, the arguments
in~\cite{Bentmann-Meyer:More_general} work in the same way if the
set~\(X\)
is countable. Here we treat this more general case right away.
Let~\(\Abel^X\)
be the category of all functors from~\(X\)
to the category of countable \(\mathbb Z/2\)-graded
Abelian groups. Equivalently, an object of~\(\Abel^X\)
is a family of countable \(\mathbb Z/2\)\nb-graded
Abelian groups~\(G_x\)
for \(x\in X\)
with grading-preserving group homomorphisms
\(\gamma_{y,x}\colon G_x \to G_y\)
for all \(x,y\in X\)
with \(x\to y\).
Morphisms \((G_x,\gamma_{y,x}) \to (H_x,\eta_{y,x})\)
in~\(\Abel^X\)
are families of grading-preserving group homomorphisms
\(t_x\colon G_x \to H_x\)
that satisfy \(t_y\circ \gamma_{y,x} = \eta_{y,x}\circ t_x\)
for all \(x,y\in X\) with \(x\to y\). We define the functor
\[
F^X_\textup{K}\colon \KKcat^X \to \Abel^X
\]
by mapping a \(\textup C^*\)\nb-algebra
over~\(X\)
to the diagram of \(\mathbb Z/2\)\nb-graded
Abelian groups \(\textup{K}_*(A(U_x))\)
for \(x\in X\)
with the maps induced by the inclusion maps \(A(U_x) \idealin A(U_y)\)
for \(x\to y\).
The target category~\(\Abel^X\)
is a paired, stable Abelian category, where
\(\Abel^X_\pm \subseteq \Abel^X\)
are the subcategories of \(\mathbb Z/2\)\nb-graded
groups where the odd or even part vanishes, respectively.
And~\(F^X_\textup{K}\)
is a stable homological functor. Let \(\Ideal^X_\textup{K}\)
be its kernel on morphisms. Equation~\eqref{eq:KKX_out_Fix2} implies
\[
\textup{KK}^X_0(i_z(\mathbb C),B)
\cong \textup{KK}_0\bigl(\mathbb C,B(U_z)\bigr)
\cong \textup{K}_0\bigl(B(U_z)\bigr)
\cong \Hom_{\Abel^X}\bigl(j_z(\mathbb C),F^X_\textup{K}(B)\bigr).
\]
Hence the partial adjoint~\((F^X_\textup{K})^\lad\)
of~\(F^X_\textup{K}\)
is defined on \(j_z(\mathbb C)\)
for all \(z\in X\).
As in Section~\ref{sec:UCT_X}, any object of~\(\Abel^X\)
is a quotient of a direct sum of objects of the form \(j_z(\mathbb C)\)
for \(z\in X\).
Hence~\((F^X_\textup{K})^\lad\)
is defined on all projective objects of~\(\Ideal^X_\textup{K}\)
and \(F^X_\textup{K} \circ (F^X_\textup{K})^\lad(P)=P\)
for all projective objects~\(P\)
of~\(\Abel^X\);
this is proved like the corresponding statement about~\(F^\mathbb Z_\textup{K}\)
in Section~\ref{sec:Z-action}. Therefore, the functor
\(F^X_\textup{K}\colon \KKcat^X \to \Abel^X\)
is the universal \(\Ideal^X_\textup{K}\)\nb-exact
stable homological functor into an exact category or into an Abelian
category by Theorem~\ref{the:characterise_universal_homological} and
Remark~\ref{rem:universal_to_exact}. And~\(F^X_\textup{K}\)
restricts to an equivalence of categories between the
\(\Ideal^X_\textup{K}\)\nb-projective
objects in~\(\KKcat^X\)
and the projective objects in~\(\Abel^X\).
Let \(\Boot^X \subseteq\KKcat^X\)
be the localising subcategory generated by the
\(\Ideal^X_\textup{K}\)\nb-projective
objects. This is the analogue of the bootstrap class in~\(\KKcat^X\).
\begin{lemma}
\label{lem:Ext_2_over_X}
Let \(G=(G_x,\gamma_{y,x})\)
and \(H=(H_x,\eta_{y,x})\) be objects of~\(\Abel^X\).
\begin{enumerate}
\item There is a projective resolution for~\(G\)
of length~\(2\) in~\(\Abel^X\).
\item There is \(A\in\Boot^X\) with \(F^X_\textup{K}(A) \cong G\).
\item Write \(\Ext\)
for the \(\Ext^1\)
of Abelian groups. The group \(\Ext^2_{\Abel^X}(G,H)\)
is naturally isomorphic to the cokernel of the map
\begin{multline}
\label{eq:cokernel_Ext_2_X}
\prod_{x\in X} \Ext(G_x,H_x)
\to \prod_{x\to y} \Ext(G_x,H_y),\\
(t_x)_{x\in X} \mapsto
(\eta_{y,x}\circ t_x - t_y\circ \gamma_{y,x})_{x\to y},
\end{multline}
\end{enumerate}
\end{lemma}
\begin{proof}
Since \(\Ideal^X \subseteq \Ideal^X_\textup{K}\),
any \(\Ideal^X\)\nb-exact
chain complex in~\(\KKcat^X\)
is also \(\Ideal^X_\textup{K}\)\nb-exact.
In particular, the \(\Ideal^X\)\nb-projective
resolution in~\eqref{eq:projective_resolution_ups} is
\(\Ideal^X_\textup{K}\)\nb-exact.
We claim that its entries have \(\Ideal^X_\textup{K}\)\nb-projective
resolutions of length~\(1\).
Hence there is a projective resolution of~\(F^X_\textup{K}(A)\)
of length~\(2\)
as in~\eqref{eq:resolution_A}. To prove the same for all
objects~\(G\)
of~\(\Abel^X\), we carry over~\eqref{eq:projective_resolution_ups}.
Let \(J_z(B)\mathrel{\in\in}\Abel^\mathbb Z\)
for a countable \(\mathbb Z/2\)\nb-graded
Abelian group~\(B\)
be the diagram with \(J_z(B)_x = B\)
if \(z\succeq x\)
and \(J_z(B)_x = 0\)
otherwise, with the identity map \(J_z(B)_x \to J_z(B)_y\)
for \(z\succeq x\to y\) and the zero map otherwise. Then
\begin{equation}
\label{eq:Abel_X_jz_adjunction}
\Hom_{\Abel^X}\bigl(J_z(B),H\bigr) \cong \Hom\bigl(B,H_z\bigr),
\end{equation}
where the isomorphism simply restricts a morphism in~\(\Abel^X\)
to the object \(z\in X\).
In particular, if \(y\succeq z\),
then \(J_y(B)_z = B\)
and so there is a canonical map
\(J_{y\succeq z}(B)\colon J_z(B) \to J_y(B)\)
in~\(\Abel^X\).
Explicitly, this map is the identity map on \(J_z(B)_x\)
if \(z\succeq x\)
and the zero map on \(J_z(B)_x=0\)
if \(z\nsucceq x\).
The proof that~\eqref{eq:projective_resolution_ups} is exact also
proves the exactness of
\begin{equation}
\label{eq:projective_resolution_ups_Abel}
0 \to
\bigoplus_{x\to y} J_y(G_x)
\xrightarrow{\psi} \bigoplus_{x\in X} J_x(G_x)
\xrightarrow{q} G
\to 0;
\end{equation}
here~\(\psi\)
restricted to the summand \(J_y(G_x)\)
for \(x\to y\)
is the map \((J_{x\to y}(G_x),J_y(\gamma_{y,x}))\)
to \(J_x(G_x) \oplus J_y(G_y)\),
and~\(q\)
maps \(J_x(G_x)\)
to~\(G\)
by the adjunct of the identity map on~\(G_x\)
under the adjunction in~\eqref{eq:Abel_X_jz_adjunction}.
Explicitly, the entry of \(\bigoplus_{x\in X} J_x(G_x)\)
at \(z\in X\)
is \(\bigoplus_{x\succeq z} G_x\),
which is mapped to~\(G_z\)
by \((\gamma_{z,x})_{x\succeq z}\).
The proof that~\eqref{eq:projective_resolution_ups_Abel} is exact
shows that the chain complexes of \(\mathbb Z/2\)\nb-graded Abelian groups
\[
0 \to
\bigoplus_{x\to y} J_y(G_x)_z
\xrightarrow{\psi_z} \bigoplus_{x\in X} J_x(G_x)_z
\xrightarrow{q_z} G_z
\to 0
\]
are contractible for all \(z\in X\).
For each \(x\in X\), there is a resolution
\begin{equation}
\label{eq:resolve_Gx}
P_{x,1} \xinto{d_{x,1}} P_{x,0} \xprto{d_{x,0}} G_x
\end{equation}
of~\(G_x\)
of length~\(1\)
by countable free \(\mathbb Z/2\)\nb-graded
Abelian groups. The homomorphism \(\gamma_{y,x}\colon G_x \to G_y\)
for \(x \to y\) lifts to a morphism of extensions
\begin{equation}
\label{eq:resolve_Gx_Gy}
\begin{tikzcd}
P_{x,1} \arrow[r, rightarrowtail, "d_{x,1}"] \arrow[d, "\gamma^1_{y,x}"]&
P_{x,0} \arrow[r, twoheadrightarrow, "d_{x,0}"] \arrow[d, "\gamma^0_{y,x}"]&
G_x \arrow[d, "\gamma_{y,x}"]\\
P_{y,1} \arrow[r, rightarrowtail, "d_{y,1}"] &
P_{y,0} \arrow[r, twoheadrightarrow, "d_{y,0}"] &
G_y
\end{tikzcd}
\end{equation}
The construction~\(J_z\)
above is an exact functor, and it maps free Abelian groups to
projective objects of~\(\Abel^X\)
by the adjunction~\eqref{eq:Abel_X_jz_adjunction}. Hence
\begin{align*}
\bigoplus_{x\to y} J_y(P_{x,1})
\rightarrowtail \bigoplus_{x\to y} J_y(P_{x,0})
\twoheadrightarrow \bigoplus_{x\to y} J_y(G_x),\\
\bigoplus_{x\in X} J_x(P_{x,1})
\rightarrowtail \bigoplus_{x\in X} J_x(P_{x,0})
\twoheadrightarrow \bigoplus_{x\in X} J_sx(G_x)
\end{align*}
are projective resolutions of length~\(1\)
in~\(\Abel^X\).
The resolution in~\eqref{eq:projective_resolution_ups_Abel}
gives a projective resolution of~\(G\)
of length~\(2\)
as in~\eqref{eq:resolution_A}. This proves the first assertion.
Now \cite{Bentmann-Meyer:More_general}*{Lemma 2.4} shows that there
is \(A\mathrel{\in\in}\Boot^X\)
with \(G=F^X_\textup{K}(A)\);
the property that \(\textup{KK}^X_0(A,B)=0\)
for all \(\Ideal^X_\textup{K}\)\nb-contractible
\(B\mathrel{\in\in}\KKcat^X\) is equivalent to \(A\mathrel{\in\in}\Boot^X\).
The projective resolution of~\(G\) built above has the form
\[
0 \to
\bigoplus_{x\to y} J_y(P_{x,1})
\to \bigoplus_{x\to y} J_y(P_{x,0}) \oplus \bigoplus_{x\in X} J_x(P_{x,1})
\to \bigoplus_{x\in X} J_x(P_{x,0})
\to G,
\]
where the maps
\(\bigoplus_{x\to y} J_y(P_{x,i}) \to \bigoplus_{x\in X}
J_x(P_{x,i})\)
for \(i=0,1\)
use \(\gamma_{y,x}^i \colon P_{x,i} \to P_{y,i}\)
in~\eqref{eq:resolve_Gx_Gy}. We use this resolution to compute the
group \(\Ext^2_{\Abel^X}(G,H)\).
An element of \(\Ext^2_{\Abel^X}(G,H)\)
is represented by a map \(\bigoplus_{x\to y} J_y(P_{x,1}) \to H\).
By the adjunction~\eqref{eq:Abel_X_jz_adjunction}, this corresponds
to a family of maps \(f_{x\to y}\colon P_{x,1} \to H_y\).
This family represents~\(0\)
in \(\Ext^2_{\Abel^X}(G,H)\)
if and only if the corresponding map
\(\bigoplus_{x\to y} J_y(P_{x,1}) \to H\)
factors through
\(\bigoplus_{x\to y} J_y(P_{x,0}) \oplus \bigoplus_{x}
J_x(P_{x,1})\).
Using the adjunction~\eqref{eq:Abel_X_jz_adjunction} again, a map on
this direct sum corresponds to families of maps
\(g_{x\to y}\colon P_{x,0} \to H_y\)
and \(h_x\colon P_{x,1} \to H_x\).
The resulting map \(\bigoplus_{x\to y} J_y(P_{x,1}) \to H\)
corresponds to the family of maps
\[
-g_{x\to y} \circ d_{x,1} + \eta_{y,x} h_x - h_y \gamma_{y,x}^1\colon
P_{x,1} \to H_y.
\]
The resolutions~\eqref{eq:resolve_Gx} compute \(\Ext(G_x,D)\)
for any \(\mathbb Z/2\)\nb-graded
Abelian group~\(D\).
So each \(f_{x\to y}\colon P_{x,1} \to H_y\)
represents an element of \(\Ext(G_x,H_y)\),
and it represents the zero element if and only if it is of the form
\(g_{x\to y} \circ d_{x,1}\)
for some \(g_{x\to y}\colon P_{x,0} \to H_y\).
The elements~\(h_x\)
above represent elements of \(\Ext(G_x,H_x)\).
If they represent the zero element of \(\Ext(G_x,H_x)\),
then the term \(\eta_{y,x} h_x - h_y \gamma_{y,x}^1\)
above may be rewritten in the form \(-g_{x\to y} \circ d_{x,1}\).
Now we get the formula for \(\Ext^2_{\Abel^X}(G,H)\)
in the third statement in the lemma.
\end{proof}
\begin{theorem}
\label{the:classification_X_obstruction}
Let \(A\)
and~\(B\)
belong to~\(\Boot^X\).
An isomorphism \(t\colon F^X_\textup{K}(A) \xrightarrow\sim F^X_\textup{K}(B)\)
in~\(\Abel^X\)
lifts to an invertible element in \(\textup{KK}^X_0(A,B)\)
if and only if \(t\delta_A = \delta_B t\)
holds in \(\Ext^2_{\Abel^X}\bigl(\Sigma F^X_\textup{K}(A),F^X_\textup{K}(B)\bigr)\),
where \(\delta_A\)
and~\(\delta_B\) are the obstruction classes of \(A\) and~\(B\).
\end{theorem}
\begin{proof}
The lemma verifies all the conditions to apply the classification
method of~\cite{Bentmann-Meyer:More_general}.
\end{proof}
The proof of the lemma also gives all the ingredients needed in
Section~\ref{sec:triangulated}. So we may now compute obstruction
classes:
\begin{theorem}
\label{the:obstruction_class_X}
Let~\(B\)
be an object of~\(\Boot^X\).
Let \(B_x \mathrel{\vcentcolon=} B(U_x)\)
for \(x\in X\).
Let \(\beta_{y,x} \in \textup{KK}_0(B_x,B_y)\)
be the \(\textup{KK}\)\nb-class
of the inclusion map \(B_x \hookrightarrow B_y\).
Split \(\beta_{y,x} = \beta_{y,x}^0 + \beta_{y,x}^1\)
with a parity-preserving part
\(\beta_{y,x}^0 \in \Hom\bigl(\textup{K}_*(B_x),\textup{K}_*(B_y)\bigr)\)
and a parity-reversing part
\(\beta_{y,x}^1 \in \Ext\bigl(\textup{K}_{1+*}(B_x),\textup{K}_*(B_y)\bigr)\).
The obstruction class of~\(B\)
is the class in the cokernel of the map
in~\eqref{eq:cokernel_Ext_2_X} that is represented by
\[
(\beta_{y,x}^1)_{x\to y} \in \prod_{x\to y} \Ext\bigl(\textup{K}_{1+*}(B_x),\textup{K}_*(B_y)\bigr).
\]
\end{theorem}
\begin{proof}
The projective resolution~\eqref{eq:projective_resolution_ups}
in~\(\KKcat[X]\) lifts to an exact triangle
\[
\bigoplus_{x\to y} i_y(B_x) \xrightarrow{\varphi}
\bigoplus_{x\in X} i_x(B_x) \xrightarrow{p}
B \to \Sigma \bigoplus_{x\to y} i_y(B_x)
\]
in \(\KKcat^X\)
by \cite{Bentmann-Meyer:More_general}*{Proposition 2.3}. This is
how the Universal Coefficient Theorem in
\cite{Meyer-Nest:Homology_in_KK}*{Theorem~66} is proved. Here the
map~\(\varphi\)
restricted to the summand~\(i_y(B_x)\)
is the difference of two maps: the map \(i_y(\beta_{y,x})\)
to~\(i_y(B_y)\)
and the canonical map \(i_{x,y}(B_x)\colon i_y(B_x) \to i_x(B_x)\)
that is the adjunct of the identity map \(B_x \to i_x(B_x)_y = B_x\)
under the adjunction~\eqref{eq:KK_out_ix}. And the map~\(p\)
restricted to~\(i_x(B_x)\)
is the adjunct of the identity map \(B_x \to B_x\)
under the adjunction~\eqref{eq:KK_out_ix}.
Both \(\bigoplus_{x\to y} i_y(B_x)\)
and \(\bigoplus_{x\in X} i_x(B_x)\)
belong to~\(\Boot^X\)
and have \(\Ideal^X_\textup{K}\)\nb-projective
resolutions of length~\(1\)
(see~\eqref{eq:resolve_Gx}). So we are in the situation
of~\eqref{eq:triangle_B1B0A}. We split each \(B_x \mathrel{\vcentcolon=} B(U_x)\)
into its even and odd parity part \(B_x = B_x^+ \oplus B_x^-\).
Then \(i_y(B_x^+)\)
and \(i_y(B_x^-)\)
are of even or odd parity, respectively. Split
\(\beta_{y,x} = \beta_{y,x}^0 + \beta_{y,x}^1\)
into a parity-preserving and a parity-reversing part. So
\(\beta_{y,x}^0 \in \Hom\bigl(\textup{K}_*(B_x),\textup{K}_*(B_y)\bigr)\)
and \(\beta_{y,x}^1 \in \Ext\bigl(\textup{K}_{1+*}(B_x),\textup{K}_*(B_y)\bigr)\)
by the Universal Coefficient Theorem for \(\textup{KK}\),
see the discussion after Theorem~\ref{the:Z_classify}. The induced
maps \(i_y(\beta_{y,x}^0)\)
and \(i_y(\beta_{y,x}^1)\)
are parity-preserving and parity-reversing, respectively. And the
map \(i_{x,y}(B_x)\)
is parity-preserving. So the parity-reversing part~\(\varphi^1\)
of~\(\varphi\)
is the map that restricts to
\(i_y(\beta_{y,x}^1)\colon i_y(B_x) \to i_y(B_y)\)
on the summand for \(x\to y\).
The composite \(p\circ \varphi^1\)
is the map \(\bigoplus_{x\to y} i_y(B_x) \to B\)
whose restriction to the summand \(i_y(B_x)\)
is adjunct to \(\beta_{y,x}^1\colon B_x \to B_y\).
These maps define an element of
\(\prod_{x\to y} \Ext\bigl(\textup{K}_{1+*}(B_x),\textup{K}_*(B_y)\bigr)\).
The obstruction class of~\(B\)
is its image under the boundary map to
\(\Ext^2_{\Abel^X}\bigl( \Sigma F^X_\textup{K}(B),F^X_\textup{K}(B)\bigr)\)
by Theorem~\ref{the:compute_obstruction}. We have described
\(\Ext^2_{\Abel^X}(G,H)\)
in Lemma~\ref{lem:Ext_2_over_X} in such a way that this boundary map
becomes tautological: it simply maps an element of
\(\prod_{x\to y} \Ext(G_x,H_y)\)
to its class in the cokernel of the map
in~\eqref{eq:cokernel_Ext_2_X}. In particular, the obstruction
class of~\(B\)
is represented by
\((\beta_{y,x}^1)_{x\to y} \in \prod_{x\to y}
\Ext\bigl(\textup{K}_{1+*}(B_x),\textup{K}_*(B_y)\bigr)\).
\end{proof}
Theorem~\ref{the:obstruction_class_X} allows us to compare the
classification theorems for \(\textup C^*\)\nb-algebras
in~\(\Boot^X\)
that use the invariant \(F^X(B)\)
or \(F^X_\textup{K}(B)\)
with the obstruction class. Let \(A,B\mathrel{\in\in}\Boot^X\).
An isomorphism
\[
t^0\colon F^X_\textup{K}(A) \xrightarrow\sim F^X_\textup{K}(B)
\]
is equivalent to a family of isomorphisms
\[
t^0_x\colon \textup{K}_*(A(U_x)) \xrightarrow\sim \textup{K}_*(B(U_x))
\]
that make the diagrams
\[
\begin{tikzcd}
\textup{K}_*(A(U_x)) \arrow[d, "\textup{K}_*(\alpha_{y,x})"]
\arrow[r, "t^0_x", "\cong"'] &
\textup{K}_*(B(U_x)) \arrow[d, "\textup{K}_*(\beta_{y,x})"] \\
\textup{K}_*(A(U_y)) \arrow[r, "t^0_y", "\cong"'] &
\textup{K}_*(B(U_y))
\end{tikzcd}
\]
commute for all edges \(x\to y\).
Here the maps \(\textup{K}_*(\alpha_{y,x})\)
and \(\textup{K}_*(\beta_{y,x})\)
are induced by the inclusion maps of our \(\textup C^*\)\nb-algebras
over~\(X\).
The obstruction class for the isomorphism~\((t^0_x)_{x\in X}\)
vanishes if and only if there are
\(t^1_x\in \Ext\bigl(\textup{K}_{1+*}(A(U_x)),\textup{K}_*(B(U_x))\bigr)\)
for \(x\in X\) such that
\begin{equation}
\label{eq:t_condition_X}
t^1_y \circ \textup{K}_*(\alpha_{y,x}) - \textup{K}_*(\beta_{y,x}) \circ t^1_x
= \beta_{y,x}^1 \circ t^0_x - t^0_y \circ \alpha_{y,x}^1
\end{equation}
for all edges \(x\to y\).
Here
\[
\alpha_{y,x}^1 \in \Ext\bigl(\textup{K}_{1+*}(A(U_x)),\textup{K}_*(A(U_y))\bigr),\qquad
\beta_{y,x}^1\in \Ext\bigl(\textup{K}_{1+*}(B(U_x)),\textup{K}_*(B(U_y))\bigr)
\]
are the parity-reversing parts of the \(\textup{KK}\)\nb-classes
of the inclusion maps. The elements \(t^0_x\)
and~\(t^1_x\)
together form \(t_x\in \textup{KK}_0(A(U_x),B(U_x))\).
And~\eqref{eq:t_condition_X} is equivalent to
\(t_y\circ \alpha_{y,x} = \beta_{y,x} \circ t_x\)
in \(\textup{KK}_0(A(U_x),B(U_y))\).
Corollary~\ref{cor:classify_KKX} says that any family of
\(\textup{KK}\)-equivalences
\(t_x\in \textup{KK}_0(A(U_x),B(U_x))\)
with \(t_y \circ \alpha_{y,x} = \beta_{y,x} \circ t_x\)
for all edges \(x\to y\)
lifts to an invertible element in \(\textup{KK}^X_0(A,B)\).
The classification using the obstruction class says that a family of
isomorphisms \(t^0_x\colon \textup{K}_*(A(U_x)) \xrightarrow\sim \textup{K}_*(B(U_x))\)
lifts to an invertible element in \(\textup{KK}^X_0(A,B)\)
if and only if there exist~\(t^1_x\)
satisfying~\eqref{eq:t_condition_X}. Corollary~\ref{cor:classify_KKX}
makes a slightly stronger assertion because it says that any choice of
elements~\(t^1_x\)
satisfying~\eqref{eq:t_condition_X} may be realised by an invertible
element in \(\textup{KK}^X_0(A,B)\).
The result in Corollary~\ref{cor:classify_KKX} about the existence of
liftings is also slightly stronger than the corresponding result using
the invariant~\(F^X_\textup{K}\)
and the obstruction class because it does not require the
\(\textup C^*\)\nb-algebras~\(A(U_x)\) to belong to the bootstrap class.
\section{Filtrated K-theory for totally ordered spaces}
\label{sec:filtrated_K}
Now we consider the special unique path space
\[
X = \{1 \leftarrow 2 \leftarrow \dotsb \leftarrow n\}
\]
for some \(n\in\mathbb N_{\ge2}\).
We are going to compare the classification for \(\textup C^*\)\nb-algebras
over~\(X\)
that follows from the Universal Coefficient Theorem
in~\cite{Meyer-Nest:Filtrated_K} to the classifications in
Section~\ref{sec:compare_ups}. The invariant used
in~\cite{Meyer-Nest:Filtrated_K} is filtrated \(\textup{K}\)\nb-theory.
This is the diagram of \(\textup{K}\)\nb-theory
groups formed by \(\textup{K}_*(A(S))\)
for all locally closed subsets \(S\subseteq X\).
Here a subset is locally closed if and only if it is of the form
\[
[a,b] \mathrel{\vcentcolon=} \{a,a+1,\dotsc,b-1,b\},\qquad
1\le a \le b \le n.
\]
The maps in the filtrated \(\textup{K}\)\nb-theory
diagram are those that come from natural transformations. We are
going to describe these below.
Let~\(\Ideal^X_\mathrm{fil}\)
be the kernel on morphisms of the filtrated \(\textup{K}\)\nb-theory
functor, that is, \(f\in\textup{KK}^X_0(A,B)\)
belongs to \(\Ideal^X_\mathrm{fil}(A,B)\)
if and only if it induces the zero map
\(\textup{K}_*(A[a,b]) \to \textup{K}_*(B[a,b])\)
for all \(1 \le a \le b \le n\).
This is a stable homological ideal. It is shown
in~\cite{Meyer-Nest:Homology_in_KK} that any object \(A\mathrel{\in\in}\Boot^X\)
has an \(\Ideal^X_\mathrm{fil}\)\nb-projective resolution
\begin{equation}
\label{eq:fil_K_resolution}
0 \to P_1 \xrightarrow{\varphi} P_0 \xrightarrow{f} A
\end{equation}
of length~\(1\).
Then it follows that~\(A\)
is isomorphic to the cone of~\(\varphi\).
And there is a Universal Coefficient Theorem for objects
of~\(\Boot^X\)
based on Hom and Ext groups between their filtrated \(\textup{K}\)\nb-theory
diagrams. Thus any isomorphism between the filtrated \(\textup{K}\)\nb-theory
diagrams of \(A,B\mathrel{\in\in}\Boot^X\)
lifts to an equivalence in \(\textup{KK}^X_0(A,B)\).
The minimal open subset~\(U_x\)
containing~\(x\)
is \([x,n]\)
for each \(x\in X\).
Since \(\textup{K}_*(A(U_x))\)
is part of the filtrated \(\textup{K}\)\nb-theory
of~\(A\),
the ideal~\(\Ideal^X_\mathrm{fil}\)
is contained in the ideal~\(\Ideal^X_\textup{K}\)
that is used above for a general unique path space. So any
\(\Ideal^X_\mathrm{fil}\)\nb-exact
chain complex is also \(\Ideal^X_\textup{K}\)\nb-exact.
If \(P\mathrel{\in\in}\Boot^X\)
is \(\Ideal^X_\mathrm{fil}\)\nb-projective,
then the Abelian groups \(\textup{K}_*(P(S))\)
are free for all locally closed subsets \(S\subseteq X\)
(this follows from \cite{Meyer-Nest:Filtrated_K}*{Theorem~3.12} and
will become manifest below). The computation of \(\Ext^2_{\Abel^X}\)
in Lemma~\ref{lem:Ext_2_over_X} shows that
\(\Ext^2_{\Abel^X}(F^X_\textup{K}(P),D)=0\)
for all \(D\mathrel{\in\in}\Abel^X\)
if \(P(U_x)\)
is free for each \(x\in X\).
Hence \(P_1\)
and~\(P_0\)
have \(\Ideal^X_\textup{K}\)\nb-projective
resolutions of length~\(1\).
So we are in the situation of~\eqref{eq:triangle_B1B0A} and
Theorem~\ref{the:compute_obstruction} computes the obstruction class
from the parity-reversing part of the map~\(\varphi\)
in~\eqref{eq:fil_K_resolution}. This computation is, however, quite
non-trivial. We must first recall how the natural transformations in
the filtrated \(\textup{K}\)\nb-theory
diagram look like. Going beyond the results
of~\cite{Meyer-Nest:Filtrated_K}, we then build an explicit
\(\Ideal^X_\mathrm{fil}\)\nb-projective
resolution. Next, we observe which parts of the map~\(\varphi\)
are parity-reversing. This gives a class in \(\Ext^2_{\Abel^X}\).
To translate it into the setting of
Theorem~\ref{the:obstruction_class_X}, we still have to compare the
resolution used there with the one coming
from~\eqref{eq:fil_K_resolution}. This requires most of the work.
We first recall the description of the \(\mathbb Z/2\)\nb-graded
Abelian groups \(\mathcal{NT}_*([a,b],[c,d])\)
of natural transformations \(\textup{K}_*(A([a,b])) \to \textup{K}_*(A([c,d]))\)
in~\cite{Meyer-Nest:Filtrated_K}.
Let \(1\le a \le b \le n\).
The functor \(\KKcat^X\to\Ab^{\mathbb Z/2}\),
\(A\mapsto \textup{K}_*(A([a,b]))\),
is represented by an object~\(\mathcal{R}_{[a,b]}\)
of~\(\KKcat^X\),
which is described in~\cite{Meyer-Nest:Filtrated_K}. Let
\(1\le a\le b \le n\)
and \(1\le c\le d \le n\).
By the Yoneda Lemma,
\[
\mathcal{NT}_*([a,b],[c,d])
\cong \textup{KK}^X_*(\mathcal{R}_{[c,d]},\mathcal{R}_{[a,b]})
\cong \textup{K}_*(\mathcal{R}_{[a,b]}([c,d])).
\]
These groups are computed in
\cite{Meyer-Nest:Filtrated_K}*{Equation~(3.1)}:
\begin{equation}
\label{eq:NT_explicit}
\mathcal{NT}_*([a,b],[c,d]) \cong
\begin{cases}
\mathbb Z_+&\text{if }c\le a \le d \le b,\\
\mathbb Z_-&\text{if }a+1 \le c \le b+1 \le d,\\
0&\text{otherwise.}
\end{cases}
\end{equation}
We write \([a,b] \to [c,d]\)
in the first two cases, that is, when there is a non-zero natural
transformation in \(\mathcal{NT}_*([a,b],[c,d])\).
The groups \(\mathcal{NT}_*([a,b],[c,d])\)
form a \(\mathbb Z/2\)\nb-graded
ring~\(\mathcal{NT}\).
The filtrated \(\textup{K}\)\nb-theory
of a separable \(\textup C^*\)\nb-algebra
over~\(X\)
is a \(\mathbb Z/2\)\nb-graded,
countable module over it, which we denote by \(FK(A)\).
The computations in~\cite{Meyer-Nest:Filtrated_K} interpret the
elements of \(\mathcal{NT}_*([a,b],[c,d])\)
as follows. Let~\(A\)
be a separable \(\textup C^*\)\nb-algebra
over~\(X\)
and let \(M \mathrel{\vcentcolon=} FK(A)\).
If \(a< b \le c\),
then \([b,c]\)
is relatively open in \([a,c]\)
with complement \([a,b-1]\).
This induces the following natural six-term exact sequence in
\(\textup{K}\)\nb-theory:
\begin{equation}
\label{eq:exact_module_NT}
\dotsb \to M[b,c] \xrightarrow{i} M[a,c] \xrightarrow{r} M[a,b-1]
\xrightarrow[\mathrm{odd}]{\delta} M[b,c] \to \dotsb
\end{equation}
where the maps \(i,r\)
preserve the \(\mathbb Z/2\)\nb-grading
and~\(\delta\)
reverses it. Any natural transformation \(M[a,b] \to M[c,d]\)
is an integer multiple of a product of the maps \(i,r,\delta\)
above. More precisely, if \(c\le a \le d \le b\),
then there is a commuting square
\begin{equation}
\label{def:tautr_even}
\begin{tikzcd}
M[a,b] \arrow[r,"i"] \arrow[d, "r"] \arrow[dr, "\tautr{a,b}{c,d}" description] &
M[c,b] \arrow[d, "r"] \\
M[a,d] \arrow[r, "i"'] & M[c,d],
\end{tikzcd}
\end{equation}
and its diagonal map~\(\tautr{a,b}{c,d}\)
generates \(\mathcal{NT}_0([a,b],[c,d]) \cong\mathbb Z\).
And if \(a+1 \le c \le b+1 \le d\), then there is a commuting square
\begin{equation}
\label{def:tautr_odd}
\begin{tikzcd}
M[a,b] \arrow[r,"\delta"] \arrow[d, "r"] \arrow[dr, "\tautr{a,b}{c,d}" description] &
M[b+1,d] \arrow[d, "i"] \\
M[a,c-1] \arrow[r, "\delta"'] & M[c,d],
\end{tikzcd}
\end{equation}
and its diagonal map~\(\tautr{a,b}{c,d}\)
generates \(\mathcal{NT}_1([a,b]),[c,d])\cong\mathbb Z\).
We have defined a generator~\(\tau_{[a,b]}^{[c,d]}\)
for \(\mathcal{NT}_*([a,b],[c,d])\)
whenever \([a,b] \to [c,d]\),
that is, whenever \(\mathcal{NT}_*([a,b]),[c,d])\neq0\)
by~\eqref{eq:NT_explicit}. It is convenient to define
\(\tau_{[a,b]}^{[c,d]}=0\)
if \(\mathcal{NT}_*([a,b]),[c,d])=0\).
By the Yoneda Lemma, the natural transformations~\(\tautr{a,b}{c,d}\)
correspond to arrows
\[
\bigl(\tautr{a,b}{c,d}\bigr)^*\colon
\mathcal{R}_{[c,d]} \to \mathcal{R}_{[a,b]}.
\]
\begin{remark}
\label{rem:exact_NT}
An \(\mathcal{NT}\)\nb-module
is called \emph{exact} if the sequences~\eqref{eq:exact_module_NT}
are exact for all \(a< b \le c\).
The exact \(\mathcal{NT}\)\nb-modules
form a stable exact subcategory of the stable Abelian category of
all \(\mathcal{NT}\)\nb-modules,
and the filtrated \(\textup{K}\)\nb-theory
of any separable \(\textup C^*\)\nb-algebra
over~\(X\)
is exact as an \(\mathcal{NT}\)\nb-module.
The results in~\cite{Meyer-Nest:Filtrated_K} imply that any exact
\(\mathcal{NT}\)\nb-module
has a projective resolution of length~\(1\).
Hence it lifts to an object of the bootstrap class~\(\Boot^X\).
So the image of the filtrated \(\textup{K}\)\nb-theory
functor is equal to the class of exact, countable
\(\mathcal{NT}\)\nb-modules.
And the filtrated \(\textup{K}\)\nb-theory
functor, viewed as a functor to the subcategory of exact, countable
\(\mathcal{NT}\)\nb-modules,
is the universal \(\Ideal^X_\mathrm{fil}\)\nb-exact
functor to an exact category. So~\(\Ideal^X_\mathrm{fil}\)
has the property that its universal exact functors to an Abelian and
to an exact category are different.
\end{remark}
Now we study the multiplication in~\(\mathcal{NT}\).
We begin with decomposing the generators in~\(\mathcal{NT}\)
further. We may rewrite the natural
transformations~\(\tautr{a,b}{c,d}\)
defined above as products of the special natural transformations
\begin{alignat*}{2}
i = \tautr{a+1,b}{a,b}&\colon [a+1,b] \to [a,b],&\qquad&a+1\le b \le n,\\
r = \tautr{a,b+1}{a,b}&\colon [a,b+1] \to [a,b],&\qquad&a\le b \le n-1,\\
\delta = \tautr{1,a-1}{a,n}&\colon [1,a-1] \to [a,n],&\qquad&2 \le a \le n.
\end{alignat*}
This is clear in the even case. In the odd case, we use the
naturality of boundary maps to rewrite the boundary map
\(\delta\colon [a,b] \to [b+1,d]\)
as the product of \(i\colon [a,b] \to [1,b]\),
the boundary map \(\delta\colon [1,b] \to [b+1,n]\),
and \(r\colon [b+1,n] \to [b+1,d]\).
The generating natural transformations defined above form a commuting
diagram as in Figure~\ref{fig:filtrated_K}.%
\begin{figure}
\caption{Natural transformations on filtrated \(\textup{K}\)\nb-theory for general~\(n\)}
\label{fig:filtrated_K}
\rotatebox{90}{\begin{tikzcd}[column sep=small,ampersand replacement=\&]
M[n,n] \arrow[r, "i"] \arrow[d] \&
M[n-1,n] \arrow[r, "i"] \arrow[d, "r"] \&
M[n-2,n] \arrow[r, "i"] \arrow[d, "r"] \&
\dotsb \arrow[r, "i"] \arrow[d, "r"] \&
M[2,n] \arrow[r, "i"] \arrow[d, "r"] \&
M[1,n] \arrow[r] \arrow[d, "r"] \&
0 \arrow[d] \\
0 \arrow[r] \&
M[n-1,n-1] \arrow[r, "i"] \arrow[d] \&
M[n-2,n-1] \arrow[r, "i"] \arrow[d, "r"] \&
\dotsb \arrow[r, "i"] \arrow[d, "r"] \&
M[2,n-1] \arrow[r, "i"] \arrow[d, "r"] \&
M[1,n-1] \arrow[r, "\delta"] \arrow[d, "r"] \&
M[n,n] \arrow[d,"i"] \\
\& 0 \arrow[r] \&
M[n-2,n-2] \arrow[r, "i"] \arrow[d] \&
\dotsb \arrow[r, "i"] \arrow[d, "r"] \&
M[2,n-2] \arrow[r, "i"] \arrow[d, "r"] \&
M[1,n-2] \arrow[r, "\delta"] \arrow[d, "r"] \&
M[n-1,n] \arrow[d,"i"] \\
\&\&0\arrow[r]\&\vdots\arrow[r]\&
\vdots\arrow[r] \arrow[d,"r"]\&
\vdots\arrow[r] \arrow[d,"r"]\&
\vdots \arrow[d,"i"]\\
\&\&\&0\arrow[r]\&
M[2,2] \arrow[r, "i"] \arrow[d] \&
M[1,2] \arrow[r, "\delta"] \arrow[d, "r"] \&
M[3,n] \arrow[d,"i"] \\
\&\&\&\&0\arrow[r]\&
M[1,1] \arrow[r, "\delta"] \arrow[d] \&
M[2,n] \arrow[d,"i"] \\
\&\&\&\&\&0\arrow[r]\&
M[1,n]
\end{tikzcd}}
\end{figure}%
\begin{figure}
\caption{Complete diagram of natural transformations on filtrated \(\textup{K}\)\nb-theory for \(n=3\)}
\label{fig:filtrated_K_3}
\rotatebox{90}{\begin{tikzcd}[column sep=small,ampersand replacement=\&]
M[3,3] \arrow[r, "i"] \arrow[d] \&
M[2,3] \arrow[r, "i"] \arrow[d, "r"] \&
M[1,3] \arrow[r] \arrow[d, "r"] \&
0 \arrow[d] \\
0 \arrow[r] \&
M[2,2] \arrow[r, "i"] \arrow[d] \&
M[1,2] \arrow[r, "\delta"] \arrow[d, "r"] \&
M[3,3] \arrow[d,"i"] \arrow[r] \&
0 \arrow[d] \\
\& 0 \arrow[r] \&
M[1,1] \arrow[r, "\delta"] \arrow[d] \&
M[2,3] \arrow[d,"i"] \arrow[r, "r"] \&
M[2,2] \arrow[d,"i"] \arrow[r] \& 0\arrow[d]\\
\&\&0\arrow[r]\&
M[1,3] \arrow[r, "r"] \&
M[1,2] \arrow[r, "r"] \&
M[1,1] \\
\end{tikzcd}}
\end{figure}
The last column and the first row in the diagram are the same. So the
diagram repeats when we put a reflected copy of it next to it.
Figure~\ref{fig:filtrated_K_3} shows the full diagram for \(n=3\).
We claim that all relations among the generating natural
transformations are given by this extended commuting diagram. In
particular, a composite of \(i,r,\delta\)
vanishes if and only if it factors through one of the objects~\(0\)
on the boundary of the extended diagram.
Let \(1 \le a \le b \le n\)
and \(1 \le c \le d \le n\).
A product of the generators of type \(i,r\)
from~\([a,b]\)
to~\([c,d]\)
exists if and only if \(c \le a\)
and \(d \le b\).
Figure~\ref{fig:filtrated_K_3} shows that all such products are equal.
Equation~\eqref{eq:NT_explicit} shows that this product is~\(0\)
unless \(a \le d\),
so that \(c \le a \le d \le b\).
Then it is equal to \(\tautr{a,b}{c,d}\)
by the definition in~\eqref{def:tautr_even}. As a consequence,
\(\tautr{e,g}{c,d} \cdot \tautr{a,b}{e,g} = \tautr{a,b}{c,d}\)
if \(c \le e \le d \le g\)
and \(e \le a \le g \le b\);
this is non-zero if and only if also \(a \le d\)
or, equivalently, \(c \le e \le a \le d \le g \le b\).
Now consider a product of \(i,r,\delta\)
going from~\([a,b]\)
to~\([c,d]\)
and containing exactly one factor of~\(\delta\).
Using the diagram in Figure~\ref{fig:filtrated_K}, we may rearrange
this product in such a way that we first go right and then go down in
the extended diagram as in Figure~\ref{fig:filtrated_K_3}. If this
goes through the zeros outside the drawn region, the product is~\(0\).
If not, we may combine consecutive~\(i\)
and consecutive~\(r\) to bring the product into the following form:
\[
[a,b] \xrightarrow{i}
[1,b] \xrightarrow{\delta}
[b+1,n] \xrightarrow{r}
[b+1,d] \xrightarrow{i}
[c,d].
\]
The combination \(r\circ\delta\circ i\)
in the beginning is the boundary map
\(\delta\colon [a,b] \to [b+1,d]\).
So we get~\(\tautr{a,b}{c,d}\)
if \(c+1 \le a \le d+1 \le b\)
and~\(0\)
otherwise by \eqref{def:tautr_odd} and~\eqref{eq:NT_explicit}. Since
we may rewrite all even~\(\tautr{a,b}{c,d}\)
in terms of \(i,r\),
we can now compute \(\tautr{e,g}{c,d}\cdot \tautr{a,b}{e,g}\)
if one of the transformations \(\tautr{e,g}{c,d}\)
and \(\tautr{a,b}{e,g}\)
is even and the other one is odd. Namely, the product is
\(\tautr{a,b}{c,d}\)
if \(a+1 \le c \le b+1 \le d\),
and~\(0\)
otherwise. In more detail, the assumption that exactly one of the
transformations \(\tautr{e,g}{c,d}\)
and \(\tautr{a,b}{e,g}\)
is even means that \(e+1 \le c \le g+1 \le d\)
and \(e \le a \le g \le b\),
or \(c \le e \le d \le g\)
and \(a+1 \le e \le b+1 \le g\).
The assumption \(a+1 \le c \le b+1 \le d\)
becomes \(e \le a < c \le g+1 \le b+1 \le d\)
or \(a< c \le e \le b+1 \le d \le g\)
in these two cases, respectively.
Finally, any product with more than two factors~\(\delta\)
vanishes because it may be deformed in the extended diagram in
Figure~\ref{fig:filtrated_K_3} so as to factor through one of the
zeros on the boundary. We sum up our results about the multiplication
in~\(\mathcal{NT}\):
\begin{equation}
\label{eq:compose_tau}
\tautr{e,g}{c,d} \cdot \tautr{a,b}{e,g} = \tautr{a,b}{c,d} \neq0
\iff
\left\{
\begin{array}{c}
c \le e \le a \le d \le g \le b,\\
e \le a \le c-1 \le g \le b < d,\\
a< c \le e \le b+1 \le d \le g,
\end{array}
\right.
\end{equation}
and \(\tautr{e,g}{c,d} \cdot \tautr{a,b}{e,g}=0\)
otherwise. We write
\[
[a,b] \to [e,g] \to [c,d] \iff
\tautr{e,g}{c,d} \cdot \tautr{a,b}{e,g} = \tautr{a,b}{c,d} \neq0.
\]
It can happen that \([a,b] \to [e,g]\),
\([e,g] \to [c,d]\)
and \([a,b] \to [c,d]\),
but not \([a,b] \to [e,g] \to [c,d]\);
that is, \(\tautr{e,g}{c,d} \cdot \tautr{a,b}{e,g} =0\)
although \(\tautr{a,b}{c,d},\tautr{a,b}{e,g}, \tautr{e,g}{c,d}\neq0\).
Given \(1\le a \le b \le n\),
there are one or two proper natural transformations to \(M[a,b]\)
that are shortest in the sense that all others factor through them.
If \(a<b<n\),
these are \(i\colon M[a+1,b] \to M[a,b]\)
and \(r\colon M[a,b+1]\to M[a,b]\).
If \(1< a<b=n\),
then~\(r\)
above is replaced by \(\delta\colon M[1,a-1] \to M[a,n]\).
One of these maps is missing if \(a=b\)
or if \((a,b) = (1,n)\),
that is, on the two outer diagonals in the diagram in
Figure~\ref{fig:filtrated_K_3} (now for general~\(n\)).
We define \(M[a+1,a]\defeq0\)
for \(0\le a \le n\) to make this a special case of the generic case.
Now we build an \(\Ideal^X_\mathrm{fil}\)\nb-projective
resolution of~\(A\)
of length~\(1\).
This has not yet been done in~\cite{Meyer-Nest:Filtrated_K}, where
only the existence of such a resolution is proven.
For \(1\le a \le b \le n\),
let \(M[a,b]_\mathrm{ss}\)
be the quotient of~\(M[a,b]\)
by the images of all proper natural transformations to \(M[a,b]\)
or, equivalently, by the images of the two shortest natural
transformations:
\[
M[a,b]_\mathrm{ss} \mathrel{\vcentcolon=}
\begin{cases}
M[a,b]\bigm/ (i(M[a+1,b]) + r(M[a,b+1]))&\text{if }b<n,\\
M[a,b]\bigm/ (i(M[a+1,b]) + \delta(M[1,a-1]))&\text{if }b=n
\end{cases}
\]
(compare \cite{Meyer-Nest:Filtrated_K}*{Definition~3.7 and
Lemma~3.8}). Choose a resolution
\begin{equation}
\label{eq:Mab_ss_resolution}
\begin{tikzcd}
Q_1[a,b] \arrow[r, rightarrowtail, "d_1"] &
Q_0[a,b] \arrow[r, twoheadrightarrow, "d_0"] &
M[a,b]_\mathrm{ss}
\end{tikzcd}
\end{equation}
of~\(M[a,b]_\mathrm{ss}\)
by countable \(\mathbb Z/2\)\nb-graded
free Abelian groups. For \(i=0,1\), let
\[
\hat{Q}_i[a,b] \mathrel{\vcentcolon=} \mathcal{R}_{[a,b]} \otimes_\mathbb Z Q_i[a,b],
\]
where the tensor product is defined as
in~\eqref{def:tensor_with_Cstar}.
Since~\(\mathcal{R}_{[a,b]} \otimes_\mathbb Z Q_i[a,b]\)
is a direct sum of copies of suspensions of~\(\mathcal{R}_{[a,b]}\),
the definition of~\(\mathcal{R}_{[a,b]}\)
as a representing object implies
\begin{equation}
\label{eq:KK_out_of_RY}
\textup{KK}^X_0(\hat{Q}_i[a,b], B) \cong \Hom\bigl(Q_i[a,b], \textup{K}_*(B[a,b])\bigr)
\end{equation}
for all \(\textup C^*\)\nb-algebras~\(B\)
over~\(X\);
here~\(\Hom\)
means grading-preserving group homomorphisms.
This property characterises~\(\hat{Q}_i[a,b]\)
uniquely up to isomorphism in~\(\KKcat^X\).
Equation~\eqref{eq:KK_out_of_RY} implies a similar description of
the \(\mathbb Z/2\)\nb-graded Abelian group
\(\textup{KK}^X_*(\hat{Q}_i[a,b], B)\),
replacing \(\Hom\)
by group homomorphisms that need not respect the grading.
Given a group homomorphism \(g\colon Q_i[a,b] \to \textup{K}_*(B[a,b])\),
let~\(g^\#\)
denote the corresponding element of
\(\textup{KK}^X_*(\hat{Q}_i[a,b], B)\).
Since the
\(\mathbb Z/2\)\nb-graded
Abelian group \(Q_0[a,b]\)
is free, the homomorphism~\(d_0\)
in~\eqref{eq:Mab_ss_resolution} lifts to a grading-preserving
homomorphism
\begin{equation}
\label{eq:Q0_to_M}
f[a,b]\colon Q_0[a,b] \to M[a,b] = \textup{K}_*(A[a,b]).
\end{equation}
Let \(f[a,b]^\# \in \textup{KK}^X_0(\hat{Q}_0[a,b], A)\)
correspond to~\(f[a,b]\) by~\eqref{eq:KK_out_of_RY}. Let
\[
P_i \mathrel{\vcentcolon=} \bigoplus_{1\le a\le b\le n} \hat{Q}_i[a,b]
\]
for \(i=0,1\). The objects \(\hat{Q}_i[a,b]\)
and~\(P_i\)
for \(i=0,1\)
are \(\Ideal^X_\mathrm{fil}\)-projective
because of~\eqref{eq:KK_out_of_RY}. There is
a unique element \(f \in \textup{KK}^X_0(P_0,A)\)
that restricts to~\(f[a,b]^\#\)
on the summand \(\hat{Q}_i[a,b]\).
\begin{lemma}
\label{lem:surjective_filtrated_K}
The map \(FK(f)\colon FK(P_0) \to M \mathrel{\vcentcolon=} FK(A)\)
is surjective. Its kernel is isomorphic to~\(FK(P_1)\)
as an \(\mathcal{NT}\)-module.
\end{lemma}
\begin{proof}
Let \(f_*\mathrel{\vcentcolon=} FK(f)\).
The \(\mathcal{NT}\)\nb-module \(M_\mathrm{ss}\)
is defined as the quotient \(M/\mathcal{NT}_\mathrm{nil}\cdot M\)
for a certain ideal~\(\mathcal{NT}_\mathrm{nil}\)
in~\(\mathcal{NT}\).
It follows that the functor \(M\mapsto M_\mathrm{ss}\) is right exact.
Even more, it is isomorphic to the tensor product functor with the
right \(\mathcal{NT}\)\nb-module
\(\mathcal{NT}_\mathrm{ss}\).
This follows from the extension of
\(\mathcal{NT}\)\nb-modules
\(\mathcal{NT}_\mathrm{nil}\rightarrowtail\mathcal{NT}
\twoheadrightarrow\mathcal{NT}_\mathrm{ss}\)
and the right exactness of tensor product functors.
Right exactness implies
\((\coker f_*)[a,b]_\mathrm{ss} = \coker (f_*[a,b]_\mathrm{ss})\)
for all \(1 \le a \le b \le n\).
The projective \(\mathcal{NT}\)\nb-module~\(FK(P_0)\)
has \(FK(P_0)[a,b]_\mathrm{ss} \cong Q_0[a,b]\).
Hence \((f_*)_\mathrm{ss}\)
is surjective by construction of~\(f\),
and the \(\mathcal{NT}\)\nb-module
\(\coker f_*\)
satisfies \((\coker f_*)_\mathrm{ss} = 0\).
This implies \(\coker f_*=0\)
by \cite{Meyer-Nest:Filtrated_K}*{Proposition~3.10}. That is,
\(f_*\)
is surjective. Let \(N\mathrel{\vcentcolon=} \ker FK(f)\).
So there is an extension \(N \rightarrowtail FK(P_0) \twoheadrightarrow M\)
of \(\mathcal{NT}\)\nb-modules.
Since \(M\)
and \(FK(P_0)\)
are exact \(\mathcal{NT}\)\nb-modules,
they satisfy
\(\Tor^1_\mathcal{NT}(\mathcal{NT}_\mathrm{ss}, M)=0\)
and \(\Tor^1_\mathcal{NT}(\mathcal{NT}_\mathrm{ss}, FK(P_0))=0\)
by \cite{Meyer-Nest:Filtrated_K}*{Lemma~3.13}. Hence
\[
N[a,b]_\mathrm{ss}
\to FK(P_0)[a,b]_\mathrm{ss}
\to M[a,b]_\mathrm{ss}
\]
is a short exact sequence for all \(1 \le a \le b \le n\)
and \(\Tor^1_\mathcal{NT}(\mathcal{NT}_\mathrm{ss}, N)=0\).
Since \(FK(P_0)[a,b]_\mathrm{ss} \cong Q_0[a,b]\),
this implies \(N[a,b]_\mathrm{ss} \cong Q_1[a,b]\).
Now \cite{Meyer-Nest:Filtrated_K}*{Theorem~3.12} shows that~\(N\)
is a projective \(\mathcal{NT}\)-module.
In fact, the proof of this theorem shows that \(N \cong FK(P_1)\).
More precisely, the quotient maps
\(N[a,b] \twoheadrightarrow N[a,b]_\mathrm{ss}\)
split because \(N[a,b]_\mathrm{ss} \cong Q_1[a,b]\)
is free. Let
\begin{equation}
\label{eq:section_hab}
\varphi[a,b]\colon Q_1[a,b] \cong N[a,b]_\mathrm{ss}
\to N[a,b] \subseteq FK(P_0)[a,b]
\end{equation}
be sections. They induce an \(\mathcal{NT}\)\nb-module
homomorphism \(FK(P_1) \to N\)
by the universal property of the ``free''
\(\mathcal{NT}\)\nb-module~\(FK(P_1)\).
And the proof of \cite{Meyer-Nest:Filtrated_K}*{Theorem~3.12} shows
that it is an isomorphism.
\end{proof}
Disregarding the \(\mathbb Z/2\)\nb-grading, we may write
\[
FK(P_0)[a,b] = \bigoplus_{[c,d]\to[a,b]} Q_0[c,d],
\]
that is, the sum runs over all \(1\le c \le d \le n\)
with \(a\le c \le b \le d\)
or \(c+1\le a \le d+1 \le b\).
So the map~\(\varphi[a,b]\) in~\eqref{eq:section_hab} has components
\[
\varphi_{[a,b]}^{[c,d]}\colon Q_1[a,b] \to Q_0[c,d]
\]
for \([c,d] \to [a,b]\).
Since~\(\varphi[a,b]\)
is even, the map \(\varphi_{\7{a,b}}^{\7{c,d}}\)
has the same parity as \(\tautr{c,d}{a,b}\),
that is, it is grading-preserving if \(a \le c \le b \le d\)
and grading-reversing if \(c+1 \le a \le d+1 \le b\).
The image of \(\varphi[a,b]\)
for \(1\le a \le b \le n\)
is contained in \(N = \ker FK(f)\),
that is, \(FK(f) \circ \varphi[a,b]=0\)
as a map \(Q_1[a,b] \to \textup{K}_*(A[a,b])\).
Unravelling the definition of \(FK(f)\), this becomes
\begin{equation}
\label{eq:tau_f_phi_relation}
\sum_{[c,d]\to[a,b]}
\tau_{[c,d]}^{[a,b]} \circ f[c,d] \circ \varphi_{[a,b]}^{[c,d]}
= 0 \colon Q_1[a,b] \to \textup{K}_*(A[a,b]).
\end{equation}
The objects \(P_0\) and~\(P_1\)
are \(\Ideal^X_\mathrm{fil}\)-projective,
and \(FK\)
is fully faithful on \(\Ideal^X_\mathrm{fil}\)-projective objects.
So the arrow \(FK(P_1) \to FK(P_0)\)
with components \(\varphi_{[a,b]}^{[c,d]}\)
for \([c,d]\to[a,b]\)
lifts uniquely to an arrow \(\varphi\in \textup{KK}^X_0(P_1,P_0)\).
More precisely, the map~\(\varphi\)
is given by a matrix of maps
\(\mathcal{R}_{[c,d]} \otimes Q_1[c,d] \to
\mathcal{R}_{[a,b]} \otimes Q_0[a,b]\)
for all \(1\le c\le d\le n\)
and \(1\le a\le b\le n\).
The entries of this matrix are
\(\Bigl(\bigl(\tautr{a,b}{c,d}\bigr)^* \otimes
\varphi_{\7{c,d}}^{\7{a,b}}\Bigr)\), that is,
\begin{multline}
\label{eq:varphi_components}
\varphi = \Bigl(\bigl(\tautr{a,b}{c,d}\bigr)^* \otimes
\varphi_{\7{c,d}}^{\7{a,b}}\Bigr)_{[c,d]\to[a,b]}\colon\\
\bigoplus_{1\le c\le d\le n} \mathcal{R}_{[c,d]} \otimes Q_1[c,d] \to
\bigoplus_{1\le a\le b\le n} \mathcal{R}_{[a,b]} \otimes Q_0[a,b].
\end{multline}
Here we use the convention that \(\tautr{a,b}{c,d}=0\)
if not \([a,b]\to [c,d]\).
Lemma~\ref{lem:surjective_filtrated_K} says
that~\eqref{eq:fil_K_resolution} with \(f\)
and~\(\varphi\)
as above is \(\Ideal^X_\mathrm{fil}\)-exact,
that is, \(FK\)
applied to~\eqref{eq:fil_K_resolution} is an exact sequence. The
\(\Ideal^X_\mathrm{fil}\)-exactness
of~\eqref{eq:fil_K_resolution} says that the functor
\(B\mapsto \textup{K}_*(B[a,b])\)
maps it to an exact sequence for each \(1 \le a \le b \le n\).
In fact, this gives projective resolutions. We write them down
explicitly:
\begin{lemma}
\label{lem:resolution_KBab}
Let \(1 \le a \le b \le n\). Then
\begin{equation}
\label{eq:varphi_f_exact_sequence_ab}
\bigoplus_{[c,d] \to [a,b]} Q_1[c,d]
\xrightarrow{\bigl(\varphi_{\7{c,d}}^{\7{e,g}}\bigr)_{\7{c,d}\to \7{e,g}\to \7{a,b}}}
\bigoplus_{[e,g] \to [a,b]} Q_0[e,g]
\xrightarrow{\bigl(\tau_{\7{e,g}}^{\7{a,b}}\circ f\7{e,g}\bigr)}
\textup{K}_*(A[a,b])
\end{equation}
is a free resolution. Here
\(\bigl(\varphi_{\7{c,d}}^{\7{e,g}}\bigr)_{\7{c,d}\to \7{e,g}\to
\7{a,b}}\)
means that the matrix entry is \(\varphi_{\7{c,d}}^{\7{e,g}}\)
if \(\7{c,d}\to \7{e,g}\to \7{a,b}\)
as described in~\eqref{eq:compose_tau}, and~\(0\) otherwise.
\end{lemma}
The boundary maps in~\eqref{eq:varphi_f_exact_sequence_ab} are
inhomogeneous, that is, the matrix entries of the maps may have even
or odd degree.
\begin{proof}
Equation~\eqref{eq:NT_explicit} computes the group
\(\textup{K}_*(\mathcal{R}_{[c,d]}([a,b])) \cong
\mathcal{NT}_*([c,d],[a,b])\):
it is~\(\mathbb Z\)
in even or odd degree if \([c,d] \to [a,b]\)
and~\(0\)
otherwise. Therefore,
\(P_i[a,b] \cong \bigoplus_{[c,d] \to [a,b]} Q_i[c,d]\)
for \(i=0,1\),
disregarding the grading. The map
\(\Bigl(\bigl(\tautr{e,g}{c,d}\bigr)^* \otimes
\varphi_{\7{c,d}}^{\7{e,g}}\Bigr)\)
between the summands \(\mathcal{R}_{[c,d]} \otimes Q_1[c,d]\)
in~\(P_1\)
and \(\mathcal{R}_{[e,g]} \otimes Q_0[e,g]\)
in~\(P_0\)
induces the map
\(\varphi_{\7{c,d}}^{\7{e,g}}\colon Q_1[c,d] \to Q_0[e,g]\)
if \([c,d] \to [e,g] \to [a,b]\),
and~\(0\)
otherwise, compare~\eqref{eq:compose_tau}. The map
\(\textup{K}_*(P_0[a,b]) \to \textup{K}_*(A[a,b])\) corresponds to a family of maps
\[
\textup{K}_*(\mathcal{R}_{[c,d]}[a,b]) \otimes Q_0[c,d] \cong
\textup{K}_*(\mathcal{R}_{[c,d]}[a,b] \otimes Q_0[c,d]) \to \textup{K}_*(A[a,b])
\]
for \(1 \le c \le d \le n\).
As above, \(\textup{K}_*(\mathcal{R}_{[c,d]}[a,b])\neq0\)
only if \([c,d] \to [a,b]\),
and then the map \(Q_0[c,d] \to K_*(A[a,b])\)
induced by \(f\colon P_0 \to A\)
is \(\tautr{c,d}{a,b} \circ f[c,d]\).
\end{proof}
We have reached the first milestone in the computation of the
obstruction class: the \(\Ideal_\mathrm{fil}^X\)-projective
resolution~\eqref{eq:fil_K_resolution}. It is explicit enough to
express the obstruction class that comes from filtrated
\(\textup{K}\)\nb-theory
in the terms of Theorem~\ref{the:obstruction_class_X}, namely, as
being represented by a family of elements in
\(\Ext^1\bigl(\textup{K}_*(A[e+1,n]),\textup{K}_*(A[e,n])\bigr)\)
for \(e=1,\dotsc,n-1\).
We compute these Ext-groups with the resolutions
in~\eqref{eq:varphi_f_exact_sequence_ab}. So the obstruction class
corresponds to a sequence of maps
\[
\delta_e\colon
\bigoplus_{[a,b] \to [e+1,n]} Q_1[a,b] \to \textup{K}_*(A[e,n]),\qquad
e=1,\dotsc,n-1.
\]
In turn, each~\(\delta_e\)
is given by maps \(\delta_e^{[a,b]}\colon Q_1[a,b] \to \textup{K}_*(A[e,n])\)
for all \(1 \le a \le b \le n\)
with \([a,b] \to [e+1,n]\).
The following theorem computes these maps~\(\delta_e^{[a,b]}\).
It is the main result of this section.
Section~\ref{sec:proof_filtrated_theorem} is dedicated to its proof.
\begin{theorem}
\label{the:compare_obstruction_filtered}
Let
\[
\delta_e^{[a,b]} \mathrel{\vcentcolon=}
\begin{cases}
\displaystyle\sum_{[c,d] \to [a,n] \to [a,b]}
\tautr{c,d}{e,n} \circ f[c,d] \circ \varphi_{\7{a,b}}^{\7{c,d}}&
\text{if }a=e \text{ and }b<n,\\
0&\text{otherwise.}
\end{cases}
\]
The resulting family of maps \((\delta_e)_{1 \le e <n}\)
lifts the obstruction class of~\(A\)
to an element of
\(\prod_{e=1}^{n-1} \Ext^1\bigl(\textup{K}_*(A[e+1,n]),\textup{K}_*(A[e,n])\bigr)\).
\end{theorem}
Any \(\Ideal_\mathrm{fil}^X\)-projective
resolution of~\(A\)
is isomorphic to one of the form above because any
\(\Ideal_\mathrm{fil}^X\)-epic
map from an \(\Ideal_\mathrm{fil}^X\)-projective
object to~\(A\)
is isomorphic to a map \(f\in \textup{KK}^X_0(P_0,A)\)
as above. Since \(\Ideal_\mathrm{fil}^X\)-projective
resolutions of~\(A\)
are equivalent to projective resolutions of \(FK(A)\),
the maps~\(\delta_e^{[a,b]}\)
may, in principle, be computed from~\(FK(A)\)
by choosing a projective resolution. This gives the maps \(f[c,d]\)
and \(\varphi_{[a,b]}^{[c,d]}\).
Such a computation may, of course, be difficult in practice.
\subsection{Proof of the obstruction class formula}
\label{sec:proof_filtrated_theorem}
First we examine the smaller invariant
\(F^X_\textup{K}\colon \KKcat^X \to \Abel^X\).
This takes the part of filtrated \(\textup{K}\)\nb-theory
consisting of \(\textup{K}_*(A[a,n])\)
for \(1 \le a \le n\)
with the maps~\(i\)
between them because the minimal open subset containing~\(a\)
is \(U_a = [a,n]\).
So the diagram~\(F^X_\textup{K}(A)\)
is simply the first row in the diagram in
Figure~\ref{fig:filtrated_K}. We have
\(i_a \mathbb C \cong \mathcal{R}_{[a,n]}\)
for \(1\le a \le n\)
because both objects represent the same functor
\(A\mapsto \textup{K}_*(A[a,n])\).
So~\(\mathcal{R}_{[a,n]}\)
for \(1 \le a \le n\)
is \(\Ideal^X_\textup{K}\)\nb-projective,
and \(F^X_\textup{K}(\mathcal{R}_{[a,n]})\) is the diagram
\begin{equation}
\label{eq:projectives_in_AX}
\mathcal{P}_{[a,n]} \mathrel{\vcentcolon=} F^X_\textup{K}(\mathcal{R}_{[a,n]}) = \Bigl(
\underbrace{\mathbb Z = \mathbb Z = \dotsb = \mathbb Z}_{a \text{ times}} \leftarrow
\underbrace{0 = \dotsb = 0}_{n-a \text{ times}}
\Bigr).
\end{equation}
in~\(\Abel^X\).
If \(1 \le a \le b \le n-1\),
then \(F^X_\textup{K}(\mathcal{R}_{[a,b]})\) is the diagram
\begin{equation}
\label{eq:generators_in_AX}
\mathcal{P}_{[a,b]} \mathrel{\vcentcolon=} F^X_\textup{K}(\mathcal{R}_{[a,b]}) = \Bigl(
\underbrace{0 = \dotsb = 0}_{a \text{ times}} \leftarrow
\underbrace{\mathbb Z_- = \dotsb = \mathbb Z_-}_{b+1-a \text{ times}} \leftarrow
\underbrace{0 = \dotsb = 0}_{n-1-b \text{ times}}
\Bigr)
\end{equation}
because of the formula for
\(\textup{K}_*(\mathcal{R}_{[a,b]}([c,n]))\)
in~\eqref{eq:NT_explicit}. Thus the objects~\(\mathcal{R}_{[a,b]}\)
for \(1\le a \le b \le n\) are even if \(b=n\) and odd if \(b<n\).
The formula for the obstruction class in
Theorem~\ref{the:compute_obstruction} uses the parity-reversing part
of \(\varphi\in\textup{KK}^X_0(P_1,P_0)\).
This is described by the following lemma:
\begin{lemma}
\label{lem:fil_K_parity-reversing}
The component
\(\bigl(\tautr{c,d}{a,b}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\)
of~\(\varphi\)
is parity-reversing if and only if \([c,d] \to [a,n]\to[a,b]\)
and \(b<n\).
\end{lemma}
\begin{proof}
If \(\tautr{c,d}{a,b}\neq0\),
then either \(a \le c \le b \le d\)
or \(c+1 \le a \le d+1 \le b\).
The map
\(\bigl(\tautr{c,d}{a,b}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\)
always belongs to \(\textup{KK}^X_0(\hat{Q}_1[a,b],\hat{Q}_0[c,d])\).
So \(\varphi_{\7{a,b}}^{\7{c,d}}\colon Q_1[a,b] \to Q_0[c,d]\)
is parity-preserving in the first case and parity-reversing in the
second case. The object~\(\mathcal{R}_{[a,b]}\)
is even if \(b=n\)
and odd if \(b<n\).
First let \(a \le c \le b \le d\).
If \(b=n\),
then \(d=n\)
as well, so that
\(\mathcal{R}_{[a,b]}\) and~\(\mathcal{R}_{[c,d]}\)
have the same parity, and~\(\varphi_{\7{a,b}}^{\7{c,d}}\)
preserves parity. So we get a parity-preserving component
of~\(\varphi\).
For the same reasons, we get a parity-preserving component if
\(a \le c \le b\le d<n\),
and a parity-reversing component if
\(a \le c \le b < d=n\).
Now let \(c+1 \le a \le d+1 \le b\),
so that~\(\varphi_{\7{a,b}}^{\7{c,d}}\)
reverses parity. If \(b=n\),
then \(d<n\),
so that \(\mathcal{R}_{[a,b]}\)
and~\(\mathcal{R}_{[c,d]}\)
have opposite parity. Hence
\(\bigl(\tautr{c,d}{a,b}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\)
is parity-preserving altogether. If \(b<n\),
then \(d<n\)
and so \(\mathcal{R}_{[a,b]}\)
and~\(\mathcal{R}_{[c,d]}\)
have the same parity. Thus
\(\bigl(\tautr{c,d}{a,b}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\)
is parity-reversing. Inspection shows that the parity-reversing
components are exactly those for which \([c,d] \to [a,n] \to [a,b]\)
as in~\eqref{eq:compose_tau} and \(b<n\).
\end{proof}
Let \(1 \le a \le b < n\). The exact triangle
\begin{equation}
\label{eq:projective_triangle_Rab}
\Sigma \mathcal{R}_{[a,n]} \xrightarrow{i^*}
\Sigma \mathcal{R}_{[b+1,n]} \xrightarrow{\delta^*}
\mathcal{R}_{[a,b]} \xrightarrow{r^*}
\mathcal{R}_{[a,n]}
\end{equation}
in~\(\KKcat^X\)
is \(\Ideal^X_\textup{K}\)\nb-exact,
that is, \(F^X_\textup{K}(r^*)=0\).
Since \(\Sigma \mathcal{R}_{[a,n]}\)
and \(\Sigma \mathcal{R}_{[b+1,n]}\)
are \(\Ideal^X_\textup{K}\)\nb-projective,
it is an \(\Ideal^X_\textup{K}\)\nb-projective
resolution of~\(\mathcal{R}_{[a,b]}\)
of length~\(1\).
Since we allow both odd and even arrows in diagrams, we may drop the
suspensions in~\eqref{eq:projective_triangle_Rab}. For
\(1 \le a \le b \le n\), let
\[
M(a,b) \mathrel{\vcentcolon=}
\begin{cases}
a&\text{if }b=n,\\
b+1&\text{if }b<n.
\end{cases}
\]
Then
\(\bigl(\tau_{[a,b]}^{[M(a,b),n]}\bigr)^*\colon
\mathcal{R}_{[M(a,b),n]} \to \mathcal{R}_{[a,b]}\)
is an \(\Ideal^X_\textup{K}\)\nb-epimorphism
both for \(b=n\)
and \(b<n\).
Its cone is \(\mathcal{R}_{[a,n]}\)
if \(b<n\)
and~\(0\)
if \(b=n\).
Now we write down \(\Ideal^X_\textup{K}\)\nb-projective
resolutions of~\(P_i\) for \(i=0,1\). Let
\begin{align}
P_{i0} &\mathrel{\vcentcolon=}
\bigoplus_{1\le a\le b\le n} \mathcal{R}_{[M(a,b),n]} \otimes Q_i[a,b],\\
P_{i1} &\mathrel{\vcentcolon=}
\bigoplus_{1\le a\le b< n} \mathcal{R}_{[a,n]} \otimes Q_i[a,b].
\end{align}
Then the following is an \(\Ideal^X_\textup{K}\)\nb-projective resolution:
\begin{equation}
\label{eq:resolution_Pi}
0 \to
P_{i1} \xrightarrow{\bigoplus \tau^*\otimes\textup{id}}
P_{i0} \xrightarrow{\bigoplus \tau^*\otimes\textup{id}}
P_i \to 0.
\end{equation}
Here \(\bigoplus \tau^* \otimes \textup{id}\)
means the direct sum of the maps
\(\bigl(\tau_{[z,w]}^{[x,y]}\bigr)^* \otimes \textup{id}_{Q_i[a,b]}\)
between the summands for fixed \(a,b\),
with the appropriate \(x,y,z,w\).
So~\eqref{eq:resolution_Pi} is the direct sum of the
resolutions~\eqref{eq:projective_triangle_Rab}, tensored
with~\(Q_i[a,b]\),
over all \(1\le a\le b<n\),
and the trivial resolutions
\(0 \to \mathcal{R}_{[a,n]} \to \mathcal{R}_{[a,n]}\),
tensored with~\(Q_i[a,b]\),
over all \(1 \le a \le b=n\).
The maps in~\eqref{eq:resolution_Pi} are inhomogeneous, that is, some
components are in \(\textup{KK}^X_0\) and others in \(\textup{KK}^X_1\).
The resolution~\eqref{eq:resolution_Pi} allows us to compute
\(\textup{KK}^X_0(P_1,P_0)\)
with the Universal Coefficient Theorem for the invariant~\(F^X_\textup{K}\).
First, the long exact sequence for the direct sum of the exact
triangles~\eqref{eq:projective_triangle_Rab} implies a natural
extension of Abelian groups
\begin{multline*}
\coker \bigl(\textup{KK}^X_*(\Sigma P_{10},P_0) \to \textup{KK}^X_*(\Sigma P_{11},P_0)\bigr)
\rightarrowtail \textup{KK}^X_*(P_1,P_0)
\\\twoheadrightarrow \ker \bigl(\textup{KK}^X_*(P_{10},P_0) \to \textup{KK}^X_*(P_{11},P_0)\bigr).
\end{multline*}
We may rewrite the kernel and cokernel here as \(\Hom\)
and \(\Ext\)
in~\(\Abel^X\),
using that~\eqref{eq:resolution_Pi} is an
\(\Ideal^X_\textup{K}\)\nb-projective
resolution. The extension above splits unnaturally, giving the
decomposition of~\(\varphi\)
into its parity-preserving and -reversing parts \(\varphi^+\)
and~\(\varphi^-\), respectively.
\begin{lemma}
\label{lem:odd_phi}
The image of the parity-reversing part~\(\varphi^-\)
of~\(\varphi\) in \(\Ext^1_{\Ideal^X_\textup{K}}(P_1,P_0)\) is the map
\[
P_{11}
= \bigoplus_{1 \le a \le b <n} \mathcal{R}_{[a,n]}\otimes Q_1[a,b]
\to \bigoplus_{1 \le c \le d \le n} \mathcal{R}_{[c,d]}\otimes Q_0[c,d]
= P_0
\]
with matrix coefficients
\(\bigl(\tautr{c,d}{a,n}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\)
if \([c,d] \to [a,n] \to [a,b]\) and \(b<n\), and~\(0\) otherwise.
\end{lemma}
\begin{proof}
In terms of the matrix description of~\(\varphi\),
each matrix entry
\(\bigl(\tautr{c,d}{a,b}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\)
has even or odd parity and thus belongs to either \(\Hom\)
or \(\Ext\),
respectively. By Lemma~\ref{lem:fil_K_parity-reversing}, the entry
belongs to~\(\varphi^-\)
if and only if~\(\tautr{c,d}{a,b}\)
factors through
\(r^*\colon \mathcal{R}_{[a,b]} \to \mathcal{R}_{[a,n]}\).
In this case, it factors as
\(\Bigl(\bigl(\tautr{c,d}{a,n}\bigr)^* \otimes
\varphi_{\7{a,b}}^{\7{c,d}}\Bigr) \circ (r^*\otimes \textup{id})\).
Since~\(r^*\)
is the boundary map in~\eqref{eq:projective_triangle_Rab}, this
exhibits a map \(P_{11} \to P_0\).
The map from \(\Ext^1_{\Abel^X}\)
to \(\textup{KK}^X_0\)
in the UCT is defined by composing with the boundary map in the
exact triangle that contains the given resolution. So the map
\(P_{11} \to P_0\)
found above is the relevant component of~\(\varphi\).
The formula in the lemma follows.
\end{proof}
According to the recipe in Theorem~\ref{the:compute_obstruction}, the
obstruction class in \(\Ext^2_{\Ideal^X_\textup{K}}(\Sigma A, A)\)
is the composite of the parity-reversing part of~\(\varphi\),
viewed as an element of \(\Ext^1_{\Ideal^X_\textup{K}}(P_1,P_0)\),
with \(f\in \textup{KK}^X_0(P_0, A)\)
and with the class of the extension~\eqref{eq:fil_K_resolution} in
\(\Ext_{\Ideal^X_\textup{K}}(A,P_1)\).
Composing the two extensions gives a length-\(2\) resolution
\begin{equation}
\label{eq:resolution_1_for_A}
P_{11} \rightarrowtail P_{10} \to P_0 \twoheadrightarrow A.
\end{equation}
The component \(\mathcal{R}_{[a,n]}\otimes Q_1[a,b] \to A\)
in the composite map \(P_{11} \to P_0 \to A\)
is the map \(\mathcal{R}_{[a,n]}\otimes Q_1[a,b] \to A\)
that corresponds to
\begin{equation}
\label{eq:odd_phi_explicit_filtered}
\sum_{[c,d] \to [a,n] \to [a,b]}
\tau_{[c,d]}^{[a,n]}\circ f[c,d] \circ \varphi_{[a,b]}^{[c,d]}
\colon Q_1[a,b] \to \textup{K}_*(A[a,n])
\end{equation}
under the isomorphism~\eqref{eq:KK_out_of_RY}. Here the sum runs only
over those \([c,d]\)
with \([c,d] \to [a,n] \to [a,b]\)
as in Lemma~\ref{lem:odd_phi}. In contrast, the sum over all
\([c,d]\)
with \([c,d] \to [a,b]\) is~\(0\) by~\eqref{eq:tau_f_phi_relation}.
In a sense, we have now computed the obstruction class. The
length-\(2\)
resolution in~\eqref{eq:resolution_1_for_A} is, however, different
from the one that is implicitly used in
Theorem~\ref{the:obstruction_class_X} to compute the relevant
\(\Ext^2\)-group
and the obstruction class in it. To translate the formula for the
obstruction class that we get from filtrated \(\textup{K}\)\nb-theory
into the setting of Theorem~\ref{the:obstruction_class_X}, we must
compare the underlying length-\(2\)
resolutions. First, we replace the resolution
in~\eqref{eq:resolution_1_for_A} by one that is
\(\Ideal^X_\textup{K}\)\nb-projective.
The entries \(P_{10}\)
and~\(P_{11}\)
are already \(\Ideal^X_\textup{K}\)\nb-projective,
and~\eqref{eq:resolution_Pi} is an \(\Ideal^X_\textup{K}\)\nb-projective
resolution of~\(P_0\).
The objects \(P_i\)
and~\(P_{ij}\)
are all sums over \(1\le a \le b \le n\),
with summands of the form \(\mathcal{R}_{[x,y]} \otimes Q_i[a,b]\)
for suitable \(x,y\)
depending on \(a,b\);
the summands in~\(P_{i1}\)
are~\(0\)
for \(b=n\).
Let \((\tau^* \otimes \varphi)\)
denote the map between these sums for \(i=1\)
to those for \(i=0\) with matrix entries
\[
\bigl(\tau_{[z,w]}^{[x,y]}\bigr)^* \otimes \varphi_{[a,b]}^{[c,d]}\colon
\mathcal{R}_{[x,y]} \otimes Q_1[a,b]
\to \mathcal{R}_{[z,w]} \otimes Q_0[c,d].
\]
As usual, \(\tau_{[z,w]}^{[x,y]}=0\) if \(\mathcal{NT}_*([z,w],[x,y]) =0\).
\begin{lemma}
\label{lem:commuting_P_diagram}
There is a commuting diagram
\[
\begin{tikzcd}[column sep = huge]
P_{11} \arrow[r, rightarrowtail, "\bigoplus \tau^*\otimes \textup{id}"]
\arrow[d, "(\tau^* \otimes \varphi)"] &
P_{10}
\arrow[r, twoheadrightarrow, "\bigoplus \tau^*\otimes \textup{id}"]
\arrow[d, "(\tau^* \otimes \varphi)"] &
P_1 \arrow[d, "(\tau^* \otimes \varphi)"] \\
P_{01} \arrow[r, rightarrowtail, "\bigoplus \tau^*\otimes \textup{id}"'] &
P_{00}
\arrow[r, twoheadrightarrow, "\bigoplus \tau^*\otimes \textup{id}"'] &
P_0
\end{tikzcd}
\]
\end{lemma}
\begin{proof}
We compare maps between direct sums by comparing their matrix
coefficients. For the two composite maps \(P_{11} \to P_{00}\),
these are maps
\(\mathcal{R}_{[a,n]} \otimes Q_1[a,b] \to \mathcal{R}_{[M(c,d),n]}
\otimes Q_0[c,d]\)
for \(1 \le a \le b <n\)
and \(1 \le c \le d \le n\).
The composite map through~\(P_{10}\)
is
\(\bigl(\tautr{M(c,d),n}{a,n}\bigr)^* \otimes
\varphi_{[a,b]}^{[c,d]}\)
if \([a,n] \leftarrow [M(a,b),n] \leftarrow [M(c,d),n]\),
and~\(0\)
otherwise; and the composite map through~\(P_{01}\)
is
\(\bigl(\tautr{M(c,d),n}{a,n}\bigr)^* \otimes
\varphi_{[a,b]}^{[c,d]}\)
if \([a,n] \leftarrow [c,n] \leftarrow [M(c,d),n]\)
and \(d<n\),
and~\(0\)
otherwise; the condition \(d<n\)
comes in because~\(P_{01}\)
contains only summands \(\mathcal{R}_{[c,n]} \otimes Q_0[c,d]\)
with \(1 \le c \le d <n\).
In both cases, the map vanishes unless \([c,d] \to [a,b]\)
because of the factor~\(\varphi_{[a,b]}^{[c,d]}\).
We claim that if \([c,d] \to [a,b]\)
and \(b<n\),
then \([a,n] \leftarrow [M(a,b),n] \leftarrow [M(c,d),n]\)
if and only if \([a,n] \leftarrow [c,n] \leftarrow [M(c,d),n]\)
and \(d<n\);
here \(M(a,b)=b+1\)
because \(b<n\).
Indeed, if \(d=n\),
then \(M(c,d)=c\),
and \([a,n] \leftarrow [b+1,n] \leftarrow [c,n]\)
means \(a \le b+1 \le c\),
which contradicts \([c,n] \to [a,b]\).
If \(d<n\),
then \(M(c,d)=d+1\).
Then \([a,n] \leftarrow [c,n] \leftarrow [d+1,n]\)
and \([a,n] \leftarrow [M(a,b),n] \leftarrow [M(c,d),n]\)
are equivalent to \(a \le c \le d+1\)
and \(a \le b+1 \le d+1\),
respectively. If \([c,d] \to [a,b]\),
both conditions say that we are in the case \(a \le c \le b \le d\).
The computations above show that the two maps \(P_{11} \to P_{00}\)
are equal.
Now consider the two maps \(P_{10} \to P_0\).
Its matrix coefficients are maps
\[
\mathcal{R}_{[M(a,b),n]} \otimes Q_1[a,b] \to \mathcal{R}_{[c,d]} \otimes Q_0[c,d],
\qquad
1 \le a \le b \le n,\ 1 \le c \le d \le n.
\]
As above, the composite maps through~\(P_{00}\)
and~\(P_1\)
are
\(\bigl(\tautr{c,d}{M(a,b),n}\bigr)^* \otimes
\varphi_{[a,b]}^{[c,d]}\)
or~\(0\).
For the maps through \(P_1\)
and~\(P_{00}\),
the former case occurs if
\([M(a,b),n] \leftarrow [a,b] \leftarrow [c,d]\)
or \([M(a,b),n] \leftarrow [M(c,d),n] \leftarrow [c,d]\),
respectively. We may assume \([a,b] \leftarrow [c,d]\)
and \([M(a,b),n] \leftarrow [c,d]\)
because otherwise \(\varphi_{[a,b]}^{[c,d]} = 0\)
or \(\tautr{c,d}{M(a,b),n}=0\).
Under these assumptions,
\([M(a,b),n] \leftarrow [a,b] \leftarrow [c,d]\)
always holds by~\eqref{eq:compose_tau}. And
\([c,d] \to [M(a,b),n]\)
implies \([M(a,b),n] \leftarrow [M(c,d),n] \leftarrow [c,d]\)
because any natural transformation \(\textup{K}_*(A[c,d]) \to \textup{K}_*(A[e,n])\)
for some \(1 \le e \le n\)
factors through \(\tautr{c,d}{M(c,d),n}\).
So the two maps \(P_{10} \to P_0\) are equal as well.
\end{proof}
Using also the resolution~\eqref{eq:fil_K_resolution} of~\(A\),
we get the following \(\Ideal^X_\textup{K}\)\nb-projective
resolution of~\(A\):
\[
P_{11} \xrightarrow{\begin{pmatrix}
-(\tau^* \otimes \varphi)\\ \bigoplus \tau^* \otimes \textup{id}
\end{pmatrix}}
P_{01} \oplus P_{10}
\xrightarrow{\begin{pmatrix} \bigoplus \tau^* \otimes \textup{id}&
(\tau^* \otimes \varphi)
\end{pmatrix}}
P_{00} \xrightarrow{\bigoplus f[a,b]^\# \circ (\tau^* \otimes \textup{id})}
A.
\]
The computation of the obstruction class in
Theorem~\ref{the:classification_X_obstruction} starts with the
following \(\Ideal^X\)\nb-projective
resolution of length~\(1\) in~\(\KKcat^X\):
\begin{equation}
\label{eq:exact_sequence_AjbA}
\bigoplus_{b=1}^{n-1} \mathcal{R}_{[b,n]} \otimes A[b+1,n]
\rightarrowtail
\bigoplus_{b=1}^n \mathcal{R}_{[b,n]} \otimes A[b,n]
\twoheadrightarrow
A.
\end{equation}
Since \(\mathcal{R}_{[b,n]} = i_b(\mathbb C)\)
in the notation of Section~\ref{sec:UCT_X}, we have
\(\mathcal{R}_{[b,n]} \otimes A[b,n] \cong i_b(A[b,n])\).
The restriction of the second map in~\eqref{eq:exact_sequence_AjbA} to
this direct summand is the one that corresponds to the identity map
on~\(A[b,n]\)
under the isomorphism in~\eqref{eq:KK_out_ix}. The first map
in~\eqref{eq:exact_sequence_AjbA}, restricted to the summand
\(\mathcal{R}_{[b,n]} \otimes A[b+1,n]\),
is the difference of the two maps
\begin{alignat*}{2}
(\tautr{b+1,n}{b,n})^*\otimes \textup{id} &&\colon
\mathcal{R}_{[b,n]}\otimes A[b+1,n]
&\to \mathcal{R}_{[b+1,n]}\otimes A[b+1,n],\\
\textup{id}\otimes \tautr{b+1,n}{b,n}&&\colon
\mathcal{R}_{[b,n]} \otimes A[b+1,n]
&\to \mathcal{R}_{[b,n]} \otimes A[b,n],
\end{alignat*}
where \(\tautr{b+1,n}{b,n}\)
denotes the inclusion of~\(A[b+1,n]\)
into~\(A[b,n]\);
we could have written~\(i\)
for~\(\tautr{b+1,n}{b,n}\)
as in Figure~\ref{fig:filtrated_K}. It is shown in
Section~\ref{sec:UCT_X} that this sequence is \(\Ideal^X\)\nb-exact.
And it is easy to prove this directly.
The projective resolutions of Abelian groups
in~\eqref{eq:varphi_f_exact_sequence_ab} imply that there is an
\(\Ideal^X_\textup{K}\)\nb-projective resolution
\[
\bigoplus_{[c,d] \to [b+1,n]} \mathcal{R}_{[b,n]} \otimes Q_1[c,d]
\rightarrowtail
\bigoplus_{[c,d] \to [b+1,n]} \mathcal{R}_{[b,n]} \otimes Q_0[c,d]
\twoheadrightarrow \bigoplus_{b=1}^{n-1} \mathcal{R}_{[b,n]} \otimes A[b+1,n].
\]
Splicing it with the resolution in~\eqref{eq:exact_sequence_AjbA}
gives an \(\Ideal^X_\textup{K}\)\nb-exact chain complex
\[
W_1 \rightarrowtail W_0
\to \bigoplus_{b=1}^n \mathcal{R}_{[b,n]} \otimes A[b,n]
\twoheadrightarrow A
\]
with
\begin{equation}
\label{eq:def_Wi}
W_i \mathrel{\vcentcolon=} \bigoplus_{b=1}^{n-1} \bigoplus_{[c,d] \to [b+1,n]}
\mathcal{R}_{[b,n]} \otimes Q_i[c,d],\qquad i=0,1.
\end{equation}
Next we are going to compare the two \(\Ideal^X_\textup{K}\)\nb-exact
chain complexes built above. We are going to build maps
\(\gamma_{ij}\)
and~\(\delta\)
for \(0\le i,j \le 1\)
that make the following diagram commute, and such that~\(\delta\)
gives the obstruction class:
\begin{equation}
\label{eq:compare_resolutions}
\begin{tikzcd}[column sep = large]
P_{11} \arrow[r, rightarrowtail]
\arrow[d, "\gamma_{11}"] &
P_{01} \oplus P_{10}
\arrow[r]
\arrow[d, "(\gamma_{01}\ \gamma_{10})"] &
P_{00} \arrow[r, twoheadrightarrow] \arrow[d, "\gamma_{00}"] &
A \arrow[d, equal] \\
W_1
\arrow[r, rightarrowtail] \arrow[d, "\delta"] &
W_0
\arrow[r, twoheadrightarrow] &
\bigoplus_{e=1}^n \mathcal{R}_{[e,n]} \otimes A[e,n] \arrow[r] &
A\\
A
\end{tikzcd}
\end{equation}
We describe maps between direct sums through matrices of maps between
the direct summands. Recall that~\(W_i\)
is defined in~\eqref{eq:def_Wi} and that
\[
P_{i0} \mathrel{\vcentcolon=} \bigoplus_{1\le a\le b\le n}
\mathcal{R}_{[M(a,b),n]} \otimes Q_i[a,b],\qquad
P_{i1} \mathrel{\vcentcolon=} \bigoplus_{1\le a\le b< n} \mathcal{R}_{[a,n]} \otimes Q_i[a,b]
\]
for \(i=0,1\). The matrix coefficients of~\(\gamma_{00}\) are maps
\[
\gamma_{00}^{e,[a,b]}\colon
\mathcal{R}_{[M(a,b),n]} \otimes Q_0[a,b] \to \mathcal{R}_{[e,n]}
\otimes A[e,n]
\]
for \(1 \le a \le b \le n\)
and \(1 \le e \le n\).
We let \(\gamma_{00}^{e,[a,b]}=0\)
if \(e\neq M(a,b)\).
Let \(e=M(a,b)\). Then \(\gamma_{00}^{e,[a,b]}\) corresponds to a map
\[
(\gamma_{00}^{e,[a,b]})^\flat\colon
Q_0[a,b] \to \textup{K}_*(\mathcal{R}_{[e,n]}[e,n] \otimes A[e,n])
\]
under the isomorphism~\eqref{eq:KK_out_of_RY}. We have already used
above that \(\mathcal{R}_{[e,n]} \otimes A[e,n] \cong i_{e}(A[e,n])\);
so
\begin{equation}
\label{eq:Ren_Aen}
\mathcal{R}_{[e,n]}[e,n] \otimes A[e,n] \cong A[e,n].
\end{equation}
Using this isomorphism implicitly, we let
\[
(\gamma_{00}^{e,[a,b]})^\flat \mathrel{\vcentcolon=}
\tautr{a,b}{e,n}\circ f[a,b]\colon
Q_0[a,b] \xrightarrow{f[a,b]}
\textup{K}_*(A[a,b]) \xrightarrow{\tautr{a,b}{e,n}}
\textup{K}_*(A[e,n]).
\]
As usual, this is~\(0\)
unless \([a,b] \to [e,n]\).
The map \(\gamma_{01}\colon P_{01} \to W_0\)
is given by a matrix of maps
\[
\gamma_{01}^{e,[c,d],[a,b]}\colon
\mathcal{R}_{[a,n]} \otimes Q_0[a,b] \to \mathcal{R}_{[e,n]} \otimes
Q_0[c,d]
\]
for \(1 \le a \le b < n\),
\(1 \le e < n\),
and \(1 \le c \le d \le n\) with \([c,d] \to [e+1,n]\). We let
\[
\gamma_{01}^{e,[c,d],[a,b]}\mathrel{\vcentcolon=}
\begin{cases}
\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \textup{id}_{Q_0[a,b]}
&\text{if } [a,b]=[c,d],\\
0&\text{otherwise.}
\end{cases}
\]
If \(a=c\)
and \(b=d<n\),
then \([c,d] \to [e+1,n]\)
if and only if \(a \le e \le b\),
so that \(\tautr{e,n}{a,n}\neq0\) in this formula.
The map \(\gamma_{10}\colon P_{10} \to W_0\) is given by a matrix of maps
\[
\gamma_{10}^{e,[c,d],[a,b]} \colon
\mathcal{R}_{[M(a,b),n]} \otimes Q_1[a,b] \to \mathcal{R}_{[e,n]}
\otimes Q_0[c,d]
\]
for \(1 \le a \le b \le n\),
\(1 \le e < n\),
and \(1 \le c \le d \le n\) with \([c,d] \to [e+1,n]\). We let
\[
\gamma_{10}^{e,[c,d],[a,b]} \mathrel{\vcentcolon=}
\begin{cases}
\bigl(\tautr{e,n}{M(a,b),n}\bigr)^* \otimes \varphi_{[a,b]}^{[c,d]}
&\text{if }M(a,b) \le e < M(c,d),\\
0&\text{otherwise.}
\end{cases}
\]
The map \(\gamma_{11}\colon P_{11} \to W_1\)
is given by a matrix of maps
\[
\gamma_{11}^{e,[c,d],[a,b]}\colon
\mathcal{R}_{[a,n]} \otimes Q_1[a,b] \to
\mathcal{R}_{[e,n]} \otimes Q_1[c,d]
\]
for \(1 \le a \le b < n\),
\(1 \le e < n\),
and \(1 \le c \le d \le n\) with \([c,d] \to [e+1,n]\). We let
\[
\gamma_{11}^{e,[c,d],[a,b]}\mathrel{\vcentcolon=}
\begin{cases}
\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \textup{id}_{Q_1[a,b]}
&\text{if }a=c,\ b=d,\\
0&\text{otherwise}.
\end{cases}
\]
The map~\(\delta\)
is given by a family of maps
\(\delta_{b,[c,d]}\colon \mathcal{R}_{[b,n]} \otimes Q_1[c,d] \to A\)
for \(1 \le b <n\)
and \([c,d] \to [b+1,n]\).
These correspond to maps
\(\delta_{b,[c,d]}^\flat \colon Q_1[c,d] \to \textup{K}_*(A[b,n])\)
by the isomorphism~\eqref{eq:KK_out_of_RY}. We define~\(\delta\)
so that the maps \(\delta_{b,[c,d]}^\flat\)
are the maps denoted by~\(\delta_b^{[c,d]}\)
in Theorem~\ref{the:compare_obstruction_filtered}.
Now we must prove that the squares in the diagram commute. We begin
on the right, comparing the two maps \(P_{00} \to A\).
Its restrictions \(\mathcal{R}_{[M(a,b),n]} \otimes Q_0[a,b] \to A\)
correspond to maps \(Q_0[a,b] \to \textup{K}_*(A[M(a,b),n])\)
under the isomorphism~\eqref{eq:KK_out_of_RY}. This map is
\(\tautr{a,b}{M(a,b),n} \circ f[a,b]\)
both for the direct boundary map \(P_{00} \to A\)
and for the map through
\(\bigoplus_{e=1}^n \mathcal{R}_{[e,n]} \otimes A[e,n]\).
So this square commutes.
Next we compare the two maps from \(P_{10}\oplus P_{01}\)
to~\(\bigoplus_{b=1}^n \mathcal{R}_{[b,n]} \otimes A[b,n]\).
We first consider the restriction to~\(P_{01}\),
then to~\(P_{10}\).
The matrix coefficients of the map on~\(P_{01}\)
are maps
\(\mathcal{R}_{[a,n]} \otimes Q_0[a,b] \to \mathcal{R}_{[e,n]} \otimes
A[e,n]\)
for \(1 \le a \le b < n\)
and \(1 \le e \le n\).
Such maps correspond to group homomorphisms
\(Q_0[a,b] \to \textup{K}_*(\mathcal{R}_{[e,n]}[a,n] \otimes A[e,n])\).
Recall that \(\mathcal{R}_{[e,n]}[a,n] = \mathbb C\)
if \(a \le e\)
and~\(0\)
otherwise. So we may assume without loss of generality that
\(a \le e\),
and then we get corresponding maps \(Q_0[a,b] \to \textup{K}_*(A[e,n])\).
The map~\(\gamma_{00}\)
picks out the summand with \(e = M(a,b) = b+1\),
and \(\bigl(\tautr{M(a,b),n}{a,n}\bigr)^*\)
induces the identity map
\(\mathbb Z \cong \textup{K}_*\bigl(\mathcal{R}_{[a,n]}([b+1,n])\bigr) \to
\textup{K}_*\bigl(\mathcal{R}_{[b+1,n]}([b+1,n])\bigr) \cong \mathbb Z\).
Therefore, the map in the square through~\(\gamma_{00}\)
contributes the map
\[
\delta_{e,b+1} \tautr{a,b}{e,n}\circ f[a,b]\colon Q_0[a,b]
\to \textup{K}_*(A[e,n]).
\]
When we map through~\(\gamma_{01}\)
instead, then we first map \(\mathcal{R}_{[a,n]} \otimes Q_0[a,b]\)
to the direct sum of \(\mathcal{R}_{[g,n]} \otimes Q_0[a,b]\)
over all \(g\in [a,b]\)
using \(\tau^*\otimes\textup{id}\)
and then apply the boundary map on~\(W_0\).
This gives a contribution in \(\textup{K}_*(A[e,n])\)
if \(g=e\)
or \(g=e-1\),
and these two contributions cancel each other for \(a<e \le b\).
For \(e=b+1\),
we get the same term as for the map that goes through~\(P_{00}\).
And we get~\(0\)
for \(e=a\)
because
\(\tautr{a+1,n}{a,n} \circ \tautr{a,b}{a+1,n} = \tautr{a,b}{a,n} =
0\). So the two maps are equal on~\(P_{01}\).
The matrix coefficients of the two maps on~\(P_{10}\)
are maps
\(\mathcal{R}_{[M(a,b),n]} \otimes Q_1[a,b] \to \mathcal{R}_{[e,n]}
\otimes A[e,n]\)
for \(1 \le a \le b \le n\)
and \(1 \le e \le n\).
As above, we may assume \(M(a,b) \le e\)
because otherwise any such map is zero. And then maps
\(\mathcal{R}_{[M(a,b),n]} \otimes Q_1[a,b] \to \mathcal{R}_{[e,n]}
\otimes A[e,n]\)
correspond to maps \(Q_1[a,b] \to \textup{K}_*(A[e,n])\).
We shall examine the difference of the map through~\(P_{00}\)
and the map through~\(W_0\).
We first consider the map through~\(P_{00}\).
It first applies the matrix \(\tau^* \otimes \varphi\),
going to the direct sum of
\(\mathcal{R}_{[M(c,d),n]} \otimes Q_0[c,d]\)
for \(1 \le c \le d \le n\).
The map \(Q_1[a,b] \to \textup{K}_*(A[e,n])\)
for the composite map through
\(\mathcal{R}_{[M(c,d),n]} \otimes Q_0[c,d]\)
is
\(\delta_{M(c,d),e} \tautr{c,d}{e,n} \circ f[c,d]\circ
\varphi_{[a,b]}^{[c,d]}\).
So we get the sum of these terms over all \(1\le c \le d \le n\).
When we apply \(\gamma_{10}\colon P_{10}\to W_0\),
then we apply the maps
\(\bigl(\tautr{g,n}{M(a,b),n}\bigr)^* \otimes
\varphi_{[a,b]}^{[c,d]}\)
to the direct summands \(\mathcal{R}_{[g,n]} \otimes Q_0[c,d]\)
of~\(W_0\),
where \(1 \le c \le d \le n\)
and \(1 \le g < n\)
are such that \([c,d] \to [g+1,n]\)
and \(M(a,b) \le g < M(c,d)\).
The condition \([c,d] \to [g+1,n]\)
is equivalent to \(c > g\)
if \(d=n\)
and \(c \le g < d+1\)
if \(d<n\).
So the set of~\(g\)
that are allowed is an interval \([x,y]\)
or empty. The upper bound is always \(y \mathrel{\vcentcolon=} M(c,d)-1\).
The lower bound~\(x\)
is \(M(a,b)\)
if \(d=n\)
or the maximum of \(c\)
and~\(M(a,b)\)
if \(d<n\).
By convention, we redefine \(x \mathrel{\vcentcolon=} M(c,d)\)
if the lower bound is bigger than~\(M(c,d)\).
So~\(g\)
runs through the interval~\([x,y]\)
if \(x\le y\),
and otherwise \(x=y+1 = M(c,d)\)
and the set of possible~\(g\) is empty.
We must compose~\(\gamma_{10}\)
with the boundary map on~\(W_0\).
As above, this only contributes to the map
\(\mathcal{R}_{[M(a,b),n]} \otimes Q_1[a,b] \to \mathcal{R}_{[e,n]}
\otimes A[e,n]\)
if \(g=e\)
or \(g=e-1\).
And the contribution to the corresponding map
\(Q_1[a,b] \to \textup{K}_*(A[e,n])\)
is \(-\tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]}\)
if \(g=e\)
and \(+\tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]}\)
if \(g=e-1\).
The contributions for \(g=e\)
and \(g=e-1\)
cancel if both occur. Therefore, when we sum over all~\(g\)
in the interval~\([x,y]\)
above, we get
\(-\tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]}\)
if \(e=x\),
\(\tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]}\)
if \(e=y+1\), and~\(0\) otherwise. So we get the map
\[
(\delta_{e,y+1}-\delta_{e,x}) \cdot
\tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]}\colon
Q_1[a,b] \to \textup{K}_*(A[e,n]).
\]
This formula remains correct if no~\(g\)
is allowed because then \(x=y+1\).
Since \(y+1 = M(c,d)\),
the map involving~\(\delta_{e,y+1}\)
is equal to the one that we get from the map through~\(P_{00}\).
So when we take the difference of the two maps in the square, this
term is cancelled. We remain with
\begin{equation}
\label{eq:sum_to_vanish}
\sum_{[c,d]\to [a,b]} \delta_{e,x} \cdot \tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]},
\end{equation}
where~\(x\)
depends on~\(a,b,c,d\)
as above. Recall that we only need the case \(e \le M(a,b)\).
We are going to prove that the sum in~\eqref{eq:sum_to_vanish}
vanishes under this assumption. First we have to study the lower
bound~\(x\)
for the different order relations among \(a,b,c,d\).
We may assume \([c,d] \to [a,b]\)
because otherwise \(\varphi_{[a,b]}^{[c,d]} = 0\).
So either \(a \le c \le b \le d\)
or \(c+1 \le a \le d+1 \le b\).
It is also important whether \(b=n\)
or \(d=n\).
First assume \(b=n\).
So \(M(a,b) = a\).
If \(a \le c \le b = d = n\),
then \(M(c,d) = c\)
and so \(x= a\).
If \(c+1 \le a \le d+1 \le b = n\),
then \(d<n\).
So \(M(c,d) = d+1\)
and \(x=a\)
as well. Therefore, \(x=a\) and \([e,n] = [a,b]\)
whenever \(b=n\).
In this case, the sum in~\eqref{eq:sum_to_vanish} vanishes because
of~\eqref{eq:tau_f_phi_relation}.
Now assume \(b<n\),
so \(M(a,b) = b+1\).
If \(a \le c \le b < d = n\),
then \(M(c,d) = c < b+1\).
So \(x= M(c,d)\)
and the summand in~\eqref{eq:sum_to_vanish} vanishes because
\(e=x < M(a,b)\).
And \(\tautr{a,b}{b+1,n} \cdot \tautr{c,d}{a,b} =0\)
as well. If \(a \le c \le b \le d < n\),
then \(M(c,d) = d+1\)
and \(x=b+1\).
In this case, we have
\(\tautr{c,d}{b+1,n} = \tautr{a,b}{b+1,n} \cdot \tautr{c,d}{a,b} \neq
0\). So we may rewrite the sum in~\eqref{eq:sum_to_vanish} as
\[
\sum_{[c,d]\to [a,b]} \delta_{e,x} \cdot \tautr{c,d}{e,n} \circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]}
= \sum_{[c,d]\to [a,b]} \tautr{a,b}{b+1,n} \circ \tautr{c,d}{a,b}
\circ f[c,d]\circ \varphi_{[a,b]}^{[c,d]},
\]
and this vanishes by~\eqref{eq:tau_f_phi_relation}. This proves the
vanishing of~\eqref{eq:sum_to_vanish} in all cases and finishes the
proof that the square of maps
\(P_{01} \oplus P_{10} \to \bigoplus \mathcal{R}_{[e,n]} \otimes
A[e,n]\) commutes.
Next, we consider the maps \(P_{11} \to W_0\).
We look at the matrix coefficient
\(\mathcal{R}_{[a,n]} \otimes Q_1[a,b] \to \mathcal{R}_{[e,n]} \otimes
Q_0[c,d]\)
of the maps \(P_{11} \to W_0\)
through \(W_1\),
\(P_{10}\)
and~\(P_{01}\)
for fixed \(1 \le a \le b < n\),
\(1 \le e < n\),
\([c,d] \to [e+1,n]\).
In~\(W_1\),
we have summands \(\mathcal{R}_{[e,n]} \otimes Q_1[a,b]\)
for those~\(e\)
with \([a,b] \to [e+1,n]\),
which is equivalent to \(a \le e \le b\)
because \(b<n\).
The map~\(\gamma_{11}\)
maps the summands \(\mathcal{R}_{[a,n]} \otimes Q_1[a,b]\)
in~\(P_{11}\)
to each of these summands through
\(\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \textup{id}\).
The boundary map \(W_1 \to W_0\)
is obtained by tensoring the boundary map
in~\eqref{eq:varphi_f_exact_sequence_ab} with
\(\textup{id}_{\mathcal{R}[e,n]}\).
So we get the contribution
\(\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \varphi_{[a,b]}^{[c,d]}\)
to our matrix coefficient if \([c,d] \to [a,b] \to [e,n]\),
and~\(0\)
otherwise. The boundary map to~\(P_{10}\)
maps the summands \(\mathcal{R}_{[a,n]} \otimes Q_1[a,b]\)
in~\(P_{11}\)
to the summand \(\mathcal{R}_{[b+1,n]} \otimes Q_1[a,b]\)
through \(\bigl(\tautr{b+1,n}{a,n}\bigr)^* \otimes \textup{id}\).
When we continue with~\(\gamma_{10}\),
we get the contribution
\(\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \varphi_{[a,b]}^{[c,d]}\)
to our matrix coefficient if and only if \(M(a,b) \le e < M(c,d)\).
The map through~\(P_{01}\)
gives the contribution
\(-\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \varphi_{[a,b]}^{[c,d]}\)
to our matrix coefficient if \(d<n\)
and \(a \le c \le e\)
(recall that the summands \(\mathcal{R}_{[c,n]} \otimes Q_0[c,d]\)
of~\(P_{01}\) only run over \(1 \le d < n\)).
We must show that the sum of these terms is~\(0\).
Again we look at different cases regarding the order among
\(a,b,c,d,e\).
We may assume \(b<n\)
and \([c,d] \to [e+1,n]\)
because our matrix coefficient is only defined in this case. And we
may assume \([e,n] \to [a,n]\)
and \([c,d] \to [a,b]\)
because otherwise
\(\bigl(\tautr{e,n}{a,n}\bigr)^* \otimes \varphi_{[a,b]}^{[c,d]}=0\).
Besides \(b<n\),
these assumptions mean, first, that \(e+1 \le c \le n = d\)
or \(c+1 \le e+1 \le d+1 \le n\)
holds; secondly, \(a \le e\);
and, thirdly, \(a \le c \le b \le d\)
or \(c+1 \le a \le d+1 \le b\).
Assume first that \(d=n\).
Then our assumptions imply \(a \le e < c \le b < d = n\).
In this case, none of the three maps \(P_{11} \to W_0\)
give a non-zero contribution because \([c,d] \to [a,b] \to [e,n]\)
is impossible, \(M(a,b) = b+1 > c = M(c,d)\)
and \(d=n\).
So we may assume \(d<n\)
from now on. If \(c+1 \le a\),
then it follows that \(c+1 \le a \le e< d+1 \le b<n\).
Again, none of the three maps \(P_{11} \to W_0\)
give a non-zero contribution in this case. So we may assume
\(a \le c\).
Then either \(a \le c \le e \le b \le d < n\)
or \(a \le c \le b < e \le d < n\).
In the first case, the maps through \(W_1\)
and~\(P_{01}\)
give contributions that cancel each other, and the map
through~\(P_{10}\)
gives no contribution because \(e < M(a,b) = b+1\).
In the second case, the maps through \(P_{10}\)
and~\(P_{01}\)
give contributions that cancel each other, and the map through~\(W_1\)
vanishes because \(b<e\).
Hence we get~\(0\)
in all cases, as needed. This finishes the proof that the square of
maps \(P_{11} \to W_0\) commutes.
Finally, we compute the composite map
\(\delta\circ \gamma_{11} \colon P_{11} \to A\)
in our commuting diagram. Consider the restriction to
\(\mathcal{R}_{[a,n]} \otimes Q_1[a,b]\)
for some \(1 \le a \le b <n\).
This map corresponds to a map \(Q_1[a,b] \to \textup{K}_*(A[a,n])\)
by~\eqref{eq:KK_out_of_RY}. The map
\(\delta_{e,[c,d]}^\flat = \delta_e^{[c,d]}\)
vanishes unless \(e=c\),
and the matrix coefficient~\(\gamma_{00}^{c,[c,d],[a,b]}\)
vanishes for \([a,b] \neq [c,d]\)
and is the identity map if \([a,b]=[c,d]\).
So the composite map corresponds simply to the map
\[
\sum_{[c,d] \to [a,n] \to [a,b]} \tautr{c,d}{a,n} \circ f[c,d] \circ \varphi_{\7{a,b}}^{\7{c,d}}
\colon Q_1[a,b] \to \textup{K}_*(A[a,n]).
\]
This is exactly the formula for the obstruction class
in~\eqref{eq:odd_phi_explicit_filtered}. This finishes the proof of
Theorem~\ref{the:compare_obstruction_filtered}.
\subsection{The case of extensions}
\label{sec:extensions}
We now specialise to the case \(n=2\).
Then an object of \(\KKcat^X\)
is equivalent to a \(\textup C^*\)\nb-algebra
extension
\[
I\overset{i}{\rightarrowtail} A \overset{r}{\twoheadrightarrow} A/I,
\]
where \(I = A[2]\),
\(A = A[1,2]\)
and \(A/I = A[1]\)
and the maps are those in~\eqref{eq:exact_module_NT}. Here we
abbreviate \([1]=[1,1]\)
and \([2]=[2,2]\).
The filtrated \(\textup{K}\)\nb-theory
is the six-periodic exact chain complex
\begin{equation}
\label{eq:K_exact_sequence}
\begin{tikzcd}
\textup{K}_0(I) \arrow[r, "i_*"] &
\textup{K}_0(A) \arrow[r, "r_*"] &
\textup{K}_0(A/I) \arrow[d, "\delta"] \\
\textup{K}_1(A/I) \arrow[u, "\delta"] &
\textup{K}_1(A) \arrow[l, "r_*"] &
\textup{K}_1(I) \arrow[l, "i_*"]
\end{tikzcd}
\end{equation}
The morphisms between the filtrated \(\textup{K}\)\nb-theory
invariants are grading-preserving chain maps (morphisms of six-term
exact sequences).
The invariant in Theorem~\ref{the:Z_classify} is the \(\textup{KK}\)\nb-class
\([i]\in \textup{KK}_0(I,A)\).
The invariant in Theorem~\ref{the:classification_X_obstruction} is the
induced map \(i_*\colon \textup{K}_*(I) \to \textup{K}_*(A)\),
together with the obstruction class. To compute the latter, let
\(i^- \in \Ext\bigl(\textup{K}_{1+*}(I), \textup{K}_*(A)\bigr)\)
be the parity-reversing part of~\([i]\)
in the Universal Coefficient Theorem for \(\textup{KK}_0(I,A)\).
The obstruction class is the image of~\(i^-\)
in the cokernel of the map
\begin{multline}
\label{eq:obstruction_cokernel}
\Ext\bigl(\textup{K}_{1+*}(I), \textup{K}_*(I)\bigr) \oplus \Ext\bigl(\textup{K}_{1+*}(A), \textup{K}_*(A)\bigr)
\to \Ext\bigl(\textup{K}_{1+*}(I), \textup{K}_*(A)\bigr),\\
(t_I,t_A) \mapsto i\circ t_I + t_A\circ i.
\end{multline}
It follows from our theory that~\(i_*\)
and the image of~\(i^-\)
in the cokernel of~\eqref{eq:obstruction_cokernel} determine an object
of~\(\Boot^X\)
uniquely up to \(\textup{KK}^X\)-equivalence.
The cokernel comes in because there are isomorphisms of
\(\textup C^*\)\nb-algebra
extensions that act identically on \(\textup{K}_*(I)\)
and~\(\textup{K}_*(A)\),
but have non-trivial components in
\(\Ext\bigl(\textup{K}_{1+*}(I), \textup{K}_*(I)\bigr)\)
or \(\Ext\bigl(\textup{K}_{1+*}(A), \textup{K}_*(A)\bigr)\).
So the isomorphism class of an object in~\(\KKcat^X\)
does not determine~\(i^-\)
uniquely. Only its image in the cokernel
of~\eqref{eq:obstruction_cokernel} is unique. And our theory shows
that isomorphism classes of pairs consisting of
\(i_*\in \Hom\bigl(\textup{K}_*(I),\textup{K}_*(A)\bigr)\)
and an element in the cokernel of~\eqref{eq:obstruction_cokernel} are
in bijection with isomorphism classes of objects in the bootstrap
class \(\Boot^X \subseteq \KKcat^X\).
We are going to compare this classification result with the filtrated
\(\textup{K}\)\nb-theory
classification by the long exact sequences
in~\eqref{eq:K_exact_sequence}. The long exact sequence
in~\eqref{eq:K_exact_sequence} contains~\(i_*\) and the extension
\begin{equation}
\label{eq:extension_from_six-term}
\coker\bigl(i\colon \textup{K}_*(I) \to \textup{K}_*(A)\bigr)
\rightarrowtail \textup{K}_*(A/I)
\twoheadrightarrow \ker\bigl(i\colon \textup{K}_{*+1}(I) \to \textup{K}_{*+1}(A)\bigr),
\end{equation}
and we may reconstruct the long exact sequence from these two pieces.
Two extensions as in~\eqref{eq:extension_from_six-term} have the same
class in Ext if and only if the long exact sequences associated to
them (for the same~\(i_*\))
are isomorphic with an isomorphism that is the identity on \(\textup{K}_*(I)\)
and~\(\textup{K}_*(A)\).
Therefore, the filtrated \(\textup{K}\)\nb-theory
invariant is equivalent to the pair consisting of
\(i_*\colon \textup{K}_*(I) \to\textup{K}_*(A)\)
and a class in \(\Ext\bigl(\ker(i_*), \coker(i_*)\bigr)\).
Now the following proposition clarifies the relationship between our
different invariants:
\begin{proposition}
\label{pro:obstruction_class_for_extensions}
The cokernel in~\eqref{eq:obstruction_cokernel} is naturally
isomorphic to the group
\[
\Ext^2_{\Abel^X}(\Sigma A,A) \cong
\Ext\Bigl(\ker\bigl(\textup{K}_{*+1}(I) \xrightarrow{i_*} \textup{K}_{*+1}(A)\bigr),
\coker\bigl(\textup{K}_*(I) \xrightarrow{i_*} \textup{K}_*(A)\bigr)\Bigr).
\]
And the obstruction class is the class that corresponds to minus the
extension in~\eqref{eq:extension_from_six-term}.
\end{proposition}
\begin{proof}
Let~\(G\)
be a \(\mathbb Z/2\)\nb-graded
Abelian group. By the long exact sequence for Hom and Ext,
the restriction map
\(\Ext\bigl(\textup{K}_{*+1}(A), G\bigr) \to \Ext\bigl(i_*(\textup{K}_{*+1}(I)),G\bigr)\)
is surjective and
\begin{multline*}
\dotsb \to
\Ext\bigl(i_*(\textup{K}_{*+1}(I)),G\bigr)
\to \Ext\bigl(\textup{K}_{*+1}(I),G\bigr)
\\\to \Ext\bigl(\ker(i_*\colon \textup{K}_{*+1}(I) \to \textup{K}_{*+1}(A)),G\bigr)
\to 0
\end{multline*}
is exact. Hence the cokernel of the map
\(i^*\colon \Ext\bigl(\textup{K}_{*+1}(A),G\bigr) \to \Ext\bigl(\textup{K}_{*+1}(I),G\bigr)\)
is \(\Ext(\ker(i_*),G)\).
If we let \(G \mathrel{\vcentcolon=} \textup{K}_*(A)\),
then we may identify the cokernel of the map
in~\eqref{eq:obstruction_cokernel} with the cokernel of the map
\[
\Ext\bigl(\textup{K}_{*+1}(I),\textup{K}_*(I)\bigr) \xrightarrow{\text{restrict}}
\Ext\bigl(\ker(i_*),\textup{K}_*(I)\bigr)
\xrightarrow{i_*} \Ext\bigl(\ker(i_*), \textup{K}_*(A)\bigr).
\]
Since the first map is surjective, this is equal to the cokernel of
\[
i_*\colon \Ext\bigl(\ker(i_*),\textup{K}_*(I)\bigr) \to \Ext\bigl(\ker(i_*), \textup{K}_*(A)\bigr).
\]
A variant of the proof above for the second variable identifies this
cokernel with the group \(\Ext\bigl(\ker(i_*), \coker(i_*)\bigr)\)
as claimed. Given a class
\(\delta\in\Ext\bigl(\textup{K}_{1+*}(I), \textup{K}_*(A)\bigr)\),
the map to \(\Ext\bigl(\ker(i_*), \coker(i_*)\bigr)\)
simply applies the bifunctoriality of Ext for the quotient map
\[
\textup{K}_*(A)\to \coker\bigl(i_*\colon \textup{K}_*(I) \to \textup{K}_*(A)\bigr)
\]
and the inclusion map
\[
\ker\bigl(i_*\colon \textup{K}_{*+1}(I) \to \textup{K}_{*+1}(A)\bigr) \to \textup{K}_{*+1}(I).
\]
It remains to compute the image of the obstruction class in
Theorem~\ref{the:compare_obstruction_filtered} in
\(\Ext\bigl(\ker(i_*),\coker(i_*)\bigr)\).
We recall the data used in
Theorem~\ref{the:compare_obstruction_filtered} in our special case.
The semi-simple part of the filtrated \(\textup{K}\)\nb-theory consists of
\begin{align*}
\textup{K}_*(A[2])_\mathrm{ss} &\cong \frac{\textup{K}_*(I)}{\delta(\textup{K}_{*-1}(A/I))}
\cong i_*(\textup{K}_*(I)) \subseteq \textup{K}_*(A),\\
\textup{K}_*(A[1,2])_\mathrm{ss} &\cong \frac{\textup{K}_*(A)}{i_*(\textup{K}_*(I))}
= \coker(i_*),\\
\textup{K}_*(A[1])_\mathrm{ss} &\cong \frac{\textup{K}_*(A/I)}{r_*(\textup{K}_*(A))}
\cong \delta_*(\textup{K}_*(A/I)) = \ker(i_*)\subseteq \textup{K}_{*-1}(I).
\end{align*}
Our construction is based on free resolutions of these
\(\mathbb Z/2\)\nb-graded
Abelian groups. In particular, we use a free resolution
\(Q_1[1] \rightarrowtail Q_0[1] \twoheadrightarrow \ker(i_*)\).
We lift these resolutions to free resolutions
\begin{gather*}
Q_1[2] \oplus Q_1[1] \rightarrowtail Q_0[2] \oplus Q_0[1] \twoheadrightarrow \textup{K}_*(A[2]),\\
Q_1[1,2] \oplus Q_1[2] \rightarrowtail Q_0[1,2] \oplus Q_0[2] \twoheadrightarrow \textup{K}_*(A[1,2]),\\
Q_1[1] \oplus Q_1[1,2] \rightarrowtail Q_0[1] \oplus Q_0[1,2] \twoheadrightarrow \textup{K}_*(A[1]),
\end{gather*}
which contain the maps \((\tau f[a,b])^\#\)
and \(\varphi_{[a,b]}^{[c,d]}\).
We shall need the maps in the third extension and put them into a
larger diagram, which commutes because of the construction of the
maps \((\tau f[a,b])^\#\) and \(\varphi_{[a,b]}^{[c,d]}\):
\begin{equation}
\label{eq:extension_filtered_K_theory_extension}
\begin{tikzpicture}[baseline=(current bounding box.west)]
\matrix(m)[cd,column sep = 6em]{
Q_1[1] \oplus Q_1[1,2]&
Q_0[1] \oplus Q_0[1,2]&
\textup{K}_*(A/I)\\
Q_1[1]&
Q_0[1]&
\frac{\textup{K}_*(A/I)}{r_*(\textup{K}_*(A))}\\
r_*(\textup{K}_*(A))&
\textup{K}_*(A/I)&
\ker(i_*)\\};
\draw[cdar,>->](m-1-1) -- node[yshift=2.2ex] {\(\scriptstyle\left( \begin{array}{@{}ll@{}\)} (m-1-2);
\draw[cdar, ->>] (m-1-2) -- node {\(\scriptstyle(f\7{1}\ r_* f\7{1,2})\)} (m-1-3);
\draw[cdar,>->](m-2-1) -- node {\(\scriptstyle\varphi_{\7{1}}^{\7{1}}\)} (m-2-2);
\draw[cdar, ->>] (m-2-2) -- node {\(\scriptstyle d_0\)} (m-2-3);
\draw[cdar,>->](m-3-1) -- node {\(\scriptstyle \text{incl.}\)} (m-3-2);
\draw[cdar, ->>] (m-3-2) -- node {\(\scriptstyle \delta_*\)} (m-3-3);
\draw[cdar, ->>] (m-1-1) -- node {\(\scriptstyle\mathrm{pr}_1\)} (m-2-1);
\draw[cdar, ->>] (m-1-2) -- node {\(\scriptstyle\mathrm{pr}_1\)} (m-2-2);
\draw[cdar, ->>] (m-1-3) -- node {\(\scriptstyle\mathrm{can.}\)} (m-2-3);
\draw[cdar, ->>] (m-2-1) -- node
{\(\scriptstyle f\7{1}\circ \varphi_{\7{1}}^{\7{1}}\)} (m-3-1);
\draw[cdar, ->>] (m-2-2) -- node {\(\scriptstyle f\7{1}\)} (m-3-2);
\draw[cdar] (m-2-3) -- node[swap] {\(\scriptstyle\cong\)} node {\(\scriptstyle\delta\)} (m-3-3);
\end{tikzpicture}
\end{equation}
The matrix coefficient~\(\delta_e^{[a,b]}\)
in Theorem~\ref{the:compare_obstruction_filtered} is defined only if
\(e = a \le b<n\).
In our case \(n=2\),
there is only one such matrix coefficient, namely,
\(\delta_1^{[1]} \colon Q_1[1] \to \textup{K}_*(A[1,2]) = \textup{K}_*(A)\).
The sum defining it has only one summand, which is indexed by
\([1,2] \to [1,2] \to [1]\). So
\[
\delta_1^{[1]} = f[1,2] \circ \varphi_{[1]}^{[1,2]} \colon Q_1[1] \to \textup{K}_*(A).
\]
We compose this map with the quotient map
\(\textup{K}_*(A) \twoheadrightarrow \coker(i_*)\).
The exact sequence~\eqref{eq:K_exact_sequence} shows that~\(r_*\)
induces an isomorphism from \(\coker(i_*)\)
onto \(r_*(\textup{K}_*(A)) \subseteq \textup{K}_*(A/I)\).
So we may as well compose with the map
\(r_*\colon \textup{K}_*(A) \to r_*(\textup{K}_*(A))\).
The composite of the two maps in the top row
in~\eqref{eq:extension_filtered_K_theory_extension} is~\(0\).
Thus \(r^* \circ \delta_1^{[1]} = - f[1] \circ \varphi_{[1]}^{[1]}\)
(compare~\eqref{eq:tau_f_phi_relation}). The commuting
diagram~\eqref{eq:extension_filtered_K_theory_extension} shows that
the group extension of~\(\ker(i_*)\)
in~\eqref{eq:extension_from_six-term} belongs to the map
\(f[1] \circ \varphi_{[1]}^{[1]}\).
The obstruction class belongs to the negative of this map.
\end{proof}
\begin{bibdiv}
\begin{biblist}
\bibselect{references}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
Encryption is an important tool for ensuring confidentiality of data. Recently,
the ECCHC -- a combination of elliptic-curve cryptography and Hill cipher --
was proposed by Dawahdeh, Yaakob and bin Othman \cite{ECCHC}.
The proposal is specifically aimed at image encryption.
The authors justify the security of the ECCHC by evaluating selected statistical
properties on sample plaintext and ciphertext images. The properties are entropy,
peak signal to noise ratio, and unified average changing intensity.
\paragraph{Our conribution.} We analyze the security of the ECCHC scheme and show
its various weaknesses that render the scheme unusable for any security sensitive
application. Evaluating some selected set of statistical
properties is not an adequate replacement for analyzing how cryptanalytic
attacks apply to an encryption scheme. In order to accentuate this point, we propose a
toy example -- the Deliberately Weak Cipher (DWC). The DWC is completely weak
(with only $8$ bit key and other serious vulnerabilities), but it attains
comparable or even better statistical parameters (depending on input image type)
than the ECCHC.
\section{The ECCHC scheme}
The ECCHC scheme combines public key cryptography based on elliptic curves with symmetric
encryption in a straightforward way. The authors tailored their proposal to encryption
of $8$-bit grayscale images of size $256\times 256$ pixels, although it can be easily
extended to other image types. To keep our presentation simple, we also use this
particular image type. The weaknesses identified in the scheme are relevant for
possible extension as well.
\subsection{Overview of the scheme}
The ECC part of the scheme is basically a Diffie-Hellman key agreement. Let
$(E,\cdot)$ be a group of points on some elliptic curve, generated by a generator
$G$. Let $|E| = p$ be a prime. A user $U$ can generate his/her private key $n_U$ by
randomly choosing from $\{0,1,\dots,p-1\}$. The corresponding public key is
$P_U =n_U G$.
Two users, $A$ and $B$, can agree on a shared point $K_I = (x,y) \in E$ by computing
$K_I=n_AP_B = n_An_BG$ (for user $A$) or $K_I=n_BP_A = n_An_BG$ (for user $B$). The
point $K_I$ is used to obtain a $2\times 2$ matrix $K$:
$$
K=\begin{bmatrix} k_{11} & k_{12} \\ k_{21} & k_{22} \end{bmatrix},
$$
where $(k_{11},k_{12}) = xG$ and $(k_{21},k_{22}) = yG$. A $4\times 4$ self-invertible
matrix $K_m$ is formed as follows ($I$ denotes a $2\times 2$ identity matrix):
$$
K_m=\begin{bmatrix} K & I-K \\ I+K & -K \end{bmatrix},
$$
Since the matrix is used for encrypting plaintexts with the alphabet of size $256$
($8$-bit grayscale) the elements are transformed mod $256$.
The second part of the scheme is Hill cipher used in the ECB (Electronic Codebook)
mode to sequentially encrypt $4\times 1$ vectors from a plaintext image.
We denote by $P_1,\ldots, P_n$ the sequence of $4\times 1$ vectors that
the image is split into (for $256\times 256$ image we have $n=256^2/4=16384$).
The ciphertext is a sequence of blocks
$C_1,\ldots,C_n$, where the $i$-th block/vector $C_i$ is computed by simple
matrix-vector multiplication: $C_i = K_m\cdot P_i$. Given that $K_m$ is
self-invertible, a decryption uses the same multiplication: $P_i = K_m\cdot C_i$.
\subsection{Weaknesses of the ECCHC scheme}
\label{sec_HCprop}
We discuss some properties of the ECCHC that, in our opinion, make the scheme insecure
and unsuitable for any security sensitive application.
\paragraph{Low entropy of the symmetric key.} The matrix $K_m$ is uniquely determined
by any of its quadrants, e.g. $4$ elements of $K \bmod 256$ are sufficient to
reconstruct $K_m$. Hence the worst case complexity of brute-force attack is very
low $\sim 2^{4\cdot 8}= 2^{32}$.
\paragraph{Known plaintext attack.} The proposal inherits the linearity of Hill cipher.
This makes it vulnerable to known plaintext attack. For $4\times 1$ plaintext
block and corresponding ciphertext block we obtain four linear equations with
$4$ unknowns (values of the matrix $K$). For random plaintext block we can find
a unique solution with high probability. Multiple distinct plaintext-ciphertext pairs
of blocks yield the key with certainty.
\paragraph{Using the symmetric transformation in the ECB mode.} Opting for the ECB mode
has a well known weakness: equal plaintext blocks are encrypted to equal ciphertext
blocks, $P_i=P_j \; \Rightarrow \; C_i=C_j$. This property can be particularly unfortunate
for images containing patterns or drawings, where equal $4\times 1$ blocks are expected.
It is certainly possible to modify the symmetric part of the ECCHC to employ a counter
or using other modes, however the proposal does not address this issue in any way.
\paragraph{Fixed points.} The structure of self-invertible matrix $K_m$ guarantees that
vectors of the form $P=(p,p,p,p)^T$, for any $p$, are fixed points regardless of
actual values in $K_m$:
$$
K_m\cdot P = \begin{bmatrix} K & I-K \\ I+K & -K \end{bmatrix} \cdot
\begin{bmatrix} p \\ p \\ p\\ p \end{bmatrix}
= \begin{bmatrix} K \cdot \begin{bmatrix} p\\p \end{bmatrix} +
(I-K) \cdot \begin{bmatrix} p\\p \end{bmatrix} \\
(I+K) \cdot \begin{bmatrix} p\\p \end{bmatrix}
-K \cdot \begin{bmatrix} p\\p \end{bmatrix} \end{bmatrix}
= \begin{bmatrix} p \\ p \\ p \\ p \end{bmatrix}.
$$
Let us illustrate the ECB mode and the fixed points issues. A checkerboard image, see Figure \ref{fig1},
is an example of plaintext image that remains intact after encrypting with any $K_m$.
Encrypted drawing with uniform color areas might be comprehensible (only slightly distorted
along edges). An example is shown in Figure \ref{fig2}.
\begin{figure}[h]
\centering\includegraphics[width=90px]{checkerboard.pdf}
\caption{Checkerboard -- invariant image for any symmetric key $K_m$.}
\label{fig1}
\end{figure}
\begin{figure}[h]
\centering\includegraphics[width=90px]{draw-plain.pdf}\hskip6em
\includegraphics[width=90px]{draw-ciph.pdf}
\caption{ECCHC -- Plaintext and ciphertext of a drawing.}
\label{fig2}
\end{figure}
\paragraph{Inadequate security analysis.} The security analysis included
in the proposal contains evaluation of these three statistical properties on few
sample images:
\begin{itemize}
\item Entropy of the encrypted image considering frequencies of pixel values.
\item Peak Signal to Noise Ratio (PSNR) using original and encrypted images.
\item Unified Average Changing Intensity (UACI) using original and encrypted images.
\end{itemize}
\noindent Such analysis neglects other cryptographic properties the symmetric cipher is expected
to satisfy. More importantly, it ignores all potential attacks (some weaknesses were discussed
in previous paragraphs). To further illustrate this point we propose an intentionally weak
cipher (weaker than the symmetric part of the ECCHC) and show that we can obtain comparable
or even better values of above properties, see Section \ref{sec-DWC}.
\paragraph{Unclear public key part of the ECCHC.} The elliptic curve key agreement part
of the ECCHC is under-specified. The proposal does not define what kind of elliptic curves
should be used, how large the parameters should be (small entropy of symmetric key $K_m$ indicates
that using standardized elliptic curves is pointless), how exactly are $xG$ and $yG$
values transformed into matrix $K$ (when to apply mod $256$ operation), etc.
\section{Deliberately weak cipher}
\label{sec-DWC}
We designed the Deliberately Weak Cipher (DWC for short) to illustrate the inadequacy
of statistical measures for assessing the strength of encryption algorithms.
Suitable values of statistical properties are necessary but by no means sufficient
condition for secure encryption. We start by reviewing properties used in the ECCHC
proposal.
\subsection{Remarks on statistical properties}
\label{sec-statprop}
Entropy for $8$-bit grayscale image $X$ is calculated as $\text{Entropy}(X) = -\sum_{i=0}^{255}
p_i \log_2 p_i$, where $p_i$ denotes a fraction of pixels with value $i$. The maximum entropy
in this case is $8$, and we expect that random noise image has entropy close to $8$.
The Peak Signal to Noise Ratio (PSNR) is used to measure quality between signal with a noise
and the original signal. In our case, let $A = (a_{i,j})_{i,j=1}^{n}$ and $B=(b_{i,j})_{i,j=1}^{n}$
be two $8$-bit grayscale images. The PSNR, expressed in decibels, is computed as follows:
\[
\text{PSNR}(A,B) = 20\cdot \log_{10} \frac{255}{\sqrt{\text{MSE}(A,B)}}\, ,
\]
where $\text{MSE}(A,B)=n^{-2}\cdot \sum_{i,j=1}^{n} (a_{i,j}-b_{i,j})^2$ is the Mean Square Error
between these two images. It is easy to calculate the expected value of the PSNR between all-black
(or all-white) image and a random noise image -- the MSE is $21717.5$,
and the corresponding PSNR is $\sim 4.7627$. Similarly, the expected value of the PSNR between two
random noise images is $\sim 7.7476$.
The Unified Average Changing Intensity (UACI) measures the average distance among pixel values
using following formula:
\[
\text{UACI}(A,B) = \frac{1}{n^2}\cdot \sum_{i,j=1}^{n} \frac{|a_{i,j}-b_{i,j}|}{255}\;\times 100\%.
\]
Trivially, the UACI of all-black (or all-white) image and a random noise image is $50$\%.
Calculating the expected value of the UACI for two random noise images yields $1/3 + 1/(3\cdot 255)
\sim 33.46$\%.
We are interested in two cases: (1) a pair of two random noise images, and (2) a pair of
monochrome black (white) image and a random noise image. The distinction between these cases
is important in the evaluation of encryption algorithms. Consider a cryptographically
strong encryption algorithm and pair of plaintext and ciphertext images. We expect that
a statistical properties for this pair are close to the first case if the plaintext is
a photograph, painting, etc.; and close to the second case if the plaintext is a drawing,
checkerboard, or simply an image with a vast majority of black/white pixels.
\subsection{The DWC algorithm}
The DWC encryption uses $8$-bit key $k$, obviously too short for any serious application.
To make it similar to the ECCHC, a core transformation used in the DWC takes $4\times 1$ input
vector. Let $P=(p_0, p_1, p_2, p_3)^T$ be an input vector of four bytes. The core transformation
is defined as follows:
\[
\text{CT}(P) = \mathbf{M} \odot (S(p_0), S(p_1), p_2, S(p_3))^T,
\]
where $S$ is an $8$-bit s-box and $\mathbf{M}\, \odot$ computes a fixed matrix
multiplication, borrowed from the AES SubBytes and MixColumns transformations,
respectively \cite{FIPS-197}. Please note that
our matrix multiplication operates on a single vector (column), instead of four
columns transformation MixColumns in the AES.
The output of CT is again $4\times 1$ vector.
The core transformation can be easily inverted by multiplying with inverse matrix
and performing three lookups to the table representing $S^{-1}$. Although we do not
care about the performance of the DWC, various enhancements in this area can be
obtained from vast literature on the implementation of the AES.
To mimic the ECCHC we split an image into a sequence of $4\times 1$ vectors
$P_1,\ldots, P_n$. The ciphertext is a sequence of blocks $C_1,\ldots,C_n$:
\[
C_i = \text{CT}(P_i \oplus i \oplus ((k \oplus \text{lsb}(i)) \ll 24)), \qquad \text{for } i=1,\ldots,n,
\]
where $\oplus$ denotes bitwise XOR operation, lsb returns the least significant byte, and
$x \ll 24$ shifts the byte $x$ to the left for $24$ position. Thus $(k \oplus \text{lsb}(i))
\ll 24)$ yields a 32-bit vector with the most significant byte being $k \oplus \text{lsb}(i)$
and other three bytes being $0$.
The decryption is straightforward: $P_i = \text{CT}^{-1}(C_i) \oplus i \oplus
((k \oplus \text{lsb}(i)) \ll 24))$, for $i=1,\ldots,n$.
\subsection{Properties of the DWC algorithm}
The DWC algorithm is very weak. The most important vulnerability is small key space
(overall, only $256$ keys), and thus being susceptible to brute-force attack.
Another weaknesses follow from fact that the core transformation CT is fixed and
does not depend on the key in any way. Therefore anyone can compute
$\text{CT}^{-1}$ on ciphertext blocks and obtain
$P_i \oplus i \oplus ((k \oplus \text{lsb}(i)) \ll 24))$. Since $i$ is known (it is
just a counter) one can easily compute $P_i \oplus (k \ll 24))$. This is the original
plaintext block with the most significant byte XOR-ed with $k$. Hence anyone can
recover $75$\% of the image without knowing the key.
Design of the DWC is aimed at obtaining good values of statistical properties, such as the entropy,
the PSNR and the UACI. We use four sample images -- two standard photographs (lena, baboon),
checkerboard, and drawing (both used in Section \ref{sec_HCprop} as well).
Figure \ref{fig3} shows a visual overview of plaintext and corresponding ciphertext
images. There are no fragments of original images, visible patterns or obvious
regularities in the ciphertext images.
\begin{figure}[h]
\centering
\includegraphics[width=90px]{lena.pdf}\hskip6em
\includegraphics[width=90px]{lena-ct.pdf} \\[1em]
\includegraphics[width=90px]{baboon.pdf}\hskip6em
\includegraphics[width=90px]{baboon-ct.pdf} \\[1em]
\includegraphics[width=90px]{checkerboard.pdf}\hskip6em
\includegraphics[width=90px]{checkerboard-ct.pdf} \\[1em]
\includegraphics[width=90px]{draw-plain.pdf}\hskip6em
\includegraphics[width=90px]{drawing-ct.pdf}
\caption{DWC on sample images.}
\label{fig3}
\end{figure}
Table \ref{tab-compare} compares the numerical values of the entropy, the PSNR and the UACI
for the DWC and the ECCHC for our sample images. The results for lena and baboon images
are very close, and we can consider the DWC and the ECCHC as equal in these types of images.
Image types like checkerboard or drawings pose a problem for the ECCHC (as already
discussed in Section \ref{sec_HCprop}). On the other hand, the DWC produces
ciphertext images with much superior statistics -- close to values we expect
for those image types, see Section \ref{sec-statprop}.
We can summarize -- the DWC is very weak cipher that excels in the entropy, the PSNR,
and the UACI properties.
\begin{table}[!ht]
\centering
\begin{tabular}{llrrr}
\toprule
algorihtm & image & Entropy & PSNR & UACI [\%] \\
\midrule
ECCHC
& lena & $7.9947$ & $9.3709$ & $28.1994$ \\
& baboon & $7.9965$ & $9.4589$ & $28.0132$ \\
& checkerboard & $1.0000$ & $\infty$ & $0.0000$ \\
& drawing & $1.6929$ & $16.2244$ & $3.9406$ \\
\midrule
DWC
& lena & $7.9974$ & $9.4180$ & $28.1236$ \\
& baboon & $7.9971$ & $9.5001$ & $27.8896$ \\
& checkerboard & $7.9979$ & $4.7623$ & $50.0049$ \\
& drawing & $7.9979$ & $4.9161$ & $48.8540$ \\
\bottomrule
\end{tabular}
\caption{Comparison of ECCHC and DWC for sample images}
\label{tab-compare}
\end{table}
\noindent\emph{Remarks.} (1) The key length of the DWC could be even smaller.
We can fix $k$ to some constant and get an encoding scheme good statistical properties
and no security at all. \\
(2) The DWC is by no means a unique construction. Various approaches can be used to
fulfill the same goals. For example, many lightweight stream ciphers with
severely reduced key length would be similarly vulnerable, while still having
good values of the entropy, the PSNR and the UACI for image encryption.
\section{Conclusion}
We showed various weaknesses of the ECCHC image encryption scheme, despite
good statistical properties published in the proposal \cite{ECCHC}.
More importantly, we illustrated insufficiency of such approach to analyzing
the strength of encryption algorithm by proposing the Deliberately Weak Cipher.
The DWC has comparable or even better statistical parameters (depending on
input image type) than the ECCHC, while being completely weak.
For any encryption scheme proposal, the real cryptanalytic assessment
of the scheme should be conducted. Evaluating some selected set of statistical
properties is not an adequate replacement for analyzing how cryptanalytic
attacks apply to the scheme.
The ECCHC should not be used for image (and other data) encryption. Use
standard encryption techniques. In case of resource-constrained devices,
the area of lightweight cryptography offers alternative algorithms with better
security. For recent survey on lightweight cryptography, see \cite{LC17}.
|
\section{Introduction}
Diabetic Retinopathy (DR) is a leading disabling chronic disease and one of the main causes of blindness and visual impairment in developed countries for diabetic patients. Studies reported that 90\% of the cases can be prevented through early detection and treatment. Eye screening through retinal images is used by physicians to detect the lesions related with this disease. Due to the increasing number of diabetic patients, the amount of images to be manually analyzed is becoming unaffordable. Moreover, training new personnel for this type of image-based diagnosis is long, because it requires to acquire expertise by daily practice \citep{pedro2010prevalence}, \citep{romero2011managing}.
Deep Learning (DL) \citep{nature-deep-learning}, \citep{Schmidhuber-nn}, is a subfield of Machine Learning that allow the automatic model construction of very effective image classifiers using a parametric model. These models are able to identify and extract the statistical regularities present in data that are important for optimizing a defined loss function, with the final objective of mapping a high-multidimensional input into a smaller multi-dimensional output (f: $\mathbb{R}^{n} \mapsto \mathbb{R}^{m}, n \gg m$). This mapping allows the classification of multi-dimensional objects into a small number of categories. The model is composed by many neurons that are organized in layers in a hierarchical way. Every neuron receives the input from a predefined set of neurons. Every connection has a parameter that corresponds to the weight of the connection. The function of every neuron is to make a transformation of the received inputs into a calculated output value. For every incoming connection, the weight is multiplied by the input value received by the neuron. The aggregation of all inputs passes through an activation function that calculates the neuron output. Parameters are usually optimized using a stochastic gradient descent algorithm that minimizes a predefined loss function. These hierarchical models are able to learn multiple levels of representation that correspond to different levels of abstraction, which enables the representation of complex concepts in a compressed way \citep{Bengio:2013:RLR:2498740.2498889}, \citep{bengio-2009}.
Convolutional neural networks (CNN) are a deep learning subfield that proved to be very effective for image classification, detection and segmentation. The first successful CNN was presented in \citep{LeCun:98}, and it was designed for hand-written digit recognition. This early CNN implementation used a combination of convolution, pooling and non-linearity that has been the key feature of DL since now. The DL breakthrough took place with the publication of \citep{NIPS2012_4824}, where for the first time a CNN won the Imagenet\citep{imagenet_cvpr09} classification competition by a large margin. In that paper a set of innovative techniques where introduced including data augmentation, the use of rectified linear units (ReLUs) as activation function, the use of dropout for avoiding overfitting, overlapping max-pooling for avoiding the averaging effects of avg-pooling and the use of GPUs for speeding up the training time. Later on, different CNN improvements where also published. In \citep{vggnet}, it was the first time that small 3x3 convolution filters where used and combined as a sequence of convolutions. In \citep{he2016deep} the authors introduced residual networks. This networks used a combination of 3x3 convolutional layers with a by-pass of the input every two layers that was summed up to the output. Such bypasses improved the gradient propagation through the network allowing the design of deeper networks and improving the classification capabilities.
DL models have been also successfully applied in many medical classification tasks. In \citep{esteva2017dermatologist} a DL classifier was designed achieving dermatologist-level accuracy for skin cancer detection. In \citep{wentao2018deeplung} a 3D CNN for automated pulmonary nodule detection and classification was designed. In \citep{wang2018classification} a CNN Alzheimer's disease classifier with high performance was also described. In \citep{doi:10.1001/jama.2016.17216} the authors presented a diabetic retinopathy classifier with better performance than human experts in the detection of the most severe cases of the disease.
Ordinal regression is a term used for multi-class classification for the cases where a underlying property can be used for prestablishing an ordering of the classes. Several quality measures exist in the literature of machine learning and statistics for ordinal regression \citep{mehdiyev2016evaluating}. Kappa is a well-known statistic coefficient defined by Cohen \citep{cohen1960coefficient} to measure inter-rater agreement in such cases. Weighted Kappa \citep{cohen1968weighted} is another index used for measuring the goodness of a ordinal regression classification. In this last coefficient, disagreements are penalized proportionally to a power of the distance between classes. The penalization most commonly used is the quadratic. In such cases, the index is commonly referred as Quadratic Weighted Kappa (referred as QWK or $\kappa$ indistinctly in this paper).
In medical diagnosis tasks is important not only the accuracy of predictions but also the reasons behind decisions. Self-explainable models enable the physicians to contrast the information reported with their own knowledge, increasing the probability of a good diagnostic, which may have a significant influence in patient's treatment.
Different attempts have been done for interpreting the results reported by neural networks. In \citep{zeiler2014visualizing} a network propagation technique is used for input-space feature visualization. After this work, \citep{bach2015pixel} used a pixel-wise decomposition for classification decision. This decomposition could be done in two ways: considering the network as a global function, disregarding its topology (functional approach) or using the natural properties of the inherent function topology for applying a message passing technique, propagating back into the pixel space the last layer output values. After this work, in \citep{montavon2017explaining} they used a so named Deep Taylor decomposition technique to replace the inherently intractable standard Taylor decomposition using a multitude of simpler analytically tractable Taylor decompositions.
Independent Component Analysis (ICA)\citep{hyvarinen2000independent} is a statistical method for the separation of a multi-dimensional random signal into a linear combination of components that are statistically as independent from each other as possible. The theoretical foundation of ICA is based on the Central Limit Theorem, which establishes that the distribution of the sum (average or linear combination) of N independent random variables approaches a gaussian as $N \rightarrow \infty$. When ICA method is applied, it is assumed that such separation exist, ie. that is possible to express the signal as a linear combination of independent components (IC). Perfect independence between random variables is achieved when mutual information between them is zero. Mutual information can be expressed as the Kullback-Leibler divergence between the joint distribution and the product of the distributions of each variable. Mutual information can be decomposed, under linear transforms, as the sum of two terms: one term expressing the decorrelation of the components and one expressing their non-Gaussianity \citep{cardoso2003dependence}. ICA uses optimization to calculate its components. Two types of optimization objectives can be used: minimize the mutual information or maximize the non-gaussianess of each component.
In this paper we study a technique that allows the identification, separation and visualization in the input and hidden space of the IC responsible of a particular DR classification decision taken by a DL classifier given a certain eye fundus image. This is done by calculating the minimum number of IC that are able to encode the maximum information about the particular classification. Identifying such components, we reduce the redundancy of feature space and identify the components that share the minimum mutual information between each other. In that way, under the supposition that the feature extraction phase of the deep learning model has been able to disentangle completely the feature information, we are able to separate the independent elements causing the disease.
We use the pixel-wise visualization method to identify in pixel space such independent causes. We use the assessment of experts clinicians for interpreting such IC.
The paper is structured as follows: in Section \ref{sec:related} the current work on DL applied to DR is briefly introduced, then, the main works on interpretation of DL are presented. Section \ref{sec:methods} we present the methods, Section \ref{sec:results} presents the results showing samples of the kind of visual interpretations given by the model and finally Section \ref{sec:conclusions} present the final conclusions of our work.
\section{Related Work}\label{sec:related}
Many DL based DR classifiers have been published in the last years \citep{pratt2016convolutional}, \citep{DELATORRE2017}, \citep{doi:10.1001/jama.2016.17216}.
In \citep{doi:10.1001/jama.2016.17216} a binary DL classifier was published for the detection of the most severe cases of DR (grouping classes 0 and 1 of DR on one side, and classes 2, 3 and 4 on another). This model was trained using an extended version of the EyePACS dataset mentioned before with a total of 128,175 images and improving the proper tagging of the images using a set of 3 to 7 experts chosen from a panel of 54 US expert Ophtalmologists. This model surpassed the human expert capabilities, reaching at the final operating point approximately 97\% sensitivity and 93.5\% specificity in the test sets of about 10,000 images for detecting the worse cases of DR. The strength of this model was its ability to predict the more severe cases with a sensitivity and specificity greater than human experts.
In our previous work \citep{DELATORRE2017} a DL classifier was published for the prediction of five different disease grades. This model was trained using the public available EyePACS dataset. The training set had 35,126 images and the test set 53,576. The quadratic weighted kappa (QWK) evaluation metric \citep{cohen1968weighted} over the test set using a unique DL model without ensembling was close to the reported by human experts.
The drawback of all these implementations, as of many other DL based models, is its lack of interpretability. In last years different approximations have been proposed for converting the initial DL black box classifiers into interpretable ones. Between the more successful interpretation models existing today we have the sensitivity maps \citep{DBLP:journals/corr/SimonyanVZ13}, layer-wise relevance propagation \citep{bach2015pixel} and Taylor type decomposition models \citep{montavon2017explaining}.
All these methods use different strategies to backpropagate the classification scores to the input space and distribute the value of the final classification between input pixels. With this score distribution is possible to identify the most relevant pixels for a particular classification.
Layer-wise relevance propagation models allow not only to report the predicted class but also to evaluate the importance of every input pixel of the image in the final classification decision. In such a way, it is possible to determine which pixels are more important in the final decision and facilitate the human experts the construction of rationale explanations based on the interpretation of such maps.
\section{Methods}\label{sec:methods}
DL models are organized in layers, being the inputs of each one a combination of the outputs of previous ones. We design the output layer to be a linear combination of last layer feature space components. In this way we are forcing the model to disentangle the important features that, combined in a linear way, allow the achievement of a maximum possible classification score. These components (or other obtained as a linear combination of them, like with ICA analysis), are easy to analyze due to the linear nature of its relationship with the classification scores.
\subsection{ICA based interpretation procedure}
In this paper we go a step forward in the interpretation of score maps. Instead of generating directly the pixel maps associated with a particular class, we try to identify and separate independent elements associated with the disease. Our new contribution comes from the identification, separation and visualization of the IC that explain a particular classification decision. We hypothesize that in order to achieve high performance classification scores, the network has to encode the information required to make the classification. Human experts base its decisions in the number and types of lesions present in the image. That's why in some way, such information has to be present in a disentangled form in last layer feature space, previous to the output layer. Instead of directly visualizing the more important pixels under a classification decision, we split the information of such last feature layer into independent features using a Independent Component Analysis (ICA). A posteriori we use a pixel-wise relevance propagation derived method to visualize such independent component in input space. In this way, we can, not only generate importance pixel maps, but also differentiate between the underlying independent causes of the disease.
We study the last layer feature space, previous to the output layer linear combination in order to identify its properties and try to isolate the independent elements that are causing a particular classification. For this purpose we use a principal component analysis (PCA) \citep{pearson1901principal} to appraise the redundancy of this space and a ICA \citep{hyvarinen2000independent} using different number of components to identify the minimum of them required to achieve a classification score close enough to the achieved without such a dimensional reduction. ICA allows to find a linear representation of non-Gaussian data so that the components are statistically independent, or as independent as possible. Such a representation seems to capture the essential structure of the data in many applications, including feature extraction \citep{hyvarinen2000independent}. When the data is not Gaussian, there are higher order statistics beyond variance that are not being taken into account by PCA. While PCA captures only uncorrelated components, these uncorrelated components are not necessarily independent for general distributions. ICA minimizes the mutual information (or relative Kullback-Leibler divergence) of non-Gaussian data because two distributions with zero mutual information are statistically independent \citep{comon1992independent}.
For finding the optimal number of IC, we apply ICA to the last layer feature space training set vectors. Using different number of IC and comparing the classification performance of the original model with the obtained using a linear combination of the reduced number of calculated components, it is possible to find the optimal number of components ($n$) that does not significantly reduce the classification performance of the original model. We modify the original model adding a new layer after the last layer feature space to calculate online the components of every analyzed image. This layer is a linear transformation and acts as a dimensionality reduction layer (see fig. \ref{fig:models}). The final classification is achieved linearly combining the low dimensional IC layer.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=0.7\textwidth]{figures/initial_classifier.pdf}
\caption{Initial classification network}
\end{subfigure}
\hfill
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/ica_classifier.pdf}
\caption{Modified for visualization of independent underlying classification causes}
\label{fig:models_ica}
\end{subfigure}
\caption{Model changes done for improving explainability}
\label{fig:models}
\end{figure}
After identifying the optimal $n$, we use the receptive field and pixel-wise explanation model \citep{de2017deep} to visualize the independent scores in the input space. In this way we are visualizing not only a score map explaining a classification but also differentiating and visualizing the mathematically IC responsible of a particular classification.
\subsection{Mathematical formalization}
Let $F_{train}$ be the set of all feature vectors of training set:
\begin{equation}
F_{train} = \{\boldsymbol{f^{(i)}} : i = 1 .. T\}, \quad \boldsymbol{f^{(i)}} = (f^{(i)}_1, f^{(i)}_2, ... f^{(i)}_m)
\end{equation}
being $T$ the number of elements of the training set, and $m$ the dimension of feature space vector.
Let $S_{train}$ the set of IC calculated from $F_{train}$:
\begin{equation}
S_{train} = \{\boldsymbol{s^{(i)}} : i = 1 .. T\}, \quad \boldsymbol{s^{(i)}} = (s^{(i)}_1, s^{(i)}_2, ... s^{(i)}_n)
\end{equation}
being $n$ the number of IC, $n < m$.
Every $\boldsymbol{s^{(i)}}$ can be expressed as a linear combination of $\boldsymbol{f^{(i)}}$:
\begin{equation}
\boldsymbol{s^{(i)}} = \boldsymbol{W} \boldsymbol{f^{(i)}}
\end{equation}
Where $\boldsymbol{W}$ is calculated using a optimization method, minimizing the mutual information between $n$ IC ($\min_{S} I(S)$) \citep{hyvarinen1999fast}.
The ordinal regression problem solved using $F_{train}$ can be expressed as:
\begin{equation}
\max_{\boldsymbol{A}} \big[ \kappa_{val} (C_{train}) \big]
\end{equation}
being:
\begin{equation}
C_{train} = \{ \boldsymbol{A} \boldsymbol{f^{(i)}}, \forall \boldsymbol{f^{(i)}} \in F_{train} \}
\end{equation}
being $\boldsymbol{c^{(i)}} = \boldsymbol{A} \boldsymbol{f^{(i)}}$ the predicted class vector and $\kappa_{val}$ the evaluation function used in the training process of the neural network calculated for the validation set.
We want to solve the same problem using the reduced ICA components (feature space compressed version):
\begin{equation}
C'_{train} = \{ \boldsymbol{B} \boldsymbol{s^{(i)}}, \forall \boldsymbol{s^{(i)}} \in S_{train} \}
\end{equation}
being $\boldsymbol{c'^{(i)}} = \boldsymbol{B} \boldsymbol{s^{(i)}}$ the predicted class vector using $n$ IC.
The optimal number of IC to select is the one that minimizes the difference in performance between both models:
\begin{equation}
\min_{n} \big[ \kappa_{val} (C'_{train}) - \kappa_{val} (C_{train}) \big]
\end{equation}
\section{Results}\label{sec:results}
\subsection{Data}
In this study we use the EyePACS dataset of the Diabetic Retinopathy Detection competition hosted on the internet Kaggle Platform. For every patient right and left eye images are reported. All images are classified by ophthalmologists according to the standard severity scale presented before in \citep{diaclass}. The images are taken in variable conditions: by different cameras, illumination conditions and resolutions.
The training set contains a total of $75,650$ images; $55,796$ of class 0, $5,259$ of class 1, $11,192$ of class 2, $1,805$ of class 3 and $1,598$ of class 4. The validation set used for hyper-parameter optimization has $3,000$ images; $2,150$ of class 0, $209$ of class 1, $490$ of class 2, $61$ of class 3 and $90$ of class 4. The test set, contains a total of $10,000$ images; $7,363$ of class 0, $731$ of class 1, $1,461$ of class 2, $220$ of class 3 and $225$ of class 4.
\subsection{Baseline Model}
Our baseline model uses a 3x640x640 input image obtained from a minimal preprocessing step where only the external background borders are trimmed and later resized to the required input size. It is a CNN of 391,325 parameters, divided in 17 layers. Layers are divided in two groups: the feature extractor and the classifier. The feature extraction has 7 blocks of 2 layers. Every layer is a stack of a 3x3 convolution with stride 1x1 and padding 1x1 followed by a batch normalization and a ReLU activation function. Between every block a 2x2 max-pooling operation of stride 2x2 is applied. Afterwards, the classification phase takes place using a 2x2 convolution. A 4x4 average-pooling reduces the dimensionality to get a final 64 feature vector that are linearly combined to obtain the output scores of every class. A soft-max function allows the conversion of scores to probabilities to feed the values to the proper cost function during the optimization process. The feature extractor has 16 filters in the first block, 32 in the second and 64 in all the other.
\subsection{Model Modifications}
For all the training set we calculate the last layer feature space, obtaining a 64-dimensional vector as a representation of each image. We observe that this vector is highly redundant. After applying a PCA, with only 10 components it is possible to explain 99\% of the variance.
\begin{figure}[h]
\centering
\includegraphics[width=0.65\textwidth]{figures/ica_class_contribution.eps}
\caption{Contribution of each ICA component in the classification final score}
\label{fig:ica_contribution}
\end{figure}
Using the 64-dimensional feature-space vector ($m=64$) of all the training set, we calculate a set of ICAs using different $n$ values. With each one, we train a linear classifier to calculate the evaluation metric obtained over a validation set (fig. \ref{fig:models_ica}). We choose the minimal $n$ that allows achieving maximum performance. The optimal $n$ for this problem is $3$, achieving a $QWK_{val} = 0.790$ not far from the achieved by the original model without dimensionality reduction ($QWK_{val} = 0.800$).
Fig. \ref{fig:ica_contribution} shows the contribution of each component to the score of each class. We can see that ICA values are differentiating between extreme classes 0, 2 and 4. As we are training the network using as a loss function $QWK$, the optimization takes place as an ordinal regression. We expect the network to find the underlying causes present in the image that produce the predefined sorting of the classes, ie. the types of lesions and the number of them. Class 0 score contributions come from $ICA_0 > 0$, $ICA_1 > 0$ and $ICA_2 > 0$; being the class markers of the presence of disease $ICA_0 < 0$, $ICA_1 < 0$ and $ICA_2 < 0$. Analyzing the pixels with higher negative signals in the three components will give us the points that are contributing the most to the signaling of a possible presence of the disease. Backpropagating the scores of each one of this negative components will give a richer explanation with distinction between three possible independent causes of the final diagnostic given by the model.
We use a two-dimensional t-SNE visualization \citep{maaten2008visualizing} for qualitative evaluation of the difference between the quality of the separation using the 64 feature vector and of the reduced version with only 3 ICA components. In fig. \ref{fig:tsne} we show the 2D t-SNE visualization for the original feature space and for the ICA reduced space. We can see how for both spaces class 0, 2 and 4 classes are clearly separated. Class 0 and 1 are not properly separated and in the case of 3 and 4, although the separation is not perfect, it is possible to distinguish a different location of both classes for both spaces. As the quantitative index QWK also show, there is not a qualitative separation capability difference between the t-SNE map of the original features and the map of the ICA 3 dimensional compressed feature map. For such $1.25\%$ difference in the classification performance index ($\kappa_{orig} = 0.800$ vs $\kappa_{ICA} = 0.790$) we can conclude that the independent component analysis has been able to find a adequate compressed expression of the information encoded in the network.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/tsne2d_p75.eps}
\caption{Original feature space (64 components)}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/tsne2d_ica_p75.eps}
\caption{ICA space (3 components)}
\end{subfigure}
\caption{Comparison between the 2D t-SNE visualization of validation set using the original feature space and the final 3-dimensional ICA space}
\label{fig:tsne}
\end{figure}
\subsection{Score components contribution for a test sample}
We use a pixel-wise relevance propagation derived method for visualizing each ICA component independently. In this way, it is possible to visualize the mathematically independent contributions, enhancing the localization of different types of primary elements causing the disease. Figure \ref{fig:ica_components_class2} shows the intermediate score maps generated using a receptive field of 61x61. Figure \ref{fig:ica_components_class4} shows the same intermediate scores for a class 4 image. These hidden layer maps are useful when a general map of the lesion locations is enough.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/img_t2_p2/rf61/ica0.png}
\caption{$ICA_0 < - 2 \sigma$}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/img_t2_p2/rf61/ica1.png}
\caption{$ICA_1 < - 2 \sigma$}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/img_t2_p2/rf61/ica2.png}
\caption{$ICA_2 < - 2 \sigma$}
\end{subfigure}
\hfill
\caption{ICA score maps generated for a class 2 image (predicted also as class 2) using a receptive field of 61x61}
\label{fig:ica_components_class2}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/img_t4_p4/rf61/ica0.png}
\caption{$ICA_0 < - 2 \sigma$}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/img_t4_p4/rf61/ica1.png}
\caption{$ICA_1 < - 2 \sigma$}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/img_t4_p4/rf61/ica2.png}
\caption{$ICA_2 < - 2 \sigma$}
\end{subfigure}
\hfill
\caption{ICA score maps generated for a class 4 image (predicted also as class 4) using a receptive field of 61x61}
\label{fig:ica_components_class4}
\end{figure}
Input-space pixel scores are the most suitable when pixel detail is required for detecting the individual lesions causing the disease.
Figure \ref{fig:ica_components_c3} show the input space final scores for $ICA_2$. We show the points having contributions higher than a prestablished limit, in this case with values higher than two standard deviations.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/c3/retina_mICA2.png}
\end{subfigure}
\caption{Visualization of $ICA_2$ score map for a class 3 image ($-ICA_2 < 2 \sigma_2$). The classification model scores are $C_0 = -661.0$, $C_1 = -294.1$, $C_2 = -10.3$, $C_3 = 70.3$, $C_4 = 26.3$; $ICA_0 = 0.0174$, $ICA_1 = -0.0181$, $ICA_2 = -0.0237$}
\label{fig:ica_components_c3}
\end{figure}
$ICA_2$ score maps have an almost perfect match with real lesions having a very low rate of false positives and false negatives. $ICA_0$ and $ICA_1$ include statistical regularities not directly related with relevant lesions. From such interpretation we conclude that ICA acts as a filter separating lesion information present in images from blink artifacts, noise and other statistical regularities present in images.
\section{Conclusions}\label{sec:conclusions}
In this paper we studied some of the feature space properties of a already trained diabetic retinopathy deep learning classifier of 17 layers. We saw how, even being a small network, the redundancy of the feature space is very high. We hypothesized that if the network is able to achieve human-level classifications, in some way, it has been able to identify the important statistical regularities present in the image that are important for the classification. As last layer is a linear classifier, such properties has to be disentagled, ie. expressed in a linear way in such last layer. We applied a independent component analysis over such last layer with the method explained in the paper in order to find the optimal number of components that maximize the classification capabilities of such compressed version of the features. We experimentally proved that reducing to only 3 components is possible to achieve almost $99\%$ of the evaluation metric. Such value experimentally prove that the ICA compression has been able to extract all the important elements required for the classification. With such elements, we applied a visualization technique for identifying the elements in the image of each one of the components, using a pixel-wise derived visualization technique. We are able to generate three maps for every image each one identifying independent statistical regularities important for the classification.
Our method allows not only the classification of retinographies but also the identification and localization in the image of the mathematically independent signs of the disease. The presented ICA score model is of general applicability and can be easily adapted for the usage in other image classification tasks thus, as future work we plan to test it in other types medical images (eg. cancer detection in mammograms).
\section*{Acknowledgements}
This work is supported by the URV grants 2017PFR-URV-B2-60, and the Spanish research project PI15/01150 (Instituto de Salud Carlos III) and Fondos FEDER. The authors would like to thank to Kaggle and EyePACS for providing the data used in this paper.
\section*{References}
\input{ica.bbl}
\end{document}
|
\section{Introduction}
This paper continues a series started in Carraro et al. \citeyear{2016Car} that aims at expanding the current sample of Galactic disk intermediate age and old open clusters for which good metallicity and abundance ratios measurements are available. As for the Open Cluster Chemical Abundances from Spanish Observatories survey (OCCASO) (Casamiquela et al. \citeyear{2016C}) we are targeting clusters in the northern hemisphere, which are out of reach of the Gaia-ESO Spectroscopic Public Survey (Gilmore et al. \citeyear{2012G}). In Carraro et al. \citeyear{2016Car} we presented the first abundance study of NGC 7762, a nearby old open cluster. Here we focus on two open clusters which are located at larger distance from the Sun, toward the anti-centre direction: NGC\,6940 and Tombaugh\,5 (hereafter To\,5).\\
We aim at finding and characterising open clusters in the Galacto-centric distance range (10-12 kpc) at which the Galactic radial abundance gradient has an abrupt change of slope (Twarog et al 1997, Sales Silva et al. 2015), and keeps flat outwards. In this way, we can expand the sample of fully characterised open clusters observable in the Northern hemisphere and improve the observational constraints of chemical evolution models.\\
\noindent
In this work, we present a spectroscopic investigation of the open clusters NGC\,6940 and To\,5. For the latter no radial velocity and chemical abundance analysis have been performed to date, while for the former abundances were available for a few stars only at the time of the observing proposal preparation.
In addition, a literature scan indicates discrepant values
for To\,5 basic parameters (especially distance).
We secured spectra for seven stars in both clusters, close to the expected location of red clump stars. We derived radial velocities and metallicities, as well as abundances ratios of $\alpha$ and iron-peak elements. The selected stars are shown in the colour magnitude diagrams (CMD) in Fig.\ref{Fig1} (highlighted in red) and their properties are summarised in Tab.\ref{Tab1}.\\
\noindent
The paper is organised as follows: in Sect. 2 we present a brief summary of previous studies on NGC~6940 and To\,5, while in Sect.3 we present the observational material and the basic reduction strategy. Sect.4 is devoted to radial velocities measurements, while in Sect.5 we present the estimates of the stellar atmospheric parameters. The abundance analysis is presented in Sect.6 and Sect.7. Moreover, in Sect.8 we derive updated estimates of the clusters' fundamental parameters. Finally, Sect.9 provides some discussion and summarises our findings.
\begin{table*}
\tabcolsep 0.3truecm
\caption{Basic information of the stars for which we obtained spectroscopic data. The signal-to-noise ratios ($\mathrm{\frac{S}{N}}$) in the last column have been measured using a few \AA~ continuum region at 6300\AA.}
\label{Tab1}
\centering
\begin{tabular}{lccccccccl}
\hline
\hline
ID & RA(J2000.0) & Dec(J2000.0) & V & B-V & J & H & K & $\mathrm{t_{exp}}$ (sec) & $\mathrm{\frac{S}{N}}$ \\
\hline
\textbf{NGC\,6940} & & & & & & & & &\\
28 & 20:33:25.0 & +28:00:46.9 & 11.56 & +1.12 & 9.565 & 9.049 & 8.929 & 2400 & 68 \\
60 & 20:33:59.6 & +28:03:01.7 & 11.56 & +1.11 & 9.593 & 9.103 & 8.974 & 2400 & 92 \\
67 & 20:34:04.1 & +28:16:48.6 & 11.18 & +1.13 & 9.134 & 8.608 & 8.518 & 2000 & 135\\
108 & 20:34:25.6 & +28:13:41.5 & 11.19 & +1.04 & 9.275 & 8.857 & 8.696 & 2000 &110 \\
130 & 20:34:38.7 & +28:20:22.7 & 11.39 & +1.07 & 9.453 & 8.966 & 8.853 & 2400 & 103 \\
132 & 20:34:40.1 & +28:26:38.9 & 10.97 & +1.10 & 9.039 & 8.491 & 8.391 & 1800 & 125 \\
139 & 20:34:47.6 & +28:14:47.3 & 11.38 & +1.08 & 9.464 & 8.942 & 8.804 & 2000 & 94 \\
\hline
\textbf{To\,5} & & & & & & & & & \\
0006 & 03:48:17.46 & +59:11:19.68 & 12.198 & 1.714 & 8.888 & 8.222 & 7.960 & 2700 & 118\\
5274 & 03:46:55.60 & +58:55:34.79 & 11.982 & 1.646 & 8.993 & 8.360 & 8.171 & 2700 & 139\\
5521 & 03:48:47.41 & +59:02:28.18 & 12.238 & 1.660 & 9.231 & 8.483 & 8.279 & 2700 & 165\\
7701 & 03:47:30.98 & +59:02:50.89 & 12.011 & 1.933 & 8.531 & 7.797 & 7.581 & 2700 & 111\\
7834 & 03:47:49.92 & +58:56:20.84 & 12.769 & 1.585 & 9.624 & 9.015 & 8.782 & 3600 & 105\\
8080 & 03:48:29.20 & +59:00:36.68 & 12.726 & 1.677 & 9.628 & 8.969 & 8.752 & 3600 & 86\\
8099 & 03:48:32.98 & +59:15:16.63 & 12.239 & 1.944 & 8.651 & 7.936 & 7.699 & 2300 & 76\\
\hline\hline
\end{tabular}
\end{table*}
\begin{figure}
\includegraphics[width=\columnwidth]{cmdN.pdf}\quad\includegraphics[width=\columnwidth]{cmdT.pdf}
\caption{Color magnitude diagrams of NGC\,6940 (Larsson-Leander (\citeyear{1960lar})) and To\,5 (Maciejewski \& Niedzielski (\citeyear{2007mac})). Red symbols indicate the stars for which we obtained spectroscopic data.}
\label{Fig1}
\end{figure}
\section{Overview of the literature on NGC~6940 and Tombaugh~5}
\subsection{NGC~6940}
NGC\,6940 is a rather well-known intermediate-age open cluster (see Tab.\ref{Tab2}).
The first membership assessment was conducted by Vasilevskis \& Rach (\citeyear{1957vas}), who identified 216 possible members using proper motions. Later, Sanders (\citeyear{1972san}) confirmed the membership of 104 stars, implementing a new procedure based on the maximum likelihood method.\\
The first extensive photometric study was conducted by Larsson-Leander (\citeyear{1964lar}), who cross-correlated several previous studies and confirmed the estimates by Walker (\citeyear{1958wal}) and Johnson et al. (\citeyear{1961joh}) of the apparent distance modulus ((m-M)$_V$=9.5\,mag) and the reddening (E(B-V)=0.26, both of which
were later confirmed by Hartwick \& McClure (\citeyear{1972hart})). Jennens \& Helfer (\citeyear{1975jen}) derived the very first age estimate of the cluster assuming previous estimates of distance modulus and reddening, and finding an age of $\sim$500 Myr.\\
The first detailed spectroscopic study was conducted by Geyer \& Nelles (\citeyear{1985gey}) who derived radial velocities of 31 member stars, one of which is included in our sample. Geisler (\citeyear{1988geis}) confirmed those estimates and determined a mean radial velocity of the cluster of 5.7$\pm$1.5\,km/s. Mermilliod \& Mayor (\citeyear{1989merm}) analyzed high-resolution spectra of member stars in different open clusters, among which NGC\,6940, confirming the existence in that region of at least six spectroscopic binaries, among which our \#130. Later, Mermilliod et al. (\citeyear{2008merm}) analysed 26 possible red giants stars, confirming 21 of them and deriving a mean cluster radial velocity of 7.89$\pm$0.14\,km/s.
We used all these works to compare our radial velocity measurements, as reported in Tab.\ref{Tab3}.\\
As for spectroscopic metallicity, the first study was conducted by Thogersen et al. (\citeyear{1993thor}), who analyzed medium-resolution spectra of six red clump stars, confirming the previous estimates of the radial velocity and deriving a metallicity of -0.06$\pm$0.13\,dex. Among the most recent studies on metallicity of NGC\,6940,
Blanco-Cuaresma et al. (\citeyear{2015bla}) and B$\mathrm{\ddot{o}}$cek-Topcu et al. (\citeyear{2016boc}) conducted a very detailed high-resolution spectroscopic study, obtaining values of the metallicity very similar, of respectively +0.04$\pm$0.09\,dex (1 star) and +0.04$\pm$0.02\,dex (12 stars). Reddy et al. (\citeyear{2016red}) confirmed their measurements for three stars. We decided to observe NGC\,6940 since
at the time of our proposal preparation only Reddy et al.(\citeyear{2016red}) and
Blanco-Cuaresma et al. (\citeyear{2015bla}) studies were available.\\
\begin{table}
\tabcolsep 0.15truecm
\caption{Reddening, apparent distance modulus and typical red clump stellar masses used to derived the atmospheric parameters. The values are taken from WEBDA database.}
\label{Tab2}
\centering
\begin{tabular}{lccccc}
\hline\hline
Cluster & l & b & E(B-V) & ${(m-M)_{\rm V}}$ & m$_{\star}$ \\
&\footnotesize{($^{\circ}$)}& \footnotesize{($^{\circ}$)} & & & \footnotesize{(M$_{\odot}$)}\\
\hline
NGC\,6940 & 69.860 & -7.147 & 0.21 & 10.10 & 2.05\\
To\,5 & 143.942 & 3.573 & 0.80 & 13.70 & 3.50\\
\hline\hline
\end{tabular}
\end{table}
\begin{figure*}
\centering
\subfigure[Spectra of NGC\,6940 stars.]{{\includegraphics[scale=0.25]{offN.pdf}}}%
\qquad
\subfigure[Spectra of To\,5 stars.]{{\includegraphics[scale=0.25]{offT.pdf}}}%
\caption{An excerpt of the individual spectra in the indicated wavelength range. Spectra are arbitrarily shifted in the vertical axis, for illustration purposes.}
\label{Fig2}
\end{figure*}
\subsection{Tombaugh\,5}
Unlike NGC\,6940, To\,5 is a poorly known star cluster, and its fundamental parameters are still unsettled.
It was discovered by Tombaugh (\citeyear{1941tom}), who provided the first estimate of the angular diameter of 17\,arcmin.\\
Only photometric analyses are available for it.
There is a general agreement on the cluster age ( $\sim$ 200-250\,Myr )
and color excess E(B-V) (0.80\,mag, Lata et al.(\citeyear{2004lat}), Maciejewski \& Niedzielski (\citeyear{2007mac}), Majaess, Turner \& Lane (\citeyear{2008maj}), Zdanavicius et al.(\citeyear{2011zda})).
On the other hand, distance estimates crowd around two marginally discrepant values.
According to Lata et al.(\citeyear{2004lat}) BV study, the apparent distance modulus is around 13.7\,mag (which implies a heliocentric distance of 1.7\,kpc). This estimate was later confirmed by Zdanavicius et al.(\citeyear{2011zda}), who re-estimated the basic cluster parameters using photometric data in the Vilnius system.\\
Maciejewski \& Niedzielski (\citeyear{2007mac}) argue that Lata et al.(\citeyear{2004lat}) CMD was highly contaminated by field stars. By selecting the cluster putative members in a statistical way, they obtained an apparent distance modulus of 13.10\,mag (which corresponds to a value of d$\sim$1.3\,kpc) using BV photometry as well. This value was later confirmed by Kharchenko et al. (\citeyear{2013kha}) via an
automated fit of 2MASS photometry.
A more careful analysis of the 2MASS CMD from Majaess, Turner \& Lane (\citeyear{2008maj}) yields a distance of 1.7\,kpc.\\
No spectroscopic study is available to our knowledge. An estimate of the cluster metallicity, together with the recently released \textit{Gaia} DR2\footnote{This work has made use of data from the European Space Agency (ESA) mission
{\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia}
Data Processing and Analysis Consortium (DPAC,
\url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC
has been provided by national institutions, in particular the institutions
participating in the {\it Gaia} Multilateral Agreement.} (Gaia Collaboration et al. \citeyear{2016gaia}; Gaia Collaboration et al. \citeyear{2018gaia}) astrometric and photometric data, should help to alleviate the aforementioned distance discrepancy.
\begin{table*}
\centering
\tabcolsep 0.55truecm
\caption{Radial velocity measurements for the program stars in NGC\,6940. Literature values are from Geyer \& Nelles (\citeyear{1985gey}, GN85), Geisler \citeyear{1988geis} (G88), Mermilliod \& Mayor \citeyear{1989merm} (M89), Thogersen et al. \citeyear{1993thor} (T93), Mermilliod et al. \citeyear{2008merm} (MMU08) and B$\mathrm{\ddot{o}}$cek Topcu et al. \citeyear{2016boc} (BT16).}
\label{Tab3}
\begin{tabular}{lccccccc}
\hline
\hline
ID & $V_r$ & $V_{R,GN85}$ & $V_{R,G88}$ & $V_{R,M89}$ & $V_{R,T93}$ & $V_{R,MMU08}$ & $V_{R,BT16}$ \\
\hline
& \footnotesize{(km/s)} & \footnotesize{(km/s)}&\footnotesize{(km/s)}&\footnotesize{(km/s)}&\footnotesize{(km/s)}&\footnotesize{(km/s)}&\footnotesize{(km/s)} \\
\hline
28 & 6.47$\pm$1.90 & - & 3.2$\pm$3.2 & 8.12$\pm$0.28 & - & 7.99$\pm$0.16 & 8.90$\pm$0.08 \\
60 & 8.41$\pm$1.30 &29$\pm$2.1 & - & 7.24$\pm$0.28 & - & 7.27$\pm$0.18 & 7.66$\pm$0.08\\
67 & 8.38$\pm$0.47 & - & 10.6$\pm$3.5 & 7.67$\pm$0.21 & 7 & 7.69$\pm$0.19 & -\\
108 & 6.91$\pm$0.21 & - & 3.8$\pm$2.2 & 6.99$\pm$0.19 & - & 6.76$\pm$0.13 & 7.39$\pm$0.09\\
130 & 6.49$\pm$1.30 & - & - & 7.33$\pm$0.20 & - & 7.25$\pm$0.18 & - \\
132 & 8.87$\pm$0.19 &- & - & 7.27$\pm$0.22 & - & 7.17$\pm$0.14 & 7.76$\pm$0.15\\
139 & 6.57$\pm$1.20 &- & - & 7.42$\pm$0.26 & 10 & 7.12$\pm$0.16 & 7.53$\pm$0.08\\
\hline\hline
\end{tabular}
\end{table*}
\section{Observation and Data Reduction}
We observed NGC\,6940 and To\,5 stars with the Main Stellar Spectrograph (MSS)\footnote{www.sao.ru/hq/lizm/mss/en/index.html} of the 6 meter telescope BTA (Big Telescope Alt-azimuth) at the Special Astrophysical Observatory in Nizhny Arkhyz, Russia. This instrument is installed in the Nasmyth-2 focus of the telescope. It is essentially a long-slit spectrograph, which in its standard mode is equipped with a differential circular polarization analyser, and combined with a double image slicer designed by Chountonov (\citeyear{2004choun}) and a rotating l/4 phase plate, to study mainly stellar magnetic fields.\
For the purpose of our observations, the instrument was set-up in its basic long slit mode. In this set-up, the MSS instrument allows to cover the wavelength range 5840$-$6690\AA~ with a resolution $\lambda / \Delta \lambda$ of $\sim$ 13000 .\\
Observations were carried out on the nights of September 7 and 8, 2017. The nights were clear and stable, with typical seeing in the range 1.5$-$3.0 arcsecs.
Beside the spectra of scientific targets, we collected bias and dome flat frames (two sets of 10 images obtained at the beginning and at the end of the night) and a set of ThAr spectra, for wavelength calibration purposes.\\
\noindent
Spectra were reduced in a standard way, which consists of the following steps: 1) construction of the master bias and subtraction of it from the scientific and calibration frames, 2) correction of data for stray light, 3) searching for the location of individual slices in 2D images, 4) extraction of 1D spectra, 5) continuum normalisation of the spectra, 6) correction of wavelengths for the Earth's motion. All these procedures were carried out within ESO-MIDAS package and its Zeeman extension. Spectra were normalized by the continuum level using the task \textit{continuum} within the \texttt{IRAF} package\footnote{\texttt{IRAF} is distributed by the National
Optical Astronomy Observatory, which is operated by the Association of
Universities for Research in Astronomy, Inc., under cooperative agreement
with the National Science Foundation}. Wavelength solution was defined using the arc spectrum obtained closest in time to the scientific exposures.\\
An excerpt of the stars spectra is shown in Fig.\ref{Fig2}.\\
\section{Radial velocities}
Radial velocities were measured using the task \texttt{fxcor} of \texttt{IRAF}, which performs a Fourier cross-correlation between the observed spectra and a synthetic one. We chose a synthetic spectrum of a typical giant star with solar metallicity and more or less the same resolution (T$\mathrm{_{eff}}$=4750 K, log(g)=2.25 dex, $\xi$=2 km/s and R$\sim$12000) (Munari et al. \citeyear{2005mun}). The typical errors given by the task is of the order of 1.0 km/s.\\
In Tab.\ref{Tab3} and \ref{Tab4} the velocity measurements are reported for NGC\,6940 and To\,5, respectively. In the tables we reported also the values found in literature for a comparison. For all the other stars, our measurements are all in good agreement with the previous studies.\\
In the case of To\,5 stars, there are no previous sprectroscopic studies to compare our measurements with. However, two of them (\#5274 and \#8099) seem to be field stars. According to Zdanavi$\mathrm{\check{c}}$ius et al.(\citeyear{2011zda}), in the cluster region there are seven possible red giant candidates: five of them are in common with the ones we observed, except for \#5274 and \#8099. So we decided to exclude them from the chemical abundances analysis.\\
The mean radial velocities for the two clusters are V$\mathrm{_{r,NGC6940}}$=8.0$\pm$0.2 km/s, in good agreement with the values found in literature, and V$\mathrm{_{r,To5}}$=-22.8$\pm$0.4 km/s.\\
\begin{table}
\tabcolsep 0.65truecm
\caption{Radial velocity measurements for the program stars in To\,5.}
\label{Tab4}
\centering
\begin{tabular}{lc}
\hline\hline
ID & $V_r$ \\
\hline
& \footnotesize{(km/s)}\\
\hline
0006 & -22.9$\pm$0.7\\
5274 & -16.0$\pm$0.7\\
5521 & -23.3$\pm$1.3\\
7701 & -23.3$\pm$0.6\\
7834 & -20.9$\pm$1.1\\
8080 & -21.2$\pm$1.8\\
8099 & -18.7$\pm$2.5\\
\hline\hline
\end{tabular}
\end{table}
\begin{table*}
\tabcolsep 0.4truecm
\caption{Atmospheric parameters for member stars in the two clusters. The first five columns report the values estimated from photometric data, while the last three are the values derived from our spectroscopic investigation.}
\label{Tab5}
\centering
\begin{tabular}{lcccccccc}
\hline\hline
ID & T(B-V) & T(V-K) & $\mathrm{T_{eff,phot}}$ & $\mathrm{log(g)_{phot}}$ & $\xi\mathrm{_{phot}}$ & $\mathrm{T_{eff,spec}}$& $\mathrm{log(g)_{spec}}$ & $\xi\mathrm{_{spec}}$\\
& \footnotesize{(K)} & \footnotesize{(K)} & \footnotesize{(K)} & \footnotesize{(dex)} & \footnotesize{(km/s)} & \footnotesize{(K)} & \footnotesize{(dex)} & \footnotesize{(km/s)}\\
\hline
\textbf{NGC\,6940} &\\
28 & 4996 & 5068 & 5032 & 2.81 & 1.32 & 4980 & 2.91 & 1.32\\
60 & 5017 & 5096 & 5057 & 2.83 & 1.31 & 5057 & 2.85 & 1.31\\
67 & 4975 & 5034 & 5005 & 2.65 & 1.37 & 5005 & 2.55 & 1.37\\
108 & 5171 & 5225 & 5198 & 2.75 & 1.33 & 5100 & 2.70 & 1.33\\
130 & 5104 & 5175 & 5140 & 2.80 & 1.32 & 5140 & 2.80 & 1.32\\
132 & 5017 & 5127 & 5072 & 2.59 & 1.39 & 5000 & 2.70 & 1.39\\
139 & 5082 & 5153 & 5118 & 2.77 & 1.39 & 4980 & 2.95 & 1.39\\
\hline
\textbf{To\,5} &\\
0006 & 4979 & 5068 & 5024 & 2.36 & 1.46 & 5024 & 3.00 & 1.46\\
5521 & 5095 & 5398 & 5247 & 2.43 & 1.44 & 5150 & 3.40 & 1.44\\
7701 & 4558 & 4863 & 4711 & 2.05 & 1.56 & 4710 & 2.10 & 1.56\\
7834 & 5266 & 5363 & 5315 & 2.72 & 1.35 & 5215 & 2.80 & 1.35\\
8080 & 5058 & 5380 & 5219 & 2.61 & 1.38 & 5220 & 3.40 & 1.38\\
\hline\hline
\end{tabular}
\end{table*}
\section{Atmospheric parameters}
Atmospheric parameters were obtained as follows.\\
First, T$_{\mathrm{eff,phot}}$ estimates were derived from the B-V and V-K colours using the relations by Alonso et al. (\citeyear{1999alo}). In the chemical abundances analysis we used the mean temperature between T(B-V) and T(V-K) as input.\\
Surface gravities (log(g)) were obtained from the canonical equation:
\begin{center}
\begin{equation*}
\begin{split}
log(g)=& log(g)_{\odot}+log\left(\frac{m_{\star}}{m_{\odot}}\right)+4\cdot log\left(\frac{T_{\rm eff}}{T_{\rm eff,\odot}}\right)+\\
&+0.4\cdot(M_V+BC_V-M_{BC,\odot})
\end{split}
\end{equation*}
\end{center}
where T$_{\rm eff,\odot}$=5777\,K, log(g)$_{\odot}$=4.44\,dex, M$_{BC,\odot}$=4.74.\\
The bolometric correction (BC$_V$) for each star was derived by adopting the relations
from Alonso et al. (\citeyear{1999alo}).
We adopted initial values for reddening E(B-V) and apparent distance modulus ${(m-M)_{\rm V}}$, from WEBDA, and inferred stellar masses (m$_{\star}$) from the Padova suite of isochrones (Marigo et al. \citeyear{2017mar}). The values used for each cluster are reported in Tab. \ref{Tab2}.\\
Micro-turbulent velocity ($\xi$) was obtained from the
relation by Gratton et al. (\citeyear{1996grat}):\\
\begin{center}
{\rm $\xi$= 2.22 - 0.322$\cdot$log(g)}
\end{center}
\noindent
The input metallicity needed to obtain T$_{\rm {eff,phot}}$ from colors and for the
isochrone fitting was assumed to be solar, which was later confirmed by the spectroscopic analysis (see below). The derived atmospheric parameters are listed in Tab.\ref{Tab5}.\\
\section{Elemental abundances}
\begin{table*}
\tabcolsep 0.65truecm
\caption{Derived abundances ratios for single stars in NGC\,6940 and To\,5. $\sigma_1$ and $\sigma_2$ are respectively the errors estimated from the EW measurements and from the dependences of the atmospheric parameters.}
\label{other}
\centering
\begin{tabular}{lcccccccc}
\hline\hline
ID & [Si/Fe]$\pm\,\sigma_1$ $\pm\,\sigma_2$ & [Ca/Fe]$\pm\,\sigma_1$ $\pm\,\sigma_2$ & [Ti/Fe]$\pm\,\sigma_1$ $\pm\,\sigma_2$ & [Ni/Fe]$\pm\,\sigma_1$ $\pm\,\sigma_2$\\
\hline
\textbf{NGC\,6940} & \\
28 & 0.18$\pm$0.08 $\pm$0.17 & 0.09$\pm$0.35 $\pm$0.04 & -0.13$\pm$0.24 $\pm$0.11 & -0.01$\pm$0.17 $\pm$0.05\\
60 & -0.02$\pm$0.05 $\pm$0.15& -0.07$\pm$0.34$\pm$0.03& -0.11$\pm$0.19$\pm$0.10& -0.2$\pm$0.18$\pm$0.14\\
67 & 0.07$\pm$0.04$\pm$0.18& 0.19$\pm$0.09$\pm$0.12& 0.03$\pm$0.13$\pm$0.09& -0.1$\pm$0.09$\pm$0.05\\
108 & 0.14$\pm$0.07$\pm$0.16 &0.07$\pm$0.12$\pm$0.01& 0.04$\pm$0.14$\pm$0.10& 0.14$\pm$0.11$\pm$0.05\\
130 &0.13$\pm$0.06$\pm$0.14 &0.09$\pm$0.1$\pm$0.02& 0.04$\pm$0.13$\pm$0.09& 0.23$\pm$0.09$\pm$0.03\\
132 &0.2$\pm$0.05$\pm$0.16& 0.08$\pm$0.04$\pm$0.02 & 0.06$\pm$0.07$\pm$0.11& 0.04$\pm$0.06$\pm$0.04\\
139 &0.12$\pm$0.14$\pm$0.16& -0.15$\pm$0.18$\pm$0.04 &-0.22$\pm$0.1$\pm$0.14 & -0.03$\pm$0.07$\pm$0.05\\
\hline
\textbf{To\,5}&\\
0006 & 0.13 $\pm$0.13 $\pm$0.15 & 0.04 $\pm$0.03 $\pm$0.05 & -0.02 $\pm$0.09 $\pm$0.12 & 0.05 $\pm$0.14 $\pm$0.04\\
5521 & 0.10 $\pm$0.03 $\pm$0.15 & 0.16 $\pm$0.17 $\pm$0.04 & 0.13 $\pm$0.05 $\pm$0.11 & -0.03 $\pm$0.05 $\pm$0.03\\
7701 & 0.04 $\pm$0.07 $\pm$0.18 & -0.07 $\pm$0.02 $\pm$0.08 & -0.09 $\pm$0.12 $\pm$0.11 & 0.08 $\pm$0.12 $\pm$0.09\\
7834 & -0.17 $\pm$0.04 $\pm$0.19 & 0.12 $\pm$0.11 $\pm$0.03 & 0.10 $\pm$0.16 $\pm$0.06 & 0.01 $\pm$0.19 $\pm$0.05\\
8080 & 0.00 $\pm$0.10 $\pm$0.15 & -0.09 $\pm$0.12 $\pm$0.07 & 0.04 $\pm$0.18 $\pm$0.08 & -0.04 $\pm$0.04 $\pm$0.03\\
\hline\hline
\end{tabular}
\end{table*}
\begin{figure*}
\centering
\begin{minipage}[c]{19cm}
\centering
\includegraphics[scale=0.24]{Conf3.pdf}\hfil
\includegraphics[scale=0.24]{Conf2.pdf}
\caption{Comparison between the atmospheric parameters (y-axis) and the values found by B$\mathrm{\ddot{o}}$cek Topcu et al. (\citeyear{2016boc}) (BT16, in x-axis). On the right panel we show the comparison between the other elemental abundances derived in this work (y-axis) and the same authors. A part for the micro-turbolence velocities, for which we considered valid the photometric values, the measurements are in agreement.}
\label{Fig3}
\end{minipage}
\end{figure*}
\begin{table}
\tabcolsep 0.25truecm
\caption{Derived metallicities for single stars in NGC\,6940 and To\,5. $\sigma_1$ and $\sigma_2$ are respectively the errors estimated from the EW measurements and from the dependences of the atmospheric parameters.}
\label{metN}
\centering
\begin{tabular}{lc}
\hline\hline
ID & [Fe/H] $\pm \sigma_1$ $\pm \sigma_2$ \\
\hline
\textbf{NGC\,6940}&\\
28 & -0.02 $\pm$0.04 $\pm$0.17\\
60 & 0.11 $\pm$0.05 $\pm$0.16\\
67 & 0.15 $\pm$0.03 $\pm$0.2\\
108 & 0.05 $\pm$0.04 $\pm$0.18\\
130 & 0.04 $\pm$0.04 $\pm$ 0.16\\
132 & 0.18 $\pm$0.03 $\pm$0.18\\
139 & 0.10 $\pm$0.06 $\pm$ 0.16\\
\hline
\textbf{To\,5}&\\
0006 & 0.11 $\pm$0.04 $\pm$0.16\\
5521 & -0.05 $\pm$0.04 $\pm$0.16\\
7701 & 0.21 $\pm$0.05 $\pm$0.2\\
7834 & -0.04 $\pm$0.06 $\pm$0.2\\
8080 & 0.06 $\pm$0.11 $\pm$0.18\\
\hline\hline
\end{tabular}
\end{table}
To derive the spectroscopic atmospheric parameters and chemical abundances of the stars confirmed members, we used the local thermodynamic equilibrium (LTE) line analysis and synthetic spectrum code \texttt{MOOG} (Sneden \citeyear{sne73}, Soebeck et al \citeyear{sob11} \footnote{https://www.as.utexas.edu/~chris/moog.html}.
We used 1D model atmospheres linearly interpolated from the ATLAS9 model atmosphere grid of Castelli \& Kurucz (\citeyear{2003cast}).\\
Si\,I, Ca\,I, Ti\,I, Fe\,I, Fe\,II and Ni\,I abundances were estimated using the equivalent width (EQW) method by running the \textit{abfind} driver of \texttt{MOOG}. EQWs for a selected line list were measured manually using the task \texttt{splot} in \texttt{IRAF}. The line list comes from D'Orazi et al. \citeyear{2017dor}. We kept only lines which are isolated and unblended, fitting a gaussian profile and/or by direct integration, depending on the feature under consideration. We considered only lines with EQWs$<$150\,m\AA. \\
\noindent
The final temperatures were derived by removing the trend between the iron abundances (log(Fe\,I)) of the single lines and the excitational potential ($\chi$). We changed the temperature until the slope of the trend was compatible with a flat trend within 1$\sigma$. In most cases, the input values were close to the final ones, so no changes were applied to the temperatures. As for the final value of the surface gravity, we checked the difference between log(Fe\,I) and log(Fe\,II) and changed it until it reached a value below 0.05 dex. As for micro-turbolence velocities, we did not change the input value since the numbers of Fe\,I lines were not sufficient to calculate a meaningful linear trend between the abundances and the reduced equivalent widths (REWs). The final metallicity were computed with respect to the solar value and reported in Tab.\ref{metN}.\\
We calculated also the abundances of all the others elements and then the abundance ratios with respect to Fe\,I. The final values are reported in Tab.\ref{other}.\\
\noindent
In order to estimate the uncertainties of the metallicities and abundance ratios, we considered mainly errors due to EQWs measurements and to internal uncertainties of the stellar parameters, defined as $\sigma_1$ and $\sigma_2$ in Tab.\ref{metN}.\\
The first source of error is well represented by the standard deviation from the mean abundances considering all the lines divided by the root square of the number of lines. The $\sigma_1$ values derived for [X/Fe] (reported in Tab.\ref{other}) were calculated by quadratically adding the $\sigma_1$ value of [Fe/H] and the one of [X/H].\\
The $\sigma_2$ values were estimated by varying T$_{eff}$ by 150\,K, log(g) by 0.2\,dex and $\xi$ by 0.3\,dex. For each new model, we calculated the difference between the new [X/Fe] and that obtained with the final parameters. Then we added all the contributions quadratically.\\
\begin{figure}
\centering
\includegraphics[width=.45\textwidth]{grad.pdf}
\caption{Galactic radial metallicity gradient from Netopil et al. (\citeyear{2016neto}), considering only open clusters with spectroscopic determination of the metallicity (red dots stand for low quality spectroscopic data, while blue ones for high quality spectroscopic data.). Starred symbols indicate the location of our two clusters.}
\label{Fig5}
\end{figure}
\section{Results of the chemical analysis}
The final values of the atmospheric parameters and the individual abundances are reported in Tabs. \ref{Tab5} and \ref{other}. In order to evaluate the validity of our method, we checked the correlation between the derived metallicities and the atmospheric parameters. Overall we could not find any evident trend.\\
In the case of NGC\,6940, we could compare our results with the values derived by B$\mathrm{\ddot{o}}$cek Topcu et al.(\citeyear{2016boc}) for five stars in common. The mean values of the cluster metallicities and abundances ratios are reported in Tab.\ref{Tab9}, in which we also reported the literature values for a comparison.\\
\begin{table*}
\tabcolsep 0.3truecm
\caption{Comparison of mean metallicity and abundances ratios derived in this work with the literature.}
\label{Tab9}
\centering
\begin{tabular}{lcccccccc}
\hline\hline
ID &$<[Fe/H]>$& $<[Si/Fe]>$& $<[Ca/Fe]>$ &$<[Ti/Fe]>$& $<[Ni/Fe]>$\\
\hline
\textbf{NGC\,6940} &\textbf{0.09$\pm$0.06} & \textbf{0.11$\pm$0.07}& \textbf{0.04$\pm$0.11}& \textbf{-0.04$\pm$0.10}& \textbf{0.01$\pm$0.13}\\
Blanco-Cuaresma et al.(2015)& 0.04$\pm$0.09& 0.13$\pm$0.13 & 0.03$\pm$0.09 & -0.01$\pm$ 0.13& -0.05$\pm$0.10\\
Bocek-Topcu (2016) &0.04$\pm$0.02& 0.12$\pm$0.06 &0.08$\pm$0.04 &-0.04$\pm$0.04& -0.01$\pm$0.02\\
Reddy(2016) &-0.06$\pm$0.03& 0.14$\pm$0.03& 0.00$\pm$0.03& -0.04$\pm$0.02 &0.00$\pm$0.03\\
\hline
\textbf{To\,5}& \textbf{0.06$\pm$0.11} & \textbf{0.02$\pm$0.11}& \textbf{0.03$\pm$0.10}& \textbf{0.03$\pm$0.08}& \textbf{0.01$\pm$0.05}\\
\hline\hline
\end{tabular}
\end{table*}
\subsection{NGC\,6940}
In Fig.\ref{Fig3} (left panel) we show the comparison between the atmospheric parameters derived in this work (y-axis) and the values found by B$\mathrm{\ddot{o}}$cek Topcu et al.(\citeyear{2016boc}) (BT16, in x-axis). The agreement is quite good: each measurement is compatible with the authors', except for the micro-turbulence velocities. We kept as valid the photometric estimates of these values, since we had too few lines to calculate the trend between the individual abundances and the reduced equivalent widths.\\
In the right panel of Fig.\ref{Fig3}, instead, we compared the abundance ratios. Some values we found are compatible with B$\mathrm{\ddot{o}}$cek Topcu et al.(\citeyear{2016boc}), but the dispersion of our measurements is larger. This can be explained with the lower resolution of our spectra and the limited spectral coverage. Moreover, most of the lines do not have a pure gaussian profile, so we measured the EQWs with other methods.\\
The mean value of metallicity and abundance ratios, as reported in Tab.\ref{Tab9}, are in very good agreement with the values found in literature. NGC\,6940 has a nearly solar metallicity of +0.09$\pm$0.06\,dex, consistent with the values expected for a cluster at that Galacto-centric distance, as Fig.\ref{Fig5} shows. he Galacto-centric distance has bee computed adopting $d_{\odot}$ =8.5 kpc, for homogeneity with previous studies and for the sake of comparison.\\
\subsection{To\,5}
Basing on previous estimates of the distance, this cluster appears to be quite interesting, since it would be located close to the critical Galacto-centric distance at which the radial abundance gradient for [Fe/H] exhibits a
significant change of slope.
For this cluster we provided the very first spectroscopic estimate of the photospheric parameters for all the sample stars but for $\#$8080, for which we had to rely on the mean photometric estimates from the other sample stars.
We derived for To\,5 a roughly solar metallicity of +0.06$\pm$0.11\,dex, which lies close to the upper envelope of the thin disk abundance gradient, but is still compatible with
the other clusters located in the same Galacto-centric distance range, as shown in Fig.\ref{Fig5}.
\section{Cluster basic parameters revisited}
\begin{figure*}
\centering
\includegraphics[scale=0.35]{vector.pdf}
\caption{Vector point diagrams of NGC\,6940 and To\,5: the black dots represent the data taken from Gaia DR2, while the red ones represent our conservative selection of cluster members. }
\label{Fig6}
\end{figure*}
\begin{table*}
\tabcolsep 0.3truecm
\caption{Proper motion components, parallaxes and distances derived from \textit{Gaia} DR2, and corresponding errors, for the stars in common with this study.}
\label{Tab10}
\centering
\begin{tabular}{lcccccccc}
\hline\hline
ID &$\mu_{\alpha} cos\delta \pm \sigma_{\alpha}$& $\mu_{\delta} \pm \sigma_{\delta}$& $\varpi \pm \sigma_{\varpi}$ &$d \pm \sigma_d$\\
\hline
\textbf{NGC\,6940} \\
67 & -1.758$\pm$0.061 & -9.514$\pm$0.046& 1.026$\pm$0.038& 975$\pm$ 36 \\
108 & -1.534$\pm$ 0.046 & -9.235$\pm$ 0.041 & 0.883$\pm$ 0.033 & 1132$\pm$ 42\\
130 & -2.178$\pm$ 0.049 & -9.182$\pm$ 0.047 & 0.934$\pm$ 0.035 & 1071$\pm$ 4\\
132 & -2.015$\pm$ 0.051 & -9.315$\pm$ 0.047 & 1.005$\pm$ 0.039 & 995$\pm$ 38\\
139 & -2.046$\pm$ 0.045 & -9.523$\pm$ 0.046 & 0.975$\pm$ 0.031 & 1026$\pm$ 33\\
\hline
\textbf{To\,5} \\
0006 & -7.478$\pm$0.862& 1.528$\pm$0.748& -3.061$\pm$0.497& 327$\pm$55\\
5521 &0.270$\pm$0.046& 4.233$\pm$0.042& 0.999$\pm$0.031& 1000$\pm$31 \\
7701 & 0.383$\pm$ 0.065 & -2.297$\pm$ 0.064 & 0.522$\pm$ 0.039 & 1917$\pm$ 145\\
7834 & -3.368$\pm$0.048& -0.923$\pm$0.048&1.007$\pm$0.031&993$\pm$31\\
8080 & 23.889$\pm$1.727 & -20.752$\pm$1.632& -1.178$\pm$1.155&849$\pm$130\\
\hline\hline
\end{tabular}
\end{table*}
\begin{figure}
\includegraphics[scale=0.3]{fitN6940.pdf}
\caption{Color magnitude diagrams of NGC\,6940 from the BV study of Larsson-Leander (\citeyear{1960lar}). Isochrones are super-imposed for a distance of 1.0 kpc and a reddening of 0.21 mag, and ages of 0.9 (green), 1.0 (red), and 1.1 (green) Gyr. See text for additional details.}
\label{Fig7}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{fitTo5.pdf}
\caption{Color magnitude diagrams of To\,5 from the BV study of Maciejewski \& Niedzielski (\citeyear{2007mac})). Isochrones are super-imposed for a distance of 1.7 kpc, a reddening of 0.80 mag, and ages of 200 (green) , 250 (red), and 300 (blue) Myr. See text for additional details. }
\label{Fig8}
\end{figure}
In order to estimate the cluster basic parameters, we complement available photometric and spectroscopic data with the recent \textit{Gaia} DR2 data.
As in Fig. \ref{Fig6}, we plotted proper motion $\mu_{\alpha}cos(\delta)$ (mas/yr) vs $\mu_{\delta}$ (mas/yr) and selected visually only the stars within the cluster radius. Then we calculated the distance from the parallax measured by \textit{Gaia} only for the proper-motion-selected stars.\\
As one can appreciate from Fig. \ref{Fig6}, the distributions have a gaussian profile around the mean value. To get a distance estimate we selected all the stars within 3$\sigma$ from the mean to include the largest number of candidate members. We note,
however, that two NGC~6940 stars observed by us are not included \textit{Gaia} DR2 (namely \#28 and \#60), while the majority of the stars in the red clump of To\,5 are missing, either because their distances are not within 3$\sigma$ from the mean (like \#0006) or because they have peculiar proper motion (like \#8080). All of them, however, are confirmed members according to radial velocities and metallicity.
The values of proper motion components, parallaxes, derived distances and the mean cluster values are reported in Tab.\ref{Tab10}.\\
A few recent studies have noted that the zero-point of Gaia DR2 parallaxes should probably be incremented by 0.05-0.08 mas (Stassun \& Torres. \citeyear{sta18} , Zinn et al. \citeyear{zin18} . We will investigate this possibility for To5, for which distance is not yet well-constrained.\\
\noindent
To provide an updated estimate of the clusters age, we super-imposed isochrones for different ages on top the CMD built with optical BV photometry as in Fig.~1. The isochrone fitting is performed
fixing the metallicity to the new value found in this work and the reddening to the value found in literature (see Fig.\ref{Fig7} and \ref{Fig8} for NGC~6940 and To\,5, respectively). \\
\noindent
In the case of NGC\,6940 the fit
yields an age of 1.0$\pm$0.1\,Gyr (for a metallicity Z=0.0187) in nice agreement with B$\mathrm{\ddot{o}}$cek Topcu et al. (\citeyear{2016boc})) adopting the \textit{Gaia} distance estimate of 1.03$\pm$0.17\,kpc , which is slightly higher than the values found in literature and reported in WEBDA. This corresponds to an apparent distance modulus of (m-M)$_V$=10.72$\pm$0.36.\\
\noindent
The distance estimate of To\,5 reported in WEBDA database is confirmed by \textit{Gaia} DR2, from which we derive an estimate of 1.7$\pm$0.1\,kpc. If we apply an offset of 0.05 mas, the distance becomes 1.68, well within the error bars.
This was calculated using MS proper-motion-selected stars brighter than 16.0 in the \textit{Gaia} passband $G_{bp}$, since for our sample of clump stars \textit{Gaia} provides extremely different proper motion components and parallaxes. We do not have a clear explanation for this, since these stars have the same radial velocity and metallicity. Since To\,5 field is severely contaminated by field stars, one possibility is that there are mismatches in \textit{Gaia} DR2, which will probably be solved with future releases of the catalog. Another possibility is that these 5 stars have similar properties by chance, and they do not have anything to do with the cluster. We would exclude this scenario since their position in the cluster CMD is compatible with the expected locations of red, evolved stars in their He-burning phase. To better clarify this issue,
we run a few simulations with the Trilegal code (Girardi \citeyear{leo16}), and found that for a single stellar populations (SSP) of the age nad estimated mass ($\sim 350 M_{\odot}$) of To5, the expected number of evolved stars is 4$\pm$3, well in agreement with available photometric data (see Fig~1).\\
\noindent
Using this distance, our new estimate of metallicity (which corresponds to Z=0.0175), and the reddening estimate provided by Majaess et al. (\citeyear{2008maj}) we derived a new age estimate using the isochrone fitting method, as shown in Fig. \ref{Fig8},
where a CMD based on Maciejewski \& Niedzielski (\citeyear{2007mac}) has been plotted.
We find that an age of 250$\pm$50 Myr nicely fits the overall stars' distribution, thus supporting earlier investigations.
The Gaia DR2 distance of 1.7 kpc places To\,5 in the close vicinity of the Perseus arm, as defined by the position of star forming regions
in this Galactic sector derived from maser distances as in Reid et al. (\citeyear{reid09}).
\section{Summary and Conclusions}
In this work we presented a spectroscopic study of two open clusters: NGC\,6940 and To\,5. While the former was studied before, for the latter these are the first spectroscopic measurements to-date. Our findings can be summarised as follows:
\begin{itemize}
\item We observed seven candidate red clump stars in the region of the two clusters. All the observed stars in NGC\,6940 turn out to be cluster members. The cluster mean value is V$\mathrm{_{NGC\,6940}}$=8.0$\pm$0.2 km/s. In the case of To\,5, we could confirm the membership of five of the observed stars and the mean radial velocity is V$\mathrm{_{r,To5}}$=-22.8$\pm$0.4 km/s.
\item We then derived abundances for the following elements: FeI, SiI, CaI, TiI and NiI. Overall, the two clusters exhibit solar composition, with mean metallicity [Fe/H]$\mathrm{_{NGC\,6940}}$=+0.09$\pm$0.06\,dex, and [Fe/H]$\mathrm{_{To\,5}}$=+0.06$\pm$0.11\,dex. The estimate we found for NGC\,6940 is in good agreement with the most recent literature. On the other hand, the metallicity measured for To\,5 confirms the large spread (almost 0.5 dex) in [Fe/H] at about10 kpc from the Galactic centre.
\item We used \textit{Gaia} DR2 to select the cluster members based on the analysis of proper motions and to calculate the cluster distances from the corresponding parallaxes. We derived for NGC\,6940 a heliocentric distance of 1.0$\pm$0.2\,kpc. This is somewhat larger than the values found in literature, but provides a generally good fit of the star distribution in the CMD. For To\,5 we found a distance of 1.7$\pm$0.1\,kpc, confirming the values reported in WEBDA, and thus ruling out previous suggestions of a shorter distance of $\sim$1.3 kpc.
\item With our new estimates of metallicities and an updated distances from \textit{Gaia} DR2, we revised the cluster fundamental parameters, in particular the age. Using the isochrone fitting technique we found an age of 1.0$\pm$0.1\,Gyr for NGC\,6940, and 250$\pm$50\,Myr for To\,5.
\end{itemize}
\acknowledgments
G. Carraro thanks the SAO staff, in particular Gennady Valyavin, for the kind hospitality. The authors also thanks Ulisse Munari for enlightening discussions.
We made extensive use of the WEBDA databased, maintained by E. Paunzen at Masaryk University in Brno. We acknowledge the useful input of the anonymous referee.
\vspace{5mm}
\facilities{CAO}
\software{IRAF ( Tody \citeyear{tody93,tody96}), MIDAS (Banse et al. \citeyear{ban83}; Grosbol \& Ponz \citeyear{gro90}) }
|
\section{Introduction}
Single-component molecular conductor, [Pd(dddt)$_2$], which is insulating at ambient pressure but becomes conducting and shows almost temperature independent resistivity under a pressure of 12.6 GPa,~\cite{Kato_JACS} has been studied using first-principles density-functional theory~(DFT) calculations,~\cite{Kohn_Sham} and linearly dispersed bands with Dirac cones were found at 8~GPa.~\cite{Kato_JACS, Tsumuraya2018}
An extended H\"uckel calculation and a tight-binding analysis of the DFT optimized structure at 8 GPa have shown nodal line semimetal.~\cite{Kato2017_JPSJ}.
Relevant physical quantities have been examined to understand
the characteristics of such
new massless Dirac electron systems.~\cite{Suzumura_Pddddt2_JJAP2017, Suzumura_Conduc_2017, Suzumura2018_JPSJ_T, Suzumura_Yamakage_JPSJ}
In fact, the multi-orbital nature due to an interplay between HOMO (highest occupied molecular orbital) and LUMO (lowest unoccupied molecular orbital) on different molecular sites is essential
for the emergence of the Dirac electron.~\cite{Kato_JACS}
Such a state, where the contact points of the Dirac cones form a closed line (loop) called nodal line semimetal,~\cite{Kim_Rappe_NLS_2015, Weng_Line_node, KieranPRL, Fang_NLS, Liu_Balents2017}
is quite different from that of massless Dirac electrons in a two-dimensional molecular conductor.~\cite{Katayama2006_JPSJ75}
In 1937, Herring~\cite{Herring_PR.52.365}
claimed accidental degeneracies in the energy bands
followed by a closed/open circuit (nodal line) for a system with inversion symmetry.
The product of parity eigenvalues at the eight time reversal invariant momentum (TRIM) determines even or odd number of circuits (nodal loop).
This formula is quite similar to Fu-Kane's criteria of $\mathbb{Z}_2$ topological insulator,~\cite{Fu_Kane_PRB.76.045302} despite assuming the absence of spin-orbit coupling (SOC).
However, since it should work well for weak SOC materials with
light elements such as molecular conductors,
the shape of the nodal line and the number of nodal loops would be useful to
evaluate the band topology and design new molecular topological materials.
The nodal line semimetal is a recent topic
related to the new concept associated with the topological property of Dirac electrons.~\cite{Murakami2007,Burkov2011,Hirayama2017,Hirayama2018}
Among them, single nodal loop has been found
in pnictides CaAg$X$ ($X$ = P and As)\cite{Yamakage_CaAgX_Dirac}, alkaline-earth compounds of $A$$X_2$ ($A$ = Ca, Sr, Ba; $X$ = Si, Ge, Sn),~\cite{Liu2016PRB} CaAs$_3$\cite{Diracloop_CaAs3_2017}, and Ag$_2$S~\cite{Ag2S_PRB.96.115106}. In fact, all these systems and [Pd(dddt)$_2$] turn out to be a strong topological insulator\cite{Fu_Kane_Mele2007} when SOC is on.~\cite{Tsumuraya2018, Kato2017_JPSJ, Liu2018}
With the regard to the Dirac electrons in [Pd(dddt)$_2$],
it is useful to describe the nodal line semimetal in terms of the effective Hamiltonian with a two-band model.
For many nodal line semimetals, first-principles calculations have characterized the shape of the
nodal line, and an effective Hamiltonian is analyzed only close to the
$\Gamma$ point due to perturbation.~\cite{Liu2016PRB, Liu2018}
However, such a Hamiltonian is insufficient to describe
the large nodal line away from TRIM due to the non-coplanar~\cite{Kato2017_JPSJ}
and snake-like loop structures.~\cite{Liu2016PRB}
Further, the model is required to show
a reasonable energy band for the Dirac cone along the nodal line
in order to examine the physical property,
since the chemical potential (i.e., the Fermi energy) crosses the nodal line.
Indeed, the model, which provides
two kinds of velocity fields associated with the Dirac cone, is useful as shown by the calculation of the topological property of the Berry phase.
\cite{Suzumura_Yamakage_JPSJ}
In single-band molecular conductors, first-principles band structures near the Fermi level
are well reproduced by a small number of hopping parameters,~\cite{Kino2006_JPSJ, Koretsune_Hotta_PRB2014}
but the corresponding parameters for a multi-band system of [Pd(dddt)$_2$] are huge and complicated.
Therefore, a new method for the direct derivation of the model
from the DFT calculation is needed to describe
the Dirac cone reasonably for all the Dirac points along the loop.
In this letter, on the basis of first-principles DFT calculations,
we calculate a number of Dirac points that form
a single non-coplanar nodal loop centered at the $\Gamma$-point
in the first Brillouin zone (BZ).
The characteristics of the two crossing bands near the Dirac points
and the velocities of the Dirac cone
are examined by focusing on two typical Dirac points on the loop.
Then, we demonstrate a new method to obtain the effective Hamiltonian for
the nodal line semimetal
where the base is not on the TRIM, but on the respective
Dirac point.~\cite{Luttinger1955,Kobayashi2007_JPSJ}
A model with analytical matrix elements
is derived using the knowledge of
the general relation between the nodal line
and the velocity fields of the Dirac cone.\cite{Suzumura_Yamakage_JPSJ}
The crystal structure of [Pd(dddt)$_2$], which is determined by x-ray diffraction at ambient pressure, belongs to a monoclinic system
with the space group of $P2_1$/$a$.\cite{Kato_JACS}
In our previous study, using first-principles calculations, we discovered anisotropic Dirac cones of [Pd(dddt)$_2$] at 8GPa by performing structural optimization under symmetry operations.~\cite{Kato_JACS}.
However, the three-dimensional characteristics of the electronic state and the effective model have not yet been reported from first-principles.
The present first-principles DFT calculations are performed based on an all-electron full-potential linearized plane wave (FLAPW) method as implemented in QMD-FLAPW12.~\cite{Wimmer1981, L_KA, Weinert}
To determine accurately the anisotropy of the Dirac cone, Fermi velocities, and contact points of Dirac cones along the nodal line, we calculate the eigenvalues around the Dirac cone by using the highly precise FLAPW method. We used a high-dense $ \bm{k}$-mesh for plotting the 3D band structures.
The exchange-correlation functional that is used in the present calculation is the generalized gradient approximation proposed by Perdew, Burke, and Ernzerhof.~\cite{GGA_PBE} The $ \bm{k}$-point sampling grid of
4 $\times$ 8 $\times$ 4 was used.
The BZ integration was performed with an improved tetrahedron method.~\cite{Imp_Tetrahedron1994}
Figures \ref{fig:structure}(a) and \ref{fig:structure}(b) show
the band structure along the $\Gamma$-Y and
along X'(-0.1967, 0, 0)-Z'(-0.1967, 0, 0.5) directions, respectively
where the band crossing suggests
two typical Dirac points.
We refer to the former Dirac point as Dirac point (I)
and the later as Dirac point (II).
In general,
the crossings of two linear bands with different characteristics
occurs forming a Dirac cone.
By comparing the characteristics of the HOMO and LUMO for the isolated [Pd(dddt)$_2$] monomer
in Figs.~\ref{fig:structure}(e) and \ref{fig:structure}(f),
we discuss the wavefunction near Dirac point (I).
As plotted in Figs.~\ref{fig:structure}(a)--\ref{fig:structure}(d),
the convex upwards band
predominately consists of HOMO like wavefunctions in Layer 1, and
the convex downwards band is made up of LUMO like wavefunctions in Layer 2.
The shape of the Dirac cone
depends on the position of the Dirac point along the nodal line.
In the following, we use a wavevector scaled by $2\pi$,
lattice constants of the reciprocal lattice taken as unity,
and energy in the unit of eV.
Figures \ref{fig:cone}(a) and \ref{fig:cone}(b) show the Dirac cone with
Dirac point (I), $ \bm{k}_{\rm 0}$ = (0, 0.086, 0), where
the axis of the cone is parallel to
a vector (--2$^{-1/2}$, 0, 2$^{-1/2}$).
The principal axes of the plane, which are perpendicular to
the axis of the Dirac cone,
are given by \textsl{\textbf{a*}}+\textsl{\textbf{c*}} ($k_{x+z}$) and
\textsl{\textbf{b*}} ($k_y$).
Dirac cone (I) shown in Fig.~\ref{fig:cone}(a) is symmetric with respect to $k_{x+z}$ while
that shown in Fig.~\ref{fig:cone}(b) is tilted along the positive $k_y$.
Figures \ref{fig:cone}(c) and (d) show the Dirac cone of Dirac point (II),
$ \bm{k}_{\rm 0}$ = (-0.1967, 0.000, 0.3924),
where the axis of the cone is parallel to \textsl{\textbf{b*}} ($k_y$).
The former shows a bird's eye view with tilting along both
$k_x$ and $k_z$ axes.
The latter denotes the gap function around Dirac point (II),
which is the difference in energy between the valence and conduction bands, $E_c( \bm{k})-E_v( \bm{k})$.
The principal axes of the Dirac cone (II) is rotated counterclockwise
with an angle $\theta \sim \tan^{-1}(0.2)$ and
the tilting does not appear.
The difference in behavior of
the Dirac cone between Dirac points (I) and (II)
is discussed in terms of the effective Hamiltonian.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{17502Fig1.eps}
\caption{(Color online)
Band structure of [Pd(dddt)$_2$] at 8GPa (without SOC) along the $\Gamma$-Y direction (a) and along the X'(-0.1967, 0, 0)-Z'(-0.1967, 0, 0.5) direction (b)
on which Dirac points (I) and (II) are found, respectively.
Wavefunction at a fixed $ \bm{k}$: charge densities $|\Psi_{i, \bm{k}}|^2$ close to Dirac point (I) for the bottom of the conduction bands (c),
and those at the top for the valence bands (d).
The notations \textsl{\textbf{a}}, \textsl{\textbf{b}}, and \textsl{\textbf{c}} denote the respective
lattice vectors of the unit cell.
Contour plot of the orbital (charge) density of the HOMO (e) and LUMO (f)
of isolated [Pd(dddt)$_2$] monomer.
}
\label{fig:structure}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{17502Fig2.eps}
\caption{(Color online) Band structure
close to Dirac point (I) on the $k_{x+z}$-$k_y$ plane,
where (a) and (b) are
the side views with the horizontal axis given by
\textsl{\textbf{b*}} with $0.06~<~k_y~<~0.122$ and
by \textsl{\textbf{a*}}+\textsl{\textbf{c*}} with $-0.2 < k_{x+z} < 0.2$,
respectively.
(c) Bird's eye view of the Dirac cone
with apex at Dirac point (II) on the $k_x$-$k_z$ plane,
where $-0.22 < k_x < -0.15$ and $0.3 < k_z < 0.5$.
(d) 2D contour plots corresponding to (c)
for the difference in energy between the top of the valence bands, $E_v$($ \bm{k}$), and the bottom of conduction bands, $E_c$($ \bm{k}$),
on the \textsl{\textbf{a*}}-\textsl{\textbf{c*}} plane,
where $k_x$ and $k_z$ represent
\textsl{\textbf{a*}} and \textsl{\textbf{c*}}, respectively, which are taken as the orthogonal axes, for simplicity.
}
\label{fig:cone}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8cm]{17502Fig3.eps}
\caption{(Color online)
(a) Nodal line of accidental degeneracies in [Pd(dddt)$_2$] at 8GPa. The loop is formed by the trajectory of the contact points of Dirac cones in the three-dimensional momentum space, $(k_x, k_y, k_z)$ in the absence of spin-orbit coupling.
The locations of the respective points on the line, which
are scaled by $2\pi$
$ \bm{k}_{\rm 0} [= ( k_{\rm 0x}, k_{\rm 0y}, k_{\rm 0z})]$ are given by
(0, 0.086, 0) (1),
(-0.042, 0.083, 0.05) (2),
(-0.078, 0.0755, 0.1) (3),
(-0.109, 0.066, 0.15) (4),
(-0.131, 0.057, 0.2) (5),
(-0.151, 0.047, 0.25) (6),
(-0.168, 0.037, 0.3) (7),
(-0.176, 0.0315, 0.325) (8),
(-0.184, 0.0245, 0.35) (9),
(-0.191, 0.0165, 0.375) (10), and
(-0.1967, 0.000, 0.3924) (11)
where we mainly examine the symmetric points
(1) and (11) defined by (I) and (II) respectively.
(b) The corresponding energy
dispersion of the nodal line, $E( \bm{k})$ [eV]. The origin of the energy is set to be the energy at Dirac point (I).
The energy is the same for Dirac points shown by
$ \bm{k}_{\rm 0}$ = $( k_{\rm 0x}, \pm k_{\rm 0y}, k_{\rm 0z})$ and
$(- k_{\rm 0x}, \pm k_{\rm 0y}, - k_{\rm 0z})$.
}
\label{fig3}
\end{figure}
Figure \ref{fig3}(a)
shows the nodal line (trajectory of Dirac points) of [Pd(dddt)$_2$]
at 8~GPa, which consists of 11 independent Dirac points $ \bm{k}_{\rm 0}$ calculated
using the first-principles calculations.
Note that the present DFT calculations show a
non-coplanar loop within the first BZ, while
the loop of the previous tight-binding model~\cite{Kato2017_JPSJ}
extends over the first BZ.
For the estimation of the effective Hamiltonian, we focus on Dirac points (I) and (II) represented by the numbers 1 and 11, respectively.
The trajectory of the Dirac points suggests an accidental degeneracy,
which occurs without the symmetry of the crystal.
Note that these Dirac points are symmetric with respect to the $\Gamma$-point and the plane
of $k_y=0$~($k_x$--$k_z$ is a mirror plane), respectively.
Using Fig.~\ref{fig3}(a), we examine the effective Hamiltonian
of a two-band model,
where the base is given by the respective Dirac point $ \bm{k}_{\rm 0}$.
Note that this method describes the Dirac cone correctly,
as shown in Refs.~\citen{Luttinger1955} and ~\citen{Kobayashi2007_JPSJ}.
Two energies consisting of
different kinds of parities degenerate on the Dirac point,
where arbitrary linear combination of these two wavefunctions
becomes possible.
The eigenfunctions with $ \bm{k}$, which is just away from the Dirac point
is determined uniquely,
and their linear combination rotates around the Dirac point.\cite{Katayama2009_EPL}
In the present case, we can choose two eigenfunctions of
HOMO and LUMO with different symmetry at the Dirac point.
Now we derive the effective Hamiltonian of the two band model explicitly.
We start with the full 8 $\times$ 8 Hamiltonian that is
represented at the $\Gamma$-point. Such a Hamiltonina
is given by
\begin{eqnarray}
H( \bm{k})
&=&
\begin{pmatrix}
H^{\rm HH}( \bm{k}) & H^{\rm HL}( \bm{k}) \\
H^{\rm LH}( \bm{k}) & H^{\rm LL}( \bm{k})
\end{pmatrix} \ ,
\label{eq:Heff}
\end{eqnarray}
where the 4 $\times$ 4 matrix $H^{\rm HH}( \bm{k})$ ( $H^{\rm LL}( \bm{k})$) consists of only HOMO (LUMO) and $H^{\rm HL}( \bm{k})$ ($H^{\rm HL}( \bm{k})$) denotes
the interaction between HOMO and LUMO.
The former is the even function of $ \bm{k}$ due to the time reversal symmetry,
while the latter is the odd function due to interaction with
different parity.
Among the energy bands of the four HOMOs and four LUMOs,
we choose
two bands given by the top HOMO bands and the bottom LUMO bands,
which are assumed to be isolated from the other bands.
The Scr\"odinger equation for these two bands
$E_{\rm H}( \bm{k}$) and $E_{\rm L}( \bm{k})$ is given by
\begin{eqnarray}
H^0( \bm{k}) |H (L)( \bm{k})> = E_{\rm H (L)}( \bm{k}) |H (L)( \bm{k})> \; ,
\end{eqnarray}
where $H^0({ \bm{k}})$ denotes Eq.~(\ref{eq:Heff})
without $H^{\rm HL}( \bm{k})$ and $H^{\rm LH}( \bm{k})$.
By using $|H ( \bm{k})>$ and $|L ( \bm{k})>$,
the reduced Hamiltonian of the 2 $\times$ 2 matrix is written as
\begin{eqnarray}
H^{\rm red} & =& \sum_{\alpha, \beta = {\rm H}, {\rm L}}
|\alpha( \bm{k})> <\alpha( \bm{k})|
H( \bm{k}) |\beta( \bm{k})> <\beta( \bm{k})| \nonumber \\
& & = \sum_{\alpha, \beta = {\rm H}, {\rm L}} |\alpha( \bm{k})> H_{\rm eff} <\beta( \bm{k})| \; .
\label{eq:eq3}
\end{eqnarray}
By defining
\begin{eqnarray}
f_0 &=& (E_{\rm H}( \bm{k}) + E_{\rm L}( \bm{k})|/2 \; ,
\nonumber \\
f_3 &=& |E_{\rm H}( \bm{k}) - E_{\rm L}( \bm{k})|/2 \; ,
\nonumber \\
f_2 &= & i <{\rm H}( \bm{k})| H |{\rm L}( \bm{k})>\; ,
\label{eq:eq4}
\end{eqnarray}
$H_{\rm eff}( \bm{k})$ in Eq.~(\ref{eq:eq3}) is expressed as
\begin{eqnarray}
{H_{\rm eff}(\bm{k})}
&=&
\begin{pmatrix}
f_0( \bm{k}) + f_3( \bm{k}) & - i f_2( \bm{k}) \\
i f_2 ( \bm{k}) & f_0( \bm{k})- f_3( \bm{k})
\end{pmatrix} \ ,
\label{eq:eq5}
\end{eqnarray}
where $ \bm{k} = (k_x, k_y, k_z)$ and $f_0(0)$ is taken as zero.
The energy is given by
$E_{\pm} =f_0 \pm \sqrt{f_2^2 + f_3^2}$,
where $E_+ (E_-) = E_c (E_v)$.
The Dirac point $ \bm{k}_{\rm 0}$ on the nodal line is obtained from
\begin{subequations}
\begin{eqnarray}
f_2( \bm{k}_{\rm 0}) &=& 0 \; ,
\label{eq:eq6a}
\\
f_3( \bm{k}_{\rm 0}) &=& 0 \; .
\label{eq:eq6b}
\end{eqnarray}
\end{subequations}
Note that $f_0( \bm{k})$ and $f_3( \bm{k})$ are
the even functions of $ \bm{k}$ due to time reversal symmetry.
The quantity $f_2( \bm{k})$, which is also real due
to the presence of inversion symmetry around the Pd atom,~\cite{Kato2017_JPSJ}
is the odd function of $ \bm{k}$ due to
HOMO and LUMO having different parities.
The function $f_2( \bm{k})$ is estimated by projecting the nodal line on the $k_x$ - $k_z$ plane in Fig.~\ref{fig4}(a), while
$f_3( \bm{k})$ is estimated by projecting the nodal line on the $k_x$ - $k_y$ plane in Fig.~\ref{fig4}(b).
This method reproduces the Dirac points well as shown in Fig.~\ref{fig3}(a).
We empirically examined the form under the constraint that
the Dirac points (0,~0.086,~0) (I) and
(-0.1967, 0.000, 0.3924) (II) satisfy Eqs.~(\ref{eq:eq6a})~and (\ref{eq:eq6b})
These functions are obtained as
\begin{subequations}
\begin{eqnarray}
f_2( \bm{k}) & \simeq & C_2(k_z + k_x + 40 k_x^3 - 380 k_x^5) \; ,
\label{eq:eq7a}
\\
f_3( \bm{k}) & \simeq & C_3((k_x/0.1967)^2 + (k_y/0.086)^2
\nonumber \\
& & +(k_xk_y/0.027)^2 - 1 ) \; ,
\label{eq:eq7b}
\\
f_0( \bm{k})& \simeq & b_xk_x^2 + b_yk_y^2 + b_zk_z^2 + C_0\; .
\label{eq:eq7c}
\end{eqnarray}
\end{subequations}
where
the velocity of the cone at the Dirac point
$ \bm{k}_{\rm 0}$ is obtained
as $\bm{v}_2 =\nabla_{ \bm{k}_{\rm 0}}f_2$, $\bm{v}_3 =\nabla_{ \bm{k}_{\rm 0}}f_3$ and
$\bm{v}_0 =\nabla_{ \bm{k}_{\rm 0}}f_0$ corresponding to the tilting.
Using $ \bm{k}_{\rm 0} = ( k_{\rm 0x}, k_{\rm 0y}, k_{\rm 0z})$, $\bm{v}_j$
is calculated as
\begin{subequations}
\begin{eqnarray}
\bm{v}_2 & = & \nabla_{ \bm{k}_{\rm 0}}f_2 \simeq
C_2(1 + 120 k_{\rm 0x}^2 - 2280 k_{\rm 0x}^4, 0, 1),
\label{eq:eq8a} \\
\bm{v}_3 & = & \nabla_{ \bm{k}_{\rm 0}}f_3 \simeq
C_3(2 k_{\rm 0x}/0.196^2+2 k_{\rm 0x} k_{\rm 0y}^2/0.027^2,
\nonumber \\
& &2 k_{\rm 0y}/0.086^2 + 2 k_{\rm 0x}^2 k_{\rm 0y}/0.027^2, 0)
\; , \label{eq:eq8b} \\
\bm{v}_0 & = & \nabla_{ \bm{k}_{\rm 0}}f_0 \simeq (2b_x k_{\rm 0x},2b_y k_{\rm 0y}, 2b_z k_{\rm 0z}) \; .
\label{eq:eq8c}
\end{eqnarray}
\end{subequations}
Note that the tangent of the nodal line is parallel to
$\bm{v}_2 \times \bm{v}_3$.\cite{Suzumura_Yamakage_JPSJ}
In the previous work,\cite{Kobayashi2007_JPSJ}
$\bm{v}_j$ was estimated directly from Eq.~(4)
using the wavefunction close to $ \bm{k}= \bm{k}_{\rm 0}$.
From $f_j( \bm{k}) - f_j( \bm{k}_{\rm 0}) \simeq \bm{v}_j \cdot \delta \bm{k}$, ($j$ = 2, 3, 0) with $\delta \bm{k} = \bm{k} - \bm{k}_{\rm 0}$,
the energy of the Dirac cone is given by
$E_{\pm}( \bm{k}) - E( \bm{k}_{\rm 0}) \simeq \bm{v}_0 \cdot \delta \bm{k}$
$\pm
\sqrt{ (\bm{v}_2 \cdot \delta \bm{k})^2 + (\bm{v}_3 \cdot \delta \bm{k})^2}$.
The coefficients of $k_x^3$, $k_x^5$ and $(k_xk_y)^2$
in Eqs.~(\ref{eq:eq7a}) and (\ref{eq:eq7b})
are determined by fitting the line to the other $ \bm{k}_{\rm 0}$.
As shown in Figs.~\ref{fig4}(a) and \ref{fig4}(b),
good coincidence is obtained for both $f_2( \bm{k})$ and $f_3( \bm{k})$ where
$f_2( \bm{k})$ reflects the symmetry of $k_y$=0.
The arrow in the inset of Fig.~\ref{fig4}(a) represents $\bm{v}_2$ at Dirac point
(I) which is perpendicular to $f_2( \bm{k})$ (solid line).
Coefficients $C_2$, $C_3$, $b_x$, $b_y$, and $b_z$
are calculated as follows.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{17502Fig4.eps}
\caption{(Color online) Dirac points (symbols) projected on the $k_x$ - $k_z$ plane (a) and
$k_x$ - $k_y$ plane (b), which are used to obtain
the fitting lines
for $f_2( \bm{k})$ and $f_3( \bm{k})$
(Eqs.~(\ref{eq:eq7a}) and (\ref{eq:eq7b})), respectively.
The inset shows $f_2( \bm{k})$
around the $\Gamma$-point.
The arrow, which depicts $\bm{v}_2$ at
Dirac point (I) with arbitrary scale,
corresponds
to the horizontal axis \textsl{\textbf{a*}}+\textsl{\textbf{c*}}
in Fig.~\ref{fig:cone}(a).
}
\label{fig4}
\end{figure}
First, we examine the Dirac cone at Dirac point (I)
where
$\bm{v}_2 = C_2(1,0,1)$,
$\bm{v}_3 = C_3(0,1,0)$, and
$\bm{v}_0 = 2 b_y k_{\rm 0y} (0,1,0)$.
Noting $\bm{v}_2 \cdot \bm{v}_3$ =~0,
the principal axes are given by
$k_x$+$k_z$ and $k_y$.
By comparing the velocities with
those of Figs.~\ref{fig:cone}(a)
and \ref{fig:cone}(b), which give
\begin{subequations}
\begin{eqnarray}
v_2 &=& \bm{v}_2 \cdot (2^{-1/2}, 0, 2^{-1/2}) = 0.21 \; ,
\label{eq:eq9a} \\
v_y &=& \bm{v}_3 \cdot (0, 1,0) \pm \bm{v}_0 \cdot (0,1,0) = 1.25 \pm 0.45
\; , \label{eq:eq9b}
\nonumber \\
\end{eqnarray}
\end{subequations}
we obtain $C_2 \simeq 0.148$, $b_y \simeq - 2.6$, and $C_3 \simeq 0.053$.
Coefficients $C_2$ and $C_3$ can also be obtained
from the gap $E_c( \bm{k})-E_v( \bm{k})$ at the $\Gamma$-point and Z-point,
which are estimated as 0.0994 and 0.1117, respectively.
The resultant quantities
$C_2$ = 0.137 and $C_3$ = 0.050 are compatible with those of
Eqs.~(\ref{eq:eq9a}) and (\ref{eq:eq9b}),
which suggests that the effective Hamiltonian of Eq.~(\ref{eq:eq5})
is valid not only for $ \bm{k}$ close to the nodal line but also for the
extended regions including the $\Gamma$- and Z-points.
Next, we examine the Dirac cone at Dirac point (II)
to verify the validity of Eqs.~(\ref{eq:eq7a}) and (\ref{eq:eq7a})
with $C_2 =0.148$ and $C_3 = 0.053$.
The Dirac cone is shown on the plane of $k_x$ - $k_z$,
but the principal axes do not coincide with that of $\bm{v}_2$ and $\bm{v}_3$
since $\bm{v}_2 \cdot \bm{v}_3 \not= 0$
from Eqs.~(\ref{eq:eq8a}) and (\ref{eq:eq8a}).
We estimate the gap function given by
$ 2 \Delta ( \bm{k}) = E_c( \bm{k})-E_v( \bm{k})$
$ = 2\sqrt{ (\bm{v}_2 \cdot \delta \bm{k})^2 + (\bm{v}_3 \cdot \delta \bm{k})^2}$
with $\delta \bm{k} =(x,0,z)$, which is calculated as
$(E_c( \bm{k})-E_v( \bm{k}))^2/4 = 0.45 x^2 + 0.117 x z + 0.02 z^2$
$(= Ax^2 + 2C xz + Bz^2)$.
This shows the rotation of the principal axes
from $(x,z)$ to $(X,Z)$, where
$x = X \cos \theta - Z \sin \theta$, and
$z= X \sin \theta + Z \cos \theta$. In terms of
$(X,Z)$, the gap function is expressed as
\begin{eqnarray}
& & \Delta( \bm{k})^2 = \Delta(\delta \bm{k} + \bm{k}_{\rm 0}) = A x^2 + 2C xz + B z^2 =
\nonumber \\
& & \frac{1}{2}\left( A+B + \sqrt{(A-B)^2 + 4C^2}\right) X^2
\nonumber \\
& & + \frac{1}{2}\left( A+B - \sqrt{(A-B)^2 + 4C^2}\right) Z^2
\; ,
\nonumber \\
& & \tan (2 \theta) = \frac{2C}{A-B} \; .
\label{eq:eq10}
\end{eqnarray}
The cross section of $\Delta( \bm{k}) = E_0 $
shows an ellipse with the radius of the minor [major] axis given by
$a=2^{1/2}E_0/\sqrt{A+B+((A-B)^2+4C^2)^{1/2}}$
[$b$ = $2^{1/2}E_0/\sqrt{A+B-((A-B)^2+4C^2)^{1/2}}$],
which is
rotated by an angle $\theta = 2^{-1}\tan^{-1}[2C/(A-B))]$.
The radius and angle for Dirac point (II)
are obtained as $a = 1.5 E_0$, $b=8.5 E_0$, $b/a \simeq 5.6$ and $\tan \theta = 0.13$, which
correspond well to the numerical result in Fig.~2(d), i.e.,
$b/a \simeq 7$ and $\tan \theta \simeq 0.2$.
Therefore, our Hamiltonian may be applied to all the Dirac points
between points (I) and (II).
Further, from Eq.~(\ref{eq:eq10}), we estimate
the area of the ellipse $S$ with the gap $2 E_0$
for an arbitrary Dirac point, where
the axis of the cone is parallel to the tangent of the nodal line
( i.e., $\propto \bm{v}_2 \times \bm{v}_3$)
and the wavefunction of the Dirac cone
is determined by $\bm{v}_2$ and $\bm{v}_3$.
By taking the axis of the cone as $y$ and the plane of the cone
as the $x$ - $z$ plane,
we obtain the area of the ellipse $S$ with the gap $2 E_0$
as
$S( \bm{k}_{\rm 0}) = \pi ab$ $= \pi E_0^2/ \sqrt{AB - C^2}
= \pi E_0^2 / |\bm{v}_2( \bm{k}_{\rm 0}) \times \bm{v}_3( \bm{k}_{\rm 0})|$,
where
$\bm{v}_2 = (v_{2x}, 0, v_{2z})$,
$\bm{v}_3 = (v_{3x}, 0, v_{3z})$,
$A= v_{2x}^2+v_{3x}^2$,
$B= v_{2z}^2+v_{3z}^2$, and
$C=v_{2x}v_{2z} v_{3x}v_{3z}$.
We note that
$\bm{v}_2 \cdot \bm{v}_3 \not= 0$
is generally expected
except for Dirac point (I), which is located
on the symmetric line of $\Gamma$ - Y.
Finally, we calculate
$b_x$ and $b_z$ in Eq.~(\ref{eq:eq7c})
from Dirac cone (II),
which gives $v_x = 0.36 \pm 0.12$ and $v_z = 0.09 \pm 0.06$.
In a way similar to Eqs.~(\ref{eq:eq9a}) and (\ref{eq:eq9b}),
we obtain $b_x \simeq -0.30$ and $b_z \simeq 0.077 $
suggesting that the tilting occurs
towards the negative (negative) direction for the $k_x$ ($k_z$)
axis.
This is qualitatively consistent with Fig.~2(c).
As shown in Fig.~\ref{fig3}(b),
our DFT calculation of the energy difference
measured from that of Dirac point (I) exhibits
nonmonotonical behavior, e.g.,
-0.0031, -0.0061, -0.0059, -0.0047, and -0.0032~eV
for the nodal points No. 3, 5, 7, 9, and 11 in Fig.~\ref{fig3}(a), respectively.
However, it is quite complicated to reproduce the dispersion
quantitatively from Eq.~(\ref{eq:eq7c}) with a constant $C_0$,
and such a problem becomes an issue to be solved in the next step.
Here, it should be noted that the SOC was ignored in the present calculation.
Since the crystal structure is centrosymmetric,
the band structures are in Kramers degeneracy.
Then, the loop degeneracies are destroyed by the SOC,
leading to an energy gap with $\simeq$ 3 meV at Dirac point (I),~\cite{Tsumuraya2018} which will be examined further.
We comment on the experiment in which the present model is useful
for performing the analysis.
Since all the directions for the nodal line can be treated
on the same footing,
the effect of magnetic field is of interest to examine the characteristic
behavior of the Landau level. The Hall coefficient is
also expected to exhibit anisotropic behavior.
In summary, we performed first-principles calculations for [Pd(dddt)$_2$] at 8GPa, and found a non-coplanar nodal loop within the first BZ.
The present effective Hamiltonian obtained
from the Dirac points in Fig.~\ref{fig3}(a)
describes well the electron close to the nodal line.
The explicit form of the velocity is obtained by
Eqs.~(\ref{eq:eq8a})-(\ref{eq:eq8c}), with
$C_2 \simeq 0.148,
C_3 \simeq 0.05,
b_x \simeq -0.31,
b_y \simeq - 2.6$, and
$b_z \simeq 0.077$,
which are useful for future calculations of
transport.
\acknowledgements
The authors thank T. Miyazaki, H. Kino, A. Yamakage, F. Ishii, H. Sawahata, T. Shishidou, and T. Kariyado for useful discussions.
The computations were mainly carried out using the computer facilities of ITO
at the Research Institute for
Information Technology, Kyushu University and HOKUSAI-GreatWave at RIKEN.
This work was supported by JSPS KAKENHI Grant, JP16H06346 and 16K17756.
|
\section{CASE STUDY: COMPARING ATTRIBUTE- WITH STRUCTURE-BASED APPROACH}
As our case study, proposed evaluation mechanisms will be applied to compare human-oriented effectiveness of Attribute-Based Approach (ABA) and Structure-Based Approach (SBA).
These approaches are the baseline of most plagiarism detection approaches \cite{Al-Khanjari2010,Karnalim2017IAENG}.
We are aware that a work proposed in \cite{Verco1996} also compares those categories.
Yet, human perspective is not directly linked to plagiarism detection approaches in their evaluation.
Both ABA and SBA work by accepting two source codes to return a similarity degree.
At first, inputted source codes are tokenized using ANTLR \cite{Parr2013} (with an exclusion of comment tokens).
Later, resulted token sequences will be compared to each other using specific similarity algorithm.
ABA will utilize Vector Space Model and Cosine Similarity \cite{Croft2010} (where each vector refers to token occurrence frequency).
Whereas, SBA will utilize Running-Karp-Rabin Greedy-String-Tiling algorithm \cite{Wise1995}, with two as its minimum matching threshold and average similarity equation \cite{Wise1992} as its normalization technique.
Fifteen lecturer assistants are involved in this evaluation.
They are experienced in terms of checking students' programming assignments.
Hence, it is expected that resulted evaluation findings will be considerably objective.
It is true that the number of respondents are somewhat limited.
However, since the main goal of this case study is to prove the applicability of proposed evaluation mechanism, we would argue such limited number of respondents are enough.
\subsection{Methodology}
Proposed evaluation mechanisms will be merged as one in this case study.
Fig. \ref{fig:methodology} shows that such combination generates five phases: artificial cases generation, contradicting plagiarism pairs selection, respondent-answering, describing, and analysis phase.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.2\textwidth]{./images/methodology}}
\caption{Combined phases from proposed evaluation mechanisms}
\label{fig:methodology}
\end{figure}
Artificial cases (for aspect-oriented evaluation mechanism) are generated by considering token order as a differentiating aspect;
such aspect is only considered by SBA while determining similarity.
Four artificial cases are considered in this study, covering various scopes of order-changed token sequences: single-instruction, multiple-instructions, method, and class scope.
All of them are written in Java programming language.
A case with single-instruction scope has 4 plagiarized codes that are formed by swapping \textit{N} instructions (where \textit{N} in these codes are 0, 1, 3, and 5 respectively).
Remaining cases contain 6 plagiarized codes each, that are formed by changing the order of three subsequences in combinatoric manner.
Generated artificial cases are valid to be used in our evaluation;
their SBA's and ABA's result are statistically different (in terms of similarity degree) when measured using two-tailed paired t-test. It generates p=0.02, which is lower than 0.05.
Contradicting plagiarism pairs (for empirical evaluation mechanism) are selected from a Java-based plagiarism dataset \cite{Karnalim2018Icoict} by clustering plagiarized codes per original code and plagiarism level \cite{FaidhiJ.A.W;Robinson1987}, resulting 42 clusters (plagiarism level is added to our consideration for providing more comprehensive findings).
Forty-five contradicting pairs are selected for our survey;
they are formed from level-2 to level-6 plagiarism categories where each category contributes 5, 9, 11, 10, 10 contradicting pairs respectively.
Those pairs lead to a statistically significant difference between ABA and SBA when measured using two-tailed paired t-test.
It generates p-value=1.59E-78, which is lower than 0.05.
Hence, they are valid to be used in this evaluation.
Regarding artificial cases and contradicting plagiarism pairs, respondents will be asked to rank them in similar manner as defined in aspect-oriented and empirical evaluation mechanism.
However, considering ranking 45 contradicting pairs (for empirical evaluation mechanism) at once may generate biased result due to human fatigue, our respondents are divided to three groups of five where the member of each group is only responsible to rank one-third of total pairs.
After ranking artificial cases and contradicting plagiarism pairs, describing phase will be conducted.
At this phase, respondents will be asked to write down their manual detection technique as detail as possible.
Survey results collected from respondent-answering and describing phase will be further analyzed in qualitative and quantitative manner.
These analysis works exactly similar as in our proposed evaluation mechanisms.
\subsection{Results}
According to our evaluation, token order (which is SBA's signature aspect) is considered by respondents when detecting plagiarism cases.
Plagiarized code without such change (i.e., a verbatim copy of original code) is always assigned with the highest rank when compared to its counterpart with such change.
Further, resulted negated ranks are strongly correlated with SBA's similarity degrees in positive manner;
their correlation degree is 0.833 when measured using Pearson correlation (where 1 represents the strongest positive correlation).
Consequently, it can be stated that SBA is preferred than ABA from aspect-oriented evaluation perspective.
It is important to note that the correlation between negated ranks and ABA's similarity degrees is immeasurable due to Pearson correlation's nature.
It cannot correlate two sequences when one of them has no variability (which is ABA in our case study, it generates resemblant similarity degrees for all artificial cases).
When perceived from empirical evaluation, SBA is preferred than ABA in both general and plagiarism level perspective.
Fig. \ref{fig:preference_empirical} shows that SBA's percentage of respondent-preferred codes is far higher than ABA's.
Further, SBA's similarity degrees are more positively correlated with negated respondent-assigned ranks despite their low correlation.
As seen in Fig. \ref{fig:correlation_empirical}, on all categories, SBA generates positive correlation while ABA generates the negative one.
Low correlation between similarity degrees and respondent-assigned ranks is natural considering plagiarized codes are not specifically designed to accentuate the difference between evaluated plagiarism detection approaches.
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.45\textwidth]{./images/preference_empirical}}
\caption{The percentage of respondent-preferred favoring codes on contradicting plagiarism pairs. Each number is generated by counting how many favoring codes are assigned with the $1^{st}$ rank and then normalized based on the number of involved plagiarism pairs.}
\label{fig:preference_empirical}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[width=0.45\textwidth]{./images/correlation_empirical}}
\caption{Pearson correlation degree between similarity degrees and negated respondent-assigned ranks on contradicting plagiarism pairs.}
\label{fig:correlation_empirical}
\end{figure}
SBA generates the strongest correlation on level-2 plagiarism level category (which is about identifier renaming).
Hence, it can be stated that respondents tend to detect plagiarism as SBA when source code structure is unchanged.
In contrast, it generates the weakest correlation on level-4 plagiarism level category.
In other words, method structure change (i.e., level-4 signature attack) makes respondents focus less on source code structure.
Respondents' descriptions about their manual detection technique show that six aspects are considered while suspecting plagiarism cases.
Table \ref{tab:think_aloud_considered_aspects} displays those aspects, including their occurrences (calculated based on the number of respondents mentioning it) and their relationship to evaluated plagiarism detection approaches.
Most of mentioned aspects are related to SBA.
Hence, it can be stated that SBA is more effective than ABA from think-aloud evaluation perspective.
\begin{table}[htbp]
\caption{Considered aspects while suspecting plagiarism cases}
\label{tab:think_aloud_considered_aspects}
\centering
\begin{tabular}{| D{0.08\textwidth}| C{0.07\textwidth}| D{0.26\textwidth}|}
\hline
\bfseries \hspace{0.2cm} Aspect & \bfseries Occurrences & \bfseries Relationship to Evaluated Approaches\\
\hline
Statement order
&
11
&
Related to SBA since order is a part of source code structure.
\\
\hline
Semantic
&
5
&
Related to SBA since semantic is preserved based on source code structure.
\\
\hline
Identifier name
&
3
&
Related to ABA since identifier name is a source code characteristic.
\\
\hline
Structure
&
2
&
Obviously related to SBA.
\\
\hline
Output
&
1
&
Related to SBA since output is defined based on source code structure.
\\
\hline
Line of code
&
1
&
Related to ABA since line of code is a source code characteristic.
\\
\hline
\end{tabular}
\end{table}
To sum up, from human perspective, SBA is more effective than ABA according to three rationales.
First, SBA's signature aspect (i.e., token order) is considered by respondents when detecting plagiarism cases.
Second, SBA's favoring cases is preferred than ABA's.
Third, SBA is related to more human-oriented aspects for suspecting source code plagiarism.
\section{CONCLUSION AND FUTURE WORK}
In this paper, three human-oriented evaluation mechanisms in source code plagiarism detection have been proposed.
Unique to these mechanisms, they directly relate human perspective with evaluated plagiarism detection approaches.
These mechanisms have been tested on a case study toward attribute- and structure-based plagiarism detection approach.
According to such test, structure-based approach is more effective; its signature aspect and similarity degree is more related to respondents' preferences.
Further, it is related to most aspects that are mentioned by respondents for suspecting plagiarism cases.
For future works, we plan to evaluate human-oriented effectiveness of frequently-used features on source code plagiarism detection (such as method linearization \cite{Karnalim2017ICSESS,Karnalim2018}) using aspect-oriented evaluation mechanism.
Further, we also plan to systematize how human suspects source code plagiarism (with and without machine learning algorithm).
The results of such work (e.g., features with high human-oriented effectiveness degree and systematic steps to suspect plagiarism) will be used to develop source code plagiarism detection approach that is more aligned to human perspective.
\subsection{Think-Aloud Evaluation Mechanism}
Think-aloud evaluation mechanism compares the effectiveness of two source code plagiarism detection approaches based on respondents' descriptions about how manual detection works.
It is suitable to be used when only a limited number of plagiarism cases exists for evaluation and the main goal is to observe from process perspective (i.e., factors considered while detecting plagiarism).
This mechanism is inspired from think-aloud protocols used in Empirical Software Engineering \cite{Singer2008}.
Think-aloud evaluation mechanism consists of three phases.
First of all, respondents are asked to manually suspect plagiarized source codes for reminding them how manual detection works.
Second, they should describe how they suspected plagiarized codes as detail as possible in natural language sentences.
Third, each aspect from resulted descriptions will be qualitatively linked to evaluated plagiarism detection approaches;
an approach is more effective than another if its characteristics are frequently mentioned on respondents' descriptions.
To mitigate bias, no clue should be given to respondents in regard to evaluated plagiarism detection approaches.
It is important to note that this mechanism might not capture some human-oriented aspects due to human limitations namely low interpretation skill and awareness.
Low interpretation skill occurs when the respondent cannot formalize what is in their mind while suspecting plagiarized source codes;
whereas, low awareness occurs when the respondent does not aware about some parts of detection process (which have been unconsciously conducted).
\subsection{Aspect-Oriented Evaluation Mechanism}
Aspect-oriented evaluation mechanism relies on one differentiating aspect between two source code plagiarism detection approaches to measure human-oriented effectiveness.
If that aspect affects (or positively correlates more with) humans' preferences, it can be stated that an approach with such aspect is more effective.
Otherwise, the approach without such aspect is preferred.
This mechanism is suitable to be used when there is a salient differentiating aspect between both plagiarism detection approaches and the main goal is to observe from result perspective (i.e., resulted similarity degree).
In general, this mechanism works by asking respondents to rank plagiarism cases from human perspective wherein the results will be further analyzed in qualitative and quantitative perspective.
It is inspired from ranking concept in Information Retrieval \cite{Croft2010} where relevancy (i.e., similarity in our case) is defined based on ranks.
Prior to asking respondents, survey cases (that accentuate the impact of differentiating aspect) should be artificially developed.
For each case, an original source code is plagiarized to several codes where the only modification is the differentiating aspect with various degree.
To guarantee that the impact of differentiating aspect is shown, on such cases, the difference between both evaluated approaches in terms of similarity degree should be statistically significant.
The significance itself can be measured using t-test.
For each survey case, respondents are asked to rank plagiarized codes in descending order based on their similarity degree toward original code from human perspective.
Each plagiarized code will be assigned with a rank where the first rank refers to the highest similarity;
if two or more codes are equally similar, they could be assigned with similar rank.
To provide more objective results, each survey case should be rated by more than one respondent.
After all survey cases are ranked by the respondents, human-oriented effectiveness of evaluated approaches can be analyzed from both qualitative and quantitative perspective.
From qualitative perspective, an approach with aforementioned differentiating aspect is more effective if human-assigned ranks are changed as the degree of involved differentiating aspect is modified.
In contrast, from quantitative perspective, the effectiveness of evaluated plagiarism detection approaches is defined through Pearson correlation between similarity degrees and averaged human-assigned ranks;
a plagiarism detection approach is more effective than another if its resulted similarity degrees (measured on artificially-plagiarized codes toward their originals) generate higher correlation toward averaged human-assigned ranks.
Considering similarity degree's result interpretation (where higher value is preferred) contradicts human-assigned rank's (where lower value is preferred), averaged human-assigned ranks will be negated before correlating to make its result interpretation proportional to similarity degree's.
\subsection{Empirical Evaluation Mechanism}
Empirical evaluation mechanism relies on humans' preferences toward contradicting plagiarism pairs (i.e., pairs where each of their members exclusively favors a plagiarism detection approach), selected from existing plagiarism dataset, to evaluate human-oriented effectiveness.
This mechanism is suitable to be used when the main goal is to observe from result perspective (i.e., resulted similarity degree) and a publicly-available plagiarism dataset is involved.
It is inspired from ranking concept in Information Retrieval \cite{Croft2010} (where relevancy is defined based on ranks) and A/B testing in Web Analytics\cite{Kohavi2017} (where human should select one between two instances based on their qualitative perspective).
Contradicting plagiarism pairs are selected from a plagiarism dataset in fourfold.
First, plagiarized codes are clustered based on their original code (one cluster for each original code).
Second, per plagiarism detection approach, plagiarized codes will be compared with their originals, resulting a set of similarity degrees.
Third, for each cluster and a plagiarism detection approach, plagiarized codes are ranked in descending similarity order.
Fourth, contradicting plagiarism pairs are selected for each cluster.
Two plagiarized source codes (e.g., \textit{A} and \textit{B}) are considered as a contradicting plagiarism pair \textit{iff} returned value from either (\ref{eq:contradiction1}) or (\ref{eq:contradiction2}) is true.
\textit{P1} returns a rank resulted from the first plagiarism detection approach while \textit{P2} returns a rank resulted from another.
\begin{equation}
Con1(A, B) = (P1(A) > P1(B)) \land (P2(A) < P2(B))
\label{eq:contradiction1}
\end{equation}
\begin{equation}
Con2(A, B) = (P1(A) < P1(B)) \land (P2(A) > P2(B))
\label{eq:contradiction2}
\end{equation}
To provide more insight, suppose we have three plagiarized codes (\textit{A}, \textit{B}, and \textit{C});
their similarity degree toward original code---measured using two detection approaches: \textit{P1} and \textit{P2}---can be seen on Table \ref{tab:example}.
According to given example, \textit{P1}'s and \textit{P2}'s ranking order are \textit{\{A, C, B\} }and \textit{\{C, A, B\}} respectively.
As a result, \textit{A} and \textit{C} are considered as a contradicting pair;
they satisfy (\ref{eq:contradiction1}) since \textit{P1(A)} $>$ \textit{P1(C)} and \textit{P2(A)} $<$ \textit{P2(C)}.
\begin{table}[htbp]
\caption{Similarity degrees for illustrating how contradicting plagiarism pairs are selected}
\label{tab:example}
\centering
\begin{tabular}{| C{0.075\textwidth}| C{0.15\textwidth}| C{0.15\textwidth}|}
\hline
\bfseries Plagiarized Code & \bfseries P1's Similarity Degree & \bfseries P2's Similarity Degree\\
\hline
A
&
70\%
&
50\% \\
\hline
B
&
50\%
&
40\% \\
\hline
C
&
60\%
&
95\% \\
\hline
\end{tabular}
\end{table}
Two things should be considered while selecting contradicting pairs.
First, plagiarized codes in each contradicting pair should be explicitly different to each other when perceived by human.
Hence, it is preferable to select pairs that generate high delta similarity degree between evaluated approaches.
Second, the difference between evaluated approaches should not be coincidental on selected contradicting pairs.
Hence, it is important to assure that similarity degrees resulted from both approaches are significantly different to each other; wherein the significance can be measured using t-test.
For each contradicting pair, each respondent should select one plagiarized code that seems more similar to original code;
preferred code will be assigned with the $1^{st}$ rank while another will be assigned with the $2^{nd}$ rank.
The result of this evaluation can be quantitatively analyzed by correlating averaged human-assigned ranks with resulted similarity degrees (measured on plagiarized codes with their originals per evaluated plagiarism detection approach) using Pearson correlation;
higher correlation refers to higher effectiveness degree.
Similar to aspect-oriented mechanism, averaged human-assigned ranks should be negated beforehand to make its result interpretation proportional to similarity degree's (where higher value is preferred).
Further, the effectiveness of evaluated plagiarism detection approaches can also be naively measured by counting how many favoring codes are assigned with the $1^{st}$ rank for each approach.
Qualitative analysis is discouraged to be conducted.
Plagiarized codes are not specifically designed to accentuate the difference between evaluated plagiarism detection approaches.
It will therefore be difficult to qualitatively collect findings regarding why a favoring code is preferred than another.
\section{HUMAN-ORIENTED EVALUATION MECHANISMS}
Three human-oriented evaluation mechanisms (i.e., evaluation mechanisms that directly relate human perspective with evaluated plagiarism detection approaches) are proposed: think-aloud, aspect-oriented, and empirical evaluation mechanism.
To mitigate qualitative bias, human-oriented effectiveness is determined by comparing plagiarism detection approaches to each other through questionnaire survey.
Considering such comparison is best performed with only two instances, proposed evaluation mechanisms are designed to accept two approaches per execution.
However, if necessary, they could compare numerous approaches by adapting tournament selection from Genetic Algorithm \cite{russell2003artificial} (with an assumption that each approach refers to one instance and fitness function is the human-oriented effectiveness degree).
\input{evaluation_mechanism_1}
\input{evaluation_mechanism_2}
\input{evaluation_mechanism_3}
\section{INTRODUCTION}
Source code plagiarism occurs when a student claims another student's source code as their original work \cite{Cosma2008}.
In programming courses, it is considered as an illegal behavior;
the correlation between student's knowledge and grade will be weakened, resulting more complicated recruitment process for industry.
Mitigating such illegal behavior is a crucial task for programming examiners (e.g., lecturers or assistant lecturers).
To mitigate human effort, automated plagiarism detection approaches have been developed \cite{Lancaster2004}.
Using such approaches, plagiarism-suspected pairs could be detected in a no time.
However, most approaches (where some of them are implemented on publicly available tools such as JPlag \cite{Prechelt2002}) could be either too sensitive or insensitive when compared to human.
Some of them could generate numerous false positives (i.e., non-plagiarized pairs which are considered as plagiarism-suspected pairs) while the others could generate numerous false negatives (i.e., plagiarized pairs which are not considered as plagiarism-suspected pairs).
The existence of false results would complicate accusation process: examiner needs to double-check each result comprehensively to avoid mis-detection.
One of the solutions for reducing false results is to utilize a plagiarism detection approach which behaves like humans (i.e., examiners), considering correct detection results are originally defined by them.
Therefore, it is natural to evaluate plagiarism detection approaches from human perspective.
Nevertheless, to the best of our knowledge, existing evaluation mechanisms only indirectly consider human perspective.
We would argue that direct relation between human perspective and evaluated plagiarism detection approaches are needed to favor human-like approach in evaluation.
This paper proposes two contributions.
First, three evaluation mechanisms that directly relate human perspective to evaluated plagiarism detection approaches are proposed.
Such mechanisms could become evaluation alternatives in addition to conventional mechanism (which relies on human-annotated dataset).
Second, using proposed evaluation mechanisms, a comparative study between attribute- and structure-based approach (i.e., two well-known approach categories in plagiarism detection \cite{Al-Khanjari2010,Karnalim2017IAENG}) is presented.
This study draws the applicability of proposed evaluation mechanisms.
\section*{Acknowledgment}
\bibliographystyle{IEEEtran}
\section{RELATED WORKS}
In general, source code plagiarism detection approaches work in two phases: conversion and comparison phase.
Conversion phase translates given source codes and/or their features to intermediate representations for more effective and efficient comparison phase.
Some of frequently-used intermediate representations are: source code token sequence \cite{Prechelt2002,Engels2007,ElBachirMenai2010,Ottenstein1976,FaidhiJ.A.W;Robinson1987,Acampora2015,Inoue2012,Ohmann2015,Mozgovoy2005,Moussiades2005,Al-Khanjari2010,Burrows2007,Poon2012,Duric2013,Bejarano2015,Kustanto2009}, source code word sequence (i.e., a sequence generated from natural language tokenizer) \cite{flores2015cross,Cosma2012}, low-level token sequence (i.e., a sequence generated from executable file of given source code) \cite{Karnalim2016,Rabbani2017,Karnalim2017IAENG,Karnalim2017ICSESS,Karnalim2018}, program dependency graph \cite{Liu2006}, and abstract syntax tree\cite{Fu2017,Ganguly2017}.
After translated, resulted intermediate representations are then passed to comparison phase---that measures similarity degree and defines which pairs are suspected as plagiarism cases.
With regard to this phase, plagiarism detection approaches can be further classified into three categories: attribute-based, structure-based, and hybrid approach \cite{Al-Khanjari2010,Karnalim2017IAENG}.
Attribute-based approach determines plagiarism-suspected pairs based on shared source code characteristics (e.g., the number of unique operators \cite{Ottenstein1976} and token occurrences \cite{Acampora2015}).
Initially, similarity measurement on such approach relies on attribute-counting mechanism;
two codes are considered similar if both of them share the same characteristic occurrence frequencies \cite{Ottenstein1976,FaidhiJ.A.W;Robinson1987,Vamplew2010}.
Later, advanced similarity measurements are incorporated.
These measurements are adapted from Fuzzy Logic (e.g., Fuzzy C-Means \cite{Acampora2015}), Machine Learning (e.g., Random Forest \cite{Ganguly2017}), and Information Retrieval (e.g., Cosine Similarity \cite{Inoue2012,Ohmann2015}, Jaccard Coefficient \cite{Moussiades2005}, Language Model \cite{Ganguly2017}, and Latent Semantic Analysis \cite{Cosma2012,flores2015cross}).
Structure-based approach determines plagiarism-suspected pairs based on shared source code structure.
This approach typically relies on string-matching algorithms---such as Running-Karp-Rabin Greedy-String-Tiling \cite{Prechelt2002,Kustanto2009,Karnalim2016,Karnalim2017ICSESS,Karnalim2017IAENG,Karnalim2018,Bejarano2015,Duric2013}, local alignment \cite{Rabbani2017}, and Winowing algorithm \cite{Duric2013}---that have been modified to handle token streams.
However, considering the limitation of those algorithms, some graph-matching algorithms (e.g., tree kernel algorithm \cite{Fu2017} and graph isomorphism algorithm \cite{Liu2006}) are proposed as alternative comparison algorithms.
Hybrid approach combines both attribute- and structure-based approach for determining plagiarism-suspected pairs.
Such combination aims to enhance either effectiveness \cite{Poon2012,Engels2007,ElBachirMenai2010} or efficiency \cite{Burrows2007}.
For enhancing effectiveness, these combined approaches either show the results of both conventional approaches at once \cite{ElBachirMenai2010} or treat the result of one approach as an input of another (e.g., using the result of structure-based approach as an attribute for learning algorithm \cite{Engels2007} or clustering algorithm \cite{Poon2012}).
On the other, for enhancing efficiency, plagiarism-suspected pairs from attribute-based approach are used as the inputs of structure-based approach \cite{Burrows2007}.
In such manner, not all source pairs are compared using structure-based approach (which is typically costly in terms of processing time).
In the context of effectiveness evaluation, two mechanisms are frequently used in source code plagiarism detection: statistical and domain-specific mechanism.
Statistical mechanism utilizes metrics commonly used in statistic (such as precision) to determine the performance of given plagiarism detection approaches \cite{Acampora2015,Kustanto2009,flores2015cross,Inoue2012,Ohmann2015,Burrows2007,Prechelt2002,Moussiades2005,Al-Khanjari2010,Engels2007,Ganguly2017,Fu2017,Cosma2012,Duric2013,Mozgovoy2005}.
This mechanism requires an evaluation dataset covering all statistical result categories (i.e., true positive, true negative, false positive, and false negative).
In contrast, domain-specific mechanism utilizes metrics specifically designed for plagiarism detection (e.g., similarity degree) to determine plagiarism detection approaches' performance \cite{Rabbani2017,Karnalim2016,Karnalim2017IAENG,Karnalim2018Icoict,Karnalim2017ICSESS,Karnalim2018,Bejarano2015,Liu2006,Poon2012,ElBachirMenai2010}.
Unique to this mechanism, it only relies on true positives (that are frequently created in controlled environment).
In other words, it is more practical to be used (since only true positives are required) yet it is less related to real environment (since other result categories are not considered).
One of the objectives in developing automated source code plagiarism detection is to let computer imitates manual plagiarism detection done by human.
Hence, it is natural to evaluate the effectiveness of plagiarism detection approaches from human perspective.
However, existing evaluation mechanisms (i.e., statistical and domain-specific mechanism) may not sufficiently reflect human perspective; considering such perspective is not directly related to evaluated approaches;
they are only related in indirect manner through human-annotated dataset.
Such dissonance could complicate accusation process due to the existence of false results. |
\section{Introduction}
In the study of higher dimensional Fano varieties, it is interesting to find examples
with special geometric structures.
The recent papers \cite{AM} and \cite{LP} investigate log Fano manifolds obtained by blowing up points in projective spaces
or products of projective spaces.
In \cite{Tsukioka}, the author considered weak Fano manifolds
having small contractions (i.e. birational morphisms whose exceptional loci has codimension greater than or equal to 2).
We recall the construction of a small contraction using successive blow-ups. Let $Y$ be a smooth projective variety of dimension $n \geq 3$. Let $S$ and $C$ be smooth subvarieties of codimension $\geq 2$
with $\dim S +\dim C\leq n-1$. Assume that $S$ and $C$ intersect transversally at points. Let $X$ be the blow-up along $C$ and $\widetilde{X}$ the blow-up along the strict transform of $S$. Then, $\widetilde{X}$ has exceptional loci of codimension $\geq 2$ over the intersections of $S$ and $C$ (this is a slight generalization of \cite{Kawamata} Example 2.6).
The problem is to determine the triples $(Y,S,C)$ such that $\widetilde{X}$ is (weak) Fano or more generally log Fano. In this note, we consider the case in which $Y$ is a product of two projective spaces,
$S$ is a fiber of a projection, and $C$ is a hypersurface in a fiber of the other projection.
We work over the field of complex numbers.
\begin{theorem} Let $Y=\mathbb{P}^{n-k}\times \mathbb{P}^{k}$ with $n\geq 3$ and $k\in \{2,\cdots, n-1\}$.
Let $S$ be a fiber of the projection $Y\to \mathbb{P}^{k}$. Let $C$ be a smooth hypersurface of degree
$d\geq 1$ in a fiber of the other projection $Y\to \mathbb{P}^{n-k}$. Assume $S\cap C\neq \emptyset$.
Let $\pi:X\to Y$ be the blow-up along $C$. Let $S'$ be the strict transform of $S$ by $\pi$.
Let $\widetilde{X}$ be the blow-up of $X$ along $S'$. Then, the anti-canonical divisor of $\widetilde{X}$ is big.
If $n=3$, $\widetilde{X}$ is log Fano for any $d\geq 1$,
is weak Fano for $d\in\{1,2,3\}$, and is not Fano for any $d\geq 1$.
Assume $n\geq 4$.
If one of the following holds, then $\widetilde{X}$ is log Fano.
\begin{itemize}
\item[(1)] $d=1$, $n$ is even and $k=\frac{n}{2}$
\item[(2)] $d=1$, $n$ is even and $k=\frac{n+2}{2}$
\item[(3)] $d=1$, $n$ is odd and $k\in\{\frac{n-1}{2},\frac{n+1}{2}\}$
\item[(4)] $(n,k,d)=(4,2,2)$ or $(4,3,2)$
\item[(5)] $(n,k,d)=(5,3,2)$.
\end{itemize}
Moreover, $\widetilde{X}$ is Fano if and only if (1) holds, and
$\widetilde{X}$ is weak Fano if and only if (1),(3) or (5) holds.
\end{theorem}
Remark: In Section \ref{proof}, we give explicit log Fano pairs for the cases (1)--(5). We also find a condition for divisors in a specific form to be boundary divisors (Lemma \ref{log Fano pair}). The author do not know whether $\widetilde{X}$ is log Fano or not in the other cases.
\
Definition: Let $X$ be a smooth projective variety.
We call $X$ {\it Fano} (resp. {\it weak Fano}) if the anti-canonical divisor $-K_{X}$ is ample (resp. nef and big).
A {\it log Fano pair} $(X,\Delta)$ is a klt pair such that $-(K_{X}+\Delta)$ is ample.
We call $X$ {\it log Fano} if there exists an effective $\mathbb{Q}$-divisor $\Delta$ such that $(X,\Delta)$ is a log Fano pair. The divisor $\Delta$ is called {\it boundary divisor}. We have the following implications:
\begin{center}
Fano $\Rightarrow$ weak Fano $\Rightarrow$ log Fano $\Rightarrow$ $-K_{X}$ is big.
\end{center}
The first implication is obvious. For the second one, see \cite{AM} Lemma 2.5.
The last one is due to Kodaira's lemma (see \cite{Lazarsfeldbook} Corollary 2.2.7).
\
This note is organized as follows.
First, we find a nef divisor on a special prime divisor $\widetilde{D}\subset \widetilde{X}$
(Section \ref{divisor}).
Then, we determine the structure of the nef cone of $\widetilde{X}$
by using the restriction map $\operatorname{Nef} (\widetilde{X})\to \operatorname{Nef} (\widetilde{D})$.
We also compute the effective cone of $\widetilde{X}$ (Section \ref{cone}).
Once the nef cone of $\widetilde{X}$ is described explicitly, it is easy to find conditions for $-(K_{\widetilde{X}}+\Delta)$ to be ample.
We prove Theorem by solving a system of linear inequalities in the coefficients of the divisor $\Delta$ and
the values $n$, $k$ and $d$ (Section \ref{proof}).
\
The numerical equivalence class of an $\mathbb{R}$-divisor $D$ is denoted by $[D]$.
Let $\mathbb{R^{+}}$ be the set of non-negative real numbers. We put
$$\mathbb{R^{+}}[D_{1},\cdots,D_{k}]:=\mathbb{R^{+}}[D_{1}]+\cdots +\mathbb{R^{+}}[D_{k}]. $$
\section{A nef divisor on a special prime divisor}\label{divisor}
The following is useful to determine the structure of
nef cones.
\begin{lemma} (\cite{Tsukioka} Lemma 3.2)\label{induction}
Let $X$ be a smooth projective variety and
let $D$ be a smooth prime divisor on $X$.
Let $N$ be a divisor on $X$.
If $[N] |_{D} \in \operatorname{Nef} (D)$ and $[N-D]\in \operatorname{Nef} (\widetilde{X})$,
then $[N]\in \operatorname{Nef} (X)$.
\end{lemma}
\begin{lemma}\label{point}
Let $\alpha: X\to \mathbb{P}^{n}$ ($n\geq 2$) be the blow-up at one point.
Then, we have
\begin{eqnarray*}
\operatorname{Nef} (X)=\mathbb{R^{+}}[\alpha^{*}{\cal O}_{\mathbb{P}^{n}}(1),\alpha^{*}{\cal O}_{\mathbb{P}^{n}}(1)-\operatorname{Exc} (\alpha)].
\end{eqnarray*}
In particular, $\alpha^{*}{\cal O}_{\mathbb{P}^{n}}(1)-\operatorname{Exc} (\alpha)$ is nef.
\end{lemma}
{\it Proof.} Note that we have $\rho(X)=2$. Hence, the cone $\operatorname{Nef} (X)$ has exactly two extremal rays. Hence, the statement follows from
the fact that the linear system
$|\alpha^{*}{\cal O}_{\mathbb{P}^{n}}(1)-\operatorname{Exc} (\alpha)|$ is base point free. \qed
\
We fix the following notation (valid only in this section):
Let $m$ and $k$ be natural numbers such that $m\geq 1$ and $k\geq 2$.
Let $V$ be a smooth hypersurface of degree $d\geq 1$ in $\mathbb{P}^{k}$.
Put $D:=\mathbb{P}^{m}\times V$.
Let $pr_{1}: D\to \mathbb{P}^{m}$ and $pr_{2}: D\to V$ be the projections.
Let $C$ (resp. $S$) be a fiber of $pr_{1}$ (resp. $pr_{2}$).
Note that $C$ (resp. $S$) is isomorphic to $V$ (resp. $\mathbb{P}^{m}$).
If $m=1$ (resp. $k=2$), then $C$ (resp. $S$) is a divisor on $D$.
Put $H:=pr_{1}^{*}{\cal O}_{\mathbb{P}^{m}}(1)$
and
$L:=pr_{2}^{*}{\cal O}_{V}(1)$ where ${\cal O}_{V}(1):={\cal O}_{\mathbb{P}^{k}}(1)|_{V}$.
We define the morphisms $\pi:D'\to D$ and $\beta:\widetilde{D}\to D'$ as follows.
\begin{itemize}
\item If $C$ in not a divisor on $D$,
let $\pi:D'\to D$ be the blow-up along $C$.
We put
$E:=\operatorname{Exc} (\pi)$, $H':=\pi^{*}H$, $L':=\pi^{*}L$ and $S':=\pi_{*}^{-1} S$.
If $C$ is a divisor on $D$,
we put
$\pi:=id$, $D':=D$, $E:=C$, $H':=H$, $L':=L$ and $S'=S$.
\item If $S'$ is not a divisor on $D'$,
let $\beta :\widetilde{D}\to D'$ be the blow-up along $S'$. We put
$F:=\operatorname{Exc} (\beta)$, $\widetilde{H}:=\beta^{*}H'$, $\widetilde{L}:=\beta^{*}L'$ and $\widetilde{E}:=\beta^{*}E$.
If $S'$ is a divisor on $D'$, we put
$\beta:=id$, $\widetilde{D}:=D'$, $F:=S'$, $\widetilde{H}:=H'$, $\widetilde{L}:=L'$ and $\widetilde{E}:=E$.
\end{itemize}
\begin{proposition}\label{nef} We have
$[\widetilde{H}+d\widetilde{L}-\widetilde{E}-F]\in \operatorname{Nef}(\widetilde{D}).$
\end{proposition}
{\it Proof}. Since $\widetilde{H}+d\widetilde{L}-\widetilde{E}-F=(d-1)\widetilde{L}+(\widetilde{H}-\widetilde{E})+(\widetilde{L}-\widetilde{F})$,
it is sufficient to show that $\widetilde{H}-\widetilde{E}$ and $\widetilde{L}-F$ are nef.
Case $m=1$:
$C$ is a divisor on $D=\mathbb{P}^{1}\times V$ and
we have $H'=H=pr_{1}^{*}{\cal O}_{\mathbb{P}^{1}}(1)\sim C=E$.
Hence, $\widetilde{H}-\widetilde{E}=\beta^{*}(H'-E)$ is linearly equivalent to $0$, in particular nef.
Consider the blow-up
$\mu:\operatorname{Bl}_{s}(\mathbb{P}^{k})\to \mathbb{P}^{k}$ at the point
$s:=pr_{2}(S)$. Put $\widetilde{V}:=\mu_{*}^{-1} V$.
Then, $\mu |_{\widetilde{V}}:\widetilde{V}\to V$ is the blow-up at the point $s\in V$.
Let $\widetilde{pr_{2}}:\widetilde{D}=\mathbb{P}^{1}\times \widetilde{V}\to\widetilde{V}$ be the projection.
Then, we have the commutative diagram:
\begin{eqnarray*}
\xymatrix{
\widetilde{D}\ar[d]_{\widetilde{pr_{2}}} \ar[r]^{\beta} & D \ar[d]^{pr_{2}} \\
\widetilde{V}\ar[r]_{\mu |_{\widetilde{V}}} & V.
}
\end{eqnarray*}
This yields
\begin{eqnarray*}
\widetilde{L}&=&\beta^{*}L
=\beta^{*}pr_{2}^{*}{\cal O}_{V}(1)
=\widetilde{pr_{2}}^{*}\left(\mu^{*}{\cal O}_{\mathbb{P}^{k}}(1)|_{\widetilde{V}}\right), \\
F&=&\widetilde{pr_{2}}^{*}\operatorname{Exc} (\mu |_{\widetilde{V}})
=\widetilde{pr_{2}}^{*}\left(\operatorname{Exc} (\mu)|_{\widetilde{V}}\right).
\end{eqnarray*}
By Lemma \ref{point}, $\mu^{*}{\cal O}_{\mathbb{P}^{k}}(1)-\operatorname{Exc} (\mu)$ is nef.
Hence, so is $\widetilde{L}-F$.
Case $m\geq 2$:
Consider the blow-up $\varepsilon: \operatorname{Bl}_{c}(\mathbb{P}^{m})\to \mathbb{P}^{m}$ at the point $c:=pr_{1}(C)$.
Let $pr_{1}':D'\simeq \operatorname{Bl}_{c}(\mathbb{P}^{m})\times V\to \operatorname{Bl}_{c}(\mathbb{P}^{m})$ be the projection.
Then, we have the commutative diagram:
\begin{eqnarray*}
\xymatrix{
D' \ar[d]_{\pi} \ar[r]^{pr_{1}'} & \operatorname{Bl}_{c}(\mathbb{P}^{m}) \ar[d]^{\varepsilon} \\
D \ar[r]_{pr_{1}} & \mathbb{P}^{m}.}
\end{eqnarray*}
By Lemma \ref{point} again,
$H'-E=(pr_{1}')^{*}(\varepsilon^{*}{\cal O}_{\mathbb{P}^{m}}(1)-\operatorname{Exc} (\varepsilon))$
is nef. Thus, so is $\widetilde{H}-\widetilde{E}=\beta^{*}(H'-E)$.
Now, we show $\widetilde{L}-F$ is nef. Fix $m\geq 2$ and use the induction on $k\geq 2$.
If $k=2$, then $\dim V=1$ and $S$ is a divisor on $D$.
Since $S$ is a fiber of $pr_{2}:D\to V$, we have
$L=pr_{2}^{*}({\cal O}_{\mathbb{P}^{k}}(1) |_{V})\equiv dS$ and $L-S\equiv (d-1)S$.
Thus, $\widetilde{L}-F=L'-S'=\pi^{*}(L-S)$ is nef.
From now on, we put the index $k$ to divisors and varieties,
i.e. $V,D,S,\widetilde{D}$ etc. are written as $V_{k},D_{k},S_{k},\widetilde{D}_{k}$ etc.
Recall that $V_{k}$ is a smooth hypersurface of degree $d\geq 1$
in $\mathbb{P}^{k}$ and $D_{k}=\mathbb{P}^{m}\times V_{k}$.
Assume $[\widetilde{L}_{k}-F_{k}]\in \operatorname{Nef} (\widetilde{D}_{k})$.
Consider a smooth divisor $M_{k+1}$ on $D_{k+1}=\mathbb{P}^{m}\times V_{k+1}$
such that $M_{k+1}\sim L_{k+1}$ and $S_{k+1}\subset M_{k+1}$.
Put $\widetilde{M}_{k+1}:=(\pi\circ \beta)_{*}^{-1} M_{k+1}$.
Since $M_{k+1}$ is isomorphic to $\mathbb{P}^{m}\times V_{k}=D_{k}$
as variety, so is
$\widetilde{M}_{k+1}$ to $\widetilde{D}_{k}$.
The identification $\widetilde{M}_{k+1}=\widetilde{D}_{k}$ yields
\begin{eqnarray*}
\widetilde{L}_{k+1}|_{\widetilde{M}_{k+1}}=\widetilde{L}_{k},\ \ \
F_{k+1}|_{\widetilde{M}_{k+1}}=F_{k}.
\end{eqnarray*}
Hence, by inductive hypothesis,
$[\widetilde{L}_{k+1}-F_{k+1}]|_{\widetilde{M}_{k+1}}\in \operatorname{Nef} (\widetilde{M}_{k+1})$.
Since $\widetilde{M}_{k+1}\sim \widetilde{L}_{k+1}-F_{k+1}$, we have
\begin{eqnarray*}
[\widetilde{L}_{k+1}-F_{k+1}-\widetilde{M}_{k+1}]=0
\in \operatorname{Nef}(\widetilde{D}_{k+1}).
\end{eqnarray*}
By Lemma \ref{induction}, $[\widetilde{L}_{k+1}-F_{k+1}]\in \operatorname{Nef} (\widetilde{D}_{k+1})$.
Therefore, $[\widetilde{L_{k}}-F_{k}]\in \operatorname{Nef}(\widetilde{D}_{k})$ for any $k\geq 2$.
\qed
\section{Structure of nef cone and effective cone}\label{cone}
From now on, we fix the following.
Notation (*):
Let $n$ and $k$ be natural numbers such that $n\geq 3$ and $2\leq k\leq n-1$.
Put $Y:=\mathbb{P}^{n-k}\times \mathbb{P}^{k}$.
Let $p:Y\to \mathbb{P}^{n-k}$ and $q:Y\to \mathbb{P}^{k}$ be the two projections.
Put $H:=p^{*}{\cal O}_{\mathbb{P}^{n-k}}(1)$ and $L:=q^{*}{\cal O}_{\mathbb{P}^{k}}(1)$.
Let $P_{0}$ be a fiber of $p:Y\to \mathbb{P}^{n-k}$.
Let $C$ be a smooth hypersurface of degree $d\geq 1$ in $P_{0}\simeq \mathbb{P}^{k}$.
Let $S$ be a fiber of $q:Y\to \mathbb{P}^{k}$ such that $S\cap C\neq \emptyset$.
Put $y_{0}:=S\cap C =S\cap P_{0}$.
Let $h$ be a line in a fiber of $p$ such that $h\cap S=\emptyset$ and $h\cap C=\emptyset$.
Let $l$ be a line in a fiber of $q$ such that $l\cap S=\emptyset$ and $l\cap C=\emptyset$.
Let $h_{0}$ be a line in $P_{0}\simeq \mathbb{P}^{k}$ such that $h_{0}\not\subset C$ and $y_{0}\not\in h_{0}$.
Let $l_{0}$ be a line in a fiber of $q$ different from $S$ such that $l_{0}\cap C\neq \emptyset$.
Let $H_{0}$ be an element of the the linear system $|H|$ such that $P_{0}\subset H_{0}$.
Let $L_{0}$ be an element of the linear system $|L|$ such that $S\subset L_{0}$. If $d=1$, we assume that $L_{0}$ does not contain $C$.
Put $D:=q^{-1}(q(C))$. Note that we have $D\sim dL$.
Let $\pi:X\to Y$ be the blow-up along $C$.
Put $E:=\operatorname{Exc} (\pi)$ and $E_{0}:=\pi^{-1}(y_{0})$.
Let $e_{0}$ be a line in $E_{0}\simeq \mathbb{P}^{n-k}$.
Let $e$ be a line in a fiber different from $E_{0}$
of the $\mathbb{P}^{n-k}$-bundle $\pi |_{E}:E\to C$
(if $k=n-1$, then $e$ and $e_{0}$
are fibers of the $\mathbb{P}^{1}$-bundle $\pi |_{E}$).
Put $S':=\pi_{*}^{-1} S$ and $P_{0}':=\pi_{*}^{-1} P_{0}$.
Note that
we have $S\cap P_{0}=y_{0}$ while $S'\cap P_{0}'=\emptyset$.
Put $H':=\pi^{*}H$ and $L':=\pi^{*}L$.
Let $H_{0}',L_{0}',D',h',l',h_{0}'$ and $l_{0}'$
be the strict transforms by $\pi$ of
$H_{0},L_{0},D,h,l,h_{0}$ and $l_{0}$, respectively.
Let $\beta:\widetilde{X}\to X$ be the blow-up along $S'$.
Put $F:=\operatorname{Exc} (\beta)$.
Let $f$ be a line in a fiber of the $\mathbb{P}^{k-1}$-bundle $\beta |_{F}:F\to S'$
(if $k=2$ then $f$ is a fiber of the $\mathbb{P}^{1}$-bundle $\beta |_{F}$).
Put
$\widetilde{H}:=\beta^{*}H'$,
$\widetilde{L}:=\beta^{*}L'$
and $\widetilde{E}:=\pi^{*}E$.
Finally, let
$\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{D},
\widetilde{h},\widetilde{l},\widetilde{h_{0}},\widetilde{l_{0}},\widetilde{e}$ and $\widetilde{e_{0}}$
be the strict transforms by $\beta$ of
$H_{0}',L_{0}',D',h',l',h_{0}',l_{0}',e$ and $e_{0}$, respectively.
\begin{lemma}\label{intersection}
We have the following table of intersection numbers.
\begin{eqnarray*}
\begin{array}{c|cccc}
&\widetilde{H}&\widetilde{L}&\widetilde{H}+d\widetilde{L}-\widetilde{E}&\widetilde{H}+d\widetilde{L}-\widetilde{E}-F \\
\hline
\widetilde{l_{0}}&1&0&0&0 \\
\widetilde{h_{0}}&0&1&0&0 \\
\widetilde{e_{0}}&0&0&1&0 \\
f&0&0&0&1
\end{array}
\end{eqnarray*}
\end{lemma}
{\it Proof}.
We have $e\equiv e_{0}$ in $X$. Note that $e_{0}$ intersects $S'$ transversally at one point.
Hence $\widetilde{e}\equiv \widetilde{e}_{0}+f$ in $\widetilde{X}$.
We have $l \equiv l_{0}$ in $Y$.
Since $l_{0}$ intersects $C$ transversally at one point,
we have $l'\equiv l_{0}'+e$ in $X$.
This implies $\widetilde{l}\equiv \widetilde{l_{0}}+\widetilde{e}$ in $\widetilde{X}$.
Note that $h\equiv h_{0}$ and $(C\cdot h_{0})_{P_{0}}=d$.
Hence $h'\equiv h'_{0}+de$, i.e. $\widetilde{h}\equiv \widetilde{h_{0}}+d\widetilde{e}$
because $S' \cap h_{0}'=\emptyset$.
The following intersection numbers are obvious.
\begin{eqnarray*}
\begin{array}{c|cccc}
&\widetilde{H}&\widetilde{L}&\widetilde{E}&F \\
\hline
\widetilde{l} &1&0&0&0 \\
\widetilde{h} &0&1&0&0 \\
\widetilde{e} &0&0&-1&0 \\
f &0&0&0&-1
\end{array}
\end{eqnarray*}
Therefore, we have the following table.
\begin{eqnarray*}
\begin{array}{c|cccc|cccc}
&\widetilde{H}&\widetilde{L}&\widetilde{E}&F &\widetilde{H}&\widetilde{L}&\widetilde{H}+d\widetilde{L}-\widetilde{E}&\widetilde{H}+d\widetilde{L}-\widetilde{E}-F \\
\hline
\widetilde{l_{0}}\equiv \widetilde{l}-\widetilde{e} &1&0&1&0&1&0&0&0 \\
\widetilde{h_{0}}\equiv \widetilde{h}-d\widetilde{e} &0&1&d&0&0&1&0&0 \\
\widetilde{e_{0}}\equiv \widetilde{e}-f &0&0&-1&1&0&0&1&0 \\
f &0&0&0&-1&0&0&0&1
\end{array}
\end{eqnarray*}
\qed
\begin{proposition}\label{nefcone} We have
\begin{eqnarray*}
\operatorname{Nef} (\widetilde{X})
=\mathbb{R^{+}} [
\widetilde{H},\widetilde{L},\widetilde{H}+d\widetilde{L}-\widetilde{E},
\widetilde{H}+d\widetilde{L}-\widetilde{E}-F].
\end{eqnarray*}
\end{proposition}
{\it Proof}. We have $[\widetilde{H}],[\widetilde{L}]\in\operatorname{Nef}(\widetilde{X})$ and
$[\widetilde{l_{0}}],[\widetilde{h_{0}}],[\widetilde{e_{0}}],[f]\in \overline{\operatorname{NE}}(\widetilde{X})$.
Recall $\rho (\widetilde{X})=4$.
By Lemma \ref{intersection}, it is sufficient to show that
$\widetilde{H}+d\widetilde{L}-\widetilde{E}$ and $\widetilde{H}+d\widetilde{L}-\widetilde{E}-F$ are nef (see \cite{Tsukioka} Lemma 3.1).
First, we show $[\widetilde{H}+d\widetilde{L}-\widetilde{E}-F]\in\operatorname{Nef} (\widetilde{X})$.
Put $V:=q(C)$. We have $D\simeq \mathbb{P}^{n-k}\times V$. Consider the projections
$pr_{1}:=p|_{D}:D\to \mathbb{P}^{n-k}$ and $pr_{2}:=q|_{D}:D\to V$. Note that
$C$ is a fiber of $pr_{1}$ and $S$ is a fiber of $pr_{2}$.
If we put $m:=n-k$,
then our divisors $\widetilde{H}|_{\widetilde{D}},\widetilde{L}|_{\widetilde{D}},\widetilde{E}|_{\widetilde{D}},F|_{\widetilde{D}}$
correspond exactly to the divisors $\widetilde{H},\widetilde{L},\widetilde{E},F$ in Proposition \ref{nef}.
Hence $[\widetilde{H}+d\widetilde{L}-\widetilde{E}-F]|_{\widetilde{D}}\in\operatorname{Nef} (\widetilde{D})$.
Note that $\widetilde{D}\sim d\widetilde{L}-\widetilde{E}-F$.
By Lemma \ref{induction},
we conclude
$[\widetilde{H}+d\widetilde{L}-\widetilde{E}-F]\in \operatorname{Nef} (\widetilde{X})$.
Now, we show $[\widetilde{H}+d\widetilde{L}-\widetilde{E}]\in\operatorname{Nef}(\widetilde{X})$.
If $n-k=1$, we have $S'\simeq S\simeq \mathbb{P}^{1}$,
$H'|_{S'}\simeq {\cal O}_{\mathbb{P}^{1}}(1)$ and $E|_{S'}\simeq {\cal O}_{\mathbb{P}^{1}}(1)$.
Hence $H'|_{S'}-E|_{S'}$ is linearly equivalent to $0$, in particular nef.
If $n-k\geq 2$, then $\pi |_{S'}:S'\to S=\mathbb{P}^{n-k}$ is the blow-up at the point $y_{0}$.
By Lemma \ref{point}, $H'|_{S'}-E|_{S'}=(\pi |_{S'})^{*}{\cal O}_{\mathbb{P}^{n-k}}(1)-\operatorname{Exc} (\pi |_{S'})$ is nef.
Since $L|_{S}\sim 0$, we have $L'|_{S'}\sim 0$.
Hence,
$[H'+dL'-E]|_{S'}=[H'-E]|_{S'}\in\operatorname{Nef}(S')$,
which implies
$$[\widetilde{H}+d\widetilde{L}-\widetilde{E}]|_{F}
=(\beta |_{F})^{*}\left([H'+dL'-E]|_{S'}\right)\in\operatorname{Nef} (F).$$
Since $[\widetilde{H}+d\widetilde{L}-\widetilde{E}-F]\in\operatorname{Nef}(\widetilde{X})$
as shown above,
we have by Lemma \ref{induction}
$[\widetilde{H}+d\widetilde{L}-\widetilde{E}]\in \operatorname{Nef}(\widetilde{X})$.
\qed
\
Remark: $H'-E$ itself is not nef because $(H'-E)|_{P_{0}'}\simeq {\cal O}_{\mathbb{P}^{k}}(-d)$.
\begin{proposition}\label{effective cone} We have
\begin{eqnarray*}
\operatorname{Eff} (\widetilde{X})
=\mathbb{R^{+}} [\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{E},F, \widetilde{D}].
\end{eqnarray*}
\end{proposition}
{\it Proof}. The divisors $\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{E},F,\widetilde{D}$ are all effective.
Hence, the inclusion
$\mathbb{R^{+}} [\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{E},F, \widetilde{D}] \subset \operatorname{Eff}(\widetilde{X})$ holds.
Let $A\subset \widetilde{X}$ be a prime divisor. Suppose $A\not\in\{ \widetilde{E},F, \widetilde{D}\}$.
Since $B:=(\pi\circ \beta)(A)$ is a prime divisor on $Y=\mathbb{P}^{n-k}\times \mathbb{P}^{k}$,
there exist $a,b\in \mathbb{Z}_{\geq 0}$ such that $B\sim aH+bL$.
Put $\mu:=\operatorname{mult}_{C}B$ and $\nu:=\operatorname{mult}_{S}B$. Since $A=(\pi\circ\beta)_{*}^{-1} B$, we have
$A\sim a\widetilde{H}+b\widetilde{L}-\mu\widetilde{E}-\nu F$.
If $q^{-1}(u)\subset B$ for any $u\in q(C)$,
we have $D=q^{-1}(q(C)))\subset B$, which implies $B=D$, a contradiction to the choice of $A$.
Hence, there exists a point $u\in q(C)$
such that $q^{-1}(u)\not\subset B$.
Put $Y_{u}:=q^{-1}(u)$ and $c_{u}:=C\cap Y_{u}$.
Note that there exists a line
$l_{u} \subset Y_{u}\simeq \mathbb{P}^{n-k}$
such that $c_{u}\in l_{u}$ and $l_{u}\not\subset B$
(if $n-k=1$, we put $l_{u}:=Y_{u}\simeq \mathbb{P}^{1}$).
Since $B\sim aH+bL$ and
$l_{u}\equiv l$, we have $B\cdot l_{u}=a$
(if $n-k=1$,
we have $B |_{Y_{u}}\cdot l_{u}=\deg (B |_{Y_{u}})$).
Therefore,
$$
a=B\cdot l_{u}=B |_{Y_{u}}\cdot l_{u}\geq \operatorname{mult}_{c_{u}}(B |_{Y_{u}})\geq \operatorname{mult}_{C}B=\nu.
$$
Since $B$ is a prime divisor on $Y=\mathbb{P}^{n-k}\times \mathbb{P}^{k}$, there exists a point $v\in \mathbb{P}^{n-k}$
such that $p^{-1}(v)\not\subset B$. Put $P_{v}:=p^{-1}(v)$ and $s_{v}:=S\cap P_{v}$.
Note that there exists a line $h_{v}\subset P_{v}\simeq \mathbb{P}^{k}$ such that
$s_{v}\in h_{v}$ and $h_{v}\not\subset B$. Since $B\sim aH+bL$ and $h_{v}\equiv h$, we have
$B\cdot h_{v}=b$. Hence,
$$
b=B\cdot h_{v}=B |_{P_{v}}\cdot h_{v}\geq \operatorname{mult}_{s_{v}}(B |_{P_{v}})\geq \operatorname{mult}_{S}B=\nu.
$$
Recall that $\widetilde{H}\sim \widetilde{H_{0}}+\widetilde{E}$ and $\widetilde{L}\sim \widetilde{L_{0}}+F$.
We conclude
$$
[A]=[a\widetilde{H}+b\widetilde{L}-\mu\widetilde{E}-\nu F]
=[a\widetilde{H_{0}}+b\widetilde{L_{0}}+(a-\mu)\widetilde{E}+(b-\nu)F]
\in \mathbb{R^{+}}[\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{E},F].
$$
\qed
Remark:
Since $\widetilde{D}\sim d\widetilde{L}-\widetilde{E}-F\sim d\widetilde{L_{0}}-\widetilde{E}+(d-1)F$,
we have
$[\widetilde{D}]\not\in\mathbb{R^{+}}[\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{E},F]$
for any $d\geq 2$.
If $d=1$,
we have
$\widetilde{L_{0}}\sim \widetilde{D}+\widetilde{E}$ and
$\mathbb{R^{+}}[\widetilde{H_{0}},\widetilde{L_{0}},\widetilde{E},F,\widetilde{D}]=\mathbb{R^{+}}[\widetilde{H_{0}},\widetilde{E},F,\widetilde{D}]$.
\section{Proof of Theorem}\label{proof}
We continue to use the notation (*). First, we find conditions for $-(K_{\widetilde{X}}+\Delta)$ to be ample
where $\Delta$ is a $\mathbb{Q}$-divisor on $\widetilde{X}$.
\begin{lemma}\label{inequalities}
Consider a divisor $\Delta=\alpha\widetilde{H}+\beta\widetilde{L}+\gamma\widetilde{E}+\delta F$ with $\alpha,\beta,\gamma,\delta\in\mathbb{Q}$.
Then, $-(K_{\widetilde{X}}+\Delta)$ is ample if and only if
\begin{eqnarray*}
\begin{cases}
\alpha+\gamma &<1 \\
\beta+d\gamma &<k+1-(n-k)d \\
-\gamma+\delta &<n-2k+1 \\
-\delta &<k-1.
\end{cases}
\end{eqnarray*}
\end{lemma}
{\it Proof}. Since $Y=\mathbb{P}^{n-k}\times \mathbb{P}^{k}$, we have
$-K_{Y}\sim (n-k+1)H+(k+1)L$.
By the adjunction formula for the blow-ups $\pi$ and $\beta$,
\begin{eqnarray*}
-K_{\widetilde{X}}
&\sim& \beta^{*}(-K_{X})-(k-1)F \\
&\sim& \beta^{*}(\pi^{*}(-K_{Y})-(n-k)E)-(k-1)F \\
&\sim& (n-k+1)\widetilde{H}+(k+1)\widetilde{L}-(n-k)\widetilde{E}-(k-1)F.
\end{eqnarray*}
Hence, $-(K_{\widetilde{X}}+\Delta)$ is linearly equivalent to
\begin{eqnarray*}
&&(n-k+1-\alpha)\widetilde{H}+(k+1-\beta)\widetilde{L}-(n-k+\gamma)\widetilde{E}-(k-1+\delta)F \\
&=&(1-\alpha-\gamma)\widetilde{H}+(k+1+dk-dn-\beta-d\gamma)\widetilde{L} \\
&&+(n-2k+1+\gamma-\delta)(\widetilde{H}+d\widetilde{L}-\widetilde{E})
+(k-1+\delta)(\widetilde{H}+d\widetilde{L}-\widetilde{E}-F).
\end{eqnarray*}
By Proposition \ref{nefcone}, we conclude that $-(K_{\widetilde{X}}+\Delta)$ is ample (i.e. the numerical equivalence class is an interior point of the nef cone) if and only if
\begin{eqnarray*}
\begin{cases}
1-\alpha-\gamma &> 0 \\
k+1+dk-dn-\beta-d\gamma &>0 \\
n-2k+1+\gamma -\delta &>0 \\
k-1+\delta &>0.
\end{cases}
\end{eqnarray*}
\qed
Remark: We obtain the condition for $-(K_{\widetilde{X}}+\Delta)$ to be nef if we replace ``$<$" by ``$\leq$".
\begin{lemma}\label{ample nef}
$-K_{\widetilde{X}}$ is ample (resp. nef) if and only if
\begin{eqnarray*}
\begin{cases}
0 &< k+1-(n-k)d \\
0 &< n-2k+1
\end{cases} \ \
\left( \mbox{resp.}
\begin{cases}
0 &\leq k+1-(n-k)d \\
0 &\leq n-2k+1
\end{cases}
\right).
\end{eqnarray*}
\end{lemma}
{\it Proof}. We put $\Delta=0$ in Lemma \ref{inequalities}. \qed
\
In view of Lemma \ref{effective cone}, it is natural that we should consider boundary divisors
in the form:
$$\Delta=x\widetilde{H_{0}}+y\widetilde{L_{0}}+z\widetilde{E}+wF+u\widetilde{D} \ \ \ (x,y,z,w,u\in\mathbb{Q}^{+}).$$
\begin{lemma}\label{log Fano pair}
$(\widetilde{X},\Delta)$ is a log Fano pair if and only if
\begin{eqnarray*}
\begin{cases}
0\leq x,y,z,w,u &<1\\
z-u &<1 \\
y-dx+dz &<k+1-(n-k)d \\
x-y-z+w &<n-2k+1 \\
y-w+u &<k-1.
\end{cases}
\end{eqnarray*}
\end{lemma}
{\it Proof}. We see that $\operatorname{Supp}\Delta=\widetilde{H_{0}}\cup\widetilde{L_{0}}\cup\widetilde{E}\cup F\cup\widetilde{D}$
is a simple normal crossing divisor. Hence, the pair $(\widetilde{X},\Delta)$ is klt if and only if
$0\leq x,y,z,w,u<1$.
On the other hand,
\begin{eqnarray*}
\Delta
&\sim&x(\widetilde{H}-\widetilde{E})+y(\widetilde{L}-F)+z\widetilde{E}+wF+u(d\widetilde{L}-\widetilde{E}-F) \\
&=&x\widetilde{H}+(y+du)\widetilde{L}+(-x+z-u)\widetilde{E}+(-y+w-u)F.
\end{eqnarray*}
If we put $\alpha:=x$, $\beta:=y+du$, $\gamma:=-x+z-u$ and $\delta:=-y+w-u$,
then,
$\alpha+\gamma=z-u$, $\beta+d\gamma=y-dx+dz$,
$-\gamma+\delta=x-y-z+w$ and $-\delta=y-w+u$. Thus, the statement follows from Lemma \ref{inequalities}. \qed
\
Now, we prove Theorem, which is divided into the following Propositions.
\begin{proposition}\label{big} The anti-canonical divisor $-K_{\widetilde{X}}$ is big for any $n\geq 3$, $k\in \{ 2,\cdots, n-1\}$ and $d\geq 1$.
\end{proposition}
{\it Proof}. Recall $-K_{\widetilde{X}}\sim (n-k+1)\widetilde{H}+(k+1)\widetilde{L}-(n-k)\widetilde{E}-(k-1)F$.
We consider the divisors
\begin{eqnarray*}
A&:=& \left(\frac{1}{2}+\frac{1}{d}\right)\widetilde{H}
+\left(1+\frac{1}{2d}\right)\widetilde{L}-\frac{1}{d}\widetilde{E}-\frac{1}{2d}F, \\
B&:=& \left(n-k+\frac{1}{2}-\frac{1}{d}\right)\widetilde{H}
+\left(k-\frac{1}{2d}\right)\widetilde{L}
-\left(n-k-\frac{1}{d}\right)\widetilde{E}-\left(k-1-\frac{1}{2d}\right)F.
\end{eqnarray*}
Then, we have $-K_{\widetilde{X}}\sim A+B$.
On the other hand,
\begin{eqnarray*}
A&\sim& \frac{1}{2}\widetilde{H}+\frac{1}{2d}\widetilde{L}
+\frac{1}{2d}\left(\widetilde{H}+d\widetilde{L}-\widetilde{E}\right)
+\frac{1}{2d}\left(\widetilde{H}+d\widetilde{L}-\widetilde{E}-F\right), \\
B&\sim& \left(n-k+\frac{1}{2}-\frac{1}{d}\right)\widetilde{H_{0}}
+\left(k-\frac{1}{2d}\right)\widetilde{L_{0}}+\frac{1}{2}\widetilde{E}+F.
\end{eqnarray*}
By Proposition \ref{nefcone}, $A$ is ample.
Since $n-1\geq k\geq 2$, we have $n-k+\frac{1}{2}-\frac{1}{d}>0$
and $k-\frac{1}{2d}>0$.
Hence, $B$ is effective. The statement follows from \cite{Lazarsfeldbook} Corollary 2.2.7. \qed
\
Remark: It follows that $\widetilde{X}$ is weak Fano if and only if
$-K_{\widetilde{X}}$ is nef.
\begin{proposition}\label{n=3}
Assume $n=3$ (hence $k=2$). Then, $\widetilde{X}$ is log Fano for any $d\geq 1$. Moreover
$\widetilde{X}$ is weak Fano if and only if $d\in \{1,2,3\}$, and is not Fano for any $d\geq 1$.
\end{proposition}
{\it Proof}. If $d=1$, we put $z=\frac{1}{2}$ and $x=y=w=u=0$.
If $d\geq 2$, we put
\begin{eqnarray*}
x=y=\frac{d-2}{d-1}, \
z=\frac{1}{2d} \
\mbox{ and } \
w=u=0.
\end{eqnarray*}
Then, the inequalities in Lemma \ref{log Fano pair} are all satisfied.
Hence, $\widetilde{X}$ is log Fano for any $d\geq 1$.
By Lemma \ref{ample nef},
$-K_{\widetilde{X}}$ is nef if and only if $d\in \{1,2,3\}$, and is not ample
for any $d\geq 1$. \qed
\
From now on, we assume $n\geq 4$.
\begin{proposition}\label{weakFano}
$\widetilde{X}$ is weak Fano if and only if
\begin{eqnarray*}
(n,k,d)=(2k-1,k,1),(2k,k,1),(2k+1,k,1) \ \mbox{or} \ (5,3,2).
\end{eqnarray*}
$\widetilde{X}$ is Fano if and only if $(n,k,d)=(2k,k,1)$.
\end{proposition}
{\it Proof}. By Lemma \ref{ample nef} and the remark to Proposition \ref{big},
$\widetilde{X}$ is weak Fano if and only if
\begin{eqnarray*}
\begin{cases}
dn & \leq (d+1)k+1 \\
2k-1 & \leq n.
\end{cases}
\end{eqnarray*}
This yields $d(2k-1)\leq dn \leq (d+1)k+1.$
In particular, we have $d(2k-1)\leq (d+1)k+1$, which implies
\begin{eqnarray*}
d\leq \frac{k+1}{k-1}=1+\frac{2}{k-1}\leq 1+2=3.
\end{eqnarray*}
If $d=1$, we have $2k-1\leq n \leq 2k+1$.
If $d=2$, we have $4k-2\leq 2n \leq 3k+1$.
Since $n\geq 4$, we have $8\leq 3k+1$, i.e. $k\geq 3$.
On the other hand, we have $4k-2\leq 3k+1$, i.e. $k\leq 3$.
Thus, $k=3$ and $n=5$.
If $d=3$, we have $6k-3\leq 3n\leq 4k+1$.
Since $n\geq 4$, we have $12\leq 4k+1$, i.e. $k\geq 3$.
On the other hand, we have $6k-3\leq 4k+1$, i.e. $k\leq 2$, a contradiction.
Therefore, we have
\begin{eqnarray*}
(n,k,d)=(2k-1,k,1),(2k,k,1),(2k+1,k,1) \ \mbox{or} \ (5,3,2).
\end{eqnarray*}
In any of these cases, $k+1-(n-k)d$ and $n-2k+1$ are non-negative, and strictly positive only for the case $(n,k,d)=(2k,k,1)$. \qed
\begin{proposition}\label{log Fano}
There exist rational numbers $x,y,z,w,u$ such that
$(\widetilde{X},\Delta=x\widetilde{H_{0}}+y\widetilde{L_{0}}+z\widetilde{E}+wF+u\widetilde{D})$ is a log Fano pair
if and only if one of the following holds:
\begin{itemize}
\item $n\in\{2k-2,2k-1,2k,2k+1\}$ and $d=1$,
\item $(n,k,d)=(4,2,2),(4,3,2) \mbox{ or } (5,3,2)$.
\end{itemize}
\end{proposition}
{\it Proof}. Assume that $(\widetilde{X},\Delta)$ is a log Fano pair.
By Lemma \ref{log Fano pair}, we have $0\leq x,y,z,w<1$ and
\begin{eqnarray}\label{logfano1}
\begin{cases}
y-dx+dz &<k+1-(n-k)d \\
x-y-z+w&<n-2k+1.
\end{cases}
\end{eqnarray}
We have
$-d<-dx\leq y-dx+dz<k+1-nd+kd$, i.e. $nd<(k+1)(d+1)$.
On the other hand, $-2<-y-z\leq x-y-z+w<n-2k+1$, i.e. $2k-3<n$.
Thus,
\begin{eqnarray}\label{logfano2}
2k-3<n<\frac{(k+1)(d+1)}{d}.
\end{eqnarray}
If $d=1$, we have $2k-3<n<2k+2$, i.e. $n\in\{ 2k-2,2k-1,2k,2k+1\}$.
Consider the case $d\geq 2$. Since $w\geq 0$,
we have $x-y-z\leq x-y-z+w$.
Hence, by (\ref{logfano1}), we obtain:
\begin{eqnarray*}
\begin{cases}
y-d(x-z) &<k+1-nd+kd \\
-y+(x-z) &<n-2k+1.
\end{cases}
\end{eqnarray*}
The sum of two inequalities gives $(d-1)(n-k)-2<(d-1)(x-z)$,
i.e. $n-k-\frac{2}{d-1}<x-z$.
On the other hand, we have
$x-z<n-2k+1+y<n-2k+2$. Therefore, $n-k+\frac{2}{1-d}<n-2k+2$, i.e.
\begin{eqnarray}\label{logfano3}
k<2+\frac{2}{d-1}.
\end{eqnarray}
Since $d,k\geq 2$, this implies $k=2$ or $3$.
If $k=2$, by (\ref{logfano2}), we have $n<\frac{3(d+1)}{d}$. Thus, $n=4$ and $d=2$.
If $k=3$, by (\ref{logfano3}), we have $d=2$. This implies $3<n<6$
by (\ref{logfano2}). Thus, $n=4$ or $5$.
Therefore, we have
\begin{eqnarray*}
(n,k,d)=(4,2,2),(4,3,2) \mbox{ or } (5,3,2).
\end{eqnarray*}
In each case, we put $(x,y,z,w,u)$ as follows:
\begin{center}
\begin{tabular}{l|c}
\ \ \ \ \ \ $(n,k,d)$ & $(x,y,z,w,u)$ \\
\hline
Case: $n\geq 4, d=1$ & \\
\ \ $n=2k-2$ & $(0,\frac{1}{2},\frac{3}{4},0,0)$ \\
\ \ $n=2k-1$ & $(0,\frac{1}{2},0,0,0)$ \\
\ \ $n=2k$ & $(0,0,0,0,0)$ \\
\ \ $n=2k+1$ & $(\frac{1}{2},0,0,0,0)$ \\
\hline
$(n,k,d)=(4,2,2)$ & $(\frac{3}{4},0,0,0,0)$\\
$(n,k,d)=(4,3,2)$ & $(\frac{1}{8},\frac{1}{2},\frac{3}{4},0,0)$\\
$(n,k,d)=(5,3,2)$ & $(\frac{1}{2},\frac{1}{2},\frac{1}{8},0,0)$
\end{tabular}
\end{center}
Then, the inequalities in Lemma \ref{log Fano pair} are satisfied. \qed
|
\section{Introduction}
Liquid rare-gas scintillation detectors are employed extensively in many different fields such as particle and nuclear physics and medical sciences.
The major difficulty with the usage of liquid rare-gas as the scintillator medium
is the detection of the scintillation light in the vacuum ultra violet (VUV) range;
the emission peak at $\lambda=175\,\mathrm{nm}$ for liquid xenon (LXe)~\cite{Nakamura:LXeEmissionSpectrum},
$147\,\mathrm{nm}$ for liquid krypton (LKr)~\cite{Aprile:2008bga} and
$128\,\mathrm{nm}$ for liquid argon (LAr)~\cite{Aprile:2008bga}.
VUV enhanced photomultiplier tubes (PMTs) or PMTs with wavelength shifting materials have been often
employed to detect the liquid rare-gas scintillation light.
The silicon photomultiplier (SiPM) is a rapidly emerging photosensor technology
as a possible replacement of the PMT.
The SiPM is now a major candidate for the photosensor of
scintillation detectors.
The LXe scintillation detector for the MEG II experiment~\cite{MEG-design-paper}
is designed to precisely measure the energy, position and time of
the 52.8\,MeV photon from $\megsign$, based on $900\,\ell$ LXe with a high-granularity scintillation readout by 4092 SiPMs.
A large-area Multi-Pixel Photon Counter (MPPC)
\footnote{MPPC is the product name of the SiPM produced by Hamamatsu Photonics K.K.}
with an enhanced VUV sensitivity has been
developed for the LXe detector
since the currently commercially available SiPMs are not sensitive to the VUV light.
A prototype model of the VUV-MPPC had been successfully tested
and shown an excellent performance in LXe as previously reported in Ref.~\cite{Ootani:2015cia}.
A production model
to be used for the LXe detector
has been developed where the main improvement is a significant reduction of the optical cross-talk.
This article describes in detail the design of the VUV-MPPC and the performance measured in LXe mainly for the production model.
The results from the performance tests at room temperature for the 4180 VUV-MPPCs produced for the LXe detector are also presented.
\section{Design of VUV-enhanced MPPC}
\label{sec:Design of VUV-enhanced MPPC}
Good resolutions of the MEG II LXe detector for the 52.8\,MeV photon from $\megsign$
can be achieved by a highly granular and uniform scintillation readout
with VUV-sensitive MPPCs.
The front side of the detector of $0.92\,\mathrm{m}^2$ on which
the signal photon impinges is tiled with 4092 VUV-MPPCs
with a total active area of $0.57\,\mathrm{m}^2$.
The major requirements for the VUV-MPPC of the MEG II experiment
can be summarised as
the photon detection efficiency (PDE) for the LXe scintillation light higher than 10\%,
the gain higher than $5\times 10^5$,
the probability of the optical cross-talk and after-pulsing lower than 15\% each,
the dark count rate lower than $10\,\mathrm{Hz}/\mathrm{mm}^2$ at LXe temperature
and the capability of single photoelectron resolution.
Fig.\,\ref{fig:S10943-4372} shows the photograph and
the schematic view of the production model of the VUV-MPPC
(Hamamatsu Photonics K.K. S10943-4372).
It is a discrete array of four bare MPPC chips, each of which has an active area
of $5.95\times 5.85\,\mathrm{mm}^2$ with a micro-cell pitch of $50\,\mu\mathrm{m}$.
The four chips are glued on a ceramic base
with low-temperature-resistant conductive adhesive with a spacing of $0.5\,\mathrm{mm}$
and can be readout individually.
In the MEG II LXe detector, the four chips are connected in series in a readout printed circuit board (PCB)
as described in Sec.\,\ref{sec:Performance of Large-area MPPC} to form a single sensor
with a total active area of $139\,\mathrm{mm}^2$
\footnote{
``Chip'' and ``sensor'' shall hereafter mean the single MPPC chip
with an active area of $5.95\times 5.85\,\mathrm{mm}^2$
and the sensor package composed of the four MPPC chips, respectively,
unless otherwise noted.}.
The outline dimension of the sensor is
$15\,\mathrm{mm}\times 15\,\mathrm{mm}\times 2.5\,\mathrm{mm}$.
The four chips in the sensor are covered by a $0.5\,\mathrm{mm}$-thick VUV-transparent quartz window for protection,
which is a non-hermetic window with thin slits at the two sides of the package.
The thin gap of $0.5\,\mathrm{mm}$ between the window and the chips
is filled with LXe when the sensor is operated in LXe.
The refractive index of LXe at the emission peak
of the LXe scintillation light is $1.64$~\cite{NakamuraLXeRefractiveIndex},
which is close to that of the quartz window ($1.60$).
The reasonable matching of the refractive indices helps to minimise an
undesirable reflection at the boundary between the quartz window and LXe.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=6cm]{VUVMPPC_photo.jpg}
\includegraphics[width=7cm]{mppc.pdf}
\end{center}
\caption{Large-area VUV-sensitive MPPC developed for the MEG II LXe detector
(Hamamatsu Photonics K.K., S10943-4372). }
\label{fig:S10943-4372}
\end{figure}
The sensitivity to the VUV light of the commercial SiPM is quite low
since the VUV light is absorbed in a thick protection coating on top of the SiPM chip,
usually made of epoxy resin.
Even without the protection coating,
the VUV light is absorbed in the top contact layer of the SiPM chip
before reaching the active layer
due to the extremely short attenuation length of the LXe scintillation light
in silicon of about 5\,nm.
The high sensitivity of the VUV-MPPC to the LXe scintillation light
is achieved by removing the protection coating,
optimising the optical matching between the LXe and the chip surface,
and thinning down the top contact layer.
The contact layer is, however, still
much thicker than the attenuation length of the LXe scintillation light.
The photo-generated charge carriers in the contact layer
diffuse and a part of the charge carriers can reach the active layer
and trigger the avalanche amplification.
There are a few other improvements for the enhanced VUV-sensitivity.
The dopant concentration in the contact layer is adjusted
to have low but non-zero electric field,
causing a moderate drift of the charge carriers to the active layer.
The processing of the contact layer is optimised to minimise the lattice defects
where the charge carriers are trapped.
More than one electron-hole pairs can be generated by a single VUV photon
since the energy of the VUV photon is twice higher
than the average energy required to produce an electron-hole pair,
which is estimated to be 3.65\,eV~\cite{Kowalski-Nuclear-Electronics}.
That also contributes to the high VUV-sensitivity of the VUV-MPPC.
The correlated noises of the MPPC such as the optical cross-talk and the after-pulsing
increase the excess noise factor and thus, worsen the resolutions.
The after-pulsing was already significantly suppressed in the prototype model
as reported in Ref.~\cite{Ootani:2015cia}.
The optical cross-talk has also been greatly reduced
in the production model described in this article
by introducing a trench structure
between the adjacent cells to prevent the luminescence photons
generated in the avalanche in the primary cell
from firing the neighbouring cells.
The substantial suppression of both the correlated noises enables one
to operate it
at a higher over-voltage
\footnote{The over-voltage ($\Delta V$) is defined as the excess voltage
with respect to the breakdown voltage ($V_\mathrm{bd}$).}
up to 7--8\,V as compared with the prototype model,
which improves the PDE, the gain and the temperature stability.
The VUV-MPPC employs metal-based quench resistors
with a much smaller temperature coefficient
than that for the polysilicon-based quench resistors
used in the previous models of the VUV-MPPC.
The resistance of the quench resistor at LXe temperature of 165\,K
was found to be increased only by 20\% compared to that at room temperature.
The pulse-decay-time is, therefore, kept reasonably short at LXe temperature.
\section{Performance of VUV-MPPC}
\label{sec:Performance of VUV-MPPC}
The performance of the VUV-MPPC was measured in detail at LXe temperature
for several sensors.
In addition, operational tests were done at room temperature
for all the produced sensors before the installation to the detector.
\subsection{Setup of Performance Measurements in LXe}
\label{sec:Setup of Performance Measurement in LXe}
Fig.\,\ref{fig:Small chamber setup} shows a typical setup used
for the detailed performance measurements in LXe for several sensors.
The whole setup was housed in an inner vessel of a low temperature cryostat
where $2\,\ell$-LXe is liquefied with a pulse tube refrigerator (Iwatani PDC08).
A few sensors were mounted on each top and bottom stage
facing a spot $\alpha$-source on a thin tungsten wire stretched in between.
The spot $\alpha$-source was developed
for the MEG experiment~\cite{Baldini:2006}.
A $1.5\,\mu\mathrm{m}$-thick gold foil,
in which a radioactive matrix of Am-241 in a chemical form of Americium Oxide
with an activity of about 100\,Bq is homogeneously incorporated,
is wound around a thin gold-plated tungsten wire
with a diameter of $100\,\mu\mathrm{m}$
and is covered by a 1.5\,$\mu\mathrm{m}$-thick gold foil.
The linear dimension of the spot source is about $500\,\mu\mathrm{m}$.
The energy of the $\alpha$-rays emitted from the spot source was measured in vacuum
with a silicon surface barrier detector in a different setup.
The measured energy is 4.8\,MeV with a spread of 4.0\% (in sigma)
\footnote{The measured energy spectrum has a low energy tail and the spread was evaluated
at the right side of the peak.},
which is lower than the initial energy of $\alpha$-rays emitted from Am-241 of 5.5\,MeV
due to the energy loss in the gold foil surrounding the source.
The sensors on the top and bottom stages were illuminated
by the LXe scintillation light generated by the spot $\alpha$-source,
which can be considered as a point light source
because of the range of $\alpha$-ray in LXe as short as $40\,\mu\mathrm{m}$.
A few blue LEDs (OSA Opto Light, OCU-400 UE390)
were also installed in the setup for a calibration purpose.
The LXe volume was surrounded by a cylindrical wall
with a VUV anti-reflecting coating (Acktar black~\cite{AcktarBlack})
to minimise the influence of the reflection
of the scintillation light from the $\alpha$-source
in the PDE measurement.
A reliable prediction of the number of scintillation photons
impinging on the sensors is required in the PDE measurement as described later.
The signal from the sensor was transmitted through a vacuum feedthrough
to an amplifier placed outside the cryostat and then to a waveform digitizer (DRS4~\cite{Ritt:2010zz}).
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=7cm]{SmallChamberSetup.pdf}
\end{center}
\caption{Typical set-up for the performance measurements in LXe.}
\label{fig:Small chamber setup}
\end{figure}
\subsection{Performance of Single Chip in LXe}
\label{sec:Performance of Single MPPC Chip}
The detailed properties of the VUV-MPPC were measured in LXe
for individual chips of $5.95\times 5.85\,\mathrm{mm}^2$.
Fig.\,\ref{fig:MPPC signal 6x6} shows a typical waveform
for the single photoelectron signal.
The rise and decay times of the single photoelectron signal
were measured by fitting exponential functions to be $2.8\,\mathrm{ns}$ and $86\,\mathrm{ns}$, respectively.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{Waveform6x6.pdf}
\end{center}
\caption{Typical single photoelectron signal from a single chip
of $5.95\times 5.85\,\mathrm{mm}^2$ operated in LXe.}
\label{fig:MPPC signal 6x6}
\end{figure}
The charge distribution obtained with a single chip operated
at $\Delta V=7\,\mathrm{V}$ for low-level light pulses
from the blue LED is shown
in Fig.\,\ref{fig:MPPC single photoelectron distribution 6x6}.
The peaks for successive numbers of photoelectrons are clearly resolved.
The tail at the right side of each photoelectron peak comes from the pileup due to after-pulsing.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{charge_6mm_run1709.pdf}
\end{center}
\caption{
Charge distribution obtained for low-level LED light pulses
with the single MPPC chip of $5.95\times 5.85\,\mathrm{mm}^2$
operated in LXe at $\Delta V=7\,\mathrm{V}$.
}
\label{fig:MPPC single photoelectron distribution 6x6}
\end{figure}
The dark count rate as a function of the over-voltage was measured at LXe temperature
by counting the random single photoelectron events
as shown in Fig.\,\ref{fig:DCR vs OV}.
The scintillation light events caused by environmental radio-activity and cosmic-rays
were vetoed by eliminating coincident events over multiple chips.
Some of the scintillation light events with low light level, however, were left unvetoed
because of the non-negligible probability of zero photoelectron events.
Furthermore, the measured rate includes the single photoelectron events due to the after-pulsing.
The typical dark count rate was estimated to be $5\,\mathrm{Hz}/\mathrm{mm}^2$ at $\Delta V = 7\,\mathrm{V}$,
although it was somewhat overestimated for the reasons mentioned above.
This extremely low dark count rate, which is about five orders of magnitude lower than the rate at room temperature,
enables one to operate the VUV-MPPC
with a good single photoelectron resolution
even with a large area of $139\,\mathrm{mm}^2$ as described in Sec.\,\ref{sec:Performance of Large-area MPPC}.
The probabilities of the correlated noises
were evaluated with low-level LED light pulses
by means of the method described in Ref.~\cite{T2K:NIMA2009}.
Firstly, the probability of the true one photoelectron event
without the correlated noises is estimated from the probability
of the zero photoelectron event assuming the Poisson distribution.
The observed probability of the one photoelectron event
is lower than estimated from the zero photoelectron probability
because some of the true one photoelectron events are observed
as the events with two or more photoelectrons
due to the overlap of the correlated noises.
The probabilities of the correlated noises
are then calculated from the difference
between the observed and estimated probabilities of the one photoelectron event.
The overall probability of the optical cross-talk and the after-pulsing
is estimated with a charge integration time of 150\,ns
which is long enough to contain most of the after-pulsing with a delayed timing,
while the probability of the optical cross-talk is separately estimated
with a shorter integration time of 30\,ns.
Fig.\,\ref{fig:cross-talk} and Fig.\,\ref{fig:after-pulsing}
show the measured probabilities
of the optical cross-talk and the after-pulsing
as a function of the over-voltage, respectively.
In Fig.\ref{fig:cross-talk} the cross-talk measured for the prototype model is also shown for comparison.
The cross-talk is drastically reduced for the production model,
which enables one to operate it at a much higher over-voltage up to 7--8\,V,
where the overall probability of the correlated noises is still as low as $30$\%.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{DCR.pdf}
\end{center}
\caption{
Measured dark count rate as a function of the over-voltage
for two different single MPPC chips operated in LXe.
}
\label{fig:DCR vs OV}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{CTvsOV.pdf}
\end{center}
\caption{
Measured probability of the optical cross-talk
as a function of the over-voltage (red filled circle).
The probability for the prototype model
is also shown for comparison (black open circle).
}
\label{fig:cross-talk}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{APvsOV.pdf}
\end{center}
\caption{
Measured probability of the after-pulsing
as a function of the over-voltage.
}
\label{fig:after-pulsing}
\end{figure}
The PDE of the VUV-MPPC was measured
as the ratio of the observed number of photoelectrons
to the expected number of photons impinging on the active area
of the MPPC chip.
The contributions from the correlated noises are subtracted from the observed number of photoelectrons
using the measured probabilities.
The expected number of photons was evaluated from
the solid angle subtended by the active area of the chip to the spot $\alpha$-source
assuming the scintillation light yield of the $\alpha$-ray of
$(51\pm5)\times10^3\,\mathrm{photons}/\mathrm{MeV}$~\cite{Doke:1999ku}.
The contribution from the reflected light on the surrounding wall
with the VUV anti-reflecting coating was evaluated to be negligibly small,
while the reflection on the surface of the gold foil
surrounding the spot $\alpha$-source was found to be non-negligible
and taken into account in the estimation of the PDE.
Fig.\,\ref{fig:PDE} shows the measured PDE
for two chips as a function of the over-voltage
where the systematic error of the measurement comes mainly from
the error in the scintillation light yield
for $\alpha$-rays (10\%~\cite{Doke:1999ku})
and the uncertainty in the estimation of the effect of the reflection
on the surface of the spot $\alpha$-source (5\%).
The PDE was measured to be 14--21\% depending on the over-voltage,
which sufficiently fulfils
the requirement for the MEG II LXe detector.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{mppc_pde.pdf}
\end{center}
\caption{
Measured PDE for two VUV-MPPC chips as a function of the over-voltage.
The error bars and the strips represent the statistical and the systematic errors, respectively.
The contributions from the correlated noises are corrected for using the measured probabilities.
}
\label{fig:PDE}
\end{figure}
\subsection{Performance of Large-area Sensor in LXe}
\label{sec:Performance of Large-area MPPC}
While a highly granular scintillation readout is required for the MEG II LXe detector,
the maximum size of the active area of the standard MPPC
($6\times 6\,\mathrm{mm}^2$)
is even too small to manage the corresponding enormous number of readout channels.
The optimal sensor size for the LXe detector is found to be about $10\times 10\,\mathrm{mm}^2$.
The large sensor capacitance of 5\,nF expected for that sensor size will be an issue due to
the resultant long pulse-decay-time.
The signal readout by transimpedance amplifiers with a low input-impedance to mitigate the long pulse-decay-time
is not possible for the LXe detector because of the non-negligible impedance of the signal cable as long as 12\,m
between the sensor and the amplifier.
The long cable is required to transmit the signal
from the inside of the detector cryostat to the readout electronics placed outside the cryostat.
A single sensor with a large active area of $139\,\mathrm{mm}^2$ is, therefore, formed
by connecting four independent small MPPC chips ($5.95\times 5.85\,\mathrm{mm}^2$ each) in series
to avoid the increase in the sensor capacitance.
The scheme of the series connection, which is made on a signal readout PCB underneath the sensor,
is illustrated in Fig.\,\ref{fig:series connection}.
The four chips are connected in series with decoupling capacitors in between
such that the four chips are biased by a common voltage.
This connection scheme is more advantageous than the simple series connection.
The bias voltage is the same as the one for the single chip
in contrast to the four times higher bias voltage required
for the simple series connection.
There is no voltage difference between the adjacent chips,
which eliminates the risk of the discharge between the adjacent chips.
Four chips with almost the same breakdown voltage are selected
to assemble each sensor in the production by the company
such that the four chips within each sensor have the same gain when biased by a common voltage.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=7cm]{MPPCSerialConnectionHybridOnly.pdf}
\end{center}
\caption{
Scheme of the series connection of the four MPPC chips on the VUV-MPPC.
}
\label{fig:series connection}
\end{figure}
The typical waveform for the signal with two photoelectrons
(primary photoelectron overlapped with optical cross-talk)
from the four chips connected in series is shown
in the bottom of Fig.\,\ref{fig:MPPC signal 12x12}.
The waveform from the four chips connected in parallel, which is equivalent
to a single large sensor of $139\,\mathrm{mm}^2$, is also shown for comparison
in the top of Fig.\,\ref{fig:MPPC signal 12x12}.
The pulse-decay-time for the series connection was measured
to be $25\,\mathrm{ns}$,
which is five times shorter than that for the parallel connection.
Fig.\,\ref{fig:MPPC single photoelectron distribution 12x12}
shows the charge distribution obtained for low-level light pulses from the blue LED
for the four chips connected in series operated in LXe at $\Delta V=7\,\mathrm{V}$.
The photoelectron peaks are clearly resolved
with a single photoelectron resolution as good as 12\%.
Fig.\,\ref{fig:MPPC gain 12x12} shows
the gain computed from the single photoelectron charge measured
as a function of the over-voltage.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{Waveform_4p2pe.pdf}
\includegraphics[width=8cm]{Waveform_4s2pe.pdf}
\end{center}
\caption{
Typical waveforms for the signal with two photoelectrons
(primary photoelectron overlapped with optical cross-talk)
from the VUV-MPPC with the four chips connected
in parallel (top) and in series (bottom).
The VUV-MPPCs were operated in LXe.
}
\label{fig:MPPC signal 12x12}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{charge_12mm_run2001.pdf}
\end{center}
\caption{
Charge distribution for a low-level LED light
for the VUV-MPPC with the four chips connected in series,
operated in LXe at $\Delta V=7\,\mathrm{V}$.
}
\label{fig:MPPC single photoelectron distribution 12x12}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{GainvsOV.pdf}
\end{center}
\caption{
Gain for the VUV-MPPC with the four chips connected in series,
operated in LXe as a function of the over-voltage.
}
\label{fig:MPPC gain 12x12}
\end{figure}
The energy resolutions of the large-area VUV-MPPC were measured
for visible light and LXe scintillation light.
The sensor was illuminated by light pulses from a blue-LED with a constant intensity
and the resolution was measured as the standard deviation
of the Gaussian function fitted to the charge distribution.
The measurement was repeated for different light intensities
by varying the voltage to the LED to study
the dependence of the resolutions on the number of photoelectrons.
The sensor was operated at a low over-voltage of 1.5\,V
to eliminate the effect of the worsening
of the resolution due to the correlated noises.
Fig.\,\ref{fig:energy resolution visible light} shows the measured resolutions
as a function of the detected number of photoelectrons ($N_\mathrm{p.e.}$)
where the contribution expected from the Poisson statistics ($1/\sqrt{N_{\mathrm{p.e.}}}$) is also shown for comparison.
The measured resolution approaches the Poisson statistics with a large number of photoelectrons.
The deviation from the Poisson statistics with a smaller number of photoelectrons
is considered to be due to the instability of the light intensity of the LED.
The energy resolution for the LXe scintillation light was measured
using the spot $\alpha$-source.
The prototype model of the VUV-MPPC was used in the measurements.
As described in Sec.\,\ref{sec:Performance of VUV-MPPC},
the spread of the energy of $\alpha$-ray emitted from the spot source
is too large to estimate the energy resolution from the spread of the charge spectrum for $\alpha$-ray events.
The resolution was, therefore, evaluated from the distribution
of the difference in the numbers of photoelectrons observed
by two groups, each of which consists
of one or two sensors placed at the same distance
from the $\alpha$-source to subtract the effect
of the energy spread of the emitted $\alpha$-rays.
The relative resolution is estimated
as $\sigma_{N_1-N_2}/PV_{N_1+N_2}$ and $\sigma_{N_1+N_2-N_3-N_4}/PV_{N_1+N_2+N_3+N_4}$
for the cases of two and four sensors, respectively,
where $N_i$ is the number of photoelectrons observed by the $i$-th sensor for the same $\alpha$-ray event,
$\sigma_N$ is the standard deviation of the Gaussian function fitted to the distribution of $N$ and
$PV_N$ is the peak value of the distribution of $N$.
The measurements were performed with several sensors placed
at different distances from the $\alpha$-source
to study the dependence of the resolutions on the number of photoelectrons.
The sensors were operated at a low over-voltage of 1.5\,V
to minimise the effect of the correlated noise.
Fig.\,\ref{fig:energy resolution VUV} shows the measured energy resolutions
as a function of the total number of photoelectrons observed
by the two or four sensors used in the estimation of the resolution.
The energy resolution for the LXe scintillation light was found to be worse
than the Poisson statistics similarly to that for the visible light
but the deviation from the Poisson statistics is larger than that for the visible light.
The larger deviation is not understood yet,
but could be related to the special detection process for the VUV light
as described in Sec.\,\ref{sec:Design of VUV-enhanced MPPC}
which is quite different from that for the visible light.
The time resolution for the LXe scintillation light was estimated
from the time difference between the two sensors placed at the same
distance from the $\alpha$ source
where about 2000 photoelectrons were observed on average.
The time of the MPPC signal is computed by a waveform analysis
with the constant fraction method to eliminate the time walk effect,
where the signal time is defined as the leading edge reaches a fixed fraction of the pulse height.
Fig.\,\ref{fig:timing resolution} shows
the measured time resolutions for different over-voltages
as a function of the fraction threshold in the constant fraction method.
The time resolution improves for higher over voltages because of the higher signal-to-noise ratio and the higher PDE.
For each over-voltage, the best time resolution is obtained
at a different fraction threshold.
The best overall time resolution of about 40\,ps
is obtained at $\Delta V = 3\,\mathrm{V}$ with a fraction threshold of 5\%
and is good enough for the MEG II LXe detector.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{EresVsStat.pdf}
\end{center}
\caption{
Measured energy resolutions for visible light pulses from a blue-LED as a function of the detected number of photoelectrons.
The contribution expected from the Poisson statistics ($1/\sqrt{N_{\mathrm{p.e.}}}$) is also shown as a red solid line for comparison.
}
\label{fig:energy resolution visible light}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{EresvsNphe1.pdf}
\end{center}
\caption{
Measured energy resolution for the LXe scintillation light
as a function of the total number of photoelectrons observed
by the two or four sensors used in the estimation of the resolution.
The contribution expected from the Poisson statistics ($1/\sqrt{N_{\mathrm{p.e.}}}$) is also shown as a red solid line for comparison.
}
\label{fig:energy resolution VUV}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{TresvsThr.pdf}
\end{center}
\caption{
Measured time resolutions for the LXe scintillation light as a function of the fraction threshold in the constant fraction method.
}
\label{fig:timing resolution}
\end{figure}
The potential issue in the scheme of the series connection
shown in Fig.\,\ref{fig:series connection} is its moderate high-rate performance.
The bias voltage can vary due to the voltage drop in the resistors in the series connection
and in the protection resistor of the bias circuit
with the increased sensor current in a high rate environment.
The rate-dependent variation of the bias voltage would cause
an instability of the sensor gain.
The gain variation was measured in LXe by monitoring the signal charge for light pulses from a blue LED
with an illumination of a background light from another blue LED,
which simulates the background light in a high rate environment.
The gain was measured at different intensities of the background light
by changing the frequency of the light pulses.
The overall resistance of $23.6\,\mathrm{k}\Omega$
(two resistors of $10\,\mathrm{k}\Omega$
each in the series connection and one resistor of $3.6\,\mathrm{k}\Omega$ in the bias circuit)
and a decoupling capacitance of 10\,nF were used in this measurement.
Fig.\,\ref{fig:Gain variation with BG light} shows the relative charge of the primary signal
as a function of the increase of the sensor current due to the background light,
where the gain variation was measured to be $0.8\%/\mu\mathrm{A}$.
The average increase of the sensor current
due to the background light expected
in the MEG II experiment is about $2\,\mu\mathrm{A}$ and thus,
the gain degradation is estimated to be only a few percent.
A larger variation of the background light would, however, be caused
by an unexpectedly large instability of the muon beam used in the experiment.
A smaller resistance ($2\,\mathrm{k}\Omega$
each in the series connection)
is, therefore, adopted in the final version of the readout PCB
as shown in Fig.\,\ref{fig:series connection} to minimise the voltage drop.
The resistance for the final version of the bias circuit is also supposed to be smaller
although an optimal resistance is still under study because the voltage across the resistance
is also used to monitor the sensor current.
A larger decoupling capacitance (22\,nF) is used
to suppress the signal leakage to the smaller resistance.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{highRate.pdf}
\end{center}
\caption{
Relative gain measured in LXe as a function of the sensor current increased by
background light.
}
\label{fig:Gain variation with BG light}
\end{figure}
\subsection{Test of Sensors Produced for MEG II LXe Detector at Room Temperature}
We produced 4200 VUV-MPPCs including 108 spares for the MEG II LXe detector
and tested 4180 sensors at room temperature.
The test includes measurements of the current-voltage characteristic ($I$-$V$ curve)
for the individual 16720 ($=4180\times4$) chips and
measurements of the relative signal gain for visible light and the waveform
for the 4180 sensors.
The breakdown voltage and the dark current were estimated from the $I$-$V$ curve.
Only a small fraction of the sensors
(0.2\% of all the chips and 0.8\% of all the sensors)
show a slightly odd behaviour near the breakdown voltage in the $I$-$V$ curve such as a small current offset
and an unusual shape of the curve
although most of them can still be operational
in the LXe detector.
The breakdown voltage was estimated by fitting the $I$-$V$ curve
with a linear and a parabolic functions before and after
the breakdown voltage, respectively.
Fig.\,\ref{fig:Mass test Vbd} shows the distribution of the measured breakdown voltages for the 16720 chips,
which ranges from 50.5\,V to 52.5\,V.
Although the spread of the breakdown voltages is not very small,
it is not an issue
since the four chips with almost the same breakdown voltage
were selected for the production of each sensor.
In fact, the spread of the breakdown voltages for the four chips within each sensor,
which is defined as the difference between the maximum and minimum, was measured to be
only a few hundred mV as shown in Fig.\,\ref{fig:Mass test Vbd max-min}.
This spread of the breakdown voltage within each sensor
corresponds to a gain variation of only a few percent over the four chips
when they are operated at a common over-voltage of 7\,V.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{vbd.pdf}
\end{center}
\caption{
Distribution of the measured breakdown voltage for the 16720 MPPC chips.
}
\label{fig:Mass test Vbd}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{vbd_diff.pdf}
\end{center}
\caption{
Distribution of the spread of the breakdown voltages
for the four MPPC chips within each MPPC.
}
\label{fig:Mass test Vbd max-min}
\end{figure}
The response to short light pulses from a blue LED was measured
for all the sensors mounted on PCBs with the circuitry
for the series connection as shown in Fig.\,\ref{fig:series connection}.
The relative signal gain for the LED light pulses was measured by taking a ratio
of the output charge to that observed for a reference MPPC.
The black filled circle
in Fig.\,\ref{fig:relative gain and pulse decay time at mass test}
shows the relative signal gain measured at the same over-voltage
as a function of the serial number of the VUV-MPPC
where a variation of 20\% and a clear dependence on the production lot are seen.
The sensor-by-sensor variation of the signal gain
is, however, considered to be due mainly to the difference
in the probabilities of the correlated noises instead of the difference in the PDE.
The pulse-decay-time, which can vary depending on the probability
of the after-pulsing, was also measured to verify this hypothesis.
In fact the same production lot dependence as in the relative signal gain
was observed in the measured pulse-decay-time
as shown as the red open circle
in Fig.\,\ref{fig:relative gain and pulse decay time at mass test}.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{charge_and_tail.pdf}
\end{center}
\caption{
Relative signal gain (black filled circle) and pulse-decay-time (red open circle) measured for all 4180 MPPCs.
}
\label{fig:relative gain and pulse decay time at mass test}
\end{figure}
\section{Summary and Conclusion}
A large-area VUV-sensitive MPPC has been developed for the LXe scintillation detector
of the MEG II experiment.
The properties of the production model of the VUV-MPPC were thoroughly studied
in a series of measurements in LXe.
It was found to have a high PDE for the LXe scintillation light up to 21\%,
a high gain, low dark count rate and a good single photoelectron resolution.
The overall probability of the correlated noises was also found to be
sufficiently low,
which enables one to operate it at a high over-voltage up to 7--8\,V.
A large active area of $139\,\mathrm{mm}^2$ was realised
by connecting four small MPPC chips
($5.95\times 5.85\,\mathrm{mm}^2$ each) in series
to avoid the large sensor capacitance and to have a short pulse-decay-time.
Performance tests of 4180 VUV-MPPCs produced
for the MEG II LXe detector were carried out at room temperature,
where the $I$-$V$ curves were measured for all 16720 chips individually
and the relative signal gains and the pulse shapes were checked for 4180 sensors.
It turned out that all the sensors with very few exceptions work properly.
The MPPC technology for the enhanced VUV-sensitivity presented here
can also be applied
to the detection of the scintillation light
for other liquid rare-gas such as LAr and LKr.
In fact, it was reported in Ref.~\cite{Igarashi:2015cma}
that a MPPC based on the same technology showed a good performance in LAr
to detect the scintillation light from LAr.
Following our development, another vendor has also started to develop SiPMs with VUV sensitivities~\cite{7182368,nEXO-FBKSiPM}.
The VUV-MPPCs tested here were successfully installed in the MEG II LXe detector, which is now being commissioned.
\section*{Acknowledgements}
This work was supported by JSPS KAKENHI Grant Numbers JP22000004, JP25247034 and JP26000004. K.~Ieki was supported by Grant-in-Aid for JSPS Fellows (Grant Number 2014-0006).
\section*{References}
|
\section{Introduction}
Several asymptotic notions play a fundamental role in the theory of generalized functions. The subject has been studied by several authors and applications have been elaborated in areas such as mathematical physics, Tauberian theorems for integral transforms, number theory, and differential equations. See the monographs \cite{estrada2012distributional, p-s-t, stevan2011asymptotic, vladimirov1988tauberian} for an overview of results and the articles \cite{drozhzhinov2016,P-V2014,yang-estrada} for recent contributions.
The purpose of this article is to present a detailed structural study of the so-called quasiasymptotics of ultradistributions. The concept of quasiasymptotic behavior for Schwartz distributions was introduced by Zav'yalov in \cite{Zavyalov} and further developed by him, Drozhzhinov, and Vladimirov in connection with their powerful multidimensional Tauberian theory for Laplace transforms \cite{vladimirov1988tauberian}. A key aspect in the understanding of this concept is its description via so-called structural theorems and complete results in that direction were achieved in \cite{StructThQADistrInf,StructThQADistrOrigin} (cf. \cite{lojasiewicz57,stevan2011asymptotic}). In \cite{10.2307/44095755} Pilipovi{\'c} and Stankovi{\'c} naturally extended the definition of quasiasymptotic behavior to the context of one-dimensional ultradistributions and studied its basic properties. We shall obtain here complete structural theorems for quasiasymptotics of non-quasianalytic ultradistributions that generalize their distributional counterparts. Our main goal is thus to characterize those ultradistributions having quasiasymptotic behavior as infinite sums of derivatives of functions satisfying classical pointwise asymptotic relations.
The paper is organized as follows. In Section \ref{preliminaries section} we explain some notions and tools that will play a role in our arguments. Section \ref{Section structure at infinity} studies the quasiasymptotic behavior at infinity. A key idea we apply here will be to connect the quasiasymptotic behavior with the so-called $S$-asymptotic behavior \cite{stevan2011asymptotic}, for which structural theorems are available, via an exponential change of variables. The nature of the problem under consideration requires to split our treatment in two cases, depending on whether the degree of the quasiasymptotic behavior is a negative integer or not. We obtain in Section \ref{structure quasiasymptotic at the origin} structural theorems for the quasiasymptotic behavior at the origin. Our technique there is based on a reduction to the results from Section \ref{Section structure at infinity} by means of a change of variables and then regularization. Our method also yields asymptotic properties of regularizations at the origin of ultradistributions having prescribed asymptotic properties, generalizing results for distributions from \cite{VindasReg2011}. It is also worth mentioning that our approach here differs from the one employed in the literature to deal with Schwartz distributions, and in fact can be used to produce new proofs for the classical structural theorems for the quasiasymptotic behavior of distributions. We conclude the article by studying extensions of quasiasymptotics to new ultradistributions spaces of Gelfand-Shilov type that we shall introduce in Section \ref{quasiasymptotics Gelfand-Shilov}.
\section{Preliminaries}
\label{preliminaries section}
Throughout this article we fix
a weight sequence of positive numbers $\{ M_{p} \}_{p \in \naturals}$
and assume it satisfies $(M.1)$, $(M.2)$, and $(M.3)$;
the meaning of all these conditions is very well explained in \cite{ultradistributions1}. Let $\Omega\subseteq\mathbb{R}$.
As customary \cite{ultradistributions1}, $\mathcal{E}^{\ast}(\Omega)$ and $\mathcal{D}^{\ast}(\Omega)$ stand, respectively, for the spaces of all ultradifferentiable functions and compactly supported ultradifferentiable functions of class $\ast$ on $\Omega$, where we employ $\ast$ as
the common notation for $\beurlingM$ and $\roumieuM$. For statements needing a separate treatment we will always talk first about the Beurling case, followed by the corresponding assertion for the Roumieu case in parenthesis. When $\Omega = \reals$, we simply write $\udspacereal{\beurou} = \udspacereals{\beurou}{}$ and $\uespacereal{\beurou} = \uespacereals{\beurou}{}$. The strong duals $\uddspacereal{\beurou}(\Omega)$ and $\udespacereal{\beurou}(\Omega)$ are the spaces of ultradistributions and compactly supported ultradistributions, respectively, on $\Omega$.
The main subject of study of this article is the quasiasymptotic behavior of ultradistributions, which is defined via asymptotic comparison with regularly varying functions. A real-valued measurable function $L$ is called \textit{slowly varying at infinity} \cite{bingham1989regular} if $L$ is positive for large arguments and
$L(ax)\sim L(x)$ as $x\to\infty$,
for any $a > 0$.
We are only interested in the terminal behavior of $L$, so \cite{bingham1989regular} we may always assume $L$ to be defined, positive, and locally bounded (or even continuous) on $[0, \infty)$. Finally, we say that a function $L$ on $(0, \infty)$ is \emph{slowly varying at the origin} if $\widetilde{L}(x):=L(x^{-1})$ is slowly varying at infinity.
In accordance to \cite{10.2307/44095755,stevan2011asymptotic}, we define the quasiasymptotic behavior of an ultradistribution at infinity or at the origin as follows.
\begin{definition}
Let $L$ be a slowly varying function at infinity (at the origin, resp.). We say that $f \in \uddspacereal{\beurou}$ has \textit{quasiasymptotic behavior at infinity (at the origin) in $\uddspacereal{\beurou}$ with respect to $\lambda^{\alpha} L(\lambda)$}, $\alpha \in \reals$, if for some $g \in \uddspacereal{\beurou}$ and every $\phi \in \udspacereal{\beurou}$,
\begin{equation}
\label{eq:quasiasympinfinity}
\lim_{\lambda \rightarrow \infty} \ev{\frac{f(\lambda x)}{\lambda^{\alpha} L(\lambda)}}{\phi(x)} = \ev{g(x)}{\phi(x)} \qquad \left(\mbox{resp., } \lim_{\lambda \rightarrow 0^{+}}\right) .
\end{equation}
If (\ref{eq:quasiasympinfinity}) holds, we also say that $f$ has \textit{quasiasymptotics of order $\alpha$ at infinity (at the origin) with respect to $L$} and write in short: $f(\lambda x) \sim \lambda^{\alpha} L(\lambda) g(x)$ in $\uddspacereal{\beurou}$ as $\lambda \rightarrow \infty$ (resp., $\lambda\to0^{+}$).
\end{definition}
If $f(\lambda x) \sim \lambda^{\alpha} L(\lambda) g(x)$ in $\uddspacereal{\beurou}$ as $\lambda \rightarrow \infty$ (as $\lambda \rightarrow 0^{+}$), it can easily be shown \cite{estrada2012distributional,stevan2011asymptotic} that this forces $g$ to be a homogeneous ultradistribution of degree $\alpha$. An adaptation of the proof of \cite[Theorem 2.6.1]{estrada2012distributional} shows that all homogeneous ultradistributions are exactly the homogeneous distributions, which we denote exactly as in \cite{estrada2012distributional}.
We mention that we will employ the notation $H(x)=x^{0}_{+}$ for the Heaviside function.
In addition, we shall make use of the special (non homogeneous!) distributions $\operatorname*{Pf}(H(\pm) x^{-k})$, $k\in\mathbb{Z}_{+}$, where $\operatorname*{Pf}$ stands for Hadamard finite part regularization \cite{estrada2012distributional}.
\begin{remark} Naturally \cite{stevan2011asymptotic}, the quasiasymptotic behavior may be defined in other spaces of generalized functions $\mathcal{F}'$ by asking (\ref{eq:quasiasympinfinity}) to hold for any $\phi \in \mathcal{F}$, whenever the dialation operators act continuously on the test function space.
\end{remark}
\section{The structure of quasiasymptotics at infinity}
\label{Section structure at infinity}
This section is devoted to studying the quasiasymptotic behavior at infinity. Our main results are Theorem \ref{t:structalpha>-1} and Theorem \ref{t:fstructnid}, where we provide a full description of the structure of quasiasymptotics at infinity. Some auxiliary lemmas used in their proofs are shown in Subsection \ref{Subsection some lemmas}. Throughout this section $L$ stands for a slowly varying function at infinity.
\subsection{Some lemmas}
\label{Subsection some lemmas}
We start with the ensuing useful estimates for the weight sequence $M_p$, which we shall often exploit throughout the article. Hereafter $S(n,k)$ stand for the \textit{Stirling numbers of the second kind} (see e.g. \cite{Johnson02thecurious}).
\begin{lemma}
\label{l:Msumfacgrowth}
For any $\ell > 0$ there is $C_{\ell} > 0$ (independent of $p$) such that
\begin{equation}
\label{eq:Msumfacgrowth}
\sum_{k = p}^{\infty} \frac{k! \ell^{k}}{M_{k}} \leq C_{\ell} \frac{p!}{M_{p}} \ell^{p}
\end{equation}
and
\begin{equation}
\label{eq:Msumgrowth}
\sum_{k = p}^{\infty} S(k + 1, p + 1) \frac{\ell^{k}}{M_{k}} \leq C_{\ell} \frac{(2\ell)^{p}}{M_{p}} .
\end{equation}
\end{lemma}
\begin{proof}
Clearly, it is enough to show \eqref{eq:Msumfacgrowth} just for sufficiently large $p$. Using \cite[Lemma 4.1, p. 55]{ultradistributions1}, there is $p_0$ such that for any $p\geq p_0$ we have $p / m_{p}:=p M_{p-1}/M_{p} \leq (2 \ell)^{-1}$. Hence, it follows that for $p$ in this range
\begin{align*}
\sum_{k = p}^{\infty} \frac{k! \ell^{k}}{M_{k}} &= \frac{p!}{M_{p}} \left( \ell^{p} + \sum_{k = p + 1}^{\infty} \frac{(p + 1) \cdot \ldots \cdot k \cdot \ell^{k}}{m_{p + 1} \cdot \ldots \cdot m_{k}} \right)
\leq 2 \frac{p!}{M_{p}} \ell^{p} .
\end{align*}
For \eqref{eq:Msumgrowth},
in view of \cite[Theorem 3]{StirlingNumbersOfSecondKind}, we have
$S(k + 1, p + 1) \leq 2^{k + 1} (p + 1)^{k - p} \leq 2^{k + 1} k!/p! $
for $k \geq p$. The rest follows by application of \eqref{eq:Msumfacgrowth}.
\end{proof}
In \cite{StructThQADistrInf}, the structure of distributional quasiasymptotics at infinity was found by noting that certain primitives preserve the asymptotic behavior, being of a higher degree, and using the fact that eventually the primitives are continuous functions. As the latter part does not hold in general for ultradistributions, a more careful analysis is needed, although we may carry over some of the distributional results. In fact, one may retread the proofs from \cite[Section 2]{StructThQADistrInf} (see also \cite[Section 2.10]{stevan2011asymptotic})
to obtain,
\begin{lemma}
\label{l:qaprimitiveinfinity}
Let $f \in \uddspacereal{\beurou}$. Suppose $f$ has quasiasymptotics with respect to $\lambda^{\alpha} L(\lambda)$.
\begin{enumerate}[(i)]
\item If $\alpha \notin \integers_{-}$: for any $n \in \naturals$ and any $n$-primitive $F_{n}$ of $f$ there exists a polynomial $P$ of degree at most $n - 1$ such that $F_{n} + P$ has quasiasymptotics with respect to $\lambda^{\alpha + n} L(\lambda)$ in $\uddspacereal{\beurou}$.
\item If $\alpha = -k$, $k \in \integers_{+}$: there is some $(k - 1)$-primitive $F$ of $f$ such that $F$ has quasiasymptotics with respect to $\lambda^{-1}L(\lambda)$ in $\uddspacereal{\beurou}$.
\end{enumerate}
\end{lemma}
The previous lemma roughly speaking shows that in order to find the structure of quasiasymptotics for arbitrary degree, it suffices to discover the structure for degrees $\geq -1$, where extra care is needed for the case $-1$. It should also be noticed that the converse statements for $(i)$ and $(ii)$ from Lemma \ref{l:qaprimitiveinfinity} trivially hold true.
The next lemma, a direct consequence of well-known moment asymptotic expansion \cite{estrada2012distributional,Schmidt2005}, states that the quasiasymptotic behavior of degree $>-1$ is a local property at infinity, which in some arguments enables us to remove the origin from the support of the ultradistribution in our analysis.
\begin{lemma}
\label{l:qalocalproperty}
Suppose that $f_{1}, f_{2} \in \uddspacereal{\beurou}$ and that for some $a > 0$, $f_{1}$ and $f_{2}$ coincide on $\reals \setminus [-a, a]$. Suppose that $f_{1}(\lambda x)\sim \lambda^{\alpha}L(\lambda) g(x)$ in $\uddspacereal{\beurou}$ as $\lambda\to\infty$, where $\alpha > -1$. Then, also $f_{2}(\lambda x)\sim \lambda^{\alpha}L(\lambda) g(x)$ in $\uddspacereal{\beurou}$.
\end{lemma}
\subsection{Structural theorem for $\alpha \notin \integers_{-}$}
We study in this subsection quasiasymptotics of degree $\alpha\notin\mathbb{Z}_{-} $. Part of our analysis reduces the general case to that when $\alpha>-1$, i.e., the case when the quasiasymptotic behavior is local. Consequently, we may restrict our discussion to those ultradistributions whose support lie in the complement of some zero neighborhood. As both the negative and positive half-line can be treated symmetrically, it is natural to start the analysis with ultradistributions that are supported on the positive half-line. In the next crucial lemma we further normalize the situation by assuming that our ultradistribution is supported in $(e,\infty)$.
\begin{lemma}
\label{l:quasiStrucSup} Let $\alpha\in\mathbb{R}$ and let $f \in \uddspacereal{\beurou}$ be such that $\operatorname*{supp} f\subset (e,\infty)$ and $f$ has quasiasymptotic behavior at infinity with respect to $\lambda^{\alpha}L(\lambda)$ in $\uddspacereal{\beurou}(0,\infty)$. Then, there are continuous functions $f_{m}$ such that $\operatorname*{supp} f_{m}\subset (e,\infty)$,
$
f = \sum_{m = 0}^{\infty} f_{m}^{(m)} ,
$
the limits
\begin{equation*}
\lim_{x \rightarrow \infty} \frac{f_{m}(x)}{x^{\alpha + m} L(x)}
\end{equation*}
exist, and furthermore, for some $\ell > 0$ (any $\ell > 0$) there is a $C=C_{\ell} > 0$ such that,
\begin{equation*}
\left| f_{m}(x) \right| \leq C \frac{\ell^{m}}{M_{m}} x^{\alpha + m} L(x), \qquad m\in\mathbb{N},\ x>0. \end{equation*}
\end{lemma}
\begin{proof}
Suppose $f(\lambda x) \sim \lambda^{\alpha} L(\lambda) g(x)$ in $\uddspacereal{\beurou}(0, \infty)$ as $\lambda \rightarrow \infty$. Since composition with a real analytic function induces continuous mappings between spaces of ultradifferentiable functions (see e.g. \cite[Prop. 8.4.1, p. 281]{hormander1990analysis}), we obtain that the composition $f(e^{x})$ is an element of $\uddspacereal{\beurou}$. Also, $\psi \in \udspacereal{\beurou}$ if and only if $\psi(x) = \varphi(e^{x})$ with $\varphi\in \udspacereal{\beurou}(0,\infty)$.
These key observations allow us to make a change of variables in order to apply the structural theorem for $S$-asymptotics \cite[Theorem~1.10, p.~46]{stevan2011asymptotic}.
In fact, we set $u(x) := f(e^{x})$, $w(x) := g(e^{x})$ and $c(h) := e^{\alpha h} L(e^{h})$. (Notice that $w$ has actually the form $w(x)=Be^{\alpha x}$ for some $B>0$.) A quick computation shows that that $u$ has the $S$-asymptotic behavior
$u(x + h) \sim c(h) w(x)$ in $\uddspacereal{\beurou} $ as $h \rightarrow \infty.$ The quoted structure theorem yields the existence of an ultradifferential operator $P(D)$ of class $\beurou$ and continuous functions $u_1$ and $u_2$ with supports on $(1,\infty)$ such that $u = P(D) u_{1} + u_{2} $
on $(0, \infty)$ and $\lim_{h \rightarrow \infty} u_{i}(x + h)/c(h)$ exist uniformly for $x$ on compacts of $(0,\infty)$.
Take any $\varphi \in \udspacereal{\beurou}(0, \infty)$ and put $\psi(x) = e^{x} \varphi(e^{x})$, then the substitution $y = e^{x}$ yields
\begin{align*}
\ev{f(y)}{\varphi(y)} &= \ev{f(y)}{\frac{\varphi(y)}{y} y} = \ev{u(x)}{\psi(x)}
= \ev{P(D) u_{1} (x)}{\psi(x)} + \ev{u_{2} (x)}{\psi(x)} .
\end{align*}
Let us consider both terms of the sum individually. The latter is simply \[ \ev{u_{2}(x)}{\psi(x)} = \int_{1}^{\infty} u_{2}(x) \psi(x) dx = \int_{e}^{\infty} u_{2}(\log y) \varphi(y) dy . \]
Setting $f_{2}(y) := u_{2}(\log y)$, we get
$ \ev{u_{2}(x)}{\psi(x)} = \ev{f_2(y)}{\varphi(y)} ,$
and the existence of
$\lim_{y \rightarrow \infty} y^{-\alpha}f_2(y)/ L(y) .
$ For the first term, we will need to explicitly calculate the derivatives of $\psi$. Using the Fa\`{a} di Bruno formula \cite[Eq. (2.2)]{Johnson02thecurious},
\begin{align*}
\psi^{(n)}(x) &= e^{x} \sum_{k = 0}^{n} {n \choose k} \frac{d^{k}}{dx^{k}} \left( \varphi(e^{x}) \right)
= e^{x} \sum_{m = 0}^{n} S(n + 1, m + 1) e^{mx} \varphi^{(m)}(e^{x}) ,
\end{align*}
where we have applied \cite[Theorem 5.3.B]{AdvancedCombinatorics}.
Then,
\begin{align*}
\int_{1}^{\infty} u_{1}(x) \psi^{(n)}(x) dx
= \sum_{m = 0}^{n} S(n + 1, m + 1) \int_{e}^{\infty} u_{1}(\log y) \varphi^{(m)}(y) y^{m} dy .
\end{align*}
If $P(D) = \sum_{n = 0} a_{n} D^{n}$, then by \eqref{eq:Msumgrowth} from Lemma \ref{l:Msumfacgrowth}, we may consider the following constants,
\begin{equation*}
c_{m} = (-1)^{m} \sum_{n = m}^{\infty} (-1)^{n} a_{n} S(n + 1, m + 1) ,
\end{equation*}
and it follows that $c_{m} \leq C \mu^{m} / M_{m}$ for some $\mu > 0$ (for any $\mu > 0$) and some $C_{\mu} = C > 0$. Collecting everything together, we obtain
\begin{equation*}
\ev{P(D) u_{1}(x)}{\psi(x)} = \sum_{m = 0}^{\infty} (-1)^{m} c_{m} \int_{e}^{\infty} u_{1}(\log y) \varphi^{(m)}(y) y^{m} dy .
\end{equation*}
So if we define $f_{1, m}(y) := u_{1}(\log y) y^{m}$, $m \in \naturals$, we get
\begin{equation*}
\ev{P(D) u_{1}(x)}{\psi(x)} = \sum_{m = 0}^{\infty} c_{m} \ev{f_{1, m}^{(m)}(y)}{\varphi(y)} ,
\end{equation*}
and the limits
$ \lim_{y \rightarrow \infty} y^{-\alpha-m}f_{1, m}(y)/L(y)$
exist. This completes the proof of the lemma.
\end{proof}
We are ready to discuss the general case.
\begin{theorem}
\label{t:structalpha>-1}
Suppose $\alpha \notin \integers_{-}$ and let $k \in \naturals$ be the smallest non-negative integer such that $-(k + 1) < \alpha$. Then, an ultradistribution $f \in \uddspacereal{\beurou}$ has quasiasymptotic behavior
\begin{equation}\label{eqquaalpha>-1}
f(\lambda x)\sim \lambda^{\alpha}L(\lambda)(c_{-}x_{-}^{\alpha}+c
_{+}x_{+}^{\alpha}) \quad \mbox{ in }\uddspacereal{\beurou} \mbox{ as }\lambda\to\infty
\end{equation}
if and only if there exist continuous functions $f_{m}$ on $\reals$, $m \geq k$, such that
\begin{equation}
\label{eq:fstructalpha>-1}
f = \sum_{m = k}^{\infty} f_{m}^{(m)} ,
\end{equation}
the limits
\begin{equation}
\label{eq:fstructalpha>-1asymp}
\lim_{x \rightarrow \pm \infty} \frac{f_{m}(x)}{x^{m}|x|^{\alpha } L(|x|)}=c^{\pm}_{m}, \qquad m \geq k,
\end{equation}
exist, and for some $\ell > 0$ (any $\ell > 0$) there is a $C=C_{\ell} > 0$ such that
\begin{equation}
\label{eq:fstructalpha>-1bound}
\left| f_{m}(x) \right| \leq C \frac{\ell^{m}}{M_{m}} (1+|x|)^{\alpha + m} L(|x|) , \qquad x\in\mathbb{R},
\end{equation}
for all $m \geq k$. Furthermore, in this case we have
\begin{equation}
\label{eq:strucalpha>-1constants}
c_{\pm}= \sum_{m=k}^{\infty} c^{\pm}_{m} \frac{\Gamma(\alpha+m+1)}{\Gamma(\alpha+1)}.
\end{equation}
\end{theorem}
\begin{proof} In view of Lemma \ref{l:qaprimitiveinfinity}(i), we may assume that $\alpha>-1$ so that $k=0$.
Suppose then first that $f $ has quasiasymptotic behavior \eqref{eqquaalpha>-1}. We write $f=f_{-}+f_{c}+f_{+}$, where $f_{c} \in \udespacereal{\beurou}$ coincides with $f$ on an open interval containing $[-e, e]$ and $\operatorname*{supp}f_{-}\subset (-\infty,-e)$ and $\operatorname*{supp}f_{+}\subset (e,\infty)$. Then, by Lemma \ref{l:qalocalproperty} each $f_{\pm}$ has quasiasymptotic behavior with respect to $\lambda^{\alpha} L(\lambda)$ in $\uddspacereal{\beurou}(-\infty, 0)$ and $\uddspacereal{\beurou}(0,\infty)$, respectively. Using Lemma \ref{l:quasiStrucSup}, we find continuous functions $f_{1, m}^{\pm}$, $m \in \naturals$ with supports in $(-\infty,-e)$ and $(e, \infty)$, respectively, such that the identities
$ f_{\pm} = \sum_{m = 0}^{\infty} ( f_{1, m}^{\pm} )^{(m)}$
hold, the limits
$$
c^{\pm}_{m}=(-1)^{m}\lim_{x\to\infty} \frac{f_{1,m}^{\pm}(\pm x)}{x^{\alpha+m}L(x)}
$$
exist, and the bounds $|f_{1,m}^{\pm}(x)|\leq C' \ell^{m} |x|^{\alpha + m} L(|x|)/ M_{m} $ are satisfied for some $\ell > 0$ (any $\ell > 0$) and some $C'=C'_{\ell} > 0$. Applying Komatsu's first structural theorem\footnote{The first structural theorem even holds true in the quasianalytic case under mild conditions, see \cite[Proposition 4.1 and Proposition 4.7]{debrouwere-vindasCousin}.} for ultradistributions \cite[Theorem 8.1 and Theorem 8.7]{ultradistributions1} one can also find continuous functions $g_m$, whose supports lie in some (arbitrarily chosen) neighborhood of $\supp f_{c}$, such that $f_{c} = \sum_{m=0}^{\infty}g^{(m)}_{m}$ in $\uddspacereal{\beurou}$ and $\sup_{x\in\mathbb{R}}|g_{m}(x)|\leq C'' \ell^{m}/M_m$ for some $\ell>0$ (for every $\ell>0$) and $C''=C''_{\ell}>0$. The functions
$ f_{m} = g_{m} + f_{1, m}^{-} + f_{1, m}^{+} $
satisfy all sought requirements. We verify the relation \eqref{eq:strucalpha>-1constants} below.
Conversely, assume that $f$ satisfies all of the conditions above. Take any $\phi \in \udspacereal{\beurou}$ and suppose that for some $R > 1$ we have $\supp \phi \subseteq [-R, R]$. Pick $\gamma > 0$ such that $\alpha - \gamma> -1$.
Using Potter's estimate \cite[Theorem 1.5.4]{bingham1989regular}, we may assume that
\begin{equation}
\label{eq:slowlyvaryingsubpolynomialbound}
\frac{L(\lambda x)}{L(\lambda)} \leq C_{\gamma} \max\{ x^{- \gamma}, x^{\gamma} \}
\end{equation}
holds for all $x,\lambda>0$. Since $\phi \in \udspacereal{\beurou}$, for any $h > 0$ (for some $h > 0$) there exists a $C_{\phi,h}$ such that for all $m \in \naturals$ and $x \in \reals$ we have $|\phi^{(m)}(x)| \leq C_{\phi,h} h^{m} M_{m}$. Due to (\ref{eq:fstructalpha>-1bound}), we now have for any $m \in \naturals$ and $\lambda>1$
\begin{align*}
&\left| \frac{1}{\lambda} \int_{- \infty}^{\infty} \frac{f_{m}(x)}{\lambda^{\alpha} L(\lambda)} \frac{\phi^{(m)}(x / \lambda)}{\lambda^{m}} dx \right|
\\
&
\leq C_{\ell} \frac{(2\ell)^{m}}{M_{m}} \left( \int_{-\lambda^{-1}}^{\lambda^{-1}} \frac{|\phi^{(m)}(x)|}{\lambda^{m + \alpha} L(\lambda)} dx + \int_{|x|\geq 1/ \lambda} \frac{L(\lambda |x|)}{L(\lambda)} |x|^{\alpha + m} |\phi^{(m)}(x)| dx \right)
\\
&
\leq 2 C_{\ell} C_{\phi,h} (2h \ell)^{m} \left( \frac{1}{\lambda^{m + 1 + \alpha} L(\lambda)} + C_{\gamma} R^{m + \alpha + \gamma + 1} + \frac{C_{\gamma}}{\alpha - \gamma + m + 1} \right)
\leq C (2h\ell R)^{m},
\end{align*}
and, as $2h \ell$ may be chosen freely, this is absolutely summable over $m \in \naturals$.
It follows by applying the Lebesgue dominated convergence theorem twice that
\begin{align*}
\lim_{\lambda \rightarrow \infty } \ev{\frac{f(\lambda x)}{\lambda^{\alpha} L(\lambda)}}{\phi(x)}
&= \lim_{\lambda \rightarrow \infty} \frac{1}{\lambda} \sum_{m = 0}^{\infty} (-1)^{m} \int_{- \infty}^{\infty} \frac{f_{m}(x)}{\lambda^{\alpha} L(\lambda)} \frac{\phi^{(m)}(x / \lambda)}{\lambda^{m}} dx
\\
&= c_{-} \int_{- \infty}^{0} |x|^{\alpha} \phi(x) dx + c_{+} \int_{0}^{\infty} x^{\alpha} \phi(x) dx ,
\end{align*}
with $c_{-}$ and $c_{+}$ given by \eqref{eq:strucalpha>-1constants}.
\end{proof}
\subsection{Structural Theorem for negative integral degrees}
\label{subsection Structural theorem for negative integral degree}
We now address the case of quasiasymptotics of degree $\alpha\in\mathbb{Z}_{-}$. The next structural theorem is the second main result of this section.
\begin{theorem}
\label{t:fstructnid}
Let $k \in \integers_{+}$ and $f \in \uddspacereal{\beurou}$. Then, $f$ has the quasiasymptotic behavior
\begin{equation}
\label{eq:strucquasi-k} f(\lambda x)\sim \frac{L(\lambda)}{\lambda^{k}}( \gamma \delta^{(k-1)}(x) + \beta x^{-k}) \quad \mbox{ in } \uddspacereal{\beurou} \mbox{ as }\lambda\to\infty
\end{equation}
if and only if there exist continuous functions $f_{m}$ on $\reals$, $m \geq k - 1$, such that
\begin{equation}
\label{eqext:fstructnid}
f = \sum_{m = k - 1}^{\infty} f_{m}^{(m)} ,
\end{equation}
the limits
\begin{equation}
\label{eq:fstructnidasymp}
\lim_{x \rightarrow \pm \infty} \frac{f_{m}(x)}{x^{ m-k} L(|x|)}=c_{m}^{\pm}, \qquad m \geq k - 1,
\end{equation}
and
\begin{equation}
\label{eq:fstructnidasymp2}
\lim_{x \rightarrow\infty} \frac{1}{ L(x)}\int_{-x}^{x}f_{k-1}(t)dt=c_{k-1}^{\ast}
\end{equation}
exist, and for some $\ell > 0$ (any $\ell > 0$) there is $C=C_{\ell} > 0$ such that
\begin{equation}
\label{eq:fstructnidbound}
\left| f_{m}(x) \right| \leq C \frac{\ell^{m}}{M_{m}}(1+ |x|)^{m-k} L(|x|), \qquad x\in\mathbb{R},
\end{equation}
for all $m \geq k $. Furthermore, we must have
\begin{equation}
\label{eq:constantsk}
\gamma= c_{k-1}^{\ast}+\sum_{m=k}^{\infty}(c^{+}_{m}-c^{-}_{m}) \quad \mbox{and} \quad \beta= (-1)^{k-1}(k-1)!c^{+}_{k-1}=(-1)^{k-1}(k-1)!c^{-}_{k-1}.
\end{equation}
\end{theorem}
\begin{proof} In view of Lemma \ref{l:qaprimitiveinfinity}(ii) we may assume that $k=1$.
\emph{Necessity.} We start showing the necessity of the conditions if $f$ has the quasiasymptotic behavior \eqref{eq:strucquasi-k}.
Take a compactly supported ultradistribution $f_c$ that coincides with $f$ on $[-e,e]$ and consider $\tilde{f}=f-f_{c}$, so that $\supp (f-f_{c})\cap [-e,e]=\emptyset$.
We set $g(x) = x(f(x)-f_{c}(x))$, which
has quasiasymptotic behavior
$g(\lambda x)\sim \beta L(\lambda)$ in $\uddspacereal{\beurou}$ as $\lambda\to\infty. $
Splitting $g$ as the sum of two distributions supported on $(-\infty,-e)$ and $(e,\infty)$ respectively, we can apply Lemma \ref{l:quasiStrucSup} to obtain its structure as
$ g = \sum_{m = 0}^{\infty} g_{m}^{(m)} , $
where each of the functions has support in $(-\infty, -e)\cup (e, \infty)$, satisfies the corresponding bounds implied by the lemma, and is such that the limits $\lim_{x\to \pm \infty}x^{-m}g_{m}(x)/L(|x|)$ exist. Define, for any $j \in \naturals$, the following continuous functions
\[ \tilde{f}_{j}(x) = \frac{x^{j - 1}}{j!} \sum_{m = j}^{\infty} m! g_{m}(x) x^{-m} , \qquad x \neq 0 , \]
and $\tilde{f}_{j}(0)=0$.
Let us verify they satisfy the requirements that the $f_j$ should satisfy. First of all, for some $\ell > 0$ (any $\ell > 0$) and $C=C_{\ell}>0$,
\[ \left| \tilde{f}_{j}(x) \right| \leq C \frac{|x|^{j - 1}}{j!} L(|x|) \sum_{m = j}^{\infty} \frac{m! \ell^{m}}{M_{m}} \leq C' |x|^{j - 1} \frac{\ell^{j}}{M_{j}} L(|x|) , \]
by \eqref{eq:Msumfacgrowth} from Lemma \ref{l:Msumfacgrowth}. This not only shows that each $\tilde{f}_{j}$ is well-defined and continuous on $\reals$, but also provides the bounds \eqref{eq:fstructnidbound} for them. From dominated convergence we infer the existence of
\[ \lim_{x \rightarrow \pm \infty} \frac{\tilde{f}_{j}(x)}{x^{j - 1} L(|x|)} = \lim_{x \rightarrow \pm \infty} \frac{1}{j!} \sum_{m = j}^{\infty} \frac{m! g_{m}(x)}{x^{m} L(|x|)} = \frac{1}{j!} \sum_{m = j}^{\infty} \lim_{x \rightarrow \pm\infty} \frac{m! g_{m}(x)}{x^{m} L(|x|)} .\]
Take an arbitrary $\phi \in \udspacereal{\beurou}$ and let $\varphi \in \udspacereal{\beurou}$ be another corresponding test function that coincides with $\phi$ on $\mathbb{R}\setminus{(-e,e)}$, while its support does not contain the origin. We then have
\begin{align*}
\ev{\tilde{f} (x)}{\phi(x)} &= \ev{g(x)}{\frac{\varphi(x)}{x}}
= \sum_{m = 0}^{\infty} \sum_{j = 0}^{m} (-1)^{m} {m \choose j} \ev{g_{m}(x)}{(-1)^{m - j} (m - j)! \frac{\varphi^{(j)}(x)}{x^{m - j + 1}}} \\
&= \sum_{j = 0}^{\infty} \frac{(-1)^{j}}{j!} \sum_{m = j}^{\infty} \ev{m! x^{j- 1} \frac{g_{m}(x)}{x^{m}}}{\varphi^{(j)}(x)}
= \sum_{j = 0}^{\infty} \ev{\tilde{f}_{j}^{(j)}}{\phi(x)} .
\end{align*}
Applying the first structural theorem to $f_c$ as in the proof of Theorem \ref{t:structalpha>-1}, we obtain compactly supported continuous functions $g_{m}$ such that $f_m=\tilde{f}_m+g_m$ satisfy \eqref{eqext:fstructnid}, \eqref{eq:fstructnidasymp}, and \eqref{eq:fstructnidbound}.
The necessity of \eqref{eq:fstructnidasymp2} follows from \eqref{eq:asymptotically homogeneous in quasi} below. That \eqref{eq:constantsk} must necessarily hold will also be shown below in the proof of the converse.
\emph{Sufficiency.} Conversely, assume that \eqref{eqext:fstructnid} holds with $f_m$ fulfilling \eqref{eq:fstructnidasymp}, \eqref{eq:fstructnidasymp2} and \eqref{eq:fstructnidbound} (recall we work with the reduction $k=1$). We assume without loss of generality that $L(x)$ is everywhere continuous and vanishes for $x\leq 1$. We consider
$
g= \sum_{m = 1}^{\infty} f_{m}^{(m - 1)}.
$
It follows from Theorem \ref{t:structalpha>-1} that $g$ has quasiasymptotic behavior of degree 0 with respect to $L(\lambda)$, and differentiation then yields
$$
f(\lambda x)-f_{0}(\lambda x)= g'(\lambda x)\sim (\gamma-c_{0}^{*}) \frac{L(\lambda)}{\lambda}\delta(x) \quad \mbox{ in } \uddspacereal{\beurou} \mbox{ as }\lambda\to\infty,
$$
with $\gamma$ precisely given as in (\ref{eq:constantsk}). It thus remains to determine the quasiasymptotic properties of $f_{0}$. Write $F(x)= \int_{0}^{x} f_{0}(t)dt$. Since $f_{0}(\pm x)\sim \pm c^{\pm}_{0}L(x)/x$, $x\to\infty,$ one readily shows that
\begin{align*}
F(\lambda x)H(\pm x)&= F(\pm \lambda)H(\pm x)+ c_{0}^{\pm} \int_{\lambda}^{\pm \lambda x}\frac{L(t)}{t}dt+ o\left(L(\lambda)\right)
\\
&= F(\pm \lambda)H(\pm x)+ c^{\pm}_{0}L(\lambda)H(\pm x)\log |x| + o\left(L(\lambda)\right), \quad \lambda \to\infty,
\end{align*}
uniformly for $x$ on compact intervals, and in particular the relation holds in $\uddspacereal{\beurou}$. Differentiating
\[
F(\lambda x)= F(-\lambda)H(-x)+ F(\lambda) H(x)+ L(\lambda)\left(c^{-}_{0}H(-x)+ c^{+}_{0} H(x)\right)\log |x|+o
\left(L(\lambda)\right) ,
\]
we conclude that
\begin{equation}
\label{eq:asymptotically homogeneous in quasi}
f_{0}(\lambda x)= \frac{F(\lambda)-F(-\lambda)}{\lambda} \delta(x)+ \frac{L(\lambda)}{\lambda}\left(c_{0}^{-}\operatorname*{Pf}\left(\frac{H(-x)}{x}\right)+ c_{0}^{+}\operatorname*{Pf}\left(\frac{H(x)}{x}\right)\right)+ o\left(\frac{L(\lambda)}{\lambda}\right),
\end{equation}
whence the result follows.
\end{proof}
\subsection{Extension from $\reals\setminus\{0\}$ to $\reals$}
\label{subsec:extension}
The methods employed in the previous two subsections also allow us to study the following question. Suppose that the restriction of $f\in \uddspacereal{\beurou}$
to $\mathbb{R}\setminus\{0\}$ is known to have quasiasymptotic behavior in $\uddspace{\beurou}{\mathbb{R}\setminus\{0\}}$, what can we say about the quasiasymptotic properties of $f$? In view of symmetry considerations, it is clear that it suffices to restrict our attention to ultradistributions supported on $[0,\infty)$.
\begin{theorem}
\label{t:extensioninftyalpha}
Suppose that $f \in \uddspacereal{\beurou}$ is supported in $[0,\infty)$ and has quasiasymptotic behavior
$
f(\lambda x)\sim c \lambda^{\alpha} L(\lambda)x^{\alpha}$ in $\uddspacereal{\beurou} (0,\infty)$ as $\lambda\to\infty.$
\begin{itemize}
\item[(i)] If $\alpha > -1$, then $f(\lambda x)\sim c \lambda^{\alpha} L(\lambda)x_{+}^{\alpha}$ in $\uddspacereal{\beurou}$ as $\lambda\to\infty$.
\item [(ii)] If $\alpha < -1$ and $N \in \naturals$ is such that $-(N + 1) < \alpha < -N$, then there exist constants $a_{0}, \ldots, a_{N - 1}$ such that
\[ f(\lambda x) - \sum_{n = 0}^{N - 1} a_{n} \frac{\delta^{(n)}(x)}{\lambda^{n+1}} \sim c \lambda^{\alpha} L(\lambda)x_{+}^{\alpha} \quad \mbox{ in } \uddspacereal{\beurou} \mbox{ as } \lambda\to\infty, \]
\item [(iii)] If $\alpha=-k\in\mathbb{Z}_{-}$, then there is a function $b$ satisfying\footnote{Such functions are called associate homogeneous of degree 0 with respect to $L$ in \cite{stevan2011asymptotic,StructThQADistrInf}. They coincide with functions of the so-called De Haan class \cite{bingham1989regular}.} for each $a>0$
\begin{equation}
\label{eq:asshomogeneous}
b(ax) = b(x) + c\frac{(-1)^{k - 1}}{(k - 1)!} L(x) \log a + \littleoh{L(x)},
\end{equation}
$x \rightarrow \infty$, and constants $a_{0}, \ldots, a_{k - 1}$ such that
\begin{equation}
\label{eq:extensioninftynid}
f(\lambda x) = c \frac{L(\lambda)}{\lambda^{k}} \operatorname*{Pf}\left(\frac{H(x)}{x^{k}}\right) + \frac{b(\lambda)}{\lambda^{k}} \delta^{(k - 1)}(x) + \sum_{j = 0}^{k - 1} a_{j} \frac{\delta^{(j)}(x)}{\lambda^{j + 1}} + \littleoh{\frac{L(\lambda)}{\lambda^{k}}} ,
\end{equation}
in $\uddspacereal{\beurou}$ as $\lambda \rightarrow \infty$.
\end{itemize}
\end{theorem}
\begin{proof}
The moment asymptotic expansion \cite{Schmidt2005} says that we may assume that, say, $\supp f\subset (e,\infty)$ by removing a neighborhood of the origin. So, we can apply exactly the same argument as in the proof of Theorem \ref{t:structalpha>-1} (via Lemma \ref{l:quasiStrucSup} and Lemma \ref{l:qaprimitiveinfinity}(i)) to show parts (i) and (ii). For (iii), we assume without loss of generality that $k=1$ (Lemma \ref{l:qaprimitiveinfinity}(ii)) and apply the same argument as in the proof of Theorem \ref{t:fstructnid} to conclude that
$f(\lambda x)= f_{0}(\lambda x)+ \gamma L(\lambda)\delta(\lambda x)+o\left(L(\lambda)/\lambda\right)$ in $\uddspacereal{\beurou},$
where the continuous function $f_0$ has also support in $(e,\infty)$ and satisfies $f_{0}(x)\sim c L(x)/x$, $x\to \infty$, in the ordinary sense. At this point the result can be derived from \cite[Theorem 4.3]{StructThQADistrInf} (see also \cite[Theorem 2.38, p. 155]{stevan2011asymptotic}), but we might argue directly as follows. In fact, we proceed in the same way we arrived at (\ref{eq:asymptotically homogeneous in quasi}). Set $b(x)=\int_{1}^{x}f_{0}(t)dt$,
then, uniformly for $x$ in compact subsets of $(0,\infty)$,
$$
b(\lambda x)=b(\lambda)H(x)+c\int_{\lambda}^{\lambda x}\frac{L(t)}{t}dt+ o(L(\lambda))= b(\lambda)H(x)+cL(\lambda)H(x) \log x+ o(L(\lambda)),
$$
so that differentiation finally shows
$$
f_{0}(\lambda x)= \frac{b(\lambda)}{\lambda} \delta(x)+ c\frac{L(\lambda)}{\lambda}\operatorname*{Pf}\left(\frac{H(x)}{x}\right)+ o\left(\frac{L(\lambda)}{\lambda}\right) \qquad \mbox{in } \mathcal{D}'.
$$ \end{proof}
\section{The structure of quasiasymptotics at the origin}\label{structure quasiasymptotic at the origin}
We now focus our attention on quasiasymptotic behavior at the origin. The reader should notice that Lemma \ref{l:qaprimitiveinfinity} holds for quasiasymptotics at the origin as well. Furthermore, it is a simple consequence of the definition that quasiasymptotics at the origin is a local property, in the sense that two ultradistributions that coincide in a neighborhood of the origin must have precisely the same quasiasymptotic properties. Throughout this section $L$ stands for a slowly varying function at the origin and we set $\tilde{L}(x)=L(1/x)$. From now on, by convention the parameters $\varepsilon\to 0^{+}$ and $\lambda\to \infty$.
We will reduce the analysis of the structure of quasiasymptotics at the origin to that of the quasiasymptotics at infinity the via the change of variables $x\leftrightarrow 1/x$. We therefore need to see how this substitution acts on derivatives.
\begin{lemma}
\label{l:faadibrunoreciprocal}
Let $\phi \in C^{\infty}(\reals\setminus\{0\})$ and set $\psi(x) := x^{-2} \phi(1 / x)$. Then for any $m \in \mathbb{N}$, there exist constants $c_{m, 0}, \ldots, c_{m, m}$ such that
\begin{equation}
\label{eq:faadibrunoreciprocal}
\frac{d^{m}}{dx^{m}} \left( \psi(x) \right) = \sum_{j = 0}^{m} c_{m, j} \frac{\phi^{(j)}(1 / x)}{x^{m + j + 2}} ,
\end{equation}
where we have the bounds
\begin{equation}
\label{eq:faadibrunoreciprocalbound}
|c_{m, j}| \leq \frac{m!}{j!} 4^{m} , \qquad 0 \leq j \leq m .
\end{equation}
\end{lemma}
\begin{proof}
Applying the Fa\`{a} di Bruno formula \cite[Eq. (2.2)]{Johnson02thecurious},
\begin{equation*}
\frac{d^{k}}{dx^{k}} \left( \phi(1/x) \right)
= \sum_{j = 1}^{k} (-1)^{k} x^{-(k + j)} \phi^{(j)}(1/x) B_{k, j}(1!, 2!, \ldots, (k - j + 1)!) ,
\end{equation*}
where $B_{k, j}$ are the Bell polynomials; from their generating function identity \cite[ (3a'), p. 133]{AdvancedCombinatorics} we infer that
\[ B_{k, j}(1!, \ldots, (k - j + 1)!) = \left. \frac{d^{k}}{dt^{k}} \left( \frac{1}{j!} \left( \frac{t}{1 - t} \right)^{j} \right) \right|_{t = 0} = \frac{k!(k - 1)!}{j! (j - 1)!(k-j)!} . \]
Therefore, we obtain that \eqref{eq:faadibrunoreciprocal} holds with
$$
c_{m,0}= (-1)^{m} (m + 1)! \quad \mbox{ and } \quad c_{m,j}= (-1)^{m}\frac{m!}{j!} \sum_{k=j}^{m} (m-k+1) \binom{k-1}{j-1}
$$
when $0<j\leq m$, whence one readily obtains the bound \eqref{eq:faadibrunoreciprocalbound}.
\end{proof}
\begin{theorem}
\label{t:structoriginalpha>-1}
Let $\alpha \notin \integers_{-}$ and let $k \in \naturals$ be the smallest integers such that $-(k + 1) < \alpha$. Then, $f \in \uddspacereal{\beurou}$ has quasiasymptotic behavior
\begin{equation}
\label{eqquasiorigin noninteger}
f(\varepsilon x)\sim \varepsilon^{\alpha}L(\varepsilon)(c_{-}x_{-}^{\alpha}+c
_{+}x_{+}^{\alpha}) \quad \mbox{ in }\uddspacereal{\beurou} \mbox{ as }\varepsilon\to0^{+}
\end{equation}
if and only if there exist functions $f_m\in L^{1}(-1,1)$, $m \geq k$, that are continuous on $[-1,1]\setminus\{0\}$ such that \eqref{eq:fstructalpha>-1} holds on $(-1,1)$,
\begin{equation}
\label{eq:structoriginalpha>-1limit}
c^{\pm}_{m}=\lim_{x \rightarrow 0\pm} \frac{f_{m}(x)}{x^{m}|x|^{\alpha} L(|x|)}, \qquad m \geq k ,
\end{equation}
exist, and furthermore, for some $\ell > 0$ (for any $\ell > 0$) there is a $C > 0$ such that
\begin{equation}
\label{eq:structoriginalpha>-1bound}
|f_{m}(x)| \leq C \frac{\ell^{m}}{M_{m}} |x|^{\alpha + m} L(|x|) , \qquad 0 < |x| \leq 1 ,
\end{equation}
for all $m \geq k$. Moreover, the relation
\eqref{eq:strucalpha>-1constants} must hold.
\end{theorem}
\begin{proof}
The proof of sufficiency can be done analogously as in Theorem \ref{t:structalpha>-1}. Hence we are only left with necessity. If we can show the theorem for degree larger than $-1$, then the full structure theorem will follow from Lemma \ref{l:qaprimitiveinfinity}(i), hence we assume that $\alpha > -1$ (hence $k=0$). If $f$ has quasiasymptotic behavior with respect to $\varepsilon^{\alpha} L(\varepsilon)$, then $\widetilde{f}(x) := f(1 / x)$ has quasiasymptotic behavior in $\uddspace{\beurou}{\reals\setminus\{0\}}$ with respect to $\lambda^{- \alpha} \widetilde{L}(\lambda)$. Then by Theorem \ref{t:structalpha>-1} or Theorem \ref{t:fstructnid} if $\alpha \in \integers_{+}$ and keeping in mind our observations from Section \ref{subsec:extension}, there exist continuous $\widetilde{f}_{m}$ in $\reals\setminus\{0\}$, $m \geq 0$, that satisfy (\ref{eq:fstructalpha>-1}), (\ref{eq:fstructalpha>-1asymp}) and (\ref{eq:fstructalpha>-1bound}). Consider now for any $m \geq 0$,
\[ f_{m}(x) := \sum_{k = m}^{\infty} (-1)^{k + m } c_{k, m} \widetilde{f}_{k}(1 / x) x^{m + k} , \]
where the $c_{k, m}$ are as in Lemma \ref{l:faadibrunoreciprocal}. By (\ref{eq:fstructalpha>-1bound}) and (\ref{eq:faadibrunoreciprocalbound}) it follows that for some $\ell > 0$ (for any $\ell > 0$) and any $0 < |x| \leq 1$,
\begin{align*}
|f_{m}(x)| &= \left| \sum_{k = m}^{\infty} (-1)^{k + m } c_{k, m} \widetilde{f}_{k}(1 / x) x^{m + k} \right|
\leq \sum_{k = m}^{\infty} \frac{k!}{m!} 4^{k} \cdot C \frac{(2\ell)^{k}}{M_{k}} |x|^{\alpha - k} L(|x|) |x|^{m + k} \\
&= C |x|^{\alpha + m} L(|x|)\frac{1}{m!} \sum_{k = m}^{\infty} k! \frac{(8\ell)^{k}}{M_{k}} \leq C C_{8\ell} \frac{(8\ell)^{m}}{M_{m}} |x|^{\alpha + m} L(|x|) ,
\end{align*}
by \eqref{eq:Msumfacgrowth} from Lemma \ref{l:Msumfacgrowth}. This not only shows existence and continuity in $[-1, 1]\setminus\{0\}$, but also shows that the $f_{m}$ satisfy (\ref{eq:structoriginalpha>-1bound}). By (\ref{eq:fstructalpha>-1asymp}) and dominated convergence, it also follows that for these functions the limits (\ref{eq:structoriginalpha>-1limit}) exist. Now take any $\phi \in \udspace{\beurou}{\reals\setminus\{0\}}$ with $\supp \phi \subseteq (-1, 1)$ and set $\psi(x) := \phi(1 / x) x^{-2}$. Then,
\[ \ev{f(x)}{\phi(x)} = \ev{\widetilde{f}(x)}{\psi(x)} = \sum_{k = 0}^{\infty} \ev{\widetilde{f}_{k}(x)}{(-1)^{k} \psi^{(k)}(x)} . \]
Since for any $k \in \naturals$, by Lemma \ref{l:faadibrunoreciprocal},
\begin{align*}
\int_{- \infty}^{\infty} \widetilde{f}_{k}(x) \psi^{(k)}(x) dx
= \sum_{m = 0}^{k} c_{k, m} \int_{- \infty}^{\infty} \widetilde{f}_{k}(1 / x) \phi^{(m)}(x) x^{m + k} dx ,
\end{align*}
it follows by switching the order of summation that
$f = \sum_{m = 0}^{\infty} f_{m}^{(m)} ,$
in $\uddspacereal{\beurou}((-1, 1)\setminus\{0\})$. Now as $\alpha > -1$, the latter sum is an element of $\uddspacereal{\beurou}$, so that there is some $g \in \uddspacereal{\beurou}$ with $\supp g \subseteq \{0\}$ for which
$ f = \sum_{m = 0}^{\infty} f_{m}^{(m)} + g , $
in $\uddspacereal{\beurou}(-1, 1)$. Since we have already shown sufficiency, the sum has quasiasymptotics with respect to $\varepsilon^{\alpha} L(\varepsilon)$, implying that the same holds for $g$. If $g\neq 0$, we can find an ultradifferentiable operator $P(D) = \sum_{n \geq n_0} a_{n} D^{n}$ of type $\beurou$ such that $g = P(D) \delta$ and $a_{n_0}\neq 0$. Then, for any $\phi \in \udspacereal{\beurou}$,
\[ \ev{\frac{g(\varepsilon x)}{\varepsilon^{\alpha} L(\varepsilon)}}{\phi(x)} = \sum_{n = n_0}^{\infty} (-1)^{n} a_{n} \frac{\varepsilon^{- n - \alpha - 1}}{L(\varepsilon)} \phi^{(n)}(0) . \]
But if $\phi(x)=x^{n_0}$ in a neighborhood of 0, we conclude that
$$
\infty=\lim_{\varepsilon \to 0^{+}}\frac{1}{\varepsilon^{n_0 +\alpha+ 1} L(\varepsilon)}= \frac{(-1)^{n_0}}{a_{n_0}n_0!} \lim _{\varepsilon\to 0^{+}}\ev{\frac{g(\varepsilon x)}{\varepsilon^{\alpha} L(\varepsilon)}}{\phi(x)},
$$
leading to a contradiction. Therefore, $g$ must be identically 0 and this completes the proof of the theorem. \end{proof}
The structure for negative integral degree can be described as follows.
\begin{theorem}
\label{t:structoriginnid} Let $f \in \uddspacereal{\beurou}$ and $k \in \integers_{+}$. Then, $f$ has quasiasymptotic behavior
\begin{equation}
\label{eq:strucquasi-k-0} f(\varepsilon x)\sim \frac{L(\varepsilon)}{\varepsilon^{k}}( \gamma \delta^{(k-1)}(x) + \beta x^{-k}) \quad \mbox{ in } \uddspacereal{\beurou} \mbox{ as }\varepsilon\to0^{+}
\end{equation}
if and only if there are continuous functions $F$ and $f_{m}$ on $[-1, 1]\setminus\{0\}$, $m \geq k$, such that
\begin{equation}
\label{eq:structoriginnid}
f = F^{(k)}+\sum_{m = k }^{\infty} f_{m}^{(m)} \qquad \text{ on } (-1, 1) ,
\end{equation}
the limits
\begin{equation}
\label{eq:structoriginnidasymp}
c^{\pm}_{m}=\lim_{x \rightarrow 0\pm} \frac{f_{m}(x)}{x^{ m-k} L(|x|)} , \qquad m \geq k ,
\end{equation}
exist,
for some $\ell > 0$ (for any $\ell > 0$) there exists $C =C_{\ell}> 0$ such that
\begin{equation}
\label{eq:structoriginnidbound}
|f_{m}(x)| \leq C \frac{\ell^{m}}{M_{m}} |x|^{ m-k} L(|x|) , \qquad 0 < |x| \leq 1 ,
\end{equation}
for all $m \geq k$, and for any $a > 0$ the limit
\begin{equation}
\label{eq:structoriginnidprimitive}
\lim_{x \rightarrow 0+} \frac{F(ax) -F(-x)}{L(x)} = c^{\ast}_{1} + c^{\ast}_{2} \log a
\end{equation}
exists. In this case,
\begin{equation}
\label{eq:constantskzero}
\gamma= c^{\ast}_1+\sum_{m=k}^{\infty}(c^{+}_{m}-c^{-}_{m}) \quad \mbox{and} \quad \beta= (-1)^{k-1}(k-1)!c^{\ast}_{2}.
\end{equation}
\end{theorem}
\begin{proof}
For the sufficiency, applying Theorem \ref{t:structoriginalpha>-1} to the series $\sum_{m=k}^{\infty}f^{(m-1)}_{m}$, one deduces
$
f(\varepsilon x)- F^{(k)}(\varepsilon x)\sim (\gamma- c^{\ast}_{1}) \delta^{(k-1)}(\varepsilon x)$ in $\uddspacereal{\beurou}$ as $\varepsilon\to0^{+}
$.
In view of \cite[Theorem 5.3]{StructThQADistrOrigin} (see also \cite[Theorem 2.33, p. 149]{stevan2011asymptotic}), we have
$
F^{(k)}(\varepsilon x)\sim L(\varepsilon)( c^{\ast}_{1} \delta^{(k-1)}(\varepsilon x) + \beta (\varepsilon x)^{-k})$ in $\mathcal{D}'$ as $\varepsilon\to0^{+},$
which yields the result.
For the necessity, we may assume that $k=1$. We now apply Theorem \ref{t:structoriginalpha>-1} to $xf(x)$. Using the same reasoning as in the proof of Theorem \ref{t:fstructnid}, one can write $f(x)=f_{0}+\sum_{m=1}^{\infty}f_{m}^{(m)}$ on $(-1,1)\setminus\{0\}$, with continuous functions $f_0, f_1,\dots$ on $[-1,1]\setminus\{0\}$ such that the limits \eqref{eq:structoriginnidasymp} exist including the case $m=0$. Applying again Theorem \ref{t:structoriginalpha>-1} to the series $\sum_{m=1}^{\infty}f_{m}^{(m-1)}$, we deduce that $f_{0}$ has an extension $g_0$ to $\mathbb{R}$ with quasiasymptotic behavior of order $-1$ with respect to $L(\varepsilon)$. Let $F$ be a first order primitive of $g_0$. Due to the fact that $F'=f_0$ off the origin and the quasiasymptotic behavior of $F'$, it is clear that $F$ is integrable at the origin and that it must have the form
\[F(x)=
- H(x) \left(\int^{1}_{x} f_{0}(t)dt + C_{+}\right) + H(-x)\left(\int_{-1}^{x} f_{0}(t)dt + C_{-}\right) .
\]
Similarly as in the proof of Theorem \ref{t:fstructnid}, we conclude that
$$
c_1^{\ast}= \lim_{x\to 0^{+}} \frac{F(x)-F(-x)}{L(x)}
$$
must exist
by comparing with the quasiasymptotics of $g_{0}$. Hence, for each $a>0$
\begin{equation*}
\lim_{x \rightarrow 0+} \frac{F(ax) -F(-x)}{L(x)} = c_{1}^{\ast}+ \lim_{x \rightarrow 0+} \frac{1}{L(x)}\int_{x}^{ax}f_{0}(t)dt
=c_{1}^{\ast}+c_{0}^{+}\log a.
\end{equation*}
\end{proof}
Our method also yields:
\begin{theorem}
\label{t:extensionzeroalpha}
Suppose that $f_{0} \in \uddspacereal{\beurou}(0,\infty)$ has quasiasymptotic behavior
\[
f_0(\varepsilon x)\sim c \varepsilon^{\alpha} L(\varepsilon)x^{\alpha} \quad \mbox{in } \uddspacereal{\beurou} (0,\infty) \mbox{ as } \varepsilon\to0^{+}.
\]
Then $f_0$ admits extensions to $\mathbb{R}$. Let $f \in \uddspacereal{\beurou}$ be any of such extensions with support in $[0,\infty)$. Then:
\begin{itemize}
\item [(I)] If $\alpha \notin \mathbb{Z}_{-}$, then there is $g\in \uddspacereal{\beurou}$ with $\operatorname*{supp} g\subseteq \{0\}$ such that
\[ f( \varepsilon x) - g(\varepsilon x) \sim c \varepsilon^{\alpha} L( \varepsilon)x_{+}^{\alpha} \quad \mbox{ in } \uddspacereal{\beurou} \mbox{ as } \varepsilon\to0^{+}. \]
\item [(II)] If $\alpha=-k\in\mathbb{Z}_{-}$, then there are a function $b$ satisfying \eqref{eq:asshomogeneous} as $x\to0^{+}$ for each $a>0$
and an ultradistribution $g\in \uddspacereal{\beurou}$ with $\operatorname*{supp} g\subseteq \{0\}$ such that
\begin{equation*}
f( \varepsilon x) = c \frac{L( \varepsilon)}{ \varepsilon^{k}} \operatorname*{Pf}\left(\frac{H(x)}{x^{k}}\right) + \frac{b( \varepsilon)}{ \varepsilon^{k}} \delta^{(k - 1)}(x) + g(\varepsilon x) + \littleoh{\frac{L( \varepsilon)}{ \varepsilon^{k}}} \quad\mbox{in } \uddspacereal{\beurou} \:\mbox{as } \varepsilon \rightarrow 0^{+}.
\end{equation*}
\end{itemize}
\end{theorem}
\section{Quasiasymptotic behavior in $\mathcal{Z}^{\prime\beurou}$}
\label{quasiasymptotics Gelfand-Shilov}
As an application of our structural theorems, we now discuss some other extension results for quasiasymptotics of ultradistributions. For distributions, the connection between tempered distributions and the quasiasymptotic behavior has been extensively studied \cite{stevan2011asymptotic, P-V2014,StructThQADistrInf,StructThQADistrOrigin,Zavyalov89}. The following properties are well known:
\begin{enumerate}
\item[1.] If $f\in\mathcal{D}'$ has quasiasymptotic behavior at infinity, then $f\in \mathcal{S}'$ and it has the same quasiasymptotic behavior in $\mathcal{S}'$.
\item [2.]If $f\in\mathcal{S}'$ has quasiasymptotic behavior at the origin in $\mathcal{D}'$, then it has the same quasiasymptotic behavior in $\mathcal{S}'$.
\end{enumerate}
Our goal is to obtain ultradistributional analogs of these results. For this, we introduce new ultradistibution spaces $\mathcal{Z}^{\prime\beurou}$, somewhat reminiscent of the Gelfand-Shilov type spaces \cite{carmichael2007boundary} and at the same time generalizing $\mathcal{S}'$. They are defined as follows. For any $n \in \naturals$ and $h > 0$, $\mathcal{Z}^{M_{p}, h}_{n}$ denotes the Banach space of all $\varphi \in C^{\infty}$ for which the norm
\[ \norm{\varphi}_{M_{p}, n, h} := \sup_{x \in \reals, m \in \naturals} \frac{(1 + |x|)^{n + m} |\varphi^{(m)}(x)|}{h^{m} M_{m}} \]
is finite. Then we consider the following locally convex spaces
\[ \mathcal{Z}^{(M_{p})}_{n} = \varprojlim_{h \rightarrow 0^{+}} \mathcal{Z}^{M_{p}, h}_{n}, \qquad \mathcal{Z}^{\{M_{p}\}}_{n} = \varinjlim_{h \rightarrow \infty} \mathcal{Z}^{M_{p}, h}_{n} , \]
corresponding to the Beurling and Roumieu case, where we use $\mathcal{Z}^{\beurou}_{n}$ as a common notation for these two cases, and finally we define
\[ \mathcal{Z}^{\beurou} = \varprojlim_{n \in \naturals} \mathcal{Z}^{\beurou}_{n} . \]
The aim of this section is to show that quasiasymptotic behavior in $\mathcal{D}^{\prime\beurou}$ naturally extends to quasiasymptotic behavior in $\mathcal{Z}^{\prime\beurou}$. Let us first consider the case at infinity.
\begin{theorem}
If $f \in \mathcal{D}^{\prime\beurou}$ has quasiasymptotic behavior with respect to $\lambda^{\alpha} L(\lambda)$, with $L$ slowly varying at infinity and $\alpha \in \reals$, then $f \in \mathcal{Z}^{\prime\beurou}$ and it has the same quasiasymptotic behavior in $\mathcal{Z}^{\prime\beurou}$.
\end{theorem}
\begin{proof}
Let $k \in \naturals$ be the smallest natural number such that $-(k + 1) \leq \alpha$. Then by either Theorem \ref{t:structalpha>-1} or Theorem \ref{t:fstructnid} we find some $\ell > 0$ (for any $\ell > 0$) and a $C=C_{\ell} > 0$ such that \eqref{eq:fstructalpha>-1}
and \eqref{eq:fstructalpha>-1bound} hold. Wet set $n=\lceil \alpha + 1 \rceil$. Employing Potter's estimate \eqref{eq:slowlyvaryingsubpolynomialbound} (with $\gamma=\lambda=1$), we find that for any $\varphi \in \mathcal{Z}^{\beurou}$ and any $m \geq k$ we have
\[ \left| \int_{- \infty}^{\infty} f_{m}(x) \varphi^{(m)}(x) dx \right| \leq C \frac{\ell^{m}}{M_{m}} \int_{\reals} (1 + |x|)^{ m +n} | \varphi^{(m)}(x) | dx \leq C' \|\varphi\|_{M_p,n + 2,h} (h \ell^{m}) , \]
and as $h \ell$ may be chosen freely, it follows that this is absolutely summable over $m \geq k$. Consequently, $f=\sum_{m=k}^{\infty} f_m^{(m)} \in \mathcal{Z}^{\prime\beurou}$.
For the quasiasymptotic behavior of $f$, the case where $\alpha$ is not a negative integer can be shown in a similar fashion as the sufficiency proof of Theorem \ref{t:structalpha>-1}. For $\alpha = - k \in \integers_{-}$, it is clear that we only need to treat the case $k = 1$, as the general case then automatically follows by differentiating. By Theorem \ref{t:fstructnid}, there exist continuous functions $f_{m}$, $m \in \naturals$, satisfying (\ref{eq:fstructnidasymp}), (\ref{eq:fstructnidasymp2}), and (\ref{eq:fstructnidbound}) such that
$f = f_{0} + \sum_{m = 1}^{\infty} f^{(m)}_{m} . $
The infinite sum in the previous identity clearly has a primitive with quasiasymptotic behavior with respect to $L(\lambda)$, so that its quasiasymptotic behavior may be extended to the whole of $\mathcal{Z}^{\prime\beurou}$, and in turn its derivative $\sum_{m = 1}^{\infty} f^{(m)}_{m}$ has quasiasymptotic behavior with respect to $\lambda^{-1} L(\lambda)$ in $\mathcal{Z}^{\prime\beurou}$. By (\ref{eq:fstructnidasymp}) and \eqref{eq:fstructnidasymp2}, $f_{0}$ has quasiasymptotic behavior with respect $\lambda^{-1} L(\lambda)$ in ${\mathcal{D}}'$, hence, by \cite[Remark 3.1]{StructThQADistrInf} (see also \cite[Theorem 2.41, p. 158]{stevan2011asymptotic}), it has the same quasiasymptotic behavior in ${\sss}'$, hence certainly also in $\mathcal{Z}^{\prime\beurou}$. Therefore, the same also holds for $f$.
\end{proof}
Let us now turn our attention to the case at the origin. The next lemma proves that the quasiasymptotic at the origin in $\mathcal{Z}^{\prime\ast}$ is a local property.
\begin{lemma}
\label{p:quasiasymporiginlocalZ}
Let $L$ be a slowly varying function at the origin and $\alpha \in \reals$. Suppose $f_{1}, f_{2} \in \mathcal{Z}^{\prime\beurou}$ are such that for some $a > 0$, $f_{1}$ and $f_{2}$ coincide on $\reals \setminus [-a, a]$. Suppose that $f_{1}(\varepsilon x) \sim \varepsilon^{\alpha} L(\varepsilon) g(x)$ in $\mathcal{Z}^{\prime\beurou}$ as $\varepsilon \rightarrow 0^{+}$, then, also $f_{2}(\varepsilon x) \sim \varepsilon^{\alpha} L(\varepsilon) g(x)$ in $\mathcal{Z}^{\prime\beurou}$.
\end{lemma}
\begin{proof}
We only show the Beurling case; the Roumieu case can be shown analogously by employing a projective description for $\mathcal{Z}^{\{M_{p}\}}$ obtained similarly as in \cite{carmichael2007boundary}. It suffices to show that if $f \in \mathcal{Z}^{\prime(M_p)}$ vanishes near the origin, then $f(\varepsilon x) \sim \varepsilon^{N} \cdot 0$ for all $N \in \naturals$. Let $f$ be as described, then there exist $0 < R < 1$, $n \in \naturals$, $\ell, C > 0$ such that
\[ \left| \ev{f(\varepsilon x)}{\phi(x)} \right| \leq C \sup_{|x| \geq R, m \in \naturals} \frac{|x|^{n + m} |\phi^{(m)}(x)|}{\ell^{m} M_{m}}, \qquad \phi \in \mathcal{Z}^{(M_{p})} . \]
Taking $\phi(x) = \varepsilon^{-1} \varphi(x / \varepsilon)$ with $\varphi \in \mathcal{Z}^{(M_{p})}$ and arbitrary $0 < \varepsilon < 1$ we have for $N \geq n$
\begin{align*}
\varepsilon^{-N} \left|\ev{f(\varepsilon x)}{\varphi}\right| &\leq C \sup_{|x| \geq R, m \in \naturals} \frac{|x|^{n + m} |\varphi^{(m)}(x / \varepsilon)|}{\varepsilon^{N + m + 1} \ell^{m} M_{m}} \\
&\leq CR^{-N+n-1} \sup_{|x| \geq R / \varepsilon, m \in \naturals} \frac{|x|^{N + m + 1} |\varphi^{(m)}(x)|}{\ell^{m} M_{m}} \rightarrow 0 ,
\end{align*}
as $\varepsilon \rightarrow 0^{+}$.
\end{proof}
\begin{theorem}
Suppose $f \in \mathcal{Z}^{\prime\beurou}$ has quasiasymptotic behavior in $\mathcal{D}^{\prime\beurou}$ with respect to $\varepsilon^{\alpha} L(\varepsilon)$, with $L$ slowly varying at the origin and $\alpha \in \reals$, then $f$ has the same quasiasymptotic behavior in $\mathcal{Z}^{\prime\beurou}$.
\end{theorem}
\begin{proof}
By Lemma \ref{p:quasiasymporiginlocalZ} we may assume that $\supp f \subset [-1, 1]$. Suppose first that $\alpha \notin \integers_{-}$ and let $k \in \naturals$ be the smallest integer such that $-(k + 1) < \alpha$. From Theorem \ref{t:structoriginalpha>-1} we find continuous functions $f_{m}$ on $[-1, 1] \setminus \{0\}$, satisfying (\ref{eq:fstructalpha>-1}), (\ref{eq:structoriginalpha>-1limit}) and (\ref{eq:structoriginalpha>-1bound}). Take any $\psi \in \mathcal{Z}^{\prime\beurou}$ and decompose it as $\psi = \psi_{-} + \psi_{c} + \psi_{+}$ where $\supp \psi_{-} \subseteq (-\infty, -1]$, $\psi_{c}$ has compact support and $\supp \psi_{+} \subseteq [1, \infty)$. Then by the hypothesis
\[ \lim_{\varepsilon \rightarrow 0+} \ev{\frac{f(\varepsilon x)}{\varepsilon^{\alpha} L(\varepsilon)}}{\psi_{c}(x)} = c_{+} \ev{x_{+}^{\alpha}}{\psi_{c}(x)} + c_{-} \ev{x_{-}^{\alpha}}{\psi_{c}(x)} . \]
It suffices to show that the same limit holds for $\psi_{\pm}$ placed instead of $\psi_{c}$. As the two cases are symmetrical, we only look at $\psi_{+}$. It follows from (\ref{eq:slowlyvaryingsubpolynomialbound}), (\ref{eq:structoriginalpha>-1bound}) and the Lebesgue dominated convergence theorem that for any $m \geq k$,
\begin{align*}
\lim_{\varepsilon \rightarrow 0+} \ev{\frac{f_{m}^{(m)}(\varepsilon x)}{\varepsilon^{\alpha} L(\varepsilon)}}{\psi_{+}(x)}
&
= \lim_{\varepsilon \rightarrow 0+} \int_{1}^{1 / \varepsilon} \frac{L(\varepsilon x)}{L(\varepsilon)} \left( \frac{f_{m}(\varepsilon x)}{(\varepsilon x)^{\alpha + m} L(\varepsilon x)} \right) x^{\alpha + m} \psi_{+}^{(m)}(x) dx \\
&= c^{+}_{m} \int_{0}^{\infty} x^{\alpha} \psi_{+}(x) dx .
\end{align*}
Then another application of dominated convergence shows that
\[ \lim_{\varepsilon \rightarrow 0+} \ev{\frac{f(\varepsilon x)}{\varepsilon^{\alpha} L(\varepsilon)}}{\psi_{+}(x)} = c_{+} \ev{x^{\alpha}_{+}}{\psi_{+}(x)} . \]
This shows the case for $\alpha \notin \integers_{-}$. The case of negative integral degree can then be done as in the proof of \cite[Theorem 6.1]{StructThQADistrOrigin}.
\end{proof}
\subsection*{Acknowledgement}
We thank Andreas Debrouwere for his useful comments that led to improvements in the results of Section \ref{quasiasymptotics Gelfand-Shilov}.
|
\section{Introduction}
Much research has been devoted in recent years to Brascamp-Lieb and related inequalities, we refer to \cite{BL}, \cite{BCCT},
\cite{BBBF}, \cite{BBCF} and the references therein.
Brascamp-Lieb inequalities are $\L^p$ estimates for certain multilinear forms on functions on
Euclidean spaces. The forms consist of integrating the tensor product of the input functions over a subspace
of the direct sum of the domain spaces.
Following general conventions, we parameterize the subspace of integration by $\mathbb{R}^m$ and write the corresponding Brascamp-Lieb inequality
\begin{equation}\label{brasliebineq}
\Big | \int_{\mathbb{R}^m} \Big ( \prod_{i=1}^n F_i(\Pi_i x)\Big )\, dx \Big |
\leq C \prod_{i=1}^n \|F_i\|_{p_i}
\end{equation}
with suitable surjective linear maps
$$\Pi_i:\mathbb{R}^m\to \mathbb{R}^{k_i} .$$
Here the constant $C$ is independent of the measurable functions
$F_i$ on $\mathbb{R}^{k_i}$, and integrability on the left-hand side being implied by finiteness of the right-hand side.
It is well understood, under which conditions the Brascamp-Lieb inequality holds.
Bennett, Carbery, Christ, and Tao \cite{BCCT} prove a necessary and sufficient dimensional condition, namely that
\begin{equation}\label{bcct}
\dim(V)\le \sum_{i=1}^n \frac 1{p_i} \dim(\Pi_i V)
\end{equation}
for every subspace $V$ of $\mathbb{R}^m$, with equality if $V=\mathbb{R}^m$. Necessity of inequality \eqref{bcct} is easily seen by testing
the Brascamp-Lieb inequality on certain characteristic functions $F_i$. These functions have minimal support such that the integrand on the left-hand side of \eqref{brasliebineq} is nonzero on a one-neighborhood in $\mathbb{R}^m$ of an arbitrarily large ball in $V$. Necessity of the reverse inequality in case $V=\mathbb{R}^m$ is obtained by using similarly an arbitrarily small ball in $\mathbb{R}^m$.
In this paper, we focus on singular Brascamp-Lieb inequalities. This variant has also seen much development
in recent years, but still lacks a general criterion
mirroring the condition \eqref{bcct}.
A singular Brascamp-Lieb inequality incorporates a Calder\'on-Zygmund kernel
on the left hand side:
\begin{equation}\label{singbraslieb}
\Big | \int_{\mathbb{R}^m} \Big( \prod_{i=1}^n F_i(\Pi_i x)\Big) K(\Pi x)\, dx \Big | \leq C \prod_{i=1}^n \|F_i\|_{p_i} .
\end{equation}
Here $\Pi: \mathbb{R}^m\to \mathbb{R}^k$ is a surjective linear map, and by Calder\'on-Zygmund kernel we mean in this paper a
tempered distribution $K$ on $\mathbb{R}^k$ whose
Fourier transform $\widehat{K}$, called the multiplier associated with $K$, is a measurable function satisfying the symbol estimates
\begin{equation}\label{czsymbol}|\partial^\alpha \widehat{K}(\xi)|\le |\xi|^{-|\alpha|}
\end{equation}
for all $\xi\neq 0$ and all multi-indices $\alpha$ up to suitably large order.
A necessary condition for the singular Brascamp-Lieb inequality \eqref{singbraslieb} can be obtained by specifying $K$ to be the Dirac delta, that is $\widehat{K}=1$. In this case, \eqref{singbraslieb} can be recognized
as a classical Brascamp-Lieb inequality \eqref{brasliebineq} with integration over the kernel of $\Pi$. Condition \eqref{bcct} then yields the necessary condition
\begin{equation}\label{newbcct}
\dim(V) \le \sum_{i=1}^n \frac 1{p_i} \dim(\Pi_i|_{\ker \Pi} (V))
\end{equation}
for all $V\subseteq \ker \Pi$, with equality if $V=\ker \Pi$.
Lacking a general necessary and sufficient condition, the theory of singular Brascamp-Lieb inequalities
remains at the stage of a case-by-case study.
Here, for the first time, we study a sufficiently general family to expose a non-trivial role of the condition \eqref{newbcct}.
We focus on a case that features the following cubical structure. For a parameter $m\ge 1$ we consider $\mathbb{R}^{2m}$ with coordinates
$$(x_1^0,\ldots ,x_m^0,x_1^1,\ldots ,x_m^1)^T,$$
which we also combine as pair of vectors $(x^0,x^1)^T$ or
we write as vector $x$.
Define the cube $Q$ to be the set of functions
$$j:\{1,2,\dots, m\} \to \{0,1\}.$$
For $j\in Q$ define the projection $\Pi_j:\mathbb{R}^{2m} \rightarrow \mathbb{R}^m$ by
\begin{align*
\Pi_j x = (x_1^{j(1)},x_2^{j(2)},\ldots ,x_m^{j(m)})^T.
\end{align*}
Our main theorem states that for these particular projections $\Pi_j$
and for the exponents $p_j=2^m$, inequalities \eqref{newbcct} provide a sufficient condition on an otherwise arbitrary surjective linear map $\Pi:\mathbb{R}^{2m}\to \mathbb{R}^m$ for the singular Brascamp-Lieb inquality to hold.
\begin{theorem}
\label{mainthm}
Given $m\ge 1$, there is an $N\ge 0$ such that for all
surjective linear maps $\Pi:\mathbb{R}^{2m}\to \mathbb{R}^m$
the following are equivalent.
\begin{enumerate}
\item For all subspaces $V\subset {\ker \Pi}$ we have
\begin{equation}
\label{cubebcct}
\dim(V) \le \sum_{j\in Q} 2^{-m} \dim(\Pi_j |_{\ker \Pi} (V)),
\end{equation}
with equality if $V=\ker \Pi$.
\item
For all $j\in Q$, the composed map $\Pi \Pi_j^T$
is regular.
\item There is a constant $C$ such that for all Calder\'on-Zygmund
kernels $K$ satisfying the symbol estimates \eqref{czsymbol} for all multi-indices up to degree $N$, and for all
tuples of Schwartz functions $(F_j)_{j\in Q}$ we have
\begin{equation}\label{maininequality}
\Big| \int_{\mathbb{R}^{2m}} \Big( \prod_{j\in Q} F_j(\Pi_j x)\Big ) K(\Pi x)\, dx\, \Big | \le C \prod_{j\in Q} \|F_j\|_{2^m} .
\end{equation}
\end{enumerate}
\end{theorem}
Condition (1) of Theorem \ref{mainthm} is the necessary condition
derived from that of Bennett, Carbery, Christ, and Tao. In the
present setting it can immediately be simplified.
For $V=\ker \Pi$, the left-hand side of \eqref{cubebcct} is at least $m$, while the right-hand side is at most $m$, because each summand is at most $2^{-m}m$ and there are $2^m$ summands.
Assuming that inequality \eqref{cubebcct} holds for this $V$, we
conclude actual equality for this $V$. We further conclude that
the restriction of $\Pi_j$ to $\ker \Pi$ is injective for each $j$,
and therefore equality in \eqref{cubebcct} holds for all
subspaces $V$ of $\ker \Pi$. Thus condition (1) in Theorem \ref{mainthm} is equivalent to the single instance with $V=\ker \Pi $, which in turn
is equivalent to all $\Pi_j$ being injective on $\ker \Pi $.
It is now easy to see that conditions (1) and (2) in Theorem \ref{mainthm}
are equivalent.
Namely, let $j$ and $l$ be any two opposite corners of the
cube. Then the range of $\Pi_j^T$ is obviously the kernel of
$\Pi_l$ and $2^m$ dimensional. Hence regularity of $\Pi \Pi_j^T$
is the same as injectivitiy of $\Pi$ on the kernel of $\Pi_l$.
By the above discussion, conditions (1) and (2) are equivalent.
We have already argued that (3) implies (1), hence the main content of
the Theorem \ref{mainthm} is that (2) implies (3).
While the projections $\Pi_j$ of Theorem \ref {mainthm} may appear rather particular, they provide no loss of generality up
to change of variables after fixing
their combinatorial datum, that is the set of integer tuples $(\dim(\Pi_j(V))_{j\in Q}$ with $V$ a subspace of $\mathbb{R}^{2m}$. For each $1\le i\le m$, there exist one-dimensional subspaces $V$ and $W$ of $\mathbb{R}^{2m}$, each spanning a certain standard coordinate axis, such that $\dim(\Pi_j(V))=j(i)$ and $\dim(\Pi_j(W))=1-j(i)$ for all $j$.
Conversely, consider any collection of linear maps $(\widetilde{\Pi}_j)_{j\in Q}$ defined on $\mathbb{R}^{2m}$ with $m$ dimensional range, such that for each $1\le i\le m$ there are spaces $V$ and $W$ with combinatorial datum analoguous as above. Then these spaces necessarily are one dimensional and together span $\mathbb{R}^{2m}$. A suitable linear transformation of $\mathbb{R}^m$ will turn these vector spaces into the standard coordinate axes. Together with a suitable choice of basis for the range of each of the maps $\widetilde{\Pi}_j$, these maps will be identified as the above maps $\Pi_j$.
The role of the cubical structure of the form in this theorem
is to allow for a symmetrization process in the tuple of functions $F_j$. Indeed, the main Lemma \ref{indstep} stated in Section \ref{sec:inductivest} is an induction over the number
of axis parallel symmetry planes of this cube that the tuple $F_j$ respects, in the sense of \eqref{reflectct}.
This symmetrization procedure, sometimes called twisted technology, originates in a series of papers such as \cite{vk:tp}, \cite{vk:bell}, \cite{pd:Lp}.
Theorem \ref{mainthm} in the case $m=2$ generalizes estimates in \cite{pd:L4} and \cite{dkst}.
Further generalizations of Theorem \ref{mainthm} appear desirable, but are beyond the scope of the present paper, except for a mild vector-valued generalization in Lemma \ref{indstep}.
Most naturally, one could seek an extension to other exponents $p_j$ and ask for an optimal range of exponents. One may also seek generalizations in which the index set is a subset of the cube. This can sometimes be achieved by setting some functions $F_j$ constantly equal to one, provided one has bounds with $p_j=\infty$.
A further question concerns the exact dependence on $\Pi$
of the bounds in the theorem.
To elaborate some of the difficulties in the absence of the cubical structure, we briefly discuss a singular Brascamp-Lieb integral
with three input functions. We take $m=4$ and $k=2$, $k_i=2$ and $p_i=3$ for $i=1,2,3$.
The projections $\Pi$ and $\Pi_i$ are then given by $2\times 4$ matrices, which we write as block matrices $(B \ A)$ and $(B_i \ A_i )$ with quadratic blocks.
Choosing coordinates suitably on domain and range of $\Pi$, we may assume that
$$\Pi=(\begin{array}{cc} 0 & I\end{array})$$
with the identity matrix $I$. In order to not violate \eqref{newbcct} with $V$ equal to $\ker\Pi$, the matrices $B_i$ need
to be regular. Changing coordinates on the range of $\Pi_i$, we may assume $B_i=I$ for each $1\le i\le 3$.
Warchalski, in his PhD thesis \cite{mw}, classifies the possiblilities for the remaining parameters
$A_1$, $A_2$, $A_3$ into nine cases.
Most cases can be normalized such that $A_1=0$ and $A_2=I$,
leaving only $A_3$ as indetermined matrix.
A trivial case occurs if $A_3=0$ or $A_3=I$, this results in a reduction of the complexity of the integral by combining
$F_3$ with one of the other functions by a pointwise product. The case that all eigenvalues of $A_3$ are different
from $0$ and $1$ is the generic two dimensional version
of the bilinear Hilbert transform \cite{lt}. The known proofs of the singular Brascamp-Lieb inequality in this case require the technique of time-frequency analysis, which is somewhat
different from the technique in the present paper. The case that one eigenvalue of $A_3$ is equal to $0$ or $1$ and the
other eigenvalue is different from $0$ and $1$ is an interesting hybrid case discussed in \cite{dt:2dbht}. The case when $A_3$ has both $0$ and $1$ as eigenvalue
is called the twisted paraproduct and is an instance of the forms in Theorem \ref{mainthm} with $m=2$, albeit with the fourth
function set constant equal to $1$.
The only case in Warchalski's thesis where the singular Brascamp-Lieb inequality is not known to hold is the one where the first columns of all three
matrices $A_1,A_2,A_3$ vanish, while the second columns are $(0,0)^T$, $(0,1)^T$, $(1,0)^T$, respectively.
Thanks to the vanishing first columns, one variable integrates out trivially and one reduces to a one-dimensional Calder\'on-Zygmund kernel.
The paradigmatic example in this case is the conjectured inequality
$$\Big | \int_{\mathbb{R}^3} F(x,y) G(y,z) H(z,x) \frac 1{x+y+z} \, dxdydz\Big | \le
C \|F\|_3\|G\|_3\|H\|_3,$$
where the left-hand side is called the triangular Hilbert transform. Proving the displayed a priori bound is one of the most
intriguing open problems in the area of singular Brascamp-Lieb inequalities. Partial progress on this problem can be found in \cite{pz:splx}
based on the approach in \cite{tt}, and in \cite{dkt}, \cite{ktz:tht}.
A more detailed survey of singular Brascamp-Lieb inequalities appears in
\cite{dt}.
\section{Symmetry considerations and the inductive statement}
\label{sec:inductivest}
Theorem \ref{mainthm} will be proven by induction. The inductive statement is the content of Lemma \ref{indstep} below. In this section we further discuss certain symmetries of the singular Brascamp-Lieb integrals \eqref{maininequality}, which will be needed in the proof of the inductive statement.
For the rest of the paper, we consider a higher-dimensional generalization of the singular Brascamp-Lieb inequality \eqref{maininequality}, motivated by the related paper \cite{patterns} on certain patterns in positive density subsets of the Euclidean space. We write vectors as column vectors and identify $x\in (\mathbb{R}^{d})^{2m}$ with a vector of vectors as
$$(x_1^0,\ldots ,x_m^0,x_1^1,\ldots ,x_m^1)^T,$$
where $x_1^0,\ldots ,x_m^0,x_1^1,\ldots ,x_m^1 \in \mathbb{R}^{d}$, which we also combine into
a pair of vectors $(x^0,x^1)^T$.
For $j\in Q$, let $\Pi_j: (\mathbb{R}^{d})^{2m} \rightarrow (\mathbb{R}^d)^m$ be given by
\begin{align*}
\Pi_j x = (x_1^{j(1)},x_2^{j(2)},\ldots ,x_m^{j(m)})^T.
\end{align*}
We define an action of an $m \times m$ matrix $A$ on a vector $y=(y_1,\ldots, y_m)^T\in (\mathbb{R}^{d})^m$ by the Kronecker product of the matrix $A$ with the $m\times m$ identity matrix $I$
$$Ay:= (A\otimes I)y,\ \ (Ay)_i = \sum_{j=1}^m a_{ij}y_j $$
for $1\leq i \leq m$.
In similar fashion, we identify $\Pi_j: (\mathbb{R}^{d})^{2m} \rightarrow (\mathbb{R}^d)^m$ with $m\times 2m$ matrices.
We also restrict attention to those projections
$\Pi:(\mathbb{R}^d)^{2m}\to (\mathbb{R}^d)^m$ which are given as analoguous block matrix product as
\begin{equation}
\label{bact}\Pi x =(\begin{array}{cc} B & A\end{array}) x,
\end{equation}
where $A$ and $B$ are $m\times m$ matrices and $x\in (\mathbb{R}^{d})^{2m}$.
This setup makes our higher-dimensional generalization a
very simple extension of the one-dimensional theory.
It is no restriction to assume that all functions $F_j$
in Theorem \ref{mainthm} are real valued.
Schwartz functions in this section will map $F_j:(\mathbb{R}^{d})^m$ to $\mathbb{R}$
and multipliers $\widehat{K}$ will map $(\mathbb{R}^d)^m \setminus~ \{0\}$
to $\mathbb{C}$.
\begin{lemma}[{Symmetries of \eqref{maininequality}}]
\label{lemma:symmetries}
The following two statements hold.
\begin{enumerate}
\item \label{sym:scale}
Let $D$ be an $m\times m$ diagonal matrix of rank $m$. Let $\widetilde{D}$ be a $2m\times 2m$ matrix which decomposes into four blocks of size $m\times m$, the two blocks on the diagonal being $D$ and the two off-diagonal blocks being $0$.
Let $1\leq p_j \leq \infty$ with $\sum_{j\in Q} \frac{1}{p_j} =1$. Then
\begin{align}\label{FFtilde}
\int_{(\mathbb{R}^d)^{2m}}\Big ( \prod_{j\in Q} F_j(\Pi_j x) \Big ) K(\Pi x)\, dx = \int_{(\mathbb{R}^d)^{2m}} \Big ( \prod_{j\in Q} \widetilde{F}_j({\Pi}_{j}x) \Big )\widetilde{K}(\widetilde{\Pi}x) dx
\end{align}
holds with
$$\widetilde{F}_j(y):=\det(D)^\frac{d}{p_j} F_j(D y), \ \widetilde{K}(y): = \det(D)^d K(Dy),\ \widetilde{\Pi} := {D}^{-1} \Pi \widetilde{D}.$$
\item \label{sym:permute}
Let $P$ be a permutation of $m$ elements, which we also identify with the $m\times m$ matrix in which the $ij$--th entry equals $\delta_{P(i)j}$ in the Kronecker delta notation.
Let $\widetilde{P}$ be a $2m\times 2m$ matrix which decomposes into four blocks of size $m\times m$, the two blocks on the diagonal being $P$ and the two off-diagonal blocks being $0$.
Then \eqref{FFtilde} holds with
$$ \widetilde{F}_j(y) := F_{j\circ P}(Py),\
\widetilde{K}(y):=K(Py),\ \widetilde{\Pi}:= P^{-1} \Pi \widetilde{P}.$$
\end{enumerate}
\end{lemma}
\begin{proof}
Proof of \eqref{sym:scale}. Changing variables by $\widetilde{D}$ we have for the left-hand side of \eqref{FFtilde}
\begin{align*}
\det(\widetilde{D})^d \int_{(\mathbb{R}^d)^{2m}}
\Big( \prod_{j\in Q} F_j(\Pi_j \widetilde{D}x) \Big ) K(\Pi \widetilde{D}x)\, dx.
\end{align*}
Using $D\Pi_j=\Pi_j \widetilde{D}$
thanks to the special structure of the projections $\Pi_j$, and using
$$\det(\widetilde{D})^d=\det(D)^{2d}$$
and
$\sum_{j\in Q} \frac{1}{p_j}=1$, the previous display equals
\begin{align}\nonumber
& \det(D)^d\int_{(\mathbb{R}^{d})^{2m}}
\Big( \prod_{j\in Q} \det(D)^\frac{d}{p_j} F_j(D\Pi_j x)\Big ) K(\Pi \widetilde{D}x)\, dx.
\end{align}
With notation as in \eqref{sym:scale} of the lemma, this becomes the right-hand side of \eqref{FFtilde}.
Proof of (2). We compute similarly as above
\begin{align*}
&\int_{(\mathbb{R}^{d})^{2m}}\Big( \prod_{j\in Q} F_j(\Pi_j x)\Big ) K(\Pi x)\, dx
= \int_{(\mathbb{R}^{d})^{2m}} \Big( \prod_{j\in Q} {F}_j(\Pi_j\widetilde{P}x) \Big )K(\Pi \widetilde{P}x) \, dx \\
= & \int_{(\mathbb{R}^{d})^{2m}} \Big( \prod_{j\in Q} {F}_j(P {\Pi}_{j\circ P^{-1}}x) \Big )\widetilde{K}(\widetilde{\Pi}x) dx
= \int_{(\mathbb{R}^{d})^{2m}} \Big( \prod_{j\in Q} {F}_{j\circ P}(P{\Pi}_{j}x)\Big )\widetilde{K}(\widetilde{\Pi}x) dx,
\end{align*}
with notation as in \eqref{sym:permute} of the lemma.
\end{proof}
To prove Theorem \ref{mainthm} it suffices to consider the singular Brascamp-Lieb integral
\begin{equation}\label{sblA}
\Lambda(K,A):=\int_{(\mathbb{R}^{d})^{2m}} \Big ( \prod_{j\in Q} F_j(\Pi_j x)\Big ) K((I\ A)x)\, dx.
\end{equation}
This is justified as follows.
Note that if $K$ is a Calder\'on-Zygmund kernel on $\mathbb{R}^{dm}$,
then so is a certain nonzero scalar multiple of
$\widetilde{K}$ defined by
$$\widetilde{K}(Eu)=K(u)$$
for some regular matrix $E$. Hence
$$\widetilde{K}(\widetilde{\Pi}u)=K(\Pi u)$$
with $\widetilde{\Pi}:=E\Pi$.
Regularity of $\widetilde{\Pi}\Pi_j ^T$ is equivalent to
regularity of ${\Pi}\Pi_j ^T$, so we may use this flexibility
to replace the matrix
$(B \ A )$ in \eqref{bact} by $(EB \ EA)$ and therefore assume that
the matrix $B$ is diagonal and idempotent.
Regularity of all matrices $\widetilde{\Pi}\Pi_j ^T$ then
requires $B$ to be the identity matrix.
Let $1\le i\le m$ act by reflection $j\mapsto i*j$ on the cube $Q$,
where
$$(i* j)(k):= j(k)$$
if $i\neq k$ and
$$(i *j)(i):= 1-j(i).$$
Denote the Gaussian on $\mathbb{R}^s$ by $g(x):=e^{-\pi |x|^2}$ and write $g_t(x):=t^{-s}g(t^{-s}x)$, where $s$ is to be understood from the context, typically $s=1$,
$d$, $md$, $2md$ or $(2m-2)d$.
By $\partial_jf$ we denote the $j$-th partial derivative of a function $f$.
Recall that the Hilbert-Schmidt norm $\|A\|_{HS}$ of a matrix $A$
is monotone in each of its arguments and dominates the operator norm
$\|A\|$.
\begin{lemma}[The inductive statement]\label{indstep}
Let $m\ge 1$, $d\geq 1$. Let $0\le l\le m$. Let $0<\epsilon<1$.
There exists a constant $C$ depending on these parameters
such that the following holds.
Let $A$ be an $m\times m$ matrix such that
\begin{equation}
\label{aepsilonct}
|\det ((I \ A)\Pi_j^T) |>\epsilon \quad \textup{and} \quad \|A\|_{HS}\le \epsilon^{-1}
\end{equation}
for all $1\le j\le m$.
Assume that the first $l$ rows of $A$ coincide with the first
$l$ rows of $-I$. Let $(F_j)_{j\in Q}$ be a tuple of real valued Schwartz functions with
\begin{align}\label{reflectct}
F_j=F_{i*j} \quad \textup{and} \quad \|F_j\|_{2^m}=1
\end{align}
for all $j\in Q$ and all $1\le i\le l$. Then the following two estimates hold for \eqref{sblA}.
\begin{enumerate}
\item Let $K$ be a kernel such that
$$|\partial^\alpha \widehat{K}(\xi)|\le |\xi|^{-|\alpha|}$$
for all multi-indices $\alpha \in \mathbb{N}_0^{dm}$ with $|\alpha|\le 3dm$ and
\begin{align}
\label{vanishing}
\widehat{K}(\xi_1,\dots, \xi_l, 0,\dots ,0)\equiv 0,
\end{align}
that is, $\widehat {K}$ vanishes for all $0\neq (\xi_1,\ldots , \xi_m)\in (\mathbb{R}^d)^m$ with $\xi_k=0$ for $k>l$. Then
$$|\Lambda(K,A)|\le C.$$
\item Let $l<i\le m$ and $1\leq k_1,k_2 \leq d$. Let $u \in \mathbb{R}^{dm}$ and let $c\in \L^\infty(0,\infty)$ with $\|c\|_\infty=1$. Let $K$ be the kernel defined by
$$\widehat{K}(\xi) = \int_0^\infty c_t(u)\, \widehat{g_{i,k_1,k_2}} ((I\ A)^T (t\xi)) e^{2\pi i u\cdot t\xi} \frac{dt}{t},$$
where $g_{i,k_1,k_2} := \partial_{(i-1)d+k_1} \partial_{(i+m-1)d+k_2} g$.
Then
$$|\Lambda(K,A)|\le C(1+\|u\|)^{2d(m-1)}.$$
\end{enumerate}
\end{lemma}
Note that the case $l=m$ of (1) is trivially true since then
$K=0$. On the other hand, (2) is void for $l=m$ since then $l<i\leq m$ does not exist.
The case $l=0$ of (1) implies the desired Theorem
\ref{mainthm}. We will therefore do an induction on $l$,
proving Lemma \ref{indstep} assuming that we have already established the lemma for all $l<l'\le m$.
We will reduce (1) at level $l$ to (2) at the same level $l$, and we will reduce (2) at level $l$ to (1) at the level $l+1$. These two reductions will be performed in the following two sections.
Note that in the case $m=1$ we are dealing with a one-dimensional Calder{\'on}-Zygmund kernel and the claim follows from the standard Calder{\'o}n-Zygmund theory. We shall therefore assume $m\geq 2$.
\section{Proof of (1) of Lemma \ref{indstep} }
Consider $m,d,l,\epsilon$ as in Lemma \ref{indstep}. We shall
prove existence of a constant $C$ such that (1) holds,
under the hypothesis that for the same $m,d,l,\epsilon$
there is a constant $C$ such that (2) holds.
Let $A$, $(F_j)_j$ and $K$ be given as in (1) of Lemma \ref{indstep}.
Our aim is to decompose ${K}$ into a convergent sum and integral of kernels defined in $(2)$ of Lemma \ref{indstep}.
We will perform a cone decomposition of $\widehat{K}$. The matrix $A$
determines certain subspaces of $(\mathbb{R}^d)^m$, and each cone will be small enough to avoid some of these subspaces, as elaborated in the following lemma.
In this section we use the notational convention
$$\xi=(\xi',\xi'')\in \mathbb{R}^{dl}\times \mathbb{R}^{d(m-l)}=\mathbb{R}^{dm}.$$
\begin{lemma}\label{geoct}
There is a number $\delta>0$ depending on $\epsilon$, $d$, and $m$, such that the following holds. For $\gamma$ a unit vector in $\mathbb{R}^{d(m-l)}$ define the stick
$$S=\left\{(0,\xi'') \in \mathbb{R}^{dm}: \frac 12 \le \|\xi''\|\le 1,
\left\|\frac {\xi''}{\|\xi''\|} -\gamma \right\|\le \delta\right\}.$$
Then there is $l< i\leq m$ and some $1\leq k_1,k_2 \leq d$ such that for all $\eta\in S$ we have
\begin{equation}\label{geolemmact}
\min( |\eta_{ik_1}|,|(A^T\eta)_{ik_2}| ) > \delta,
\end{equation}
where we write $\eta=(\eta_1,\ldots, \eta_m)^T$, $\eta_j=(\eta_{j1},\ldots, \eta_{jd})^T$ for $1\leq j \leq m$, and analogously we write the coordinates of $A^T\eta$.
\end{lemma}
\begin{proof}
We first claim that $S$ contains a point $\xi$
such that there is $l< i\leq m$ with
\begin{align}\label{cancellativecond1ct}
\min(\|\xi_i\|, \|(A^T\xi)_i\|) > 4d\max(1,\|A\|)\delta .
\end{align}
Assume to get a contradiction that the claim is false.
For every $\xi\in S$ we choose $j\in Q$ such that for
$l < i\le m$ the value of $j(i)$ corresponds to which term on the left hand side of \eqref{cancellativecond1ct} is less than or equal to the right-hand side. Hence we obtain
\begin{align}\label{nonconcellativect}
\|(A_j^T\xi)_i\|\leq 4d\max(1,\|A\|)\delta,
\end{align}
where we have denoted $A_j:=(I\ A)\Pi_j^T.$
By pigeonholing with respect to the $2^m$ elements of $Q$, there exists $j\in Q$ and $S'\subseteq S$ of size $|S'|\geq 2^{-m}|S|$ such that
\eqref{nonconcellativect} holds for this same $j$ and
all $\xi \in S'$ and $l< i \leq m$.
To obtain a contradiction, we compare the volume of $\Psi S'$,
where $\Psi$ is projection onto the $d(m-l)$ dimensional space spanned by the last components,
with that of the linear image
$$\Psi A_j^T \Psi^T \Psi S'=\Psi A_j^T S'.$$
We obtain
\begin{equation}\label{stickballct}
c(d,m) \delta^{d(m-l)-1} \leq |\Psi S'|
= |\det (\Psi A_j^T \Psi^T) |^{-d} |\Psi A_j^TS'|
\leq C(d,m,\epsilon) \delta^{d(m-l)}
\end{equation}
with positive constants $c(d,m)$ and $C(d,m,\epsilon)$.
On the left hand side we used the growth in $\delta$
of the volume of the stick. On the right hand side
we used that the first $l$ rows of $A$ equal those of $-I$ and thus
$$\epsilon< |\det(A_j^T)|=|\det(\Psi A_j^T \Psi^T)|, $$
and we estimated the size of the ball with radius $\delta$
in $\mathbb{R}^{d(m-l)}$ that contains $\Psi A_j^T S'$ by virtue of \eqref{nonconcellativect}.
Choosing $0<\delta<0.1$ small enough depending on $d,m,\epsilon$,
inequality \eqref{stickballct} is a contradiction, thereby proving the claim.
By the triangle inequality, the $\xi$ obtained via the claim also satisfies
\begin{equation*
\min(|\xi_{ik_1}|,|(A^T\xi)_{ik_2}|) \ge 4\max(1,\|A\|)\delta
\end{equation*}
for some
$1\leq i\leq m$ and $1\leq k_1,k_2 \leq d$.
To prove the desired lower bound \eqref{geolemmact} for every $\eta\in S$, since $1/2\le \|\eta\|,\|\xi\| \le 1$, it suffices by scaling to show the analoguous bounds with $2\delta$ on the right-hand side under the assumption that $\|\eta\|=\|\xi\|$.
Then $|\eta-\xi|\leq \delta$ and $|A(\eta-\xi)|\leq \|A\|\delta$. Thus
\begin{align*}
|(A^T\eta)_{ik_2}| \geq |(A^T\xi)_{ik_2}| - |(A^T(\eta-\xi))_{ik_2}| \geq 4\max(1,\|A\|)\delta - \|A\|\delta > 2\delta ,
\end{align*}
and similarly
\begin{align*}
|\eta_{ik_1}| \geq |\xi_{ik_1}| - |\eta_{ik_1}-\xi_{ik_1}| > 2\delta .
\end{align*}
This completes the proof of Lemma \ref{geoct}.
\end{proof}
We proceed to decompose $K$. Let $\delta$ be as in the above Lemma \ref{geoct}.
Consider a maximal set $\Gamma$ of $\delta/6$-separated vectors
of unit length in $\mathbb{R}^{d(m-l)}$. By volume considerations on
the unit sphere, there are at most $C(d,m) \delta^{-d(m-l)}$
elements in $\Gamma$. The balls of radius $\delta/2$ centered around
these points cover the sphere.
For $\gamma\in \Gamma$, let $\rho_\gamma$ be a smooth nonnegative bump function in $\mathbb{R}^{d(m-l)}$ supported on a ball of radius $\delta$ about $\gamma$ and constant one on ball of radius $\delta/2$ about $\gamma$.
Then evidently $\sum_{\gamma \in \Gamma} \rho_\gamma$ is uniformly bounded below on the unit sphere and we may consider the partition
of unity of $\mathbb{R}^{md}\setminus \{ \mathbb{R}^{dl} \times \{0\}\}$ by the functions
$$f_\gamma(\xi):=\frac{\rho_\gamma(\xi''/\|\xi''\|)}
{\sum_{\gamma' \in \Gamma} \rho_{\gamma'}(\xi''/\|\xi''\|)}.$$
Note the derivative bounds
$$
|\partial^\alpha f_\gamma(\xi'')|\leq C_\alpha \|\xi''\|^{-|\alpha|}
$$
for all $\xi''\neq 0$. We write
\begin{align*}
\widehat{K}(\xi) =\sum_{\gamma} \widehat{K}(\xi)f_\gamma(\xi'')
= \sum_{\gamma} \widehat{K_\gamma}(\xi).
\end{align*}
Since the number of summands $K_\gamma$ depends only on $d$ and $m$,
we may restrict attention to an individual summand and prove
$$ |\Lambda(K_\gamma,A)|\le C. $$
Let $\psi: \mathbb{R}^{dl} \rightarrow \mathbb{R}$ and $\phi: \mathbb{R}^{d(m-l)}\rightarrow \mathbb{R}$ be radial Schwartz functions supported in the annuli
$\{1/2\leq |\eta| \leq 1\}$ in $\mathbb{R}^{dl}$ and $\mathbb{R}^{d(m-l)}$, respectively. We normalize them such that
\begin{align*}
1=\int_0^\infty \psi(t\xi') \frac{dt}{t} = \int_0^\infty \phi(t\xi'') \frac{dt}{t}
=\int_0^\infty \int_0^\infty \psi(s t\xi') \phi(t\xi'')\frac{ds}{s} \frac{dt}{t}
\end{align*}
for every $\xi',\xi'' \neq 0$.
Then for each $\xi$ with $\xi''\neq 0$ we decompose $\widehat{K}_\gamma(\xi)$
according to the small and large values of $s$ as
\begin{align}\label{k2ct}
\int_0^\infty \int_1^\infty \widehat{K_\gamma}(\xi) {\psi}(st\xi') \phi(t\xi'') \frac{ds}{s} \frac{dt}{t}
\end{align}
\begin{align}\label{k1ct}
+ \int_0^\infty \int_0^1 \widehat{K_\gamma}(\xi) {\psi}(st\xi') \phi(t\xi'') \frac{ds}{s} \frac{dt}{t} .
\end{align}
We estimate the effect of the multipliers \eqref{k2ct} and \eqref{k1ct} separately. For \eqref{k2ct} we integrate in $s$ and note that
$$\rho(\xi'):=\int_1^\infty {\psi}(s\xi')\frac{ds}{s}$$
extends to a smooth bump function with compact support in $\|\xi'\|<2$.
We then fix $t$ and rescale the
corresponding portion of the multiplier back as on the left-hand side of the following display \eqref{definektct}. Moreover, we define the multiplier $\widehat{K_t}$ by
\begin{align}\label{definektct}
\widehat{K_\gamma}(t^{-1}\xi) {\rho}(\xi') \phi(\xi'')
=: \widehat{K_t}(\xi) \widehat{g_{i,k_1,k_2}}((I \ A)^T \xi),
\end{align}
where $g_{i,k_1,k_2}$ is defined in (2) of Lemma \ref{indstep}
for suitable $i,k_1,k_2$.
To make sure that $\widehat{K_t}$ is well defined and well behaved, we need that the second factor on the right-hand side is bounded away from $0$ on the compact support of the left-hand side. By Lemma \ref{geoct}, there exist $l+1\leq i\leq m$ and $1\leq k_1,k_2 \leq d$ such that for each $\xi$ in the support of
the left-hand side of \eqref{definektct} we have
\begin{align*
|\xi_{ik_1}| > \delta \quad \textup{and} \quad |(A^T\xi)_{ik_2}| > \delta .
\end{align*}
Since $\widehat{g_{i,k_1,k_2}}$ vanishes only at
$\xi_{ik_1}=0$ and $(A^T\xi)_{ik_2}=0$, it
is bounded uniformly away from $0$ on the support of the left-hand side of \eqref{definektct}. Therefore, the function $\widehat{K_t}$ is well defined, smooth, and satisfies some uniform bounds
$$|\partial^\alpha {\widehat{K_t}}(\xi)|\leq C $$
uniformly in $t$ for all $|\alpha| \leq 3dm$.
We expand it into its Fourier integral
\begin{align*}
\widehat{K_t}(\xi) = \int_{\mathbb{R}^{dm}} K_t(u)e^{2\pi i u \cdot \xi} du.
\end{align*}
Integrating by parts, using the derivative estimates up to order $3dm$ and bounding the size of the support of $\widehat{K_t}$ by an absolute constant times $\delta^{d m-1}$, we obtain the bound
\begin{align}\label{bd-fourier1}
|K_t(u)| \leq C (1+\|u\|)^{-3dm}.
\end{align}
Combining \eqref{definektct} and \eqref{bd-fourier1}, and rescaling back, we see that
it suffices to consider the multiplier
\begin{align*}
\int_{\mathbb{R}^{dm}} (1+\|u\|)^{-3dm}
\Big ( \int_0^\infty
\big( K_t(u) (1+\|u\|)^{3dm} \big )
\widehat{g_{i,k_1,k_2}}((I \ A)^T t \xi)
e^{2\pi i u \cdot t \xi}
\frac{dt}{t}\Big ) du .
\end{align*}
Using (2) of Lemma \ref{indstep} at level $l$ to estimate the singular Brascamp-Lieb integral
associated with the multiplier in the bracket for a fixed $u$ and integrating in $u$ we obtain
the desired bound for \eqref{k2ct}.
It remains to consider the part \eqref{k1ct}. Here we fix $0<s<1$ and consider
$$\widehat{K_s}(\xi):=\
\int_0^\infty \widehat{K_\gamma}(\xi) {\psi}(st\xi') \phi(t\xi'') \frac{dt}{t}. $$
We will prove a bound on $\Lambda(K_s,A)$ that is proportional to $s$, so that
we will be able to integrate against $ds/s$ and obtain a good bound for
the form associated with \eqref{k1ct}.
Let $D$ be the $m\times m$ diagonal matrix with $d_{ii}=s^{}$ for $i\leq l$ and
$d_{ii} = 1$ for $i>l$.
By (1) of Lemma \ref{lemma:symmetries} we have
$$\Lambda(K_s,A)=\Lambda(\widetilde{K}_s,\widetilde{A}),$$
where
$$\widetilde{K}_s(\xi)=\det(D)^{d} K_s (D\xi),\ \widetilde{A}=D^{-1}AD.$$
Recall that the first $l$ rows of $A$ coincide with the first $l$ rows of $-I$,
hence we may view $A$ as lower triangular block matrix relative to the splitting
$$\mathbb{R}^{dl}\times \mathbb{R}^{d(m-l)}.$$ The matrix $\widetilde{A}$ arises by multiplying the non-trivial
off diagonal block by $s\le 1$. Hence
\begin{align*}
\|\widetilde{A}\|_{HS}\leq \|A\|_{HS} \leq 1/\epsilon, \quad |\det ((I\ \widetilde{A})\Pi_j)|=|\det ((I\ {A})\Pi_j)| > \epsilon.
\end{align*}
We thus plan to apply (2) of Lemma \ref{indstep} with the matrix $\widetilde{A}$.
We note
$$\widehat{\widetilde{K}_s}(\xi)=\
\int_0^\infty \widehat{K_\gamma}(D^{-1} \xi) {\psi}(t\xi') \phi(t\xi'') \frac{dt}{t} .$$
Now we fix in addition $t$ and rescale similarly to \eqref{definektct}. We set
\begin{equation}
\label{defktsct}
\widehat{K_\gamma}(t^{-1}D^{-1}\xi) \psi(\xi')\phi(\xi'')=:\widehat{{K}_{t,s}}(\xi)
\widehat{g_{i,k_1,k_2}} ((I\ \widetilde{A})^{T} \xi)
\end{equation}
with some suitable $l+1\leq i \leq m$ and $1\leq k_1,k_2 \leq d$ from Lemma \ref{geoct}.
Similarly as in the discussion of \eqref{definektct}, on the compact
support of the left-hand side, $\xi'\sim 1$ and $\xi''\sim 1$, the second factor on the right hand side is bounded below, so the function $\widehat{{K}_{t,s}}$ is well defined.
We now claim that
$$|\partial^\alpha \widehat{K_{t,s}}(\xi)|\le s C $$
uniformly in $t$ for all multi-indices $\alpha$ up to order $3dm-1$. To see this, we
need to show the analoguous estimate for the left hand side of \eqref{defktsct}.
Applying a partial derivative on the left-hand side, we apply the Leibniz rule
and consider the terms separately.
By $\partial_{ik}f$ we denote the $((i-1)d+k)$-th partial derivative of a function $f$ on $\mathbb{R}^{md}$, $1\leq i \leq m, 1\leq k\leq d$.
If one derivative $\partial_{ik}$ with $i>l$, $1\leq k \leq d$,
falls on $\widehat{K_\gamma}(t^{-1}D^{-1}\xi)$, we estimate
$$\partial_{ik}(\widehat{K_\gamma}(t^{-1}D^{-1}\xi))\le C t^{-1}(t^{-1}s^{-1}\|\xi'\|+ t^{-1}\|\xi''\|)^{-1}\le Cs$$
since both $\xi'$ and $\xi''$ can be assumed of unit length. Similarly we estimate
if more than one derivative $\partial_{ik}$ with $i>l$
falls on $\widehat{K_\gamma}(t^{-1}D^{-1}\xi)$. If no such derivative falls on $\widehat{K_\gamma}(t^{-1}D^{-1}\xi)$, then only partial derivatives $\partial_{ik}$
with $i\le l$ fall on $\widehat{K}(t^{-1}D^{-1}\xi)$. Restricting attention to one such derivative we use the vanishing condition \eqref{vanishing} to obtain with the fundamental theorem of calculus
\begin{align*}\partial_{ik} (\widehat{{K_\gamma}}(t^{-1}D^{-1} \xi)) & = \partial_{ik}
\int_0^1 \partial_h ( \widehat{{K_\gamma}}((ts)^{-1}\xi',h t^{-1}\xi'') ) \, dh\\
&
= (ts)^{-1} \int_0^1 t^{-1}\xi''\cdot (\nabla \partial_{ik} \widehat{K_\gamma}) ((ts)^{-1}\xi',h t^{-1}\xi'') dh,
\end{align*}
where $\nabla$ denotes the gradient in the last $d(m-l)$ components.
The desired estimate now follows through derivative estimates for $\widehat{K_\gamma}$
with one degree higher than $|\alpha|$, note the gain of the factor $s$ comes from the length of $\xi''$ relative to the length of $\xi$ in the relevant support.
As before, we expand the Fourier integral
\begin{align*}
\widehat{{K}_{t,s}}(\xi) = \int_{\mathbb{R}^{dm}} {K}_{t,s}(u)e^{2\pi i u \cdot \xi} du
\end{align*}
and we observe the bound
\begin{align*
|{K}_{t,s}(u)| \leq C s (1+\|u\|)^{-3dm+1}.
\end{align*}
It suffices to consider the multiplier
\begin{align*}
\int_{\mathbb{R}^{dm}} (1+\|u\|)^{-3dm+1}
\Big ( \int_0^\infty
\big( {K}_{t,s}(u) (1+\|u\|)^{3dm-1} \big )
\widehat{g_{i,k_1,k_2}}((I \ \widetilde{A})^T t \xi)
e^{2\pi i u \cdot t \xi}
\frac{dt}{t}\Big ) du.
\end{align*}
We again apply (2) of Lemma \ref{indstep} at level $l$ and integration in $v$ and $s$ to obtain the desired bound.
\section{Proof of (2) of Lemma \ref{indstep} }
Consider $m,d,l,\epsilon$ as in Lemma \ref{indstep}. We shall
prove existence of a constant $C$ such that (2) holds,
under the hypothesis that for the same $m,d$ but for
$l$ replaced by $l+1$ and for $\epsilon$ replaced by
possibly much smaller $\tilde{\epsilon}$ depending on $d,m,\epsilon$,
there is a constant $C$ such that (1) holds.
Let $A$ be as in Lemma \ref{indstep}. Recall that
the first $l$ rows of $A$ coincide with the first $l$ rows of $-I$.
We shall assume $l<m$ because the case $l=m$ is void.
With $i,k_1,k_2$ as in (2) of Lemma \ref{indstep}, we need to estimate
the form associated with the multiplier
$$\widehat{K}(\xi) = \int_0^\infty c_t(u)\, \widehat{g_{i,k_1,k_2}} ((I\ A)^T (t\xi)) e^{2\pi i u\cdot t\xi} \frac{dt}{t}.$$
Let us first compute the kernel and the form on the spatial side. We have
\begin{align*}
K((I\ A)x) & = \int_{(\mathbb{R}^d)^{m}} \widehat{K}(\xi)e^{2\pi i \xi\cdot ((I\ A)x)} d\xi\\
& = \int_0^\infty c_t(u) \int_{(\mathbb{R}^d)^{m}} \widehat{g_{i,k_1,k_2}}((I\ A)^T(t\xi))e^{2\pi i u\cdot t\xi} e^{2\pi i ((I\ A)^T\xi) \cdot x} d\xi \frac{dt}{t} \\
& = \int_0^\infty c_t(u) \int_{(\mathbb{R}^d)^{m}} (g_{i,k_1,k_2})_t(x+(-Ap+ut, p)) dp \frac{dt}{t} ,
\end{align*}
where we write $f_t(\cdot )=t^{-{2dm}}f(t^{-1} \cdot )$ for a function $f$ in dimension $2dm$.
The last equality is verified noting that the right-hand side is the integral of the function
$$y\mapsto (g_{i,k_1,k_2})_t(x+y+(ut,0))$$
over the subspace $\{(-A^T\ I)^Tp:p\in (\mathbb{R}^d)^m\}$, while the left-hand side is the integral
of the Fourier transform of this function over the orthogonal subspace
$$\{(I\ A)^T\xi: \xi\in (\mathbb{R}^d)^m\}.$$
Using the definition of $g_{i,k_1,k_2}$ and Fubini, we obtain for the associated form $\Lambda(K,A)$
\begin{align}\label{lambdakact}
\nonumber
&\int_0^\infty c_t(u) \int_{(\mathbb{R}^d)^m} \int_{(\mathbb{R}^{d})^{2m-2}}
\Big( \int_{\mathbb{R}^d} \Big(\prod_{j(i)=0} F_j(\Pi_j x)\Big)
(\partial_{k_1} g)_{t}(x_i^0+(-Ap+ut)_i)
\, dx_i^0\Big )\\ \nonumber
& \Big( \int_{\mathbb{R}^d} \Big( \prod_{j(i)=1} F_j(\Pi_j x)\Big)
(\partial_{k_2} g)_{t}(x_i^1+p_i)
\, dx_i^1\Big )\\
& g_{t}\big ((x^0-Ap+ut)_{h\neq i}, (x^1+p)_{h\neq i}\big )\,
d ((x^0)_{h\neq i}, (x^1)_{h\neq i} )
\, dp \frac {dt}t .
\end{align}
We next prove a particular case of the desired inequality. The
particular case is defined by the assumptions $1\le i\le l+1$, $k_1=k_2=:k$,
$c_t=1$ for all $t>0$, $u=0$, and in addition to the symmetries
stated in the lemma, also $F_{(l+1)*j}=F_j$ for all $j\in Q$,
and the $(l+1)$-st row of $A$ also coincides with the $(l+1)$-st row of $-I$. Note all assumptions are more specific than in (2) of
Lemma \ref{indstep}, except that we on purpose allow $i\le l$ here.
We then recognize that the first bracket in the last display becomes equal to the second bracket by the conditions on $i$, $u$, and $A$, and the reflection symmetries of the tuple $(F_j)_{j\in Q}$. The two brackets therefore form a square. As Gaussians are positive and $c_t(u)$ is positive, the entire form becomes non-negative. This holds for all $1\le i\le l+1$ and all $k$. Therefore, instead of proving bounds for each of these terms, it suffices to estimate the sum of all these terms
over $1\le i\le l+1$ and $k$, which has better algebraic properties.
To identify the good properties of this sum, note it is associated with the multiplier
$$\widehat{K}_\Sigma(\xi):= \sum_{i=1}^{l+1} \sum_{k=1}^{d} \int_0^\infty \widehat{g_{i,k,k} }((I\ A)^T (t\xi)) \frac{dt}{t}.$$
We will add and subtract $\pi$ from this multiplier. We will estimate by hand the form associated with $\pi$, and we will apply the induction hypothesis to $\widehat{K}_\Sigma(\xi)-\pi$.
The form associated with $\pi$ on the spatial side is $\pi$ times
$$\Lambda(\delta_0 ,A) = \int_{(\mathbb{R}^d)^{2m}}\Big ( \prod_{j\in Q} F_j(\Pi_j x)\Big )\delta_0((I\ A)x)dx ,$$
where $\delta_0$ denotes the Dirac delta distribution. This is a standard Brascamp-Lieb integral. Applying the arithmetic-geometric mean inequality at every point $x$ and pulling the arithmetic mean out of the integral, we bound the last display by
\begin{align*}
2^{-m}\sum_{j\in Q} \Big ( \int_{(\mathbb{R}^d)^{2m}} F_j(\Pi_j x)^{2^m} \delta_0((I\ A)x)dx \Big ).
\end{align*}
This is an average over $j\in Q$, and it suffices to prove bounds for fixed $j$ as follows
\begin{align*}
& \int_{(\mathbb{R}^d)^{2m}} F_j(\Pi_j(x^0,x^1)^T)^{2^m}\delta_0(x^0 + Ax^1)dx^0dx^1\\
&=\int_{(\mathbb{R}^d)^m}F_j(\Pi_j (-Ax^1,x^1)^T )^{2^m}dx^1 \\
&= |\det (\Pi_j (-A^T\ I)^T )|^{-d}\|{F}_j\|^{2^m}_{2^m} \le \epsilon^{-d}\|{F}_j\|^{2^m}_{2^m}.
\end{align*}
In the last inequality we used the assumption on $A$ and that the absolute value of the determinant in this display is equal to
$$ |\det (\Pi_\ell (I \ A)^T )| = |\det ( (I \ A) \Pi_\ell^T )|\ge \epsilon, $$
where $\ell$ is the corner of the cube opposite to $j$, that is $j(i)+\ell(i)=1$ for all $i$.
This completes the bound for the multiplier $\pi$.
To estimate the form associated with $\widehat{K}_\Sigma-\pi$, we apply (1) of Lemma \eqref{indstep} for $l+1$ and $\epsilon$ to a suitably normalized kernel. Most assumptions of (1) are straightforward, the main
difficulty is the vanishing condition \eqref{vanishing}.
Using
\begin{align*}
\widehat{g'}(-\eta) = \overline{\widehat{g'}(\eta)}, \quad \widehat{g}(0)=1
\end{align*}
for a Gaussian $g$ on $\mathbb{R}$ and the assumption that the first $l+1$ rows of $A$ are
equal to the first $l+1$ rows of $-I$, we obtain
\begin{align}\label{kxizeroct}
\widehat{K}_\Sigma(\xi_1,\ldots \xi_{l+1},0,\ldots,0) = \sum_{i=1}^{l+1} \sum_{k=1}^{d} \int_0^\infty |\widehat{\partial_{k} g}(t\xi_{i})|^2 \Big ( \prod_{j=1, j\neq i}^{l+1} |\widehat{g}(t\xi_{j})|^2 \Big ) \frac{dt}{t}.
\end{align}
Observe the elementary identity
\begin{align}
\label{heat}
-t\partial_t |\widehat{g}(t\eta)|^2 = \frac{1}{\pi} |\widehat{g'}(t\eta)|^2
\end{align}
valid for a one-dimensional Gaussian.
Since $\widehat{g}(t\xi_i)$ is a product of one-dimensional Gaussians $\widehat{g}(t\xi_{i1})\cdots \widehat{g}(t\xi_{id})$, together with the Lebniz rule the identity \eqref{heat} implies
\begin{align}
\label{heat2}
-t \partial_t|\widehat{g}(t\xi_i)|^2 = \frac{1}{\pi}\sum_{k=1}^d |\widehat{\partial_k g} (t\xi_i)|^2 .
\end{align}
By \eqref{heat2}, the fundamental theorem of calculus and another application of
the Leibniz rule, we equate \eqref{kxizeroct} with
\begin{align*
-\pi \int_0^\infty t\partial_t \Big ( \prod_{j=1}^{l+1} |\widehat{g}(t\xi_j)|^2\Big ) \frac{dt}{t} = \pi.
\end{align*}
This completes verification of \eqref{vanishing} for $\widehat{K}_\Sigma-\pi$ and establishes the desired estimate for the associated form.
To round up the discussion, we present a derivation of the elementary identity \eqref{heat} from the heat equation
$$ \partial_t g_t (s)= \frac {t}{2\pi } \partial_s^2 g_t(s)$$
and the convolution identity
$$g_{\sqrt{2}t}(s_1-s_0)=\int_{\mathbb{R}} g_{t}(s_1-p) g_{t}(s_0-p)\, dp.$$
Indeed, integrating by parts in $p$ we obtain
\begin{align*}
& \partial_t g_{\sqrt{2}t}(s_1-s_0)\\
& =\frac t{2\pi }\int_{\mathbb{R}} \partial_p^2 g_{t}(s_1-p) g_{t}(s_0-p)\, dp+\frac t{2\pi }\int_{\mathbb{R}} g_{t}(s_1-p) \partial_p^2 g_{t}(s_0-p)\, dp\\
& = -\frac t{\pi }\int_{\mathbb{R}} \partial_p g_{t}(s_1-p) \partial_p g_{t}(s_0-p)\, dp .
\end{align*}
This can be turned into \eqref{heat} by taking the Fourier transform.
We have completed the estimate of the form associated with \eqref{lambdakact}
in the particular case. It remains to reduce the general case to the particular case. We will reduce to the particular case with $A$ replaced
by different matrices, which may satisfy \eqref{aepsilonct} with different $\tilde{\epsilon}$. These different $\tilde{\epsilon}$
however only depend on $m,d,\epsilon$.
We shall first reduce the general case to the case $i\le l+1$. This is done by a permutation of the coordinates if needed. If $i>l+1$, let $P$ be the involution that switches $i$ and $l+1$.
Applying (2) of Lemma \ref{lemma:symmetries} reduces the to new data which still satisfy
our assumptions of $(2)$ of Lemma \ref{indstep}. Henceforth we assume $i\le l+1$.
Next, we symmetrize the tuple $F_j$ and the pair $k_1,k_2$. We pull $c_t(u)$ into one of the brackets, apply Cauchy-Schwarz, and then estimate $c_t(u)$ by a constant. This bounds \eqref{lambdakact} by the geometric mean of
\begin{align}\nonumber
& \int_0^\infty \int_{(\mathbb{R}^d)^m} \int_{(\mathbb{R}^d)^{2m-2}}
\Big (\int_{\mathbb{R}^d} \Big(\prod_{j(i)=0} F_j(\Pi_j x)\Big)
(\partial_{k_1} g)_{t}(x_i^0+(-Ap+ut)_i)
\, dx_i^0\Big)^2\\
& g_{t}\big ((x^0-Ap+ut)_{h\neq i}, (x^1+p)_{h\neq i}\big)\,
d ((x^0)_{h\neq i}, (x^1)_{h\neq i})
\, dp \frac {dt}t \label{cs1ct}
\end{align}
and
\begin{align} \nonumber
&\int_0^\infty \int_{(\mathbb{R}^d)^m} \int_{(\mathbb{R}^d)^{2m-2}}
\Big(\int_{\mathbb{R}^d} \Big( \prod_{j(i)=1} F_j(\Pi_j x)\Big)
(\partial_{k_2} g)_{t}(x_i^1+p_i)
\, dx_i^1\Big)^2\\
& g_{t}\big((x^0-Ap+ut)_{h\neq i}, (x^1+p)_{h\neq i}\big )\,
d ((x^0)_{h\neq i}, (x^1)_{h\neq i})
\, dp \frac {dt}t \label{cs2ct}.
\end{align}
It suffices to bound both terms separately and we begin with
\eqref{cs2ct}. To get rid of $u$, we dominate a non-centered Gaussian by a centered Gaussian as in
\begin{align*
g (s+v) \leq 10 g \Big (\frac {s}{2+2\|v\|} \Big ).
\end{align*}
Let $v$ be the vector $u$ with the $i$-th $d$-dimensional component replaced by $0$.
Let $D$ the $m\times m$ diagonal matrix with $d_{hh}=2(1+\|v\|)$ for $h\neq i$, and $d_{ii} =1$. Using the above
domination we estimate \eqref{cs2ct} by
\begin{align*}
& \int_0^\infty \int_{(\mathbb{R}^d)^m} \int_{(\mathbb{R}^{d})^{2m-2}}
\Big ( \int_{\mathbb{R}^d} \Big(\prod_{j(i)=0} F_j(\Pi_j x)\Big)
(\partial_{k_2} g)_{t}((D^{-1}x^1)_i+p_i)
\, dx_i^0\Big )^2\\ \nonumber
& g_{t}\big ((D^{-1}x^0-D^{-1}A p)_{h\neq i}, (D^{-1}x^1+D^{-1}p)_{h\neq i}\big)\,
d ((x^0)_{h\neq i}, (x^1)_{h\neq i} )
\, dp \frac {dt}t.
\end{align*}
Replacing variables $p$ by $Dp$, $x^0$ by $Dx^0$, $x^1$ by $Dx^1$ and
using $\widetilde{F}_j$ as in (1) of Lemma~\ref{lemma:symmetries}
turns this into
\begin{align*}
\det(D)^{2d} & \int_0^\infty \int_{(\mathbb{R}^d)^m} \int_{(\mathbb{R}^{d})^{2m-2}}
\Big ( \int_{\mathbb{R}^d} \Big(\prod_{j(i)=0} \widetilde {F}_j(\Pi_j x)\Big)
(\partial_{k_2} g)_{t}(x^1_i+p_i)
\, dx_i^0\Big )^2 \\ \nonumber
& g_{t}\big ((x^0-D^{-1}ADp)_{h\neq i}, (x^1+p)_{h\neq i}\big )\,
d ((x^0)_{h\neq i}, (x^1)_{h\neq i} )
\, dp \frac {dt}t .
\end{align*}
To obtain the desired bound, it suffices to apply the particular case of
\eqref{lambdakact} with the matrix $D^{-1}\widetilde{A} D$ in place of $A$,
where $\widetilde{A}$ is the matrix whose $i$-th row is that of $-I$ and
whose other rows equal those of $A$.
In particular, the first $l+1$ rows of the matrix ${D}^{-1}\widetilde{A}{D}$
coincide with the first $l+1$ rows of $-I$, and we have
\begin{align*}
\|{D}^{-1}\widetilde{A}{D}\|_{HS}\le \|\widetilde{A}\|_{HS} \leq \|A\|_{HS}+1 \le \epsilon^{-1}+1,
\end{align*}
\begin{align*}
|\det ((I\ {D}^{-1}\widetilde{A}{D})\Pi_j^T) | \ge \inf_{\tilde{j}}| \det((I\ {D}^{-1}{A}{D})\Pi_{\tilde{j}}^T) | >\epsilon.
\end{align*}
Note that we have the upper bound
$$\det(D)^{2d}\le (1+\|u\|)^{2d(m-1)},$$
which is the additional factor in (2) of Lemma \ref{indstep}.
This concludes the estimate of the term \eqref{cs2ct}.
It remains to estimate the term \eqref{cs1ct}. We reduce it to
the previous case \eqref{cs2ct} by a $t$-dependent affine linear change of variables
$$\widetilde{p}= -A p + ut . $$
This reduces \eqref{cs1ct} to
\begin{align*}\nonumber
|\det(A)|^{-d} & \int_0^\infty \int_{(\mathbb{R}^d)^m} \int_{(\mathbb{R}^d)^{2m-2}}
\Big (\int_{\mathbb{R}^d} \Big(\prod_{j(i)=0} F_j(\Pi_j x)\Big)
(\partial_{k_1} g)_{t}(x_i^0+\widetilde{p}_i )
\, dx_i^0\Big)^2\\
& g_{t}\big ((x^0+\widetilde{p})_{h\neq i}, (x^1-A^{-1}\widetilde{p}+A^{-1}ut)_{h\neq i}\big)\,
d ((x^0)_{h\neq i}, (x^1)_{h\neq i} )
\, d\tilde{p} \frac {dt}t .
\end{align*}
Interchanging the roles of $0$ and $1$ in the range of $j$
reduces this to the previous case with an additional factor
$|\det(A)|^{-d}$, $A$ replaced by $A^{-1}$ and with $u$ replaced by $A^{-1}u$.
As the first $l+1$ rows of $A^{-1}$ coincide with the first $l+1$ rows of $-I$, it remains to show the conditions \eqref{aepsilonct} for $A^{-1}$
for some $\tilde{\epsilon}$ depending on $\epsilon, m, d$.
The entries of $A^{-1}$ can be estimated by Cramer's rule
by
$$\|A\|_{HS}^{m-1}\det(A)^{-1}\le \epsilon^{-m}$$
and hence
$$\|A^{-1}\|_{HS}\le m\epsilon^{-m}.$$
Estimating the determinants of
$(I \ A^{-1})\Pi_j^T$ in absolute value from below is tantamount
to estimating determinants of submatrices of $A^{-1}$ obtained
by deleting any number of pairs of matching rows and columns.
Considering block decompositions with squares on the diagonal
$$\ A = \left ( \begin{array}{cc}
A_{11} & A_{12}\\
A_{21}& A_{22}
\end{array} \right ),\quad A^{-1} = \left ( \begin{array}{cc}
X_{11} & X_{12}\\
X_{21}& X_{22}
\end{array} \right ),
$$
we will show a lower bound on $\det (X_{11})$. The general case, when we delete arbitrary rows and columns of $A^{-1}$ can be deduced similarly after permuting rows and columns.
Note that $A_{22}$ is invertible, since \eqref{aepsilonct} gives a lower bound on its determinant when choosing suitable $\Pi_j$.
We successively compute
\begin{align*}
& A_{21}X_{11} +A_{22}X_{21}=0,\\
& A_{12}A_{22}^{-1}A_{21}X_{11} +A_{12}X_{21}=0,\\
& A_{12}A_{22}^{-1}A_{21}X_{11} -A_{11}X_{11}=-I.
\end{align*}
A lower bound on $\det(X_{11})$ follows from an upper bound on the determinant of
$$A_{12}A_{22}^{-1}A_{21} -A_{11}.$$
Such bound follows from an upper bound on the norm of this matrix.
Upper bounds on the norms of $A_{12}$, $A_{22}$, $A_{11}$ are obtained
using the bound on the Hilbert Schmidt norm of $A$, while the bound on the norm
of $A_{22}^{-1}$ uses Cramer's rule as above and the lower bound on the determinant of $A_{22}$. Note finally that
$$(1+\|A^{-1}u\|)^{2d(m-1)}\le (m^d\epsilon^{-md})^{2dm}(1+\|u\|)^{2d(m-1)},$$
which is up to a constant dominated by the factor in (2) of Lemma \ref{indstep}.
\section*{Acknowledgments}
The authors thank Vjekoslav Kova\v{c} and Kristina Ana \v{S}kreb for inspiring discussions
aided by the bilateral DAAD-MZO
grant {\em Multilinear singular integrals and applications}.
The second author acknowledges support
by the Hausdorff Center for Mathematics and the Deutsche Forschungsgemeinschaft through the Collaborative Research Center 1060.
The authors thank the anonymous referee for a number of thoughtful comments.
|
\section{Question initiale}\label{t22}
L'origine du texte qui suit est l'étude de la question suivante.
\begin{ques}\label{t20}
Soit $c : \, {\mathbb Q}^{+*} \rightarrow {\mathbb C}$ une fonction telle que
\[
c(1/r) =\overline{c(r)} \quad ( r \in {\mathbb Q}^{+*}),
\]
et telle que, pour tout $ n \in {\mathbb N}$, le déterminant
\[
D_n =\det\big(c(i/j)\big)_{1\leqslant i,j \leqslant n}
\]
soit strictement positif (par convention, $D_0=1$).
La suite
\[
D_n^{1/n} \quad ( n \in {\mathbb N}^*)
\]
a-t-elle une limite quand $n$ tend vers l'infini?
\end{ques}
\smallskip
Si $c : {\mathbb Q}^{+*} \rightarrow {\mathbb C}$ est une fonction quelconque (ne vérifiant pas nécessairement la condition~$c(1/r) =~\overline{c(r)}$), la matrice Toeplitz-multiplicative infinie associée à $c$ est
\[
M=\big(c(i/j)\big)_{ i,j\geqslant 1}
\]
Cette définition, due à Hilberdink (2006, cf. \cite{zbMATH05082236}), est analogue à celle, classique, de matrice de Toeplitz (additive)
\[
\big(c_0(i-j)\big)_{ i,j\geqslant 1}
\]
où $c_0$ est définie sur ${\mathbb Z}$ (on numérote alors parfois les lignes et colonnes de la matrice à partir de~$0$, mais nous garderons la même convention pour les deux cas).
\`A la fonction $c$ on peut associer la suite des matrices $M_n$ obtenues à partir de $M$ en ne retenant que les lignes et colonnes d'indices $\leqslant n$, où $n \in {\mathbb N}^*$. On pose
\[
D_n=\det M_n=\det\big(c(i/j)\big)_{1\leqslant i,j\leqslant n} \quad (n \in {\mathbb N}^*).
\]
Nous dirons que la suite $(D_n)$ ainsi définie est une suite de déterminants Toeplitz-multiplicatifs.
\smallskip
La condition $c(1/r) =\overline{c(r)}$ assure de plus que les matrices $M_n$ sont hermitiennes, et un critère classique de Sylvester montre que la question posée concerne le cas où les matrices $M_n$ sont toutes définies positives.
Cette question est motivée par le fait que son analogue pour les déterminants de Toeplitz {\og additifs\fg} a une réponse positive. On a même un résultat plus fort, signalé par Fekete à Szeg\H o, et mentionné dans le premier article de ce dernier (1915, \cite{zbMATH02616579}, \S 3, p. 494-495) : si la fonction $c_0$ vérifie $c_0(-i)=\overline{c_0(i)}$ pour tout $i \in {\mathbb Z}$, et si les déterminants
\[
\Delta_n=\det\big(c_0(i-j)\big) _{1 \leqslant i,j \leqslant n}
\]
sont tous strictement positifs, alors la suite $\Delta_n/\Delta_{n-1}$ est décroissante, donc convergente. La démonstration classique de ce fait est rappelée au \S\ref{t3}. Il en résulte que la suite $\Delta_n^{1/n}$ est aussi convergente. Dans le cas des déterminants Toeplitz-multiplicatifs hermitiens positifs, nous verrons au même paragraphe \S\ref{t3} une propriété analogue de décroissance {\og multiplicative\fg} ; avec les hypothèses et les notations de la question ci-dessus, on a
\[
k \mid n \Rightarrow D_n/D_{n-1} \leqslant D_k/D_{k-1} \quad (k,n \in {\mathbb N}^*).
\]
Dès lors, en oubliant provisoirement les spécificités des déterminants Toeplitz-multiplicatifs, il est légitime de s'interroger, dans un cadre plus général, sur les conséquences, pour le comportement asymptotique d'une suite donnée, de l'hypothèse de monotonie multiplicative ; c'est l'objet des paragraphes \S\S\ref{t23}-\ref{t24}. Le résultat principal est l'existence d'une valeur moyenne logarithmique, démontrée à la proposition \ref{t8} au \S\ref{t26}. Après ce détour, je reviendrai aux déterminants Toeplitz-multiplicatifs au \S\ref{t3} ; j'y donnerai un début de réponse à la question initiale, qui restera néanmoins ouverte.
\section{Fonctions arithmétiques multiplicativement monotones}\label{t23}
Au lieu du terme {\og suite\fg}, nous pouvons employer le terme {\og fonction arithmétique\fg}. Il s'agit formellement des mêmes objets ; le changement de vocabulaire induit simplement un changement de point de vue, ici pertinent.
\subsection{Définition}
Nous dirons d'une fonction arithmétique $f$ à valeurs réelles qu'elle est \emph{multiplicativement croissante} si elle vérifie
\[
k \mid n \Rightarrow f(k) \leqslant f(n) \quad (k,n \in {\mathbb N}^*).
\]
La fonction $f$ est \emph{multiplicativement décroissante} si la propriété analogue a lieu avec la dernière inégalité changée de sens ; elle est \emph{multiplicativement monotone} si elle est multiplicativement croissante ou multiplicativement décroissante.
\subsection{Exemples}\label{t4}
Toute fonction arithmétique monotone au sens usuel est multiplicativement monotone. Voici trois autres classes d'exemples de fonctions multiplicativement monotones.
\subsubsection{Ensembles de multiples}
Les fonctions arithmétiques, multiplicativement croissantes, et à valeurs dans $\{0,1\}$, sont exactement les fonctions indicatrices des \emph{ensembles de multiples}. Rappelons que l'ensemble des multiples de la partie $A$ de ${\mathbb N}^*$ est défini par
\[
{\mathcal M}(A)=\{ka, \, a\in A, \, k \in {\mathbb N}^*\}.
\]
\subsubsection{Fonctions dont la dérivée au sens de Bougaïef est de signe constant}\label{t9}
Bougaïef a défini la {\og dérivée numérique\fg} d'une fonction arithmétique $f$ par la formule
\[
Df=f*\mu,
\]
(où $*$ désigne la convolution des fonctions arithmétiques, et $\mu$ la fonction de Möbius), et développé la théorie des fonctions arithmétiques à partir de cette notion dans les années 1870 (cf. \cite{bugaev1876} pour un résumé en français). Sa terminologie est imparfaite, notamment parce que cette {\og dérivation\fg} ne vérifie pas de règle de Leibniz\footnote{Il ne faut pas confondre la dérivée au sens de Bougaïef avec la dérivée $f'(n)=f(n)\ln n$ de la théorie des nombres premiers, qui, elle, vérifie la règle de Leibniz $(f*g)'=f'*g+f*g'$.}. Néanmoins, la relation inverse
\[
f(n)=\sum_{d \mid n} (Df)(d),
\]
est décrite par Bougaïef comme le fait que $f$ est l'{\og intégrale numérique\fg} de $Df$ ; ses deux définitions ont donc une certaine cohérence.
Si $Df\geqslant 0$, alors $f$ est multiplicativement croissante. En effet,
\[
k \mid n \Rightarrow f(n)=\sum_{d \mid n} (Df)(d) \geqslant \sum_{d \mid k} (Df)(d)=f(k).
\]
La réciproque est fausse : la dérivée au sens de Bougaïef de la fonction indicatrice de l'ensemble des multiples de $2$ et $3$ vaut $-1$ pour $n=6$.
\smallskip
\subsubsection{Déterminants Toeplitz-multiplicatifs hermitiens positifs}
Comme nous l'avons signalé au \S \ref{t22}, les quotients $D_n/D_{n-1}$, formés à partir d'une suite de déterminants Toeplitz-multiplicatifs hermitiens positifs, constitue une suite multiplicativement croissante. Cela sera démontré au \S\ref{t3}.
\subsection{Premières propriétés}
Nous énonçons dans ce paragraphe quelques propriétés générales, et immédiates, de la monotonie multiplicative, en nous limitant au cas de la croissance, celui de la décroissance étant analogue.
\smallskip
{\bf 1.} Une fonction arithmétique $f$ multiplicativement croissante est bornée inférieurement. En effet, on a $f(n)\geqslant~f(1)$.
{\bf 2.} Si $f$ est une fonction arithmétique multiplicativement croissante et $\alpha$ une fonction à valeurs réelles, définie et croissante sur une partie de ${\mathbb R}$ contenant l'image de $f$, alors $\alpha\circ f$ est multiplicativement croissante.
{\bf 3.} L'ensemble des fonctions arithmétiques multiplicativement croissantes est un cône convexe.
{\bf 4.} Si $(f_i)_{i \in i}$ est une famille de fonctions arithmétiques multiplicativement croissantes telle que
\[
f(n)=\sup_{i\in I}f_i(n) <\infty \quad (n \in {\mathbb N}^*),
\]
alors $f$ est multiplicativement croissante.
{\bf 5.} Si $f$ est une fonction arithmétique à valeurs réelles, il existe une plus petite fonction arithmétique $g$, multiplicativement croissante, et telle que $f \leqslant g$. Elle est définie par
\[
g(n)=\max\{f(d), \, d \mid n\} \quad ( n \in {\mathbb N}^*).
\]
Si $f$ est la fonction indicatrice d'une partie $A$ de ${\mathbb N}^*$, alors $g$ est la fonction indicatrice de l'ensemble de multiples ${\mathcal M}(A)$.
{\bf 6.} Si $f$ et $g$ sont deux fonctions arithmétiques à valeurs positives ou nulles, et si l'une d'elles, disons $g$, est multiplicativement croissante, alors $f*g$ est multiplicativement croissante. Cela généralise l'exemple du \S\ref{t9} :
\[
k \mid n \Rightarrow f*g(n)=\sum_{d \mid n} f(d)g(n/d) \geqslant \sum_{d \mid k} f(d)g(n/d) \geqslant \sum_{d \mid k} f(d)g(k/d)=f*g(k).
\]
\section{Facteurs directs de l'ensemble des nombres entiers positifs}
Indépendamment de l'hypothèse de monotonie multiplicative, nous rappelons dans ce paragraphe un cas important d'existence de la limite des moyennes de Cesàro d'une fonction arithmétique. Cette condition suffisante s'appuie sur la notion de \emph{facteur direct}, et sera employée dans la démonstration des propositions \ref{t8} et \ref{t30}.
Une partie $A$ de l'ensemble ${\mathbb N}^*$ des nombres entiers positifs est un facteur direct s'il existe une partie $B \subset {\mathbb N}^*$ telle que tout nombre entier $n \geqslant 1$ admette une décomposition unique,
\[
n=a\cdot b \quad (a \in A, \, b \in B).
\]
La partie $B$, dite \emph{complémentaire} de $A$, est manifestement aussi un facteur direct, et $A$ est complémentaire de $B$.
Un théorème d'Erd\H os, Saffari et Vaughan (cf. \cite{zbMATH03520506}, \cite{zbMATH03636131} et \cite{zbMATH03636132}) affirme que tout facteur direct admet une densité asymptotique, égale à la somme des inverses des éléments de son complémentaire :
\begin{equation}\label{t32}
\frac 1x\sum_{\substack{b \in B\\b \leqslant x}} 1 \rightarrow \Big(\sum_{a\in A}\frac 1a\Big)^{-1} \quad ( x \rightarrow \infty)
\end{equation}
(si la série diverge, la densité est nulle).
\smallskip
Si $f$ est une fonction arithmétique et $A$, $B$ des facteurs directs complémentaires, la fonction arithmétique $A$-réduite de $f$ est définie par
\begin{equation}\label{t34}
f(n;A)= f(a) \quad ( n=ab, \, a \in A, \, b \in B).
\end{equation}
\begin{prop}\label{t31}
Soit $A$ et $B$ deux facteurs directs complémentaires de ${\mathbb N}^*$. On suppose que la somme des inverses des éléments de $A$ converge, et on pose
\[
\lambda = \Big(\sum_{a\in A}\frac 1a\Big)^{-1} (\in\, ]0,1]).
\]
Soit $f$ une fonction arithmétique à valeurs réelles, telle que
\begin{equation}\label{t35}
\sum_{a \in A} \frac{\min\big(f(a),0\big)}a > -\infty.
\end{equation}
Alors
\[
t^{-1}\sum_{n\leqslant t} f(n;A) \rightarrow \alpha(f;A) \quad ( t \rightarrow \infty),
\]
où
\[
\alpha(f;A)=\lambda \sum_{a \in A}\frac{f(a)}a \quad (\in\, ]-\infty,\infty]).
\]
\end{prop}
\noindent {\bf D\'emonstration\ }
En notant $a$ un élément générique de $A$, et $b$ un élément générique de $B$, on a
\begin{align*}
t^{-1}\sum_{n\leqslant t} f(n;A) &= t^{-1}\sum_{ab\leqslant t} f(a)\\
&=\sum_a\frac{f(a)}a (t/a)^{-1}\sum_{b\leqslant t/a}1.
\end{align*}
Si $\sum_{a}\lvert f(a)\rvert /a<\infty$, le résultat découle donc de \eqref{t32} et du théorème de Tannery (théorème de convergence dominée de Lebesgue pour les séries). Sinon, il découle de \eqref{t35} et du lemme de Fatou (ou d'un raisonnement élémentaire).\hfill$\Box$
\medskip
Par sommation partielle, on obtient aussi la relation
\begin{equation*}
(\ln t)^{-1}\sum_{n\leqslant t} \frac{f(n;A)}n \rightarrow \alpha(f;A) \quad ( t \rightarrow \infty),
\end{equation*}
si $f$ vérifie \eqref{t35}.
\section{\'Etude de la moyenne logarithmique d'une fonction multiplicativement monotone}
La proposition \ref{t8} ci-dessous généralise aux fonctions multiplicativement croissantes un résultat de Davenport et Erd\H os concernant les ensembles de multiples (1936, 1951, cf. \cite{zbMATH03023559}, \cite{zbMATH03065959}). Nos démonstrations suivent précisément l'approche élémentaire adoptée par ces auteurs dans \cite{zbMATH03065959} ; par souci d'autonomie, nous donnons néanmoins tous les détails.
\subsection{Décomposition en facteurs friable et criblé}
Soit $y>1$. Un nombre entier positif est dit $y$-friable\footnote{Dans certains textes en langue anglaise, la locution \emph{$y$-smooth} est utilisée pour désigner un tel nombre entier.} si tous ses diviseurs premiers sont $\leqslant y$. L'ensemble de ces nombres sera noté ${\mathcal S}(y)$. Un nombre entier positif est dit $y$-criblé si tous ses diviseurs premiers sont $>y$. L'ensemble de ces nombres sera noté ${\mathcal E}(y)$. Ces deux ensembles sont des facteurs directs complémentaires : tout nombre entier positif $n$ a une unique décomposition
\begin{equation*}
n=ab, \text{ où } a \in {\mathcal S}(y) \text{ et }b \in {\mathcal E}(y).
\end{equation*}
Nous dirons que $a$ est le facteur $y$-friable de $n$, et $b$ son facteur $y$-criblé.
Dans ce cas, le théorème d'Erd\H os, Saffari et Vaughan est immédiat ; on a les relations
\begin{align}
\sum_{a \in {\mathcal S}(y)} \frac 1a&= \prod_{p \leqslant y} (1-1/p)^{-1}\label{t19}\\
t^{-1} \sum_{\substack{b\leqslant t\\b\in {\mathcal E}(y}}1 &\rightarrow \prod_{p \leqslant y} (1-1/p) \quad (t \rightarrow \infty).\label{t10}
\end{align}
Si $z\geqslant y>1$, le facteur $y$-friable de $n$ divise son facteur $z$-friable, le facteur $z$-criblé de $n$ divise son facteur $y$-criblé. De plus, si
\begin{equation}\label{t12}
n=a'b', \text{ où } a' \in {\mathcal S}(z) \text{ et }b' \in {\mathcal E}(z),
\end{equation}
alors
\begin{equation*}
a'=ab'', \text{ où } b''=(a',b) \in {\mathcal E}(y).
\end{equation*}
En particulier, $n$ et $a'$ ont même facteur $y$-friable.
\smallskip
Si ${\mathcal S}(y,z)$ désigne l'ensemble des nombres entiers dont tous les diviseurs premiers $p$ vérifient l'encadrement $y<p\leqslant z$, tout élément de ${\mathcal S}(z)$ est, de façon unique, le produit d'un élément de~${\mathcal S}(y)$ et d'un élément de ${\mathcal S}(y,z)$. L'identité suivante est une variante de \eqref{t19} :
\begin{equation}\label{t16}
\sum_{n\in {\mathcal S}(y,z)}\frac 1n=\prod_{y<p\leqslant z} (1-1/p)^{-1}.
\end{equation}
\subsection{Réduction friable d'une fonction arithmétique}
Si $f$ est une fonction arithmétique, et $y>1$, la \emph{réduite $y$-friable} de $f$ est la fonction arithmétique définie par
\begin{equation}\label{t11}
f(n;y)= f(a) \quad ( n\in {\mathbb N}^* ; \, n=ab, \,a \in {\mathcal S}(y), \, b \in {\mathcal E}(y)),
\end{equation}
autrement dit, $f(n;y)=f\big(n;{\mathcal S}(y)\big)$, avec la notation \eqref{t34}.
\smallskip
Si $z\geqslant y>1$ et si $g$ est la réduite $y$-friable de $f$, alors, la réduite $z$-friable de $g$ est égale à $g$. En effet, avec les notations de \eqref{t12}, on a
\begin{equation}\label{t14}
g(n;z)=g(a')=f(a';y)=f(n;y)=g(n),
\end{equation}
puisque $n$ et $a'$ ont le même facteur $y$-friable.
\smallskip
Explicitons le cas particulier de la proposition \ref{t31} obtenu en prenant $A={\mathcal S}(y)$ et $B={\mathcal S}(y)$.
\begin{prop}\label{t15}
Soit $f$ une fonction arithmétique à valeurs réelles, et soit~$y>1$. On suppose que
\begin{equation}\label{t13}
\sum_{a \in {\mathcal S}(y)} \frac{\min\big(f(a),0\big)}a > -\infty.
\end{equation}
Alors
\[
t^{-1}\sum_{n\leqslant t} f(n;y) \rightarrow \alpha(f;y) \quad ( t \rightarrow \infty),
\]
où
\[
\alpha(f;y)=\prod_{p\leqslant y} \Big(1-\frac 1p\Big)\sum_{a \in {\mathcal S}(y)}\frac{f(a)}a \quad (\in\, ]-\infty,\infty]).
\]
On a donc aussi
\begin{equation}\label{t17}
(\ln t)^{-1}\sum_{n\leqslant t} \frac{f(n;y)}n \rightarrow \alpha(f;y) \quad ( t \rightarrow \infty).
\end{equation}
\end{prop}
\begin{prop}\label{t18}
Soit $f$ une fonction arithmétique à valeurs réelles, et soit~$z \geqslant y>1$. On suppose que \eqref{t13} est vérifiée et on considère la fonction arithmétique $g(n)=f(n;y)$ définie par \eqref{t11}.
Alors, \eqref{t13} est vérifiée en remplaçant $f$ par $g$ et $y$ par $z$, et
\[
\alpha(g;z)=\alpha(f;y).
\]
\end{prop}
\noindent {\bf D\'emonstration\ }
Compte tenu de \eqref{t16}, \eqref{t14} et \eqref{t13}, on a
\begin{equation*}
\sum_{a' \in {\mathcal S}(z)} \frac{\min\big(g(a'),0\big)}{a' }=\sum_{\substack{a \in {\mathcal S}(y)\\b\in{\mathcal S}(y,z)}} \frac{\min\big(f(a),0\big)}{ab}=
\prod_{y<p \leqslant z}(1-1/p)^{-1}\sum_{a \in {\mathcal S}(y)} \frac{\min\big(f(a),0\big)}a >-\infty.
\end{equation*}
Par conséquent,
\begin{equation*}
\alpha(g;z)=\lim_{t\rightarrow\infty}t^{-1}\sum_{n\leqslant t} g(n;z)=\lim_{t\rightarrow\infty}t^{-1}\sum_{n\leqslant t} f(n;y)=\alpha(f;y).\tag*{\mbox{$\Box$}}
\end{equation*}
\subsection{Existence de la valeur moyenne logarithmique des fonctions multiplicativement monotones}\label{t26}
Comme nous l'avons signalé ci-dessus, la proposition suivante généralise aux fonctions multiplicativement croissantes un théorème de Davenport et Erd\H os concernant les ensembles de multiples.
\begin{prop}\label{t8}
Soit $f$ une fonction arithmétique multiplicativement croissante. La quantité
\[
\alpha(f;y)=\prod_{p\leqslant y} \Big(1-\frac 1p\Big)\sum_{n \in {\mathcal S}(y)}\frac{f(n)}n
\]
est une fonction croissante de $y$. Posons
\[
\alpha =\lim_{y \rightarrow \infty}\alpha(f;y) \quad (\in \, ]-\infty,+\infty]).
\]
Alors, on a
\[
\alpha=\lim_{x \rightarrow \infty} \frac{1}{\ln x} \sum_{n\leqslant x} \frac{f(n)}n=\liminf_{x \rightarrow \infty} \frac 1x\sum_{n\leqslant x}f(n).
\]
\end{prop}
\noindent {\bf D\'emonstration\ }
Si $f$ est multiplicativement croissante, elle est bornée inférieurement, donc la condition \eqref{t13} est vérifiée et $\alpha(f;y)$ est bien défini pour tout $y>1$. C'est une fonction croissante de $y$. En effet, soit $z\geqslant y>1$. Comme $f$ est multiplicativement croissante, on a
\begin{equation*}
f(n;z) \geqslant f(n;y) \quad (n \in {\mathbb N}^*),
\end{equation*}
d'où $\alpha(f;z) \geqslant \alpha(f;y)$, d'après la proposition \ref{t15}.
\smallskip
Posons maintenant
\begin{align*}
\alpha &=\lim_{y\rightarrow \infty}\alpha(f;y)=\sup_{y >1}\alpha(f;y)\\
\beta&=\liminf_{x\rightarrow \infty}\frac 1x\sum_{n \leqslant x} f(n)\\
\gamma &=\liminf_{x\rightarrow \infty}\frac 1{\ln x}\sum_{n \leqslant x} \frac{f(n)}n\\
\delta &=\limsup_{x\rightarrow \infty}\frac 1{\ln x}\sum_{n \leqslant x} \frac{f(n)}n\cdotp
\end{align*}
Les quatre limites $\alpha, \beta, \gamma, \delta$ appartiennent à $]-\infty,\infty]$. Nous allons voir qu'elles sont toutes égales en démontrant les inégalités
\[
\alpha \leqslant \beta \leqslant \gamma \leqslant \delta \leqslant \alpha.
\]
Parmi elles, l'inégalité $\gamma \leqslant \delta$ est évidente, et l'inégalité $\beta \leqslant \gamma$ est une relation, vraie pour toute fonction arithmétique $f$, qui se démontre par sommation partielle. Il reste à prouver les deux inégalités concernant $\alpha$.
\smallskip
\underline{Preuve de l'inégalité $\alpha \leqslant \beta$.}
Soit $y>1$. Pour tout $n \in {\mathbb N}^*$ on a $f(n) \geqslant f(n;y)$, donc
\[
\beta=\liminf_{t\rightarrow \infty}t^{-1}\sum_{n \leqslant t} f(n) \geqslant \lim_{t\rightarrow \infty}t^{-1}\sum_{n \leqslant t} f(n;y)=\alpha(f;y),
\]
d'après la proposition \ref{t15}. Le résultat s'obtient en faisant tendre $y$ vers l'infini.
\smallskip
\underline{Preuve de l'inégalité $\delta \leqslant \alpha$.}
Si $\alpha=\infty$, l'inégalité est évidente. Supposons donc $\alpha <\infty$.
Soit $y>1$. Posons
\begin{align*}
g(n) &=f(n;y) \\
h(n)&=f(n)-g(n).
\end{align*}
Comme $f$ est multiplicativement monotone, on a $h \geqslant 0$. Par conséquent, pour $x>y$,
\begin{align}
\prod_{p\leqslant x}(1-1/p)\sum_{n \leqslant x} \frac{h(n)}n &\leqslant\prod_{p\leqslant x}(1-1/p) \sum_{n \in {\mathcal S}(x)} \frac{h(n)}n\notag\\
&=\alpha(h;x)=\alpha(f;x)-\alpha(g;x)\notag\\
&=\alpha(f;x)-\alpha(f;y)\expli{d'après la proposition \ref{t18}}\notag\\
&\leqslant \alpha-\alpha(f;y).\label{t6}
\end{align}
Maintenant, quand $x$ tend vers l'infini,
\begin{align*}
\frac{1}{\ln x} \sum_{n\leqslant x} \frac{f(n)}n &= \frac{1}{\ln x} \sum_{n\leqslant x} \frac{g(n)}n + \frac{1}{\ln x} \sum_{n\leqslant x} \frac{h(n)}n \\
&\leqslant \alpha(f;y)+o(1)+\frac{1}{\ln x} \prod_{p\leqslant x}(1-1/p)^{-1}\big(\alpha-\alpha(f;y)\big)\expli{d'après \eqref{t17} et \eqref{t6}}\\
&\leqslant \alpha +o(1)+O\big(\alpha-\alpha(f;y)\big),
\end{align*}
où l'on a utilisé la majoration élémentaire $\prod_{p\leqslant x}(1-1/p)^{-1} \ll \ln x$. On a donc, pour tout $y$, la majoration $\delta \leqslant \alpha+O\big(\alpha-\alpha(f;y)\big)$, ce qui entraîne la conclusion en faisant tendre $y$ vers l'infini.~\hfill$\Box$
\smallskip
Si $f$ est multiplicativement monotone, nous poserons
\[
\alpha(f)=\lim_{x \rightarrow \infty}\frac{1}{\ln x} \sum_{n\leqslant x} \frac{f(n)}n\cdotp
\]
C'est la \emph{valeur moyenne logarithmique} de $f$. Notons que $\alpha(f) \geqslant f(1)$, si $f$ est multiplicativement croissante, et $\alpha(f) \leqslant f(1)$, si $f$ est multiplicativement décroissante.
\smallskip
L'égalité (8),~p.~22, de l'article \cite{zbMATH03065959} de Davenport et Erd\H os, est généralisée aux fonctions multiplicativement croissantes dans la proposition suivante.
\begin{prop}
Soit $(f_k)_{k \geqslant 1}$ une suite croissante de fonctions arithmétiques multiplicativement croissantes. On suppose que
\[
f(n)=\sup_{k \geqslant 1} f_k(n) < \infty \quad (n \in {\mathbb N}^*).
\]
Alors $\alpha(f_k)$ tend vers $\alpha(f)$ quand $k$ tend vers l'infini.
\end{prop}
\noindent {\bf D\'emonstration\ }
On sait que $f$ est multiplicativement croissante. Pour tout $y>1$, $\alpha(f_k;y)$ tend vers $\alpha(f;y)$, quand $k$ tend vers l'infini, par convergence monotone. Par suite,
\begin{align*}
\lim_{k \rightarrow \infty} \alpha(f_k) &=\sup_{k \geqslant 1} \alpha(f_k)=\sup_{k \geqslant 1}\,\sup_{y>1}\alpha(f_k;y)\\
&=\sup_{y>1}\,\sup_{k \geqslant 1}\alpha(f_k;y)=\sup_{y>1}\alpha(f;y)=\alpha(f).\tag*{\mbox{$\Box$}}
\end{align*}
\section{Remarques sur la moyenne de Cesàro des fonctions multiplicativement monotones}\label{t24}
En vertu de la proposition \ref{t8}, l'existence de la valeur moyenne d'une fonction arithmétique multiplicativement croissante, \text{c'est-\`a-dire } de la limite, éventuellement infinie, de ses moyennes de Cesàro, équivaut à l'inégalité
\begin{equation}\label{t25}
\limsup_{x \rightarrow \infty} \frac 1x\sum_{n\leqslant x}f(n) \leqslant \alpha(f).
\end{equation}
On sait depuis 1934 que cette inégalité n'est pas toujours vraie, puisque Besicovitch a montré dans \cite{zbMATH03014226} comment construire des ensembles de multiples sans densité asymptotique.
\medskip
Erd\H os a donné en 1948 (cf. \cite{zbMATH03049250}) un critère nécessaire et suffisant pour la validité de \eqref{t25} dans le cas d'un ensemble de multiples. Il serait intéressant de généraliser ce critère aux fonctions arithmétiques multiplicativement croissantes.
\medskip
Observons que l'existence de la valeur moyenne est évidente dans le cas où la dérivée au sens de Bougaïef est de signe constant :
\[
\frac 1x\sum_{n\leqslant x}f(n)=\frac 1x\sum_{m\leqslant x}Df(m)\lfloor x/m\rfloor \rightarrow \sum_{m\geqslant 1}\frac{Df(m)}m \quad ( x \rightarrow \infty).
\]
\section{Déterminants Toeplitz-multiplicatifs hermitiens positifs}\label{t3}
Toute matrice hermitienne définie positive infinie\footnote{J'entends par là que toutes les sections finies $(h_{ij})_{1\leqslant i, j \leqslant n}$ ($n \in {\mathbb N}^*$) sont définies positives.}
\[
H=(h_{ij})_{i, j \in {\mathbb N}^*}
\]
est la matrice de Gram
\[
G=(\scal{e_i}{e_j})_{i,j \in {\mathbb N}^*}
\]
d'une suite $(e_n)_{n \in {\mathbb N}^*}$ de vecteurs d'un espace de Hilbert ${\mathcal H}$, par exemple $\ell^2({\mathbb N}^*)$. Cela se démontre en appliquant l'algorithme de Cholesky à $H$.
Si $H$ est une matrice de Toeplitz (au sens usuel, additif), on a $h_{ij}=c_0(i-j)$, donc
\begin{equation}\label{t27}
\scal{e_i}{e_j} =\scal{e_{i+k}}{e_{j+k}}\quad (i,j \in {\mathbb N}^*, \, k \in {\mathbb N}).
\end{equation}
Si $H$ est une matrice Toeplitz-multiplicative, on a $h_{ij}=c(i/j)$, donc
\begin{equation}\label{t28}
\scal{e_i}{e_j} =\scal{e_{ki}}{e_{kj}}\quad (i,j,k \in {\mathbb N}^*).
\end{equation}
Notons
\[
D_{n}=\det(h_{ij})_{1 \leqslant i, j \leqslant n} \quad (n \in {\mathbb N}^*)
\]
avec $D_0=1$ par convention. Une formule classique de géométrie hilbertienne nous donne
\[
\frac{D_{n}}{D_{n-1}}=\frac{{\rm Gram}(e_1,\dots,e_n)}{{\rm Gram}(e_1,\dots,e_{n-1})}={\rm dist}_{{\mathcal H}}^2\big(e_n,V_{n-1}\big),
\]
où
\[
V_k={\rm Vect}(e_1, \dots, e_k) \quad (k \in {\mathbb N}),
\]
(on a donc $V_0={0}$).
Si $H$ est une matrice de Toeplitz au sens usuel, on a, pour $n \geqslant 2$,
\begin{align*}
\frac{D_{n}}{D_{n-1}} &={\rm dist}_{{\mathcal H}}^2\big(e_n,V_{n-1}\big)\\
&\leqslant {\rm dist}_{{\mathcal H}}^2\big(e_n,{\rm Vect}(e_2, \dots, e_{n-1})\big)\\
&=\frac{{\rm Gram}(e_2,\dots,e_n)}{{\rm Gram}(e_2,\dots,e_{n-1})}\\
&=\frac{{\rm Gram}(e_1,\dots,e_{n-1})}{{\rm Gram}(e_1,\dots,e_{n-2})} \expli{d'après \eqref{t27}}\\
&=\frac{D_{n-1}}{D_{n-2}} \cdotp
\end{align*}
C'est le résultat de Fekete mentionné au \S\ref{t22}. La convergence de la suite $D_{n}/D_{n-1}$, qui en résulte, entraîne celle de la suite~$D_n^{1/n}$, vers la même limite. Si les $c_k$ sont les coefficients de Fourier d'une fonction $f$, de période $1$, continue et positive, Szeg\H o démontra dans le même article \cite{zbMATH02616579} une conjecture de P\'olya, affirmant que cette limite est la moyenne géométrique des valeurs de $f$ :
\[
D_n^{1/n} \rightarrow \exp\Big(\int_0^1\ln f(t) \, dt\Big) \quad (n \rightarrow \infty).
\]
Ce résultat fut le premier d'une nouvelle théorie, dont l'ouvrage récent \cite{nikolski2017} est un exposé actualisé.
\smallskip
Le raisonnement ci-dessus s'adapte de la façon suivante au cas où $H$ est une matrice Toeplitz-multiplicative (hermitienne, définie positive). Si $n=km$, on a
\begin{align*}
\frac{D_{n}}{D_{n-1}} &={\rm dist}_{{\mathcal H}}^2\big(e_{km},V_{km-1}\big)\\
&\leqslant {\rm dist}_{{\mathcal H}}^2\big(e_{km},{\rm Vect}(e_m,e_{2m}, \dots, e_{(k-1)m})\big)\\
&=\frac{{\rm Gram}(e_m,e_{2m},\dots,e_{(k-1)m},e_{km})}{{\rm Gram}(e_m,e_{2m},\dots,e_{(k-1)m})}\\
&=\frac{{\rm Gram}(e_1,\dots,e_{k})}{{\rm Gram}(e_1,\dots,e_{k-1})} \expli{d'après \eqref{t28}}\\
&=\frac{D_{k}}{D_{k-1}} \cdotp
\end{align*}
La suite des quotients $D_n/D_{n-1}$ est donc multiplicativement décroissante. La suite de leurs logarithmes l'étant également, on obtient la conclusion suivante par application de la proposition~\ref{t8} à la fonction
\[
f(n) =-\ln (D_n/D_{n-1}) \quad (n \in {\mathbb N}^*).
\]
\begin{prop}\label{t29}
Avec les hypothèses et les notations de la question du \S\ref{t22}, la moyenne logarithmique
\[
\frac{1}{\ln N}\sum_{k\leqslant N} \frac 1k\ln(D_k/D_{k-1})
\]
converge vers un élément $\alpha$ de $[-\infty, \ln c(1)]$ quand $N$ tend vers l'infini. De plus,
\[
\limsup_{N \rightarrow \infty} D_N^{1/N}=e^{\alpha}.
\]
\end{prop}
Si $\alpha = -\infty$, on a donc une réponse positive à la question du \S\ref{t22} : la limite cherchée est nulle. Si~$\alpha > -\infty$, on peut récrire l'assertion sur la moyenne logarithmique sous la forme
\[
\prod_{k=1}^N (D_k/D_{k-1})^{1/k}=N^{\alpha+o(1)} \quad ( N \rightarrow \infty).
\]
\medskip
Cela étant, en toute généralité, je ne sais pas répondre à la question initiale. La construction de Besicovitch peut-elle être adaptée au cas des déterminants Toeplitz-multiplicatifs, et fournir ainsi un exemple où $D_n^{1/n}$ n'a pas de limite?
\medskip
Les propositions suivantes, \ref{t30} et \ref{t36}, indiquent des cas simples où cette limite existe.
\begin{prop}\label{t30}
Soit $A$ un facteur direct, stable par multiplication, et dont la somme des inverses converge. On pose
\begin{equation*}
R =\{a/a', \, a\in A, \, a' \in A\}
\end{equation*}
Soit $c : \, {\mathbb Q}^{+*} \rightarrow {\mathbb C}$ une fonction telle que
\begin{align*}
c(1/r) &=\overline{c(r)} \quad (r \in R)\\
c(r) &=0 \quad (r\notin R),
\end{align*}
et telle que, pour tout $n \in {\mathbb N}^*$, le déterminant
\[
D_n =\det\big(c(i/j)\big)_{1\leqslant i,j \leqslant n}
\]
soit strictement positif.
Alors la suite $D_n^{1/n}$ est convergente.
\end{prop}
\noindent {\bf D\'emonstration\ }
Soit $B$ le facteur direct complémentaire de $A$. En interprétant $D_n$ comme déterminant de Gram, et en utilisant le fait que chaque nombre entier~$j\in~{\mathbb N}^*$ s'écrit de façon unique $j=ab$, avec~$a \in A$,~$b\in B$, on a
\[
D_n={\rm Gram}(e_{ab} \, ; \, a \in A, \, b\in B, \, ab\leqslant n),
\]
où nous utilisons aussi le fait qu'un déterminant de Gram est indépendant de l'ordre des vecteurs qui le composent.
Si $b,b' \in B$, et $a,a',\alpha,\alpha' \in A$, on a
\[
\frac{ab}{a'b'}=\frac{\alpha}{\alpha'}\Rightarrow \alpha' a\cdot b= \alpha a' \cdot b' \Rightarrow b=b',
\]
puisque $A$ et $B$ sont des facteurs directs complémentaires et puisque $A$ est stable par multiplication. Comme $c$ est à support dans $R$, cela montre que
\[
\scal{e_i}{e_j}=0 \quad (i=ab\, ; \, j=a'b' \, ; \, a,a' \in A \, ; \, b,b' \in B \, ; \, b\neq b').
\]
Autrement dit, les vecteurs $e_i$ et $e_j$ sont orthogonaux si les décompositions de $i$ et $j$ en produits d'un élément de $A$ par un élément de $B$ ne font pas intervenir le même élément de $B$. Cela permet de décomposer $D_n$ en produit :
\begin{align*}
D_n&=\prod_{ b \in B, b\leqslant n} {\rm Gram}(e_{ab}\, ; \, a \in A, \, a\leqslant n/b)\\
&=\prod_{ b \in B, b\leqslant n} {\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant n/b),
\end{align*}
d'après \eqref{t28}.
En écrivant maintenant $n=\alpha\beta$, avec $\alpha\in A$, $\beta \in B$, on a
\begin{equation*}
{\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant n/b)={\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant (n-1)/b) \quad ( b\neq \beta).
\end{equation*}
Par conséquent
\begin{align}
\frac{D_n}{D_{n-1}} &=\frac{{\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant n/\beta)}{{\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant (n-1)/\beta)}\notag\\
&=\frac{{\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant \alpha)}{{\rm Gram}(e_{a}\, ; \, a \in A, \, a\leqslant \alpha -1)}\label{t37}\\
&=\frac{D_{\alpha}}{D_{\alpha-1}}\raisebox{.7mm}{,}\notag
\end{align}
puisque \eqref{t37} ne dépend pas de $\beta$.
On peut donc appliquer la proposition \ref{t31} à la fonction arithmétique
\[
f(n) =-\ln (D_n/D_{n-1}) \quad (n \in {\mathbb N}^*),
\]
puisqu'elle vérifie $f(n)=f(n;A)$ et $f(n) \geqslant -\ln c(1)$ pour tout $n$. Le résultat s'en déduit.\hfill$\Box$
\medskip
\begin{prop}\label{t36}
Soit $\sigma : {\mathbb N}^* \rightarrow {\mathbb C}$ une fonction multiplicative, et $c: {\mathbb Q}^{+*} \rightarrow {\mathbb C}$ la fonction définie par
\[
c(i/j)=\sigma\big(i/(i,j)\big)\cdot\overline{\sigma\big(j/(i,j)\big)} \quad (i,j \geqslant 1),
\]
On suppose que, pour tout $n \in {\mathbb N}^*$, le déterminant
\[
D_n =\det\big(c(i/j)\big)_{1\leqslant i,j \leqslant n}
\]
est strictement positif.
Alors la suite $D_n^{1/n}$ est convergente.
\end{prop}
\noindent {\bf D\'emonstration\ }
Au point $(ii)$ du Corollary 3.4, p. 274 de l'article \cite{zbMATH05082236}, Hilberdink donne une expression de~$D_n$ :
\[
D_n=\prod_p\prod_{k\geqslant 1} (\rho_{k+1}(p)/\rho_k(p))^{\lfloor n/p^k\rfloor},
\]
où, pour tout nombre premier $p$,
\[
\rho_{k}(p)=\frac{\Delta_k(p)}{\Delta_{k-1}(p)} \quad (k \geqslant 1),
\]
où $\big(\Delta_k(p)\big)_{k\geqslant 1}$ est une suite de déterminants de Toeplitz usuels (additifs), hermitiens et positifs.
Le théorème de Fekete montre que les quotients $\rho_{k+1}(p)/\rho_k(p)$ sont tous $\leqslant 1$. Par conséquent,
\[
\frac 1n \ln D_n=\sum_p\sum_{k\geqslant 1} \ln \big(\rho_{k+1}(p)/\rho_k(p)\big)\frac{\lfloor n/p^k\rfloor}n \rightarrow \sum_p\sum_{k\geqslant 1} \frac{\ln \big(\rho_{k+1}(p)/\rho_k(p)\big)}{p^k}\quad (n \rightarrow \infty)
\]
Dans ce cas, la dérivée de $\ln(D_n/D_{n-1})$ au sens de Bougaïef est $\leqslant 0$. \hfill$\Box$
\smallskip
Terminons par deux remarques sur cette proposition \ref{t36}. D'abord, le Theorem 3.2, p. 274 de \cite{zbMATH05082236} montre que l'hypothèse de stricte positivité des déterminants $D_n$ équivaut à la stricte positivité des déterminants $\Delta_k(p)$ (notés $\det T_k(\widehat{f_p})$ par Hilberdink). Ensuite, la proposition \ref{t36} s'applique en particulier au cas, également considéré par Hilberdink, où la fonction $\sigma$ est complètement multiplicative. Dans ce cas, les matrices $M_n$ sont définies positives si, et seulement si $\lvert \sigma(p)\rvert <1$ pour tout nombre premier $p$. Le Corollary 5.1, p. 283 de \cite{zbMATH05082236}, montre alors que
\[
D_n^{1/n} \rightarrow \prod_p\big(1-\lvert \sigma(p)\rvert ^2\big)^{1/p} \quad (n \rightarrow \infty).
\]
\bigskip
\begin{center}
{\sc Remerciements}
\end{center}
\begin{quote}
{\footnotesize
La recherche ayant conduit à ces résultats a été subventionnée par le Conseil Européen de la Recherche (ERC), dans le Septième Programme-Cadre de l'Union Européenne (FP7/2007-2013), accord de subvention ERC n$^{\text{o}}$ 670239.
L'article a été, en partie, élaboré au printemps 2017, lors d'un séjour au département de mathématiques de l'University College de Londres. Je remercie Andrew Granville et cette institution pour les conditions de travail idéales dont j'ai alors bénéficié.}
\end{quote}
\providecommand{\bysame}{\leavevmode ---\ }
\providecommand{\og}{``}
\providecommand{\fg}{''}
\providecommand{\smfandname}{et}
\providecommand{\smfedsname}{\'eds.}
\providecommand{\smfedname}{\'ed.}
\providecommand{\smfmastersthesisname}{M\'emoire}
\providecommand{\smfphdthesisname}{Th\`ese}
|
\section{Introduction}
The occurrence of nanospinodal decomposition is at the origin of the exceptional magnetic signatures in relevant semiconductors, like $e.g.$ Ge, GaAs, GaN and ZnO doped with magnetic elements and in particular with transition metal (TM) ions\,\cite{Bonanni:2010_CSR,Sato:2010_RMP,Dietl:2015_RMP}. In these material systems the aggregation of the TM ions
takes place either by preserving the crystallographic structure of the host lattice (chemical phase separation) or generating TM-rich nanocrystals (crystallographic phase separation) embedded in a TM-poor matrix. Crystallographic phase separation in particular, allows to combine the properties of the semiconducting host with those of embedded magnetic nanostrcutures, opening wide perspectives for spin detection and injection, spin valve effects, and flash-memory elements\,\cite{Coey:1999_JMMM}.
While the structural and magnetic properties of these phase-separated materials have been widely studied\,\cite{Dietl:2015_RMP}, unveiling comparable characteristics among the different systems, the understanding of the underlying (magneto)transport mechanisms is in its infancy. The transport characteristics and the magnetoresistance (MR) of semiconducting paramagnetic (PM) materials containing ferromagnetic nanocrystals (NCs)\,\cite{Akiyama:2012_APL} are dominated by spin and magnetic field dependent localization phenomena at the nanocrystal-matrix interface and significantly depend on the properties of the paramagnetic host, on the magnetic response of the embedded nanostructures and on the interaction of the electronic states of the host with the magnetic nanostructures.
In (Ga,Mn)As phase-separated systems, a large negative MR associated with ferromagnetic MnAs nanoparticles (NPs) buried in the GaAs matrix was observed at temperatures below 30\,K\,\cite{Akinaga:1998_APL}, and a positive MR at intermediate fields up to 100\,K was reported\,\cite{Michel:2008_APL}. The large negative MR was explained in the frame of spin-dependent scattering of the carriers at the MnAs NPs, that decreases when the direction of the NPs magnetization aligns with the field. In (Ge,Mn) systems containing Ge$_3$Mn$_5$ ferromagnetic clusters and nanocolumns, it was found that the doping of the surrounding matrix plays a crucial role on the overall magnetotransport properties. In particular, a positive MR with a linear field dependence for layers grown on $p$-type GaAs was detected, while a negative MR with hysteretic behavior at 5\,K for layers grown on As-rich surfaces was reported\,\cite{Yu:2010_PRB}. The observed MR in this case was interpreted in terms of tunneling magnetoresistance effects. In ZnO containing Co embedded nanocrystals and at magnetic fields below 1\,T, a similar hysteretic behavior of the MR was also detected together with large positive MR values of 20\% at 10\,K\,\cite{Hamieh:2015_PRB}. In this material system, the transport properties are determined by the ZnO matrix containing magnetic localized Co$^{2+}$ impurities. Moreover, the hysteretic behavior indicates a small contribution to the conduction process at low fields, which is attributed to the spin polarization of the Co NCs.
In the case of GaN -- whose relevance for opto-electronics, high-power electronics and spintronics has steadily increased over the recent years -- crystallographic phase separation is observed when TM ions are incorporated above the solubility limit\,\cite{Giroud:2004_EPL,Kunert:2012_APL,Bonanni:2008_PRL,Bonanni:2007_PRB}. In particular, phase separation with ferromagnetic signatures up to 540\,K was reported in GaN doped with a concentration of Fe above 0.4\%\,\cite{Bonanni:2007_PRB} cations and grown by metalorganic vapour phase epitaxy (MOVPE). While the inhomogeneous distribution of a variety of ferromagnetic (FM) and antiferromagnetic (AF) Fe$_x$N phases $(x = 2,3,4)$ in this particular material system has to-date limited its applicability into functional devices\,\cite{Navarro:2010_PRB,Navarro:2011_PRB}, the recently reported spatial localization of arrays of single-phase face-centered cubic $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs with in-plane uniaxial magnetic anisotropy has opened new perspectives for these material system\,\cite{Navarro:2012_APL,Grois:2014_Nanotech}.
Furthermore, by varying the fabrication conditions, the lattice parameter of the embedded $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs can be tuned on-demand from the one of $\gamma$'-Fe$_4$N, towards the one of $\gamma$'-GaFe$_3$N by controlling the incorporation of Ga into the $\gamma$'-Fe$_4$N lattice. Since the magnetic response of $\gamma$'-Ga$_y$Fe$_{4-y}$N ranges from strongly ferromagnetic ($y$$<$0.25) to weakly antiferromagnetic ($y$$>$0.25)\,\cite{Houben:2009_ChemMat, Burghaus:2011_JSSC}, planar arrays of $\gamma$'-Ga$_y$Fe$_{4-y}$N NC arrays embedded in GaN become suitable for FM as well as for the emerging field of AF-spintronics\,\cite{Jungwirth:2016_NatNano, Wadley:2016_Science}.
Here, the magnetotransport mechanisms in MOVPE phase-separated (Ga,Fe)N thin layers (Ga$\delta$FeN) grown on GaN buffers and containing $\gamma$'-Ga$_y$Fe$_{4-y}$N nanocrystals are investigated in the temperature range between 2\,K and 300\,K. While the main conduction channel is confirmed to be the GaN buffer layer, the magnetoresistance of the system is significantly affected by the presence of the NCs, as evidenced by anisotropic magnetoresistance (AMR) observed at all temperatures. In consideration of the relevance of capping layers and encapsulation for applications, the behavior of GaN-capped phase-separated Ga$\delta$FeN is also analysed.
\section{Experimental details}
The investigated samples are fabricated by MOVPE according to the procedure previously reported\,\cite{Navarro:2012_APL} and consist of a 1200\,nm thick GaN buffer layer -- unintentionally $n$-type with a carrier concentration $n$\,=\,(8$\times$10$^{16}$)\,cm$^{-3}$ at room-temperature (RT) -- deposited at 1040$^{\circ}$C on 2"\,$c$-[0001] sapphire substrates, followed by the growth of a 50\,nm thin Ga$\delta$FeN layer containing nanocrystals deposited at 780$^{\circ}$C, eventually capped with GaN. Specifically, the following structures are considered: (i) Sample A -- uncapped; (ii) Sample A* -- $i.e.$ Sample A upon annealing at 600$^\circ$C under N$_2$ carried out to remove the $\alpha$-Fe inclusions formed in proximity of the sample surface when the samples are left uncapped; (iii) Sample B -- with the same basic structure as Samples A and A*, but additionally capped with a nominally 70\,nm thick GaN layer grown at 1000$^{\circ}$C. A schematic representation of the samples architecture is reproduced in Figs.\,\ref{fig:struct1}(a) and (b). A bare 1200\,nm thick GaN buffer layer on sapphire is employed as reference. The relevant characteristics of the samples are provided in Table\,I.
Information on the samples structure and on the NCs crystallographic phase is obtained from high-resolution x-ray diffraction (HRXRD) rocking curves. The measurements are carried out using a PANalytical X’Pert Pro Material Research Diffractometer equipped with a hybrid monochromator with a 0.25$^{\circ}$ divergence slit and a PixCel detector using 19 channels for detection and a 5\,mm anti-scatter slit. In order to determine the distribution of the NCs in the GaN matrix, as well as to verify the crystallographic phase of the NCs, (high-resolution) transmission electron microscopy (HRTEM) measurements are performed. The cross-section TEM specimens are prepared by mechanical polishing, dimpling and ion milling in a Gatan Precision Ion Polishing System. Dark-field and bright-field measurements in conventional mode, and high-angular annular dark-field (HAADF) measurements in STEM mode are performed using a JEOL JEM\,2000\,EX system.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig1.eps}
\caption{(a) Schematic representation of Sample A -- uncapped; (b) schematic representation of Sample B -- capped with a nominally 70\,nm thick GaN layer. (c) and (d): Cross-section TEM images of Samples A and B, respectively, evidencing the spatial distribution of the NCs in a plane perpendicular to the growth direction [0001].}
\label{fig:struct1}
\end{figure}
The magnetic characteristics of the samples are measured in a Quantum Design superconducting quantum interference device (SQUID) magnetometer MPMS XL. The samples are investigated in the temperature range between 2\,K and 300\,K in magnetic fields up to 5\,T. The diamagnetic component originating from the sapphire substrate is carefully compensated according to the procedure described elsewhere\,\cite{Sawicki:2011_SST,Gas:2018_arxiv}.
Additionally, the angular dependence of the magnetization is studied by performing ferromagnetic resonance (FMR) measurements at RT in out-of-plane and in-plane configuration with a Bruker Elexsys E580 electron paramagnetic resonance spectrometer at microwave frequencies between 9.4\,GHz and 9.5\,GHz. During measurements, the static magnetic field is modulated with an amplitude of 0.5\,mT at 100\,kHz to allow lock-in detection\,\cite{Grois:2014_Nanotech}.
The magnetotransport measurements are carried out in Van der Pauw geometry using a high-sensitivity Janis Super Variable Temperature 7TM-SVM magnetotransport system in the temperature range between 2\,K and 300\,K and in an external magnetic field $H$ tunable between -6\,T and 6\,T, applied perpendicular to the film plane. Details on the fabrication of the contacts are provided in the Supplemental Material\,\cite{supplementary}. For AMR measurements, the direction of $H$ has been varied with respect to the surface sample normal from perpendicular (out-of-plane angle $\beta = 90^\circ$) to parallel ($\beta = 0^\circ$).
\section{Results and Discussion}
\subsection{Structural and magnetic properties}{\label{3a}}
\begin{figure}
\includegraphics[width=\columnwidth]{Fig2.eps}
\caption{(a) HRXRD of the samples with the identified NCs diffraction peaks. Inset to (a): HRTEM of $\alpha$-Fe and $\gamma$'-Ga$_y$Fe$_{4-y}$N in Sample A. (b) HAADF cross-section TEM image of Sample A*. (c) HRTEM of the polycrystalline FeO$_x$ formed at the surface of the area occupied by an $\alpha$-Fe NC prior to annealing of Sample A. Inset to (c): SAD pattern of GaN in proximity of the area previously occupied by $\alpha$-Fe, confirming the outward diffusion of Fe.}
\label{fig:xrd1}
\end{figure}
The structure of the nanocrystals present in the Ga$\delta$FeN layers is assessed by HRXRD: besides the 0002 diffraction peak of the GaN buffer and the 0006 peak of the Al$_2$O$_3$ substrate, two additional reflections are observed in the spectra reported in Fig.\,\ref{fig:xrd1}(a) for Sample A at 44.275$^{\circ}$$\pm$0.005$^{\circ}$ and 47.821$^{\circ}$$\pm$0.005$^{\circ}$. In Sample B the peak at around 47.8$^{\circ}$ solely is detected. From the distance between the diffraction planes ($d$-spacing), the reflections are identified as the (110) of $\alpha$-Fe and the (200) of $\gamma$'-Ga$_y$Fe$_{4-y}$N, respectively\,\cite{Rovezzi:2012_arxiv}. These two crystalline phases were already reported for phase-separated (Ga,Fe)N\,\cite{Navarro:2010_PRB, Li:2008_JCG}. From the full-width-at-half-maximum (FWHM) of the diffraction peaks and by employing the Scherrer formula\,\cite{Patterson:1939_PR}, the size of the nanocrystals along the growth direction is estimated to be (17\,$\pm$2)\,nm for the $\alpha$-Fe and (22\,$\pm$2)\,nm for the $\gamma$'-Ga$_y$Fe$_{4-y}$N nanocrystals.
The presence of the two crystallographic phases in Sample A is confirmed by the cross-section HRTEM image reported in the inset to Fig.\,\ref{fig:xrd1}(a): $\gamma$'-Ga$_y$Fe$_{4-y}$N nanocrystals buried 30\,nm below the sample surface and -- as identified by selective area diffraction pattern (SAD) analysis -- $\alpha$-Fe NCs located at the surface. It is estimated that this phase accounts for $\sim$50\% of the total NCs in the sample. The tendency of the $\alpha$-Fe to form at the sample surface was previously reported and attributed to the evaporation of nitrogen taking place during the MOVPE process\,\cite{Li:2008_JCG}. According to the cross-section image in Fig.\,\ref{fig:struct1}(d), in Sample B only embedded $\gamma$'-Ga$_y$Fe$_{4-y}$N are found located 70\,nm to 100\,nm below the sample surface, in agreement with the nominal thickness of the GaN capping layer. The NCs size obtained from TEM images ranges between 10\,nm and 30\,nm, consistent with the average size obtained from the FWHM analysis of the HRXRD spectra. Energy dispersive x-ray (EDX) scattering performed during the TEM imaging, points at a minimal concentration $x$$<$0.1\% of dilute Fe in the host matrix, suggesting that most of the Fe provided during growth is incorporated into the nanocrystals. The dilute Fe ions in the GaN matrix are known to occupy preferentially substitutional Ga-sites and be in the Fe$^{3+}$ charge state\,\cite{Bonanni:2007_PRB}.
The Fe-rich NCs embedded in previously considered phase-separated (Ga,Fe)N layers were found to be stable up to temperatures as high as 900$^\circ$C\,\cite{Navarro:2010_PRB}. This behavior is confirmed for Sample B, in which the NCs are buried below the GaN capping layer, but in the case of Sample A, post-growth annealing carried out at 600$^\circ$C under N$_2$ atmosphere -- and resulting in Sample A* -- significantly affects the system, as detailed in Fig.\,S1(a) of the Supplemental Material\,\cite{supplementary}. The XRD analysis of Sample A* ($i.e.$ Sample A annealed) points at a removal of $\alpha$-Fe NCs upon annealing, as evidenced in Fig.\,\ref{fig:xrd1}(a). This effect is corroborated by the detailed TEM analysis of Sample A* summarized in Figs.\,\ref{fig:xrd1}(b) and (c). In the HAADF cross-section image of Fig.\,\ref{fig:xrd1}(b) a $\gamma$'-Ga$_y$Fe$_{4-y}$N NC as well as a void resulting from the removal of $\alpha$-Fe and located at the sample surface are pictured. In the HRTEM image in Fig.\,\ref{fig:xrd1}(c), the region adjacent to the void is considered in detail and a SAD analysis points to the structure of GaN. Moreover -- again upon annealing -- the presence of polycrystalline FeO$_x$ at the surface is detected, suggesting out-diffusion of the Fe originally bound in $\alpha$-Fe NCs. The polycrystalline FeO$_x$ is completely removed through chemical treatment of 1 hour in HCl, as confirmed by x-ray photoemission spectroscopy (XPS) and as evidenced in Fig.\,S1(b) of the Supplemental Material \,\cite{supplementary}.
The magnetic response of the Ga$\delta$FeN layers investigated here is qualitatively similar to the one of phase-separated (Ga,Fe)N layers previously reported\,\cite{Bonanni:2007_PRB}.
The field dependency of the in-plane aerial density of magnetic moment $m/m_\mathrm{s}$ is characterized primarily by a strongly non-linear superparamagnetic-like response detected over the whole range of the studied temperatures and by a PM contribution from dilute substitutional Fe$^{3+}$ ions in the matrix dominating at low temperatures. These two magnetic components can be treated separately and quantitatively, as previously reported\,\cite{Pacuski:2008_PRL,Navarro:2010_PRB}.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig3.eps}
\caption{Normalized aerial density of magnetic moment of Samples A (stars), A* (dots) and B (squares): (a) at 2\,K and (b) at 300\,K. Inset to (b): Coercive field strength ($H_c$) \textit{vs.} temperature.}
\label{fig:magnetic}
\end{figure}
In Figs.\,\ref{fig:magnetic}(a) and (b), the normalized aerial density of magnetic moment as a function of the applied magnetic field is compared for the considered samples containing NCs, both at 2\,K and at 300\,K, respectively. In Sample A the magnetization saturates swiftly at 1\,T at room-temperature, as expected due to the presence of the soft ferromagnetic $\alpha$-Fe, known to saturate at fields as low as 10\,mT at RT\,\cite{Nie:2003_TSF}. Upon annealing -- $i.e.$ Sample A* -- the response quenches and its saturation shifts from 2\,T to 5\,T at 2\,K, as evidenced in Fig.\,\ref{fig:magnetic}(a) and from 1\,T to 2\,T at RT, as reported in Fig.\,\ref{fig:magnetic}(b).
All samples show an open hysteresis below 100\,K with coercive field values between 10\,mT and 60\,mT. The coercive field for $\alpha$-Fe is reported to be as low as 0.2\,mT at RT\,\cite{Nie:2003_TSF}, while the one of $\gamma$'-Fe$_4$N nanoparticles is in the range between 50\,mT and 80\,mT\,\cite{Wu:2004_JMMM}. These last values fit well to the ones observed in the considered samples and evidenced in the inset to Fig.\,\ref{fig:magnetic}(b), pointing at the presence of FM $\gamma$'-Fe$_4$N nanocrystals.
Taking into consideration the structural information previously discussed, the difference in magnetization between Samples A and A* is calculated. The difference signal is nearly temperature independent -- as detailed in Fig.\,S2 of the Supplemental Material\,\cite{supplementary} -- rapidly saturates with increasing $H$ and its saturation as a function of temperature is comparable to the one observed for Fe nanoparticles\,\cite{Zhang:1998_PRB}, hinting at ferromagnetic $\alpha$-Fe NCs as main responsible for the difference between the two responses. Remarkably, the amount of Fe$^{3+}$ ions contributing to the overall magnetic signal is the same for both samples, indicating that dilute Fe in the host matrix is not affected by the annealing procedure, in accordance with previous findings\,\cite{Navarro:2010_PRB}.
On the other hand, the normalized magnetization as a function of the applied field is comparable for Samples B and A* at 2\,K, while significantly differing at RT: both specimens are found to contain only $\gamma$'-Ga$_y$Fe$_{4-y}$N (and not $\alpha$-Fe) NCs, but with density, as well as composition ($y$), differing in the two samples.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig4.eps}
\caption{ Angular dependence of the FMR signal for Samples A and A*: (a) the out-of-plane FMR follows a cos$^2\theta$ dependence, while (b) the in-plane signal follows a sin(6$\phi$) behavior.}
\label{fig:FMR}
\end{figure}
The out-of-plane and in-plane angular dependence of the FMR signals provided in Fig.\,\ref{fig:FMR} confirm that the magnetic anisotropy of Sample A is equivalent to the one observed for capped samples\,\cite{Grois:2014_Nanotech}. An uniaxial $cos^2\theta$ dependence of the resonance field is observed when varying the out-of-plane angle ($\theta$), as evidenced in Fig.\,\ref{fig:FMR}(a), while the variation of the in-plane angle (azimuth $\phi$) produces a $sin(6\phi)$ dependence reported in Fig.\,\ref{fig:FMR}(b). The uniaxial out-of plane anisotropy is attributed to shape anisotropy, while the in-plane anisotropy is related to the hexagonal crystal structure of the GaN matrix. The FMR results point at three easy axes lying in the plane of the NCs.
\begin{table*}[ht]
\centering
\begin{tabular}{c| c c c c c c c}
\hline \hline
Sample & Structure & Annealing & $R_\mathrm{RT}$ & $n$ & $\mu$ & $E_\mathrm{a}$ & $T_{0}^{\mathrm{Mott}}$ \\
& & & $(\mathrm{k}\Omega$) & $(10^{16}$ cm$^{-3}$) & (cm$^2$/Vs) & (meV) & (K) \\
\hline
GaN & 1200$\mu$m GaN layer & none & 3 & 8 & 350 & 20 & $7\times10^{4}$ \\
A & uncapped Ga$\delta$FeN/GaN & none & 5 & 6 & 600 & 17 & $1\times10^{7}$ \\
A* & uncapped Ga$\delta$FeN/GaN & 600$^\circ$C/N$_{2}$/10min & 6 & 8 & 700 & 19 & $6\times10^{6}$ \\
B & capped Ga$\delta$FeN/GaN & none & 43 & 2 & 100 & 25 & $6\times10^{12}$ \\
\hline \hline
\end{tabular}
\label{tab:tab1}
\caption{Structure and characteristic transport parameters of the investigated samples. The mobility $\mu$, the carrier density $n$ and the resistance $R$ values are acquired at RT. With $R$ the average resistance across the four ohmic contacts in van der Pauw geometry). The activation energy $E_\mathrm{a}$ and the $T_{0}^{\mathrm{Mott}}$ are obtained from fitting $\rho(T)$.}
\end{table*}
\subsection{Resistivity and magnetoresistance}
For completeness and considering that the Ga$\delta$FeN are grown on GaN buffer layers with unintentional $n$-conductivity, the bare reference GaN buffer is also studied. The resistivity $\rho$ of the samples is measured as a function of temperature and reported in Fig.\,\ref{fig:rho}(a). The characteristic transport parameters obtained at RT for all samples are listed in Table I.
The resistivity of all samples increases with decreasing temperature, as expected for semiconducting thin films. While the $\rho$s of Sample A, of Sample A* and of the GaN reference are comparable down to 50\,K, the one of Sample B is enhanced by approximately one order of magnitude at temperatures down to 100\,K and is about three orders of magnitude larger below this temperature. In the inset to Fig. 5(a) the $\rho(T)$ is depicted in full scale, evidencing the large difference in resistivity between Sample B and the other measured samples. This effect is attributed to the diffusion of dilute Fe ions -- with a concentration $x$$\sim0.1$\% -- from the Ga$\delta$FeN region into the GaN capping layer of Sample B, which compensates the unintentional $n$-type conductivity of the GaN capping layer\,\cite{Heikman:2002_APL} and increases the measured resistivity.
\begin{figure}
\includegraphics[width=0.9\columnwidth]{Fig5.eps}
\caption{(a) Resistivity \textit{vs.} temperature for Samples A, A* and B compared to GaN. (b) Resistivity \textit{vs.} $T^{0.25}$ pointing at a characteristic Mott VRH type conduction below 40\,K for Sample A, for Sample A* and for the GaN reference, and for Sample B below 130\,K. Above these temperatures, an Arrhenius-like behavior is observed.}
\label{fig:rho}
\end{figure}
The temperature dependence of $\rho (T)$ points at two distinct regimes. Specifically, an Arrhenius-type conduction is observed for Sample A, for Sample A* and for the GaN reference down to 50\,K, and for Sample B down to 130\,K. The activation energies $E_\mathrm{a}$ for all samples are comparable, suggesting that the transport in this conduction regime takes place mainly in the underlying GaN buffer. The obtained $E_\mathrm{a}$ values fit well with those previously reported for GaN layers\,\cite{Goetz:1996_APL,Yildiz:2010_APA}. Furthermore, the behavior of $\rho(T)$ points at a Mott variable range hopping (VRH) conduction mechanism, as evidenced in Fig.\,\ref{fig:rho}(b), where the linear dependency of $\rho(T)$ on $T^{-0.25}$ is reported. The transition from Arrhenius to Mott VRH conduction occurs between 50\,K and 25\,K in Sample A, Sample A* and GaN, while it shifts to between 130\,K and 100\,K for Sample B.
The VRH electronic transport is characterized by conduction in an impurity band of localized states with random spatial and energy distribution\,\cite{Pollak:2002_pssb} and a constant density of states at the Fermi energy $N(E_\mathrm{F})$. In the Mott VRH model the resistivity as a function of temperature follows\,\cite{Mott:1969}:
\begin{equation}
\rho (T) \sim exp(T_{0}^{\mathrm{Mott}}/T)^{1/4},
\end{equation}
\label{eq:one}
where $T_{0}^{\mathrm{Mott}}$ is the characteristic hopping temperature, whose value depends on the electronic density of states at the Fermi level and on the localization length $\xi$.
The linear fits in the Mott regime yield the $T_{0}^{\mathrm{Mott}}$ values listed in Table I, that are in good agreement with those previously reported for Gd-doped GaN\,\cite{Pinto:2009_PRB} and for GaN\,\cite{Yildiz:2010_APA}. The value obtained for $T_{0}^{\mathrm{Mott}}$ in the GaN reference is two orders of magnitude lower than those for Samples A and A*, suggesting a larger ratio between the hopping distance $R_\mathrm{h}$ and the localization length $\xi$ in the uncapped Ga$\delta$FeN layers, likely to be related to the presence of $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs and of dilute Fe ions in these layers. Sample A and Sample A* show a comparable behavior and correspondent fitting parameters, indicating that the $\alpha$-Fe NCs present in Sample A give a negligible contribution to the transport. One way to determine the values of the localization length and of the hopping energy from the obtained $T_{0}^{\mathrm{Mott}}$, goes through the knowledge of the density of states at the Fermi level. Alternatively, these values can be obtained from magnetoresistance measurements, as shown below.
\begin{table}[ht]
\centering
\begin{tabular}{c| c c c c}
\hline \hline
& GaN & & Sample A* & \\
\hline
$T$(K) & $a\pm 0.05$ & $b\pm 0.05$ & $a\pm 0.05$ & $b\pm 0.05$ \\
\hline
50 & $0.42$ & 0.00 & $0.36$ & $0.04$ \\
100 & $0.63$ & 0.00 & $0.58$ & $0.06$ \\
150 & $0.65$ & 0.00 & $0.61$ & $0.07$ \\
200 & - & - & $0.47$ & $0.07$ \\
250 & $0.37$ & 0.00 & $0.38$ & $0.05$ \\
\hline \hline
\end{tabular}
\label{table2}
\caption{Two-band model\,\cite{Zaremba:1992_PRB} fit parameters for the MR above 50\,K in the GaN reference and in Sample A*.}
\end{table}
In order to gain insight into the contribution of the NCs to the (magneto)transport in the studied structures, the resistance of the layers is measured as a function of the applied magnetic field. The MR is defined as:
\begin{equation}
\frac{\Delta\rho}{\rho_0} = \frac{\rho(H,T)-\rho_{0}(T)}{\rho_{0}(T)},
\end{equation}
where $\rho_0$ and $\rho(H)$ are the resistivity at $H$=0 and at a field $H$$\neq$0, respectively. In Figs.\,\ref{fig:MR}(a) and (b) the MR of Sample A* as a function of the applied magnetic field is reported: as evidenced in Fig.\,\ref{fig:MR}(a), at temperatures $\geq$50\,K the MR is positive and follows a $H^2$ dependence similar to the one of the GaN reference\, presented in Fig.\,S3 of the Supplemental Material\,\cite{supplementary}. The positive MR can be explained in the frame of a two-band conduction model, in which one conduction band is the one of GaN and the second one is an impurity band induced by the diluted Fe ions in the host matrix. In this model, the MR is given by:
\begin{equation}
\mathrm{MR}=a^2(\mu_0H)^2/[1+b^2(\mu_0H)^2],
\end{equation}
where $\mu_0$ is the vacuum permeability, while $a$ and $b$ are parameters directly related to the conductivities and mobilities of each band\,\cite{Zaremba:1992_PRB}. The values of $a$ and $b$ obtained by fitting with the mentioned model the MR of Sample A* for temperatures $\geq$50\,K are collected in Table II. While for GaN the MR can be fitted considering one single band, two bands are required for Sample A*. The values of the fit parameter $a$ are comparable to those found for the GaN reference, confirming that the dominating conduction channel at high temperatures in this sample is the underlying GaN buffer.
The MR curves acquired at temperatures below 50\,K are represented in Fig.\,\ref{fig:MR}(b), and show a negative slope up to a critical field $H_\mathrm{s}$, and a positive quadratic behavior at higher magnetic fields. In the VRH regime, the MR depends on the variations in the hopping probability with the magnetic field. Specifically, for low magnetic fields, the negative magnetoresistance (NMR) is due to quantum interference of different hopping paths between initial and final states, similarly to weak-localization in metals\,\cite{Zhao:1991_PRB}. When the magnetic field is increased, the effect is overcome by the presence of a strong positive MR\,\cite{Gantmakher:1996_JETP}. The critical value of the field at which the MR changes from negative to positive allows obtaining directly the characteristic hopping parameters $R_\mathrm{h}$ and $\xi$\,\cite{Entin:1989_PRB}.
The evolution of the MR acquired at 6\,T as a function of temperature for all samples is reported in Fig.\,\ref{fig:MR2_1}(a). At temperatures above 100\,K the MR of all samples is positive and follows a $H^2$ behavior, while at lower temperatures the MR of GaN is negative, and the one of Sample A* is dominantly positive reaching 42\% at 2\,K. The NMR of the GaN reference layer is analyzed in the frame of the Mott VRH conduction\,\cite{supplementary}, similarly to the case of \textit{n}-type InP\,\cite{Abdia:2009_SSE}.
The low temperature MR of Sample A* is fitted according to the model developed by Nguyen, Spivak and Shlovskii\,\cite{Zhang:1992_PRB}, and $R_\mathrm{h}$ and $\xi$ as a function of temperature directly obtained from the critical field $H_\mathrm{s}$ are plotted in Fig.\,\ref{fig:MR2_1}(b). The values of $R_\mathrm{h}$ lie between 30\,nm and 70\,nm, while those of $\xi$ vary in the range (7-10)\,nm, fulfilling the condition $R_\mathrm{h}$$>$$\xi$ for hopping conduction\,\cite{Mott:1969}. From HRTEM plane-view measurements\,\cite{Navarro:2012_APL}, the average distance between the NCs is found to range between 30\,nm and 100\,nm, rendering hopping conduction through the NCs significant at temperatures below 10\,K.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig6.eps}
\caption{Magnetoresistance of Sample A* as a function of the applied magnetic field: (a) above 50\,K, and (b) below 50\,K. Symbols: experimental data; solid lines: fits.}
\label{fig:MR}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Fig7.eps}
\caption{(a) Comparison of the MR as a function of temperature for Samples A, A*, GaN and B at 6\,T. (b) $R_\mathrm{h}$ and $\xi$ obtained as a function of temperature for Sample A*. (c) MR and normalized density of magnetic moment as a function of the applied field for Sample A*. The field sweep directions are indicated by arrows.}
\label{fig:MR2_1}
\end{figure}
As evidenced in Fig.\,\ref{fig:MR2_1}(c), a hysteretic behavior of the MR of Sample A* is observed at low fields, similar to the one reported for the ZnO:Co phase-separated system\,\cite{Hamieh:2015_PRB} and for GeMn containing ferromagnetic Ge$_3$Mn$_5$ nanocrystals\,\cite{Yu:2010_PRB}. When sweeping the magnetic field from -6\,T to 6\,T (stars -- upward arrow) and $vice\,versa$ (diamonds -- downward arrow), the maxima of the MR curves acquired at 2\,K are shifted by 120\,mT with respect to each other. This value corresponds to twice the coercive field at this temperature, as is indicated by the hysteresis shown in Fig.\,\ref{fig:MR2_1}(c), suggesting that until the magnetic moment of the NCs is aligned with the magnetic field, the electronic transport is sensitive to the spin polarization in the NCs.
The observed behavior of the MR as a function of temperature can be explained as follows: at temperatures above 50\,K, and due to phonon scattering, the spin life time of the spin-polarized conduction electrons in the $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs is too short to span the distance between the NCs. However, for temperatures below 25\,K a spin-polarized current can be transported between the NCs, leading to a reduction of the MR. For temperatures around 10\,K, the $R_\mathrm{h}$ is too low for a hopping conduction between NCs to take place and hopping $via$ intermediate states can be ruled out, due to spin flip processes at the paramagnetic spins of the dilute Fe ions in the host matrix\,\cite{Zhao:1991_PRB}. At temperatures below 10\,K, $R_\mathrm{h}$ becomes sufficiently large for inter-NC-hopping to occur, increasing the hopping probability.
The MR of Sample B differs significantly from the one of the other investigated samples: it is positive down to 150\,K and it turns negative at 100\,K\,\cite{supplementary}. Due to the mentioned diffusion of dilute Fe from the layer containing the NCs into the GaN capping layer, the disorder of the system increases and the VRH conduction mechanism is likely to dominate already at temperatures as high as 100\,K. This conclusion is supported by the fact that a similar MR behavior is observed in semi-insulating dilute GaN:Fe layers, as reported in Fig.S3(d) of the Supplemental Material\,\cite{supplementary}. However, as the measured conductivity is a combination of the conductivities of the underlying GaN buffer, of the Ga$\delta$FeN layer containing the NCs and of the GaN(:Fe)capping layer, in this sample also the Arrhenius behavior from the buffer layer is significant in a wide temperature range.
The contribution of the embedded NCs to the MR in Sample A* is further tested by changing the direction of the applied magnetic field by the out-of-plane angle $\beta$ with respect to the surface normal from perpendicular ($\beta = 90^\circ$) to parallel ($\beta = 0^\circ$). The anisotropic magnetoresistance is defined as:
\begin{equation}
\mathrm{AMR} = \frac{\rho(\beta)-\rho(0^\circ)}{\rho(0^\circ)}.
\end{equation}
All samples investigated and containing Ga$\delta$FeN show a positive AMR throughout the entire temperature range between 2\,K and 300\,K. The AMR is not observed in the GaN reference, pointing at the observed AMR as due to the $\gamma$'-Ga$_y$Fe$_{4-y}$N nanocrystals present in the other layers. The AMR acquired at 150\,K for Sample A* is reported in Figs.\,\ref{fig:AMR}(a). The AMR follows a $cos^2(\beta)$ behavior, consistent with the out-of-plane magnetic anisotropy of the ferromagnetic Ga$_y$Fe$_{4-y}$N NCs obtained by FMR. A positive AMR has been also reported for epitaxial thin films of $\gamma$'-Fe$_4$N grown on SrTiO$_3$(001) substrates\,\cite{Nikolaev:2003_APL}.
The evolution of the AMR ratio, defined as $(\rho_{\parallel}-\rho_{\perp})/\rho_{\perp}$, where $\rho_{\parallel}$ and $\rho_{\perp}$ are the resistivities for an applied field respectively parallel or perpendicular to the sample normal, as a function of temperature is presented in Fig.\,\ref{fig:AMR}(b) for all investigated samples. The highest AMR for Sample A* is 16\% at 2\,K, while it reaches values around (2-4)\% up to 300\,K. The AMR values for Sample B are slightly larger than those of Sample A*, pointing to a greater amount of ferromagnetic NCs in this sample. This is in accordance with the faster saturation observed in the magnetization of Sample B in Fig.\,\ref{fig:magnetic}(a).
\begin{figure}
\includegraphics[width=\columnwidth]{Fig8.eps}
\caption{(a) AMR as a function of the out-of-plane angle $\beta$ between the magentic field and the surface normal (symbols) in Sample A* fitted by a $cos^2\beta$ function (solid lines). (c) AMR ratio as a function of temperature for Sample A* and Sample B.}
\label{fig:AMR}
\end{figure}
Anisotropic magnetoresistance is an effect generally dominated by spin-orbit interaction, which causes spin-mixing in the scattering processes of the conduction electrons with $3d$ orbitals, \textit{i.e.} \textit{s-d} scattering. When in $\gamma$'-Fe$_4$N the electron occupation changes as a function of the direction of the magnetic moment with respect to the cubic axes, the AMR is affected\,\cite{Tsunoda:2010_APEX}. In $\gamma$'-Fe$_4$N, the conduction electrons are preferentially scattered by the $3d$ orbitals when the magnetization is parallel to the easy axis, leading to an increased AMR when the current flows along this direction. According to the FMR measurements, the $\gamma$'-Ga$_y$Fe$_{4-y}$N nanocrystals embedded in GaN have a strong uniaxal magnetic anisotropy with three easy axes lying in the plane normal to the $c$-axis of GaN\,\cite{Grois:2014_Nanotech}. Therefore, when the magnetic field is applied parallel to the sample surface ($\beta = 0^\circ$), \textit{i.e.} parallel to the in-plane easy axis, the electron scattering is enhanced and hence, an increased resistance is observed. In contrast, when the field is applied normal to the surface ($\beta=90^\circ$), the electron scattering is reduced and the resistance diminishes.
\section{Conclusions}
The magnetotransport in phase separated (Ga,Fe)N containing $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs and grown by MOVPE on a GaN buffer deposited on $c$-sapphire is investigated.
The results show that, while the overall conduction mechanism in the studied layers occurs at the unintentionally \textit{n}-doped GaN buffer, the MR is significantly affected by the presence of the embedded $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs. This is supported by the observed MR hysteresis at magnetic fields below 1\,T. The behavior of the MR in these layers can be described in terms of: (i) an Arrhenius-like mechanism for temperatures $\geq$50\,K, and (ii) an inter-nanocrystal hopping conduction, where the spin-polarized current is transported between NCs at temperatures below 25\,K in a regime in which phonon-scattering effects are not dominant. The hopping probability is increased at temperatures below 10\,K, where $R_\mathrm{h}$ becomes sufficiently large for inter-hopping to occur.
In contrast to the negative AMR generally observed in $\gamma$-Fe$_4$N polycrystalline and epitaxial thin layers\,\cite{Ito:2012_JJAP,Tsunoda:2010_APEX}, the AMR of the $\gamma$'-Ga$_y$Fe$_{4-y}$N NCs embedded in GaN is positive with a high resistance state when a magnetic field is applied along the surface plane -- \textit{i.e.} parallel to the easy axis of the magnetization -- and with a low resistance state when the magnetic field is applied perpendicular to the surface. This is consistent with the in-plane shape anisotropy of the ferromagnetic NCs.
The observed values of the AMR ratio are $\sim$(2-3)\% at RT, $i.e.$ significantly higher than the 0.17\% so far reported for $\gamma$'-Fe$_4$N thin films\,\cite{Nikolaev:2003_APL}, and open wide perspectives for the manipulation of AMR by external electric and magnetic fields -- previously reported for dilute (Ga,Mn)N \,\cite{Sztenkiel:2016_NComm} -- in these phase-separated systems.
\begin{acknowledgments}
This work was supported by the Austrian Science Fund (FWF) through the Elise-Richter Project No.V478, and projects P26830 and P24471, and the Austrian Exchange Service (\"{O}AD) project PL-01/2017, by the National Science Centre (Poland) [project OPUS (DEC-2013/09/B/ST3/04175)] and by the European Commission through the InTechFun (POIG.01.03.01-00-159/08) grant. The authors want to acknowledge C. Atteneder, P. Lindner and S. Wimmer for technical support and P. Dluzewski for his contribution to the HRTEM measurements.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
The aromatic infrared bands (AIBs) dominate the near and mid IR spectra of many interstellar sources. These bands are generally attributed to IR fluorescence of large ($\sim$ 50 C atoms) polycyclic aromatic hydrocarbon (PAH) molecules after excitation by ultraviolet (UV) photons \citep{all89,pug89,sel84,gen98,tie13}. PAHs are expected to form in the ejecta of AGN stars and are ubiquitous and abundant, containing $\sim$ 10\% of the elemental carbon in the interstellar medium (ISM). They play an important role in the ionization and energy balance of the ISM \citep[and references therein]{tie08}. The assignment of the AIBs to PAH molecules has been discussed in several studies \citep{all89,peeters2002,tie05,tie08}. Generally, the prominent 3.3 $\mu$m band is linked to aromatic CH stretches, the 6.2 and 7.7 $\mu$m features are considered to be due to CC vibrations of aromatic rings, and the 8.6 $\mu$m as well as the longer wavelength region bands are assigned to CH in- and out--of--plane bending and/or CCC bending vibrations \citep{tie13}.
IR spectroscopy provides a powerful tool to study the molecular characteristics of the emitting species. Specifically, for the CH out$-$of$-$plane (oop) bending vibrations in the 11$-$14 $\mu$m region, a different number of adjacent CH bonds gives rise to vibrations in distinctly different wavelength regions \citep{hon01,tie08}. These spectroscopic rules on H-adjacency and the positions of the oop modes have largely been derived from experimental studies on small (substituted) aromatic systems such as benzene and naphthalene \citep{bellamy} and are supported by Density Functional Theory (DFT) calculations \citep[c.f.][]{bauschlicher2010}. Furthermore, laboratory and quantum chemical studies have revealed that the CC stretching modes in the 6$-$9 $\mu$m region are much stronger in ionized PAHs than in their neutral counterparts \citep{allamandola1999,peeters2002}. These three bands (6.2, 7.7, and 8.6 $\mu$m), which correlate well, therefore {\bf may be attributed} to PAH cations. Recent studies, using PAH ions trapped in hydrogen matrices {\citep{tsu16}} or measured in tagged ion dissociation studies \citep{rick09,har09} also show nice agreement with vibrational modes recorded for protonated PAH species. This is supported by observational studies that reveal that these bands grow in strength -- relative to the CH bands characteristic for neutral PAHs -- in regions characterized by high ionization parameters \footnote{the ionization parameter is given by the ratio of the ionization rate over the recombination rate and scales with $G_0T^{1/2}/n_e$ with $G_0$ the strength of the radiation field in terms of the average interstellar radiation field, $T$ the temperature, and $n_e$ the electron density \citep{tie05}.}; e.g., close to the illuminating source \citep{berne2007,rapacioli2005,boersma2013}. However, in such regions, PAH cations can also be further ionized, resulting in doubly, triply, or even higher charged ions. In competition with the ionization channel, PAH cations can also fragment upon energetic photo-excitation. This fragmentation follows the available pathways (i.e., dehydrogenation or loss of C$_{2}$ or C$_{2}$H$_{2}$ units), producing PAH derivative species along a top-down chemical scenario \citep{zhen2014, bou16, pet16}. Photo-ionization and photo-fragmentation compete, and their relative importance depends both on the geometry of PAH molecules and the excitation wavelengths used \citep{zhen2015, zhe2016}.
In order to elucidate possible AIB carriers, neutral and ionized PAH molecules with different size and molecular geometries have been studied, both in the laboratory and by theory. Experimentally, IR spectra of PAH neutrals, cations and other derivatives have been measured in rare gas matrices \citep[e.g.][]{szc931, szc932, hud951, hud952, mat03, ber07, tsu16}. In the gas phase IR emission spectra have been studied for neutral PAHs and PAH cations \citep{coo98, kim01, kim02}. Absorption gas phase spectra of PAH cations have been studied using IRMPD \citep{oom01, oom06} or messenger atom photo-dissociation (MAP) spectroscopy \citep{pie99, dun03, dou08, rick09}. A recent application shows that this method is even suited to study fullerene cations \citep{cam15,kuh16}. While some PAH spectral bands coincide approximately with prominent AIBs, the agreement is not always compelling. A more convincing overlap with astrophysical emission spectra is obtained with composite spectra that consist of (selected) distributions of spectra of different PAH and PAH$^{+}$ species that have been recorded using matrix isolation spectroscopy or computed using DFT calculations \citep{peeters2002, cami2011}.
For technical reasons, IR laboratory studies of PAHs have largely been focused on smaller species, containing up to 24 C-atoms. IR spectra of large interstellar relevant PAHs ($\sim$ 50 C-atoms) are generally lacking and this is also definitely the case for their cations or di-cations. In this article, we present the first IR spectra of the cation and di-cation of hexa-peri-hexabenzocoronene: C$_{42}$H$_{18}^+$ and C$_{42}$H$_{18}$$^{2+}$. The resulting IR spectra are compared with quantum chemical computations and astronomical data. HBC (Figure 1) has been selected as it is an all-benzenoid PAH with a size in the astrophysically-relevant range \citep{tie08,croiset2016} that possibly serves as a prototypical example for large(r) PAHs. Moreover, the molecular geometry of this PAH is special, as it has only CH groups with a \textquotedblleft triply-adjacent\textquotedblright or \textquotedblleft trio\textquotedblright CH-structure, as shown in Figure 1. This allows us to unambiguously identify the wavelength of the out-of-plane CH bending mode(s) for such a geometry.
\section{Experimental Methods}
\label{sec:exp}
The IR spectra of the fully-benzenoid hexa-peri-hexabenzocoronene cation (C$_{42}$H$_{18}^{\phantom{26}+}$ with $m/z = 522.14$) and its di-cation (C$_{42}$H$_{18}$$^{2+}$) with $m/z = 261.07$) have been recorded using our instrument for Photo-dissociation of PAHs (i-PoP) connected to the Free Electron Laser for InfraRed eXperiments (FELIX) at Radboud University \citep{oep95}. The apparatus comprises a quadrupole ion-trap (QIT) time-of-flight (TOF) mass spectrometer and has been described in detail in \citet{zhe14}. Below only a brief description of the relevant details is given.
Neutral HBC precursor molecules are transferred into the gas phase by heating HBC powder (Kentax, purity higher than 99.5 \%) in an oven ( $\sim$ 580K) until it slowly evaporates.
The sublimated HBC molecules are ionized using electron impact ionization (83~eV electron impact energy). The resulting HBC$^+$ and HBC$^{2+}$ cations are transported into the ion trap via an ion gate and are trapped by applying a 1~MHz radiofrequency electric field of 3150 and 2130 V$_{p.p.}$, respectively, onto the ring electrode. Helium is introduced continuously into the trap to thermalize the ion cloud through collisions.
Spectral information is obtained over the 530 to 1800~cm$^{-1}$ range by IRMPD using FELIX. This free electron laser delivers 5~$\mu$s long macropulses of light at a repetition rate of 10~Hz with a macropulse energy up to 100~mJ. The bandwidth of the laser is Fourier limited and is typically 0.6\% (FWHM) of the central wavelength. When the laser wavelength is in resonance with an allowed vibrational transition in the ion, the absorption of multiple photons takes place, causing the molecule to dissociate. The resulting fragments are then pulse-extracted from the ion trap and analyzed in a TOF mass spectrometer. All channels corresponding to photo-fragmentation of the parent are summed up and normalized to the total signal, i.e. parent plus photo-fragment ions in the trap, yielding the relative photo-fragmentation intensity. The infrared spectrum of the ion is constructed by plotting the fragment ion yield as a function of wavelength. The recorded spectra are normalized to the {\bf pulse energy} to correct for any power dependencies.
Mass spectra data for HBC$^+$ and HBC$^{2+}$ are recorded typically with steps of 5 cm$^{-1}$. Optimal irradiation times amount to 0.8 s (typically $\sim$ 8 pulses). In order to obtain a good signal-to-noise ratio, mass-spectra are averaged 25 times. In addition, blank spectra are recorded in order to subtract fragmentation peaks that are already present in the background signal.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig1.pdf}
\caption{The molecular geometry of hexa-peri-hexabenzocoronene (HBC, C$_{42}$H$_{18}$).
}
\label{fig1}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{fig2.pdf}
\caption{Mass spectrum of the photo-products resulting from irradiation of HBC$^+$ and HBC$^{2+}$: (A) The photo-dehydrogenation behavior of HBC$^+$, irradiated at different wavelength in the 762cm$^{-1}$ band range; (B) The photo-dehydrogenation behavior of HBC$^{2+}$, irradiated at different wavelengths in the 766cm$^{-1}$ band range. Additional peaks in the background trace are isotopes and fragmentations induced by the electron gun.
}
\label{fig2}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{fig3.pdf}
\caption{Measured IRMPD spectrum of gas-phase HBC$^+$ and HBC$^{2+}$. Computed vibrational normal modes scaled to correct for anharmonicities (taken from \citet{boe2014}) are represented by vertical bars.
}
\label{fig3}
\end{figure*}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{fig4.pdf}
\caption{IR spectra of HBC$^+$ and HBC$^{2+}$ (top) compared to two representative AIB emission spectra (bottom) taken towards NGC 6946 and NGC 4569 and adopted from \citet{smi07}.
}
\label{fig4}
\end{figure}
\section{Results and discussion}
\label{sec:results}
Figure 2 shows the mass spectra of the photo-products that result upon excitation of the 762~cm$^{-1}$ band of HBC$^+$ (panel A) and the 766 cm$^{-1}$ band of HBC$^{2+}$ (panel B). The mass spectra before irradiation (labeled as background) reveal a small amount of residual fragmentation due to the electron impact ionization and also include $^{13}$C isotopes that contribute to the main mass peak signal. The different graphs in Fig. 2(A) show the fragmentation patterns of trapped HBC$^+$ for a set of wavelengths between 750 and 780~cm$^{-1}$. Several dehydrogenated fragment ions can be seen that are the consequence of multiple and cumulative photon absorptions. The photo-dissociation pattern follows mainly sequential 2H (or H$_2$) loss channels, similarly to what was observed for HBC$^+$ in \citet{zhe14} upon irradiation with optical photons (532nm). Figure 2(B) shows the same data for HBC$^{2+}$, now after irradiation with wavelengths between 730 and 810 cm$^{-1}$. Similar to the photo-fragmentation behavior of HBC$^+$, a wide range of partially dehydrogenated fragment ions is found and also in the case of HBC$^{2+}$ the photo-dissociation pattern mainly follows sequential 2H (or H$_2$) losses.
Figure 3 shows the resulting IR spectra for HBC$^+$ (panel A) and HBC$^{2+}$ (panel B) from 530 to 1800 cm$^{-1}$ (5.6$-$18.9 $\mu$m). The spectra are recorded by integrating the fragmentation signals that vary with wavelength as illustrated in Figure 2. This figure shows a number of clearly resolved vibrational bands, with reasonable signal-to-noise ratios, even in the case of HBC$^{2+}$. Many of the bands are rather broad, which is commonly observed for IRMPD spectra and is attributed to anharmonic effects induced by the multiphoton process. It is also possible that broad unresolved bands hide underneath the separate bands visible here. This broadening may be less prominent in experiments using rare gas ion tagging. Such experiments, however, may come with the drawback that transitions substantially shift upon complexation, depending on where the atom tags to the PAH. So far, the existing tagging experiments are available mainly for smaller PAH-cations, and likely also hard to apply to di-cations. An extension to larger species will be needed to directly compare with the results presented here.
Also included in Figure 3 are the theoretical predictions of the vibrational band positions of HBC$^+$ and HBC$^{2+}$ \citep{boe2014} that have been scaled by a uniform factor of 0.958 to correct for anhamonicities\footnote{This global scaling factor corrects approximately for the shifts in peak position due to anharmonicity at 0K (Langhoff 1996). This is an approximation as individual bands show different anharmonic corrections in the calculations (Mackie et al 2016). We also stress that this does not account for the anharmonic shifts in peak position of highly excited species inherent to the IRMPD technique (Oomens et al 2003).}. In the case of HBC$^{2+}$ we considered both the singlet and triplet states, yet the singlet state is chosen for comparison as it is lowest in energy \citep{boe2014}. Table~1 summarizes all vibrational and computed wavelengths with their assignments. The vibrational peak positions are determined directly from the maximum intensity without fitting the overall spectrum. The latter prohibits an overfitting of our data as we do not know how many broad bands may contribute to the overall profile.
Fig. 3(A) and Table 1 compare our experimental IRMPD HBC$^+$ spectrum with the theoretical predictions \citep{boe2014}. The most prominent bands include the carbon ring distortions at 1280 (7.81), 1379 (7.25) and 1530 (6.54) cm$^{-1}$ ($\mu$m), and the CH in- or out-of-bending modes at 762 (13.12), 1060 (9.43) and 1192 (8.39) cm$^{-1}$ ($\mu$m). The intense band at 762 cm$^{-1}$ with a FWHM of $\sim$ 20 cm$^{-1}$ in the IRMPD spectrum corresponds with the 779 cm$^{-1}$ feature in the DFT spectrum, which arises from out-of-plane CH bending modes. The bands measured at 1060 and 1192 cm$^{-1}$ are close to transitions predicted at 1077 and 1223 cm$^{-1}$, which arise from in-plane CH bending modes. The experimental bands at 1280 and 1379 cm$^{-1}$ are close to the theoretically predicted bands at 1289 and 1374 cm$^{-1}$. It is possible that the 1280 cm$^{-1}$ band also includes bands at 1303 and 1308 cm$^{-1}$. All these bands have a FWHM of $\sim$ 60 cm$^{-1}$. The band at 1530 cm$^{-1}$ is broad -- with a FWHM of $\sim$ 120 cm$^{-1}$ -- and likely is composed of a number of bands predicted at 1482, 1549, 1559 and 1579 cm$^{-1}$ (Table 1).
Figure 3(B) and Table 1 provide similar information for HBC$^{2+}$. The error bars here are larger due to a lower ion signal, as fewer di-cations are trapped in the QIT. The most prominent bands include again the carbon ring distortions at 1290 (7.75), 1370 (7.30) and 1530 (6.54) cm$^{-1}$ ($\mu$m), and the CH in- or out-of-bending modes at 660 (15.15), 766 (13.05), 1054 (9.49) and 1176 (8.50) cm$^{-1}$ ($\mu$m). The bands at 660 and 766 cm$^{-1}$ with FWHMs of $\sim$ 20 cm$^{-1}$ in the IRMPD spectrum correspond with the bands at 675 and 785 cm$^{-1}$ in the DFT spectrum, {\bf which arise} from out-of-plane CH bending modes. The band measured at 1054 cm$^{-1}$ is close to the transition predicted at 1075 cm$^{-1}$, and the band at 1176 cm$^{-1}$ in the IRMPD spectrum links to bands at 1217 and 1234 cm$^{-1}$, which arise from in-plane CH bending modes. The band measured at 1290 cm$^{-1}$ corresponds to the bands computed at 1299 and 1328 cm$^{-1}$ and possibly also contains contributions from nearby bands at 1278, 1290 and 1328 cm$^{-1}$. The band at 1370 cm$^{-1}$ in the IRMPD spectrum is found in the DFT calculations at 1391 cm$^{-1}$ and possibly also contains contributions from the computed 1369 and 1408 cm$^{-1}$ bands (Table 1). All these bands have a FWHM of $\sim$ 60 cm$^{-1}$. The {\bf band at 1530 cm$^{-1}$ is broad band} with a FWHM of $\sim$120 cm$^{-1}$ and likely contains contributions from other bands calculated at 1487, 1532, 1546 1566 and 1583 cm$^{-1}$ (Table 1).
As a general point, we emphasize that, despite the fact that a scaling factor of 0.958 is already applied, all experimentally measured band positions of HBC$^+$ and HBC$^{2+}$ are red-shifted with respect to the calculated positions. We attribute this shift to the large anharmonicity associated with the high internal excitation required for the dissociation of large PAHs, which is inherent to the IRMPD technique \citep{oom03}. We note that these shifts are different for different vibrational modes. For example, from Fig. 3(A) and Table 1, it becomes clear that the 1192 cm$^{-1}$ band is about $\sim$ 30 cm$^{-1}$ situated to lower energy than predicted, whereas for the 1280 cm$^{-1}$ band the shift is only $\sim$10 cm$^{-1}$. Taking this into account, the agreement between the experimental values presented here and the available theoretical data is reasonable.
While overall, the comparison between experiments and theory is very encouraging, there are also some issues. The HBC$^+$ and HBC$^{2+}$ bands at 1060 and 1054 cm$^{-1}$, respectively, exhibit seemingly broad wings towards lower frequencies in the 850$-$1000 cm$^{-1}$ range. Theory does not predict vibrational normal modes in this region, and these broad bands may be examples of the extreme effects of anharmonicity on IRMPD spectra. As another discrepancy with theory, we note that theory predicts bands at 588 and 587 cm$^{-1}$ for HBC$^+$ and HBC$^{2+}$, respectively, which are not apparent in the experiments. Perhaps, anharmonicity is also a factor here, shifting these bands outside of the spectral range scanned with FELIX. In addition, the non-linear intensity response of IRMPD in combination with a lower output power at the low (and high side) of the accessible wavelength domain may be also responsible for this observation.
One further difference occurs at high frequencies where DFT calculations predict a strong band for HBC$^+$ at 1549 cm$^{-1}$ and for HBC$^{2+}$ at 1583 cm$^{-1}$ with a number of relatively weaker bands nearby (Table 1). In contrast, rather than one strong band, the experiments reveal weak, but broad features between 1430 and 1630 cm$^{-1}$. This difference may again reflect extreme anharmonicity for these modes. Likewise, we do detect the theoretically predicted band at 660 cm$^{-1}$ of HBC$^{2+}$ but a similarly predicted band of HBC$^+$ is missing in our spectrum. On the one hand, this may indicate an issue with the theoretical spectrum of HBC$^+$. In addition, this might result from a much larger anharmonicity in HBC$^+$ (than in HBC$^{2+}$) which would shift this band rapidly out of resonance in our experiment before fragmentation can occur. Further experiments and quantum chemical calculations are required to improve our understanding of these observations.
\begin{table*}
\centering
\caption{Infrared Transitions Measured and Calculated for HBC$^+$ and HBC$^{2+}$
\label{BenchCalcs}}
\begin{tabular}{llllllllll}
\hline
\multicolumn{3}{c}{HBC$^+$}&\multicolumn{3}{c}{HBC$^{2+}$}&\\
Experimental & Computed & Computed & Experimental & Computed & Computed & Assignment\\
band&band&band&band&band&band&&\\
cm$^{-1}$, $\mu$m & cm$^{-1}$& km$/$mol& cm$^{-1}$, $\mu$m & cm$^{-1}$& km$/$mol &\\
\hline
& 262& 33&&264& 23&\\
& 588& 94&&587& 106&\\
& 676& 82$^c$& 660 (m, A), 15.15& 675& 98&out-of-plane CH bending\\
732, (w$^a$, A$^b$), 13.66& 728& 21&720 (w, B), 13.89& 722& 20&out-of-plane CH bending\\
762 (s, B), 13.12& 779 & 267&766 (s, A), 13.05& 785& 268&out-of-plane CH bending\\
&819& 43&804 (w, B), 12.44& 819& 43&out-of-plane CH bending\\
&&&&924& 21&\\
1060 (s, B), 9.43& 1077& 457&1054 (s, B), 9.49& 1075& 367&in-plane CH bending\\
&&&&1172&26&\\
1192 (s, B), 8.39& 1223& 828&1176 (s, B), 8.50& 1217& 318&in-plane CH bending\\
&&&&1234& 234&\\
&&&&1238& 28&\\
&1258& 26&\\
&&&&1278& 172&\\
&&&&1290& 227&\\
1280 (s, B), 7.81& 1289& 800&1290 (s, B), 7.75& 1299& 625& carbon ring distortions\\
&1303& 170&\\
&1308& 259&\\
&1324& 23&&1328& 482&\\
&&&&1331& 70&\\
&&&&1369& 101&\\
1379 (s, B), 7.25& 1374& 115&1370 (s, B), 7.30& 1391& 185& carbon ring distortions\\
&1394& 32&\\
&&&&1408& 114&\\
&&&&1460& 30&\\
&1482& 140&&1487& 154&\\
&1497& 29&&1498& 76&\\
&&&&1515& 57&\\
&&&&1532& 172&\\
&&&&1546& 251&\\
1530 (m, C), 6.54& 1549& 878&1530 (m, C), 6.54& 1566& 468&carbon ring distortions\\
&1559& 188&&1583& 658&\\
&1579& 72&\\
\hline
\end{tabular}
\\
{$^a$}Relative intensities of experimental bands are indicated as w, m, s for weak, medium, strong, respectively;
{$^b$}The uncertainty in the experimental band positions is estimated - following the NIST definition - to be A (1$\sim$3 cm$^{-1}$), B (3$\sim$6 cm$^{-1}$) and C (15$\sim$30 cm$^{-1}$);
$^c$ Vibrational frequencies were scaled by a factor of 0.958 to allow comparison with the experimental data, consistent with the recommended scaling for this level of theory \citep{sco96}, and intensities (km$/$mol) are given in parentheses. We only list bands with intensities larger than 20 km$/$mol.
\end{table*}
\section{Astrophysical relevance and conclusions}
\label{sec:astro}
This paper presents the first IR spectrum of a PAH cation with an astrophysically relevant size -- C$_{42}$H$_{18}$ -- in the gas phase. In addition, this is the first time that the spectrum of a PAH di-cation has been measured experimentally in the gas phase. The experimental IR spectra for both HBC$^+$ and HBC$^{2+}$ (top two traces) are shown together with typical AIB emission spectra (bottom two traces) in Fig.~4. The astronomical emission spectra are from NGC 6946 and NGC 4569 \citep{smi07}. This interstellar spectrum shows the well-known prominent broad AIBs at 6.2, 7.7, 8.6, 11.2 and 12.7 $\mu$m. As shown in the figure, the experimental spectra of HBC$^+$ and HBC$^{2+}$ exhibit similarities to several of the key bands in the AIB spectrum. The strong bands at 7.7 and 8.6 $\mu$m in the laboratory spectrum closely match the positions of the AIB bands at these wavelengths, and the 13.1 $\mu$m band is very close to the 12.7 $\mu$m band due to $"$trio$"$ CH-structure bending modes in HBC. The 11.2 $\mu$m band that shows up in the astronomical spectrum is absent in the HBC data, as this band is due to ``solo'' CH-structure excitation \citep{hon01} and this group is not present in HBC. As such, this comparison also illustrates that the astronomical spectra may be a combination of emission features originating from many separate species, and likely HBC$^+$ and HBC$^{2+}$ will contribute but not at a level that allows an unambiguous identification.
Theoretical and experimental studies have highlighted the large spectral differences between neutral and ionized PAHs (Langhoff 1996; Allamandola et al 1999). As charging involves the delocalized $\pi$ electron system (rather than the localized $\sigma$ electrons), bond strength and mode distribution do not differ much between these charge states. The change in charge distribution does influence, though, the change in dipole moment during the vibration and this is the underlying cause for the striking spectral differences. Our experiments show that -- in contrast to the neutral-cation behavior -- there is no clear spectral signature that can lead to the unambiguous identification of di-cations in space. This is unfortunate, as a tool for determining the charge state of PAHs in space would contribute substantially to our knowledge on how PAHs behave upon irradiation. The main difference occurs in the 600-700 cm$^{-1}$ region where a band at 660 cm$^{-1}$ in HBC$^{2+}$ is present in both our experimental study and in the DFT calculations but the corresponding theoretical band in HBC$^+$ is not observed in the IRMPD study. As argued above, this may reflect a difference in anharmonicity between HBC$^{2+}$ and HBC$^+$ -- rather than a spectral signature of di-cations -- which adversely affects the measured spectra. Further studies will be required to address this issue.
The strong transitions in the 530$-$1800 cm$^{-1}$ (5.6$-$18.9 $\mu$m) IR fingerprint region have been compared to published DFT data and we conclude that, despite drawbacks that complicate the interpretation of band profiles and band intensities, the applied IRMPD technique provides useful band positions for IR active modes. Specifically, we emphasize that the measured HBC spectra support the assignment \citep{hon01} of the interstellar 12.7 $\mu$m band to the CH out-of-plane bending mode associated with trio's on (large) PAHs (although a duo assignment for the interstellar band may also be viable; \citep{bauschlicher2010}). Finally, we stress that this study demonstrates the potential of IRMPD spectroscopy for much larger species than measured before and this work paves the way to systematically study even larger PAHs ($>$60 C-atoms) with our set-up in the future.
\acknowledgments
The authors gratefully acknowledge the FELIX staff for their technical assistance during preparation and the actual beamline shifts. We are grateful to M.J.A.\ Witlox for technical support. We are grateful to Giacomo Mulas and Tao Chen for many very useful discussions. Studies of interstellar chemistry at Leiden Observatory are supported through advanced-ERC grant 246976 from the European Research Council, through a grant by the Netherlands Organisation for Scientific Research (NWO) as part of the Dutch Astrochemistry Network, and through the Spinoza premie. JB acknowledges NWO for a Veni grant (722.013.014). We acknowledge the European Union (EU) and Horizon 2020 funding awarded under the Marie Sklodowska-Curie action to the EUROPAH consortium, grant number 722346.
|
\section*{Introduction}
Direct brain-to-brain interfaces (BBIs) in humans\cite{Rao:2014,RaoPilot2013, Min:2010,Grau:2014,Stocco:2015} are interfaces which combine neuroimaging and neurostimulation methods to extract and deliver information between brains, allowing direct brain-to-brain communication. A BBI extracts specific content from the neural signals of a “Sender” brain, digitizes it, and delivers it to a “Receiver” brain. Because of ethical and safety considerations, existing human BBIs rely on non-invasive technologies, typically electroencephalography (EEG), to record neural activity and transcranial magnetic stimulation (TMS) to deliver information to the brain. For example, the first human BBI demonstrated by Rao and colleagues in 2013\cite{RaoPilot2013} decoded motor intention signals using EEG in the Sender and conveyed the intention via TMS directly to the motor cortex of the Receiver to complete a visual-motor task \cite{Rao:2014}. Stocco and colleagues\cite{Stocco:2015} extended these results by showing that a Sender and a Receiver can iteratively exchange information using a BBI to identify an unknown object from a list, using a question-and-answer paradigm akin to “20 Questions.” Grau and colleagues\cite{Grau:2014} proposed a related but offline non-iterative BBI.
Early interest in human BBIs came from the potential for expanding human communication and social interaction capabilities \cite{RaoStocco:2014,Dingemanse:2017,Kyriazis:2015,Hongladarom2015,Montague2002}. However, previous BBIs have lacked several key features of real-world human communication. First, the degree of interactivity has been minimal; for example, in the case of the ``20 Questions'' BBI \cite{Stocco:2015}, the Sender only responds to the question the Receiver chooses, and the Receiver’s performance does not affect the Sender’s decision. Second, their interface required \emph{physical} action: the Receiver touched the screen to select a question. Thus, the communication loop was completed via a motor output channel rather than a brain interface. Third, all past human BBIs have only allowed two subjects. Human communication, on the other hand, has become increasingly dominated by means such as social media that allow multiple parties to interact in a network. The potential for BBIs allowing interactions between multiple humans has previously been theorized\cite{Min:2010,Nicolelis:2011} but not demonstrated.
Here, we present BrainNet (Figure~\ref{fig:brainnet}), a next-generation BBI that addresses many of the limitations of past BBIs. First, BrainNet is designed to be a BBI for more than two human subjects; its current implementation allows two Senders and one Receiver to communicate, but it can be readily scaled up to large numbers of Senders. The Senders have the same role in observing the current state of the task and conveying their decisions to the Receiver. The Receiver has the role of integrating these independent decisions and deciding on a course of action. Second, BrainNet's design incorporates a second round of interactions between the Senders and the Receiver, so that the action of the Receiver in the first round can be perceived by the Senders, giving them a second chance to convey (potentially corrective) decisions to the Receiver. Third, the Receiver is equipped with both TMS (to receive Senders’ decisions) and EEG (to perform an action in the task), thereby completely eliminating the need to use any physical movements to convey information. We report results from five groups, each with three human subjects (henceforth, ``triad''), who successfully used BrainNet to perform a collaborative task based on a Tetris-like game.
An important feature of communication in social networks is deciding which sources of information to pay attention to when deciding on a course of action\cite{Bakshy:2012}. To investigate whether BrainNet allows such a capability, we additionally explored whether the Receiver can learn the reliability of each Sender over the course of their brain-to-brain interactions. We varied the reliability of the signal from one Sender compared to the other by injecting noise into the signals from one randomly chosen Sender. Our results show that like conventional social networks, BrainNet allows a Receiver to learn to trust the Sender who is more reliable, i.e., whose signal quality is not affected by our manipulation.
\section*{Results}
To measure the direct brain-to-brain communication capabilities of BrainNet, we asked each triad of participants to perform 16 trials of an iterative Tetris-like game. In each trial, one participant, designated as the Receiver, is in charge of deciding whether or not to rotate a block before it drops to fill a gap in a line at the bottom of the screen. Critically, the Receiver is prevented from seeing the bottom part of the screen and must rely on the counsel of the other two participants, designated as the Senders, who can see the screen in its entirety. These Senders are tasked with making the correct decision (rotate or not) based on the shape of the current block and the gap at the bottom, and informing the Receiver of the decision via the brain-to-brain interface. All members of the triad communicate their decisions through an EEG-based interface using steady state visually evoked potentials (SSVEPs; see Methods). The Senders' decisions are delivered to the Receiver through two TMS pulses delivered sequentially to the occipital cortex, eliciting a phosphene for a "yes" decision or no phosphene for a "no" rotation decision for each Sender (see Methods). Each trial is composed of two rounds: the first round is as described above; after the first round, the Senders are given the opportunity to examine the Receiver's decision, shown on their screen as the block (now potentially rotated) mid-way through its fall. The Senders are then given another chance to make new (possibly corrective) suggestions to the Receiver through the brain-to-brain interface. A successful completion of a trial thus requires accurate communication between the Senders and the Receiver across these two rounds (see Figure~\ref{fig:two-step}). Further, to examine the issue of reliability of the Senders, our software randomly chooses one Sender to be less reliable by making the decision sent to the Receiver from that Sender incorrect in ten out of sixteen trials. The order of trials requiring the block to be rotated and trials not requiring rotation was pseudo-randomized, with the constraint that each half of the session contained 4 rotation and 4 non-rotation trials. Trials 8-12 for the first triad were excluded from all analysis due to a problem with the timestamp routine. We analyzed both the EEG and behavioral data from the subjects in the remaining trials.
\subsection*{Overall Performance}
The simplest measure of overall performance of the interface is the proportion of correct block rotations (equivalently, the proportion of number of lines cleared, or the proportion of the maximum theoretical total score, i.e. 16 points, achieved) for each of the five triads of participants. Figure~\ref{fig:accuracy} shows the results. The mean accuracy across all triads was 0.8125, corresponding to 13 correct trials out of 16. A corresponding $p$-value was calculated using the binomial distribution, which confirmed that the mean performance was indeed higher than expected by chance ($p$ = .002).
Another important metric is the mean performance of participants in the SSVEP task since both Senders and the Receiver in each triad had to use this method to share information. In the task, subjects focused their attention on a 17 Hz flashing LED to indicate a ``Rotate'' decision and a 15 Hz flashing LED to indicate a ``Do Not Rotate'' decision. Figure~\ref{fig:ssvep} shows that before and after the SSVEP task, the 17 Hz and 15 Hz average power values overlap, whereas during the task, the average power of the frequency corresponding to the correct answer in the trial is significantly larger than that of the frequency corresponding to the wrong answer (two-sample $t$-test; $t$(15) = 9.709, $p$ < 0.0001 for "Rotate" signal; $t(15)$ = 10.725, $p$ < 0.0001 for "Do Not Rotate" signal).
Since our SSVEP classifier compares the magnitude of power values to decode a Sender's decision, the large difference in power values implies good performance for our EEG-based brain interfaces.
As noted in previous studies from our group \cite{Rao:2014,Stocco:2015,Losey:2016}, raw accuracy can be an inadequate measure of performance because it does not differentiate the kind of mistakes being made, i.e., whether they are misses or false positives. A better measure of performance can be obtained by calculating each triad’s Receiver Operating Characteristic (ROC) curve\cite{Fawcett:2006}, which plots the True Positive Rate versus the False Positive rate, and calculating the area under this curve (AUC; Figure~\ref{fig:roc}). Responding uniformly at random yields an AUC of 0.5 while the AUC of an ideal observer is 1.0. Because the distribution of AUC values is constrained between 0 and 1, it does not conform to the normal distribution. Thus, to properly conduct statistical tests of these values, we followed two complementary approaches. First, we conducted $t$-tests on the angular transformation (i.e., the arcsine square root transformation) of the AUC values, a common technique used to normalize data distributions \cite{rao1998statistical}. Second, we entered the raw, untransformed values in a Wilcoxon test, a non-parametric test with a continuity correction. Both tests confirmed that the mean AUC value of 0.83 across all triads of participants was significantly higher than the performance expected by chance (one-sample $t$-test on angular transformed data: $t$(4) = 11.366, $p$ < .001; one-sample Wilcoxon test: $V$ = 15, $p$ = 0.031).
As Figure~\ref{fig:roc} shows, the overall AUC value for each triad of brains is affected by the bad Sender's performance, but not by much. The overall AUC values are smaller than the AUC values of the good Senders (two-sample $t$-test on angular transformed data: $t$(4) = -2.897, $p$ = .021; Wilcoxon test: $W$ = 2, $p$ = .036) but significantly larger than those of the bad Senders (two-sample $t$-test on angular transformed data: $t$(4) = 9.184, $p$ < 0.001; Wilcoxon test: $W$ = 25, $p$ = 0.008).
\subsection*{Mutual Information Between Participants}
An important measure of a brain-to-brain interface is the mutual information (MI)\cite{Cover:2006} transmitted between subjects, which is defined as:
$$MI(R, S) = \sum_{r\in \{0, 1\}}\sum_{s\in \{0, 1\}}p_{R, S}(r, s)\log\frac{p_{R, S}(r, s)}{p_{R}(r)p_{S}(s)}$$
\noindent where $r$ represents a decision made by the Receiver (0 or 1 corresponding to "do not rotate" or "rotate"), $s$ represents a decision made by one of the Senders, $p_{R}(r)$ represents the probability of the Receiver making the decision $r$, $p_{S}(s)$ represents the probability of one of the Senders making the decision $s$, and $p_{R, S}(r, s)$ represents the {\em joint} probability of the Receiver making the decision $r$ and a Sender making the decision $s$. Note that, in this case, chance performance corresponds to $\textrm{MI} = 0.0$ while perfect communication corresponds to $\textrm{MI} = 1.0$. Because mutual information values are also constrained between 0 and 1 and, therefore, are not normally distributed, we analyzed them using the statistical methods we applied to the AUC values (i.e., $t$-test on angular transformed data and Wilcoxon test with continuity correction).
Due to our experimental design, we expect significantly higher MI values (i.e., larger amounts of information being transferred) between a good Sender and the Receiver than between a bad Sender and the Receiver. This is corroborated by our results (Figure~\ref{fig:mi}).
The information transmitted was significantly greater than the MI for chance performance for both the good Senders ($MI$ = 0.336, $t$-test on angular transformed data: $t$(4) = 5.374, $p$ =.006; Wilcoxon test: $V$ = 15, $p$ = 0.031) and the bad Senders ($MI$ = $0.051$; $t$-test on angular transformed data: $t$(4) = 3.544, $p$ =.024; Wilcoxon test: $V$ = 15, $p$ = 0.031). The difference between good and bad Senders was also statistically significant (two-sided $t$-test on angular transformed data, $t$(8) = 5.187, $p$ = .002; Wilcoxon test: $W$ = 0, $p$ = 0.031), with the good Senders transmitting, on average, more information than the bad Senders.
For consistency with previous studies \cite{RaoStocco:2014,Stocco:2015,Rao:2014}, we have reported {\it uncorrected} estimates of MI. Given the relatively small number of samples, uncorrected MI values might overestimate the true amount of information shared by two participants. For this reason, we used a recently proposed method \cite{Panzeri:2017} to calculate the amount of bias in our estimates. Under the conditions of our experiment, the bias $b$ can be approximated as $b = -N_R/[2 \times N_S \times \log(2)]$, with $N_R$ being the number of possible responses (in our case, $N_R=2)$ and $N_S$ the number of samples (in our case, $N_S = 32$ for each pair of participants). The bias thus estimated was found to be negligible ($b = -0.045$) and does not affect the results of any of our statistical tests.
\subsection*{Learning of Sender Reliability by Receiver}
The differences in accuracy and mutual information between the ``good'' and ``bad'' Senders in the previous section suggest that the Receiver successfully learned which of the two Senders is a more reliable source of information. Confirming that this is indeed the case would bring BrainNet a step closer to conventional social networks where users utilize differential weighting for different sources of information. To further investigate this issue, we divided each experimental session into four consecutive blocks of four trials each. We quantified the time course of the Receiver's learning process using two measures: (1) block-by-block estimates of the linear regression weights for the Receiver's decisions versus each Sender's decisions; and (2) the block-by-block correlation of decisions made by the Receiver and by each Sender \cite{Rao:2014}. Because of the small number of trials ($N=4$) in each block, the decision vectors for Senders and Receivers were created by concatenating the decisions of participants with the same role (Receiver, good Sender, or bad Sender) across the five triads; this procedure captures group-level behavior and is less sensitive to outliers. Thus, if $R\in\mathbb{R}^{20\times 1}$ is a decision vector for the five Receivers in a four-trial block (each decision is encoded as a 0 or 1), and $S\in\mathbb{R}^{20 \times 1}$ is a decision vector for one type of Sender ("good" or "bad"), the linear regression weights $ \beta$ can be estimated using the standard pseudoinverse method\cite{Kenney:1947} as: $ \beta = (S^TS)^{-1}S^TR$. The same concatenated vectors $R$ and $S$ were also used to estimate the Pearson correlation coefficients for each four-trial block.
As shown in Figure~\ref{fig:learning}, the time course of both the beta weights and correlation coefficients show a steep ascending trend for the good Sender, but not for the bad Sender. To test the difference between these trends, we estimated two simple linear trend models of the relationship between each measure and the block number, one for the good Sender and one for the bad Sender. The difference between the linear trend model's slope coefficients $\beta_g$ and $\beta_b$ for the good and bad Senders respectively was then tested for statistical significance using the formula derived by Paternoster and colleagues \cite{Paternoster:1998}:
$$ Z = \frac{\beta_g - \beta_b}{\sqrt{SE\beta_g^2 + SE\beta_b^2 }} $$
where $SE\beta_g^2$ and $SE\beta_b^2$ are the variances of $\beta_g$ and $\beta_b$, respectively. The difference in linear trends was statistically significant for both measures (beta weight measure: Z = 5.87, $p$ < 0.001; correlation coefficient measure: Z = 7.31, $p$ < 0.001). These results strongly suggest that Receivers were able to learn which Sender was more reliable during the course of their brain-to-brain interactions with the two Senders.
\section*{Discussion}
This paper presents, to our knowledge, the first successful demonstration of multi-person non-invasive direct brain-to-brain interactions for collaboratively solving a task. We believe our brain-to-brain interface, which we call BrainNet, improves upon previous human brain-to-brain interfaces (BBIs) on three fronts: (1) BrainNet expands the scale of BBIs to multiple human subjects working collaboratively to solve a task. (2) BrainNet is the first BBI to combine brain recording (EEG) and brain stimulation (TMS) in a single human subject, eliminating the need to use any physical movements to convey information (although we did not explicitly instruct subjects to avoid eye movements when using the SSVEP interface, other researchers have shown that an SSVEP BCI can be operated without eye movements \cite{Kelly2005,Allison2008}). With sufficient hardware, our system can be scaled to the case where every subject can both send and receive information using the brain interface. (3) Using only the information delivered by BrainNet, Receivers are able to learn the reliability of information conveyed to their brains by other subjects and choose the more reliable sender. This makes the information exchange mediated by BrainNet similar to real-life social communication, bringing us a step closer to a ``social network of brains.''
Our results on combining information from multiple users builds on previous work in the field of brain-computer interfaces (BCIs) linking the individual contributions of more than two brains to control a computer. In humans, researchers have studied “collaborative BCIs” (rather than BBIs) that pool information from multiple human brains to improve performance in a delayed saccade-or-reach task\cite{Wang:2011}; however, subjects performed the task on different days and no brain stimulation was used to convey information directly to subjects' brains. A different study\cite{Ramakrishnan2015} demonstrated that three non-human primates can jointly control a 3D virtual avatar arm using brain signals recorded with invasive electrodes implanted in the motor cortex; again, the goal was distributing a single task across multiple individuals linked to a common BCI without encoding any neural information for feedback and interaction via stimulation. More closely related to our study is the work of Pais-Vieira {\em et al.}\cite{Pais-Vieira:2015}, who used implanted electrodes to both decode information from and transmit information to the somatosensory cortices of multiple rodents to demonstrate the possibility of distributing computations across multiple brains. The brains of the rats were linked to solve several computational problems including a weather forecasting task based on weather data from a local airport. However, the animals were entirely unaware of both the actual task being solved and of their collaboration with others; by contrast, in BrainNet, the participants are completely aware of the task and are conscious of being collaborators within a ``network of brains''.
BrainNet could be improved in several ways: (1) From the first human BBI\cite{Rao:2014} to BrainNet, the level of information complexity has remained binary, i.e., only a bit of information is transmitted during each iteration of communication. Additionally, this low bit rate required a disproportionate amount of technical hardware and setup. To address the limitation of low bit rate, we are currently exploring the use of functional Magnetic Resonance Imaging (fMRI) to increase the bandwidth of human BBIs. Other approaches worth exploring include combining EEG and fMRI to achieve both high spatial and temporal resolution\cite{Huster:2012} for decoding, and using TMS to stimulate higher-order cortical areas to deliver more complex information such as semantic concepts. (2) We purposefully introduced a “bad” sender in BrainNet design to study whether the Receiver can learn which Sender is more reliable. It would be interesting to investigate whether the Receiver can learn the reliability of Senders in more natural scenarios where the unreliability originates from the noisy nature of a Sender's brain recordings or from a Sender's lack of knowledge, diminished attention, or even malicious intent. (3) From an implementation standpoint, BrainNet uses a typical server-client TCP protocol to transmit information between computers. However, the server is solely designed for BrainNet’s experimental task and is not a general purpose server. A cloud-based BBI server could direct information transmission between any set of devices on the BBI network and make it globally operable through the Internet, thereby allowing cloud-based interactions between brains on a global scale. Such BBIs, when developed within an ethically-grounded framework, have the potential to not only open new frontiers in human communication and collaboration but also provide a new scientific tool to explore questions in neuroscience and gain a deeper understanding of the human brain.
\section*{Methods}
\subsection*{Participants}
Fifteen healthy participants (aged 18–35 yrs, average 22.7 yrs, eight female), took part in a controlled laboratory experiment. All participants were recruited through word of mouth, were fully informed about the experimental procedure and its potential risks and benefits, and gave written consent prior to the beginning of the experiment according to the guidelines put forth by the University of Washington. Both the experimental and the recruitment procedures were reviewed and approved by the Institutional Review Board of the University of Washington (IRB Application \#52392). The participants were divided into five groups, with each group being a triad of one participant playing the role of the “Receiver” and two playing the roles of “Senders.” To maintain their decision to participate free of any external influence, all participants received monetary compensation that was independent of their role and proportional to the total amount of time devoted to the study.
\subsection*{Experimental Task}
During each session, a triad of three participants collaborated to play a simplified Tetris-like game. The game consisted of independent trials, each of which involved deciding whether or not to rotate a single block of particular shape by 180 degrees. At the bottom of the screen, there was a partially filled line whose gaps could be filled by either the top or bottom part of the block at the top of screen. The goal of the game was to achieve the highest possible score by making the correct decision to rotate or not rotate the current block so that when dropped at the end of the trial, it would fill the missing parts of the line at the bottom. We designed the task such that the actual player of the game, namely the Receiver, could only see the block at the top of screen and not the bottom line. The other two subjects, namely the Senders, could see both the block at top and the line at bottom (see Figure~\ref{fig:two-step}). Thus, the only way for the Receiver to achieve a high score was by integrating the decisions transmitted by both Senders and make his/her own decision for the game.
Each session was made of sixteen independent trials; in half of them the falling block had to be rotated and in the other half, it had to be left in the original orientation. The order of rotation and non-rotation trials was randomized, with the constraint that each half of the session had to contain 4 rotation and 4 non-rotation trials.
Each trial comprised of two rounds of interactions between the Senders and the Receiver. Each round offered a chance to rotate the block. After the first round, the block was rotated or remained in the same orientation based on Receiver’s decision. The block then dropped halfway and the screens shown to all three subjects were updated to show the (possibly rotated) block at the halfway location (see Figure~\ref{fig:two-step}). Note that one decision is sufficient to complete the task of filling the bottom line but because of our two-step design, the Senders receive feedback on the Receiver's action in the first round and can send the Receiver new suggestions, allowing the Receiver to potentially correct a mistake made in the first round and still successfully complete a trial.
The three participants in a triad were located in different rooms in the same building on the University of Washington campus and could only communicate with each other through the brain-to-brain interface.
\subsection*{BrainNet: Multi-Person Brain-to-Brain Interface}
Figure~\ref{fig:brainnet} depicts the architecture of BrainNet. BrainNet relies on two well-known technologies: Electroencephalography (EEG)\cite{Nunez:2005} for non-invasively recording brain signals from the scalp and transcranial magnetic stimulation (TMS)\cite{OShea:2007} for non-invasively stimulating the visual cortex. The Senders convey their decisions of "rotate" or "do not rotate" by controlling a horizontally moving cursor (Figure~\ref{fig:ssvep-screen}) using steady-state visually-evoked potentials (SSVEPs) \cite{Vialatte:2010}: to convey a ``Rotate'' decision, Senders focused their attention on a ``Yes'' LED light flashing at 17 Hz placed on the left side of their computer screen; to convey a "Do Not Rotate" decision, they focused on the “No” LED light flashing at 15 Hz placed on the right side. These LEDs are depicted as circles attached to the screens in Figure~\ref{fig:brainnet}. The cursor position provided real-time visual feedback to the Senders. The direction of movement of the cursor was determined by comparing the EEG power at 17 Hz versus 15 Hz, with a higher power at 17 Hz over that at 15 Hz moving the cursor towards the left side near the “Yes” LED, and vice-versa for the "No" LED. A ``Rotate'' (``Do Not Rotate'') decision was made when the cursor hit the side of the screen appropriately marked "YES" ("NO") (see Figure~\ref{fig:ssvep-screen}). In trials where the cursor did not reach either side of the screen due to trial time elapsing, the decision closest to the last location of the cursor was chosen as the subject's decision.
The decisions of the two Senders were sent to the Receiver’s computer through a TCP/IP network and were further translated into two pulses of transcranial magentic stimulation (TMS) delivered sequentially to the occipital cortex of the Receiver. Each TMS pulse lasted 1 ms. An eight-second delay was enforced between the two pulses to remain within the strictest safety guidelines of TMS stimulation\cite{Rossi:2009}. The intensity of the stimulation was set above or below the threshold at which the Receiver could perceive a flash of light known as a phosphene: a ``Yes'' response was translated to an intensity above the threshold, and “No” was translated to an intensity below the threshold. During each round of trials, the Receiver always received the decision from one Sender first, then the other. The screen the Receiver saw also had visual prompts to remind them whose decision the current TMS stimulation was conveying. Receivers made their decision based on whether a phosphene was perceived and conveyed their decision (rotate or do not rotate) to the game using the same SSVEP-based procedure used by both Senders. After the game state was updated, the trial moved into the second round and the above process was repeated. At the end of each trial, all three subjects received feedback on the result of the trial (Figure~\ref{fig:two-step}, bottom row).
\subsection*{Differential Reliability of Senders}
When the decisions from the two Senders do not agree with each other, the Receiver must decide which Sender to trust. To investigate whether the Receiver can learn the reliability of each Sender and choose the more reliable Sender for making decisions, we designed the system to deliberately make one of the Senders less accurate than the other. Specifically, for each session, one Sender was randomly chosen as the ``Bad'' Sender and, in 10 out of sixteen trials, this Sender's decision when sent to the Receiver was forced to be incorrect, both in the first and second round of each trial.
\subsection*{EEG Procedure for Senders}
Each Sender performed the task in a dedicated room in front of a 21" LCD monitor, with two Arduino-controlled LED lights attached to the left and right outer frames of the monitor for eliciting SSVEPs. EEG signals were recorded through an 8-channel OpenBCI Cyton system (OpenBCI: Brooklyn, NY) with a sampling rate of 250Hz at a resolution of 16 bits. Signals were acquired from gold-plated electrodes and a layer of electro-conductive paste was applied between each electrode and the participant’s scalp. For the experimental session, three electrodes were set up along the midline in a custom montage with the signal recorded from one occipital electrode (location Oz in the 10–10 placement system) and two frontal electrodes (locations AFz and FCz in the 10–10 system) used as the ground and reference, respectively.
Incoming EEG data was passed through a 4th-order Butterworth filter \cite{Sorrentino:2007} between 0 and 30 Hz to remove signal drifting and line noise. The time-series EEG data was then divided into 1-second epochs and transformed to the frequency domain using Welch's method\cite{Welch:1967}. The intention to rotate the falling block or not was decoded by comparing the power at 17 Hz and 15 Hz obtained from Welch's method. The final decision was made by tallying up the number of epochs in which the greatest power was recorded at either 17 Hz or 15 Hz over a 10-second period. Signal processing and data storage were managed through a custom software library developed by two of the authors (LJ and DL).
There is no prior training required for controlling the cursor using SSVEP. During the experiment, the Sender's monitor displays either the cursor-control interface or a gray background with a text prompt indicating that the Receiver is making a decision.
\subsection*{TMS Procedure for the Receiver}
Participants playing the role of the Receiver came in for two consecutive sessions. During the first session, as part of informed consent, they were asked to complete a TMS safety screening questionnaire, aimed at identifying potential conditions (such as family history of seizures or frequent migraines) that might represent potential risk factors for adverse side effects of TMS.
No participant was rejected for failing the safety questionnaire. In addition to the safety screening, all Receivers underwent a procedure to determine their absolute phosphene threshold, that is, the minimum amount of stimulation necessary to elicit the perception of an induced phosphene 50\% of the time. The absolute threshold was assessed using the PEST method\cite{Taylor:1967}. The absolute threshold was then used as the starting point to identify the stimulation levels associated with the binary ``Rotate'' and ``Do Not Rotate'' decisions. Starting from the absolute threshold, the stimulation intensity was first adjusted upwards in increments of 5\% until phosphenes could be elicited for 10 consecutive pulses; this value was then used for conveying a ``Rotate'' decision from a Sender. Then, starting from the absolute threshold value, the stimulation intensity was lowered in 5\% increments until no phosphene was elicited for 10 consecutive pulses. This value was then used to convey a ``Do Not Rotate'' decision from a Sender. During both the testing session and the experimental session, TMS was delivered through a 70-mm Figure-8 Alpha coil (Magstim, UK) positioned over the left occipital lobe in a location corresponding to site O1 in the 10-20 system. The coil was positioned flushed to the head, with the handle parallel to the ground and extending towards the left. The coil was attached to a SuperRapid2 magnetic stimulator (Magstim, UK).
The maximum intensity of the electric field for our TMS equipment is 530 V/m, and with our coil, the maximum intensity of the induced magnetic field is 2.0 T.
\subsection*{EEG Procedure for the Receiver}
The EEG procedure for the Receiver was identical to that used for the Senders, except that the signal was acquired from a BrainAmp system (BrainVision, Berlin, Germany) with a sampling rate of 5000 Hz and a resolution of 20 bits. The system was equipped with a DC voltage amplifier to reduce signal distortions due to the TMS pulses. Participants wore a standard 32-channel headcap, using AFz and FCz as the ground and reference, respectively. As in the case of the Senders, only data from the Oz channel was recorded. After downsampling to 500Hz, the incoming data underwent the same preprocessing steps described above for the Senders.
|
\section{Experimental Evaluation}
In the following, we study the performance of framing the multimedia indexing as a modified gist understanding task conducted on the MIR Flickr dataset, which contains around 25,000 images from Flickr with textual tags. We report the results according to the evaluation measures Mean Average Precision (MAP) and Precision of the first 50 positions of our ranking output (P@50).
Given an image-tags pair the task is to assign at least one label of the 10 general- and 19 subtopics, e.g., sky, clouds, class topics to such a query pair. Some of the topics have less image instances than the other, resulting in an instance per topic range between 116 to 10,373 for the topics baby (relevant) and people (potential), respectively.
\medskip \noindent \textbf{Multimedia Classification (Multimedia Indexing)}
We provide a comparison to state-of-the-art approaches of the multimedia indexing task compared to the results reported by~\cite{Huiskes10,Chen2016,DBM2012}, relying on their results without re-implementation.
\begin{table*}
\caption[MIR Flickr: Classification performance across all topics]{Classification performance of the different gist detection pipeline settings across all topics according to \textbf{MAP} and \textbf{P@50}, compared to the two best methods from~\cite{Huiskes10} (the two low-level settings are discarded here. Original numbers from~\cite{Huiskes10}), the random baseline, the Deep Boltzman Machine (DBM) (numbers and approach of~\cite{DBM2012}), and the 32-bit robust multi-label hashing (RMLH) of~\cite{Chen2016}.
\label{tab:miniComp}}
\begin{scriptsize}
\begin{center}
\begin{tabular}{lllllllllllllll}
\toprule
\textbf{Method} & Gist($S_{t, i},I$)& Gist($S_{t, i}$) & Gist($S_{t}$) & &LDA & SVM & Random & RMLH, 32-bit & DBM\\
\midrule
\textbf{MAP} & 61.42 & 45.56 & 35.87 & &49.2 & 47.5 & 12.4 & 65.6 & 52.6\\
\textbf{P@50} & 88.3 & 53.71 & 49.18 & &74.5& 75.8 & 12.4 & - & 79.1\\
\bottomrule
\end{tabular}
\end{center}
\end{scriptsize}
\end{table*}
The extended gist detection pipeline outperforms all of the comparison approaches in terms of precision (88.3, cf. Table~\ref{tab:miniComp}).
Additionally, all but one approach are outperformed according to Mean Average Precision (MAP: 61.42) by the extended gist detection pipeline (cf. Table. Only, the multi-label hashing method performs better (MAP: 65.6). However, there is no result reported for this approach with respect to precision.
\medskip \noindent \textbf{Comparison of concept candidates}
Finally, we compare the performance of the extended gist pipeline, when considering three different sets of concepts in the gist ranking task: concepts from the seed nodes of the tags (Gist($S_{t}$)), concepts from the seed nodes of the tags and the images (Gist($S_{t, i}$)), and concepts from the seed and the intermediate nodes (Gist($S_{t, i},I$).
The lexicon of Gist($S_{t, i},I$) consists of 1,538 different top-10 concepts. This is an average of 40.5 concepts of representing one class.
Gist($S_{t, i}$) and Gist($S_{t}$) contain 5,809 and 11,755 concepts, respectively.
The higher number of concepts for Gist($S_{t}$) underlines the diversity across the concepts using the tags only (cf. Appendix - Seed Node Linking) - which might not be the best characteristic to find semantic similarity. That Gist($S_{t, i},I$) has the smallest number of different concepts, confirms the idea of finding common concepts with the modified gist understanding pipeline.
The comparison shows that the extended candidate set performs best in terms of MAP and P@50 (MAP: 61.42 and P@50: 88.3, cf. Table~\ref{tab:miniComp}). Comparing the MAP results (MAP: 45.56 vs. 35.87) of the two simplified pipelines Gist($S_{t, i}$) and Gist($S_{t}$), indicate the benefit of including information from the image even though it might contain false positive detections. This observation is also confirmed by the result of P@50: 53.71 vs. 49.18).
This study demonstrates that all pipeline steps and the automatic object detection, which allows for an end-to-end approach, increase the performance in the multimedia indexing task. Finally, it confirms the benefit of understanding the gist of multimedia pairs, such as images and texts, in established research domains.
\section{Conclusion}
To lower the barrier of fostering novel research tasks, such as the one of gist detection and understanding, far beyond the literal meaning of images, we have shown that gist detection performs well on established research questions. We benefit from a pipeline that finds semantic diversity, while also addressing semantic relatedness. Both aspects address the characteristic of the multimedia indexing task: instances can be assigned to several classes, as they convey different semantic meanings, however, at the same time, they share semantics with instances assigned to the same class.
In nearly 40\% of the general topic and 84\% of the sub-topic cases, the extended gist detection pipeline outperforms the comparison approaches, which implies robustness of our approach.
Comparing the overall performance, the gist detection pipeline outperforms the Deep Boltzman Machine approach and is comparable to the robust multi-label hashing approach, where for the latter one no result is given for precision.
In an end-to-end approach we have demonstrated that multimodal modelling - even if one modality comes from noisy object detectors - performs better than a single modality approach.
\section*{Appendix A}
\subsection{Seed Node Linking}
\begin{table*}[h!]
\begin{center}
\caption[Number of candidates and seed nodes after concept linking (Step 1)]{Number of candidates and seed nodes after the concept linking according to Flickr Tags (text) or image as their origin.
\label{tab:entityMini}}
\begin{scriptsize}
\begin{tabular}{cccccccc}
\toprule
& \multicolumn{2}{c}{Total} && \multicolumn{2}{c}{Unique} & & Empty Instances\\
\cmidrule{2-3} \cmidrule{5-6}
& Candidates & Seed Nodes && Candidates & Seed Nodes & & \\
\midrule
Tags & 223,537 & 157,473 && 74,427 & 34,127 & & 2,128 \\
Images & 514,819 & 487,724 && 976 & 964 & & 166 \\
\bottomrule
\end{tabular}
\end{scriptsize}
\end{center}
\end{table*}
There are no gold standard entity links provided in the dataset and it is not feasible to annotate the entity links for all of the 25,000 pairs. Consequently, we provide a statistical overview on how many of the concept candidate mentions are actually linked to a concept in the knowledge base.
In Table~\ref{tab:entityMini} the total number of candidates and the finally total number of linked entities are given, which we refer to as seed nodes.
We provide the numbers separately according to the texts and the images. Furthermore, we study how many unique candidates and seed nodes are found across the dataset.
The sum of all tags for all image instances (25,000) of the dataset, is over 223,000 tags (cf. Table~\ref{tab:entityMini}), which results in an average of 9 tags per image.
However, around 2,000 images do not have a tag at all (cf. Table~\ref{tab:entityMini}, Empty Instances).
For the images around 500,000 image objects are detected and annotated by the Computer Vision API, which is an average of 20 visually recognizable objects per image.
On 166 images no visually recognizable object is detected, whereas 2 images could not be processed by the API due to an extreme format (e.g., 500px width, 49px height).
For both types of media - image and text - around half of the candidate concept mentions can be linked to actual concepts in the knowledge base. This results in 157,473 and 487,724 seed nodes for the texts and the images, respectively.
The statistics about uniqueness across the dataset reveals that the tags are more diverse than the visually recognizable objects: Even though the total number of visually recognizable objects are twice the number of tags, the image objects are from 976 semantic concepts, whereas the tags are from 74,427 semantic concepts.
\section{Introduction}\label{UseCase:Intro}
Conveying meaning by the use of different modalities is probably as old as human mankind. Nowadays, where everyone takes pictures constantly, storage is cheap, and several platforms allow for the distribution, sharing, and modification of multimodal data, the need of adequately representing the content of such multimodal data is more important than ever.
Research on multimedia data, independent of the goal - i.e., whether it is modality generation, modality retrieval, or representation of the modalities - have shown
to perform better with a joint representation of the different modalities.
Thus, multimedia researchers motivate the joint modeling by observing ("people often caption an image to say things that may not be obvious from the image itself, such as the name of the person, place, or a particular object in the picture."~\cite[p. 2950]{DBM2012}).
In this work, we focus on multimedia representation. Thus, we make use of the MIR Flickr dataset~\cite{Huiskes2008}, which provides - besides the images and their affiliated tags - a gold standard multi-label annotation, where the labels represent semantic class topics, such as 'sky' or 'people'.
A lot of research has been conducted providing a detailed analysis on the MIR Flickr dataset, these works can be mainly grouped into deep learning, auto-encoder, hashing-based and other approaches.
To the best of our knowledge, none of the previous works have made benefit from knowledge bases, such as Wikipedia.
\medskip \noindent \textbf{Task:} Given an image and its tags, represented by concepts from a knowledge base, create a mapping between the representing concepts and the target classes.
\medskip \noindent To represent images, we adapt, modify, and expand the work of~\cite{weiland16a}. Their task is to understand the \textit{gist} (equivalent to meaning or message) of an image-caption pair. Their idea is to create query-specific knowledge graphs, where both, an image-caption pair and the gist are represented by concepts from the knowledgebase. There are two modifications to their methodology: First, we need to create a mapping between the ranked list of concepts and the target class topic(s) to conduct the multimedia indexing. Second, we encode a graph creation strategy, which has shown to be useful for text-lexical understanding~\cite{NavigliLapata:10,ponzetto10b}. This modified gist pipeline is an end-to-end setting, in that the method relies on out-of-the-box object detectors and a model that has been pre-trained on a dataset with the purpose of gist understanding.
In an experimental evaluation, we study the per class and the overall performance of the modified gist pipeline. We compare these results with shallow, deep, hashing, and auto-encoder approaches. We find that not only representing the pair as concepts, but the expansion steps to collect semantically related concepts, has a positive impact to the overall performance. Overall, we confirm that the combination of tags and a deep learning based automatic object detection - even though the latter encodes noise - achieves better performance than single modality representations.
\section{Methodology}\label{sec:method}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=\textwidth, height=100.0pt]{pipeline.png}
\caption[Modified gist pipeline]{The twice-learning-to-rank pipeline for the multimedia indexing task. Image-text pairs are represented as concepts from a knowledge base (Image \#24706: Flickr/clare savory, CC BY 2.0).}
\label{fig:Modpipeline}
\end{center}
\end{figure*}
Since we base our method on the pipeline introduced by~\cite{weiland16a}, we focus on our novel changes (cf. Fig.~\ref{fig:Modpipeline}).
However, we start with an overview of the complete pipeline and indicate - where required - the modifications compared to the original pipeline.
The main idea of the approach is to represent the images and their affiliated text by concepts from a knowledge base and conduct a query-specific knowledge graph generation, which in the case of~\cite{weiland16a} represents the gist and which in our case is used to create a mapping between query image and target class.
In the knowledge base representation of Wikipedia we are using, article pages and categories are the concepts.
Besides the representation of Wikipedia as a knowledge base, we benefit from a representation as a knowledge graph, where categories and articles are nodes in the graph. The article redirects and the category links are the edges in the graph.
\medskip \noindent \textbf{Seed Node Linking}
As we aim at an end-to-end setting, we benefit from a state-of-the-art image processing API, such as Microsoft Cognitive Services~\footnote{Note: By End of 2017 the API is integrated into Microsoft Azure.}, to assign textual labels to objects, attributes, and the scenery in the images, e.g., car, yellow, outdoor, respectively.
We use the textual object labels as candidates for concept mentions.
Additionally, the Flickr tags provided in the MIRflickr25k benchmarking dataset, serve as candidates for the concept linking.
Via a string-match based concept linking, the concept mentions from both, the image and the text, are linked to concepts of a knowledge base.
The resulting set of concepts is what is called the seed nodes, consisting of nodes with origin in the image or the textual Flickr tags, in the following we refer to these concepts as \textbf{$S_{i}$} and \textbf{$S_{t}$}, or to their combination as \textbf{$S_{t, i}$}, respectively.
\medskip \noindent \textbf{Intermediate Graph Creation.}\label{sec:method_inter} Under the assumption that the image and the text convey some specific - sometimes complementary - aspect of a common meaning, the next step is to search for semantic connections between the image and its text. Given the seed nodes from the previous step, we collect all category nodes on shortest paths (up to length 4) between all pairs of seed nodes. Therefore edges representing category links are followed.
The result is a sub-graph, which consists of seed nodes, the collected category nodes (called intermediate nodes) and the edges connecting the seed and intermediate nodes. In the following we refer to this procedure as the intermediate graph creation and to concepts collected in this step to as \textbf{$I$}.
\textbf{Example:} One query image has the seed nodes volvo and car, where a node on the intermediate path is Category:Motor vehicles. Another query image has the seed nodes motorcycle and traffic, which are also connected to Category:Motor vehicles. Consequently, the instances which initially did not share a concept, now share a concept.
By applying Louvain clustering~\cite{blondel2008fast} based on the path based semantic relatedness measure of~\cite{Hulpus2015pathbased}, we address the fact that a query image is of diverse nature, in that it can be assigned to several class topics, consequently, the seed concepts are not about one semantic topic and thus, of different semantic relevance.
\medskip \noindent \textbf{Ranking the Nodes.} To conduct the ranking of concepts best representing the image-text pair, we use 16 different features, e.g., graph connectivity and content based measures, boolean and features based on the article texts of Wikipedia concepts (for further details, please refer to~\cite{weiland16a}). As there is no gold standard about the relevance of the so far collected concepts given (recall: a concept in the knowledge base has an equivalent node in the knowledge graph, but as we are not doing a graph ranking, we switch back the notation of concept), we rely on a pre-trained model using the data and gold standard of~\cite{weiland16a}to rank the concepts.
\medskip \noindent \textbf{Learning-to-rank model (l2r).} Our generated feature vectors serve as input for a list-wise learning-to-rank model~\cite{li2011letor}. In a learning-to-rank setting, the image-text pairs are the set of documents $D$, with $d_i$ as the i-th document (following the notation of ~\cite{li2011letor}).
For each instance the candidate concepts are the set of queries, denoted by $Q$, where $q_i$ is the i-th query.
The j-th image-text pair for a query $q_i$ is represented as a feature vector $x_{i,j} = \phi(q_i, d_{i,j})$ of feature functions $\phi$, such as Betweeness Centrality.
Finally, $S'={(x_i, y_i)}$ represents the training data for $q_i$, with $y$ denoting the set of labels $\{1, 2,..., 5\}$.
The objective is to find a parameter setting $\widetilde{\phi}$ that maximizes the scoring function:
$\widetilde{\pi} = \argmax_{\pi_i \in \Pi_i} S(x_i, \pi_i)$.
Specifically, we use RankLib\footnote{https://sourceforge.net/p/lemur/wiki/RankLib}, trained with respect to the target metric Mean-average precision (MAP). For optimization we use coordinate ascent with a linear kernel~\cite{Metzler2007}.
The result is a ranked list of concepts for each image-text pair.
\medskip \noindent \textbf{Image per Topic Ranking.} As the objective is a multi-label classification, the ranked list of concepts need to be mapped to the class topics.
There are several ways of creating such a mapping, such as using lexical and conceptual hierarchies (e.g., WordNet).
The issue with these methods is that each single gist concept needs to be matched to the MIR topics, but tags and images in their single instances are not completely semantically coherent with the topic of a class, e.g., a building will not be matched to the class topic people. However, the combination of the concepts make the meaning, e.g., crowded city street.
Hence, we benefit from an additional learning-to-rank approach, with a feature vector representing the ranked gist concepts of an input image-tag pair with the relevance scores from the first l2r serve as feature for the respective concept.
Across all ranked concepts we build a lexicon, where the size of the lexicon is the number of dimensions for the feature vectors, which will be created.
The danger here is that the dimension is higher than the number of training instances, however, we find that the query-specific graph creation leads to a small lexicon.
Thus, finally we formulate: Given the class as a query, the task is to rank images that fit the class highest.
\section{Related Work}\label{sec:rw}
Representing multimedia data is often approached as a classification task, where the goal is to assign the multimedia instances to one or several classes best representing the content of the instance.
In context of multimedia data, which consists of image and text, i.e., the best representing class(es) is given as the class that best describes the most salient object(s) in the image~\cite{Huiskes2008}.
The originators of the MIR Flickr benchmarking dataset presented two different approaches for classification of the data. An SVM and Linear Discriminant Analysis (LDA) based classification.
Both approaches are trained with different feature sets: one using uni-modal features (low-level features from the vision domain) and one combining the low level visual features with the textual tags as features~\cite{Huiskes10}.
There are several successors using various types of approaches, however, none benefit from query-specific knowledge graphs. ~\cite{Hare2010} presents a high-dimensional vector space representations based on a cross-lingual latent semantic indexing.
Others make use of deep-learning approaches, e.g., multimodal deep boltzman machine (DBM)~\cite{DBM2012}, regularized deep neural network (RE-DNN)~\cite{Wang2016}, deep hashing~\cite{Wu2017}.
|
\section{Introduction}
Grasping and transporting objects is a canonical robotics problem which has seen great advancements in recent years, especially with regards to detection of grasp poses for previously unseen objects using only visual information. As the performance of these visual grasp detection systems has improved, so has the difficulty of standard benchmarking tasks for evaluation, seeing a shift away from grasping relatively simple, isolated objects to grasping geometrically and visually challenging objects in cluttered environments~\cite{mahler2017binpicking}.
The use of cluttered environments and complex objects has resulted in more visually challenging scenarios for grasp detection, with the level of clutter impacting grasp detection performance~\cite{mahler2017binpicking}. Recent work has shown that
improved visual information from point cloud fusion~\cite{ten2017grasp, arruda2016active} or
viewpoint selection~\cite{gualtieri2017viewpoint} can improve the quality of grasp estimates in clutter. However, these typically are treated as separate, fixed data gathering step
and do not directly take into account grasp detections from multiple viewpoints.
At the same time, improvements in grasp detection models and computational hardware have seen the time required to visually detect grasps (especially using deep learning methods) decrease from tens of seconds~\cite{lenz2015deep} to less than a second~\cite{mahler2017dex, viereck2017learning} to small fractions of a second~\cite{morrison2018closing}. Consequently, the largest contributor to grasp execution time has become the movement of the robotic arm executing the grasp, making grasp detection from multiple viewpoints feasible with very little impact to overall execution time.
To address the added difficulties of grasping in clutter, we propose to use the act of reaching towards a grasp as a method of data collection, making it a meaningful part of the grasping pipeline, rather than just a mechanical necessity.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{hero.eps}
\vspace{-1mm}
\caption{Our Multi-View Picking (MVP) controller considers multiple informative viewpoints for an eye-in-hand camera while reaching for a grasp in clutter to reduce uncertainty in the grasp pose estimate caused by clutter and occlusions. This allows it to perform high quality grasps that may not be clearly visible from the initial viewpoint.}
\label{fig:hero}
\vspace{-4mm}
\end{figure}
To achieve this, we develop the Multi-View Picking (MVP) controller, which selects multiple informative viewpoints for an eye-in-hand camera while reaching to a grasp in order to reduce uncertainty in the grasp pose estimate caused by clutter and occlusion, resulting in an overall improvement in grasp success (\fig{hero}). Unlike previous works in active perception for grasping which employ object-specific heuristics~\cite{gualtieri2017viewpoint} or a secondary task such as point cloud reconstruction~\cite{arruda2016active, kahn2015active, ten2017grasp}, our approach directly uses entropy in the grasp pose estimation to influence control.
We validate our approach through trials of grasping from piles of 20 objects in clutter, and compare our results to baselines which represent common visual grasp detection approaches. Using our MVP controller, we achieve an 80\% success rate in grasping from clutter, a 12\% increase compared to a single-viewpoint grasp detection approach.
Additionally, our method outperforms a baseline which considers multiple viewpoints over a predefined trajectory, achieving a higher grasp success rate with fewer distinct viewpoints and reduced mean time per grasp. This highlights the advantages of using our information-gain based approach which focuses on salient areas, reducing unnecessary data collection. By varying the cost associated with data collection during reaching, we show that it is possible to trade off between success rate and execution time, allowing the system to be optimised for either raw grasp success rate or overall efficiency.
\section{Related Work}
In order to grasp and transport a wide range of unknown objects, a robotic system cannot rely on using offline information such as object models or object-specific grasp poses. Instead, it must use geometric information to compute stable grasp poses for previously unseen objects. Recently, many approaches combining visual inputs with machine-learning techniques -- primarily Convolutional Neural Networks (CNNs) -- have been widely and successfully applied to this problem~\cite{lenz2015deep, morrison2018closing, mahler2017dex, mahler2017binpicking, pinto2016supersizing, ten2017grasp, johns2016deep}, which we refer to as visual grasp detection.
The robustness of many visual grasp detection systems is sensitive to factors such as sensor noise, robot control inaccuracies and visual occlusion, which is especially prevalent in cluttered environments. While the detrimental effects of sensor noise and poor robot precision can be minimised as part of the visual grasp detection algorithm, e.g. through sensor noise injection during training~\cite{johns2016deep, mahler2017dex, mahler2017binpicking, viereck2017learning}, overcoming the issue of occlusion in clutter requires multiple camera viewpoints to be considered. For example, ten \citet{ten2017grasp} showed that computing grasp poses using a fused point cloud from many viewpoints along a predefined trajectory resulted in a 9\% increase in grasp success rate compared to using a point cloud collected from two static cameras.
Rather than rely on a fixed data collection routine, \citet{arruda2016active} use an active perception approach to choose viewpoints which specifically aid point cloud reconstruction near potential finger contact points in an efficient manner.
Broadly, active perception is defined as the situation where a robot ``adopts strategies for decisions of sensor placement or sensor configuration" in order to perform a task~\cite{chen2011active}. It is a concept which has been applied to a wide variety of robotic tasks, such as mapping~\cite{stachniss2005information, burgard2005coordinated}, object modelling~\cite{pito1999solution}, object identification~\cite{roy2004active, monica2018contour} and path planning~\cite{velez2011planning}. Common strategies for active perception focus on planning the expected next best action to efficiently minimise measurement uncertainty or maximise information gain via a metric such as Shannon entropy~\cite{vazquez2001viewpoint} or KL divergence~\cite{van2012maximally}.
Active perception approaches have been applied to robotic grasping in prior work, but rather than directly use grasp quality as a metric rely on a secondary task such as object modelling~\cite{bone2008automated, aleotti2014perception}, point cloud reconstruction while searching for graspable geometry~\cite{arruda2016active, kahn2015active} or localising previously seen objects~\cite{holz2014active} to generate next best view commands. This is partly due to the difficulty in defining a probability distribution over predicted grasp poses~\cite{gualtieri2017viewpoint}, which we overcome by using a visual grasp detection system which generates a pixelwise distribution of grasp pose estimates~\cite{morrison2018closing}.
\citet{gualtieri2017viewpoint} apply active perception directly to grasp detection. They compute a distribution of viewpoints for object classes which are likely to improve the quality of detected grasps. However, this approach requires knowledge of the object's class and doesn't easily generalise to cluttered environments or more complex objects. They also show improved grasp estimates using a heuristic approach where the camera is aligned to the best detected grasp.
\section{Multi-View Picking}
The choice of camera viewpoints plays an important role in the quality of visual grasp detection. In this work, we apply active perception techniques to compute the next best viewpoint for a robot with an eye-in-hand camera in real time while reaching for an object. Unlike previous work, our approach does not rely on any object-specific knowledge or heuristics, does not use a fixed data collection routine and uses visual grasp detection observations directly rather than a secondary metric such as point cloud reconstruction.
We develop an information-gain controller, which we call the Multi-View Picking (MVP) controller. The MVP controller combines visual grasp predictions from multiple camera viewpoints along a trajectory to a grasp pose, and seeks to minimise the uncertainty (entropy) associated with the grasp pose prediction by altering the trajectory to include informative viewpoints, overcoming the challenges associated with visual grasp detection in clutter. Our implementation of the MVP controller is described in the following sections, with an overview given in \fig{overview}.
\begin{figure}[b]
\centering
\includegraphics[width=\columnwidth]{ggcnn_fig2.eps}
\vspace{-5mm}
\caption{We use the GG-CNN grasp prediction network \cite{morrison2018closing}. Given a depth image, the GG-CNN predicts the quality, rotation around the vertical axis and gripper width for a grasp at the position corresponding to each pixel.}
\label{fig:ggcnn}
\end{figure}
\begin{figure*}[tpb]
\centering
\includegraphics[width=\textwidth]{overview.eps}
\vspace{-7mm}
\caption{An overview of our MVP controller. (a) Three time steps during an example grasping trajectory which converges to the point with the best detected grasp pose (the screwdriver handle), considering informative viewpoints corresponding to areas of high expected information gain while reaching. From left to right: The position of the robot (and camera); the collection of objects being grasped; the depth image captured from the current pose $\bp_t$; the quality output of the GG-CNN ($\mathbf{Q}$); the average quality at each grid cell ($\bar{q}$); the entropy at every grid cell ($H$); the utility of a viewpoint ($U$) above each cell in the workspace, from which the velocity command $\bv$ is calculated, and the traversed path superimposed in blue. (b) The histograms $\bm_{j,k}$ and $\bq_{j,k}$ which represent the counts of observations at, from left to right, grid cells with the best estimated grasp pose, a flat surface with low entropy and, a highly occluded area which results in a high measurement entropy.}
\label{fig:overview}
\vspace{-3mm}
\end{figure*}
\subsection{Viewpoint Trajectory}
We consider the case of a robotic arm with an antipodal gripper and an eye-in-hand camera reaching from an arbitrary starting pose to a grasp pose $\bg$. At any instance during this motion, the camera's viewpoint $\bp$ is the 3D position of the camera, centred above a point in the workspace $(x,y)$ at a height $z$, which we constrain to be parallel to the xy-plane. We define a viewpoint trajectory for grasping as a set of $K$ discrete viewpoints $\cT = \{\bp_0, ..., \bp_K\}$ from which we make a visual grasp detection observation while reaching for a grasp. The initial viewpoint $\bp_0$ is at a fixed height $z_\text{max}$, and the trajectory ends when the camera reaches a height $z_\text{min}$, at which point the best detected grasp is executed.
\subsection{Visual Grasp Detection}
For visual grasp detection, we use the real-time Generative Grasping Convolutional Neural Network (GG-CNN) from \cite{morrison2018closing}. Given a depth image as input, the GG-CNN produces a pixelwise visual grasp detection $\bx = (\bQ, \bP, \bW)$, representing the grasp quality (the chance of grasp success), angle of rotation around the vertical axis and gripper width at each pixel of the input respectively (\fig{ggcnn}). Each pixel of $\bx$ represents an antipodal grasp $\bg = (\bc, \phi, w, q)$, parameterised by the grasp's centre position $\bc = (x,y,z)$, rotation $\phi \in [0, \pi]$, gripper width $w$ and quality $q \in [0,1]$. $\bc$ is computed from the measured depth and the camera's intrinsic parameters.
Other visual grasp detection systems regress a single grasp pose~\cite{kumra2017grasp, redmon2015realtime} or perform classification on sampled grasp candidates~\cite{mahler2017dex, lenz2015deep}, which do not easily lend themselves to defining a probability distribution over grasp estimates. The GG-CNN is an ideal component of our active perception system because it directly generates a distribution of grasp estimates and also gives real-time computation (approx. 50Hz). An additional advantage of GG-CNN is that it would be trivial to mask the pixelwise grasp estimates with an off-the-shelf semantic segmentation system if object-specific grasping was required.
\subsection{Grid Map Representation}
To combine observations at time-steps along the viewpoint trajectory,
we represent the workspace of the robot as a 2D grid map, $M$, of $J \times K$ cells. Each cell corresponds an $\gcs \times \gcs$ physical area. The grid map has the advantage of being computationally efficient compared to other representations, such as Gaussian Processes, which could be used in a similar.
Within each cell $(j,k)$, grasp quality observations ($q$) are counted in a vector $\bq_{j,k}$, discretised into $N_q$ intervals, and combined grasp quality and angle observations ($q$, $\phi$) are counted in a 2-dimensional histogram $\bm_{j,k}$, discretised into $N_q \times N_\phi$ intervals respectively (\fig{overview}b). These vectors represent the distribution of observations at each point, and form the basis for our information gain approach.
A grasp at cell $(j,k)$ is parameterised by the mean of observations within that cell:
\begin{equation}
\bg_{j,k} = (\bc_{j,k}, \bar{\phi}_{j,k}, \bar{w}_{j,k}, \bar{q}_{j,k})
\end{equation}%
where $\bc_{j,k}$ is the physical position at the centre of the cell and the mean observations are calculated as per below. (We drop the $j,k$ notation for the sake of readability.)
For a single cell, the mean quality observation $\bar{q}$ is given by:
\begin{equation}
\bar{q} = \frac{1}{\sum \bq} \sum_{n_q=1}^{N_q} \frac{n_q}{N_q} \bq_{n_q}
\end{equation}%
and the mean angle $\bar{\phi}$ is the vector mean of the angle observations~\cite{mardia2014statistics} weighted by the corresponding grasp quality observations:
\begin{equation}
\bar{\phi} = \arctan \frac{
\sum_{n_\phi=1}^{N_\phi}
\sum_{n_q=1}^{N_q}
\sin(\frac{n_\phi}{N_\phi}\pi) \frac{n_q}{N_q}\bm_{n_\phi,n_q} }
{\sum_{n_\phi=1}^{N_\phi}
\sum_{n_q=1}^{N_q}
\cos(\frac{n_\phi}{N_\phi}\pi)\frac{n_q}{N_q}\bm_{n_\phi,n_q}}
\end{equation}
The mean grasp width $\bar{w}$ for a cell is simply the mean of $n$ observations: $\bar{w} = \frac{1}{n} \sum w$
\subsection{MVP Controller}
We formulate our MVP controller using an information gain approach, where
viewpoints are selected to reduce the uncertainty in the grasp pose observations. Specifically, we aim to reduce the entropy of grasp quality observations in $M$ which correspond to high quality grasps.
We can calculate the entropy of the grasp quality observations a single grid cell as:
\begin{equation}
H(\bq) = - \sum_{n_q = 1}^{N_q} \frac{\bq_{n_q}}{\sum \bq} \log \left( \frac{\bq_{n_q}}{\sum \bq} \right)
\end{equation}
Here, we calculate entropy in the quality observations $\bq$ only, rather than the full distribution of grasp quality and angle in $\bm$ because entropy in the distribution of angle measurements is not always a good indicator of uncertainty in the measurement. For example a small, spherical object can be graspable from any angle with high quality, so may have a high entropy across $\bm$ but low entropy when considering only $\bq$.
We find that entropy in the distribution of grasp quality measurements is a good candidate for predicting informative viewpoints. The fifth column of \fig{overview}a shows the entropy of measurements in a semi-cluttered scene at three time steps during a grasp. As shown in \fig{overview}, areas of high entropy are present around areas of clutter, occlusion and complex geometry where the output of the grasp detector is highly variable and dependent on viewpoint, compared to ``uninteresting" the areas with certain measurements (and hence low entropy) regardless of viewpoint such as the flat bottom surface of the workspace.
Calculating the expected information gain of an observation from a viewpoint, in terms of reduction in entropy, is an intractable problem in this case. As such, we use the common simplifying assumption that the total entropy of the observed area will provide a good approximation to computing the expected information gain~\cite{thrun2005probabilistic}. That is, viewpoints which observe areas of high entropy are likely to be more informative (i.e. reduce entropy) than observing areas which already have low entropy.
Hence, we approximate the expected information gain from an observation at a viewpoint as the weighted sum of entropy of the grid cells observed from that viewpoint:
\begin{equation}
E[I(M, \bp)] \approx \sum_{j,k \in O_\bp} P(j,k) \cdot H(\bq_{j,k})
\end{equation}%
where $O_\bp$ is the set of grid cell coordinates observable by the camera from viewpoint $\bp$, and $P(j,k)$ weights points by a Gaussian function based on their distance from the geometric centre of $O$, encouraging the controller to view high-entropy areas front-on rather than peripherally.
To predict the next best viewpoint, we calculate the utility of moving to a viewpoint centred above each cell in $M$, which represents the desirability of moving to each viewpoint as:
\begin{equation}
U(M, \bp) = E[I(M, \bp)] - \gamma \cost{\bp}
\end{equation}
where $\cost{\bp}$ is the cost associated with moving to the viewpoint $\bp$, tunable by the \textit{exploration cost} parameter $\gamma$. To encourage the controller to observe areas nearby the best detected grasp, rather than irrelevant distant points, the cost term represents the horizontal ($xy$) distance to the grid cell with the highest average grasp quality, centred at $\bc$:
\begin{equation}
\cost{\bp} = \left\Vert \bp_{xy} - \bc_{xy} \right\Vert \cdot \left( 1-\frac{\bp_z - z_\text{min}}{z_\text{max}-z_\text{min}} \right)
\end{equation}
The second term scales the cost based on the vertical position of the camera in the trajectory ($\bp_z$). At the beginning of the trajectory there is zero cost, encouraging exploration of the workspace, which increases linearly as the end-effector descends, causing the controller to converge to the best grasp.
The MVP controller generates a horizontal velocity command $\bv_{xy}$ in the direction of maximum utility, shown in the final column of \fig{overview}a. To enable direct comparison between our experiments, we scale the vertical component of velocity $\bv_z$ such that the absolute end-effector velocity $\left\vert\bv\right\vert$ is constant.
\section{Robotic Experiments}
\label{secn:experiments}
\begin{figure*}
\centering
\hfill
\subfloat[]{%
\includegraphics[width=0.35\textwidth]{setup.eps}}
\label{1a}\hfill
\subfloat[]{%
\includegraphics[width=0.3\textwidth]{adversarial.jpg}}
\label{1b}\hfill
\subfloat[]{%
\includegraphics[width=0.225\textwidth]{household.jpg}}
\label{1c}\hfill
\caption{(a) Our experimental setup. (b) The adversarial object set. (c) The household object set.}
\label{fig:setup}
\end{figure*}
We test and validate our approach through a number of trials of robotic grasping in clutter, with the setup is shown in \fig{setup}a. Our experiments are described in the following sections. Our software implementation of the system in the form of ROS nodes, primarily written in python, will be made available online.
\subsection{Equipment}
Experiments are performed using a Franka Emika Panda robot, fitted with custom, 3D-printed gripper fingers using the design from~\cite{guo2017design}. We use an Intel Realsense D435 depth camera, which is mounted to the robot's end-effector.
\subsection{Objects}
We use a set of 40 objects, comprising 20 ``adversarial'' 3D-printed objects from the DexNet 2.0~\cite{mahler2017dex} dataset\footnote{3D-printable mesh files are available from\\ \href{https://berkeleyautomation.github.io/dex-net/}{https://berkeleyautomation.github.io/dex-net/}} (\fig{setup}b), and 20 household objects (\fig{setup}c). The adversarial objects have complex geometry, making them difficult to perceive and grasp. The household objects contains objects that are a wide variety of sizes and shapes and include deformable objects and visually challenging objects which are transparent or black.
\subsection{Method}
For each experimental run, 20 objects are chosen at random, 10 from each of the adversarial and household sets, and emptied into a $30\times30$cm bin in an unstructured jumble. The robot then grasps objects one by one and places them into a second bin until all objects have been removed. A grasp is counted as a success if the object is successfully transported to the second bin. Scales on the second bin are used to record the success or failure.
We perform 9 experiments using the MVP controller, comprising 7 runs each, varying the \textit{exploration cost} $\gamma$ while keeping all other parameters listed in \tab{exp1_params} constant. Additionally, we perform baseline experiments for comparison as described in the next section.
\begin{table}[tpb]
\caption{Parameters used during experiments}
\label{tab:exp1_params}
\centering
\begin{tabular}{@{}llr@{}}
\toprule
\multicolumn{2}{l}{ \textbf{Parameter} } & \textbf{Value} \\
\midrule
$J$, $K$ & Grid Map Size & 68 \\
$\gcs$ & Grid Cell Size & 5mm \\
$N_q$ & Quality Bins & 10 \\
$N_\phi$ & Angle Bins & 18 \\
$z_\text{max}$ & Starting Height & 0.5m \\
$z_\text{min}$ & Final Height & 0.2m \\
& Controller Update Rate & 10Hz \\
$\left\vert\bv\right\vert$ & End-effector Velocity (During Reach) & 0.1m/s \\
$\gamma$ & Exploration Cost & \textbf{Varied} \\
\bottomrule
\end{tabular}
\vspace{-3mm}
\end{table}
\subsection{Baselines}
\label{secn:baselines}
We compare our results to three baselines which represent common methods in other robotic grasping work. We complete 5 runs of each baseline using the method above. Where relevant, the parameters in \tab{exp1_params} (including the end-effector velocity) are kept constant to allow for the best possible comparison.
\textbf{Single Viewpoint} Most work in visual grasp detection considers only a single viewpoint for grasp detection~\cite{mahler2017dex, mahler2017binpicking, pinto2016supersizing, lenz2015deep, johns2016deep}. In this baseline we always execute the best grasp detected by the GG-CNN from a single viewpoint centred above the workspace.
\textbf{Fixed Data Collection} ten \citet{ten2017grasp} increase their grasp success rate by performing point cloud fusion along a fixed trajectory. Because our method uses grasp estimates rather than point clouds, we use our grid map representation to combine GG-CNN predictions along a fixed, spiral trajectory (\fig{paths}), considering 25 and 50 uniformly spaced viewpoints in two experiments.
\textbf{No Exploration} \citet{gualtieri2017viewpoint} showed an increase in grasp success by aligning the camera to the axis of the best detected grasp. In this baseline we disable the exploration of our MVP controller, instead always generating a velocity command to align the camera to the best detected grasp at each time step.
\section{Results}
\begin{table*}[t]
\caption{Experimental results for grasping objects in clutter, using our MVP controller and the three baseline cases.}
\label{tab:results}
\centering
\begin{tabular}{@{}llllllllllcllll@{}}
\toprule
& \multicolumn{9}{c}{ \textbf{Ours (MVP Controller)} } && \multicolumn{4}{c}{ \textbf{Baselines} } \\
\cmidrule{2-10} \cmidrule{12-15}
& \multicolumn{9}{c}{ \textit{Exploration Cost} ($\gamma$) } && \textit{Single View} & \textit{No Expl.} & \multicolumn{2}{c}{ \textit{Fixed Data Collection} } \\
\cmidrule{14-15}
& 0.0 & 0.05 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & & \textit{} & \textit{} & 25 Views & 50 Views \\
\cmidrule{2-10} \cmidrule{12-15}
Total Attempts & 174 & 175 & 178 & 171 & 181 & 182 & 179 & 189 & 188 && 196 & 187 & 137 & 125 \\
Failures & 35 & 35 & 38 & 36 & 42 & 42 & 45 & 50 & 49 && 62 & 48 & 37 & 28 \\
Mean Viewpoints & 44 & 43 & 39 & 36 & 35 & 35 & 35 & 35 & 35 && 1 & 35 & 25 & 50 \\
\cmidrule{2-10} \cmidrule{12-15}
Success Rate (\%) & \textbf{80} & \textbf{80} & 79 & 79 & 77 & 77 & 75 & 74 & 74 && 68 & 74 & 73 & 78 \\
Mean Time (s) & 10.5 & 10.2 & 9.8 & 9.2 & 9.0 & 9.1 & 9.0 & 9.1 & 9.1 && \textbf{8.8} & 9.1 & 11.4 & 11.4 \\
MPPH & 273 & 282 & 288 & \textbf{308} & 307 & 305 & 299 & 292 & 292 && 281 & 293 & 230 & 245 \\
\bottomrule
\end{tabular}
\vspace{-2mm}
\end{table*}
The results of our experiments are shown in \tab{results}. We assess each experiment based on three metrics:
\begin{itemize}
\item \textit{Success Rate:} The overall ratio of successful grasps to grasp attempts across all runs.
\item \textit{Mean Time per Pick:} The average time per grasp attempt (in seconds), regardless of success.
\item \textit{Mean Picks Per Hour (MPPH):} The overall efficiency of the system representing the average rate of successful picks per hour calculated as $\text{3600 [s/h]} \div (\text{Mean Time per Pick [s]} ) \times (\text{Success Rate})$
\end{itemize}
It is important to note that we include the MPPH measurement as a way of comparing the different results, so are primarily concerned with the relative difference of values rather than their overall magnitude which is highly dependent on the fixed end-effector velocity that we use.
\subsection{Exploration During Grasping}
We first investigate the effect of trading off between exploration and execution time. Varying the \textit{exploration cost} from $\gamma=0.7$ (minimum exploration) to $\gamma=0.0$ (maximum exploration) results in a 6\% increase in grasp success rate, improving from 74\% to 80\%, at the expense of 1.4 seconds per grasp on average. Two example trajectories for $\gamma=0.1$ and $\gamma=0.7$ are shown in \fig{paths}.
As the increased time associated with exploration does not scale linearly with the increase in grasp success rate, the overall efficiency of the system, measured in MPPH, is maximised between the two extremes, for $\gamma=0.2$, where there is an increase in grasp success rate (79\%) but minimal extra time cost. As a result, the MVP controller can be optimised for either raw grasp success rate or overall efficiency by adjusting the cost of exploration.
As $\gamma$ is increased, the results begin to plateau in the range 0.5 to 0.7 due to the the cost term becoming dominant in this range and causing to controller to perform minimal exploration and converge directly to the grasp.
\subsection{Comparison to Baselines}
\textbf{Single Viewpoint} Our approach outperforms the single viewpoint baseline approach in terms of both success rate, improving up to 12\% (68\% vs. 80\%), and MPPH, increasing by approximately 10\% from 281 to 308 despite an increased mean time per pick. This reinforces the assertion that considering multiple viewpoints is an effective way to overcome the visual challenges associated with grasping in clutter.
\textbf{No Exploration} The \textit{No Exploration} baseline achieves similar results to our MVP controller when using a high \textit{exploration cost}, which is unsurprising since the high \textit{exploration cost} results in minimal exploration. However, for lower values of $\gamma$, our MVP controller outperforms this baseline by 6\% success rate, highlighting the added benefit of actively exploring based on uncertainty compared to using a heuristic such as aligning with the best detected grasp.
\textbf{Fixed Data Collection} The \textit{Fixed Data Collection} baseline reinforces the idea that incorporating multiple viewpoints can improve the success rate of a grasping system, with both experiments outperforming the \textit{Single Viewpoint} baseline. Furthermore, the 50-viewpoint experiment outperforms the 25-viewpoint experiment by 5\% grasp success rate. However, because the fixed trajectory always views the entire workspace uniformly, it results in a constant execution time which is longer than all other experiments and is unable to focus on salient areas of the workspace. As a result, our MVP controller outperforms this baseline with regards to all metrics, including a higher grasp success rate with fewer viewpoints. The main advantage comes from our information gain approach, which is able to focus on salient areas of the workspace to reduce unnecessary data collection.
\begin{figure}[tpb]
\centering
\includegraphics[width=\columnwidth]{paths.eps}
\vspace{-5mm}
\caption{Three viewpoint trajectories, for $\gamma=0.1$, $\gamma=0.7$ and the \textit{Fixed Data Collection} baseline.}
\label{fig:paths}
\vspace{-4mm}
\end{figure}
\subsection{Automatically Adapting to Scene Complexity}
As shown with the \textit{Fixed Data Collection} baseline, using a set of fixed viewpoints for data collection can increase grasp detection accuracy. However, it results in longer, constant execution times, and is unable to focus attention on ``interesting" parts of the scene. In contrast, our MVP controller is able to actively adapt to the complexity of the scene and provides more a more efficient data collection process. \fig{timevsobjects} shows that for our MVP controller the mean time per pick is dependent on the number of objects in the workspace. As the number of objects in the scene increases, and with it the amount of clutter and potential occlusions, so does the mean time per pick, taking on average 20\% longer when 20 objects are present compared to a single object.
\begin{figure}[tpb]
\centering
\includegraphics[width=0.97\columnwidth]{time_vs_objects.eps}
\vspace{-2mm}
\caption{Mean time per pick increases with the number of objects present, compared to the mean time for one object, averaged over all experimental runs (all values of $\gamma$). Shaded area shows one standard deviation.}
\label{fig:timevsobjects}
\vspace{-4mm}
\end{figure}
\section{Conclusion}
We presented an active perception approach to grasping in clutter, in which we consider visual grasp detections from multiple viewpoints while reaching. Our work reinforces the importance of viewpoint selection and combining data from multiple viewpoints when grasping in clutter, but in contrast to previous work our Multi-View Picking controller uses an information-gain approach to select informative viewpoints that directly seek to reduce entropy in the grasp pose estimates caused by clutter and occlusions.
We validate our approach with several experiments in grasping from clutter. Our MVP controller achieves up to 80\% grasp success while picking from cluttered piles of up to 20 objects, including adversarial objects with complex geometry, outperforming a single-viewpoint method by 12\%.
Additionally, by using an information-gain approach, our MVP controller is able to adapt to the complexity of the scene, unlike other approaches which rely on fixed data collection routines as part of a visual grasp detection pipeline. Compared to such a method, our approach results in a higher grasp success rate while also being more efficient, requiring fewer viewpoints and less time per grasp.
\printbibliography
\end{document}
|
\section{Motivation and significance}
\label{sec:motivation}
Most scientific fields produce theoretical or experimental data which is not necessarily the result of a measurement, but also of simulations or evaluations of theoretical models. This data is intrinsically complex consisting of multiple parameters or multiple observables, and thus, data-sets can be regarded as point clouds in a high-dimensional space. Often, due to the restrictions imposed by the use of paper for data visualization, e.g. a figure in a journal, the status quo is still to publish the data in a two (or three) dimensional format. A typical example is Figure~\ref{fig:example}.
However, such a two dimensional (2D) representation obscures most of the correlations within the solution space.
In order to encourage the publication of high-level data in the complete high-dimensional space, without restrictions, data visualization can be done via web-based tools which allow for, e.g. an automatic generation of multiple relevant histograms. The aim of \textsc{spot}{}~\cite{jisk_attema_2017_1003346} is to provide a flexible data visualization framework to visualize such data. \textsc{spot}{}, which is typically coupled to a database holding the data-sets is a tool to promote the use of \emph{open research data} and \emph{open science}~\cite{machado2015open}. It follows \emph{FAIR}~\cite{wilkinson2016fair} principles to allow reuse and comparison of the published data-sets (see Fig.~\ref{fig:workflow}). This paper briefly introduces \textsc{spot}{}. The source code and the documentation is available at \url{https://github.com/NLeSC/spot}.
\begin{SCfigure}
\centering
\caption{A typical scientific results from the field of High-Energy-Physics \cite{Aad:2015baa}. A model with 19 parameters (Supersymmetry) has been tested again the data to determine if parameter sets are excluded by the data or not. Here the fraction of model parameter sets (as colour code) are plotted for two observables predicted by the model as $x$- and $y$-axis. This should show how difficult it is to visualize a 19-dimensional model on a paper and how much information is then lost.}
\includegraphics[width=0.5\textwidth]{figures/N4.png}
\label{fig:example}
\end{SCfigure}
\begin{figure}[!b]
\centering
\includegraphics[trim={0 2cm 0 2cm}, width=0.4\textwidth]{figures/flow1.png}
\includegraphics[trim={0 2cm 0 2cm}, width=0.4\textwidth]{figures/flow2.png}
\caption{Workflow of \textsc{spot}{}.}
\label{fig:workflow}
\end{figure}
\section{Software description}
\label{sec:description}
\textsc{spot}{} provides users with an interactive data exploration environment for high-dimensional data-sets.
The focus is on scientific use, with the aim of facilitating open science, data sharing and reuse (see Figure~\ref{fig:workflow}). It is ideal for numerical data, but categorical (labeled) and temporal data is supported
Built on a number of concepts from the field of information visualization, it allows a user to create multiple coordinated views called \emph{charts}, showing the data from different perspectives. All charts allow direct manipulation (i.e., selecting and zooming) of the data, and provide visual clues or \emph{animations} when data changes due to user interaction.
\subsection{Software Architecture}
\label{sec:architecture}
The software consists of three components: a \emph{framework}, a \emph{frontend}, and a \emph{server}. A brief description for each of these components follows:
\paragraph{\textsc{spot}{}-framework}
The framework provides classes for data-sets, data views, partitions, aggregation and filtering.
A \emph{data-set} consists of a number of items (or rows), and each item has a set of \emph{facets}.
Facets can be used to partition the data, or they can be aggregated (\textsf{counted}). For numerical data more complex operations are possible, namely, \textsf{summation}, \textsf{averaging}, \textsf{extremes}, \textsf{standard deviation}.
One or more facets make up a \emph{filter}, and all filters combined together form the \emph{data view}.
The user interacts with the framework by setting ranges or selections for the filters, and by adding or removing filters from the data view.
After filtering, the partitioned and aggregated data is available as a simple array, which can be plotted or further processed.
All filters in a data view are linked, and a change in one filter triggers an update of the whole data view.
\paragraph{Frontend}
The frontend is a web-based application with separate pages where a user can upload and define data-sets, a dashboard page that provides the main interaction, and a page where analyses can be downloaded or shared.
The frontend is built on a number of open-source \tech{JavaScript} packages, such as \tech{Chart.js}, \tech{Vis.js}, and \tech{Sigma.js} for animated plots, and \tech{Ampersand.js} and \tech{Material Design Lite} for the interface.
The filtering is implemented using asynchronous functions, so updates can start as soon as the first results become available. The visualization libraries are \tech{HTML Canvas} based, and offer \tech{OpenGL} accelerated rendering where available.
\paragraph{Server}
The server processes requests for data and applies the necessary filtering and aggregation. When data becomes available, it is pushed to the client which can then update the charts.
We currently have two different implementations for the server component. The first one, which is included in \textsc{spot}-framework, is based on \tech{Crossfilter.js} and runs in the user's web browser without requiring any further resources or even internet access. The second implementation (\textsc{spot}-server), provides a bridge to an external \tech{PostgreSQL} database for scalability. Database queries are run in parallel, and make use of indices for extra performance. Connections to other datastores, like \tech{MySQL} or \tech{MongoDB}, can easily be achieved by extending the server component.
\subsection{Software Functionality}
\label{sec:functionality}
\paragraph{Data import and database connection}
There are two options to import a new data-set.
In the first option, users can upload data available on their own system.
The software supports most common data formats \tech{CSV} and \tech{JSON}, to make data import process easy for different scientific domains.
After the import, the data is checked to automatically detect data types, such as integers and strings. Users can then fix auto detection issues. In the second option, the data is imported from \textsc{spot}-server. The meta data for a data-set (e.g. the name and description) can easily be set in a configuration file stored on server-side.
\paragraph{Dashboard}
\label{sec:charts}
\textsc{spot}{} has eight ready-to-use chart types, namely, horizontal and vertical histograms, line chart, pie chart, bubble chart, 3-d scatter chart, radar chart and network chart.
Charts are added to the dashboard by clicking the chart icon.
The chart's filter requires one or more facets to partition over, and can take up to 4 facets to aggregate.
Charts show their configuration pane by default.
A visual feature of the chart can be linked to a specific facet by dragging a facet from the top of the screen and dropping it on a slot in the configuration pane.
A Partition or Aggregation can be further configured by clicking on its name on the configuration pane.
\paragraph{Download and share}
The dashboard generated by the user can be saved as a single file, a session file, in \tech{JSON} format. This file contains aggregated data and settings of the dashboard such as existing charts, existing filters. The session file then can be used to restore the analysis. In addition, the session file can be uploaded to a cloud storage and a link to the session file can be shared.
\section{Illustrative Examples}
\label{sec:example}
We show the applicability and the features of the \textsc{spot}{} software on two examples of data-set: (1) the Titanic data-set \cite{titanic-data}, a well-known data-set in data science, and (2) a high-dimensional data-set containing models for dark matter (see \cite{achterberg2017implications,hendriks2015description} for more information).
\begin{figure}[!h]
\centering
\includegraphics[width=0.8\textwidth]{figures/titanic}
\caption{Web-interface of \textsc{spot}{}.
\circled{1} Main menu: to navigate between different pages. Users can upload data (Datasets page). analyze it (Analyze page), share selection and charts as a \emph{session} (Share page), and read help documentation (Help page); \circled{2} Information bar: provides basic statistics about selected data; \circled{3} Chart bar: in order to add available charts to the dashboard; \circled{4} Variable bar:
lists available variables within the data-set; \circled{5} The dashboard.}
\label{fig:titanic}
\end{figure}
\subsection{Titanic data-set}
Visualisation plots of the Titanic data-set are provided in Figure \ref{fig:titanic}; the data can be viewed interactively on the website \url{http://www.idarksurvey.com} by clicking on Demo.
The top of the figure shows the chart types, each of which can be selected to make a new chart. Directly below that are the data facets, which can be dragged-and-dropped into the empty charts to create any visualisation of any parameter(s).
\subsection{Example from High-Energy Physics}
A real world example where \textsc{spot}{} can help the scientific community is visualizing e.g. models of high-energy physics. The data in this field is typically high-dimensional and even though different models have different theoretical parameters, they share the same observables.
\textsc{spot}{} allows comparison of these observables from different data-sets, providing the user an unprecedented ability to compare the model space. In Figure \ref{fig:dmmodels}, two data-sets for models predicting Dark Matter are compared for three observables: the dark matter mass, a annihilation probability $\sigma v$ and a cosmic density $\Omega h^2$.
\begin{SCfigure
\centering
\includegraphics[width=0.5\textwidth]{figures/dataset-comparison.png}
\caption{An example 3D visualization of a comparison between two dark matter model data-sets. There are three shared observables. When hovering over a data point, a pop-up tooltip will show the actual values of the model point.}
\label{fig:dmmodels}
\end{SCfigure}
\section{Impact}
\label{sec:impact}
The high-dimensional space can be stored in the \textsc{spot}{} database so that the data-set can be published along with the paper. While the paper still contains the most relevant 2D plots, a researcher can plot different variables using \textsc{spot}{} for further research. By intuitively making cuts in for example histograms, researchers can investigate the high-dimensional space in an unprecedented way. In addition, comparison between data-sets on shared observables is likely to lead to new research questions.
Intuitive data visualization is still in its infancy, especially in the field of high energy physics. \textsc{spot}{} aims to be the first tool to provide an intuitive interface for visualizing high-dimensional data-sets and as a place for researcher to store their high-dimensional data.
Thus, there are two categories of applications that are related to \textsc{spot}{}: online sharing services and visualization libraries. The most relevant to \textsc{spot}{} are Microsoft Power BI~\cite{powerbi}, Spotfire~\cite{ahlberg1996spotfire} and Tableau~\cite{heer2008graphical} but these are commercial products. The most popular sharing services include data sharing platform Zenodo~\cite{zenodo}, digital repository for sharing Figshare~\cite{figshare}, and high-energy physics specific platform Hepdata~\cite{maguire2017hepdata}. In comparison with \textsc{spot}{}, however, these services basically provide only storage that allow researchers to upload and publicly store their data and figures, but not compare different visualizations from different articles. On the other hand, visualization frameworks such as Dash~\cite{plotly} is a Python framework that gives users possibility of interactive visualization, but also it requires users to have an advance knowledge to create a desired dashboard.
\section{Conclusions}
\label{sec:conclusion}
The software has been written to provide a free and open state-of-the-art platform for researchers in many scientific domains. It helps researchers to publish, share their data-sets, and collaborate by comparing their data-sets to identify the differences. This makes \textsc{spot}{} a perfect candidate for FAIR data platforms.
\section*{Acknowledgements}
\label{sec:acks}
This work is supported by the Netherlands eScience Center under the project iDark: The intelligent Dark Matter Survey.
\bibliographystyle{unsrt}
|
\section{Summary of Fabrication and Measurements Methods}
Devices are fabricated on a single crystalline thin-film lithium niobate (LN) device layer bonded onto a silicon (Si) handle wafer with a 2 $\mu$m thick thermally grown silicon dioxide layer on top. Standard electron-beam (e-beam) lithography is used to realize optical waveguides and microresonator patterns in hydrogen silsequioxane (HSQ) e-beam resist. The patterns are then transferred into the LN layer using argon plasma etching in a standard inductively couple plasma reactive ion etching (ICP-RIE) tool. The etched depth is 350 nm leaving a 250 nm LN slab behind. The slab allows for efficient electrical field penetration into the waveguide core region. The first layer of the gold interconnects is patterned using e-beam lithography and the metals are deposited with e-beam evaporation methods and lift-off processes. Next, a 1.6 $\mu$m silicon dioxide layer is deposited on top using plasma enhanced physical vapour deposition (PECVD) method. Finally, metal vias and the top metal layer are realized using a standard photolithography followed by e-beam evaporation and lift-off processes.
The light from a tunable telecom wavelength laser (SANTEC TS510) is launched into, and collected from, the LN waveguides using a pair of lensed optical fibres. The microwave control signals are generated from an arbitrary waveform generator (AWG, TEKTRONIX 70001A), before they are sent to electrical amplifiers. Electrical circulator or isolators are used to prevent multiple electrical reflections. For the Rabi oscillation measurements, the electric field amplitude ($c_1$) in the S mode is measured by interfering the light out-coupled from the double-ring system with the pump light in the optical waveguide. The interference produces a homodyne signal for $c_1$ that is sent to a 12 GHz photodiode (Newport 1544A), and due to the optical frequency difference, the rapid interference signal between the pump light and $c_2$ can be filtered out electrically using a low-pass filter. For the Ramsey measurements, the optical power is sampled after the 2nd $\pi/2$ pulse using the fast photodiode. For the photon storage measurements, the pump light is synchronously turned off with the first $\pi/2$ pulse allowing for direct power readout of the retrieved light and prevent pump further leaking into the bright mode. The modulation on the pump signal is achieved by an external electro-optic modulator synchronized with the microwave control signals.
\section{Device and Experimental Setup Details}
\begin{figure}[ht!]
\centering
\includegraphics[angle=0,width=0.8\textwidth]{figs2.pdf}
\caption{\label{figs2}\textbf{Device details.} (a) scanning electron microscope (SEM) image of the gap between the coupled microring resonators. (b) Cross-section of the optical mode profile in the ring resonator. (c) microscope image of the full device showing the double-ring and microwave electrodes. (d) SEM image of the array of double ring devices fabricated on a single chip.}
\label{fig:devices}
\end{figure}
The ring resonators of the photonic molecule have waveguide width of 1.4 $\mu$m and vary coupling gaps of $\sim 700$ nm. The coupling waveguides are 800 nm wide by 600 nm thick with a rib height of 350 nm and slab height of 250 nm. This ratio is chosen to ensure optimum electro-optic over-lap while still maintaining a tight bending radius.
The high frequency microwave electrodes are designed to achieve differential driving of the two resonators. As shown in figure \ref{fig:devices}, the top contact pad is connected to the bottom electrodes of the left ring and the top electrodes of the right ring (and vise versa for the bottom contact pad). Since the LN thin film is $x$-cut (extraordinary axis $z$ in plane), such configuration allows the electrical field to be in the same direction on each ring, while opposite for the two rings. Here we pattern 60 programmable photonic molecules on a 10 mm by 8 mm thin film lithium niobate (LN) chip.
\begin{figure}[ht!]
\centering
\includegraphics[angle=0,width=0.8\textwidth]{figs1.pdf}
\caption{\label{figs1}\textbf{Experimental Setup.} The device is optically pumped by a tunable telecom laser centered around 1630 nm. The light is sent through an external electro-optic modulator and polarization controllers before coupling into the chip with a lensed fiber. The output optical signal, also coupled with a lensed fiber, is sent to a 12 GHz photodetector (New focus 1811A). The converted electrical signal is directed to an oscilloscope. The microwave control signals are generated by an arbitrary wave generator (AWG, Tektronix 70001A) and amplified before sent on to the device. A bias T is used to allow DC control on the microresonators. An electrical isolator is used to capture the electrical reflection from the microresonators. The oscilloscope, device drive signals and modulator drive signals are all synchronized.}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[angle=0,width=0.9\textwidth]{figs0.pdf}
\caption{\label{figs3}\textbf{Measured transmission spectrum.} Typical normalized transmission spectra of a double-ring resonator showing pairs of photonic two-level doublets, separated by $\sim$ 2.2 nm of free-spectral-range.}
\label{fig:trans}
\end{figure}
The optical properties of the devices are characterized using a tunable telecom laser (Santec TSL-550). The polarization of the light is tuned to TE (in-plane of the chip) before sending into the chip with a lensed fiber. The DC transmission is measured on a 125 MHz photodetector (New Focus 1811A) and the fast oscillation signal is measured on a 12 GHz photodetector (New Focus 1544A) and a fast real-time oscilloscope (Tektronix MSO71604C). Figure \ref{fig:trans} shows the normalized transmission spectrum of the photonic molecule, showing pairs of two-level resonances separated by the free-spectral-range of the microring resonator ($\sim$ 2.2 nm).
The microwave signals are prepared using an arbitrary waveform generator (Tektronix 70001A), which provides synchronized signals to drive the photonic molecule as well as an external modulator and the oscilloscope. The microwave driving signal for the photonic molecule is amplified and sent through an isolator to minimize reflection to the amplifier. A bias-T is used to coupled DC field into the resonator for fine tuning.
\section{Description of the photonic molecule}
\subsection{System Hamiltonian}
The nanophotonic two-level system under a coherent microwave drive can be described by a Hamiltonian with the form
\begin{equation}
\label{eq:ham}
\hat{H}=\omega_1 a_1^\dagger a_1+\omega_2 a_2^\dagger a_2 + \mu (a_1^\dagger a_2+a_1 a_2^\dagger) +\Omega (a_1^\dagger a_1-a_2 a_2^\dagger) \textrm{cos} (\omega_m t)
\end{equation}
where $a_1$ ($a_1^\dagger$) and $a_2$ ($a_1^\dagger$) are the annihilation (creation) operators of the two optical modes of the respective microresonator, $\mu$ is the coupling strength between the two optical resonators, $\Omega=g V_0$ is the interaction strength of microwave field to the optical resonator, determined by the coupling strength $g$ and microwave peak amplitude $V_0$. $\omega_m$ is the frequency of the microwave modulation.
\subsection{Identical resonator: two bright modes}
When the resonant frequencies of the two resonators are identical, i.e. $\omega_1 = \omega_2 \equiv \omega_0$, equation \eqref{eq:ham} can be expressed as
\begin{equation}
\label{eq:hamid}
\hat{H}=\omega_+ c_1^\dagger c_1+ \omega_- c_2^\dagger c_2 +\Omega (c_1^\dagger c_2+c_1 c_2^\dagger) \textrm{cos} (\omega_m t)
\end{equation}
where $\omega_{\pm}=\omega_{0}\pm \mu$, $c_1^{(\dagger)}=\frac{1}{\sqrt{2}}(a_1^{(\dagger)}+a_2^{(\dagger)})$ and $c_2^{(\dagger)}=\frac{1}{\sqrt{2}}(a_1^{(\dagger)}-a_2^{(\dagger)})$.
To simplify the system, we apply a unitary transformation $U_1=e^{(i \omega_+ c_1^{\dagger} c_1 t + \omega_- c_2^{\dagger} c_2 t)}$ and the rotating wave approximation (RWA), we obtain
\begin{equation}
U_1\hat{H}U_1^{\dagger}= \frac{\Omega}{2} (c_1^\dagger c_2 e^{i\delta t}+c_1 c_2^\dagger e^{-i \delta t})
\label{eq:hmit}
\end{equation}
where the microwave detuning $\delta=\omega_m-\Delta \omega$ and $\Delta\omega$ is the frequency difference between $c_1$ and $c_2$. Equation \ref{eq:hmit} is equivalent to the time independent Hamiltonian
\begin{equation}
\label{eq:hamid_1}
\hat{H}=\frac{\delta}{2} c_1^\dagger c_1 - \frac{\delta}{2} c_2^\dagger c_2 +\frac{\Omega}{2} (c_1^\dagger c_2+c_1 c_2^\dagger)
\end{equation}
We obtain the following equations of motion using input-output theory from equation \ref{eq:hamid_1},
\begin{equation}
\frac{\d}{\d t}
\begin{pmatrix}
c_1 \\
c_2
\end{pmatrix}
=
\begin{pmatrix}
-i \frac{\delta}{2} & - i \frac{\Omega}{2}\\
- i \frac{\Omega}{2} &i \frac{\delta}{2}
\end{pmatrix}
\begin{pmatrix}
c_1 \\
c_2
\end{pmatrix}
\label{eq:eom}
\end{equation}
\subsection{Autler-Townes Splitting and AC Stark Shifts}
\begin{figure}[ht!]
\centering
\includegraphics[angle=0,width=0.8\textwidth]{figs4.pdf}
\caption{\label{figs4}\textbf{Calculated optical transmission spectrum for Autler-Townes (AT) splitting and AC Stark shift} (a) Simulated AT spectrum with full mode described in equation \ref{eq:hamid}. (b) Simulated AC Stark shift spectrum with full mode described in equation \ref{eq:hamid}}
\label{fig:at}
\end{figure}
Under a continuous wave microwave modulation with frequency close to the optical two-mode splitting, each resonance of $c_1,c_2$ will further split into two dips with splitting frequency
\begin{equation}
W=\sqrt{(\Omega^2+\delta^2)}
\label{eq:at}
\end{equation}
obtained by diagonalizing the coupling matrix
$M=\begin{pmatrix}
-i \delta/2 & - i \Omega/2\\
- i \Omega/2 &i \delta/2
\end{pmatrix}$ from equation \ref{eq:eom}.
For the Autler-Townes splitting measurement data shown in Fig. \ref{fig2}a in the main text and simulation shown in Fig. \ref{fig:at}, the microwave modulation signal is on-resonance with the optical two-mode splitting, i.e. $\delta = 0$, with $W=\Omega=g V_0$.
When the microwave frequency detuning to the two-mode resonances is much larger than the driving strength (i.e. $\delta\gg\Omega$), the microwave modulation behaves as a weak perturbation. This perturbation induces a shift of the optical energy levels. Applying the second-order perturbation theory, we have
\[M
\approx
\begin{pmatrix}
-i \left(\frac{\delta}{2}+\frac{\Omega^2}{4\delta}\right) & \\
& -i \left(-\frac{\delta}{2}-\frac{\Omega^2}{4\delta}\right)
\end{pmatrix}
\]
In the static frame we obtain the shifted frequency:
\[
\omega_\pm^\prime=\omega_\pm \mp\frac{\Omega^2}{4\delta}
\]
In the experimental measurements in Fig. \ref{fig2}b and \ref{fig:at} we have $\delta<0$ where the two energy levels of $c_1,c_2$ moves away from each other. For a blue-detuned microwave drive signal, i.e. $\delta>0$, the two energy levels will move towards to each other.
\subsection{Optical two-level dynamics}
To study the dynamics of the photonic two-level system, we consider the equations of motion \ref{eq:eom} with optical loss and laser input fields. For a single tone optical input on resonance with $c_1$, the equations of motion are:
\begin{align}
&\dot{c_1}=\left(-i \frac{\delta}{2}-\frac{\gamma}{2}\right) c_1 - i \frac{\Omega}{2} c_2
-\sqrt{\gamma_{\textrm{ex}}/2}s_{\textrm{in}} \label{eq:motion1}
\\
&\dot{c_2}=\left(
i \frac{\delta}{2}
-\frac{\gamma}{2}\right) c_2 - i \frac{\Omega}{2} c_1
-\sqrt{\gamma_{\textrm{ex}}/2}
s_{\textrm{in}}e^{\i \omega_m t}
\label{eq:motion2}
\end{align}
where $\gamma$ is the total decay rate of the modes $c_{1,2}$, $\gamma_\textrm{ex}$ is the waveguide coupling rate to mode $a_1$ and $s_\textrm{in}$ is the input laser field. After dropping the fast rotating terms with frequency of $\omega_m$, we obtain the dynamical solution of $c_1$ and $c_2$,
\begin{equation}
\begin{aligned}
c_1&=c_{10}e^{-\frac{\gamma}{2}t} \left( \cos \frac{W}{2}t-\i \frac{\delta}{W}\sin \frac{W}{2}t\right)\\
c_2&=-\i c_{10} e^{-\frac{\gamma}{2}t}\frac{\Omega}{W} \textrm{sin}\frac{W}{2}t
\label{eq:sol1}
\end{aligned}
\end{equation}
where $c_{10}=\frac{\sqrt{\gamma_\textrm{ex}} s_\textrm{in}}{-i\Delta-\gamma/2}$ is the steady state amplitude in mode $c_1$ before the microwave is turned on and $\Delta$ is the laser detuning. Here $W=\sqrt{\Omega^2+\delta^2}$ is the Autler-Townes splitting frequency as in equation \ref{eq:at} and is also the Rabi oscillation frequency. The output power from the waveguide is then
\begin{equation}
\begin{aligned}
P_{\textrm{out}}&=|s+\sqrt{\gamma_{\textrm{ex}}}a_1|^2\\
&=|s+\sqrt{\gamma_{\textrm{ex}}/2}(c_1+c_2 e^{i \omega_m t})|^2
\label{eq:sol1out}
\end{aligned}
\end{equation}
Substitute equations \ref{eq:sol1} into the expression of the output power and set $\delta=0$, we obtain
\begin{equation}
\begin{aligned}
P_{\textrm{out}}=P_\textrm{in}\left|1-\frac{2\gamma_{ex}}{\gamma}e^{-\frac{\gamma}{2}t}\left(
\cos \frac{W}{2}t
\right)
\right|^2
\label{eq:sol1app}
\end{aligned}
\end{equation}
where $P_\textrm{in}=|s|^2$ and we made the approximation to drop the fast rotating terms of $c_2$ at $\sim 7$ GHz, as we restrict our measurement bandwidth to $\sim 2$ GHz. The intra-cavity field can therefore be directly calculated from the output intensity by $c_1=(1-\sqrt{P_{\textrm{out}}/P_\textrm{in}})\times \gamma/2\gamma_{ex}$. Therefore the real-time AC signal on the photodetector is directly related to the intracavity field $c_1$. Figure \ref{fig:dyn} shows the full numerical solution as of equation \ref{eq:hamid}, the approximated solution \ref{eq:sol1app} and the measured photodetector voltage.
\begin{figure}[ht!]
\centering
\includegraphics[angle=0,width=0.8\textwidth]{figs3.pdf}
\caption{\label{figs5}\textbf{Rabi oscillation, Ramsey interference and dark mode measurements} (a) Rabi oscillation is measured in real-time as the microwave signal is turned on. The laser entering the cavity after the turn-on of the microwave signal is negligible as the split optical mode is transparent at the original laser wavelength. Ramsey interference is measured by two time-delayed microwave $\pi/2$ pulses. The optical power is recorded at the end of the second pulse. For the photonic dark mode measurements, the laser is turned off by an external electro-optic modulator for direct detection of optical powers and prevent further coupling into the bright mode during the delay time. (b) Comparison of the measured photodetector voltage, the simulated transmission of the full system and the calculated output from equation \ref{eq:sol1app}}
\label{fig:dyn}
\end{figure}
The Ramsey interference is achieved by applying two microwave $\pi/2$ pulses separated by a time delay $\tau$. Here the microwave signal is detuned from the level-splitting by $\delta$. Using the protocol of Ramsey interference and equation \ref{eq:motion1}, \ref{eq:motion2} , we can get the solution after two $\pi/2$ pulse as a function of time delay $\tau$:
\begin{equation}
\begin{split}
c_1(2T+\tau)=e^{-\frac{\gamma}{2}(2T+\tau)}e^{-\i \frac{\delta}{2}\tau}
\left[
\frac{\Omega^2}{2W^2}(1-e^{\i \delta \tau})-\i \frac{\delta}{W}
\right]c_{10}
+e^{-\frac{\gamma}{2}T}e^{-\i \frac{\delta}{2}\tau}\left[1-e^{-\frac{\gamma}{2}\tau}\right]
\left[
\frac{1}{\sqrt 2}-\i \frac{\delta}{W} \frac{1}{\sqrt 2}
\right]c_{10}
\end{split}
\label{eq:rmc1}
\end{equation}
\begin{equation}
\begin{split}
c_2(2T+\tau)=e^{-\frac{\gamma}{2}(2T+\tau)}e^{-\i \frac{\delta}{2}\tau}
\left[
\frac{\delta \Omega}{2W^2}(e^{\i \delta \tau}-1)-\i \frac{\Omega}{2W}(e^{\i \delta \tau}+2)
\right]c_{10}
+e^{-\frac{\gamma}{2}T}e^{-\i \frac{\delta}{2}\tau}\left[1-e^{-\frac{\gamma}{2}\tau}\right](-\i \frac{\Omega}{W}c_{10})
\end{split}
\label{eq:rmc2}
\end{equation}
where from mode $c_1$. $T$ is the duration of the $\pi/2$ pulse. The first terms in equations \ref{eq:rmc1} and \ref{eq:rmc2} represent the intra-cavity fields exist in the cavity prior to the arrival of the first $\pi/2$ pulse. The second terms are the leakage of the pump signal into the cavity during the delay, which contributes to a DC term. So the oscillation of the output signal $P_\textrm{out}=\left|s+\sqrt{\gamma_{ex}/2}(c_1+c_2 e^{i \omega_m t})\right|^2$ rotating at with the frequency $\delta$ corresponds to the population of $c_1$ after dropping the fast rotating terms related to $c_2$.
\subsection{Optical dark modes}
In the case where $a_1$ and $a_2$ are far detuned by $\delta \omega\gg \mu$, as in the photon storage and retrieval measurements, we prefer to eliminate the coupling term $\mu(a_1^\ad a_2+a_1 a_2^\ad)$ by a Bogoliubov transformation. Assuming a new basis of $c_1, c_2$ satisfying:
\[
c_1=va_2-ua_1
\]
\[
c_2=ua_2+va_1
\]
Since $c_1,c_2$ needs to satisfy the bosonic commutation relationship, we have the condition $u^2+v^2=1$. So we set $u=\cos {\frac \theta 2}, v=\sin{\frac \theta 2}$ with $\tan\theta=\mu/\delta \omega$. Then result of this transformation gives us a Hamiltonian for $c_1,c_2$:
\[
\begin{split}
\hat{H}=& \omega_1 c_1^\ad c_1 +\omega_2 c_2^\ad c_2 +\Omega \cos (\omega_m t) \sin\theta (c_1 c_2^\ad+c_1^\ad c_2)\\&+ \Omega \cos (\omega_m t) \cos\theta (c_2^\ad c_2-c_1^\ad c_1)
\end{split}
\]
where $\omega_1=\omega_0-\sqrt{\mu^2+\delta \omega^2},\omega_2=\omega_0+\sqrt{\mu^2+\delta \omega^2}$. This Hamiltonian indicate that for $c_1,c_2$, the microwave modulation has a component that induces an exchange interaction $c_1 c_2^\ad+c_1^\ad c_2$ and a component that induces a frequency modulation $c_2^\ad c_2-c_1^\ad c_1$.
As for the bright mode pair case we discussed above, i.e. for a small bias $\mu\gg\delta \omega$, $a_1,a_2$ are nearly degenerate and in $c_1, c_2$ basis,
\[
c_1\approx -\frac {1} {\sqrt{2}} (a_1+a_2)
\]
\[
c_2\approx \frac {1}{ \sqrt{2}} (a_2-a_1)
\]
whereas in the case of a birght-dark optical mode pair, i.e. the bias voltage is high $\mu \ll \delta \omega$, $a_1,a_2$ have a large frequency difference and in $c_1,c_2$ basis, $c_1$ is composed by a large part of $a_1$ with small part of $a_2$ and $c_2$ has large part of $a_2$ while has only small part of $a_1$. That means:
\[
c_1\approx -\left(a_1+\frac{|\mu|}{2\delta \omega}a_2\right)
\]
\[
c_2\approx \left(a_2-\frac{|\mu|}{2\delta \omega}a_1\right)
\]
Here the conversion term is finite and is suppressed by a factor $\sin \theta=\frac{\mu}{\delta \omega}$ and the term $\Omega \cos (\omega_m t) \cos\theta (c_2^\ad c_2-c_1^\ad c_1)$ is large, meaning that $c_1,c_2$ are being frequency modulated. However, Since the modulation frequency is orders of magnitudes larger than the bandwidth of the optical modes, we can neglect this term under high-$Q$ approximation. The resulting Hamiltonian still has a similar form to the bright mode pairs, with a pre-factor $\sin{\theta}$ in conversion efficiency.
\begin{equation}
\begin{split}
\hat{H}_\textrm{dark}=& \omega_1 c_1^\ad c_1+\omega_2 c_2^\ad c_2 +\Omega \cos (\omega_m t) \sin\theta (c_1 c_2^\ad+c_1^\ad c_2)
\end{split}
\label{eqn:dk}
\end{equation}
Similar to the previous section, applying RWA and input-output theory to equation \ref{eqn:dk}, we obtain
\[
\begin{split}
\dot c_1=&(-\i \omega_1 - \frac{\gamma_1}{2})c_1 -\i \frac{\Omega}{2} \sin \theta c_2- \sqrt{\gamma_{\textrm{ex1}}} s_\textrm{in} e^{-\i \omega_L t}
\end{split}
\]
\[
\begin{split}
\dot c_2=&(-\i \omega_2 - \frac{\gamma_2}{2})c_2 -\i \frac{\Omega}{2} \sin \theta c_1- \sqrt{\gamma_{\textrm{ex2}}} s_\textrm{in} e^{-\i \omega_L t}
\end{split}
\]
where $\gamma_{1,2}=\gamma_i+\gamma_\textrm{ex1,2}, \gamma_{\textrm{ex1}}=\gamma_\textrm{ex} \cos^2\frac \theta 2, \gamma_{\textrm{ex2}}=\gamma_\textrm{ex}\sin^2\frac \theta 2$, $\gamma_i$ is the internal loss of each ring and $\gamma_\textrm{ex}$ is the external loss of ring 1 to waveguide, and $\gamma_{ex1}, \gamma_{ex2}$ are the external losses of modes $c_1,c_2$ to waveguide. From these equations of motion, we see that at large bias voltages, $\gamma_{ex2}\rightarrow 0$ and the access of mode $c_2$ to the external waveguide is effectively controlled by the microwave drive $\Omega$. At the same time, the life time for mode $c_2$ becomes closer to that of the intrinsic cavity life time. Therefore, $\pi$-pulse control sequences can be implemented as described in the main text to achieve the on-demand photon-storage and retrieval.
The bright mode can be over-coupled to the optical waveguide while the dark mode is detuned far from the bright optical mode. As an example, assuming a lithium niboate ring resonator can be fabricated close to its material loss limit ($Q\sim10^9$), set $\delta\omega=20$ GHz, $\mu=1$ GHz so $\theta=0.05$, and let $\gamma_\textrm{ex1}=2$ GHz, one could achieve a tunable delay of more than 800 ns with a minimum of 500 ps steps (limited by output coupling bandwidth) and over $95\%$ efficiency.
\end{document}
|
\section*{Introduction}
\subsection*{Motivation}
Exel's original tight groupoid construction in \cite{Exel2008} produces an ample (i.e. totally disconnected \'etale) groupoid from an inverse semigroup. Since then, this general construction has proved extremely successful in producing groupoid models for a vast array of C*-algebras. The only caveat here is that C*-algebras coming from ample groupoids inevitably have plenty of projections. Our ultimate goal is to produce combinatorial groupoid models even for projectionless C*-algebras which means we must first find a non-ample generalisation of the tight groupoid construction.
\begin{rmk*}
This task seems all the more urgent given that projectionless C*-algebras like the Jiang-Su and Jacelon-Razak algebras have come to the fore in other contexts, namely the classification program for C*-algebras. Whether these C*-algebras have any groupoid models at all was only resolved recently in \cite{DeeleyPutnamStrung2015} and \cite{AustinMitra2018}. The hope would be that an even better picture of these C*-algebras could be provided by more combinatorial groupoid models obtained from a generalised tight groupoid construction. One might even hope that important C*-algebra properties (e.g. strong self-absorption) could be detected at the inverse semigroup level.
\end{rmk*}
\subsection*{Outline}
In Exel's original tight groupoid construction, the elements of the inverse semigroup get represented as compact open bisections in the resulting groupoid. To produce more general (potentially non-ample) groupoids we should of course take the inverse semigroup elements to represent more general (potentially non-compact) open bisections. However, we still require some information about compactness in order to produce locally compact groupoids.
As in domain theory, the key idea is to obtain this information from a (potentially non-reflexive) relation $\prec$ representing `compact containment'. In compact Hausdorff spaces, this is the same as `closed containment', so in this case
\[p\prec q\qquad\text{represents}\qquad\overline{p}\subseteq q.\]
\begin{rmk*}
We will define several other abstract relations but, as above, in order to provide some intuition, we will say what each relation informally `represents' for subsets of compact Hausdorff spaces.
\end{rmk*}
At first we put the inverse semigroup structure to one side and focus just on the relation $\prec$. In \autoref{Preliminaries}, we take $\prec$ to be a binary relation on an arbitrary set $P$, from which we define cover relations $\mathrel{\mathsf{D}}$ and $\mathrel{\mathsf{C}}$. We then consider centred subsets and finish with a selection principle extending K\"onig's Lemma in \autoref{SelectionPrinciple}.
In \autoref{Pseudobases}, we move on to the topological part of our construction. With just the single weak interpolation assumption in \eqref{Shrinking}, we are able to show that the tight spectrum always produces a locally compact space in \autoref{TightLocallyCompact}. Moreover, this yields a duality between abstract and concrete `pseudobases' of locally compact Hausdorff spaces, as we show in \autoref{abstract->concrete} and \autoref{TopologyRecovery}.
In \autoref{InverseSemigroups}, we extend this to inverse semigroups, showing how the the algebraic structure passes to the topological space. This yields a duality between pseudobasic inverse semigroups and \'etale pseudobases of locally compact Hausdorff \'etale groupoids, as shown in \autoref{LCHEgroupoid} and \autoref{GroupoidRecovery}.
\subsection*{Background}
We have taken inspiration from a number sources, notably domain theory (see \cite{GierzHofmannKeimelLawsonMisloveScott2003} and \cite{Goubault2013}), set theory (see \cite{Kunen2011}), point-free topology (see \cite{PicadoPultr2012}) and its non-commutative generalisations (see \cite{Resende2007}, \cite{LawsonLenz2013} and \cite{KudryavtsevaLawson2017}). We also discovered that the earlier parts of our theory were developed independently in some of Exel's unpublished notes, namely \cite{Exel2012} and \cite{Exel2012b} (kindly provided to us by the author after presenting our work at the UFSC Operator Algebra Seminar in August 2018), as we mention at relevant points below.
Probably the best way of putting our construction in context is to view it as a generalisation of various Stone type dualities, as summarised in the diagram below. Note each duality is connected above to all those dualities it generalises and is expressed in the following form:
\vspace{10pt}
\begin{center}
Author (Year) [Paper]\\
Abstract Structure\\
Concrete (Pseudo)basis\\
Topological Space/Groupoid\\
\end{center}
\vspace{10pt}
E.g. the regular ${}^{c\circ}$-$\overline{\cup}^\circ$-$\cap$-bases below considered by De Vries refer to bases of regular open sets (i.e. $O=\overline{O}^\circ$) that are closed under regular complements (i.e. $O^{c\circ}$), regular pairwise unions (i.e. $\overline{O\cup N}^\circ$) and pairwise intersections (i.e. $O\cap N$). Also LCLH stands for `locally compact locally Hausdorff' (and for compact or Hausdorff spaces we remove the corresponding L).
\begin{rmk*}
Our construction can also be viewed as a generalization of a classical set theoretic construction of a Boolean algebra from a poset (see \cite[Lemma III.4.8]{Kunen2011} and \cite[\S7]{BiceStarling2016}). Specifically, given any poset $(P,\leq)$, the regular open sets in the Alexandroff topology form a Boolean algebra and, moreover, $P$ is dense in the subalgebra it generates, which can then be identified with its (totally disconnected) Stone space. The topological part of our construction extends this from posets $(P,\leq\nolinebreak)$ to abstract pseudobases $(P,\prec)$ where $\prec$ need not be reflexive and the resulting spaces are no longer necessarily totally disconnected.
\end{rmk*}
\null
\vspace{10pt}
\hspace{-25pt}\begin{tikzpicture}[scale=3]
\node (Stone) at (0,0) [align=center]{Stone (1936) \cite{Stone1936}\\ Boolean Algebras\\ All Clopen Subsets\\ CH Totally Disconnected Spaces};
\node (De Vries) at (-1.5,-1) [align=center]{De Vries (1962) \cite{DeVries1962}\\ Compingent Algebras\\ Regular ${}^{c\circ}$-$\overline{\cup}^\circ$-$\cap$-Bases\\ CH Spaces};
\node (Wallman) at (0,-1) [align=center]{Wallman (1938) \cite{Wallman1938}\\ Normal Lattices\\ $\cup$-$\cap$-Bases\\ CH Spaces};
\node (Lawson) at (1.5,-1) [align=center]{Lawson (2012) \cite{Lawson2012}\\ Boolean Inverse Semigroups\\ All Compact Open Bisections\\ LCH Ample Groupoids};
\node (Shirota) at (-1.5,-2) [align=center]{Shirota (1952) \cite{Shirota1952}\\ R-Lattices\\ Regular $\overline{\cup}^\circ$-$\cap$-Bases\\ LCH Spaces};
\node (BiceStarling) at (1.5,-2) [align=center]{B.-Starling (2018) \cite{BiceStarling2018}\\ Basic Inverse Semigroups\\ \'Etale $\cup$-Bases\\ LCLH \'Etale Groupoids};
\node (BiceStarling2) at (-.5,-3) [align=center]{[This Paper]\\ Pseudobasic Inverse Semigroups\\ \'Etale Pseudobases\\ LCH \'Etale Groupoids};
\node (BiceStarling3) at (1.5,-3) [align=center]{[Next Paper]\\ Quasibasic Ordered Groupoids\\ \'Etale Quasibases\\ LCLH \'Etale Groupoids};
\draw (Stone)--(De Vries)--(Shirota)--(BiceStarling2);
\draw (Shirota)--(BiceStarling3);
\draw (Stone)--(Wallman)--(BiceStarling)--(BiceStarling3);
\draw (Wallman)--(BiceStarling2);
\draw (Lawson)--(BiceStarling2);
\draw (Stone)--(Lawson)--(BiceStarling);
\end{tikzpicture}
\vspace{10pt}
One key difference we should point out is that, while the other dualities above are first order, this is no longer the case with the (generalised) tight groupoid construction. This is because we must often work with subsets rather than single elements. However, for the most part we can work with finite subsets so the theory could be expressed in terms of omitting types (as in \cite[\S5]{BiceStarling2016}, which the present paper generalises), which makes it first order in a weak sense. In particular, the abstract pseudobases we deal with would still provide a natural way of interpreting locally compact spaces in forcing extensions (Wallman's normal lattices were used for this purpose instead in \cite{Kubis2014}).
In contrast, the frames/locales/quantales appearing in point-free topology and their non-commutative generalizations are undeniably second order, making vital use of infinite joins. Even the finite joins and the corresponding distributivity required in \cite{Lawson2012} and \cite{BiceStarling2018} restricts the freedom one has in making combinatorial constructions. Thus, while these various point-free counterparts of \'etale groupoids are attractive from an abstract theory point of view, it is rather the (generalised) tight groupoid construction that is more suitable for creating specific examples.
\begin{rmk*}
A similar phenomenon occurs in set theory, where forcing is usually done with combinatorial posets rather than Boolean algebras, even though they produce the same extensions.
\end{rmk*}
\subsection*{Acknowledgements}
We would like to thank Gilles de Castro and Ruy Exel for some very useful comments and questions (see \autoref{CastroQuestion} and \autoref{ExelQuestion}) after presenting our work at the UFSC operator algebra seminar in August 2018.
\section{Preliminaries}\label{Preliminaries}
\begin{center}
\textbf{Throughout we assume $\prec$ is a binary relation on a set $P$}.
\end{center}
First let us make some general notational conventions.
\begin{dfn}
For any $Q\subseteq P$ we define
\[Q^\prec=\{p\in P:\exists q\in Q\ (q\prec p)\}.\]
We extend $\prec$ to a binary relation on $\mathcal{P}(P)=\{Q:Q\subseteq P\}$ by defining
\[Q\prec R\qquad\Leftrightarrow\qquad Q\subseteq R^\succ\qquad\Leftrightarrow\qquad\forall q\in Q\ \exists r\in R\ (q\prec r).\]
\end{dfn}
Note this convention makes $\prec$ `auxiliary' to $\subseteq$ on $\mathcal{P}(P)$, i.e.
\[Q\subseteq Q'\prec R'\subseteq R\qquad\Rightarrow\qquad Q\prec R.\]
This is also the same convention used in \cite[Definition 4.1]{Exel2012} and \cite[\S3]{Lenz2008} (at least for filters, just in reverse). Also note above that we are using the standard convention of writing $\succ$ for the opposite relation of $\prec$, i.e. $p\succ q$ means $q\prec p$. For subsets, however, $R\succ Q$ is not the same as $Q\prec R$ \textendash\, the former means $R\subseteq Q^\prec$ while the latter means $Q\subseteq R^\succ$.
\subsection{Cover Relations}
\begin{dfn}\label{Covers}
We define the relations $\mathrel{\mathsf{D}}$ and $\mathrel{\mathsf{C}}$ on $\mathcal{P}(P)$ by
\begin{align}
\label{DenseCover}\tag{Dense Cover}Q\mathrel{\mathsf{D}} R\qquad&\Leftrightarrow\qquad\forall q\prec Q\ \exists r\prec R\ (r\prec q).\\
\label{CompactCover}\tag{Compact Cover}Q\mathrel{\mathsf{C}} R\qquad&\Leftrightarrow\qquad\exists\text{ finite }F\ (Q\mathrel{\mathsf{D}} F\prec R).
\end{align}
\end{dfn}
Essentially the same relations appear in \cite[Definition 4.2]{Exel2012}. Note that we again immediately obtain the following auxiliarity properties.
\begin{align}
\label{Daux}Q\subseteq Q'\mathrel{\mathsf{D}} R'\subseteq R\qquad&\Rightarrow\qquad Q\mathrel{\mathsf{D}} R.\\
Q\subseteq Q'\mathrel{\mathsf{C}} R'\subseteq R\qquad&\Rightarrow\qquad Q\mathrel{\mathsf{C}} R.
\end{align}
Also note the dense cover relation $\mathrel{\mathsf{D}}$ can be expressed more concisely by
\begin{equation}\label{Ddef2}
Q\mathrel{\mathsf{D}} R\qquad\Leftrightarrow\qquad Q^\succ\succ R^\succ,
\end{equation}
which just means $Q^\succ\subseteq R^{\succ\prec}$. Informally, $Q\mathrel{\mathsf{D}} R$ is saying that $R$ is dense in $Q$ while $Q\mathrel{\mathsf{C}} R$ is saying that the $Q$ is compactly contained in $R$, i.e.
\begin{align*}
Q\mathrel{\mathsf{D}} R\qquad&\text{represents}\qquad\bigcup Q\subseteq\overline{\bigcup R}.\\
Q\mathrel{\mathsf{C}} R\qquad&\text{represents}\qquad\overline{\bigcup Q}\subseteq\bigcup R.
\end{align*}
(For more precise statements, see \autoref{RDQ}, \autoref{Crep}, \eqref{QDR} and \eqref{QCR}).
Note here it is important that we consider the elements of $P$ to represent \emph{non-empty} sets, i.e. $P$ will \emph{not} have a minimum $0$ (i.e. satisfying $0^\prec=P$) except in trivial cases. Indeed, in \autoref{InverseSemigroups} when we consider an inverse semigroup $S$ with $0$, we will take $P=S\setminus\{0\}$. However, $\mathcal{P}(P)$ certainly has a minimum with respect to $\subseteq$, namely the empty set $\emptyset$, which we immediately see is also a minimum with respect to $\prec$, $\mathrel{\mathsf{D}}$ and $\mathrel{\mathsf{C}}$ on $\mathcal{P}(P)$, i.e. $\emptyset\prec Q$, $\emptyset\mathrel{\mathsf{D}} Q$ and $\emptyset\mathrel{\mathsf{C}} Q$, for all $Q\subseteq P$.
\begin{rmk}
Alternatively, one could assume that $P$ always has a minimum $0$ and then modify the definitions of $\mathrel{\mathsf{D}}$ and $\mathrel{\mathsf{C}}$ accordingly. However, we would then have to say `non-zero' all the time, e.g. `for all non-zero $p$...' or `there exists non-zero $p$...', which can get somewhat tiresome.
\end{rmk}
\begin{rmk}
In the classical case when $\prec$ is a preorder $\leq$, $\mathrel{\mathsf{D}}$ is essentially the same as the Lenz arrow relation in \cite{Lawson2012}, originally from \cite{Lenz2008}. Equivalently, $Q\mathrel{\mathsf{D}} R$ is saying that $R$ is an `outer cover' of $Q^\geq$ in sense of \cite{ExelPardo2016}. There is also another more topological description of $\mathrel{\mathsf{D}}$, which provides extra motivation for calling it the `dense cover relation' (in fact the `dense' terminology is already used for singleton sets in \cite[Definition 11.10]{Exel2008}). Specifically, $Q\mathrel{\mathsf{D}} R$ means that $Q^\geq\cap R^\geq$ is dense in $Q^\geq$ in the Alexandroff topology, where the open sets are precisely the lower sets. Also, if $R\leq Q$ then, in set theoretic terminology, $Q\mathrel{\mathsf{D}} R$ means `$R$ is predense in $Q^\geq$' \textendash\, see \cite[Definition III.3.58]{Kunen2011}.
\end{rmk}
\begin{prp}\label{DCproperties}
If $\prec$ is transitive on $P$, $\mathrel{\mathsf{D}}$ and $\mathrel{\mathsf{C}}$ are transitive on $\mathcal{P}(P)$ and
\begin{align}
\label{DC=>C}Q\mathrel{\mathsf{D}} S\mathrel{\mathsf{C}} R\qquad\Rightarrow\qquad Q\mathrel{\mathsf{C}} R\qquad&\Rightarrow\qquad Q\mathrel{\mathsf{D}} R\qquad\Leftrightarrow\qquad Q\mathrel{\mathsf{D}} R^\succ.\\
\label{Dcap}Q\mathrel{\mathsf{D}} Q'\quad\text{and}\quad R\mathrel{\mathsf{D}} R'\qquad&\Rightarrow\qquad Q^\succ\cap R^\succ\mathrel{\mathsf{D}} Q'^\succ\cap R'^\succ.\\
\label{Dcup}Q\mathrel{\mathsf{D}} Q'\quad\text{and}\quad R\mathrel{\mathsf{D}} R'\qquad&\Rightarrow\qquad Q\cup R\mathrel{\mathsf{D}} Q'\cup R'.\\
\label{Ccup}Q\mathrel{\mathsf{C}} Q'\quad\text{and}\quad R\mathrel{\mathsf{C}} R'\qquad&\Rightarrow\qquad Q\cup R\mathrel{\mathsf{C}} Q'\cup R'.
\end{align}
\end{prp}
\begin{proof}
If $Q\prec R\prec S$ then
\[Q\subseteq R^\succ\subseteq S^{\succ\succ}\subseteq S^\succ,\]
the middle $\subseteq$ following from $R\subseteq S^\succ$ and the last $\subseteq$ following from transitivity of $\prec$ on $P$. Thus $\prec$ is transitive on $\mathcal{P}(P)$ and hence $\mathrel{\mathsf{D}}$ is transitive, by \eqref{Ddef2}.
\begin{itemize}
\item[\eqref{DC=>C}] Consequently, $Q\mathrel{\mathsf{D}} S\mathrel{\mathsf{D}} F\prec R$ implies $Q\mathrel{\mathsf{D}} F\prec R$, so $Q\mathrel{\mathsf{D}} S\mathrel{\mathsf{C}} R$ implies $Q\mathrel{\mathsf{C}} R$. Also $Q\mathrel{\mathsf{D}} F\prec R$ implies $Q^\succ\succ F^\succ\subseteq R^{\succ\succ}\subseteq R^\succ$, i.e. $Q\mathrel{\mathsf{C}} R$ implies $Q\mathrel{\mathsf{D}} R$. Thus $Q\mathrel{\mathsf{C}} R\mathrel{\mathsf{C}} S$ implies $Q\mathrel{\mathsf{D}} R\mathrel{\mathsf{C}} S$ and hence $Q\mathrel{\mathsf{C}} S$, i.e. $\mathrel{\mathsf{C}}$ is also transitive. Finally note that if $Q\mathrel{\mathsf{D}} R$ then, for all $q\prec Q$, we have $r\prec q\prec Q$ with $r\prec R$. By transitivity, $r\prec Q$ so we can further take $r'\prec r$ (with $r'\prec R$), so $r'\prec q$ and $r'\in R^{\succ\succ}$, i.e. $Q\mathrel{\mathsf{D}} R$ implies $Q\mathrel{\mathsf{D}} R^\succ$.
\item[\eqref{Dcap}] Assume $Q^\succ\succ Q'^\succ$ and $R^\succ\succ R'^\succ$. For any $p\in Q^\succ\cap R^\succ$, we have $p'\prec p$ with $p'\prec Q'$, as $Q^\succ\succ Q'^\succ$. Then $p'\prec p\prec R$ so $p'\prec R$, by the transitivity of $\prec$, and hence we have $p''\prec p'$ with $p''\prec R'$, as $R^\succ\succ R'^\succ$. Thus $p''\prec p'\prec Q'$ so $p''\prec Q'$ and $p''\prec p'\prec p$ so $p''\prec p$, again by the transitivity of $\prec$. Thus $p''\in R'^\succ\cap Q'^\succ$ and, arguing as above, we can further obtain $p'''\prec p''$ so $p'''\in(R'^\succ\cap Q'^\succ)^\succ$. Then
\[(R^\succ\cap Q^\succ)^\succ\subseteq(R^\succ\cap Q^\succ)\succ(R'^\succ\cap Q'^\succ)^\succ,\]
i.e. $(R^\succ\cap Q^\succ)^\succ\succ(R'^\succ\cap Q'^\succ)^\succ$, as required.
\item[\eqref{Dcup}] Simply note that $Q^\succ\succ Q'^\succ$ and $R^\succ\succ R'^\succ$ implies
\[(Q\cup R)^\succ=Q^\succ\cup R^\succ\succ Q'^\succ\cup R'^\succ=(Q'\cup R')^\succ.\]
\item[\eqref{Ccup}] Now $Q\mathrel{\mathsf{D}} F\prec Q'$ and $R\mathrel{\mathsf{D}} G\prec R'$ implies $Q\cup R\mathrel{\mathsf{D}} F\cup G\prec Q'\cup R'$.\qedhere
\end{itemize}
\end{proof}
\subsection{Round Subsets}
Other properties of $\mathrel{\mathsf{C}}$ and $\mathrel{\mathsf{D}}$ require an extra assumption.
\begin{dfn}\label{def:round}
We call $R\subseteq P$ \emph{round} if $R\succ R$, i.e.
\[\label{Round}\tag{Round}\forall r\in R\ \exists q\in R\ (q\prec r).\]
\end{dfn}
Equivalently, $R\subseteq P$ is round if iff $R$ has no strictly minimal elements (where $r\in R$ is strictly minimal if $q\not\prec r$, for any $q\in R$). The `round' terminology is common when talking about ideals, for example, in domain theory \textendash\, see \cite[Proposition 5.1.33]{Goubault2013} or \cite[Proposition III-4.3]{GierzHofmannKeimelLawsonMisloveScott2003}. As long as $P$ itself is round, $P\mathrel{\mathsf{D}} P$ and $\emptyset$ is the only (non-strict) minimum for $\mathrel{\mathsf{D}}$ and $\mathrel{\mathsf{C}}$.
\begin{prp}
Consider the following statements.
\begin{align}
\label{precMin}Q\prec\emptyset\qquad\Leftrightarrow\qquad Q=\emptyset.\\
\label{DMin}Q\mathrel{\mathsf{D}}\emptyset\qquad\Leftrightarrow\qquad Q=\emptyset.\\
\label{CMin}Q\mathrel{\mathsf{C}}\emptyset\qquad\Leftrightarrow\qquad Q=\emptyset.
\end{align}
In general, \eqref{precMin} always holds. If $P$ is round then \eqref{DMin} and \eqref{CMin} also hold.
\end{prp}
\begin{proof}\
\begin{itemize}
\item[\eqref{precMin}] $Q\prec\emptyset$ means $Q\subseteq\emptyset^\succ=\emptyset$ and hence $Q=\emptyset$.
\item[\eqref{DMin}] $Q\mathrel{\mathsf{D}}\emptyset$ means $Q^\succ\succ\emptyset^\succ=\emptyset$ and hence $Q^\succ=\emptyset$, by \eqref{precMin}. If $P$ is round then this means $Q=\emptyset$.
\item[\eqref{CMin}] $Q\mathrel{\mathsf{C}}\emptyset$ means $Q\mathrel{\mathsf{D}} F\prec\emptyset$ so $F=\emptyset$, by \eqref{precMin}, and hence $Q=\emptyset$, by \eqref{DMin}.\qedhere
\end{itemize}
\end{proof}
\begin{prp}
If $P$ is round and $\prec$ is transitive then
\begin{align}
\label{QsubR}Q\subseteq R\qquad&\Rightarrow\qquad Q\mathrel{\mathsf{D}} R.\\
\label{FprecQ}F\prec R\qquad&\Rightarrow\qquad F\mathrel{\mathsf{C}} R,\quad\text{for all finite }F\subseteq P.
\end{align}
\end{prp}
\begin{proof}\
\begin{itemize}
\item[\eqref{QsubR}] If $P$ is round then $Q\subseteq Q^{\succ\prec}$, for all $Q\subseteq P$. Replacing $Q$ with $R^\succ$ yields $R^\succ\subseteq R^{\succ\succ\prec}\subseteq R^{\succ\prec}$, as $R^\succ\subseteq R^{\succ\succ}$ because $\prec$ is transitive, i.e. $R\mathrel{\mathsf{D}} R$. Thus $Q\subseteq R\mathrel{\mathsf{D}} R$ implies $Q\mathrel{\mathsf{D}} R$, by \eqref{Daux}.
\item[\eqref{FprecQ}] By \eqref{QsubR}, $F\prec R$ implies $F\mathrel{\mathsf{D}} F\prec R$, which means $F\mathrel{\mathsf{C}} R$.\qedhere
\end{itemize}
\end{proof}
\subsection{Centred Subsets}
\begin{dfn}
We define the \emph{formal meet} $\widehat{Q}$ of any $Q\subseteq P$ by
\[\widehat{Q}=\bigcap_{q\in Q}q^\succ\]
We call $Q\subseteq P$ \emph{centred} if $\widehat{F}\neq\emptyset$, for all finite $F\subseteq Q$.
\end{dfn}
The term `centred' comes from \cite[Definition III.3.23]{Kunen2011}. It will be convenient to also define centred versions of the relations we have considered so far.
\begin{dfn}
For any relation $\sqsubset$ on $\mathcal{P}(P)$, we define a centred version $\widehat{\sqsubset}$ by
\[Q\mathrel{\widehat{\sqsubset}}R\qquad\Leftrightarrow\qquad\exists\text{ finite }F\subseteq Q\ (\widehat{F}\mathrel{\sqsubset}R).\]
\end{dfn}
We are particularly interested in the centred compact cover relation $\mathrel{\widehat{\mathsf{C}}}$, which is related to the centred dense cover relation $\mathrel{\widehat{\mathsf{D}}}$ as before, i.e.
\[Q\mathrel{\widehat{\mathsf{C}}} R\qquad\Leftrightarrow\qquad\exists\text{ finite }F\ (Q\mathrel{\widehat{\mathsf{D}}} F\prec R).\]
We also have the following auxiliarity property (note the first relation is $\supseteq$ not $\subseteq$)
\[Q\supseteq Q'\mathrel{\widehat{\mathsf{C}}} R'\subseteq R\qquad\Rightarrow\qquad Q\mathrel{\widehat{\mathsf{C}}} R\qquad\Rightarrow\qquad Q\mathrel{\mathsf{C}} R,\]
at least when $Q\neq\emptyset$ (while $\emptyset\mathrel{\widehat{\mathsf{C}}} R$ means $\widehat{\emptyset}=P\mathrel{\mathsf{C}} R$). Also, by \eqref{CMin},
\[Q\text{ is centred}\qquad\Leftrightarrow\qquad Q\not\mathrel{\widehat{\mathsf{C}}}\emptyset,\]
at least when $P$ is round. If $Q$ is round then $Q\mathrel{\widehat{\mathsf{C}}} R$ also can be interpreted topologically as saying that the intersection of the sets represented by $Q$ (or their compact closures) are covered by $R$, i.e.
\[Q\mathrel{\widehat{\mathsf{C}}} R\qquad\text{represents}\qquad\bigcap Q\subseteq\bigcup R.\]
\begin{lem}\label{CentredDense}
If $\prec$ is transitive and $F$ is finite then
\[Q\mathrel{\widehat{\mathsf{D}}} F\quad\text{and}\quad Q\not\mathrel{\widehat{\mathsf{C}}} R\qquad\Rightarrow\qquad\exists f\in F\ (Q\cup\{f\}\not\mathrel{\widehat{\mathsf{C}}} R).\]
\end{lem}
\begin{proof}
Assume we have finite $G\subseteq Q$ with $\widehat{G}\mathrel{\mathsf{D}} F$ and, for all $f\in F$, we have $Q\cup\{f\}\mathrel{\widehat{\mathsf{C}}} R$, which means we have finite $H_f\subseteq Q$ and $I_f\prec R$ with $\widehat{H}_f\cap\widehat{f}\mathrel{\mathsf{D}} I_f$. We need to show that $Q\mathrel{\widehat{\mathsf{C}}} R$. Putting $H=\bigcup_{f\in F}H_f$ and $I=\bigcup_{f\in F}I_f$ we then have, for all $f\in F$,
\[\widehat{H}\cap\widehat{f}\ \subseteq\ \widehat{H}_f\cap\widehat{f}\ \mathrel{\mathsf{D}}\ I_f\ \subseteq\ I.\]
Now take any $p\in\widehat{G}\cap\widehat{H}$. As $p\in\widehat{G}\mathrel{\mathsf{D}} F$, we have $q\prec p$ and $f\in F$ with $q\prec f$. Then $q\in\widehat{H}\cup\widehat{f}\mathrel{\mathsf{D}} I$ so we have $r\prec q\prec p$ with $r\prec I$. As $p$ was arbitrary, $\widehat{G}\cap\widehat{H}\mathrel{\mathsf{D}} I\prec R$ so $\widehat{G}\cap\widehat{H}\mathrel{\mathsf{C}} R$ and hence $Q\mathrel{\widehat{\mathsf{C}}} R$.
\end{proof}
By \eqref{CMin}, we can apply this result with $R=\emptyset$ to obtain the following corollary (note here that $P$ has to be round, otherwise we could potentially have $F=\emptyset$).
\begin{cor}\label{CentredDenseCor}
If $P$ is round, $\prec$ is transitive and $F$ is finite then
\[Q\mathrel{\widehat{\mathsf{D}}} F\text{ and $Q$ is centred}\qquad\Rightarrow\qquad Q\cup\{f\}\text{ is centred, for some }f\in F.\]
\end{cor}
\subsection{Frink Filters}
Recall that $U\subseteq P$ is a \emph{filter} if
\[\tag{Filter}u,v\in U\qquad\Leftrightarrow\qquad\exists w\in U\ (u,v\succ w).\]
The $\Leftarrow$ part is saying $U$ is upwards closed while the $\Rightarrow$ part is saying $U$ is downwards directed. In contrast to Exel's construction, our tight subsets need not be filters, at least in the usual sense. However, they will satisfy the following weaker condition.
\begin{dfn}
We call $U\subseteq P$ a \emph{Frink filter} if
\[\label{FrinkFilter}\tag{Frink Filter}u\in U\qquad\Leftrightarrow\qquad\exists w\prec u\ (U\mathrel{\widehat{\prec}}w).\hspace{50pt}\]
\end{dfn}
The $\Leftarrow$ part of \eqref{FrinkFilter} can be stated more explicitly by saying $u\in U$ whenever $u$ is above some upper bound $w$ for a formal meet $\widehat{F}$, where $F$ is a finite subset of $U$. In particular, taking singleton $F$, we see that
\[U\text{ is a Frink Filter}\qquad\Rightarrow\qquad U\text{ is upwards closed},\]
as $u\succ w\in U$ implies $\widehat{w}=w^\succ\prec w\prec u$ and hence $u\in U$.
Likewise, the $\Rightarrow$ part of \eqref{FrinkFilter} is immediate if $U$ is round because then, for any $u\in U$, we have $w\prec u$ with $w\in U$ and $\widehat{w}=w^\succ\prec w$. Conversely, if $P$ satisfies \eqref{Interpolation},
\[U\text{ is a Frink filter}\qquad\Rightarrow\qquad U\text{ is round}.\]
Indeed, $u\in U$ implies $U\mathrel{\widehat{\prec}}w\prec u$, for some $w$, and \eqref{Interpolation} then yields $v$ with $w\prec v\prec u$. Thus $v\in U$, by the $\Leftarrow$ part of \eqref{FrinkFilter}.
If $P$ is round then, for every $p\in P$, we have $q\in P$ with $(\emptyset\prec)q\prec p$ and hence $P$ is the only possible non-centred Frink filter. Indeed, as long as $P$ itself is not centred (which will be true except in trivial cases \textendash\, see \autoref{S0tight} \eqref{Ptight} below),
\[U\text{ is a centred Frink filter}\qquad\Leftrightarrow\qquad U\text{ is a proper (i.e. $\neq P$) Frink filter}.\]
\begin{rmk}\label{Frink=>Filter}
For posets, Frink filters are dual to the Frink ideals defined in \cite[\S5 Definition]{Frink1954}. For $\wedge$-semilattices, Frink filters are just filters. Indeed, if $(P,\leq)$ is a $\wedge$-semilattice and $F=\{u,v\}\subseteq U$, for some Frink filter $U\subseteq P$, then
\[\widehat{F}=(u\wedge v)^\geq\leq u\wedge v\]
so $u,v\geq u\wedge v\in U$, i.e. $U$ is a filter. To see why Frink filters might be more natural for general posets, it is instructive to consider Exel's original motivation for introducing tight filters. First note that when $B$ is a Boolean algebra, the ultrafilters(=maximal proper filters) are closed in $\mathcal{P}(B)$ when we identify each subset of $B$ with its characteristic function in $\{0,1\}^B$ (with its usual product topology). Exel observed that ultrafilters are no longer closed when $B$ is an arbitrary $\wedge$-semilattice. To obtain a compact space, one therefore needs to consider the closure of the ultrafilters, i.e. the tight filters. But when $B$ is an arbitrary poset, even the filters may not be closed and this is why one must consider more general Frink filters. So as we get more general, this shift can be represented as
\begin{align*}
\text{Ultrafilters}\qquad&\rightarrow\qquad\text{Tight Filters}&&\rightarrow\qquad\text{Tight Frink Filters}.\\
\text{Boolean Algebras}\qquad&\rightarrow\qquad\text{$\wedge$-Semilattices}&&\rightarrow\qquad\text{Posets}.\\
\text{\cite{Stone1936}}\qquad&\rightarrow\qquad\text{\cite{Exel2008}}&&\rightarrow\qquad\text{\cite{BiceStarling2016}}.
\end{align*}
In fact, our tight subsets will automatically be Frink filters \textendash\, see \autoref{Tight=>Frink}.
\end{rmk}
\subsection{Disjoint Subsets}
\begin{dfn}
We define the \emph{disjoint relation} $\perp$ on $\mathcal{P}(P)$ by
\[\label{Disjoint}\tag{Disjoint}Q\perp R\qquad\Leftrightarrow\qquad Q^\succ\cap R^\succ=\emptyset.\]
We also define $Q^\perp=\{p\in P:Q\perp p\}$ so that we always have $Q\perp Q^\perp$ and
\[Q^\perp=\bigcap_{q\in Q}q^\perp.\]
\end{dfn}
As the the name suggests, again thinking of the elements of $P$ as open subsets,
\[Q\perp R\qquad\text{represents}\qquad(\bigcup Q)\cap(\bigcup R)=\emptyset\]
(see \eqref{QperpR} below). Note $P$ is round if and only if $\perp$ is irreflexive on $P$, i.e.
\[\label{Irreflexive}\tag{Irreflexive}\forall p\ (p\not\perp p).\]
Also $\mathrel{\mathsf{D}}$ could have equivalently been defined from $\perp$ as
\[Q\mathrel{\mathsf{D}} R\qquad\Leftrightarrow\qquad Q^\succ\cap R^\perp=\emptyset.\]
Moreover, just as $\mathrel{\mathsf{D}}$ is auxiliary to $\subseteq$ \textendash\, see \eqref{Daux} \textendash\, $\perp$ is auxiliary to $\mathrel{\mathsf{D}}$.
\begin{prp}
If $\prec$ is transitive then
\[\label{perpAuxiliary}\tag{$\perp$-Auxiliary}Q\mathrel{\mathsf{D}} Q'\perp R\qquad\Rightarrow\qquad Q\perp R.\]
\end{prp}
\begin{proof}
If we had $Q\not\perp R$, then we would have $q\in Q^\succ\cap R^\succ$. Then $Q\mathrel{\mathsf{D}} Q'$ would yield $q'\prec q$ with $q'\prec Q'$ and hence $q'\in Q'^\succ\cap R^\succ$, i.e. $Q'\not\perp R$.
\end{proof}
\subsection{A Selection Principle}
Later we will need the following selection principle to obtain an certain subset from a family of finite subsets (in fact, for the most part a weaker version which could be derived from K\"{o}nig's lemma would suffice, but the full strength version will be required once in the proof of \autoref{O_FSubset} below). Roughly speaking,
\begin{enumerate}
\item $\Delta$ represents the finite families of open sets with non-empty intersection.
\item $\Gamma$ represents a round collection of finite families whose unions have the FIP (i.e. finite subsets have non-empty intersection).
\end{enumerate}
The goal is to select one open set in each family in $\Gamma$ in a coherent way so that the chosen sets still form round subset with the finite intersection property.
\begin{lem}\label{SelectionPrinciple}
Say $\Delta$ and $\Gamma$ are collections of finite subsets of $P$ such that
\begin{gather}
\label{succClosed}\tag{$\succ$-Closed}d\succ D\in\Delta\qquad\Rightarrow\qquad\{d\}\cup D\in\Delta\\
\label{precRound}\tag{$\prec$-Round}\forall F\in\Gamma\ \exists G\in\Gamma\ (G\prec F).\\
\label{DeltaCentred}\tag{$\Delta$-Centred}\forall\text{ finite }\Phi\subseteq\Gamma\ \exists D\in\Delta\ \forall F\in\Phi\ (F\cap D\neq\emptyset).
\end{gather}
Then we have round $R\subseteq P$ such that
\begin{gather}
\label{FIP}\tag{$\Delta$-FIP}\forall\text{ finite }D\subseteq R\ (D\in\Delta).\\
\label{Selector}\tag{$\Gamma$-Selector}\forall F\in\Gamma\ (R\cap F\neq\emptyset).
\end{gather}
\end{lem}
\begin{proof}
First we prove some basic facts about $\Gamma$. To start with, say we had $F\in\Gamma$ and $G'\subseteq G\in\Gamma$ such that
\[F\prec G\setminus G'.\]
We claim that we can remove $G'$ from $G$, i.e. replace $G$ with $G\setminus G'$, without destroying the $\Delta$-centred property of $\Gamma$. To see this, take any finite $\Phi\subseteq\Gamma$. As $\Gamma$ is $\Delta$-centred, we have some $D\in\Delta$ such that $D\cap F\neq\emptyset\neq D\cap H$, for all $H\in\Phi$. As $F\prec G\setminus G'$, we have some $g\in G\setminus G'$ with $g\succ D\cap F$. Thus $D\cup\{g\}\in\Delta$, by \eqref{succClosed}, and $(D\cup\{g\})\cap(G\setminus G')\supseteq\{g\}\neq\emptyset$ as well. As $\Phi$ was arbitrary, this proves the claim.
Next, take disjoint subsets $F^+,F^-\subseteq F\in\Gamma$. We claim that we can remove at least one of these subsets from $F$ again without destroying the $\Delta$-centred property of $\Gamma$. If not then we could find finite $\Phi^+\subseteq\Gamma$ and $\Phi^-\subseteq\Gamma$ such that $\Phi^+\cup\{F\setminus F^+\}$ and $\Phi^-\cup\{F\setminus F^-\}$ are not $\Delta$-centred. This means that, for any $D\in\Delta$ such that $D\cap G\neq\emptyset$, for all $G\in\Phi^+$, we have $D\cap F\subseteq F^+$. Likewise, for any $D\in\Delta$ such that $D\cap G\neq\emptyset$, for all $G\in\Phi^-$, we have $D\cap F\subseteq F^-$. Thus, for any $D\in\Delta$ such that $D\cap G\neq\emptyset$, for all $G\in\Phi^+\cup\Phi^-$, we have $D\cap F\subseteq F^+\cap F^-=\emptyset$, contradicting the $\Delta$-centred property of $\Gamma$.
Now take another $G\in\Gamma$ with $G\prec F$ and set
\[G^+=\{g\in G:g\not\prec F\setminus F^+\}\qquad\text{and}\qquad G^-=\{g\in G:g\not\prec F\setminus F^-\}.\]
As $G\prec F=(F\setminus F^+)\cup(F\setminus F^-)$, it follows that $G^+$ and $G^-$ are also disjoint. Also note that $G^+$ and $G^-$ are the smallest subsets of $G$ such that
\[G\setminus G^+\prec F\setminus F^+\qquad\text{and}\qquad G\setminus G^-\prec F\setminus F^-.\]
Thus if we can remove $G_+$ from $G$ without destroying the $\Delta$-centred property of $\Gamma$, we can also remove $F_+$ from $F$, by what we proved above (and likewise for $G^-$).
With these basic properties out of the way, let us now index $\Gamma$ by some set $\Lambda$, i.e. let $\Gamma$ be $(F_\lambda)_{\lambda\in\Lambda}$. As $\Gamma$ is $\prec$-round, we also have a function $\theta$ on $\Lambda$ such that
\[F_{\theta(\lambda)}\prec F_\lambda,\]
for all $\lambda\in\Lambda$. Now consider the collection $\Theta_\Lambda$ of all $\Delta$-centred $\Lambda$-indexed families $(G_\lambda)_{\lambda\in\Lambda}$ of finite subsets of $P$ such that $G_{\theta(\lambda)}\prec G_\lambda$, for all $\lambda\in\Lambda$. Order $\Theta_\Lambda$ by
\[(G_\lambda)_{\lambda\in\Lambda}\leq(H_\lambda)_{\lambda\in\Lambda}\qquad\Leftrightarrow\qquad\forall\lambda\in\Lambda\ (G_\lambda\subseteq H_\lambda).\]
We claim that any decreasing (potentially transfinite) sequence in $\Theta_\lambda$
\[(G^1_\lambda)_{\lambda\in\Lambda}\geq(G^2_\lambda)_{\lambda\in\Lambda}\geq\ldots\]
has a lower bound in $\Theta_\Lambda$ defined by $G_\lambda=\bigcap_\alpha G^\alpha_\lambda$. To see this, note that for each $\lambda$, $G_\lambda$ and $G_{\theta(\lambda)}$ are finite so there must be some $\alpha$ with $G_\lambda=G^\alpha_\lambda$ and $G_{\theta(\lambda)}=G^\alpha_{\theta(\lambda)}$. As $(G^\alpha_\lambda)_{\lambda\in\Lambda}$ is in $\Theta_\Lambda$, it follows that $G_{\theta(\lambda)}\prec G_\lambda$. Likewise, for any finite $\Phi\subseteq\Lambda$, we must have $\alpha$ such that $G_\lambda=G^\alpha_\lambda$, for all $\lambda\in\Phi$. As $(G^\alpha_\lambda)_{\lambda\in\Lambda}$ is $\Delta$-centred, it follows that we can find $G\in\Delta$ such that $G_\lambda\cap G\neq\emptyset$, for all $\lambda\in\Phi$. This shows that $(G_\lambda)_{\lambda\in\Lambda}$ is also $\Delta$-centred and hence in $\Theta_\Lambda$.
Thus, by the Kuratowski-Zorn lemma, we have some minimal $(G_\lambda)_{\lambda\in\Lambda}$ below the given $(F_\lambda)_{\lambda\in\Lambda}$. As $(G_\lambda)_{\lambda\in\Lambda}$ is $\Delta$-centred, each $G_\lambda$ is non-empty. In fact, we claim that each $G_\lambda$ must be a singleton. If not then some $G_\gamma$ has disjoint non-empty subsets $G^+_\gamma$ and $G^-_\gamma$. Let $\theta^0(\gamma)=\gamma$ and $\theta^{n+1}(\gamma)=\theta(\theta^n(\gamma))$ and, for $*=+$ or $-$, recursively define
\[G^*_{\theta^{n+1}(\gamma)}=\{g\in G_{\theta^{n+1}(\gamma)}:g\not\prec G_{\theta^n(\gamma)}\setminus G^*_{\theta^n(\gamma)}\}.\]
As noted above, $G^+_{\theta^n(\gamma)}$ and $G^-_{\theta^n(\gamma)}$ are disjoint and
\begin{equation}\label{Removals}
(G_{\theta^{n+1}(\gamma)}\setminus G^*_{\theta^{n+1}(\gamma)})\prec(G_{\theta^n(\gamma)}\setminus G^*_{\theta^n(\gamma)}),
\end{equation}
so if we can remove $G^*_{\theta^{n+1}(\gamma)}$ from $G_{\theta^{n+1}(\gamma)}$ then we can also remove $G^*_{\theta^n(\gamma)}$ from $G_{\theta^n(\gamma)}$, without destroying the $\Delta$-centred property of $\Gamma$. But for each $n$, we must be able to remove $G^*_{\theta^n(\gamma)}$ from $G_{\theta^n(\gamma)}$ for either $*=+$ or $-$. Thus for either $*=+$ or $-$, we can in fact remove $G^*_{\theta^n(\gamma)}$ from $G_{\theta^n(\gamma)}$ for all $n$. Then the resulting $(G'_\lambda)_{\lambda\in\Lambda}$ will not only remain $\Delta$-centred but also still satisfy $G'_{\theta(\lambda)}\prec G'_\theta$, by \eqref{Removals}. But this means $(G'_\lambda)_{\lambda\in\Lambda}$ is in $\Theta_\Lambda$, even though it is strictly below $(G_\lambda)$, contradicting the minimality of $(G_\lambda)_{\lambda\in\Lambda}$. So each $G_\lambda$ must have been a singleton, proving the claim.
Finally, let $R=\bigcup_\lambda G_\lambda$, which certainly satisfies \eqref{Selector}. Also $R$ is round, as $(G_\lambda)_{\lambda\in\Lambda}$ consists of $\prec$-round singletons. Moreover, $R$ satisfies \eqref{FIP}, as $(G_\lambda)_{\lambda\in\Lambda}$ consists of $\Delta$-centred singletons.
\end{proof}
\section{Pseudobases}\label{Pseudobases}
\subsection{The Shrinking Condition}
From now on we will focus on transitive relations $\prec$ that make $P$ round. However, there is one more condition that we need to be able to define general locally compact Hausdorff spaces from $(P,\prec)$. Specifically, we need a condition which represents the ability to `shrink' a cover to another cover, each element of which is compactly contained in some element of the original cover.
\begin{dfn}
$(P,\prec)$ is an \emph{abstract pseudobasis} if $P$ is round, $\prec$ is transitive and
\[\label{Shrinking}\tag{Shrinking}p\prec q\qquad\Rightarrow\qquad p\mathrel{\mathsf{C}} q^\succ.\]
\end{dfn}
More explicitly, this condition is saying
\[\tag{Shrinking}p\prec q\qquad\Rightarrow\qquad\exists\text{ finite }F\ (p\mathrel{\mathsf{C}} F\prec q).\]
This makes it clear that \eqref{Shrinking} is a kind of weak interpolation condition.
\begin{rmk}\label{InterpolationRemark}
By \eqref{FprecQ}, \eqref{Shrinking} weakens the more standard interpolation condition considered in domain theory (see \cite[Definition I-1.17]{GierzHofmannKeimelLawsonMisloveScott2003}) given by
\[\label{Interpolation}\tag{Interpolation}p\prec q\qquad\Rightarrow\qquad\exists r\ (p\prec r\prec q).\]
In \cite[Definition 2.1 (v)]{Exel2012} and \cite[Definition 1.2 (v)]{Exel2012b}, \eqref{Interpolation} is also considered as a kind of regularity/normality condition. In particular, any poset becomes an abstract pseudobasis when we take $\prec\ =\ \leq$, as reflexive relations are trivially round and satisfy \eqref{Interpolation}.
However, it would seem that \eqref{Shrinking} is more appropriate for our work than \eqref{Interpolation}, as \eqref{Shrinking} applies to arbitrary (pseudo)bases of locally compact Hausdorff spaces, even when they are not closed under taking finite unions. The extra freedom coming from \eqref{Shrinking} will also be convenient for constructing specific examples. In any case, \eqref{Shrinking} suffices to guarantee that we have a large supply of tight subsets, as we will soon see.
\end{rmk}
First we note \eqref{Shrinking} extends to $\mathrel{\mathsf{C}}$, allowing us to replace $\mathrel{\mathsf{D}}$ with $\mathrel{\mathsf{C}}$ in \eqref{DC=>C}.
\begin{prp}
If $(P,\prec)$ is an abstract pseudobasis then
\begin{equation}\label{RCQ}\tag{$\mathsf{C}$-Shrinking}
R\mathrel{\mathsf{C}} Q\qquad\Rightarrow\qquad R\mathrel{\mathsf{C}} Q^\succ.
\end{equation}
\end{prp}
\begin{proof}
First note that, for all finite $F\subseteq P$, \eqref{Shrinking} and \eqref{Ccup} yields
\[F\prec Q\qquad\Rightarrow\qquad F\mathrel{\mathsf{C}} Q^\succ.\]
Thus if $R\mathrel{\mathsf{C}} Q$, i.e. $R\mathrel{\mathsf{D}} F\prec Q$, for some finite $F\subseteq P$, then $R\mathrel{\mathsf{D}} F\mathrel{\mathsf{C}} Q^\succ$ and hence $R\mathrel{\mathsf{C}} Q^\succ$, by \eqref{DC=>C}.
\end{proof}
\subsection{Tight Subsets}
\begin{center}
\textbf{From now on we assume $(P,\prec)$ is an abstract pseudobasis.}
\end{center}
Again recall that we are thinking of elements of $P$ as open subsets of a topological space. It should then be possible to recover the points of the space from their open neighbourhoods. The key thing to note here is that no neighborhood of a point $x$ can be covered by subsets which do not contain $x$. Moreover, in locally compact spaces every neighborhood of $x$ compactly contains a smaller neighbourhood of $x$. Thus the neighbourhoods of $x$ in $P$ are `tight' in the following sense.
\begin{dfn}\label{TightFilter}
We call $T\subseteq P$ \emph{tight} if $T\succ T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T$.
\end{dfn}
So $T$ is tight iff $T$ is round and, for all finite $F,G\subseteq P$,
\begin{equation}\label{Tight-Round}
F\subseteq T\quad\text{and}\quad G\prec P\setminus T\qquad\Rightarrow\qquad\widehat{F}\cap G^\perp\neq\emptyset
\end{equation}
(because $\widehat{F}\cap G^\perp=\emptyset$ would imply $\widehat{F}\mathrel{\mathsf{D}} G\prec P\setminus T$ and hence $\widehat{F}\mathrel{\mathsf{C}} P\setminus T$).
\begin{prp}\label{Tight=>Frink}
If $T\subseteq P$ is tight then $T$ is a Frink filter.
\end{prp}
\begin{proof}
The $\Rightarrow$ part of \eqref{FrinkFilter} is immediate because $T$ is round. Conversely, if $T\mathrel{\widehat{\prec}}v\prec u$ then $T\mathrel{\widehat{\mathrel{\mathsf{D}}}}v\prec u$ so $T\mathrel{\widehat{\mathrel{\mathsf{C}}}}u$. Thus $u\in T$, as $T\not\mathrel{\widehat{\mathrel{\mathsf{C}}}}P\setminus T$.
\end{proof}
In particular, any tight $T$ is upwards closed and thus $T=T^\prec$, as $T$ is also round.
\begin{rmk}
If $(P,\prec)=(S\setminus\{0\},\leq)$, for some $\wedge$-semilattice $S$ with minimum $0$, then every tight subset is a filter, by \autoref{Frink=>Filter} and \autoref{Tight=>Frink}. In this case, our tight subsets are precisely the tight filters defined in \cite{ExelPardo2016}. This becomes clear from the following general characterisation of tight subsets, as \eqref{Tight'} below extends \cite[(2.10)]{ExelPardo2016}.
\end{rmk}
\begin{prp}\label{tightChars}
If $T\neq P$ then $T$ is tight iff either of the following holds.
\begin{align}
\label{Tight}\tag{Tight}t\in T\quad&\Leftrightarrow\quad\exists\text{ finite }F\subseteq T\ \exists\text{ finite }G\prec P\setminus T\ (\widehat{F}\cap G^\perp\mathrel{\mathsf{C}} t).\\
\label{Tight'}\tag{Tight$'$}\emptyset\neq H\cap T\quad&\Leftrightarrow\quad\exists\text{ finite }F\subseteq T\ \exists\text{ finite }G\prec P\setminus T\ (\widehat{F}\cap G^\perp\mathrel{\mathsf{C}} H).
\end{align}
\end{prp}
\begin{proof}
If $T$ is round then, for any $t\in T$, we have $s\in T$ with $s\prec t$ and hence $\widehat{s}\cap\emptyset^\perp\mathrel{\mathsf{C}} t$. If $T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T$ and we have finite $F\subseteq T$ and $G\prec P\setminus T$ with $\widehat{F}\cap G^\perp\mathrel{\mathsf{C}} H$ then we have finite $I$ with $\widehat{F}\cap G^\perp\mathrel{\mathsf{D}} I\prec H$ and hence $\widehat{F}\mathrel{\mathsf{D}} G\cup I$. If we had $H\cap T=\emptyset$ then this would imply $\widehat{F}\mathrel{\mathsf{D}} G\cup I\prec P\setminus T$ and hence $T\mathrel{\widehat{\mathsf{C}}} P\setminus T$, a contradiction, so $H\cap T\neq\emptyset$. Thus if $T$ is tight then $T$ satisfies \eqref{Tight'}. Also \eqref{Tight'} implies \eqref{Tight}, as the latter is just the former restricted to singleton $H$.
Now say $T$ satisfies \eqref{Tight}. For any $t\in T$, we have finite $F\subseteq T$ and $G\prec P\setminus T$ with $\widehat{F}\cap G^\perp\mathrel{\mathsf{C}} t$. By \eqref{RCQ}, $\widehat{F}\cap G^\perp\mathrel{\mathsf{C}} t^\succ$ so we have finite $H,I\subseteq P$ with $\widehat{F}\cap G^\perp\mathrel{\mathsf{D}} I\prec H\prec t$ and hence $\widehat{F}\cap(G\cup I)^\perp=\emptyset\mathrel{\mathsf{C}} p$, for all $p\in P$. Then $H\subseteq P\setminus T$ would imply $G\cup I\prec P\setminus T$ and hence $T=P$, by \eqref{Tight}, contradicting our assumption. Thus $H\cap T\neq\emptyset$, which shows that $T$ is round.
Still assuming $T$ satisfies \eqref{Tight}, say we had $T\mathrel{\widehat{\mathsf{C}}} P\setminus T$, which means we have finite $F\subseteq T$ and $G\prec P\setminus T$ with $\widehat{F}\mathrel{\mathsf{D}} G$. Then $\widehat{F}\cap G^\perp=\emptyset\mathrel{\mathsf{C}} p$, for all $p\in P$, so $T=P$, by \eqref{Tight}, again contradicting our assumption. Thus $T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T$, showing that \eqref{Tight} does indeed imply that $T$ is tight.
\end{proof}
Incidentally, $P$ itself will only be tight in trivial cases. At the other extreme, it is more common for $\emptyset$ to be tight, although we will specifically omit $\emptyset$ from the tight spectrum considered in the next section (in order to be able to obtain locally compact rather than just compact spaces).
\begin{prp}\label{S0tight}\
\begin{enumerate}
\item\label{Ptight} $P$ is tight iff $P$ is centred, in which case $P$ is the only tight subset of $P$.
\item\label{0tight} $\emptyset$ is tight iff $P\not\mathrel{\mathsf{C}} P$.
\end{enumerate}
\end{prp}
\begin{proof}\
\begin{itemize}
\item[\eqref{Ptight}] By definition, $P$ is tight iff $P\not\mathrel{\widehat{\mathsf{C}}}\emptyset$. This means $\widehat{F}\not\mathrel{\mathsf{C}}\emptyset$, for all finite $F\subseteq P$, which is just saying that $P$ is centred. In particular, this implies $P\mathrel{\mathsf{D}} p$ and hence $P\mathrel{\mathsf{C}} p$, for any $p\in P$. Thus $T\mathrel{\widehat{\mathsf{C}}} P\setminus T$ whenever $T\neq P$ so there can be no other tight subsets.
\item[\eqref{0tight}] By definition, $\emptyset$ is tight iff $\emptyset\not\mathrel{\widehat{\mathsf{C}}} P$, which means $\widehat{\emptyset}=P\not\mathrel{\mathsf{C}} P$. \qedhere
\end{itemize}
\end{proof}
Thus, instead of $T\neq P$, we could assume $T$ is centred in \autoref{tightChars}, i.e.
\[T\text{ is tight}\qquad\Leftrightarrow\qquad T\text{ is centred and satisfies \eqref{Tight}}.\]
While round centred subsets need not be tight (or even upwards closed), they always have tight extensions. Indeed, by the Kuratowski-Zorn Lemma, any such subset is contained in one that is maximal among all round centred subsets. By the following result, these subsets will necessarily be tight and, in particular, Frink filters. Thus this result generalizes \cite[Proposition 12.7]{Exel2008}, which says that ultrafilters in $\wedge$-semilattices are necessarily tight.
\begin{prp}\label{maximalroundcentredtight}
If $U$ is round and centred then
\begin{align}
\label{UltraExists}U\text{ is maximal}\qquad&\Leftrightarrow\hspace{53pt}\forall a\prec P\setminus U\ (U\mathrel{\widehat{\perp}}a)\\
\label{UltraForall}&\Leftrightarrow\qquad\forall\text{ finite }G\prec P\setminus U\ (U\mathrel{\widehat{\perp}}G)\\
\label{TightForall}\Rightarrow\quad U\text{ is tight}\qquad&\Leftrightarrow\qquad\forall\text{ finite }G\prec P\setminus U\ (U\not\mathrel{\widehat{\mathsf{D}}} G).
\end{align}
\end{prp}
\begin{proof}\
\begin{itemize}
\item[\eqref{UltraExists}] Any round extension of $U$ must contain some $a\prec P\setminus U$. So if the right side of \eqref{UltraExists} holds then such an extension could not be centred, which implies that $U$ must be maximal. Conversely, say the right side of \eqref{UltraExists} fails, so we have some $a\prec P\setminus U$ with $\widehat{F}\not\perp a$, for all finite $F\subseteq U$, i.e. $U\cup\{a\}$ is centred. By \eqref{Shrinking}, we have a sequence $(F_n)$ of finite subsets of $P$ such that, for all $n$,
\[a\mathrel{\mathsf{C}} F_{n+1}\prec F_n\prec P\setminus U.\]
In particular, for all $n$, $a\mathrel{\mathsf{D}} F_n$ which implies that $U\cup\{f\}$ is centred, for some $f\in F_n$, by \autoref{CentredDenseCor} (with $Q=U\cup\{a\}$ and $F=F_n$). Thus we can apply \autoref{SelectionPrinciple}, taking $(F_n)$ for $\Gamma$ and setting
\[\Delta=\{D\subseteq P:D\text{ is finite and $U\cup D$ is centred}\},\]
to obtain round $T$ such that $U\cup T$ is centred and $T\cap F_n\neq\emptyset$, for all $n$ (actually $F'_n=\{f\in F_n:U\cup\{f\}\text{ is centred}\}$ defines a finitely branching $\omega$-tree so in this case we could also obtain $T$ from a simple application of K\"onig's lemma \textendash\, see \cite[Lemma III.5.6]{Kunen2011}). Thus $U\cup T^\prec$ is a round centred subset containing some element of $P\setminus U$ so $U$ could not have been maximal.
\item[\eqref{UltraForall}] We immediately see that \eqref{UltraForall} implies \eqref{UltraExists}. Conversely, if \eqref{UltraExists} holds and we are given finite $G\prec P\setminus U$ then, for each $g\in G$, we have finite $F_g\subseteq U$ with $\widehat{F}_g\perp g$. Taking $F=\bigcup_{g\in G}F_g$, we then have $\widehat{F}\perp G$.
\item[\eqref{UltraForall}$\Rightarrow$\eqref{TightForall}] If \eqref{UltraForall} holds and we are given $G\prec P\setminus U$ then we have finite $F\subseteq U$ with $\widehat{F}\perp G$. For any other finite $F'\subseteq U$, we have $\emptyset\neq\widehat{F\cup F'}\perp G$, as $U$ is centred, so $\widehat{F}'\not\mathrel{\mathsf{D}} G$.
\item[\eqref{TightForall}] Just note that the right side of \eqref{TightForall} means $U\not\mathrel{\widehat{\mathsf{C}}} P\setminus U$.\qedhere
\end{itemize}
\end{proof}
The following result shows that we can even be more selective about our tight extensions. For example, we might want to find a tight extension of some round $R$ but still avoid another given subset $S$. The following result (in the $Q=\emptyset$ case) shows that this can be done as long as $R\not\mathrel{\widehat{\mathsf{C}}} S$.
\begin{rmk}
Consequently, this result is essentially an extension of Birkhoff's prime ideal theorem for distributive lattices (see \cite[Lemma I-3.20]{GierzHofmannKeimelLawsonMisloveScott2003}). Indeed, if $(P,\prec)=(L\setminus\{0\},\leq)$, for some separative distributive lattice $L$ with minimum $0$, then the tight subsets are precisely the prime filters. In this case, $R\not\mathrel{\widehat{\mathsf{C}}} S$ is saying that the filter generated by $R$ is disjoint from the ideal generated by $S$. Birkhoff's theorem says that $R$ therefore extends to a prime filter disjoint from $S$.
\end{rmk}
\begin{thm}\label{TightStretching}
For any $Q,R,S\subseteq P$,
\[R\text{ is round and }Q\cup R\not\mathrel{\widehat{\mathsf{C}}} S\qquad\Rightarrow\qquad\exists\text{ tight }T\supseteq R\ (Q\cup T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T\supseteq S).\]
\end{thm}
\begin{proof}
As $Q\cup T\not\mathrel{\widehat{\mathsf{C}}} S$ depends only on the finite subsets of $T$, we can apply the Kuratowski-Zorn lemma to obtain maximal round $T\supseteq R$ with $Q\cup T\not\mathrel{\widehat{\mathsf{C}}} S$. Looking for a contradiction, assume that $Q\cup T\mathrel{\widehat{\mathsf{C}}} P\setminus T$, which means we have finite $F\subseteq Q\cup T$ and $G\prec P\setminus T$ with $\widehat{F}\mathrel{\mathsf{D}} G$. By \eqref{FprecQ} and \eqref{RCQ}, we have a sequence $(G_n)$ of finite subsets of $P$ such that
\[\widehat{F}\mathrel{\mathsf{D}} G\mathrel{\mathsf{C}} G_{n+1}\prec G_n\prec P\setminus U,\]
and hence $\widehat{F}\mathrel{\mathsf{D}} G_n$, for all $n$. Thus for each $n$, we have $g\in G_n$ with $Q\cup T\cup\{g\}\not\mathrel{\widehat{\mathsf{C}}} S$, by \autoref{CentredDense}. Thus we can apply \autoref{SelectionPrinciple}, taking $(G_n)$ for $\Gamma$ and setting
\[\Delta=\{D\subseteq P:D\text{ is finite and }Q\cup T\cup D\not\mathrel{\widehat{\mathsf{C}}} S\}\]
to obtain round $U$ such that $Q\cup T\cup U\not\mathrel{\widehat{\mathsf{C}}} S$ (actually, again K\"onig's lemma would suffice here). But then $U^\prec\cap P\setminus T\neq\emptyset$ so $T\cup U^\prec$ would be a proper round extension of $T$ with $Q\cup T\cup U^\prec\not\mathrel{\widehat{\mathsf{C}}} S$, contradicting maximality. Thus we must have indeed had $Q\cup T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T$. In particular, $T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T$ so $T$ is tight, as $T$ is also round. Lastly note that if we had $a\in T\cap S$ then, as $T$ is round, we would have $b\prec a$ with $b\in T\subseteq Q\cup T\not\mathrel{\widehat{\mathsf{C}}} S\ni a$ and hence $b\not\mathrel{\mathsf{C}} a$, contradicting \eqref{FprecQ}, so $S\subseteq P\setminus T$.
\end{proof}
\subsection{The Tight Spectrum}
\begin{center}
\textbf{Recall our standing assumption that $P$ is an abstract pseudobasis.}
\end{center}
\begin{dfn}
The \emph{tight spectrum} $\mathcal{T}(P)$ of $P$ is the set of non-empty tight subsets of $P$ with the topology given by the following basis, for finite $F,G\subseteq P$.
\[O_F^G=\{T\in\mathcal{T}(P):F\subseteq T\text{ and }G\mathrel{\mathsf{C}} P\setminus T\}.\]
\end{dfn}
\begin{rmk}
If we allowed $\emptyset$ to be part of the tight spectrum then we would always obtain compact rather than locally compact spaces. A more serious issue arises for inverse semigroups, where we need to avoid the empty subset to ensure that multiplying tight subsets yields a groupoid operation.
\end{rmk}
For convenience, we often omit empty sets and write $O_F=O_F^\emptyset$ and $O^G=O_\emptyset^G$. Incidentally, note that $O_\emptyset^\emptyset=\mathcal{T}(P)$, which of course would be open even if we did not include it in the basis. Also note that \eqref{Ccup} yields
\begin{equation}\label{OFGdef}
O_F^G=\bigcap_{f\in F}O_f\cap\bigcap_{g\in G}O^g
\end{equation}
so the given sets are closed under finite intersections and really do form a basis.
\begin{rmk}\label{CastroQuestion}
For anyone familiar with Stone duality or point-free topology, it might be tempting to consider the weaker topology on tight subsets generated just by the sets $O_f=\{T\in\mathcal{T}(P):f\in T\}$. The problem is that, for general abstract pseudobases, the resulting space may not be Hausdorff or even $T_1$, although it will still be locally compact and $T_0$. A standard way of strengthening such a topology to make it Hausdorff is to consider the patch topology as in \autoref{PseudoPatch} below. Indeed, combined with \autoref{abstract->concrete}, this shows that the resulting patch topology is precisely the topology obtained by adding the sets $O^g=\{T\in\mathcal{T}(P):g\mathrel{\mathsf{C}} P\setminus T\}$ to the subbasis. This answers a question posed by Gilles de Castro.
\end{rmk}
\begin{rmk}
Note $G\mathrel{\mathsf{C}} P\setminus T$ implies $G\subseteq P\setminus T$, but not conversely. That is unless $\prec$ is reflexive and hence a preorder \textendash\, then $G\mathrel{\mathsf{C}} P\setminus T$ is indeed the same as $G\subseteq P\setminus T$ and the topology in this case could be viewed as coming from the product topology on the characteristic functions of the tight subsets. Thus when $(P,\prec)=(S\setminus\{0\},\leq)$, for some $\wedge$-semilattice $S$ with minimum $0$, our tight spectrum agrees with Exel's original tight spectrum \textendash\, see \cite{ExelPardo2016}.
\end{rmk}
\begin{rmk}
There is also actually no need for $G$ above to be finite. Indeed, for any $Q\subseteq P$, the definition of $\mathrel{\mathsf{C}}$, \eqref{DC=>C} and \eqref{FprecQ} shows that $Q\mathrel{\mathsf{C}} P\setminus T$ is equivalent to $Q\mathrel{\mathsf{D}} G\mathrel{\mathsf{C}} P\setminus T$, for some finite $G$. Thus $O^Q$ is open as a union of open sets, i.e.
\[O^Q=\{T\in\mathcal{T}(P):Q\mathrel{\mathsf{C}} P\setminus T\}=\bigcup_{\substack{G\text{ is finite}\\\text{and }Q\,\mathrel{\mathsf{D}}\,G}}O^G.\]
\end{rmk}
Now we consider topological properties of the tight spectrum.
\begin{prp}\label{TightHausdorff}
The tight spectrum is Hausdorff.
\end{prp}
\begin{proof}
Take distinct $T,U\in\mathcal{T}(P)$. Thus we must have $t\in T\setminus U$ (or vice versa). As $T$ is round, we have $s\in T$ with $s\prec t$, which means $T\in O_s$ and $U\in O^s$. If $s\in V\in\mathcal{T}(P)$ then $V\not\mathrel{\widehat{\mathsf{C}}} P\setminus V$ and hence $s\not\mathrel{\mathsf{C}} P\setminus V$ so $O_s\cap O^s=\emptyset$.
\end{proof}
Note the above proof used the fact that $O^s_s=\emptyset$. More generally, for any finite $F,G\subseteq P$, we can characterize when $O_F^G=\emptyset$ as follows.
\begin{prp}\label{OFGempty}
$O_F^G=\emptyset$ iff every $\mathrel{\mathsf{C}}$-cover of $G$ is a $\mathrel{\mathsf{C}}$-cover of $\widehat{F}$, i.e.
\begin{equation}\label{C=>L}
G\mathrel{\mathsf{C}} H\qquad\Rightarrow\qquad\widehat{F}\mathrel{\mathsf{C}} H.
\end{equation}
\end{prp}
\begin{proof}
Assume $G\mathrel{\mathsf{C}} H$ and $\widehat{F}\not\mathrel{\mathsf{C}} H$. By \eqref{RCQ}, we have finite $G'$ with $G\mathrel{\mathsf{C}} G'\prec H$ and hence $\widehat{F}\not\mathrel{\mathsf{D}} G'$. This means we have $p$ with $F\succ p\perp G'$. As $P$ is round, we have a sequence $p=p_1\succ p_2\succ\ldots$. By the Kuratowski-Zorn lemma, this sequence extends to some maximal round centred $T$, necessarily with $F\subseteq T$ and $T\cap G'=\emptyset$. By \eqref{TightForall}, $T$ is tight so $T\in O_F^G$, proving the `only if' part.
Conversely, if $T\in O_F^G$ then $F\subseteq T$ and $G\mathrel{\mathsf{C}} P\setminus T$. As $T$ is tight, $F\subseteq T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T$ so $\widehat{F}\not\mathrel{\mathsf{C}} P\setminus T$, proving the `if' part.
\end{proof}
This has several important corollaries, e.g. we can generalise \cite[Theorem 12.9]{Exel2008} which says that ultrafilters are dense in the tight spectrum of a $\wedge$-semilattice.
\begin{cor}
The maximal round centred subsets of $P$ are dense in $\mathcal{T}(P)$.
\end{cor}
\begin{proof}
Any non-empty open set contains some non-empty $O_F^G$. Now just note that the $T\in O_F^G$ in the proof of \autoref{OFGempty} is a maximal round centred subset.
\end{proof}
\begin{cor}
For all $p,q\in P$
\begin{equation}\label{pperpq}
p\perp q\qquad\Leftrightarrow\qquad O_p\cap O_q=\emptyset.
\end{equation}
\end{cor}
\begin{proof}
If $p\perp q$ then no centred set can contain both $p$ and $q$ so $O_p\cap O_q=\emptyset$. Conversely, if $O_{p,q}^\emptyset=O_p\cap O_q=\emptyset$ then, as $\emptyset\mathrel{\mathsf{C}}\emptyset$, we must have $\widehat{\{p,q\}}\mathrel{\mathsf{C}}\emptyset$, by \autoref{OFGempty}. Thus $\widehat{\{p,q\}}=\emptyset$, by \eqref{CMin}, i.e. $p\perp q$.
\end{proof}
In particular, for any $p\in P$, we have $p\not\perp p$, as $P$ is round, so \eqref{pperpq} yields
\begin{equation}\label{Opnonempty}
O_p\neq\emptyset.
\end{equation}
\begin{cor}\label{FDG=>nonempty}
If $\widehat{F}\mathrel{\mathsf{D}} G$ then $O_F^G=\emptyset$.
\end{cor}
\begin{proof}
If $\widehat{F}\mathrel{\mathsf{D}} G\mathrel{\mathsf{C}} H$ then $\widehat{F}\mathrel{\mathsf{C}} H$, by \eqref{DC=>C}, so $O_F^G=\emptyset$, by \autoref{OFGempty}.
\end{proof}
Conversely, if $O_F^G=\emptyset$ and $P$ is a poset then $G$ itself is a $\mathrel{\mathsf{C}}$/$\mathrel{\mathsf{D}}$-cover of $G$ so $\widehat{F}\mathrel{\mathsf{D}} G$, by \autoref{OFGempty}.
\begin{xpl}
This no longer holds when $P$ is not a poset. For example, let $P=\mathbb{N}\cup\{a\}$ and let $\prec$ be $=$ on $\mathbb{N}$ while $n\prec a$ iff $n\neq2$ and $a\not\prec a$. Then $2\perp a$ and, in particular, $2\not\mathrel{\mathsf{D}} a$ even though $a$ has no $\mathrel{\mathsf{C}}$-cover and hence $O^a_{2}=\emptyset$.
\end{xpl}
We can now show that $\mathrel{\mathsf{D}}$ has the desired representation in $\mathcal{T}(P)$.
\begin{prp}
For any $F,Q\subseteq P$ with $F$ finite,
\begin{equation}\label{FDCQ}
\widehat{F}\mathrel{\mathsf{D}} Q\qquad\Leftrightarrow\qquad O_F\subseteq\overline{\bigcup_{q\in Q}O_q}.
\end{equation}
\end{prp}
\begin{proof}
If $\widehat{F}\not\mathrel{\mathsf{D}} Q$ then we have $p\in P$ with $F\succ p\perp Q$, which means $O_F\supseteq O_p$ and $O_p\cap\bigcup_{q\in Q}O_q=\emptyset$, by \eqref{pperpq}. By \eqref{Opnonempty}, $\emptyset\neq O_p\subseteq O_F\setminus\overline{\bigcup_{q\in Q}O_q}$ so $O_F\not\subseteq\overline{\bigcup_{q\in Q}O_q}$.
Conversely, if $O_F\not\subseteq\overline{\bigcup_{q\in Q}O_q}$ then we have $T\in O^G_H\subseteq O_F\setminus\overline{\bigcup_{q\in Q}O_q}$, for some finite $G,H\subseteq P$. As $G\mathrel{\mathsf{C}} P\setminus T$, \eqref{RCQ} yields $I$ with $G\mathrel{\mathsf{C}} I\prec P\setminus T$. As $T$ is tight, $F\cup H\subseteq T$ and $I\prec P\setminus T$, we have some $p\in P$ with $F\cup H\succ p\perp I$. Thus $O_p\subseteq O^G_H$ and hence $O_p\cap\bigcup_{q\in Q}O_q=\emptyset$ so $p\perp Q$, by \eqref{pperpq}. As we also have $F\succ p$, this shows that $\widehat{F}\not\mathrel{\mathsf{D}} Q$.
\end{proof}
\begin{cor}\label{RDQ}
For any $Q,R\subseteq P$,
\[R\mathrel{\mathsf{D}} Q\qquad\Leftrightarrow\qquad\bigcup_{r\in R}O_r\subseteq\overline{\bigcup_{q\in Q}O_q}.\]
\end{cor}
\begin{proof}
Note $R\mathrel{\mathsf{D}} Q$ iff $r\mathrel{\mathsf{D}} Q$, for all $r\in R$. By \eqref{FDCQ}, this is the same as saying $O_r\subseteq\overline{\bigcup_{q\in Q}O_q}$, for all $r\in R$, i.e. $\bigcup_{r\in R}O_r\subseteq\overline{\bigcup_{q\in Q}O_q}$.
\end{proof}
Next we wish to do the same for $\mathrel{\mathsf{C}}$. First we need the following.
\begin{prp}\label{TinClosureO_F}
For any finite $F\subseteq P$ and $T\in\mathcal{T}(P)$,
\begin{equation}\label{TinClosure}
T\in\overline{O_F}\qquad\Leftrightarrow\qquad T\cup F\not\mathrel{\widehat{\mathsf{C}}} P\setminus T.
\end{equation}
\end{prp}
\begin{proof}
Note $T\notin\overline{O_F}$ iff we have a basic neighbourhood of $T$ disjoint from $O_F$, i.e. iff we have finite $F'\subseteq T$ and $G\mathrel{\mathsf{C}} P\setminus T$ with $O_{F\cup F'}^G=O_F\cap O_{F'}^G=\emptyset$. By \eqref{C=>L}, this implies $\widehat{F'\cup F}\mathrel{\mathsf{C}} P\setminus T$ and hence $T\cup F\mathrel{\widehat{\mathsf{C}}} P\setminus T$.
Conversely, if $T\cup F\mathrel{\widehat{\mathsf{C}}} P\setminus T$ then we have finite $F'\subseteq T$ with $\widehat{F'\cup F}\mathrel{\mathsf{C}} P\setminus T$ and hence we have finite $G$ with $\widehat{F'\cup F}\mathrel{\mathsf{D}} G\prec P\setminus T$. So $G\mathrel{\mathsf{C}} P\setminus T$, by \eqref{FprecQ}, and $O_F\cap O_{F'}^G=O_{F\cup F'}^G=\emptyset$, by \autoref{FDG=>nonempty}. Thus $T\in O_{F'}^G$ and hence $T\notin\overline{O_F}$.
\end{proof}
Let $\Subset$ denote the compact containment relation, i.e.
\[O\Subset N\qquad\Leftrightarrow\qquad\exists\text{ compact }C\ (O\subseteq C\subseteq N).\]
So in Hausdorff spaces, $O\Subset N$ is just saying that $\overline{O}$ is compact and $\overline{O}\subseteq N$.
\begin{thm}\label{O_FSubset}
For any $F,Q\subseteq P$ with $F$ finite,
\begin{equation}\label{FCCQ}
\widehat{F}\mathrel{\mathsf{C}} Q\qquad\Leftrightarrow\qquad O_F\Subset\bigcup_{q\in Q}O_q.
\end{equation}
\end{thm}
\begin{proof}
Assume $F$ is finite and $\widehat{F}\mathrel{\mathsf{C}} Q$. Then, for any $T\in\mathcal{T}(P)\setminus\bigcup_{q\in Q}O_q$, we have $\widehat{F}\mathrel{\mathsf{C}} Q\subseteq P\setminus T$ so $T\notin\overline{O_F}$, by \autoref{TinClosureO_F}. Thus $\overline{O_F}\subseteq\bigcup_{q\in Q}O_q$.
To prove that $\overline{O_F}$ is compact, it suffices to show that every subbasic open cover has a finite subcover, by the Alexander subbasis theorem. So assume we have some $S\subseteq P$ and a collection $\mathcal{G}$ of finite subsets of $P$ such that no finite subcollection of $(O_s)_{s\in S}$ and $(O^G)_{G\in\mathcal{G}}$ covers $\overline{O_F}$ (in particular we must have $O_F\neq\emptyset$ and hence $\widehat{F}\neq\emptyset$). We can also assume that $\mathcal{G}$ contains some $G$ with $G\mathrel{\mathsf{C}} H$, for some finite $H$. Indeed, we have $G$ with $\widehat{F}\mathrel{\mathsf{C}} G\prec Q$ and hence $\emptyset\neq G\prec H\neq\emptyset$ for some finite $H\subseteq Q$. Thus $\overline{O_F}\cap O^G=O_F^G=\emptyset$, which means we could add $G$ to $\mathcal{G}$ and there would still be no finite subcollection covering $\overline{O_F}$.
We need to show that the entire collection of $(O_s)_{s\in S}$ and $(O^G)_{G\in\mathcal{G}}$ does not cover $\overline{O_F}$ either. For this we first apply the selection principle in \autoref{SelectionPrinciple} with
\begin{align*}
\Delta&=\{H\subseteq P:H\text{ is finite and }\widehat{H\cup F}\not\mathrel{\mathsf{C}} S\}.\\
\Gamma&=\{H\subseteq P:H\text{ is finite and }\mathcal{G}\ni G\mathrel{\mathsf{C}} H\}.
\end{align*}
We immediately see that $\Delta$ satisfies \eqref{succClosed}. Also $\Gamma$ satisfies \eqref{precRound}, by \eqref{RCQ}. Moreover, we can assume $\Gamma$ contains some non-empty $H$, as mentioned above. To see that \eqref{DeltaCentred} holds, take finite $\Phi\subseteq\Gamma$. For each $H\in\Phi$, we have $G_H\in\mathcal{G}$ with $G_H\mathrel{\mathsf{C}} H$. For every finite $I\subseteq S$, the fact that $\overline{O_F}$ has no finite cover in $(O_s)_{s\in S}$ and $(O^G)_{G\in\mathcal{G}}$ yields
\[T\in\overline{O_F}\setminus\bigcup_{s\in I}O_s\cup\bigcup_{H\in\Phi}O^{G_H}.\]
For each $H\in\Phi$, the fact that $T\notin O^{G_H}$ and $G_H\mathrel{\mathsf{C}} H$ means $H\not\subseteq P\setminus T$, i.e. $T\cap H\neq\emptyset$. Thus $D=\bigcup_{H\in\Phi}T\cap H$ satisfies $D\cap H\neq\emptyset$, for each $H\in\Phi$. By \eqref{TinClosure},
\[D\cup F\subseteq T\cup F\not\mathrel{\widehat{\mathsf{C}}} P\setminus T\supseteq I.\]
As this holds for all finite $I\subseteq S$, we have $\widehat{D\cup F}\not\mathrel{\mathsf{C}} S$ and hence $D\in\Delta$. This verifies \eqref{DeltaCentred}, as required.
Applying \autoref{SelectionPrinciple} we obtain $R\subseteq P$ such that $H\cap R\neq\emptyset$, for all $H\in\Gamma$, and $D\in\Delta$, for all finite $D\subseteq R$, which means $R\cup F\not\mathrel{\widehat{\mathsf{C}}} S$. By \autoref{TightStretching}, we have tight $T\supseteq R$ with $T\cup F\not\mathrel{\widehat{\mathsf{C}}} P\setminus T\supseteq S$. Thus $T\neq\emptyset$, as $\Gamma\neq\emptyset$, and $T\in\overline{O_F}$, by \eqref{TinClosure}. Moreover $T\notin\bigcup_{s\in S}O_s$, as $S\subseteq P\setminus T$, and $T\notin\bigcup_{G\in\mathcal{G}}O^G$, as $H\cap T\neq\emptyset$, for all $H$ with $G\mathrel{\mathsf{C}} H$, for some $G\in\mathcal{G}$. Thus $T\in\overline{O_F}\setminus(\bigcup_{s\in S}O_s\cup\bigcup_{G\in\mathcal{G}}O^G)$ and hence the entirety of $(O_s)_{s\in S}$ and $(O^G)_{G\in\mathcal{G}}$ does not cover $\overline{O_F}$ either. This shows that $\overline{O_F}$ is indeed compact and hence $O_F\Subset\bigcup_{q\in Q}O_q$.
Conversely, assume $\widehat{F}\not\mathrel{\mathsf{C}} Q$. If, moreover, $\widehat{F}\mathrel{\mathsf{C}} P$ then, by \eqref{RCQ}, we have a sequence $(G_n)$ of finite subsets of $P$ such that, for all $n$,
\[\widehat{F}\mathrel{\mathsf{C}} G_{n+1}\prec G_n\prec P.\]
Using \autoref{CentredDense} and \autoref{SelectionPrinciple} (or K\"onig's lemma), we obtain round $R$ with $F\cup R\not\mathrel{\widehat{\mathsf{C}}} Q$ and $R\cap G_n\neq\emptyset$, for each $n$. In particular, $R\neq\emptyset$. By \autoref{TightStretching}, we have tight $T\supseteq R$ with $F\cup T\not\mathrel{\widehat{\mathsf{C}}} P\setminus T\supseteq Q$. By \eqref{TinClosure}, this means $T\in\overline{O_F}\setminus\bigcup_{q\in Q}O_q$ and hence $O_F\not\Subset\bigcup_{q\in Q}O_q$.
On the other hand, if $\widehat{F}\not\mathrel{\mathsf{C}} P$ then $\overline{O_F}$ is not even compact so $O_F\not\Subset\bigcup_{q\in Q}O_q$. To see this, note that $(O_p)_{p\in P^\succ}$ covers the entire tight spectrum, as each $T\in\mathcal{T}(P)$ is non-empty and round. However, for any finite $G\prec P$, $\widehat{F}\not\mathrel{\mathsf{C}} P$ yields $\widehat{F}\not\mathrel{\mathsf{D}} G$ so we have $f$ with $F\succ f\perp G$. By \eqref{Opnonempty}, we have $T\in O_f\subseteq O_F$ and hence $T\notin\bigcup_{g\in G}O_g$, by \eqref{pperpq}. As $G$ was arbitrary, this shows $(O_p)_{p\in P^\succ}$ has no finite subcover of $O_F$.
\end{proof}
\begin{cor}\label{Crep}
For any $Q,R\subseteq P$,
\[R\mathrel{\mathsf{C}} Q\qquad\Leftrightarrow\qquad\bigcup_{r\in R}O_r\Subset\bigcup_{q\in Q}O_q.\]
\end{cor}
\begin{proof}
If $R\mathrel{\mathsf{C}} Q$ then $R\mathrel{\mathsf{D}} F\prec Q$, for some finite $F\subseteq P$. Then \eqref{RDQ} yields $\bigcup_{r\in R}O_r\subseteq\overline{\bigcup_{f\in F}O_f}$, while \eqref{FprecQ} and \eqref{FCCQ} yield $O_f\Subset\bigcup_{q\in Q}O_q$, for all $f\in F$. Thus $\overline{\bigcup_{f\in F}O_f}=\bigcup_{f\in F}\overline{O_f}$ is compact and $\bigcup_{r\in R}O_r\subseteq\overline{\bigcup_{f\in F}O_f}\subseteq\bigcup_{q\in Q}O_q$, i.e.
\[\bigcup_{r\in R}O_r\Subset\bigcup_{q\in Q}O_q.\]
Conversely, if $\bigcup_{r\in R}O_r\Subset\bigcup_{q\in Q}O_q$ then, for each $T\in\overline{\bigcup_{r\in R}O_r}\subseteq\bigcup_{q\in Q}O_q$, we have some $t\in T\cap Q^\succ$. Thus $\overline{\bigcup_{r\in R}O_r}\subseteq\bigcup_{t\in Q^\succ}O_t$ and, by compactness, we have some finite $F\prec Q$ with $\overline{\bigcup_{r\in R}O_r}\subseteq\bigcup_{f\in F}O_f$ and hence $\bigcup_{r\in R}O_r\subseteq\overline{\bigcup_{f\in F}O_f}$. Thus $R\mathrel{\mathsf{D}} F\prec Q$, by \eqref{RDQ}, i.e. $R\mathrel{\mathsf{C}} Q$.
\end{proof}
Recall that a Hausdorff space is \emph{locally compact} iff each point has a compact neighbourhood (which implies that each point actually has a neighbourhood base of compact sets and is thus consistent with \autoref{LCdefinitions} below).
\begin{cor}\label{TightLocallyCompact}
The tight spectrum is locally compact.
\end{cor}
\begin{proof}
Any $T\in\mathcal{T}(P)$ is non-empty and round, so we have $s,t\in T$ with $s\prec t$. By \eqref{FprecQ}, $s\mathrel{\mathsf{C}} t$ so $T\in O_s\Subset O_t$, by \autoref{O_FSubset}. In particular, $T$ has a compact neighbourhood so, as $T$ was arbitrary, $\mathcal{T}(P)$ is locally compact.
\end{proof}
For $\mathcal{T}(P)$ to be locally compact as above, \eqref{Shrinking} is crucial, as the following example shows (this answers a question posed by Exel).
\begin{xpl}\label{ExelQuestion}
Let $P$ be the collection of all dyadic open intervals in $\mathbb{R}$, i.e.
\[P=\{(\tfrac{k}{2^n},\tfrac{k+1}{2^n}):k,n\in\mathbb{Z}\}.\]
Considering $(P,\Subset)$, we see that $\Subset$ is transitive and $P$ is round but \eqref{Shrinking} fails so $(P,\Subset)$ is not an abstract pseudobasis. Moreover, $\mathcal{T}(P)$ can be identified with the complement $\mathbb{R}\setminus D$ of the dyadic numbers $D\subseteq\mathbb{R}$. As both $D$ and $\mathbb{R}\setminus D$ are dense in $\mathbb{R}$, all compact subsets of $\mathbb{R}\setminus D$ have empty interior and hence $\mathcal{T}(P)\approx\mathbb{R}\setminus D$ is not locally compact.
\end{xpl}
\subsection{Patch Topologies}
We are primarily concerned with pseudobases (to be introduced in the next section) in locally compact Hausdorff spaces. However, it will be useful to first consider more general pseudosubbases and how they relate to more general stably locally compact $T_0$ spaces via the patch construction.
First, for any topological space $X$, let $\mathcal{O}(X)$ denote the open subsets of $X$ and let $\mathcal{C}(X)$ denote the compact saturated subsets of $X$, i.e. the compact subsets $C\subseteq X$ that are also intersections of open sets so
\[C=\bigcap_{C\subseteq O\in\mathcal{O}(X)}O.\]
Note that if $X$ is Hausdorff or even $T_1$ then arbitrary subsets are saturated so $\mathcal{C}(X)$ in this case just consists of arbitrary compact subsets.
\begin{dfn}\label{LCdefinitions}
A topological space $X$ is
\begin{enumerate}
\item \emph{locally compact} if each point has a neighbourhood base of compact sets.
\item \emph{coherent} if $X$ is locally compact and $\mathcal{C}(X)$ is closed under $\cap$, i.e.
\[\tag{$\cap$-Closed}O,N\in\mathcal{C}(X)\quad\Rightarrow\quad O\cap N\in\mathcal{C}(X).\]
\item \emph{stably locally compact} if $X$ is coherent and, for $O\in\mathcal{O}(X)$ and $\mathcal{C}\subseteq\mathcal{C}(X)$
\[\label{WellFiltered}\tag{Well-Filtered}\bigcap\mathcal{C}\subseteq O\quad\Rightarrow\quad\exists\text{ finite }F\subseteq\mathcal{C}\ (\bigcap F\subseteq O).\]
\end{enumerate}
\end{dfn}
\begin{rmk}\label{StablyLocallyCompactSubset}
These are the standard definitions \textendash\, see \cite[Definitions 4.8.1, 5.2.21 and 8.3.31]{Goubault2013} and \cite[Definitions O-5.9, VI-6.2 and VI-6.7]{GierzHofmannKeimelLawsonMisloveScott2003}. Equivalent definitions can be stated using $\Subset$ on $\mathcal{O}(X)$ instead, specifically $X$ is
\begin{enumerate}
\item \emph{locally compact} if $\{O\in\mathcal{O}(X):x\in O\}$ is $\Subset$-round, for each $x\in X$.
\item \emph{stably locally compact} if, for $O,O',N,N'\in\mathcal{O}(X)$ and $\Subset$-round $\mathcal{O}\subseteq\mathcal{O}(X)$,
\begin{align*}
O'\Subset O\ \text{and}\ N'\Subset N\quad&\Rightarrow\quad O'\cap N'\Subset O\cap N,\quad\text{and}\\
\bigcap\mathcal{O}\subseteq O\quad&\Rightarrow\quad\exists\text{ finite }F\subseteq\mathcal{O}\ (\bigcap F\subseteq O).
\end{align*}
\end{enumerate}
(see \cite[Proposition VI-7.71]{GierzHofmannKeimelLawsonMisloveScott2003} and \cite[Theorem 8.3.34]{Goubault2013}).
Also, among locally compact $T_0$ spaces, the directed version of \eqref{WellFiltered} is equivalent to being sober \textendash\, see \cite[Propositions 8.3.5 and 8.3.8]{Goubault2013}.
\end{rmk}
In a stably locally compact space $X$, both $\mathcal{C}(X)$ and $\Subset$ behave like they would in a Hausdorff space. Indeed, in this case $\mathcal{C}(X)$ and $\Subset$ are stable under the patch construction, which makes $X$ Hausdorff, as long as $X$ was originally $T_0$.
Specifically, recall that the \emph{patch topology} of a topological space is the topological join of $\mathcal{O}(X)$ and $\mathcal{C}(X)^c$, i.e. the topology generated by all open sets and all complements of compact saturated sets (see \cite[Definition V-5.11]{GierzHofmannKeimelLawsonMisloveScott2003} or \cite[Definition 9.1.26]{Goubault2013}). We denote $X$ with its patch topology by $X^\mathrm{patch}$ so
\[\mathcal{O}(X^\mathrm{patch})=\mathcal{O}(X)\vee\mathcal{C}(X)^c.\]
\begin{prp}
If $X$ is a stably locally compact $T_0$ space then $X^\mathrm{patch}$ is a locally compact Hausdorff space such that $\mathcal{C}(X)\subseteq\mathcal{C}(X^\mathrm{patch})$ and, for all $O,N\in\mathcal{O}(X)$,
\begin{equation}\label{SubsetPatch}
O\Subset N\text{ in }X\qquad\Leftrightarrow\qquad O\Subset N\text{ in }X^\mathrm{patch}.
\end{equation}
\end{prp}
\begin{proof}
First note that the patch topology of any locally compact $T_0$ space is Hausdorff, by \cite[Proposition 9.1.12 and Lemma 9.1.31]{Goubault2013}.
For any $C,D\in\mathcal{C}(X)$, $C\cap D\in\mathcal{C}(X)$, as $X$ is coherent. Thus $C^\mathrm{patch}$, the patch topology coming from the subspace topology of $C$, coincides with the subspace topology coming from $X^\mathrm{patch}$. By \cite[Proposition 9.1.27]{Goubault2013}, $C^\mathrm{patch}$ is compact so $C$ is compact in $X^\mathrm{patch}$ and hence $C\in\mathcal{C}(X^\mathrm{patch})$, as $X$ is Hausdorff.
As $X$ is locally compact, every $x\in X$ has a neighbourhood in $\mathcal{C}(X)\subseteq\mathcal{C}(X^\mathrm{patch})$. As $X^\mathrm{patch}$ is Hausdorff, it follows that $X^\mathrm{patch}$ is also locally compact.
Now take any $O,N\in\mathcal{O}(X)$. If $O\Subset N$ in $X^\mathrm{patch}$ then this remains true in $X$ because any compact subset of $X^\mathrm{patch}$ is certainly compact in the coarser original topology of $X$. Conversely, if $O\Subset N$ in $X$ then we have compact $C$ in $X$ with $O\subseteq C\subseteq N$. But the saturation of $C$ is also compact, by \cite[Proposition 4.4.14]{Goubault2013}, and is still contained in the open set $N$. Thus we may take $C\in\mathcal{C}(X)\subseteq\mathcal{C}(X^\mathrm{patch})$ and hence $O\Subset N$ in $X^\mathrm{patch}$.
\end{proof}
Conversely, to produce general stably locally compact $T_0$ spaces from locally compact Hausdorff spaces, we consider pseudosubbases.
\begin{dfn}
We call $P\subseteq\mathcal{O}(X)$ a \emph{pseudosubbasis} of a space $X$ if
\begin{align}
\label{Separating'}\tag{Separating}&\forall x,y\in X\ \exists O\in P\ (x\neq y\quad\Rightarrow\quad x\notin O\ni y\ \ \text{or}\ \ y\notin O\ni x).\\
\label{capPointRound}\tag{$\cap$-Point-Round}&\forall\text{ finite }F\subseteq P\ \forall x\in\bigcap F\ \exists\text{ finite }G\subseteq P\ (x\in\bigcap G\Subset\bigcap F).
\end{align}
\end{dfn}
\begin{rmk}
Recall that a \emph{subbasis} $S$ of a space $X$ is a collection of open sets that generates the topology of $X$, i.e. such that every element of $\mathcal{O}(X)$ is a union of finite intersections of elements of $S$. In particular, every subbasis of a locally compact $T_0$ space is a pseudosubbasis. Conversely, if $X$ has a pseudosubbasis then $X$ must be both locally compact and $T_0$.
\end{rmk}
\begin{rmk}
By \eqref{Separating}, any pseudosubbasis $P$ must cover all except possibly one point of $X$, i.e. $|X\setminus\bigcup P|\leq 1$. Taking $F=\emptyset$ in \eqref{capPointRound} we see that, for any $x\in X=\bigcap\emptyset$, we have finite $G\subseteq P$ with $x\in\bigcap G\Subset X$. If $X$ is compact then we can always take $G=\emptyset$ too but otherwise $G$ must be non-empty, i.e. if $X$ is not compact then $P$ must truly cover $X$.
\end{rmk}
\begin{prp}
Any pseudosubbasis of a stably locally compact space $X$ is a pseudosubbasis of $X^\mathrm{patch}$.
\end{prp}
\begin{proof}
Immediate because $\Subset$ is the same on $\mathcal{O}(X)$ and $\mathcal{O}(X^\mathrm{patch})$, by \eqref{SubsetPatch}.
\end{proof}
In particular, any true subbasis of a stably locally compact $T_0$ space $X$ is a pseudosubbasis of the locally compact Hausdorff space $X^\mathrm{patch}$. Conversely, we now show that any pseudosubbasis of a locally compact Hausdorff space generates a stably locally compact $T_0$ topology from which the original topology can be recovered from the patch construction. Thus we have the following duality:
\begin{gather}
\nonumber\text{pseudosubbases of locally compact Hausdorff spaces}\\
\label{pseudosubDuality}\updownarrow\\
\nonumber\text{subbases of stably locally compact $T_0$ spaces}.
\end{gather}
Given $P\subseteq\mathcal{P}(X)$, let $X_P$ denote $X$ with the topology generated by $P$.
\begin{prp}\label{PseudoPatch}
If $P$ is a pseudosubbasis of a locally compact Hausdorff space $X$ then $X_P$ is a stably locally compact $T_0$ space such that $X=X_P^\mathrm{patch}$.
\end{prp}
\begin{proof}
First replace $P$ with its closure under finite intersections. Note $P$ is still a pseudosubbasis of $X$ and now also a basis of $X_P$.
We claim that $\mathcal{C}(X_P)\subseteq\mathcal{C}(X)$. To see this, take any $C\in\mathcal{C}(X_P)$ and $x\in X\setminus C$. As $C$ is saturated in $X_P$, we have $O\in\mathcal{O}(X_P)$ with $x\notin O\supseteq C$. For every $c\in C$, we have $N\in P$ with $c\in N\subseteq O$. As $C$ is compact in $X_P$, we thus have finite $F\subseteq P$ with $C\subseteq\bigcup F\subseteq O$. As $x$ was arbitrary, this shows that $C=\bigcap_{F\in\Gamma}\bigcup F$, where $\Gamma$ is the collection of all finite $F\subseteq P$ with $C\subseteq\bigcup F$. Again by compactness in $X_P$ and \eqref{capPointRound}, $\Gamma$ is $\Subset$-round (where $\Subset$ is defined from the original topology of $X$), i.e. for all $F\in\Gamma$, we have $G\in\Gamma$ with $G\subseteq F^\Supset$. As $X$ is Hausdorff this means $C=\bigcap_{F\in\Gamma}\bigcup F=\bigcap_{F\in\Gamma}\overline{\bigcup F}$, where the closure is taken in $X$. Thus $C$ is an intersection of compact subsets in the Hausdorff space $X$ so $C\in\mathcal{C}(X)$.
As in the proof of \eqref{SubsetPatch}, this means that, for any $O,N\in\mathcal{O}(X_P)$,
\begin{equation}
O\Subset N\text{ in }X_P\qquad\Leftrightarrow\qquad O\Subset N\text{ in }X.
\end{equation}
As stable local compactness can be characterised by $\Subset$ as in \autoref{StablyLocallyCompactSubset}, and $X$ is certainly stably locally compact (being Hausdorff), it follows that $X_P$ is also stably locally compact. By \eqref{Separating}, $X_P$ is also $T_0$.
Thus $X_P^\mathrm{patch}$ is Hausdorff, by \cite[Proposition 9.1.12 and Lemma 9.1.31]{Goubault2013}. As $\mathcal{C}(X_P)\subseteq\mathcal{C}(X)$, the identity $\mathbf{id}:X\rightarrow X_P^\mathrm{patch}$ is continuous. For any $x\in X$, \eqref{capPointRound} yields $O,N\in P$ with $x\in N\Subset O$. As $\overline{N}$ is compact and $X_P^\mathrm{patch}$ is Hausdorff, $\mathbf{id}$ takes closed subsets of $\overline{N}$ to closed subsets, i.e. $\mathbf{id}$ is a homeomorphism on $\overline{N}$ and hence on $N$. As $N$ is also open in $X_P^\mathrm{patch}$, $\mathbf{id}$ is a local homeomorphism. But $\mathbf{id}$ is clearly bijective and thus a global homeomorphism.
\end{proof}
\begin{rmk}
There is also an ordering induced on $X$ by $P$, namely
\[x\leq y\qquad\Leftrightarrow\qquad\forall O\in P\ (x\in O\ \Rightarrow\ y\in O).\]
As long as $P$ is a pseudosubbasis of a locally compact Hausdorff space, one can verify that $\leq$ is closed in $X\times X$ and thus turns $X$ into a pospace. Moreover, the open up-sets in $X$ are then precisely the open sets in $X_P$ (and $\leq$ is just the specialisation order in $X_P$). By \eqref{capPointRound}, the open up-sets $N^\uparrow_x$ containing any $x\in X$ will also be $\Subset$-round. Thus an alternative way of viewing \eqref{pseudosubDuality} is
\begin{gather*}
\text{locally compact pospaces such that each $N_x^\uparrow$ is $\Subset$-round}\\
\updownarrow\\
\text{stably locally compact $T_0$ spaces}.
\end{gather*}
This generalises the duality between compact pospaces and stably compact spaces (see \cite[Proposition 9.1.27 and 9.1.34]{Goubault2013} and \cite[Proposition VI-6.23]{GierzHofmannKeimelLawsonMisloveScott2003}).
\end{rmk}
\subsection{Concrete Pseudobases}\label{ConcretePseudobases}
We have just seen in \autoref{PseudoPatch} that the topology of any locally compact Hausdorff space $X$ can be recovered from any pseudosubbasis. However, this requires that we already know the points of $X$. What we would like to do is recover both the points and the topology from the relational structure $(P,\Subset)$, for suitable $P\subseteq\mathcal{O}(X)$. Pseudosubbases are too general for this as the following elementary examples show.
\begin{xpl}
Consider the following possibilities.
\begin{align*}
X&=\{x,y\}&P&=\{\emptyset,\{x\},\{y\},\{x,y\}\}.\\
X&=\{x,y,z\}&P&=\{\emptyset,\{x\},\{y\},\{x,y,z\}\}.
\end{align*}
In both cases, $P$ is a (even $\cap$-closed) pseudosubbasis for $X$, considered as a discrete space. Moreover, the resulting relational structures $(P,\Subset)$ are immediately seen to be isomorphic, even though the spaces are certainly not homeomorphic, having different numbers of points.
\end{xpl}
This leads us to consider the following stronger structures. Note \eqref{Dense} is the key extra condition distinguishing pseudobases from pseudosubbases.
\begin{dfn}
We call $P\subseteq\mathcal{O}(X)$ a \emph{(concrete) pseudobasis} of a space $X$ if
\begin{align}
\label{Cover}\tag{Cover}&\text{Every $x\in X$ is contained in some $O\in P$}.\\
\label{Separating}\tag{Separating}&\text{The subsets in $P$ distinguish the points of $X$}.\\
\label{PointRound}\tag{Point-Round}&\text{The neighborhoods in $P$ of any fixed $x\in X$ are $\Subset$-round}.\\
\label{Dense}\tag{Dense}&\text{Every non-empty $O\in\mathcal{O}(X)$ contains some non-empty $N\in P$}.
\end{align}
\end{dfn}
Our primary goal in this section is to exhibit a duality between abstract and concrete pseudobases. First, however, we make some general observations.
To begin with, pseudobases are indeed generalisations of bases, at least when $X$ is locally compact and $T_0$. Indeed, bases of locally compact spaces satisfy \eqref{PointRound}, while bases of $T_0$ spaces satisfy \eqref{Separating} and, by definition, bases of arbitrary spaces satisfy \eqref{Cover} and \eqref{Dense}.
\begin{xpl}
Consider the one-point compactification $X=\mathbb{N}\cup\{\infty\}$ of $\mathbb{N}$ and let $P=\{\{n\}:n\in\mathbb{N}\}\cup\{X\}$. It is immediately verified that $P$ is a pseudobasis but not a basis, as it does not contain a neighbourhood base of the point $\infty$.
\end{xpl}
\begin{rmk}
Our previous notion of a pseudobasis from \cite[Definition 8.2]{BiceStarling2016} contained all the above axioms except \eqref{PointRound}. Indeed, \eqref{PointRound} holds trivially for the compact pseudobases considered in \cite{BiceStarling2016}. One immediate consequence of \eqref{PointRound} worth noting is that $\subseteq$ coincides with the lower order defined from $\Subset$, i.e. for all $O,N\in P$,
\[\tag{Lower Order}O\subseteq N\qquad\Leftrightarrow\qquad\forall M\in P\ (M\Subset O\ \ \Rightarrow\ \ M\Subset N).\]
\end{rmk}
\begin{rmk} there are several equivalent ways of stating the pseudobasis axioms. For example, $P\subseteq\mathcal{O}(X)$ is a pseudobasis if and only if the following are satisfied.
\begin{align}
\label{Cover+Round}\tag{Cover+Point-Round}O&=\bigcup_{P\ni N\Subset O}N,\text{ for all }O\in P\cup\{X\}.\\
\label{Dense'}\tag{Dense}\overline{O}&=\overline{\bigcup_{P\ni N\Subset O}N},\text{ for all }O\in\mathcal{O}(X).\\
\tag{Separating}\{x\}&=\bigcap_{x\in O\in P}O\cap\bigcap_{x\notin O\in P}X\setminus O,\text{ for all }x\in X.
\end{align}
We will primarily be concerned with pseudobases of non-empty sets. In this case, we can set $N_x=\{O\in P:x\in O\}$ and concisely state the pseudobasis axioms as
\begin{align}
\tag{Non-Empty Open+Dense}&\emptyset\notin P\subseteq\mathcal{O}(X)\subseteq P^\Subset.\\
\tag{Cover+Separating+Round}&\emptyset\neq N_x\neq N_y\subseteq N_y^\Subset,\quad\text{for all distinct $x,y\in X$}.
\end{align}
\end{rmk}
Now for the promised duality between abstract and concrete pseudobases. First we represent abstract pseudobases as concrete pseudobases.
\begin{prp}\label{abstract->concrete}
If $(P,\prec)$ is an abstract pseudobasis then $(O_p)_{p\in P}$ is a concrete pseudobasis of $\mathcal{T}(P)$.
\end{prp}
\begin{proof}\
\begin{itemize}
\item[\eqref{Cover}] As every $T\in\mathcal{T}(P)$ is non-empty, we have some $t\in T$ and hence $T\in O_t$.
\item[\eqref{PointRound}] If $T\in O_t$ then $t\in T$ so, as $T$ is round, we have $s\in T$ with $s\prec t$ and hence $T\in O_s\Subset O_t$, by \eqref{FprecQ} and \autoref{O_FSubset}.
\item[\eqref{Dense}] If $T\in O_F^G$ then $F\subseteq T$ and $G\mathrel{\mathsf{C}} S\setminus T$. By \eqref{RCQ}, we have $H$ with $G\mathrel{\mathsf{C}} H\prec P\setminus T$. By \eqref{Tight-Round}, we have $p$ with $F\succ p\perp H$ so $O_p\subseteq O_F^G$. As the $(O_F^G)$ form a basis for $\mathcal{T}(P)$, this shows that any non-empty open $O\subseteq\mathcal{T}(P)$ contains some $O_p$, which is non-empty by \eqref{Opnonempty}.
\item[\eqref{Separating}] If $T,U\in\mathcal{T}(P)$ are distinct then we have some $t\in T\setminus U$ or $t\in U\setminus T$, i.e. $U\notin O_t\ni T$ or $T\notin O_t\ni U$ so $O_t$ distinguishes $T$ from $U$. \qedhere
\end{itemize}
\end{proof}
Next we show how concrete pseudobases yield abstract pseudobases. Note the following extends \cite[Theorem 8.4]{BiceStarling2016} to non-0-dimensional spaces.
\begin{thm}\label{TopologyRecovery}
If $P\not\ni\emptyset$ is a pseudobasis of $X$ and we take $\prec\ =\ \Subset$ then
\begin{align}
\label{QperpR}Q\perp R\qquad&\Leftrightarrow\qquad(\bigcup Q)\cap(\bigcup R)=\emptyset.\\
\label{QDR}Q\mathrel{\mathsf{D}} R\qquad&\Leftrightarrow\qquad\bigcup Q\subseteq\overline{\bigcup R}.\\
\intertext{If $X$ is also Hausdorff then $P$ is an abstract pseudobasis with}
\label{QCR}Q\mathrel{\mathsf{C}} R\qquad&\Leftrightarrow\qquad\bigcup Q\Subset\bigcup R
\end{align}
and $x\mapsto T_x=\{O\in P:x\in O\}$ is a homeomorphism from $X$ onto $\mathcal{T}(P)$.
\end{thm}
\begin{proof}\
\begin{itemize}
\item[\eqref{QperpR}] If $(\bigcup Q)\cap(\bigcup R)=\emptyset$ then certainly $Q^\succ\cap R^\succ\subseteq Q^\supseteq\cap R^\supseteq=\emptyset$. If $(\bigcup Q)\cap(\bigcup R)\neq\emptyset$ then we must have $O\cap N\neq\emptyset$, for some $O\in Q$ and $N\in R$. Then \eqref{Dense} and \eqref{PointRound} yields $M\in P$ with $M\Subset O\cap N$ so $O\not\perp N$.
\item[\eqref{QDR}] If $Q\not\mathrel{\mathsf{D}} R$ then we have $O\in Q^\succ$ with $R\perp O$. Thus $O\subseteq\bigcup Q$ and $O\cap\bigcup R=\emptyset=O\cap\overline{\bigcup R}$ so $\bigcup Q\not\subseteq\overline{\bigcup R}$. Conversely, if $\bigcup Q\not\subseteq\overline{\bigcup R}$ then $N\setminus\overline{\bigcup R}\neq\emptyset$, for some $N\in Q$. Then \eqref{Dense} and \eqref{PointRound} yields $O\in P$ with $O\Subset N\setminus\bigcup R$ so $Q\not\mathrel{\mathsf{D}} R$.
\item[\eqref{QCR}] If $G$ is Hausdorff and $\bigcup Q\Subset\bigcup R$ then $\overline{\bigcup Q}$ is compact and $\overline{\bigcup Q}\subseteq\bigcup R$. Thus, for each $g\in\overline{\bigcup Q}$, we have $O\in R$ with $g\in O$. Then \eqref{PointRound} yields $N\in P$ with $g\in N\Subset O$. As $\overline{\bigcup Q}$ is compact we can therefore cover it with finite $F\subseteq O^\succ\subseteq R^\succ$ so $Q\mathrel{\mathsf{D}} F\prec R$, i.e. $Q\mathrel{\mathsf{C}} R$. Conversely, if $Q\mathrel{\mathsf{D}} F\prec R$ then $\overline{\bigcup{Q}}\subseteq\overline{\bigcup{F}}=\bigcup_{O\in F}\overline{O}\subseteq\bigcup R$, as $F$ is finite. Also $\bigcup_{O\in F}\overline{O}$ is compact, as a finite union of compact sets, and hence $\bigcup Q\Subset\bigcup R$.
\end{itemize}
Certainly $\Subset$ is transitive and $P$ is round, by \eqref{PointRound}. Moreover, as above we see that whenever $O\Subset N$ we can find $F$ with $O\mathrel{\mathsf{C}} F\prec N$, so $P$ also satisfies \eqref{Shrinking} and is thus an abstract pseudobasis.
Now each $T_x$ is non-empty and round by \eqref{Cover} and \eqref{PointRound}. We claim each $T_x$ is also tight. Indeed, for any finite $F\subseteq T_x$ and $H\prec P\setminus T_x$, note that $O\in H$ means $O\Subset N\not\ni x$, for some $N\in P$, and hence $x\notin\overline{O}$, as $X$ is Hausdorff. Thus $x\in\bigcap F\setminus\overline{\bigcup H}$ so we have some $O\in P$ with $O\Subset\bigcap F\setminus\overline{\bigcup H}$, by \eqref{Dense}, thus verifying \eqref{Tight-Round}.
Thus $x\mapsto T_x$ maps $X$ to $\mathcal{T}(P)$. By \eqref{Separating}, $x\mapsto T_x$ is injective. To see that it is also surjective, take $T\in\mathcal{T}(P)$ and note that
\[C=\bigcap T\setminus\bigcup(P\setminus T)=\bigcap_{\substack{O\in T\\ N\prec P\setminus T}}\overline{O}\setminus N,\]
as $T$ is round and $P$ is point-round (so that we can replace $N\in P\setminus T$ with $N\prec P\setminus T$ at the end). As $T$ is tight, this last collection of sets has the finite intersection property. As it also consists of compact closed subsets, the entire intersection must also be non-empty and hence $x\in C$, for some $x\in G$. By the definition of $C$, if $O\in T$ then $x\in\bigcap T\subseteq O$ so $O\in T_x$, while if $O\in P\setminus T$ then $x\notin\bigcup P\setminus T\supseteq O$ so $O\notin T_x$. Thus $T=T_x$, showing that $x\mapsto T_x$ is indeed surjective.
Then we immediately see that $x\mapsto T_x$ is a homeomorphism with respect to the topologies on $X$ and $\mathcal{T}(P)$ generated by $P$ and $(O_p)_{p\in P}$ respectively. Thus $x\mapsto T_x$ is also a homeomorphism with respect to the corresponding patch topologies, which are just the original topologies of $X$ and $\mathcal{T}(P)$, by \autoref{PseudoPatch} and \autoref{abstract->concrete}.
\end{proof}
In summary, a concrete pseudobasis of a locally compact Hausdorff space always yields an abstract pseudobasis. Moreover, if we represent this abstract pseudobasis as a concrete pseudobasis of its tight spectrum, we recover the original pseudobasis.
However, if we represent a given abstract pseudobasis $(P,\prec)$ as a concrete pseudobasis $((O_p)_{p\in P},\Subset)$ then we lose $\prec$, as $\Subset$ corresponds instead to the weaker relation $\mathrel{\mathsf{C}}$ on $P$. We could even have $O_p=O_q$ for distinct $p,q\in P$. If we want a true duality then we should really restrict to the abstract pseudobases with $\prec\ =\ \mathrel{\mathsf{C}}$ on $P$. More explicitly, these are the \emph{separative} pseudobases, i.e. satisfying
\[\label{Separative}\tag{Separative}p\not\prec q\qquad\Rightarrow\qquad\forall\text{ finite }F\prec q\ \exists r\prec p\ (r\perp F).\]
Thus the duality we obtain can be represented as
\[\text{separative abstract pseudobases}\qquad\leftrightarrow\qquad\text{pseudobases of Hausdorff spaces}.\]
\begin{rmk}
When $\prec$ is reflexive it suffices to consider $F=q$ in \eqref{Separative}, in which case this reduces to the usual notion of a separative poset (see \cite[Definition III.4.10]{Kunen2011}) and $(P,\mathrel{\mathsf{C}})$ is the separative quotient of $(P,\prec)$. In general, if $(P,\prec)$ is an abstract pseudobasis then $(P,\mathrel{\mathsf{C}})$ is again an abstract pseudobasis, which can be seen by going through the representation above or more directly using the basic properties in \autoref{DCproperties}.
\end{rmk}
\section{Inverse Semigroups}\label{InverseSemigroups}
Next we examine what happens when our abstract pseudobases have some extra inverse semigroup structure. In this case, each tight subset will be a coset which will naturally give the tight spectrum an extra \'etale groupoid structure. First, however, we need some general results about inverse semigroups and their cosets.
Recall that an \emph{inverse semigroup} (see \cite{Lawson1998}) is a semigroup $S$ in which every $s\in S$ has a unique generalised inverse $s^{-1}$, i.e. satisfying
\[ss^{-1}s=s\qquad\text{and}\qquad s^{-1}ss^{-1}=s^{-1}.\]
Actually, uniqueness here is equivalently to saying that the idempotents
\[E=\{s\in S:s=s^2\}=\{s^{-1}s:s\in S\}=\{ss^{-1}:s\in S\}\]
always commute. The \emph{canonical partial order} $\leq$ on $S$ is given by
\[s\leq t\qquad\Leftrightarrow\qquad s^{-1}s=s^{-1}t\qquad\Leftrightarrow\qquad ss^{-1}=ts^{-1}\qquad\Leftrightarrow\qquad s\in Et=tE.\]
Also, a \emph{zero} in $S$ is a (necessarily unique) element $0\in S$ such that
\[0S=S0=\{0\}.\]
Note this implies $0\in E$ and $S\geq 0$.
\subsection{Cosets}
\begin{center}
\textbf{Throughout this section we assume $S$ is an inverse semigroup.}
\end{center}
\begin{dfn}[{\cite[\S3.1]{LawsonMargolisSteinberg2013}}]
We call $C\subseteq S$ a \emph{coset} if $CC^{-1}C=C^\leq=C$.
\end{dfn}
As $s=ss^{-1}s$, for all $s\in S$, for $C$ to be a coset it suffices that
\[\tag{Coset}CC^{-1}C\subseteq C^\leq\subseteq C.\]
Moreover, even if $C$ just satisfies $CC^{-1}C\subseteq C^\leq$, we can still obtain a coset by taking the upwards closure, as $C^\leq C^{\leq-1}C^\leq\subseteq(CC^{-1}C)^\leq\subseteq C^{\leq\leq}=C^\leq$.
\begin{prp}
For any non-empty $T,U,C\subseteq S$, if $C$ is a coset then
\begin{equation}\label{T-1U}
T^{-1}U\subseteq(CC^{-1})^\leq\qquad\Leftrightarrow\qquad T^{-1}UC\subseteq C\qquad\Rightarrow\qquad(TC)^\leq=(UC)^\leq.
\end{equation}
\end{prp}
\begin{proof}
If $T^{-1}U\subseteq(CC^{-1})^\leq$ then
\[T^{-1}UC\subseteq(CC^{-1})^\leq C\subseteq(CC^{-1}C)^\leq=C^\leq=C.\]
Conversely, if $T^{-1}UC\subseteq C(\neq\emptyset$ and $T\neq\emptyset$) then
\begin{align*}
T^{-1}U&\subseteq(T^{-1}UCC^{-1})^\leq\subseteq(CC^{-1})^\leq\quad\text{and}\\
(UC)^\leq&\subseteq(TT^{-1}UC)^\leq\subseteq(TC)^\leq.
\end{align*}
But $T^{-1}U\subseteq(CC^{-1})^\leq$ implies $U^{-1}T=(T^{-1}U)^{-1}\subseteq(CC^{-1})^{\leq-1}=(CC^{-1})^\leq$ and hence $U^{-1}TC\subseteq C$ which (as $U\neq\emptyset$) likewise yields
\[(TC)^{-1}\subseteq(UU^{-1}TC)^\leq\subseteq(UC)^\leq\]
and hence $(TC)^\leq=(UC)^\leq$.
\end{proof}
\begin{prp}\label{TCcoset}
If $T^{-1}T\subseteq(CC^{-1})^\leq$ and $C$ is coset then so is $(TC)^\leq$.
\end{prp}
\begin{proof}
By \eqref{T-1U}, $T^{-1}T\subseteq(CC^{-1})^\leq$ implies $T^{-1}TC\subseteq C$ so
\[TC(TC)^{-1}TC=TCC^{-1}T^{-1}TC\subseteq TCC^{-1}C=TC.\]
Thus the same applies to $(TC)^\leq$ and hence $(TC)^\leq$ is a coset.
\end{proof}
\begin{prp}\label{IdempotentCosets}
If $C$ is a coset then $C\cap E\neq\emptyset$ iff $C=(CC^{-1})^\leq$.
\end{prp}
\begin{proof}
If $e\in C\cap E$ then $C\subseteq(Ce)^\leq\subseteq(CC^{-1})^\leq$ and
\[(CC^{-1})^\leq\subseteq((CC^{-1})^\leq e)^\leq\subseteq(CC^{-1}C)^\leq=C.\]
On the other hand, if $C=(CC^{-1})^\leq$ then $cc^{-1}\in C\cap E$, for any $c\in C$.
\end{proof}
In \cite[\S3.1]{LawsonMargolisSteinberg2013} the cosets are considered as an inverse semigroup with the product $C\otimes D$ defined to be the coset generated by $CD$. In particular, we always have $(CD)^\leq\subseteq C\otimes D$. In the cases we are interested in, equality will even hold here. Indeed, by \autoref{TCcoset} with $T=C^{-1}$, we always have
\[C^{-1}\otimes C=(C^{-1}C)^\leq.\]
Thus $C^{-1}\otimes C=D\otimes D^{-1}$ is equivalent to $(C^{-1}C)^\leq=(DD^{-1})^\leq$, in which case we again have $C\otimes D=(CD)^\leq$. Indeed, for any $c,c'\in C'$, we have $c^{-1}c'\in(DD^{-1})^\leq$ so $(cD)^\leq=(c'D)^\leq$, by \eqref{T-1U}, and hence $(CD)^\leq=(cD)^\leq$ is a coset, again by \autoref{TCcoset}.
As with any inverse semigroup, this restricted product yields a \emph{groupoid} (see \cite[\S3.1]{Lawson1998}), i.e. a set $G$ with an inverse operation and partial product operation making $G$ a small category in which all morphisms are isomorphisms.
\begin{dfn}
The \emph{coset groupoid} of $S$ is the set of cosets of $S$ with operations
\begin{align}
\tag{Inverse}C&\mapsto C^{-1}\quad\text{and}\\
\tag{Product}(C,D)&\mapsto(CD)^\leq\quad\text{iff}\quad(C^{-1}C)^\leq=(DD^{-1})^\leq.
\end{align}
\end{dfn}
\subsection{Ordered Inverse Semigroups}
We will be interested inverse semigroups with an extra (possibly non-reflexive) order relation $\prec$ which behaves well with respect to the inverse semigroup structure.
\begin{dfn}
An \emph{ordered inverse semigroup} is an inverse semigroup $S$ together with a transitive left auxiliary relation $\prec\ \subseteq\ \leq$ preserving products and inverses, i.e.
\begin{align}
\label{LeftAuxiliary}\tag{Left Auxiliary}q\prec q'\leq r\qquad&\Rightarrow\qquad q\prec r\qquad\Rightarrow\qquad q\leq r.\\
\label{Multiplicative}\tag{Multiplicative}q\prec q'\quad\text{and}\quad r\prec r'\qquad&\Rightarrow\qquad qr\prec q'r'.\\
\label{Inverses}\tag{Invertive}q\prec r\qquad&\Rightarrow\qquad q^{-1}\prec r^{-1}.
\end{align}
\end{dfn}
\begin{rmk}\label{CanonicalOrderedInverseSemigroup}
Any inverse semigroup $S$ becomes an ordered inverse semigroup when we take $\prec\ =\ \leq$. The point is that there may be many other strictly stronger relations $\prec$ which make $S$ an ordered inverse semigroup.
\end{rmk}
One immediate consequence is the following auxiliary-product property.
\begin{prp}
If $r^{-1}r\leq ss^{-1}$ and $s\prec S$ in an ordered inverse semigroup $S$,
\begin{equation}\label{ac<bc}
q\prec r\qquad\Rightarrow\qquad qs\prec rs.
\end{equation}
\end{prp}
\begin{proof}
Take $t\in S$ with $s\prec t$ and hence $s\leq t$. By \eqref{Multiplicative},
\[qs\prec rt=rr^{-1}rt=rr^{-1}rss^{-1}t=rs.\qedhere\]
\end{proof}
\begin{prp}\label{Frink=>Coset}
Every Frink filter $T$ in an ordered inverse semigroup is a coset.
\end{prp}
\begin{proof}
To see that $T=T^\leq$, note that if $u\geq t\in T$ then, as $T$ is a Frink filter, we have $w\in P$ and finite $F\subseteq T$ with $\widehat{F}\prec w\prec t\leq u$. As $\prec$ is auxiliary to $\leq$, this yields $\widehat{F}\prec w\prec u$ and hence $u\in T$, as $T$ is a Frink filter.
To see that $T=TT^{-1}T$, again note that, for any $a,b,c\in T$, we have $a',b',c'\in P$ and finite $F,G,H\subseteq T$ with $\widehat{F}\prec a'\prec a$, $\widehat{G}\prec b'\prec b$ and $\widehat{H}\prec c'\prec c$. For any $d\in\widehat{I}$, where $I=F\cup G\cup H$, \eqref{Multiplicative} and \eqref{Inverses} yields
\[d=dd^{-1}d\prec a'b'^{-1}c'\prec ab^{-1}c,\]
i.e. $\widehat{I}\prec a'b'^{-1}c'\prec ab^{-1}c$. Thus $ab^{-1}c\in T$, as $T$ is a Frink filter.
\end{proof}
\begin{center}
\textbf{For the rest of this section we assume that $(S,\prec)$ is an ordered inverse semigroup with zero and $P=S\setminus\{0\}$ is round}.
\end{center}
\begin{rmk}
If $\prec$ is not just left but also right auxiliary, i.e.
\[\label{RightAuxiliarity}\tag{Right Auxiliary}q\leq r'\prec r\qquad\Rightarrow\qquad q\prec r,\]
then, as $P$ is round, for any $p\in P$, we have $p'\in P$ with $0\leq p'\prec p$ and hence $0\prec p$. As long as we also have some $p,q\in P$ with $pq=0$ then \eqref{Multiplicative} yields $0=00\prec pq=0$ too. So in this case $0$ is not just a $\leq$-minimum but also a $\prec$-minimum (which is moreover unique, as $\prec\ \subseteq\ \leq$).
\end{rmk}
We define $\mathrel{\mathsf{D}}$, $\mathrel{\mathsf{C}}$ and $\perp$ on $P$ from $\prec$ exactly as before. For convenience, we extend these relations to $S$ by simply ignoring $0$, e.g. for $Q,R\subseteq S$,
\[Q\mathrel{\mathsf{D}} R\qquad\Leftrightarrow\qquad(Q\setminus\{0\})\mathrel{\mathsf{D}}(R\setminus\{0\}).\]
First we note that $\perp$ and $\mathrel{\mathsf{D}}$ can equivalently be defined from $\leq$ rather than $\prec$.
\begin{prp}
For all $Q,R\subseteq S$ and $s\in S$,
\begin{align}
\label{QperpR0}Q^\geq\cap R^\geq=\{0\}\qquad\Leftrightarrow\qquad&Q\perp R\qquad\Rightarrow\qquad Qs\perp Rs.\\
\label{QDR0}Q^\geq\setminus\{0\}\ \geq\ R^\geq\setminus\{0\}\qquad\Leftrightarrow\qquad&Q\mathrel{\mathsf{D}} R\qquad\Rightarrow\qquad Qs\mathrel{\mathsf{D}} Rs.
\end{align}
\end{prp}
\begin{proof}\
\begin{itemize}
\item[\eqref{QperpR0}] If $q,r\geq p\neq0$ then, as $P$ is round, we have $p'\neq0$ with $q,r\geq p\succ p'$ and hence $q,r\succ p'$, by \eqref{LeftAuxiliary}, so $q\not\perp r$. Conversely, if $q\not\perp r$ then we have $p\neq0$ with $q,r\succ p$ and hence $q,r\geq p$, as $\prec\ \subseteq\ \leq$. This proves the $\Leftrightarrow$ part. For $\Rightarrow$, just note $q\wedge r=0$ implies $qs\wedge rs=(q\wedge r)s=0$, by \cite[Proposition 1.4.19]{Lawson1998}.
\item[\eqref{QDR0}] If $Q\mathrel{\mathsf{D}} R$, i.e. $Q^\succ\setminus\{0\}\succ R^\succ\setminus\{0\}$, then
\[Q^\geq\setminus\{0\}\ \succ\ Q^\succ\setminus\{0\}\ \succ\ R^\succ\setminus\{0\}\ \subseteq\ R^\geq\setminus\{0\}\]
and hence $Q^\geq\setminus\{0\}\geq R^\geq\setminus\{0\}$, as $P$ is round and \eqref{LeftAuxiliary} holds. Conversely, if $Q^\geq\setminus\{0\}\geq R^\geq\setminus\{0\}$ then likewise
\[Q^\succ\setminus\{0\}\ \subseteq\ Q^\geq\setminus\{0\}\ \geq\ R^\geq\setminus\{0\}\ \succ\ R^\succ\setminus\{0\}\]
and hence $Q^\succ\setminus\{0\}\succ R^\succ\setminus\{0\}$.
Now assume $Q\mathrel{\mathsf{D}} R$ and $0\neq p\leq qs$, for some $q\in Q$. Then $0\neq p=pp^{-1}p\leq qsp^{-1}p$ so $sp^{-1}\neq0$. Thus $0\neq ps^{-1}\leq qss^{-1}\leq q$. As $Q\mathrel{\mathsf{D}} R$, we have non-zero $r\leq ps^{-1}$ with $r\leq R$. Then again $0\neq rs\leq ps^{-1}s\leq p$, showing that $Qs\mathrel{\mathsf{D}} Rs$.\qedhere
\end{itemize}
\end{proof}
\begin{rmk}
If $S$ is separative, i.e. if $\mathrel{\mathsf{C}}$ and $\prec$ coincide on singletons, then \eqref{RightAuxiliarity} holds. Indeed \eqref{DC=>C}, \eqref{FprecQ}, \eqref{QDR0} and separativity yield
\[q\leq r'\prec r\quad\Rightarrow\quad q\mathrel{\mathsf{D}} r'\mathrel{\mathsf{C}} r\quad\Rightarrow\quad q\mathrel{\mathsf{C}} r\quad\Rightarrow\quad q\prec r.\]
\end{rmk}
\begin{cor}
For any $Q,Q',R,R'\subseteq S$,
\begin{align}
\label{QRDQ'R'}Q\mathrel{\mathsf{D}} Q'\quad\text{and}\quad R\mathrel{\mathsf{D}} R'\qquad&\Rightarrow\qquad QR\mathrel{\mathsf{D}} Q'R'.\\
\label{QRCQ'R'}Q\mathrel{\mathsf{C}} Q'\quad\text{and}\quad R\mathrel{\mathsf{C}} R'\qquad&\Rightarrow\qquad QR\mathrel{\mathsf{C}} Q'R'.
\end{align}
\end{cor}
\begin{proof}\
\begin{itemize}
\item[\eqref{QRDQ'R'}] If $Q\mathrel{\mathsf{D}} Q'$ and $R\mathrel{\mathsf{D}} R'$ then $QR\mathrel{\mathsf{D}} Q'R\mathrel{\mathsf{D}} Q'R'$, by \eqref{QDR0} and \eqref{Dcup} (or rather an extension of \eqref{Dcup} to infinite unions).
\item[\eqref{QRCQ'R'}] If $Q\mathrel{\mathsf{C}} Q'$ and $R\mathrel{\mathsf{C}} R'$ then $Q\mathrel{\mathsf{D}} F\prec Q'$ and $R\mathrel{\mathsf{D}} G\prec R'$, for some finite $F,G\subseteq S$, and hence $QR\mathrel{\mathsf{D}} FG\prec Q'R'$, by \eqref{QRDQ'R'} and \eqref{Multiplicative}.\qedhere
\end{itemize}
\end{proof}
\begin{prp}
If $s\prec S$ and $ss^{-1}\in(T^{-1}T)^\leq$ then
\begin{equation}\label{TsTight}
T\in\mathcal{T}(P)\qquad\Rightarrow\qquad(Ts)^\leq\in\mathcal{T}(P).
\end{equation}
\end{prp}
\begin{proof}
If $T$ is tight then $T$ is a round Frink filter and hence a coset, by \autoref{Frink=>Coset}. Also, for any $t\in T$, \eqref{T-1U} yields $tss^{-1}\in Tss^{-1}\subseteq T$. As $T$ is round, we have $u\in T$ with $u\prec tss^{-1}$ and then \eqref{ac<bc} yields $us\prec tss^{-1}s=ts$. This shows that $Ts$ is round and hence so is $(Ts)^\leq$, by \eqref{LeftAuxiliary}.
Now we need to show that $(Ts)^\leq\not\mathrel{\widehat{\mathsf{C}}} S\setminus(Ts)^\leq$. For this it suffices to show that
\[Ts\mathrel{\widehat{\mathsf{C}}} H\qquad\Rightarrow\qquad H\cap(Ts)^\leq\neq\emptyset.\]
So take finite $F\subseteq Ts$ with $\widehat{F}\mathrel{\mathsf{C}} H$. Then \eqref{ac<bc} yields $\widehat{Fs^{-1}}=\widehat{F}s^{-1}\mathrel{\mathsf{C}} Hr^{-1}$, for any $r\succ s$, by \eqref{FprecQ} and \eqref{QRCQ'R'}. As $T$ is tight and $Fs^{-1}\subseteq Tss^{-1}\subseteq T$, we have $h\in H$ with $hr^{-1}\in T$. Thus $h\geq hr^{-1}r\geq hr^{-1}s\in Ts$ so $H\cap(Ts)^\leq\neq\emptyset$.
\end{proof}
\begin{cor}\label{TightGroupoid}
$\mathcal{T}(P)$ is a subgroupoid of the coset groupoid of $S$.
\end{cor}
\begin{proof}
Take any $T,U\in\mathcal{T}(P)$ with $(T^{-1}T)^\leq=(UU^{-1})^\leq$. As $U$ is round, we have $u\in U$ with $u\prec U\subseteq S$. Thus $(TU)^\leq=(Tu)^\leq\in\mathcal{T}(P)$, by \eqref{T-1U} and \eqref{TsTight}. Thus $\mathcal{T}(P)$ is closed under the restricted product. By \eqref{Inverses}, $\mathcal{T}(P)$ is also immediately seen to be closed under taking inverses.
\end{proof}
\begin{dfn}
We call $\mathcal{T}(P)$ the \emph{tight groupoid} of $S$.
\end{dfn}
\subsection{\'Etale Groupoids}
Recall that an \emph{\'etale groupoid} $G$ is a groupoid together with a topology that behaves well with respect to the groupoid operations, specifically
\begin{gather*}
g\mapsto g^{-1}g\quad\text{is a local homeomorphism and}\\
(g,h)\mapsto gh\quad\text{and}\quad g\mapsto g^{-1}\quad\text{are continuous}.
\end{gather*}
However, we will use an equivalent more algebraic definition. First let
\begin{align*}
G^{(0)}&=\{g^{-1}g:g\in G\}.\\
G^{(2)}&=\{(g,h):g^{-1}g=hh^{-1}\}.\\
\mathcal{B}(G)&=\{B\subseteq G:B^{-1}B,BB^{-1}\subseteq G^{(0)}\}.\\
\mathcal{OB}(G)&=\mathcal{O}(G)\cap\mathcal{B}(G).
\end{align*}
So $G^{(0)}$ is the unit space of $G$, $G^{(2)}$ is the set of composable pairs, $\mathcal{B}(G)$ denotes the bisections of $G$ and $\mathcal{OB}(G)$ denotes the open bisections of $G$.
\begin{dfn}
We call a collection of bisections $\mathcal{B}\subseteq\mathcal{B}(G)$ an \emph{\'etale family} if $\mathcal{B}$ covers $G$ and is closed under pointwise products and inverses, i.e.
\[\tag{\'Etale Family}G=\bigcup\mathcal{B}\quad\text{and}\quad(O,N\in\mathcal{B}\quad\Rightarrow\quad ON\in\mathcal{B}\text{ and }O^{-1}\in\mathcal{B}).\]
An \emph{\'etale (pseudo)(sub)basis} is a (pseudo)(sub)basis that is also an \'etale family.
\end{dfn}
\begin{prp}\label{EtaleSubbasis}
$G$ is an \'etale groupoid if and only if $G$ has an \'etale subbasis.
\end{prp}
\begin{proof}
By \cite[Proposition 6.6]{BiceStarling2018}, $G$ is \'etale if and only if $G$ has an \'etale basis. To show that an \'etale subbasis suffices, one only has to note that the pointwise product preserves intersections of bisections, i.e. for any $\mathcal{A},\mathcal{B}\subseteq\mathcal{B}(G)$.
\begin{equation}\label{IntersectionPreserving}
(\bigcap\mathcal{A})(\bigcap\mathcal{B})=\bigcap\mathcal{A}\mathcal{B}(=\bigcap_{\substack{A\in\mathcal{A}\\ B\in\mathcal{B}}}AB).
\end{equation}
This follows from the fact that, for any $B\in\mathcal{B}(G)$ and $\mathcal{A}\subseteq\mathcal{P}(G)$,
\[B\bigcap\mathcal{A}=\bigcap B\mathcal{A}(=\bigcap_{A\in\mathcal{A}}BA).\]
Indeed, $B\bigcap\mathcal{A}\subseteq\bigcap B\mathcal{A}$ is immediate, while $B^{-1}B\subseteq G^{(0)}$ yields
\[\bigcap B\mathcal{A}\subseteq BB^{-1}\bigcap B\mathcal{A}\subseteq B\bigcap B^{-1}B\mathcal{A}\subseteq B\bigcap\mathcal{A}.\qedhere\]
\end{proof}
In Hausdorff groupoids even \'etale pseudosubbases suffice, as we now show.
\begin{prp}\label{CoherentEtale}
If $G$ is a coherent \'etale groupoid then $G^\mathrm{patch}$ is also \'etale.
\end{prp}
\begin{proof}
First note saturated subsets are closed under taking products (even in an arbitrary \'etale groupoid). Indeed, as $\mathcal{OB}(G)$ covers $G$, for any $Q,R\subseteq\mathcal{O}(G)$,
\begin{align*}
\big(\bigcap Q\big)\big(\bigcap R\big)&=\big(\bigcup_{A\in\mathcal{OB}(G)}A\cap\bigcap Q\big)\big(\bigcup_{B\in\mathcal{OB}(G)}B\cap\bigcap R\big)\\
&=\big(\bigcup_{A\in\mathcal{OB}(G)}\bigcap_{C\in Q}A\cap C\big)\big(\bigcup_{B\in\mathcal{OB}(G)}\bigcap_{D\in R}B\cap D\big)\\
&=\bigcup_{A,B\in\mathcal{OB}(G)}\big(\bigcap_{C\in Q}A\cap C\big)\big(\bigcap_{D\in R}B\cap D\big)\\
&=\bigcup_{A,B\in\mathcal{OB}(G)}\bigcap_{\substack{C\in Q\\ D\in R}}(A\cap C)(B\cap D),
\end{align*}
by \eqref{IntersectionPreserving}. As $G$ is \'etale, each $(A\cap C)(B\cap D)$ is open and hence each intersection $\bigcap_{C\in Q,D\in R}(A\cap C)(B\cap D)$ is saturated. As open sets are closed under arbitrary unions, so are saturated subsets and hence the product is indeed saturated.
Next we claim that products of compact saturated subsets are again compact. To see this, take $C,D\in\mathcal{C}(G)$. For each $c\in C$, we have $O,N\in\mathcal{OB}(G)$ with $c\in O\Subset N$, as $\mathcal{OB}(G)$ covers $G$ and $G$ is locally compact. As $C$ is compact, we can cover it with finitely many such $O$. This yields a finite $F\subseteq\mathcal{C}(G)\cap\mathcal{OB}(G)^\Supset$ with $C\subseteq\bigcup F$, i.e. each $C'\in F$ is compact, saturated and contained in some open bisection. As $G$ is coherent, $C'\cap C\in\mathcal{C}(G)$, for all $C'\in F$, and $C=\bigcup_{C'\in F}C'\cap C$. The same applies to $D$ and thus $CD$ can be expressed as a finite union of products of compact saturated subsets, each of which is contained in some open bisection. As $\mathcal{C}(G)$ is closed under finite unions, it suffices to consider the case that $C$ and $D$ are already contained in open bisections $O$ and $N$ respectively.
Now consider the maps $s(g)=g^{-1}g$ and $r(g)=gg^{-1}$ restricted to $O$ and $N$ respectively, which are homeomorphisms. In particular, as they are continuous, $C^{-1}C$ and $DD^{-1}$ are compact and also saturated, as mentioned above. As $G$ is coherent, the same applies to their intersection, i.e. $E=C^{-1}C\cap DD^{-1}\in\mathcal{C}(G)$. Now $p(g)=s^{-1}(g)r^{-1}(g)$ restricted to $O^{-1}O\cap NN^{-1}$ is also a homeomorphism and hence $CD=p[E]$ is also compact and saturated, i.e. $CD\in\mathcal{C}(G)$, as required.
Lastly, note that $G^\mathrm{patch}$ has a basis of elements of the form $O\setminus C$ where $O\in\mathcal{O}(G)\cap\mathcal{OB}(G)^\Supset$ and $C\in\mathcal{C}(G)$, so each $O$ is open and contained in some compact bisection $B_O$. But then, for any such $O\setminus C$ and $N\setminus D$,
\[(O\setminus C)(N\setminus D)=ON\setminus(OD\cup CN)=ON\setminus(B_OD\cup CB_N).\]
As $\mathcal{C}(G)$ is closed under products, this shows that our basis is closed under products too, and it is immediately seen to be closed under inverses as well.
\end{proof}
\begin{cor}\label{EtalePseudosubbasis}
If $G$ is Hausdorff with an \'etale pseudosubbasis $P$ then $G$ is \'etale.
\end{cor}
\begin{proof}
By \autoref{EtaleSubbasis}, $G_P(=$ $G$ with the topology generated by $P)$ is \'etale. By \autoref{PseudoPatch} and \autoref{CoherentEtale}, $G=G_P^\mathrm{patch}$ is \'etale.
\end{proof}
Later we will need the following elementary result on \'etale families.
\begin{prp}
If $\mathcal{B}\subseteq\mathcal{B}(G)$ is \'etale, $T_g=\{O\in\mathcal{B}:g\in O\}$ and $\leq\ =\ \subseteq$ on $\mathcal{B}$,
\begin{equation}\label{Functor}
T_{g^{-1}}=T_g^{-1}\qquad\text{and}\qquad T_{gh}=(T_gT_h)^\leq,\quad\text{for all }(g,h)\in G^{(2)}.
\end{equation}
\end{prp}
\begin{proof}
We immediately see that $T_g^{-1}=T_{g^{-1}}$ and $T_gT_h\subseteq T_{gh}$ and hence $(T_gT_h)^\leq\subseteq T_{gh}$, as $gh\in O\subseteq N$ implies $gh\in N$. Conversely, take any $O\in T_g$ and note that, as $O$ is a bisection, $OO^{-1}N\subseteq N$, for any $N\in T_{gh}$, so
\[T_{gh}\subseteq(OO^{-1}T_{gh})^\leq\subseteq(T_gT_{g^{-1}}T_{gh})^\leq\subseteq(T_gT_{g^{-1}gh})^\leq\subseteq(T_gT_h)^\leq.\qedhere\]
\end{proof}
\subsection{Pseudobasic Inverse Semigroups}
Now we can finally show that, for suitable inverse semigroups, our tight groupoid is necessarily \'etale.
\begin{dfn}
We call $(S,\prec)$ a \emph{pseudobasic inverse semigroup} if $S$ is an ordered inverse semigroup with zero and $P=S\setminus\{0\}$ is an abstract pseudobasis.
\end{dfn}
\begin{thm}\label{LCHEgroupoid}
If $S$ is a pseudobasic inverse semigroup then $\mathcal{T}(P)$ is a locally compact Hausdorff \'etale groupoid with an \'etale pseudobasis $(O_s)_{s\prec S}$.
\end{thm}
\begin{proof}
By \autoref{TightHausdorff}, \autoref{TightLocallyCompact} and \autoref{TightGroupoid}, $\mathcal{T}(P)$ is locally compact Hausdorff groupoid. By \autoref{abstract->concrete}, $(O_p)_{p\in P}$ and hence $(O_s)_{s\prec S}$ is a pseudo(sub)basis for $\mathcal{T}(P)$. In particular, $(O_s)_{s\prec S}$ covers $\mathcal{T}(P)$ so, by \autoref{EtalePseudosubbasis}, it suffices to verify that $(O_s)_{s\prec S}$ is an \'etale family.
First note that, for any $s\prec S$,
\begin{equation}\label{ss-1}
O_sO_s^{-1}=O_{ss^{-1}}.
\end{equation}
Indeed, $O_sO_s^{-1}\subseteq O_{ss^{-1}}$ is immediate. Conversely, if $T\in O_{ss^{-1}}$ then $U=(Ts)^\leq\in\mathcal{T}(P)$, by \eqref{TsTight}. Thus $U\in O_s$, $U^{-1}\in O_{s^{-1}}$ and $(UU^{-1})^\leq=(Tss^{-1}T^{-1})^\leq=(Tss^{-1}ss^{-1})^\leq=T$ so $T\in O_sO_{s^{-1}}$, as required. As $ss^{-1}\in E$, \autoref{IdempotentCosets} yields $O_{ss^{-1}}\in\mathcal{T}(P)^{(0)}$. Likewise, $O_s^{-1}O_s=O_{s^{-1}s}\in\mathcal{T}(P)^{(0)}$ so $O_s\in\mathcal{OB}(\mathcal{T}(P))$.
We also immediately have $O_{s}^{-1}=O_{s^{-1}}$ and $O_sO_t\subseteq O_{st}$, for any $s,t\in S$. Conversely, as long as $s\prec S$, \eqref{ss-1} yields
\[O_{st}=O_{st}O_{t^{-1}s^{-1}}O_{st}\subseteq O_{stt^{-1}s^{-1}}O_{st}\subseteq O_{ss^{-1}}O_{st}=O_sO_{s^{-1}}O_{st}\subseteq O_sO_{s^{-1}st}\subseteq O_sO_t.\]
Thus $(O_s)_{s\prec S}$ is also closed under products, as required.
\end{proof}
Exel's original tight groupoid construction was obtained from the groupoid of germs coming from the canonical partial action of $S$ on the tight spectrum of $E$. In the Hausdorff case, this agrees with our tight groupoid when we consider $S$ as a pseudobasic inverse semigroup with $\prec\ =\ \leq$ (see \autoref{InterpolationRemark} and \autoref{CanonicalOrderedInverseSemigroup}). Indeed, it follows from the Hausdorff characterization in \cite[Theorem 3.16]{ExelPardo2016} that in this case every one of our tight subsets will be a true filter, not just a Frink filter, and these can be identified with germs just like in Lenz's alternative construction of Paterson's universal groupoid (see \cite{Lenz2008} and \cite[\S3.3]{LawsonMargolisSteinberg2013}). In the non-Hausdorff case, our tight groupoid will instead be a `Hausdorffification' of Exel's original tight groupoid. To illustrate this, we give a simple example.
\begin{xpl}
Let $G=\mathbb{N}\cup\{\infty,\infty'\}$ be the two-point compactification of $\mathbb{N}$, so the neighborhoods of $\infty$ are of the form $C\cup\{\infty\}$, for some cofinite $C\subseteq\mathbb{N}$, and likewise for the neighborhoods of $\infty'$. We make $G$ an ample \'etale groupoid by letting the unit space be $G^0=\mathbb{N}\cup\{\infty\}$ and defining $\infty'\cdot\infty'=\infty$. Let
\[S=\{\{n\}:n\in\mathbb{N}\}\cup\{e\}\cup\{s\}\cup\{\emptyset\},\]
where $e=\mathbb{N}\cup\{\infty\}$ and $s=\mathbb{N}\cup\{\infty'\}$. So $S$ is a pseudobasis of compact open bisections and an inverse semigroup under pointwise operations with all elements idempotent except $s$. If we applied Exel's original tight groupoid construction to $S$, we would recover $G$. However, our tight groupoid construction identifies $\infty$ and $\infty'$ because every $T\in\mathcal{T}(S\setminus\{\emptyset\})$ must contain both $s$ and $e$, as $s^\perp=e^\perp=\{\emptyset\}$.
\end{xpl}
\begin{rmk}
If we want a faithful extension of Exel's original contruction, even in the non-Hausdorff case, then this can still be done in a similar way. Basically, the idea is to modify the definition of (both abstract and concrete) pseudobases and work with `local' versions of tight subsets and the tight spectrum, as we discuss in a follow up paper. However, the Hausdorff construction presented here may be of some interest in its own right, even in the original $\prec\ =\ \leq$ case. Indeed, it would be interesting to see if our Hausdorff tight groupoid might be a suitable alternative to some of the non-Hausdorff tight groupoids already considered in the literature.
\end{rmk}
We can also turn our tight groupoid construction into a duality:
\begin{gather*}
\text{separative pseudobasic inverse semigroups}\\
\updownarrow\\
\text{\'etale pseudobases of Hausdorff groupoids}.
\end{gather*}
Indeed, \autoref{LCHEgroupoid} has already provided the $\downarrow$ part (although strictly speaking should also assume \eqref{ac<bc} holds for arbitrary $s\in S$, not just $s\prec S$, so that the entirety of $(O_s)_{s\in S}$ is an \'etale family). The $\uparrow$ part and its reversibility comes from the following extension of the Hausdorff case of \cite[Theorem 4.8]{Exel2010} (from bases to pseudobases but, more importantly, even for non-ample groupoids).
\begin{prp}\label{GroupoidRecovery}
If $G$ is Hausdorff with \'etale pseudobasis $P\not\ni\emptyset$ then $(P\cup\{\emptyset\},\Subset)$ is a pseudobasic inverse semigroup. Moreover, $G$ is isomorphic to $\mathcal{T}(P)$ via
\[g\mapsto T_g=\{O\in P:g\in O\}.\]
\end{prp}
\begin{proof}
It is immediately verified that $P\cup\{\emptyset\}$ is inverse semigroup where $\leq\ =\ \subseteq$ on $P$. As $\Subset$ is a stronger relation than $\subseteq$, \eqref{LeftAuxiliary} and \eqref{Inverses} for $\prec\ =\ \Subset$ immediately follows. As $G$ is Hausdorff, \eqref{Multiplicative} holds, by \cite[Proposition 6.4]{BiceStarling2018} (or note products of compact subsets are compact as in the proof of \autoref{CoherentEtale}). By \autoref{TopologyRecovery}, $P$ is an abstract pseudobasis and $g\mapsto T_g$ is a homeomorphism onto $\mathcal{T}(P)$. By \eqref{Functor}, $g\mapsto T_g$ is also functor. Also, if $g^{-1}g\neq hh^{-1}$ then, by \eqref{Separating},
\[(T_g^{-1}T_g)^\leq=T_{g^{-1}g}\neq T_{hh^{-1}}=(T_hT_h^{-1})^\leq,\]
i.e. $(g,h)\in G^{(2)}$ iff $(T_g,T_h)\in\mathcal{T}(P)^{(2)}$. So $G$ is isomorphic to $\mathcal{T}(P)$ via $g\mapsto T_g$.
\end{proof}
\begin{xpl}
Say $G$ is a discrete group acting on some locally compact Hausdorff space $X$. Given any pseudobasis $P$ for $X$, we can obtain a larger pseudobasis $P'$ by taking its closure under finite intersections and images of elements of $G$ (and if $X$ is second countable then we can even take $P$ to be a countable (pseudo)basis which will remain countable when enlarged in this way, as long as $G$ is also countable). Then $S=(g,O)_{g\in G,O\in P'}$ will form a \'etale pseudobasis of the transformation groupoid $(G,X)$. Applying \autoref{GroupoidRecovery} allows us to recover $(G,X)$ just from the inverse semigroup structure of $S$.
\end{xpl}
\newpage
|
\section{Hecke modules and their centralizers}\label{section3}
\subsection{Fundamental properties}
We follow ideas in \cite[Chapter 9.1]{DDPW08}. Let $\pi$ be an index set. Suppose that we are given a map $\pi \rightarrow \{ J \mid J \subset \{ 0,1,\ldots,d-1 \} \}$. We denote by $I_\lambda$ the image of $\lambda \in \pi$ under this map. For each $\lambda \in \pi$, for simplicity, we will denote $W_{I_\lambda}$, $w_{I_\lambda}$, $x_{I_\lambda}$, etc. by $W_\lambda$, $w_{\lambda}$, $x_{\lambda}$, etc.
\begin{defi}\normalfont
Associated with $\pi$, we define a right $\Hc$-module $\bbT(\pi) := \bigoplus_{\lambda \in \pi} x_\lambda \Hc$, and its centralizer algebra $\bbS(\pi) := \End_\Hc(\bbT(\pi))$; we let $\bbS(\pi)$ act on $\bbT(\pi)$ from the left.
\end{defi}
It is obvious that $\bbT(\pi)$ has two bases $\{ {}^\lm C_w \mid \lm \in \pi,\ w \in {}^\lm W \}$ and $\{ {}^\lm D_w \mid \lm \in \pi,\ w \in {}^\lm W \}$; we call them the Kazhdan-Lusztig basis and dual Kazhdan-Lusztig basis, respectively.
For each $m = \sum_{\lambda \in \pi} m_\lambda \in \bbT(\pi)$ with $m_\lambda \in x_\lambda \Hc$, we define $\ol{m} \in \bbT(\pi)$ to be $\sum_{\lambda \in \pi} \ol{m_\lambda}$. Also, for each $f \in \bbS(\pi)$, define $\ol{f} \in \bbS(\pi)$ by $\ol{f}(m) = \ol{f(\ol{m})}$ for all $m \in \bbT(\pi)$. This gives anti-linear automorphisms $\ol{\ \cdot \ }$ on $\bbT(\pi)$ and $\bbS(\pi)$.
For each $\lambda \in \pi$, define $p_\lambda\index{plambda@$p_\lambda$} \in \bbS(\pi)$ to be the composite\begin{align}
p_\lambda: \bbT(\pi) \twoheadrightarrow x_\lambda \Hc \hookrightarrow \bbT(\pi) \nonumber
\end{align}
of the projection and the inclusion. Clearly, $\{ p_\lambda \mid \lambda \in \pi \}$ is a family of orthogonal idempotents with $\sum_{\lambda \in \pi} p_\lambda = \id_{\bbT(\pi)}$. Hence, we have a decomposition
\begin{align}
\bbS(\pi) = \bigoplus_{\lambda,\mu \in \pi} p_\lambda \bbS(\pi) p_\mu, \qu p_\lambda \bbS(\pi) p_\mu = \Hom_\Hc(x_\mu \Hc, x_\lambda \Hc). \nonumber
\end{align}
Take $f \in \Hom_\Hc(x_\mu \Hc, x_\lambda \Hc)$ arbitrarily. Since $x_\mu \Hc$ is generated (as a right $\Hc$-module) by $x_\mu$, the $f$ is determined by $f(x_\mu) \in x_\lambda \Hc$. Let us write
\begin{align}
f(x_\mu) = \sum_{w \in {}^\lambda W} c_{\lambda,w,\mu}(f) x_\lambda H_w, \qu \Forsome c_{\lambda,w,\mu} \in \AZ. \nonumber
\end{align}
\begin{lem}
Let $w \in {}^\lambda W$ and $j \in I_\mu$ be such that $w < ws_j$. Then, we have
\begin{align}
c_{\lambda,w,\mu}(f) = q_j c_{\lambda,ws_j,\mu}(f). \nonumber
\end{align}
Consequently, we have
\begin{align}
f(x_\mu) = \sum_{w \in {}^\lambda W^\mu} \sum_{\substack{y \in W_\mu \\ wy \in {}^\lambda W}} q_y\inv c_{\lambda,w,\mu}(f) x_\lambda H_{wy}, \nonumber
\end{align}
and hence, $f$ is determined by $(c_{\lambda,w,\mu}(f))_{w \in {}^\lambda W^\mu} \in \AZ^{{}^\lm W^\mu}$, where ${}^\lambda W^\mu := {}^\lambda W \cap ({}^\mu W)\inv$.
\end{lem}
\begin{proof}
We have
\begin{align}
\begin{split}
q_j\inv f(x_\mu) &= f(x_\mu H_j) \\
&= f(x_\mu)H_j \\
&= \sum_{\substack{w \in {}^\lambda W \\ ws_j < w}} c_{\lambda,w,\mu}(f) x_\lambda(H_{ws_j} + (q_j\inv - q_j) H_w) + \sum_{\substack{w \in {}^\lambda W \\ ws_j > w}} c_{\lambda,w,\mu}(f) x_\lambda H_{ws_j} \\
&= \sum_{\substack{w \in {}^\lambda W \\ ws_j < w}}(c_{\lambda,ws_j,\mu}(f) + (q_j\inv - q_j)c_{\lambda,w,\mu}(f))x_\lambda H_w + \sum_{\substack{w \in {}^\lambda W \\ ws_j > w}} c_{\lambda,ws_j,\mu}(f) x_\lambda H_w.
\end{split} \nonumber
\end{align}
Comparing the coefficients of $x_\lambda H_w$, we obtain the assertion.
\end{proof}
Conversely, given $(c_{\lambda,w,\mu})_{w \in {}^\lambda W^\mu} \in \AZ^{{}^\lm W^\mu}$, there exists a unique $g \in \Hom_\Hc(x_\mu \Hc, x_\lambda \Hc)$ such that $c_{\lambda,w,\mu}(g) = c_{\lambda,w,\mu}$ for all $w \in {}^\lambda W^\mu$. Thus, we obtain an $\AZ$-linear isomorphism between $\AZ^{{}^\lambda W^\mu}$ and $\Hom_{\Hc}(x_\mu \Hc,x_\lambda \Hc)$.
\begin{lem}[{\cite[Theorem 4.18]{DDPW08}}]\label{double coset}
Let $\lm,\mu \in \pi$. For each $x \in {}^\lm W^\mu$, there exists a unique $J_x \subset \{ 0,1,\ldots,d-1 \}$ such that the multiplication map
$$
W_\lm \times \{x\} \times {}^{J_x} W_\mu \rightarrow W_\lm x W_\mu;\ (u,x,v) \mapsto uxv
$$
is a bijection, where ${}^{J_x} W_\mu := {}^{J_x} W \cap W_\mu$. Moreover, we have $\ell(uxv) = \ell(u) + \ell(x) + \ell(v)$ for all $u \in W_\lm$ and $v \in {}^{J_x} W_\mu$.
\end{lem}
For $\lambda, \mu \in \pi$ and $x \in {}^\lambda W^\mu$, define $\xi_{\lambda,x,\mu} \in \Hom_{\Hc}(x_\mu \Hc,x_\lambda \Hc)$ to be the one corresponding to $(\delta_{x,w} q_{x'})_{w \in {}^\lambda W^\mu} \in \AZ^{{}^\lambda W^\mu}$, where $x' \in W_\mu$ is the longest element in ${}^{J_x}W_\mu$ ($J_x$ is as in Lemma \ref{double coset}). Then, the next proposition is clear.
\begin{prop}
$\{ \xi_{\lambda,x,\mu} \mid \lambda,\mu \in \pi,\ x \in {}^\lambda W^\mu \}$ forms a basis of $\bbS(\pi)$.
\end{prop}
For each $\lambda,\mu \in \pi$, $x \in {}^\lambda W^\mu$, set
\begin{align}
\eta_{\lambda,x,\mu} = q_{w_\lambda xx'} \sum_{w \in W_\lambda x W_\mu} q_w\inv H_w. \nonumber
\end{align}
\begin{lem}\label{property of eta}
Let $\lambda,\mu \in \pi$, $x \in {}^\lambda W^\mu$.
\begin{enumerate}
\item $\eta_{\lambda,x,\mu}^\flat = \eta_{\mu,x\inv,\lambda}$.
\item $\xi_{\lambda,x,\mu}(x_\mu) = \eta_{\lambda,x,\mu} = \frac{1}{P_\mu}\eta_{\lambda,x,\mu} \cdot x_\mu = \frac{1}{P_\lm} x_\lm \cdot \eta_{\lm,x,\mu}$.
\item $\ol{\xi_{\lambda,e,\mu}} = \xi_{\lambda,e,\mu}$, where $e$ denotes the identity element of $W$.
\end{enumerate}
\end{lem}
\begin{proof}
\begin{enumerate}
\item By the definition of $x'$, we have $y := w_\lambda x x'$ is the longest element in $W_\lm x W_\mu$. Also, it is easily checked that the map $W \rightarrow W,\ w \mapsto w\inv$ gives a bijection $W_\lambda x W_\mu \rightarrow W_\mu x\inv W_\lambda$. Since this bijection preserves the length, $y\inv$ is the longest element in $W_\mu x\inv W_\lm$. Then, we compute $\eta_{\lambda,x\mu}^\flat$ as follows:
\begin{align}
\eta_{\lambda,x\mu}^\flat = q_y \sum_{w \in W_\lambda x W_\mu} q_w\inv H_{w\inv} = q_{y} \sum_{w \in W_\mu x\inv W_\lambda} q_{w\inv}\inv H_w = q_{y\inv} \sum_{w \in W_\mu x\inv W_\lambda} q_w\inv H_w = \eta_{\mu,x\inv,\lambda}. \nonumber
\end{align}
\item By the definition of $\xi_{\lambda,x,\mu}$, we have
\begin{align}
\begin{split}
\xi_{\lambda,x,\mu}(x_\mu) &= \sum_{\substack{y \in W_\mu \\ xy \in {}^\lambda W}} q_{x'}q_y\inv x_\lambda H_{xy} \\
&= \sum_{\substack{y \in W_\mu \\ xy \in {}^\lambda W}} \sum_{z \in W_\lambda} q_{x'}q_y\inv q_{w_\lambda} q_z\inv H_{zxy} \qu (\text{by the definition of $x_\lambda$})\\
&= \sum_{w \in W_\lambda x W_\mu} q_{w_\lambda} q_x q_{x'} q_w\inv H_w = \eta_{\lambda,x,\mu}.
\end{split} \nonumber
\end{align}
This proves the first equation. Next, we have
\begin{align}
\eta_{\lm,x,\mu} \cdot x_\mu = \xi_{\lm,x,\mu}(x_\mu^2) = P_\mu \xi_{\lm,x,\mu}(x_\mu) = P_\mu \eta_{\lm,x,\mu}, \nonumber
\end{align}
which implies the second equality. Finally, the third equality follows from the fact that $\eta_{\lm,x,\mu} = \xi_{\lm,x,\mu}(x_\mu) \in x_\lm \clH$.
\item It suffices to check that $\ol{\xi_{\lambda,e,\mu}}(x_\nu) = \xi_{\lambda,e,\mu}(x_\nu)$ for all $\nu \in \pi$. Only the non-trivial case is when $\nu = \mu$. Since we have
\begin{align}
\ol{\xi_{\lambda,e,\mu}}(x_\mu) = \ol{\xi_{\lambda,e,\mu}(x_\mu)} = \ol{\eta_{\lambda,e,\mu}}, \nonumber
\end{align}
the problem is reduced to proving that $\eta_{\lambda,e,\mu}$ is fixed under the involution $\ol{\ \cdot \ }$. One can write
\begin{align}
\eta_{\lambda,e,\mu} = \sum_{w \in W_\lambda W_\mu} q_{w_\lambda} q_{e'} q_w\inv H_w = x_\lm \sum_{y \in {}^\lm W_\mu} q_{e'} q_y\inv H_y. \nonumber
\end{align}
On the other hand, we have
$$
x_\lm x_\mu = x_\lm x_{I_\lm \cap I_\mu} \sum_{y \in {}^\lm W_\mu} q_{e'} q_y\inv H_y = P_{I_\lm \cap I_\mu} x_\lm \sum_{y \in {}^\lm W_\mu} q_{e'} q_y\inv H_y.
$$
Hence, we obtain
$$
\eta_{\lm,e,\mu} = \frac{1}{P_{I_\lm \cap I_\mu}} x_\lm x_\mu,
$$
which is invariant under $\ol{\ \cdot \ }$. Thus, the proof completes.
\end{enumerate}
\end{proof}
\begin{prop}\label{flat on bbS}
The linear map $\flat:\bbS(\pi) \rightarrow \bbS(\pi);\ \xi_{\lambda,x,\mu} \mapsto \xi_{\mu,x\inv,\lambda}$ defines an $\AZ$-algebra anti-automorphism on $\bbS(\pi)$.
\end{prop}
\begin{proof}
We have to verify that $(\xi_{\lambda,x,\mu} \cdot \xi_{\kappa,y,\nu})^{\flat} = \xi_{\nu,y\inv,\kappa} \cdot \xi_{\mu,x\inv,\lambda}$ for all $\lambda,\mu,\nu,\kappa$ and $x \in {}^\lambda W^{\mu}$, $y \in {}^\kappa W^\nu$. Since the both sides are equal to zero unless $\kappa = \mu$, we may assume that $\kappa = \mu$. Let us write
\begin{align}\label{prod of xi}
&\xi_{\lambda,x,\mu} \cdot \xi_{\mu,y,\nu} = \sum_{z \in {}^\lambda W^\nu} c_z \xi_{\lambda,z,\nu} \qu \Forsome c_z \in \AZ.
\end{align}
Applying the both sides to $x_\nu \in \bbT(\pi)$, by Lemma \ref{property of eta} $(2)$, we obtain
\begin{align}\label{prod of eta}
\frac{1}{P_\mu} \eta_{\lambda,x,\mu} \eta_{\mu,y,\nu} = \sum_{z \in {}^\lambda W^\mu} c_z \eta_{\lambda,z,\nu}.
\end{align}
To prove the assertion, we compute as follows:
\begin{align}
\begin{split}
\xi_{\nu,y\inv,\mu} \cdot \xi_{\mu,x\inv,\lambda}(x_\lambda) &= \frac{1}{P_\mu}\eta_{\nu,y\inv,\mu} \cdot \eta_{\mu, x\inv,\lambda} \qqu (\text{by Lemma \ref{property of eta} $(2)$})\\
&= (\frac{1}{P_\mu}\eta_{\lambda,x,\mu} \cdot \eta_{\mu,y,\nu})^\flat \qqu (\text{by Lemma \ref{property of eta} $(1)$})\\
&= (\sum_{z \in {}^\lambda W^\nu} c_z \eta_{\lambda,z,\nu})^\flat \qqu (\text{by equation \eqref{prod of eta}})\\
&= \sum_{z \in {}^\lambda W^\nu} c_z \eta_{\nu,z\inv,\lambda} \qqu (\text{by Lemma \ref{property of eta} $(1)$})\\
&= \sum_{z \in {}^\lambda W^\nu} c_z \xi_{\nu,z\inv,\lambda}(x_\lambda) \qqu (\text{by Lemma \ref{property of eta} $(2)$})\\
&= (\xi_{\lambda,x,\mu} \cdot \xi_{\mu,y,\nu})^\flat(x_\lambda) \qqu (\text{by equation \eqref{prod of xi}}).
\end{split} \nonumber
\end{align}
This shows that $\xi_{\nu,y\inv,\mu} \cdot \xi_{\mu,x\inv,\lambda} = (\xi_{\lambda,x,\mu} \cdot \xi_{\mu,y,\nu})^\flat$, and hence, the proof completes.
\end{proof}
Recall the bilinear form $\la \cdot \mid \cdot \ra_\lambda = \la \cdot \mid \cdot \ra_{I_\lm}$ on $x_\lambda \Hc$ defined in Section \ref{Parabolic Kazhdan-Lusztig bases}.
\begin{prop}
Let $\lambda,\mu \in \pi$, $m \in x_\lambda \Hc$, and $n \in x_\mu \Hc$. Then, for each $w \in {}^\lambda W^\mu$, we have
\begin{align}
\la m \mid \xi_{\lambda,w,\mu}(n) \ra_\lambda = \la \xi_{\lambda,w,\mu}^\flat(m) \mid n \ra_\mu. \nonumber
\end{align}
\end{prop}
\begin{proof}
We compute as follows:
\begin{align}
\begin{split}
\la m \mid \xi_{\lambda,w,\mu}(n) \ra_\lambda
&= \frac{1}{P_\lambda} \la m \mid \xi_{\lambda,w,\mu}(n) \ra \\
&= \frac{1}{P_\lambda P_\mu} \la m \mid \eta_{\lambda,w,\mu} n \ra \qqu (\text{by Lemma \ref{property of eta} $(2)$})\\
&= \frac{1}{P_\lambda P_\mu} \la \eta_{\mu,w\inv,\lambda} m \mid n \ra \qqu (\text{by Lemma \ref{property of eta} $(1)$})\\
&= \frac{1}{P_\mu} \la \xi_{\mu,w\inv,\lambda}(m) \mid n \ra \qqu (\text{by Lemma \ref{property of eta} $(2)$})\\
&= \la \xi_{\lambda,w,\mu}^\flat(m) \mid n \ra_\mu. \nonumber
\end{split}
\end{align}
This proves the proposition.
\end{proof}
Define a bilinear form $\la \cdot \mid \cdot \ra_\pi$ on $\bbT(\pi)$ by $\la m \mid n \ra_\pi := \delta_{\lambda,\mu} \la m \mid n \ra_\lambda$ for all $\lambda,\mu \in \pi$, $m \in x_\lambda \Hc$, $n \in x_\mu \Hc$.
\begin{cor}
The two bases $\{ {}^\lambda C_w \mid \lambda \in \pi,\ w \in {}^\lambda W \}$ and $\{ {}^\lambda D_{w_\lambda w w_0} \mid \lambda \in \pi,\ w \in {}^\lambda W \}$ of $\bbT(\pi)$ are dual to each other with respect to the bilinear form $\la \cdot \mid \cdot \ra_\pi$. Moreover, for all $m,n \in \bbT(\pi)$ and $x \in \bbS(\pi)$, we have $\la m \mid xn \ra_\pi = \la x^\flat m \mid n \ra_\pi$.
\end{cor}
\subsection{Cell representations}
Let $X \in L(W)$ and $x \in X$. Set
\begin{align}
\begin{split}
&C_{\leq_L X}(\pi) := \bigoplus_{\lambda \in \pi} \bigoplus_{\substack{w \in {}^\lambda W \\ w_\lambda w \leq_L x}} \AZ {}^\lambda C_w,\qqu D_{\leq_L}(\pi) := \bigoplus_{\lambda \in \pi} \bigoplus_{\substack{w \in {}^\lambda W \\ w \leq_L x}} \AZ {}^\lambda D_w, \\
&C_{<_L X}(\pi) := \bigoplus_{\lambda \in \pi} \bigoplus_{\substack{w \in {}^\lambda W \\ w_\lambda w <_L x}} \AZ {}^\lambda C_w,\qqu D_{\leq_L}(\pi) := \bigoplus_{\lambda \in \pi} \bigoplus_{\substack{w \in {}^\lambda W \\ w <_L x}} \AZ {}^\lambda D_w, \\
&C^L_X(\pi) := C_{\leq_L X}(\pi)/C_{<_L X}(\pi),\qu\qqu D^L_X(\pi) := D_{\leq_L X}(\pi) / D_{<_L X}(\pi).
\end{split} \nonumber
\end{align}
Note that these objects are independent of the choice of $x \in X$. We denote the image of $m \in C_{\leq_L X}(\pi)$ (resp., $D_{\leq_L X}(\pi)$) under the quotient map $C_{\leq_L X}(\pi) \rightarrow C^L_X(\pi)$ (resp., $D_{\leq_L X}(\pi) \rightarrow D^L_X(\pi)$) by $[m]_X$ (resp., $[m]'_X$).
\begin{prop}
Let $X \in L(W)$.
\begin{enumerate}
\item $C_{\leq_L X}(\pi)$ is a $\bbS(\pi)$-submodule of $\bbT(\pi)$.
\item $C_{<_L X}(\pi)$ is a $\bbS(\pi)$-submodule of $\bbT(\pi)$.
\item $C^L_{X}(\pi)$ is a left $\bbS(\pi)$-module having a basis $\{ [{}^\lambda C_w]_X \mid \lambda \in \pi,\ w \in {}^\lambda W \cap w_\lambda X \}$. Here, $w_\lambda X := \{ w_\lambda x \mid x \in X \}$.
\end{enumerate}
\end{prop}
\begin{proof}
We will prove only $(1)$ since the proof of $(2)$ is similar to that of $(1)$, and $(3)$ follows from $(1)$ and $(2)$. Fix $x \in X$. In order to show that $C_{\leq_L X}(\pi)$ is a $\bbS(\pi)$-submodule, it suffices to verify that $\xi_{\lambda,y,\mu} {}^\mu C_w \in C_{\leq_L X}(\pi)$ for all $\lambda,\mu \in \pi$, $y \in {}^\lambda W^\mu$, and $w \in {}^\mu W$ such that $w_\mu w \leq_L x$. By Proposition \ref{JCw} and Lemma \ref{property of eta} (2), we have
\begin{align}
\xi_{\lambda,y,\mu} {}^\mu C_w = \xi_{\lambda,y,\mu} C_{w_\mu w} = \frac{1}{P_\mu} \eta_{\lambda,y,\mu} C_{w_\mu w}. \nonumber
\end{align}
Also, by Lemma \ref{property of eta} (2), we have $\eta_{\lambda,y,\mu} = \xi_{\lambda,y,\mu}(x_\mu) \in x_\lambda \Hc$; one can write $\eta_{\lambda,y,\mu} = x_\lambda H$ for some $H \in \Hc$. Then, $H C_{w_\mu w}$ is a linear combination of $C_{w'}$, $w' \leq_L w_\mu w (\leq_L x)$. Hence, by Proposition \ref{from KL to parabolic KL}, $x_\lambda H C_{w_\mu w}$ is a linear combination of ${}^\lambda C_{w''}$ for $w'' \in {}^\lambda W$ with $w_\lambda w'' \leq_L w' (\leq_L x)$. Therefore, we have $\xi_{\lambda,y,\mu} {}^\mu C_w = \frac{1}{P_\mu} \eta_{\lambda,y,\mu} C_{w_\mu w} \in \frac{1}{P_\mu} C_{\leq_L X}(\pi)$. However, since $\xi_{\lambda,y,\mu} {}^\mu C_w \in x_\lambda \Hc = \bigoplus_{z \in {}^\lambda W} \AZ {}^\lambda C_z$, we conclude that $\xi_{\lambda,y,\mu} {}^\mu C_w \in C_{\leq_L X}(\pi)$. This completes the proof.
\end{proof}
Similarly, one can prove the following: $D_{\leq_L X}(\pi)$ and $D_{<_L X}(\pi)$ are $\bbS(\pi)$-submodules, and $D^L_X(\pi)$ is a left $\bbS(\pi)$-module having a basis $\{ [{}^\lambda D_w]'_X \mid \lambda \in \pi,\ w \in {}^\lambda W \cap X \}$.
\section{Introduction}
Let $\U = U_q(\mathfrak{sl}_{2r+1})$ be the quantum group over the field $\Q(q)$ of rational functions in one variable $q$, and $\Uj$ its coideal subalgebra such that $(\U,\Uj)$ forms a quantum symmetric pair of type AIII in the sense of \cite{Le99}. Bao and Wang \cite{BW13} introduced the notion of $\jmath$-canonical bases for the based $\U$-modules. A based $\U$-module is a $\U$-module $M$ with a bar-involution $\psi_M$ and a distinguished basis $\bfB$ satisfying some conditions (see \cite{L94} for the precise definition). One of the key ingredients for the construction of the $\jmath$-canonical bases is the intertwiner (also known as the quasi-$K$-matrix) $\Upsilon$. Using $\Upsilon$, Bao and Wang defined a new involution $\psij_M := \Upsilon \circ \psi_M$ on $M$ that is compatible with the bar-involution $\psij$ on $\Uj$. Then, for each $b \in \bfB$, there exists a unique $b^{\jmath} \in M$ such that $\psij_M(b^{\jmath}) = b^{\jmath}$ and $b^{\jmath} - b \in \bigoplus_{\substack{b' \in \bfB \\ b' \prec^{\jmath} b}} q\Q[q] b'$, where $\preceq^{\jmath}$ is a partial order on $\bfB$. Clearly, $\{ b^{\jmath} \mid b \in \bfB \}$ is a basis of $M$, which is called the $\jmath$-canonical basis of $(M,\bfB)$.
The multi-parameter version of $\Uj$ was considered in \cite{BWW16}. Thanks to the integrality of the intertwiner $\Upsilon$, the notion of $\jmath$-canonical bases can be defined analogously. The condition $b^{\jmath} - b \in \bigoplus_{\substack{b' \in \bfB \\ b' \prec^{\jmath} b}} q\Q[q] b'$ is replaced by $b^{\jmath} - b \in \bigoplus_{\substack{b' \in \bfB \\ b' \prec^{\jmath} b}} (p \Q[p,q,q\inv] \oplus q\Q[q]) b'$. The general theory of the $\jmath$-canonical bases (usually called $\imath$-canonical bases) for the general quantum symmetric pairs was developed in \cite{BW18}.
In \cite{W17}, the author classified all irreducible $\Uj$-modules in a category $\Oj$, which is an analog of the category $\Oint$ of integrable $\U$-modules, and proved that $\Oj$ is semisimple; the isomorphism classes of irreducible modules in $\Oj$ are classified by the set $\Pj$ of bipartitions of length $(r;r+1)$. When the parameters are in the asymptotic case, to each irreducible module in $\Oj$, the author associated a local basis, $\jmath$-crystal basis, which is an analog of Kashiwara's crystal basis. By the complete reducibility, every object in $\Oj$ admits a $\jmath$-crystal basis. In particular, each $\U$-module in $\Oint$, regarded as a $\Uj$-module, has a $\jmath$-crystal basis.
It should be noted that the $\jmath$-crystal basis of a $\U$-module in $\Oint$ is the localized $\jmath$-canonical basis (\cite[Section 1.3]{W17}). To be precise, let $M \in \Oint$ with a canonical basis (or global crystal basis) $\bfB$. Since $(M,\bfB)$ is a based module, it has a $\jmath$-canonical basis $\{ b^\jmath \mid b \in \bfB \}$. Set $\clL := \Span_{\Ao} \bfB$, where $\Ao := \{ f/g \in \Q(p,q) \mid f,g \in p\Q[p,q,q\inv] \oplus \Q[q],\ \lim_{q \rightarrow 0}(\lim_{p\rightarrow 0} g) \neq 0 \}$. Then, $\clB := \{ b^{\jmath} + q\clL \mid b \in \bfB \}$ is a $\Q$-basis of $\clL/q\clL$, and $(\clL,\clB)$ forms a $\jmath$-crystal basis of $M$. Hence, $b^{\jmath} + q\clL$ can be thought of as the localization of $b^{\jmath}$ at $p = q = 0$. Conversely, we may say that the $\jmath$-canonical basis of a based $\U$-module is a globalization of its $\jmath$-crystal basis.
Here arises a natural question: Does a $\Uj$-module in $\Oj$ that is not a based $\U$-module admit a globalization of its $\jmath$-crystal basis? One of the main result of this paper gives the affirmative answer to this question.
In our strategy, the multi-parameter $q$-Schur duality between $\Uj$ and the Hecke algebra of type $B$ (\cite{BWW16}), and the irreducibility of the Kazhdan-Lusztig cell representations of the asymptotic multi-parameter Hecke algebra of type $B$ (\cite{BI03}) play key roles. Let us recall the latter objects briefly. Kazhdan and Lusztig \cite{KL79} gave a partition $W = \sqcup_{X \in L(W)} X$ of a Coxeter group $W$ into the left cells; here, $L(W)$ denotes the set of left cells. To each left cell $X \in L(W)$, they associated an $\Hc$-module $C^L_X$ which is called the left cell representation corresponding to $X$. The left cell representation $C^L_X$ is defined to be the quotient of a left ideal $C_{\leq_L X}$ of $\Hc$ spanned by some Kazhdan-Lusztig basis elements by its subspace $C_{<_L X}$, which is also spanned by some Kazhdan-Lusztig basis elements. Therefore, $C^L_X$ has a basis consisting of the images of some Kazhdan-Lusztig basis elements under the canonical map $C_{\leq_L X} \twoheadrightarrow C^L_X$. It is known that each left cell representation is irreducible if $W$ is of type $A$. When $W$ is of type $B$, the irreducibility of the left cell representations depend on the choice of the parameters $p,q$. According to \cite{BI03}, the left cell representations are irreducible when the parameters are asymptotic.
By the multi-parameter $q$-Schur duality for type $B$, the tensor power $\V^{\otimes d}$ of the vector representation of $\U$ is equipped with a $(\Uj,\bfH)$-bimodule structure whose irreducible decomposition is multiplicity free, where $\bfH$ denotes the multi-parameter Hecke algebra of type $B$ over the field $\Q(p,q)$ of rational functions in two variables $p,q$. Then, for each $X \in L(W)$, the left $\Uj$-module $\V^{\otimes d} \otimes_{\bfH} \bfC^L_X$ is irreducible, where $\bfC^L_X := \Q(p,q) \otimes_{\Z[p^{\pm1},q^{\pm1}]} C^L_X$. Every irreducible $\Uj$-module can be obtained in this way as $d \geq 1$ varies. The main result of this paper states that the basis of $\V^{\otimes d} \otimes_{\bfH} \bfC^L_X$ induced from the Kazhdan-Lusztig basis of $C^L_X$ is a globalization of the $\jmath$-crystal basis.
Our approach provides the following characterization of the $\jmath$-crystal bases and its globalization of the finite-dimensional irreducible $\Uj$-modules. Let $L \in \Oj$ be irreducible and $v \in L$ a highest weight vector. Define two symmetric bilinear forms $(\cdot, \cdot)_1$ and $(\cdot,\cdot)_2$ on $L$ and an involutive anti-linear automorphism $\psij_L$ on $L$ by
\begin{align}
\begin{split}
&(v,v)_1 = 1,\ (xm,n)_1 = (m,\sigmaj(x)n)_1 \qu \Forall x \in \Uj,\ m,n \in L, \\
&(v,v)_2 = 1,\ (xm,n)_2 = (m,\tauj(x)n)_2 \qu \Forall x \in \Uj,\ m,n \in L, \\
&\psij_L(v) = v,\ \psij_L(xm) = \psij(x) \psij_L(m) \qu \Forall x \in \Uj, m \in L,
\end{split} \nonumber
\end{align}
where $\sigmaj$, $\tauj$, and $\psij$ are automorphisms of $\Uj$ defined in Proposition \ref{automorphisms}.
\begin{Thm}\label{main}
Let $\bflm \in \Pj$, $L(\bflm)$ the corresponding irreducible $\Uj$-module, $(\clL(\bflm),\clB(\bflm))$ the $\jmath$-crystal basis of $L(\bflm)$ such that $v + q\clL(\bflm) \in \clB(\bflm)$. Then, there exist $\Gj(b)$, $b \in \clB(\bflm)$ satisfying the following hold.
\begin{enumerate}
\item $\clL = \{ m \in L \mid (m,m)_2 \in \Ao \}$.
\item $\clB$ forms an orthonormal basis of $\clL/q\clL$ with respect to the symmetric bilinear form induced from $(\cdot,\cdot)_2$.
\item Set $L_{\A} := \Span_{\A}\{ \Gj(b) \mid b \in \clB \}$, where $\A := \Q[p,p\inv,q,q\inv]$. Then, $(\clL,L_{\A},\psij_L(\clL))$ forms a balanced triple with the global basis $\{ \Gj(b) \mid b \in \clB \}$.
\item $L$ has the basis dual to $\Gj(\clB)$ with respect to $(\cdot,\cdot)_1$.
\end{enumerate}
\end{Thm}
Next, we investigate basic properties of global $\jmath$-crystal basis for not necessarily irreducible $\Uj$-modules. Especially, we roughly describe the matrix coefficients of the actions of the generators of $\Uj$ with respect to a given global $\jmath$-crystal basis.
We end this paper by proving that the global $\jmath$-crystal basis for a $\U$-module (not $\Uj$-module) is compatible with the filtration coming from the dominance order of the bipartitions (see subsection \ref{subsec: global basis} for the definition of this filtration). A similar result is well-known for ordinary global crystal bases (\cite{K93}, \cite{L94}).
\begin{Thm}
Let $M$ be a $\U$-module with a global $\jmath$-crystal basis $\Gj(\clB)$. Then, for each $\bflm \in \Pj$, the subquotient $W_\bflm(M)$ of $M$ has $\{ \Gj(b) + W_{\succ \bflm} \mid I(b) = \bflm \}$ as a global $\jmath$-crystal basis. Moreover, there exists an isomorphism $L(\bflm)^{\oplus m_\bflm} \rightarrow W_\bflm(M)$ which restricts to a bijection $\{ \Gj(b) \mid b \in \clB(\bflm) \}^{\oplus m_\bflm} \rightarrow \{ \Gj(b) + W_{\succ \bflm} \mid I(b) = \bflm \}$, where $m_\bflm$ denotes the multiplicity of $L(\bflm)$ in $M$.
\end{Thm}
In particular, if we take $M$ to be an irreducible $\Uj$-module, we obtain the following.
\begin{Cor}
Let $\bflm \in \Pj$. Then, $\{ \Gj(b) \mid b \in \clB(\bflm) \}$ is a unique global $\jmath$-crystal basis for $L(\bflm)$.
\end{Cor}
This paper is organized as follows. In Section \ref{notations}, we prepare necessary notations concerning (bi)partitions and Young (bi)tableaux. In Section \ref{RT for Uj} and \ref{crystal}, we give a brief review of \cite{W17}. In Section \ref{global bases}, we introduce the notion of global $\jmath$-crystal bases, and show that the $\jmath$-canonical bases are examples of them. Sections \ref{KL bases}--\ref{existence of global jcry bases} are devoted to prove the existence theorem for the global $\jmath$-crystal bases of the finite-dimensional irreducible $\Uj$-modules. After studying basic properties of the global $\jmath$-crystal bases in Section \ref{basic properties}, we finally prove the compatibility of the $\jmath$-crystal bases and the filtration associated to the dominance order of the bipartitions in Section \ref{sec:proof of main theorem}.
\subsection*{Acknowledgement}
This work is supported by JSPS KAKENHI grant number 17J00172.
\section{Global $\jmath$-crystal bases for the irreducible $\Uj$-modules}\label{existence of global jcry bases}
\subsection{Surjection $\xi:\Uj \rightarrow \bfS(\pij)$}
Let $\pij = \{ \lambda = (\lambda_0,\ldots,\lambda_r) \in \Z_{\geq 0}^{r+1} \mid \sum_{i=0}^r \lambda_i = d \}$. For $\lambda \in \pij$, set $I_\lambda = \{ 0,1,\ldots,d-1 \} \setminus \{ \lambda_0,\lambda_{0,1},\ldots,\lambda_{0,r-1} \}$, where $\lambda_{0,k} = \sum_{i=0}^k \lambda_i$.
Let $\bfV = \bigoplus_{i = -r}^r \Q(p,q) v_i$ be the vector representation of $\U$ with $v_{-r}$ a highest weight vector. Then, $\bfV^{\otimes d}$ has a basis $\{ v_{i_1,\ldots,i_d} := v_{i_1} \otimes \cdots \otimes v_{i_d} \mid -r \leq i_1,\ldots,i_d \leq r \}$. $\bfH := \Q(p,q) \otimes_{\AZ} \clH$ acts on $\bfV^{\otimes d}$ by
\begin{align}
\begin{split}
v_{i_1,\ldots,i_d} H_0 &= \begin{cases}
v_{-i_1,i_2,\ldots,i_d} \qu & \IF i_1 > 0, \\
p\inv v_{i_1,\ldots,i_d} \qu & \IF i_1 = 0, \\
v_{-i_1,i_2,\ldots,i_d} + (p\inv - p)v_{i_1,\ldots,i_d} \qu & \IF i_1 < 0,
\end{cases} \\
v_{i_1,\ldots,i_d} H_j &= \begin{cases}
v_{\ldots,i_{j+1},i_j,\ldots} \qu & \IF i_j < i_{j+1}, \\
q\inv v_{i_1,\ldots,i_d} \qu & \IF i_j = i_{j+1}, \\
v_{\ldots,i_{j+1},i_j,\ldots} + (q\inv - q)v_{i_1,\ldots,i_d} \qu & \IF i_j > i_{j+1}.
\end{cases}
\end{split} \nonumber
\end{align}
Then, it is easily seen that $\bfV^{\otimes d}$ is isomorphic to $\bfT(\pij) := \Q(p,q) \otimes_{\AZ} \bbT(\pij)$ as a right $\bfH$-module. Setting $\bfS(\pij) := \Q(p,q) \otimes_{\AZ} \bbS(\pij)$, $\bfV^{\otimes d}$ becomes a left $\bfS(\pij)$-module. By the double centralizer property between $\Uj$ and $\bfH$ on $\bfV^{\otimes d}$ (\cite{BW13}, \cite{BWW16}), there exists a surjective algebra homomorphism $\xi:\Uj \rightarrow \bfS(\pij)$. In particular, every $\bfS(\pij)$-modules are regarded as $\Uj$-modules via $\xi$. In \cite{W17}, it is proved that for each $\bflm \in \Pj$, the irreducible highest weight module $L(\bflm)$ is isomorphic to $\bfC^L_X(\pij) := \Q(p,q) \otimes_{\AZ} C^L_X(\pij)$ for some $X \in L(W_d)$, where $d = |\bflm|$.
For $i \in \Ij$, we define two maps $\etil_i,\ftil_i: \pij \rightarrow \pij \sqcup \{ 0 \}$, where $0$ denotes a formal symbol, as follows. Let $\lm = (\lm_0,\ldots,\lm_r) \in \pij$. Then, we set
$$
\etil_i \lm = \begin{cases}
(\lm_0,\ldots,\lm_{i-2},\lm_{i-1}+1,\lm_i-1,\lm_{i+1},\ldots,\lm_r) \qu & \IF \lm_i > 0, \\
0 \qu & \IF \lm_i = 0,
\end{cases}
$$
and
$$
\ftil_i \lm = \begin{cases}
(\lm_0,\ldots,\lm_{i-2},\lm_{i-1}-1,\lm_i+1,\lm_{i+1},\ldots,\lm_r) \qu & \IF \lm_{i-1} > 0, \\
0 \qu & \IF \lm_{i-1} = 0.
\end{cases}
$$
By convention, we set $\xi_{\lm,x,\mu} = 0$ if $\lm = 0$ or $\mu = 0$.
\begin{prop}
For $i \in \Ij$, we have
\begin{align}
\begin{split}
&\xi(e_i) = \sum_{\lambda \in \pij} \xi_{\etil_i(\lambda),e,\lambda}, \\
&\xi(f_i) = \sum_{\lambda \in \pij} \xi_{\ftil_i(\lambda),e,\lambda}.
\end{split} \nonumber
\end{align}
\end{prop}
\begin{proof}
We prove only the statement for $f_1$; the proofs for $f_i$, $i \neq 1$ and for $e_i$ are similar.
Recall the comultiplication $\Delta$ of $\bfU$; we have
\begin{align}
\Delta^{(d-1)}(E_i) = \sum_{k = 1}^d 1^{\otimes k-1} \otimes E_i \otimes (K_i\inv)^{\otimes d-k},\ \Delta^{(d-1)}(F_i) = \sum_{k = 1}^d K_i^{\otimes d-k} \otimes F_i \otimes 1^{\otimes k-1}. \nonumber
\end{align}
Then, we compute as
\begin{align}
\begin{split}
f_1v_\lambda &= pq\inv q^{\lambda_0} \sum_{k=1}^{\lambda_0} q^{\lambda_0-k} v_{0^{\lambda_0-k},1,0^{k-1},1^{\lambda_1},\ldots,r^{\lambda_r}} \\
&\qu + \sum_{k=1}^{\lambda_0} q^{\lambda_0-k} v_{0^{k-1},-1,0^{\lambda_0-k},1^{\lambda_1},\ldots,r^{\lambda_r}} \\
&= pq^{\lambda_0-1}\sum_{k=1}^{\lambda_0} q^{\lambda_0-k} v_{\ftil_1(\lambda)} H_{\lambda_0-1} \cdots H_{\lambda_0-(k-1)} \\
&\qu + \sum_{k=1}^{\lambda_0} q^{\lambda_0-k} v_{\ftil_1(\lambda)} H_{\lambda_0-1} \cdots H_1 H_0 H_1 \cdots H_{k-1} \\
&= pq^{2(\lambda_0-1)} \sum_{k=1}^{\lambda_0} q^{-k+1} v_{\ftil_1(\lambda)} H_{\lambda_0-1} \cdots H_{\lambda_0-(k-1)} \\
&\qu + pq^{2(\lambda_0-1)}\sum p\inv q^{-(\lambda_0+k-2)} v_{\ftil_1(\lambda)} H_{\lambda_0-1} \cdots H_1 H_0 H_1 \cdots H_{k-1} \\
&= \xi_{\ftil_1(\lambda),e,\lambda}(v_{\lambda}).
\end{split} \nonumber
\end{align}
This proves the assertion.
\end{proof}
Here are immediate consequences.
\begin{cor}
Let $x \in \Uj$. Then, $\xi(\sigmaj(x)) = \xi(x)^\flat$, $\xi(\psij(x)) = \ol{\xi(x)}$.
\end{cor}
\begin{cor}
The bilinear form $\la \cdot \mid \cdot \ra_{\pij}$ of $\bfT(\pij)$ satisfies
\begin{align}
\la xm \mid n \ra_{\pij} = \la m \mid \sigmaj(x)n \ra_{\pij} \nonumber
\end{align}
for all $x \in \Uj$, $m,n \in \bfT(\pij)$.
\end{cor}
\subsection{Global $\jmath$-crystal basis of irreducible $\Uj$-module}
Let $X \in L(W)$. Then, $\bfC^L_X(\pij) \simeq L(\bflm)$ for some $\bflm \in \Pj$. Since $L(\bflm)$ is a highest weight module, there exists a unique $\lambda \in \pij$ and $w \in {}^\lambda W$ such that $[{}^\lambda C_w]_X \in \bfC^L_X(\pij)$ is a highest weight vector.
Recall the isomorphism $D^L_{Xw_0} \simeq C^L_X$ of left $\clH$-modules. Set $\bfC^L_X := \Q(p,q) \otimes_{\AZ} C^L_X$, and define $\bfD^L_{Xw_0}$ and $\bfD^L_{Xw_0}(\pij)$ similarly. Then, we have
$$
\bfD^L_{Xw_0}(\pij) \simeq \bfT(\pij) \otimes_{\bfH} \bfD^L_{Xw_0} \simeq \bfT(\pij) \otimes_{\bfH} \bfC^L_X \simeq \bfC^L_X(\pij)
$$
as left $\Uj$-modules. Hence, $[{}^\lambda D_{w_\lambda w w_0}]'_{Xw_0} \in \bfD^L_{Xw_0}(\pij)$ is also a highest weight vector. Thus, we obtain two isomorphisms
\begin{align}
\begin{split}
&\vphi_C: L(\bflm) \rightarrow \bfC^L_X(\pij);\ v_\bflm \mapsto [{}^\lm C_w]_X, \\
&\vphi_D: L(\bflm) \rightarrow \bfD^L_{Xw_0}(\pij);\ v_\bflm \mapsto [{}^\lm D_{w_\lm ww_0}]'_{Xw_0}
\end{split} \nonumber
\end{align}
of $\Uj$-modules, where $v_\bflm \in L(\bflm)$ is a fixed highest weight vector.
\begin{defi}\normalfont
Let $\bflm \in \Pj$ and $v_\bflm \in L(\bflm)$ be a highest weight vector. Define the bilinear form $(\cdot,\cdot)_1$ on $L(\bflm)$ by $(v_\bflm,v_\bflm)_1 = 1$ and $(xm,n)_1 = (n,\sigmaj(x)n)_1$ for all $x \in \Uj$, $m,n \in L(\bflm)$.
\end{defi}
\begin{prop}\label{(,)_1 is nondegenerate}
Let $\bflm \in \Pj$. Then, the bilinear form $(\cdot,\cdot)_1$ is nondegenerate.
\end{prop}
\begin{proof}
For $m,n \in L(\bflm)$, set $(m,n) := \la \vphi_C(m) \mid \vphi_D(n) \ra_{\pij}$. Then, we have
$$
(v_\bflm,v_\bflm) = \la [{}^\lm C_w]_X \mid [{}^\lm D_{w_\lm ww_0}]'_{Xw_0} \ra_{\pij} = 1,
$$
and
$$
(xm,n) = \la x \vphi_C(m) \mid \vphi_D(m) \ra_{\pij} = \la \vphi_C(m) \mid \sigmaj(x) \vphi_D(n) \ra_{\pij} = (m,\sigmaj(x) n).
$$
Hence, we have $(\cdot,\cdot) = (\cdot,\cdot)_1$. Then, it is clear that $\{ \vphi_C\inv([{}^\mu C_y]_X) \mid \mu \in \pij,\ y \in {}^\mu W \cap w_\mu X \}$ and $\{ \vphi_D\inv([{}^\mu D_{w_\mu yw_0}]'_{Xw_0}) \mid \mu \in \pij,\ y \in {}^\mu W \cap Xw_0 \}$ form bases which are dual to each other with respect to $(\cdot,\cdot)_1$. This proves the proposition.
\end{proof}
Recall that the set $\{ (\mu,y) \mid \mu \in \pij,\ y \in {}^\mu W \cap w_\mu X \}$ is identical to $\clB(\bflm)$. For each $b \in \clB(\bflm)$, set
$$
\Gjlow(b) := \vphi_C\inv([{}^\mu C_y]_X), \qu \Gjup(b) := \vphi_D\inv([{}^\mu D_{w_\mu yw_0}]'_{Xw_0}),
$$
where $(\mu,y)$ is the pair corresponding to $b$. Then, $\Gjlow(\bflm) := \{ \Gjlow(b) \mid b \in \clB(\bflm) \}$ and $\Gjup(\bflm) := \{ \Gjup(b) \mid b \in \clB(\bflm) \}$ are bases of $L(\bflm)$.
\begin{defi}\normalfont
Let $\bflm \in \Pj(d)$, and $v_{\bflm} \in L(\bflm)$ be a highest weight vector. Define a bilinear form $(\cdot,\cdot)_2$ on $L(\bflm)$, and a $\psij$-involution $\psij_{\bflm}$ on $L(\bflm)$ by
\begin{align}
\begin{split}
&(v_{\bflm},v_{\bflm})_2 = 1,\ (xm,n)_2 = (m,\tauj(x)n)_2 \qu \Forall x \in \Uj,\ m,n \in L(\bflm), \\
&\psij_{\bflm}(v_{\bflm}) = v_{\bflm}.
\end{split} \nonumber
\end{align}
\end{defi}
Let $(\clL(\bflm),\clB(\bflm))$ be the unique $\jmath$-crystal basis of $L(\bflm)$ such that $v_\bflm + q\clL(\bflm) \in \clB(\bflm)$.
\begin{theo}
Let $\bflm \in \Pj(d)$. Then, the following hold.
\begin{enumerate}
\item $\psij_\bflm(\Gjlow(b)) = \Gjlow(b)$ for all $b \in \clB(\bflm)$.
\item $\psij_\bflm(\Gjup(b)) = \Gjup(b)$ for all $b \in \clB(\bflm)$.
\item $\Gjlow(\bflm)$ and $\Gjup(\bflm)$ are dual bases with respect to $(\cdot,\cdot)_1$.
\item $\clL(\bflm) = \{ m \in L(\bflm) \mid (m,m)_2 \in \Ao \}$. Consequently, $(\cdot,\cdot)_2$ induces the bilinear form $(\cdot,\cdot)_0$ on $\clL(\bflm)/q\clL(\bflm)$ defined by $(m+q\clL(\bflm),n+q\clL(\bflm))_0 := \lim_{q \rightarrow 0}(\lim_{p \rightarrow 0}(m,n)_2)$.
\item $\{ \Gjlow(b) \mid b \in \clB(\bflm) \}$ forms an almost orthonormal basis with respect to $(\cdot,\cdot)_2$, i.e., we have $(\Gjlow(b),\Gjlow(b'))_2 \in \delta_{b,b'} + q\Ao$ for all $b,b' \in \clB(\bflm)$.
\item $(b,b')_0 = \delta_{b,b'}$ for all $b,b' \in \clB(\bflm)$.
\item Let $L(\bflm)_\A$ be the $\A$-span of $\Gjlow(\bflm)$. Then, $(\clL(\bflm),L(\bflm)_\A,\psij_\bflm(\clL(\bflm)))$ is balanced. Moreover, the global basis associated to $\clB(\bflm)$ is $\{ \Gjlow(b) \mid b \in \clB(\bflm) \}$. In particular, $L(\bflm)$ has a global $\jmath$-crystal basis.
\end{enumerate}
\end{theo}
\begin{proof}
Items $(1)$ and $(2)$ are obvious from the definition of $\Gjlow(b)$ and $\Gjup(b)$. Item $(3)$ follows from the proof of Proposition \ref{(,)_1 is nondegenerate}.
To prove the rest, observe that $L(\bflm)$ is realized as a subquotient of $\bfV^{\otimes d}$ by using Kazhdan-Lusztig basis elements. To be precise, let $X \in \Pj$ be such that $L(\bflm) \simeq \bfC^L_X(\pij)$ and $x \in X$. Then,
$$
\bfC^L_X(\pij) = \frac{\Span_{\Q(p,q)}\{ {}^\lm C_w \mid \lm \in \pij,\ w \in {}^\lm W,\ w_\lm w \leq_L x \}}{\Span_{\Q(p,q)}\{ {}^\mu C_y \mid \mu \in \pij,\ y \in {}^\mu W,\ w_\mu y <_L x \}}.
$$
Then, items $(4)$-$(6)$ follows from the proof of \cite[Proposition 7.4.4]{W17}. To prove item $(7)$, it suffices to show that $L(\bflm)_\A$ is a $\UjA$-module. It follows from the fact that the $\A$-submodule of $\bfV^{\otimes d}$ spanned by the Kazhdan-Lusztig basis is a $\UA$-module, and that $\UjA \subset \UA$.
\end{proof}
\section{Kazhdan-Lusztig bases}\label{KL bases}
The subsequent three sections are dedicated to prove the existence of a global $\jmath$-crystal basis and its ``dual'' basis for $L(\bflm)$, $\bflm \in \Pj$. In this section, we formulate variants of the Kazhdan-Lusztig bases following \cite{KL79}, \cite{Deo87}, and \cite{L}.
\subsection{Hecke algebra of type $B$}
Fix $d \in \Z_{> 0}$. Let $W = W_d$ be the Weyl group of type $B_d$ with simple reflections $S = \{ s_0,s_1,\ldots,s_{d-1} \}$ such that
$$
s_0s_1s_0s_1 = s_1s_0s_1s_0, \qu s_is_{i+1}s_i = s_{i+1}s_is_{i+1} \IF i \geq 1,\qu s_is_j = s_js_i \IF |i-j| > 1.
$$
\begin{defi}\normalfont
The Hecke algebra $\clH = \clH(W)$ associated to $W$ with unequal parameters $p,q$ is the associative algebra over $\AZ := \Z[p,p\inv,q,q\inv]$ generated by $\{ H_s \mid s \in S \}$ subject to the following relations:
\begin{itemize}
\item $(H_s - q_s\inv)(H_s + q_s) = 0$ for all $s \in S$, where $q_s = p$ if $s = s_0$ and $q_s = q$ otherwise.
\item $H_{s_0}H_{s_1}H_{s_0}H_{s_1} = H_{s_1}H_{s_0}H_{s_1}H_{s_0}$.
\item $H_{s_i}H_{s_{i+1}}H_{s_i} = H_{s_{i+1}}H_{s_i}H_{s_{i+1}}$ if $i \geq 1$.
\item $H_{s_i}H_{s_j} = H_{s_j}H_{s_i}$ if $|i-j| > 1$.
\end{itemize}
\end{defi}
We often write $H_i = H_{s_i}$. For each $w \in W$ with a reduced expression $w = s_{i_1} \cdots s_{i_l}$, the product $H_{i_1} \cdots H_{i_l}$ is independent of the choice of a reduced expression of $w$; we denote it by $H_w$. Similarly, $q_w := q_{s_{i_1}} \cdots q_{s_{i_l}}$ is well-defined.
Let $U,V$ be modules over $\AZ$. We say a $\Z$-linear map $f:U \rightarrow V$ is anti-linear if it satisfies $f(gu) = \ol{g} f(u)$ for all $g \in \AZ$ and $u \in U$. In the sequel, we will often use the following automorphisms, all of which are involutions, of $\clH$.
\begin{lem}\label{automorphisms on H}
\
\begin{enumerate}
\item There exists a unique anti-linear algebra automorphism $\ol{\ \cdot \ }$ of $\clH$ such that $\ol{H_w} = H_{w\inv}\inv$.
\item There exists a unique anti-linear algebra automorphism $\sgn$ of $\clH$ such that $\sgn(H_w) = (-1)^{\ell(w)}H_{w}$. Here, $\ell:W \rightarrow \Z_{\geq 0}$ denotes the length function on $W$.
\item There exists a unique $\AZ$-algebra anti-automorphism $(\cdot)^\flat$ of $\clH$ such that $H_w^\flat = H_{w\inv}$.
\end{enumerate}
Moreover, all of these automorphisms commute with each other.
\end{lem}
For $y,w \in W$, define $r_{y,w} \in \AZ$ by
\begin{align}
\ol{H_w} = \sum_{y \in W} r_{y,w} H_y. \nonumber
\end{align}
It is well-known and easily proved that $r_{w,w} = 1$ for all $w \in W$ and $r_{y,w} = 0$ unless $y \leq w$.
\subsection{Kazhdan-Lusztig bases}
Let us formulate the Kazhdan-Lusztig basis and the dual Kazhdan-Lusztig basis. Set
\begin{align}
\begin{split}
\AZ^+ &:= \AZ \cap \Ao = p\Z[p,q,q\inv] \oplus q\Z[q], \\
\AZ^- &:= \ol{\AZ^+} = p\inv\Z[p\inv,q,q\inv] \oplus q\inv \Z[q\inv].
\end{split} \nonumber
\end{align}
\begin{theo}[{\cite[Theorem 1.1]{KL79}}, {\cite[Theorem 5.2]{L}}]\label{KL-bases}
For each $w \in W$, there exists a unique $C_w \in \clH$ such that
\begin{enumerate}
\item $\overline{C_w} = C_w$.
\item $C_w = H_w + \sum_{y < w} c_{y,w} H_y$ for some $c_{y,w} \in \AZ^+$. Here, $<$ denotes the Bruhat order on $W$.
\end{enumerate}
\end{theo}
Replacing $\AZ^+$ with $\AZ^-$, we see the following: For each $w \in W$, there exists a unique $D_w \in \clH$ such that
\begin{enumerate}
\item $\overline{D_w} = D_w$.
\item $D_w = H_w + \sum_{y < w} d_{y,w} H_y$ for some $d_{y,w} \in \AZ^-$.
\end{enumerate}
\begin{rem}\label{C-basis and D-basis}\normalfont
Noting that the automorphisms $\ol{\ \cdot \ }$ and $\sgn$ commute with each other, it is easy to verify that $D_w = (-1)^{\ell(w)} \sgn(C_w)$.
\end{rem}
It is obvious from the definitions that both $\{ C_w \mid w \in W \}$ and $\{ D_w \mid w \in W \}$ form $\AZ$-bases of $\clH$. We call the former the Kazhdan-Lusztig basis, and the latter the dual Kazhdan-Lusztig basis of $\clH$.
\subsection{Left cell representations}
Let us recall from \cite{KL79} the notion of left cells of $W$ and the associated left cell representations.
\begin{defi}\normalfont
Let $y,w \in W$.
\begin{enumerate}
\item $y \rightarrow_L w$ if the coefficient of $C_y$ in $C_s C_w$ expanded in the Kazhdan-Lusztig basis is nonzero for some $s \in S$.
\item $y \leq_L w$ if there exist $y = y_0,y_1,\ldots,y_l = w \in W$ such that $y_{i-1} \rightarrow_L y_i$.
\item $y \Lcell w$ if $y \leq_L w$ and $w \leq_L y$.
\item $y <_L w$ if $y \leq_L w$ and $y \not\Lcell w$.
\item Each equivalence class of $W/\Lcell$ is called a left cell of $W$. We denote by $L(W)$ the set of left cells of $W$.
\end{enumerate}
\end{defi}
\begin{rem}\normalfont
By Remark \ref{C-basis and D-basis}, we obtain the same equivalence relation as $\Lcell$ if we replace $C_w$'s by $D_w$'s.
\end{rem}
For each $X \in L(W)$ and $x \in X$, set
\begin{align}
&C_{\leq_L X} = \bigoplus_{y \leq_L x} \AZ C_y, \qu C_{<_L X} = \bigoplus_{y <_L x} \AZ C_y, \qu C^L_X = C_{\leq_L X} / C_{<_L X}, \nonumber\\
&D_{\leq_L X} = \bigoplus_{y \leq_L x} \AZ D_y, \qu D_{<_L X} = \bigoplus_{y <_L x} \AZ D_y, \qu D^L_X = D_{\leq_L X} / D_{<_L X}. \nonumber
\end{align}
Note that these are independent of the choice of $x \in X$. We denote the image of $m \in C_{\leq_L X}$ (resp., $m \in D_{\leq_L X}$) under the quotient map $C_{\leq_L X} \rightarrow C^L_X$ (resp., $D_{\leq_L X} \rightarrow D^L_X$) by $[m]_X$ (resp., $[m]'_X$).
\begin{lem}
Let $X \in L(W)$. Then, $C_{\leq_L X}$, $C_{<_L X}$, $D_{\leq_L X}$, and $D_{<_L X}$ are left ideals of $\clH$, and consequently, $C^L_{X}$ and $D^L_{X}$ are left $\clH$-modules. Moreover, $C^L_{X}$ has a basis $\{ [C_x]_X \mid x \in X \}$, while $D^L_{X}$ has a basis $\{ [D_x]'_X \mid x \in X \}$.
\end{lem}
\begin{proof}
The assertions are obvious from the definitions.
\end{proof}
We call $C^L_X$ the left cell representation of $\clH(W)$ associated to $X \in L(W)$.
\subsection{Bilinear form on $\clH$}
Let $\clH^* := \Hom_{\AZ}(\clH,\AZ)$. $\clH^*$ has a left $\clH$-module structure given by
\begin{align}
(Hf)(H') = f(H^\flat H'), \qu \Forall f \in \clH^*,\ H,H' \in \clH. \nonumber
\end{align}
Let $\{ h_w \mid w \in W \} \subset \clH^*$ be the dual basis of $\{ H_w \mid w \in W \}$, that is, they are characterized by $h_y(H_w) = \delta_{y,w}$ for all $y,w \in W$.
\begin{lem}\label{module structure of V*}
For each $w \in W$ and $s \in S$, the following holds.
\begin{align}
H_s h_w = \begin{cases}
h_{sw} \qu & \IF w < sw, \\
h_{sw} + (q_s\inv - q_s)h_w \qu & \IF sw < w.
\end{cases} \nonumber
\end{align}
\end{lem}
\begin{proof}
For each $y \in W$, we compute as
\begin{align}
\begin{split}
(H_sh_w)(H_y) &= h_w(H_sH_y) \\
&= \begin{cases}
h_w(H_{sy}) \qu & \IF sy > y, \\
h_w(H_{sy} + (q_s\inv-q_s)H_y) \qu & \IF sy < y
\end{cases} \\
&= \begin{cases}
1 \qu & \IF sy > y \AND sy = w, \\
1 \qu & \IF sy < y \AND sy = w, \\
q_s\inv - q_s \qu & \IF sy < y \AND y = w, \\
0 \qu & \OW
\end{cases} \\
&= \begin{cases}
h_{sw}(H_y) \qu & \IF sw > w, \\
(h_{sw} + (q_s\inv - q_s)h_w)(H_y) \qu & \IF sw < w.
\end{cases}
\end{split} \nonumber
\end{align}
This implies
\begin{align}
H_sh_w = \begin{cases}
h_{sw} \qu & \IF sw > w, \\
h_{sw} + (q_s\inv - q_s)h_w \qu & \IF sw < w.
\end{cases} \nonumber
\end{align}
Thus, the proof completes.
\end{proof}
There exists an anti-linear automorphism $\ol{\ \cdot \ }$ of $\clH^*$ defined by $\ol{f}(H) = \ol{f(\ol{H})}$ for $f \in \clH^*$, $H \in \clH$.
\begin{lem}\label{bar on V*}
For each $w \in W$, we have
\begin{align}
\ol{h_w} = \sum_{y \geq w} \ol{r_{w,y}} h_y. \nonumber
\end{align}
In particular, $\ol{h_{w_0}} = h_{w_0}$, where $w_0 \in W$ denotes the longest element.
\end{lem}
\begin{proof}
Let $y \in W$. Then, we have
\begin{align}
\ol{h_w}(H_y) = \ol{h_w(\ol{H_y})} = \ol{h_w(\sum_{z \leq y} r_{z,y} H_z)} = \ol{r_{w,y}}. \nonumber
\end{align}
Since $\ol{h_w} = \sum_{y \in W} \ol{h_w}(H_y) h_y$, the assertion follows.
\end{proof}
Let $\{ C_w^* \mid w \in W \} \subset \clH^*$ denote the dual basis of $\{C_w \mid w \in W \}$.
\begin{prop}\label{characterization of C^*}
$C_w^*$ is characterized by the following two conditions:
\begin{enumerate}
\item $\ol{C_w^*} = C_w^*$.
\item $C_w^* = h_w + \sum_{z > w} c_{w,z}^* h_z$ for some $c_{w,z}^* \in \AZ^+$.
\end{enumerate}
\end{prop}
\begin{proof}
Thanks to Lemma \ref{bar on V*}, one can prove that there exists a unique $C'_w \in \clH^*$ such that $\ol{C'_w} = C'_w$ and $C'_w - h_w \in \bigoplus_{y > w} \AZ^+ h_y$ in a similar way to Theorem \ref{KL-bases}. Hence, it suffices to show that $C_w^*$ satisfies the two conditions.
The first condition is verified as follows. For each $y \in W$, we have
\begin{align}
\ol{C_w^*}(C_y) = \ol{C_w^*(\ol{C_y})} = \ol{C_w^*(C_y)} = \ol{\delta_{y,w}} = \delta_{y,w} = C_w^*(C_y). \nonumber
\end{align}
Since $\{ C_y \mid y \in W \}$ is a basis of $\clH$, we obtain $\ol{C_w^*} = C_w^*$.
Next, we prove the second condition. For each $y \in W$, we can write $H_y = C_y + \sum_{z < y} b_{z,y}C_z$ for some $b_{z,y} \in \AZ^+$. Then, we have
\begin{align}
C_w^* = \sum_{y \in W} C_w^*(H_y)h_y = h_w + \sum_{y > w} b_{w,y} h_y. \nonumber
\end{align}
This completes the proof.
\end{proof}
\begin{lem}\label{Duality of Hecke}
The linear map $d\index{dHeckealgebra@$d$ (Hecke algebra)}:\clH \rightarrow \clH^*;\ H \mapsto H \cdot h_{w_0}$ gives an isomorphism of left $\clH$-modules. Moreover, we have
\begin{enumerate}
\item $d(\ol{H_y}) = h_{yw_0}$ for all $y \in W$.
\item $d(\ol{H}) = \ol{d(H)}$ for all $H \in \clH$.
\end{enumerate}
\end{lem}
\begin{proof}
By Lemma \ref{module structure of V*}, the linear map $\vphi:\clH \rightarrow \clH^*;\ H_w \mapsto h_w$ is an isomorphism of left $\clH$-modules. On the other hand, the map $\psi:\clH \rightarrow \clH;\ H \mapsto H \cdot H_{w_0}$ is clearly an isomorphism of left $\clH$-modules. Thus, the composite map $d:= \vphi \circ \psi: \clH \rightarrow \clH^*$ is an isomorphism of left $\clH$-modules satisfying
\begin{align}
d(H) = \vphi(H \cdot H_{w_0}) = H \cdot \vphi(H_{w_0}) = H \cdot h_{w_0} \qu \Forall H \in \clH. \nonumber
\end{align}
Also, we have, for all $y \in W$,
\begin{align}
d(\ol{H_y}) = \vphi(\ol{H_y} \cdot H_{w_0}) = \vphi(H_{y\inv}\inv \cdot H_{y\inv} H_{yw_0}) = \vphi(H_{yw_0}) = h_{yw_0}. \nonumber
\end{align}
Finally, for each $H,H' \in \clH$, we have
\begin{align}
\begin{split}
&d(\ol{H})(H') = (\ol{H}\cdot h_{w_0})(H') = h_{w_0}\left( \left( \ol{H} \right)^\flat H' \right), \\
&\ol{d(H)}(H') = \ol{d(H)(\ol{H'})} = \ol{h_{w_0}(H^\flat \ol{H'})} = \ol{h_{w_0}} \left( \ol{H^\flat} H' \right).
\end{split} \nonumber
\end{align}
Then, the equality $d(\ol{H}) = \ol{d(H)}$ follows from the facts that $\ol{h_{w_0}} = h_{w_0}$ and $\left( \ol{H} \right)^\flat = \ol{H^\flat}$; the former is proved in Lemma \ref{bar on V*}, and the latter is in Lemma \ref{automorphisms on H}.
\end{proof}
Using this isomorphism, we define a bilinear form $\la \cdot \mid \cdot \ra$ on $\clH$ by
\begin{align}
\la H \mid H' \ra := d(H')(H), \qu (H,H' \in \clH). \nonumber
\end{align}
Clearly, this bilinear form satisfies $\la H' \mid HH'' \ra = \la H^\flat H' \mid H'' \ra$ for all $H,H',H'' \in \clH$.
\begin{lem}
The bilinear from $\la \cdot \mid \cdot \ra$ is symmetric.
\end{lem}
\begin{proof}
Let $H_1,H_2 \in \clH$. It suffices to show that $h_{w_0}(H_2^\flat H_1) = h_{w_0}(H_1^\flat H_2)$. Since $H_{w_0}^\flat = H_{w_0}$, it holds that $h_{w_0}(H^\flat) = h_{w_0}(H)$ for all $H \in \clH$. Then, the assertion follows if one notes $(H_2^\flat H_1)^\flat = H_1^\flat H_2$.
\end{proof}
\begin{prop}\label{duality of KL-bases}
The bases $\{ C_w \mid w \in W \}$ and $\{ D_{ww_0} \mid w \in W \}$ are dual to each other with respect to $\la \cdot \mid \cdot \ra$, that is, we have $\la C_y \mid D_w \ra = \delta_{y,ww_0}$ for all $y,w \in W$.
\end{prop}
\begin{proof}
Recall that $D_w = \sum_{y \leq w} d_{y,w} H_y$ with $d_{w,w} = 1$ and $d_{y,w} \in \AZ^-$ for all $y < w$. Then, we have
\begin{align}
\begin{split}
&\ol{d(D_w)} = d(\ol{D_w}) = d(D_w), \\
&d(D_w) = d(\ol{D_w}) = d(\sum_{y \leq w} \ol{d_{y,w}} \ol{H_y}) = \sum_{y \leq w} \ol{d_{y,w}} h_{yw_0} = \sum_{z \geq ww_0} \ol{d_{zw_0,w}} h_{z}.
\end{split} \nonumber
\end{align}
This and Proposition \ref{characterization of C^*} show that $d(D_w) = C^*_{ww_0}$. Hence, it holds that $\la C_y \mid D_w \ra = C^*_{ww_0}(C_y) = \delta_{y,ww_0}$, which proves the proposition.
\end{proof}
Here, we describe the duality between $C^L_X$'s and $D^L_X$'s.
\begin{lem}\label{left cell and w_0}
Let $y,w \in W$, $X \in L(W)$. Then, the following hold.
\begin{enumerate}
\item $y \rightarrow_L w$ if and only if $ww_0 \rightarrow_L yw_0$.
\item $y \leq_L w$ if and only if $ww_0 \leq_L yw_0$.
\item $Xw_0 := \{ xw_0 \mid x \in X \} \in L(W)$.
\end{enumerate}
\end{lem}
\begin{proof}
We first prove part $(1)$. Suppose that $y \rightarrow_L w$. Then, there exists $s \in S$ such that $\la C_s C_w \mid D_{yw_0} \ra \neq 0$. This implies that $\la C_w \mid C_s D_{yw_0} \ra \neq 0$, and hence, we obtain $ww_0 \rightarrow_L yw_0$. Replacing $y,w$ by $yw_0,ww_0$, we also have the opposite indication. This proves part $(1)$. Assertion $(2)$ is an immediate consequence of $(1)$. We prove part $(3)$. Let $x \in X$. Then, $X = \{ y \in W \mid x \leq_L y \leq_L x \}$. By part $(2)$, we have $x \leq_L y \leq_L x$ if and only if $xw_0 \leq_L yw_0 \leq_L xw_0$. This implies that $Xw_0 = \{ z \in W \mid xw_0 \leq_L z \leq_L xw_0 \}$, and it is a unique left cell of $W$ containing $xw_0$. Thus, the proof completes.
\end{proof}
\begin{lem}
The bilinear from $\la \cdot \mid \cdot \ra$ induces a non-degenerate bilinear form on $C^L_X \times D^L_{Xw_0}$. Moreover, $\{ [C_x]_X \mid x \in X \}$ and $\{ [D_{xw_0}]_{Xw_0}' \mid x \in X \}$ form bases which are dual to each other.
\end{lem}
\begin{proof}
Let $x \in X$, $y,w \in W$ be such that $y <_L x$ and $ww_0 <_L xw_0$. It suffices to show that $\la C_y \mid D_u \ra = 0$ for all $u \leq_L xw_0$ and $\la C_v \mid D_{ww_0} \ra = 0$ for all $v \leq_L x$. Both are obvious from Lemma \ref{left cell and w_0} $(2)$.
\end{proof}
\begin{prop}
Let $X \in L(W)$. Then, we have an isomorphism $D^L_{Xw_0} \simeq C^L_X$ of $\clH$-modules.
\end{prop}
\begin{proof}
It suffices to show that the characters $\ch_{D^L_{Xw_0}}$ of $D^L_{Xw_0}$ and $\ch_{C^L_X}$ of $C^L_X$ coincide with each other. For each $w \in W$, we compute as
\begin{align}
\begin{split}
\ch_{C^L_X}(H_w) &= \sum_{x \in X} \la H_w [C_x]_X \mid [D_{xw_0}]'_{Xw_0} \ra \\
&= \sum_{x \in X} \la [C_x]_X \mid H_{w\inv} [D_{xw_0}]_{Xw_0}' \ra \\
&= \ch_{D^L_{Xw_0}}(H_{w\inv}) = \ch_{D^L_{Xw_0}}(H_w).
\end{split} \nonumber
\end{align}
Thus, the proof completes.
\end{proof}
\subsection{Parabolic Kazhdan-Lusztig bases}\label{Parabolic Kazhdan-Lusztig bases}
Throughout this subsection, we fix a subset $J \subset \{ 0,1,\ldots,d-1 \}$ arbitrarily. Let $W_J$ denote the parabolic subgroup of $W$ generated by $\{ s_j \mid j \in J \}$, ${}^JW$ the set of minimal length coset representatives for $W_J \backslash W$, and $w_J \in W_J$ the longest element. Also, we set
\begin{align}
x_J := q_{w_J} \sum_{w \in W_J} q_w\inv H_w \in \clH. \nonumber
\end{align}
\begin{lem}\label{property of x_J}
Let $j \in J$. Then, the following hold.
\begin{enumerate}
\item $x_J H_j = q_{s_j}\inv x_J$.
\item $x_J^\flat = x_J$.
\item $x_J = C_{w_J}$. In particular, $\ol{x_J} = x_J$.
\end{enumerate}
\end{lem}
\begin{proof}
The assertion $(1)$ follows from a direct calculation and the fact that $W_J = \{ w \in W_J \mid w < s_jw \} \sqcup \{ w \in W_J \mid s_jw < w \}$. The assertion $(2)$ follows from the definition of $x_J$ and the facts that $W_J = \{ w\inv \mid w \in W_J \}$, and $q_{w\inv} = q_w$ for all $w \in W$. The proof of $(3)$ can be found in \cite[Proposition 1.17 (2)]{X94}.
\end{proof}
By Lemma \ref{property of x_J} $(1)$, the right ideal $x_J \clH$ of $\clH$ has a basis $\{ x_J H_w \mid w \in {}^JW \}$. Also, by Lemma \ref{property of x_J} $(3)$, $x_J \clH$ is closed under the involution $\ol{\ \cdot \ }$. Hence, we can construct analogs of the Kazhdan-Lusztig basis and the dual Kazhdan-Lusztig basis of $\clH$ in the ideal $x_J \clH$:
\begin{theo}{\cite[Proposition 3.2]{Deo87}}\label{parabolic KL basis}
\
\begin{enumerate}
\item For each $w \in {}^JW$, there exists a unique ${}^JC_w \in x_J \clH$ such that
\begin{enumerate}
\item $\ol{{}^JC_w} = {}^JC_w$.
\item ${}^JC_w = x_J(H_w + \sum_{\substack{y \in {}^JW \\ y < w}} {}^Jc_{y,w} H_y)$ for some ${}^Jc_{y,w} \in \AZ^+$.
\end{enumerate}
\item For each $w \in {}^JW$, there exists a unique ${}^JD_w \in x_J \clH$ such that
\begin{enumerate}
\item $\ol{{}^JD_w} = {}^JD_w$.
\item ${}^JD_w = x_J(H_w + \sum_{\substack{ y \in {}^JW \\ y < w}} {}^Jd_{y,w} H_y)$ for some ${}^Jd_{y,w} \in \AZ^-$.
\end{enumerate}
\end{enumerate}
\end{theo}
Clearly, $\{ {}^JC_w \mid w \in {}^JW \}$ and $\{ {}^JD_w \mid w \in {}^JW \}$ are linear bases of $x_J \clH$. We call them the parabolic Kazhdan-Lusztig basis and the dual parabolic Kazhdan-Lusztig basis of $x_J \clH$, respectively.
\begin{prop}{\cite[Proposition 3.4]{Deo87}}\label{JCw}
Let $w \in {}^JW$. Then, ${}^JC_w = C_{w_Jw}$.
\end{prop}
\begin{prop}\label{JDw}
Let $w \in {}^JW$. Then, ${}^JD_w = x_J D_w$.
\end{prop}
\begin{proof}
For each $y \in W$, define $y_J \in W_J$ and ${}^Jy \in {}^JW$ to be the unique elements satisfying $y = y_J {}^Jy$ and $\ell(y) = \ell(y_J) + \ell({}^Jy)$. Then, we have
\begin{align}
\begin{split}
x_J D_w &= x_J \sum_{y \leq w} d_{y,w} H_y \\
&= x_J\left( H_w + \sum_{y < w} d_{y,w} H_{y_J}H_{{}^Jy} \right) \\
&= x_J\left( H_w + \sum_{y < w} q_{y_J}\inv d_{y,w} H_{{}^Jy} \right) \qu (\text{by Lemma \ref{property of x_J} $(1)$})\\
&= x_J\left( H_w + \sum_{\substack{y \in {}^JW \\ y < w}} \sum_{\substack{x \in W_J \\ xy < w}} q_x\inv d_{xy,w} H_y \right).
\end{split} \nonumber
\end{align}
This shows that $x_JD_w - x_JH_w \in \bigoplus_{\substack{y \in {}^JW \\ y < w}} \AZ^- x_JH_y$. Hence, by Theorem \ref{parabolic KL basis} $(2)$, $x_JD_w$ coincides with ${}^JD_w$.
\end{proof}
For a later use, let us consider $x_J C_y$ and $x_J D_y$ for $y \in W$.
\begin{prop}\label{from KL to parabolic KL}
Let $y \in W$. Then, we have
\begin{align}
x_J C_y = \sum_{\substack{w \in {}^JW \\ w_Jw \leq_L y}} \alpha_w {}^JC_w, \nonumber
\end{align}
for some $\alpha_w \in \AZ$.
\end{prop}
\begin{proof}
Let us write
\begin{align}
x_J C_y = \sum_{w \in {}^JW} \alpha_w {}^JC_w = \sum_{w \in {}^JW} \alpha_w C_{w_Jw} \qu \Forsome \alpha_w \in \AZ. \nonumber
\end{align}
Also, by the definition of $\leq_L$, we can write
\begin{align}
x_J C_y = \sum_{z \leq_L y} \beta_z C_z \qu \Forsome \beta_z \in \AZ. \nonumber
\end{align}
This shows $\alpha_w = 0$ unless $w_Jw \leq_L y$.
\end{proof}
\begin{lem}{\cite[Theorem 6.6 $(b)$]{L}}\label{HsDw}
Let $w \in W$ and $s \in S$ be such that $sw < w$. Then, it holds that $H_s D_w = -q_s D_w$.
\end{lem}
\begin{prop}
Let $y \in W \setminus {}^JW$. Then, $x_J D_y = 0$.
\end{prop}
\begin{proof}
Since $y \notin {}^JW$, there exists $j \in J$ such that $s_jy < y$. For such $j$, we have $x_J H_j = q_j\inv x_J$ (Lemma \ref{property of x_J} (1)) and $H_j D_y = -q_j D_y$ (Lemma \ref{HsDw}). Hence, we obtain
\begin{align}
x_J D_y = q_j x_J H_j D_y = -q_j^2 x_J D_, \nonumber
\end{align}
which implies $x_J D_y = 0$, as desired.
\end{proof}
Set $P_J := q_{w_J} \sum_{x \in W_J} q_x^{-2} \in \AZ$. Note that, by Lemma \ref{property of x_J} $(1)$, it holds that $x_J^2 = P_J x_J$. Then, for each $H,H' \in \clH$, we have
\begin{align}
\la x_JH \mid x_JH' \ra = \la x_J^2 H \mid H' \ra = P_J \la H \mid H' \ra \in P_J \AZ; \nonumber
\end{align}
here, we use Lemma \ref{property of x_J} $(2)$. Hence, we can define a $\AZ$-valued bilinear form $\la \cdot \mid \cdot \ra_J$ on $x_J \clH$ by $\la \cdot \mid \cdot \ra_J := \frac{1}{P_J} \la \cdot \mid \cdot \ra$.
\begin{prop}\label{dual of parabolic KL}
The basis $\{ {}^JC_w \mid w \in {}^JW \}$ and $\{ {}^JD_{w_Jww_0} \mid w \in {}^JW \}$ are dual to each other with respect to $\la \cdot \mid \cdot \ra_J$, that is, we have $\la {}^JC_y \mid {}^JD_{w} \ra_J = \delta_{y,w_Jww_0}$ for all $y,w \in {}^JW$.
\end{prop}
\begin{proof}
Let $y,w \in {}^JW$. We compute as follows:
\begin{align}
\begin{split}
\la {}^JC_y \mid {}^JD_w \ra_J &= \frac{1}{P_J} \la {}^JC_y \mid {}^JD_w \ra \\
&= \frac{1}{P_J} \la C_{w_Jy} \mid x_J D_w \ra \qu (\text{by Proposition \ref{JCw} and \ref{JDw}}) \\
&= \la C_{w_Jy} \mid D_w \ra \qu (\text{since $C_{w_Jy} = {}^JC_y \in x_J \clH$}) \\
&= \delta_{w_Jy,ww_0} = \delta_{y,w_Jww_0} \qu (\text{by Proposition \ref{duality of KL-bases}}).
\end{split} \nonumber
\end{align}
This proves the proposition.
\end{proof}
\section{Proof of Theorem \ref{filtration}}\label{sec:proof of main theorem}
For a $\Uj$-module $M$ with a global $\jmath$-crystal basis $\Gj(\clB)$, and for $m \in M$, $b \in \clB$, let $[m:\Gj(b)]$ denote the coefficient of $\Gj(b)$ in $m$.
\subsection{The $r=1$ case}
In this subsection, we prove Theorem \ref{filtration} for $r=1$.
\begin{proof of filtration}
We proceed by descending induction on $\bflm$ with respect to $\preceq$. Assume that the statement holds for all $\bflm' \succ \bflm$. Replacing $M$ with $M/W_{\succ \bflm}(M)$, we may assume that $\bflm$ is maximal among $\Pj(M)$. Let $b_1,\ldots,b_{m_\bflm} \in \clB$ and $u_1,\ldots,u_{m_\bflm} \in \clL$ be distinct highest weight vectors of type $\bflm$ with $u_i + q\clL = b_i$, $i = 1,\ldots,m_\bflm$. By retaking the $u_i$'s if necessary, we may assume that $[u_i:\Gj(b_j)] = \delta_{i,j}$ for all $i,j$. Fix $i$ arbitrarily, and set $b := b_i$, $u := u_i$. Then, we can write
$$
u = \Gj(b) + \sum_{\substack{b' \\ I^\jmath(b') \not\succeq \bflm}} c_{b'}\Gj(b'), \qu c_{b'} \in q\Ao.
$$
We first prove that $c_{b'} = 0$ for all $b'$ with $\vep_1(b') = 0$. Assume contrary, and take $b' \in \clB \setminus \{ b \}$ such that $c_{b'} \neq 0$, $\vep_1(b') = 0$, and $\vphi_1(b')$ is minimal among $\{ \vphi_1(b'') \mid c_{b''} \neq 0,\ \vep_1(b'') = 0 \}$. Set $\bfmu := I^\jmath(b')$. Then, we have $\wtj(\bfmu) = \wtj(\bflm)$, in particular, $\bfmu_0 = \bflm_0$. Since $\bfmu \not\succeq \bflm$, we have $\vphi_1(b') = \bfmu_0-\bfmu_{-1} > \bflm_0-\bflm_{-1} = \vphi_1(b)$. Hence, we have
\begin{align}
\begin{split}
-f_1^{(\vphi_1(b)+1)}\Gj(b) &= c_{b'} \bigl( \Gj(\ftil_1^{\vphi_1(b)+1}b') + \sum_{\substack{b'' \\ \vep_1(b'') > \vphi_1(b)+1}} d_{b'',b'} \Gj(b'') \bigr) \\
&\qu + \sum_{b''' \neq b'} \sum_{\substack{b'''' \\ \vep_1(b'''') \geq \vep_1(b''') + \vphi_1(b)+1}} d_{b'''',b'''} \Gj(b''''),
\end{split} \nonumber
\end{align}
for some $d_{b_1,b_2} \in \A$. By our assumption, the coefficient of $\Gj(\ftil_1^{\vphi_1(b)+1}b')$ in the right-hand side is equal to $c_{b'}$. On the other hand, the left-hand side is fixed by $\psij_M$, and it belongs to $M_{\A}$. Therefore, we have $c_{b'} \in q\A_0 \cap \A$ and $\ol{c_{b'}} = c_{b'}$, which implies $c_{b'} = 0$.
Next, we prove that $c_{b'} = 0$ for all $b'$ with $\vep_1(b') > 0$. Assume contrary that $c_{b'} \neq 0$ for some such $b'$. Set $\bfmu := I^{\jmath}(b')$. Since $\bflm$ is maximal, we have $\bfmu_0 + \bfmu_{-1} < \bflm_0 + \bflm_{-1}$. Substituting $\bfmu_0 = \bflm_0 + \vep_1(b')$, $\bfmu_{-1} = \bflm_0 - \vphi_1(b')$, and $\bflm_0 - \bflm_{-1} = \vphi_1(b)$, we obtain $\vphi_1(b') > \vphi_1(b) + \vep_1(b')$. We may assume that $(\vep_1(b'),\vphi_1(b'))$ is minimal (with respect to the lexigographical order) among such $b'$'s. Then, for all $t = 1,\ldots, \vep_1(b')+1$, we have
\begin{align}
-f_1^{(\vphi_1(b)+t)}\Gj(b) = c_{b'}{\vep_1(b')+\vphi_1(b)+t \brack \vep_1(b')} \Gj(\ftil_1^{\vphi_1(b)+t} b') + \otherterms. \nonumber
\end{align}
This implies that $c_{b'} \in q\A_0$, $\ol{c_{b'}} = c_{b'}$, and $c_{b'}{\vep_1(b')+\vphi_1(b)+t \brack \vep_1(b')} \in \A$ for all $t = 1,\ldots, \vep_1(b')+1$. Now, it suffices to show that $c_{b'} \in \A$, which follows from next lemma.
This far, we have proved that $\Gj(b) = u$, and hence, we have $e_1 \Gj(b) = 0$ and $\Uj_1 \Gj(b) \simeq L(\bflm)$. Then, for all $n = 1,\ldots, \bflm_0-\bflm_{-1}$, we have
\begin{align}
f_1^{(n)} \Gj(b) = f_1^{(n)} u = \ftil_1^n u. \nonumber
\end{align}
The left-hand side belongs to $M_{\A}$, while the right-hand side belongs to $\clL$. Moreover, we have $\psij_M(f_1^{(n)} \Gj(b)) = f_1^{(n)} \Gj(b)$, and $\ftil_1^n u + q\clL = \ftil_1^n b$. This implies that $f_1^{(n)} \Gj(b) = \Gj(\ftil_1^n b)$. Thus, the proof completes.
\end{proof of filtration}
\begin{lem}\label{lemma}
Let $A \in \Q(p,q)$, $m \geq n \in \Z_{\geq 0}$. Suppose that $A {m+t \brack n} \in \A$ for all $t = 1,\ldots, n+1$. Then, we have $A \in \A$.
\end{lem}
\begin{proof}
Let us write $A = B/C$ for some $B,C \in \A_0 \cap \A$ that are coprime. By the hypothesis, $C$ is a common devisor of ${m+t \brack n}$, $t = 1,\ldots,n+1$. Hence, it suffices to show that the greatest common divisor of them in $\Z[q]$ is equal to $1$. This is equivalent to say that the greatest common divisor of $a_t := [m+t][m+t-1] \cdots [m+t-n+1]$, $t = 1,\ldots,n+1$ is equal to $[n]!$. Since $[l] = q^{-l} \prod_{1 \neq d | l} \Phi_d$, where $\Phi_d = \Phi_d(q^2)$ denotes the $d$-th cyclotomic polynomial in variable $q^2$, we have
$$
b_t := q^{n(m+t)-\frac{n(n-1)}{2}}a_t = \prod_{l=0}^{n-1} \prod_{1 \neq d | (m+t-l)} \Phi_d,
$$
which is the irreducible decomposition of $b_t$ in $\Z[q^2]$. Then, we have
$$
b_t = \prod_{d \geq 2} \Phi_d^{m_{d,t}}, \text{ where } m_{d,t} := |\{ 0 \leq l \leq n-1 \mid d | (m+t-l) \}|,
$$
and hence,
$$
\gcd_{1 \leq t \leq n+1}(b_t) = \prod_{d \geq 2} \Phi_d^{\min_{1 \leq t \leq n+1}(m_{d,t})}.
$$
We prove that $\min_{1 \leq t \leq n+1}(m_{d,t}) = \lfloor \frac{n}{d} \rfloor$ for all $d$. It is clear that $m_{d,t} \geq \lfloor \frac{n}{d} \rfloor$ for all $t$ since $\{ m+t,m+t-1,\ldots,m+t-(\lfloor \frac{n}{d} \rfloor d-1) \}$ contains exactly $\lfloor \frac{n}{d} \rfloor$ integers divisible by $d$. If $\min_{1 \leq t \leq n+1}(m_{d,t}) > \lfloor \frac{n}{d} \rfloor$, then $\{ m+t-\lfloor \frac{n}{d} \rfloor d, m+t-(\lfloor \frac{n}{d} \rfloor d+1), \ldots, m+t-(n-1) \}$ contains at least one multiple of $d$ for all $t$. Then, for $t=1$, there exists $l_1 \in \{ \lfloor \frac{n}{d} \rfloor d, \lfloor \frac{n}{d} \rfloor d+1,\ldots,n-1 \}$ such that $m+1-(\lfloor \frac{n}{d} \rfloor d + l_1) \in d\Z$. Set $t' := n-l_1 + 1$, and consider the integers
$$
m+t'-\lfloor \frac{n}{d} \rfloor d, m+t'-(\lfloor \frac{n}{d} \rfloor d+1), \ldots, m+t'-(n-1) = (m+1-l_1) + 1.
$$
These are $(n-\lfloor \frac{n}{d} \rfloor d)$ consecutive integers with $(m+1-l_1) + 1 = 1$ modulo $d$. Since $n-\lfloor \frac{n}{d} \rfloor d < d$, they have no multiples of $d$. Hence, we have $\min_{1 \leq t \leq n+1}(m_{d,t}) = \lfloor \frac{n}{d} \rfloor$ for all $d \geq 2$. Thus, we obtain
$$
\gcd_{1 \leq t \leq n+1}(b_t) = \prod_{d \geq 2} \Phi_d^{\lfloor \frac{n}{d} \rfloor} = \prod_{d=2}^n \Phi_d^{\lfloor \frac{n}{d} \rfloor} = \prod_{l=2}^n \left(\prod_{1 \neq d' | l} \Phi_{d'}\right) = \prod_{l=2}^n [l] = [n]!.
$$
This proves the lemma.
\end{proof}
\subsection{The $r \geq 2$ case}
Now, we are ready to prove Theorem \ref{filtration} by induction on $r$.
When $r=1$, we have already completed the proof. Let $r \geq 2$ and assume that the assertions hold for all $r' < r$.
\begin{lem}\label{weak lemma}
Let $\bflm \in \Pj(M)$ be a maximal element, $b \in \clB$ such that $I^\jmath(b) = \bflm$ and $\etil_i b = 0$ for all $i \in \Ij$. Suppose the following:
\begin{enumerate}
\item There exists a homomorphism $\xi : L(\bflm) \rightarrow M$ of $\Uj$-modules such that $\xi(\Gjlow(\Tj_{b'})) = \Gj(b')$ for all $b' \in \Cj(b)$ which is strongly connected to some $b'' \in \Cj(b)$ with $\wtj(b) <^\jmath \wtj(b'')$.
\item $\xi$ commutes with the $\psij$-involutions on $L(\bflm)$ and $M$.
\item $[\xi(\Gjlow(\Tj_b)) : \Gj(b)] = 1$.
\end{enumerate}
Then, we have
$$
\xi(\Gjlow(\Tj_b)) = \Gj(b) + \sum_{\substack{b' \in \clB \setminus \{ b \} \\ \Tj_{b'} = \Tj_b}} c_{b'} \Gj(b') + \sum_{\substack{b'' \in \Cj(b'),\ c_{b' \neq 0} \\ s(b'') < s(b')}} c_{b''} \Gj(b'').
$$
for some $c_{b'},c_{b''} \in \Ao$.
\end{lem}
\begin{proof}
Since $\Uj$-module homomorphisms preserve $\jmath$-crystal lattices, we have $\xi(\Gjlow(\Tj_b)) \in \clL$, and $\xi(\Gjlow(\Tj_b)) + q\clL = b$. Let us write
$$
\xi(\Gjlow(\Tj_b)) = \Gj(b) + \sum_{b' \in \clB \setminus \{b\}} c_{b'} \Gj(b')
$$
for some $c_{b'} \in q\Ao$. Also, since $\xi$ commutes with $\psij$-involutions, we have $\ol{c_b} = c_b$, $\ol{c_{b'}} = c_{b'}$. We claim the following: if $b' \in \clB \setminus \{b\}$ satisfies
$$
(\dagger) \qu \text{$c_{b'} \neq 0$ and $s(b')$ is maximal among $\{ s(b'') \mid b'' \in \clB \setminus \{b\} \AND c_{b''} \neq 0 \}$},
$$
then $\Tj_{b'}(-i) \geq \bflm_{-i}$ for all $i = 0,1,\ldots,r$. By the case $r = 1$, we have $I^\jmath_1(b') \succeq I^\jmath_1(b)$, which implies $\Tj_{b'}(0) = \Tj_{b}(0) = \bflm_0$, and $\Tj_{b'}(-1) \geq \Tj_b(-1) = \bflm_{-1}$. We proceed by induction on $i$. Assume that $i \geq 2$, and that $\Tj_{b'}(-(i-1)) \geq \bflm_{-(i-1)}$ for all $b'$ satisfying $(\dagger)$. Suppose that there exists $b'$ satisfying $(\dagger)$ such that $\Tj_{b'}(-i) < \bflm_{-i}$. Let $b'' \in \clB \setminus \{b\}$ be such that $s(b'') = s(b')$ and $\vphi_{-\ul{b''}}$ is minimal among such elements. Recall that $s(b'') = s(b')$ implies $\wt(b'') = \wt(b')$, and hence, $\Tj_{b''}(-i) = \Tj_{b'}(-i) < \bflm_{-i}$. Then, we have
$$
\vep_{-\ul{i}}(b'') = \vphi_{-\ul{i}}(b'') + \Tj_{b''}(-(i-1)) - \Tj_{b''}(-i) > \Tj_{b''}(-(i-1)) - \bflm_{-i} + \vphi_{-\ul{i}}(b'').
$$
By the minimality of $\vphi_{-\ul{i}}(b'')$, it holds that
$$
[f_i^{(t)} \sum_{b' \in \clB \setminus \{b\}} c_{b''} \Gj(b') : \Gj(\ftil_i^t b'')] = c_{b'} {t \brack \vphi_{-\ul{i}}(b'')} \neq 0
$$
for all $\Tj_{b''}(-(i-1)) - \bflm_{-i} +1 \leq t \leq \Tj_{b''}(-(i-1)) - \bflm_{-i} + \vphi_{-\ul{i}}(b'') + 1$. On the other hand, $f_i^{(t)} \Gjlow(\Tj_b)$ is the sum of $\Gjlow(\Tj_{\ftil_i^t b})$ and an $\A$-linear combination of $\Gjlow(\Tj_{\bhat})$ such that $\bhat \in \Cj(b_t)$ is strongly connected to $b''' \in \Cj(b_t)$ with $\wtj(b) <^\jmath \wtj(b''')$. Hence, we have
\begin{align}
\begin{split}
\xi(\Gjlow(\Tj_{\ftil_i^t b})) &= f_i^{(t)} \xi(\Gjlow(\Tj_b)) + \sum_{\bhat} a_{\bhat} \Gj(\bhat) \\
&= f_i^{(t)} \Gj(b) + f_i^{(t)} \sum_{b' \in \clB \setminus \{b\}} c_{b'} \Gj(b') + \sum_{\bhat} a_{\bhat} \Gj(\bhat)
\end{split} \nonumber
\end{align}
for some $a_{\bhat} \in \A$. Here, note that we have $\etil_j \ftil_i^tb = 0$ for all $j = 1,\ldots,i-1$, $[\xi(\Gjlow(\Tj_{\ftil_i^t b})) : \Gj(\ftil_i^t b)] = 1$, and $s(\ftil_i^t b'')$ is maximal among $\{ s(b''') \mid b''' \neq \ftil_i^tb \AND [\xi(\Gjlow(\Tj_{\ftil_i^t b})) : \Gj(b''')] \neq 0 \}$. Then, by our induction hypothesis on $i$, we obtain that $\Tj_{\ftil_i^t b''}(-(i-1)) \geq \Tj_{\ftil_i^t b}(-(i-1)) = \bflm_{-i}$, which is a contradiction since $t \geq \Tj_{b''}(-(i-1)) - \bflm_{-i} + 1$. Hence we must have $[\xi(\Gjlow(\Tj_{\ftil_i^t b})) : \Gj(\ftil_i^t b'')] = 0$. Since
$$
[\xi(\Gjlow(\Tj_{\ftil_i^t b})) : \Gj(\ftil_i^t b'')] = c_{b''} {t \brack \vphi_{-\ul{i}}(b'')} + [f_i^{(t)} \Gj(b) : \Gj(\ftil_i^t b'')] + a_{\ftil_i^t b''},
$$
and the second and the third term of the right-hand side lies in $\A$, we obtain
$$
c_{b''} {t \brack \vphi_{-\ul{i}}(b'')} \in \A
$$
for all $\Tj_{b''}(-(i-1)) - \bflm_{-i} + 1 \leq t \leq \Tj_{b''}(-(i-1)) - \bflm_{-i} + \vphi_{-\ul{i}}(b'') + 1$. By Lemma \ref{lemma}, this implies $c_{b''} = 0$.
This far, we have proved that if $b' \in \clB \setminus \{b\}$ satisfies $(\dagger)$, then we have $\Tj_{b'}(-i) \geq \bflm_{-i}$ for all $i \in \{ 0,1,\ldots,r \}$. In particular, we have $\Ij(b') = \bflm$ for such $b'$ (since $\bflm$ is maximal in $\Pj(M)$). In this case, the condition $\Tj_{b'}(-i) \geq \bflm_{-i}$ for all $i$ forces $b'$ to satisfy that $\Tj_{b'} = \Tj_b$. Hence, we have
$$
\xi(\Gjlow(\Tj_b)) = \Gj(b) + \sum_{\substack{b' \in \clB \setminus \{ b \} \\ \Tj_{b'} = \Tj_b}} c_{b'} \Gj(b') + \sum_{\substack{b'' \in \Cj(b'),\ c_{b' \neq 0} \\ s(b'') < s(b')}} c_{b''} \Gj(b''),
$$
as desired.
\end{proof}
\begin{lem}\label{strong lemma}
Let $\bflm \in \Pj(M)$ be a maximal element, $j \in \Ij \setminus \{ 1 \}$, $b \in \clB$ such that $I^\jmath(b) = \bflm$, $\etil_i b = 0$ for all $i \in \Ij$, and $\etil_{j'}(b) \neq 0$. Suppose the following:
\begin{enumerate}
\item There exists a homomorphism $\xi : L(\bflm) \rightarrow M$ of $\Uj$-modules such that $\xi(\Gjlow(\Tj_{b'})) = \Gj(b')$ for all $b' \in \Cj(b)$ which is strongly connected to some $b'' \in \Cj(b)$ with $\wtj(b) <^\jmath \wtj(b'')$.
\item $\xi$ commutes with the $\psij$-involutions on $L(\bflm)$ and $M$.
\end{enumerate}
Then, we have
$$
\xi(\Gjlow(\Tj_b)) = \Gj(b) + \sum_{\substack{b' \in \clB \setminus \{ b \} \\ \Tj_{b'} = \Tj_b}} c_{b'} \Gj(b') + \sum_{\substack{b'' \in \Cj(b'),\ c_{b' \neq 0} \\ s(b'') < s(b')}} c_{b''} \Gj(b'').
$$
for some $c_{b'},c_{b''} \in \Ao$.
\end{lem}
\begin{proof}
If we can prove that $c_b := [\xi(\Gjlow(\Tj_b)) : \Gj(b)] = 1$, then the assertion follows from the previous lemma. Hence, we aim to show $c_b = 1$.
By the same argument as before, we have
$$
[\xi(\Gjlow(\Tj_{\ftil_i^t b})) : \Gj(\ftil_i^t b'')] = c_{b''} {t \brack \vphi_{-\ul{i}}(b'')} + c_b[f_i^{(t)} \Gj(b) : \Gj(\ftil_i^t b'')] + a_{\ftil_i^t b''},
$$
for all $b'' \in \clB \setminus \{b\}$ satisfying $(\dagger)$. Here, let us assume further that $s(b') > s(b)$. Then, we have $[f_i^{(t)} \Gj(b) : \Gj(\ftil_i^t b'')] = 0$ since $f_i^{(t)} \Gj(b)$ is a linear combination of $\Gj(\btil)$ with $s(\btil) \leq s(b) + t\alpha_{-\ul{i}} < s(b'') + t \alpha_{-\ul{i}} = s(\ftil_i^t b'')$. Hence, we have $c_{b''}{t \brack \vphi_{-\ul{i}}(b'')} \in \A$, and therefore, $c_{b''} = 0$ by Lemma \ref{lemma}. In particular, we obtain that $s(b)$ is maximal. Then, we have
$$
[e_j(c_b \Gj(b) + \sum c_{b'} \Gj(b')) : \Gj(\etil_{j'}b)] = c_b[\vphi_{\ul{j}}(b) + 1].
$$
On the other hand, since $[e_j \Gjlow(\Tj_b) : \Gjlow(\etil_{j'}b)] = [\vphi_{\ul{j}}(b) + 1]$, we have
$$
[e_j(c_b \Gj(b) + \sum c_{b'} \Gj(b')) : \Gj(\etil_{j'}b)] = [\vphi_{\ul{j}}(b) + 1],
$$
and hence, $c_{b} = 1$, as desired.
\end{proof}
We prove Theorem \ref{filtration} by descending induction (with respect to $\preceq$) on $\bflm$. As in the $r=1$ case, we may assume that $\bflm$ is maximal among $P^\jmath(M)$. Then, in order to complete the proof, we have to show the following:
\begin{enumerate}
\item $I_\bflm(M)$ has a basis $\{ \Gj(b) \mid I^\jmath(b) = \bflm \}$.
\item There exists an isomorphism $\xi:L(\bflm)^{\oplus m_\bflm} \rightarrow I_\bflm(M)$ of $\Uj$-modules which sends the $\jmath$-global basis elements of $L(\bflm)^{\oplus m_\bflm}$ to those of $I_\bflm(M)$, where $m_\bflm$ denotes the multiplicity of $L(\bflm)$ in $M$.
\end{enumerate}
Let $b_1,\ldots,b_{m_\bflm} \in \clB$ and $u_1,\ldots,u_{m_\bflm} \in \clL$ be distinct highest weight vectors of type $\bflm$ with $u_t + q\clL = b_t$, $t = 1,\ldots,m_\bflm$. By retaking the $u_t$'s if necessary, we may assume that $[u_t:\Gj(b_u)] = \delta_{t,u}$ for all $t,u$. Let $\xi_t : L(\bflm) \rightarrow M$ be the $\Uj$-homomorphism which sends $v_\bflm$ to $u_t$.
\begin{lem}\label{lemma 1}
We have $\xi_t(\Gjlow(\Tj_{b_t})) = \Gj(b_t)$ for all $t = 1,\ldots,m_\bflm$.
\end{lem}
\begin{proof}
By the setting above, we can write
$$
\xi_t(\Gjlow(\Tj_{b_t})) = u_t = \Gj(b_t) + \sum_{\substack{b' \\ I^\jmath(b') \not\succeq \bflm}} c_{b'}\Gj(b'), \qu c_{b'} \in q\Ao.
$$
Then, we can apply Lemma \ref{weak lemma} to obtain $\xi_t(\Gjlow(\Tj_{b_t})) = \Gj(b_t)$ as desired.
\end{proof}
In order to complete the proof, it suffices to prove the following: For each $t = 1,\ldots,m_\bflm$ and $b \in \Cj(b_t)$, we have $\xi_t(\Gjlow(\Tj_b)) = \Gj(b)$. We prove this statement by descending induction on $\wtj(b)$ and $I^{\jmath}_{r-1}(b)$. When $\wtj(b)$ is maximal, it must hold that $b = b_t$, and in this case, we have already shown that $\xi_t(\Gjlow(\Tj_{b_t})) = \Gj(b_t)$. Suppose that $\wtj(b) <^\jmath \wtj(b_t)$, and the statement holds for all $b' \in \bigsqcup_{t=1}^{m_\bflm} \Cj(b_t)$ such that $\wtj(b') {}^\jmath > \wtj(b)$ or $\wtj(b') = \wtj(b)$ and $I^\jmath_{r-1}(b') \succ I^\jmath_{r-1}(b)$. In this case, since $b$ is not a $\Uj$-highest weight vector, the exists $i \in \IIj$ such that $\etil_i b \neq 0$.
\begin{lem}
Suppose there exists $i \in \Ij$ such that $\etil_i b \neq 0$. Then, the statement holds.
\end{lem}
\begin{proof}
Set $b' := \etil_i^{\vep_i(b)} b$. We prove the lemma by descending induction on $\vep_i(b')$. Since $\wtj(b') > \wtj(b)$, we have $\Gj(b') = \xi_t(\Gjlow(\Tj_{b'})) \in \Uj \Gj(b_i)$. We know that $\Gj(b)$ (resp., $\Gjlow(\Tj_b)$) is the sum of $\ftil_i^{(\vep_i(b))} \Gj(b')$ (resp., $\ftil_i^{(\vep_i(b))} \Gjlow(\Tj_{b'})$) and a $q\Q[q]$-linear combination of $\Gj(b'')$ (resp., $\Gjlow(\Tj_{b''})$) with $\wtj(b'') = \wtj(b)$ and $\vep_i(b'') > \vep_i(b)$. By our induction hypothesis, $\Gj(b) - \xi_t(\Gjlow(b))$ is a $q\Q[q]$-linear combination of $\Gj(b'')$'s, and is $\psij_M$-invariant. Such a vector must be zero, and hence, we obtain $\Gj(b) = \xi_t(\Gjlow(b))$.
\end{proof}
\begin{lem}\label{lemma 3}
Suppose there exists $j \in \Ij \setminus \{ 1 \}$ such that $\etil_{j'} b \neq 0$ and $\etil_i b = 0$ for all $i \in \Ij$. Then, the statement holds.
\end{lem}
\begin{proof}
Apply Lemma \ref{strong lemma}.
\end{proof}
Now, one can complete the proof by combining Lemma \ref{lemma 1}-\ref{lemma 3} since each $b \in \clB$ with $I^\jmath(b) = \bflm$ is connected to $b_t$ for some $t = 1,\ldots,m_\bflm$.
\section{Basic properties of global crystal bases}\label{basic properties}
\subsection{Global crystal bases}
In this subsection, we exposite some basic properties concerning global crystal bases of $\U$-modules in $\Oint$. Let $M \in \Oint$, $(\clL,\clB)$ a crystal basis of $M$, $\psi_M$ a $\psi$-involution, and $M_\A$ a $\UA$-submodule of $M$. Suppose that $M$ has a global basis $G(\clB)$ with the associated balanced triple $(\clL,M_\A,\psi_M(\clL))$.
\begin{prop}[{\cite{K93}}]\label{fund prop for global crystal basis}
Let $i \in \bbI$, $b \in B$ and $m \in \Z_{\geq 0}$. Then, we have the following.
\begin{enumerate}
\item $\sum_{n \geq m} F_i^{(n)}M_{\A} = \bigoplus_{\substack{b' \in \clB \\ \vep_i(b') \geq m}} \A \Gj(b')$.
\item $\sum_{n \geq m} E_i^{(n)}M_{\A} = \bigoplus_{\substack{b' \in \clB \\ \vphi_i(b') \geq m}} \A \Gj(b')$.
\item $F_i \Gj(b) = [\vep_i(b)+1]\Gj(\Ftil_i b) + \sum_{\substack{b' \in \clB \\ \vep_i(b') > \vep_i(b)+1}} \vphi^{(i)}_{b',b} \Gj(b')$ for some $\vphi^{(i)}_{b',b} \in q^{2-\vep_i(b')}\Q[q]$.
\item $E_i \Gj(b) = [\vphi_i(b)+1]\Gj(\Etil_i b) + \sum_{\substack{b' \in \clB \\ \vphi_i(b') > \vphi_i(b)+1}} \vep^{(i)}_{b',b} \Gj(b')$ for some $\vep^{(i)}_{b',b} \in q^{2-\vphi_i(b')}\Q[q]$.
\end{enumerate}
\end{prop}
For $\lm \in P(M)$, set $I_\lm(M)$ to be the sum of submodules of $M$ isomorphic to $L(\lm)$. Also, we set
\begin{align}
\begin{split}
&W_{\succeq \lm}(M) := \sum_{\mu \succeq \lm} I_\mu(M), \\
&W_{\succ \lm}(M) := \sum_{\mu \succ \lm} I_\mu(M), \\
&W_\lm(M) := W_{\succeq \lm}(M)/W_{\succ \lm}(M).
\end{split} \nonumber
\end{align}
\begin{theo}[{\cite{K93}, \cite{L94}}]
Let $M,\clL,\clB,M_\A$ be as above. Then, for each $\lm \in P(M)$, the following hold:
\begin{enumerate}
\item $W_{\succeq \lm}(M)$ has a global crystal basis $W_{\succeq \lm}(G(\clB)) := \{ G(b) \mid I(b) \succeq \lm \}$ with the associated balanced triple $(W_{\succeq \lm}(\clL),W_{\succeq \lm}(M_{\A}),W_{\succeq \lm}(\psi_M(\clL)))$, where $W_{\succeq \lm}(\clL) := W_{\succeq \lm}(M) \cap \clL$, and so on.
\item $W_{\succ \lm}(M)$ has a global crystal basis $W_{\succ \lm}(G(\clB)) := \{ G(b) \mid I(b) \succ \lm \}$ with the associated balanced triple $(W_{\succ \lm}(\clL),W_{\succ \lm}(M_{\A}),W_{\succ \lm}(\psi_M(\clL)))$, where $W_{\succ \lm}(\clL) := W_{\succ \lm}(M) \cap \clL$, and so on.
\item $W_{\lm}(M)$ has a global crystal basis $W_{\lm}(G(\clB)) := \{ G(b) + W_{\succ \lm}(M) \mid I(b) = \lm \}$ with the associated balanced triple $(W_{\lm}(\clL),W_{\lm}(M_{\A}),W_{\lm}(\psi_M(\clL)))$, where $W_{\lm}(\clL) := W_{\succeq \lm}(\clL)/W_{\succ \lm}(\clL)$, and so on.
\item There exists a $\U$-module isomorphism $\xi: L(\lm)^{\oplus m_\lm} \rightarrow W_\lm(M)$ which induces an isomorphism
$$
(\clL(\lm)^{\oplus m_\lm}, (L(\lm)_{\A})^{\oplus m_\lm}, \psi_\lm(\clL(\lm))^{\oplus m_\lm}) \simeq (W_{\lm}(\clL),W_{\lm}(M_{\A}),W_{\lm}(\psi_M(\clL))),
$$
where $m_\lm := \dim \Hom_{\U}(L(\lm),M)$ denotes the multiplicity of $L(\lm)$ in $M$.
\end{enumerate}
\end{theo}
\begin{rem}\label{filtration for frp}\normalfont
By replacing $P(M)$ with $\Pj(M)$ and $\preceq$ with $\trilefteq$, the same result holds for integrable modules over $\bfU(\frl)$ with global crystal bases.
\end{rem}
\subsection{$\jmath$-canonical bases}
Let $M \in \Oint$ be a based $\U$-module with a crystal basis $(\clL,\clB)$, a global crystal basis $G(\clB)$, a $\psi$-involution $\psi_M$, and a balanced triple $(\clL,M_{\A},\psi_M(\clL))$. Set $\psij_M := \Upsilon \circ \psi_M$. We denote by $\Gj(\clB)$ the associated $\jmath$-canonical basis. Recall that $\psij_M$ is a $\psij$-involution on $M$, and $(\clL,M_\A,\psij_M(\clL))$ is a balanced triple with the associated global basis $\Gj(\clB)$.
\begin{lem}\label{expansion of jCB in CB}
Let $b \in \clB$. Let us write as
$$
\Gj(b) = G(b) + \sum_{\substack{b' \in \clB \\ \wtj(b') = \wtj(b) \AND \wt(b') < \wt(b)}} c_{b',b} G(b')
$$
for some $c_{b',b} \in q\Ao \cap \A$. Then, we have $c_{b',b} = 0$ unless
\begin{align}\label{jcanonical condition}
I^\jmath(b) \trilefteq I^\jmath(b') \OR |I^\jmath(b')^-| < |I^\jmath(b)^-|.
\end{align}
\end{lem}
\begin{proof}
By the construction of $\Gj(b)$, it suffices to show that $\psij_M(G(b))$ is a linear combination of $G(b')$ with $b'$ satisfying \eqref{jcanonical condition}. Since $\psij_M(G(b)) = \Upsilon G(b) \in \bfU^- G(b)$, it suffices to show that for each $l \in \Z_{\geq 0}$ and $i_1,\ldots,i_l \in \bbI$, we have
$$
F_{i_l} \cdots F_{i_1} G(b) \in \Span_{\Q(p,q)} \{ G(b') \mid \text{$b'$ satisfies condition \eqref{jcanonical condition}} \}.
$$
We prove it by induction on $l$. When $l = 0$, there are nothing to prove. So, assume that $l > 0$ and that $F_{i_{l-1}} \cdots F_{i_1} G(b) \in \Span_{\Q(p,q)} \{ G(b') \mid \text{$b'$ satisfies condition \eqref{jcanonical condition}} \}$ for all $i_1,\ldots,i_{l-1} \in \bbI$. If $i_l \neq \ul{1}$, then, by Remark \ref{filtration for frp}, we have
$$
F_{i_l} G(b') \in \Span_{\Q(p,q)} \{ G(b'') \mid I^\jmath(b') \trilefteq I^\jmath(b'') \}
$$
for all $b'$ satisfying condition \eqref{jcanonical condition}. Since $|I^\jmath(b'')^-| = |I^\jmath(b')^-|$ for all $b''$ with $I^\jmath(b') \trilefteq I^\jmath(b'')$, $b''$ satisfies condition \eqref{jcanonical condition}.
If $i_l = \ul{1}$, then $\wt(F_{i_l} G(b')) = \wt(G(b')) - \alpha_{\ul{1}}$. This immediately implies that $F_{i_l} G(b') \in \Span_{\Q(p,q)} \{ G(b'') \mid |I^\jmath(b'')^-| < |I^\jmath(b')^-| \}$. Therefore, $F_{i_l} \cdots F_{i_1} G(b)$ is a linear combination of $G(b')$ with $|I^\jmath(b')^-| < |I^\jmath(b)^-|$. Thus, the proof completes.
\end{proof}
\begin{prop}\label{str const for Gj-basis}
Let $b \in \clB$ and $i \in \Ij \setminus \{1\}$. Then, we have
\begin{align}
\begin{split}
e_i \Gj(b) &= [\vphi_{\ul{i}}(b) + 1] \Gj(\Etil_{\ul{i}}b) + \sum_{\substack{b' \in \clB \setminus \{ \Etil_{\ul{i}}b \} \\ \wtj(b') = \wtj(b) + \gamma_i \AND \wt(b') \leq \wt(b) + \alpha_{\ul{i}}}} e_{b',b}^{(i)} \Gj(b'), \\
f_i \Gj(b) &= [\vphi_{-\ul{i}}(b) + 1] \Gj(\Etil_{-\ul{i}}b) + \sum_{\substack{b' \in \clB \setminus \{ \Etil_{-\ul{i}}b \} \\ \wtj(b') = \wtj(b) - \gamma_i \AND \wt(b') \leq \wt(b) + \alpha_{-\ul{i}}}} f_{b',b}^{(i)} \Gj(b')
\end{split} \nonumber
\end{align}
for some $e_{b',b}^{(i)}, f_{b',b}^{(i)} \in \A$. Moreover, $e_{b',b}^{(i)} = f_{b',b}^{(i)} = 0$ unless $I^\jmath(b) \trilefteq I^\jmath(b')$ or $|I^\jmath(b')^-| < |I^\jmath(b)^-|$.
\end{prop}
\begin{proof}
We prove the assertion only for $e_i$; the proof for $f_i$ is similar. By Lemma \ref{expansion of jCB in CB}, we can write
$$
\Gj(b) = G(b) + \sum_{b' \in \clB \setminus \{ b \}} c_{b',b} G(b')
$$
for some $c_{b',b} \in \A$ such that $c_{b',b} = 0$ unless $I^\jmath(b) \trilefteq I^\jmath(b')$ or $|I^\jmath(b')^-| < |I^\jmath(b)^-|$. Since $e_i \in U_q(\frl)$, it holds that
$$
e_i \Gj(b) \in \Span_\A \{ G(b'') \mid I^\jmath(b) \trilefteq I^\jmath(b') \OR |I^\jmath(b')^-| < |I^\jmath(b)^-| \}.
$$
Hence, it suffices to show that $[e_i \Gj(b) : \Gj(\Etil_{\ul{i}}b)] = [\vphi_{\ul{i}}(b) + 1]$. By the definitions of $e_i$ and $\Gj(b)$, $e_i \Gj(b)$ is the sum of $E_{\ul{i}} G(b)$ and a linear combination of weight vectors of $M$ of weight lower than $\wt(b) + \alpha_{\ul{i}}$. We know from Proposition \ref{fund prop for global crystal basis} $(4)$ that $[E_{\ul{i}} \Gj(b) : G(\Etil_{\ul{i}}b)] = [\vphi_{\ul{i}}(b)+1]$. Hence, we have $[e_i \Gj(b) : \Gj(\Etil_{\ul{i}}b)] = [\vphi_{\ul{i}}(b) + 1]$. This proves the assertion.
\end{proof}
\subsection{Global $\jmath$-crystal bases}\label{subsec: global basis}
Let $M \in \Oj$, $(\clL,\clB)$ a $\jmath$-crystal basis of $M$, $\psij_M$ a $\psij$-involution, and $M_{\A}$ a $\UjA$-submodule of $M$. Suppose that $M$ has a $\jmath$-global basis $\Gj(\clB)$ with the associated balanced triple $(\clL,M_{\A},\psij_M(\clL))$.
Following \cite{K02}, let us introduce modified Kashiwara operators:
\begin{defi}
For $n \in \Z$, set
\begin{align}
\begin{split}
\ftil_i^{(n)} &:= \sum_{t \geq 0,-n} f_i^{(n+t)} e_i^{(t)} A_n(t;k_i), \\
\ftil_1^{(n)} &:= \sum_{t \geq 0,-n} f_1^{(n+t)} e_1^{(t)} a_n(t;k_1),
\end{split} \nonumber
\end{align}
where
\begin{align}
\begin{split}
A_n(t;x) &:= (-1)^t q^{t(1-n)} x^t \prod_{s=0}^{t-1}(1-q^{n+2s}), \\
a_n(t;x) &:= (-1)^t p^t q^{t(1-n)} x^t \prod_{s=0}^{t-1} q^s(1-q^{n+2s}).
\end{split} \nonumber
\end{align}
\end{defi}
\begin{lem}\label{modified ftil preserves lattice}
Let $M \in \Oj$ with the $\jmath$-crystal basis $(\clL,\clB)$. For $n \in \Z$, we have $\ftil_i^{(n)} \clL \subset \clL$, and $\ftil_i^{(n)} \clL = \ftil_i^n \clL$ modulo $q\clL$.
\end{lem}
\begin{proof}
If $i \neq 1$, then the statement follows from \cite[Proposition 6.1]{K02}. Hence, we prove the case when $i=1$. It suffices to prove the following: For each $u \in \clL$ such that $e_1u = 0$, $k_1u = q^au$, $e_1f_1u = [b]\{ a-b-1 \}u$ with $a \in \Z$ and $b \in \Z_{\geq 0}$, we have $\ftil_1^{(n)} f_1^{(m)} u = cf_1^{(m+n)}$ for some $c \in 1 + q\A_0 \cap \A$. First of all, we have
\begin{align}
\ftil_1^{(n)} f_1^{(m)} u = \sum_{t \geq 0,-n} a_n(t;q^{a-3m}) {m+n \brack m-t} {b-m+t \brack t} \prod_{s=0}^{t-1} \{ a-b-m+s \} f_1^{(m+n)} u. \nonumber
\end{align}
We compute the coefficient, say $A$, of the right-hand side as follows.
\begin{align}
\begin{split}
A &= \sum_{t \geq 0,-n} A_n(t;q^{b-2m}) {m+n \brack m-t} {b-m+t \brack t} \prod_{s=0}^{t-1}(1 + p^2q^{2(a-b-m+s)}) \\
&= \sum_{t \geq 0,-n} B_t + p^2\sum_{t \geq 0,-n} B_t g_t,
\end{split} \nonumber
\end{align}
where $B_t := A_n(t;q^{b-2m}) {m+n \brack m-t} {b-m+t \brack t}$, and $g_t \in \Z[p,q,q\inv]$ with $\prod_{s=0}^{t-1}(1 + p^2q^{2(a-b-m+s)}) = 1 + p^2g_t$. By the proof of \cite[Proposition 6.1]{K02}, we have $B_t \in 1 + q\Z[q]$. Also, it is clear that $p^2\sum_{t \geq 0,-n} B_t g_t \in p^2\Z[p,q,q\inv]$. Thus, we have $\ftil_1^{(n)} f_1^{(m)} u \in \clL$ and $\ftil_1^{(n)} f_1^{(m)} u = f_1^{(m+n)} u = \ftil_1^n f_1^{(m)} u$ modulo $q\clL$. This proves the lemma.
\end{proof}
\begin{prop}
Let $i \in \Ij$, $b \in \clB$ and $m \in \Z_{\geq 0}$. Then, we have the following.
\begin{enumerate}
\item $\sum_{n \geq m} f_i^{(n)}M_{\A} = \bigoplus_{\substack{b' \in \clB \\ \vep_i(b') \geq m}} \A \Gj(b')$.
\item $\sum_{n \geq m} e_i^{(n)}M_{\A} = \bigoplus_{\substack{b' \in \clB \\ \vphi_i(b') \geq m}} \A \Gj(b')$ if $i \neq 1$.
\item $f_i \Gj(b) = [\vep_i(b)+1]\Gj(\ftil_i b) + \sum_{\substack{b' \in \clB \\ \vep_i(b') > \vep_i(b)+1}} \vphi^{(i)}_{b',b} \Gj(b')$ for some $\vphi^{(i)}_{b',b} \in q^{2-\vep_i(b')}\Q[q]$.
\item $e_i \Gj(b) = [\vphi_i(b)+1]\Gj(\etil_i b) + \sum_{\substack{b' \in \clB \\ \vphi_i(b') > \vphi_i(b)+1}} \vep^{(i)}_{b',b} \Gj(b')$ for some $\vep^{(i)}_{b',b} \in q^{2-\vphi_i(b')}\Q[q]$ if $i \neq 1$.
\end{enumerate}
\end{prop}
\begin{proof}
Since $(e_i,k_i,f_i)$, $i \neq 1$ forms an $\frsl_2$-triple, most of the assertions follows from Proposition \ref{fund prop for global crystal basis}. What we have to prove are assertions $(1)$ and $(3)$ for $i = 1$. First, we prove part $(1)$ by induction on $m$. When $m = 0$, the both sides of the equation to be proved are $0$. Assume that assertion $(1)$ holds for all $m' > m$. Let $b' \in \clB$ be such that $\vep_1(b') = m$. Set $b'_0 := \etil_1^mb$, and consider $u := \ftil_1^{(m)} \Gj(b'_0)$. By the definition of $\ftil_1^{(m)}$ and Lemma \ref{modified ftil preserves lattice}, we have
$$
u - f_1^{(m)} \Gj(b'_0) \in \sum_{n > m} f_1^{(n)} M_\A \AND u + q\clL = b'.
$$
By our inductive hypothesis, we can write
$$
u - f_1^{(m)}\Gj(b'_0) = \sum_{\substack{b'' \in \clB \\ \vep_1(b'') > m}} a_{b''} \Gj(b'')
$$
for some $a_{b''} \in \A$. Then, we can take $a'_{b''} \in q\Q[q]$ in a way such that $a_{b''} - \ol{a_{b''}} = a'_{b''} - \ol{a'_{b'''}}$. Set $v := u - \sum_{b''} a'_{b''} \Gj(b'') = f_1^{(m)}\Gj(b'_0) + \sum_{b''} (a_{b'''} - a'_{b''}) \Gj(b'')$. Then, we have $v \in M_\A \cap \clL$, $\psij_M(v) = v$, and $v + q\clL = u + q\clL = b'$. These implies that $v = \Gj(b')$, and therefore, $\Gj(b') \in \sum_{n \geq m} f_1^{(n)} M_\A$. Hence, we obtain $\sum_{n \geq m} f_i^{(n)}M_{\A} \supset \bigoplus_{\substack{b' \in \clB \\ \vep_i(b') \geq m}} \A \Gj(b')$.
We prove the opposite inclusion. For each $\lm \in \Lmj$, we have
\begin{align}
\begin{split}
(M_\A)_\lm &\subset \sum_{b \in \clB_\lm} \A \Gj(b) \\
&= \sum_{\substack{b \in \clB_\lm \\ \vep_1(b) = 0}} \A \Gj(b) + \sum_{\substack{b' \in \clB_\lm \\ \vep_1(b') \geq 1}} \A \Gj(b') \\
&\subset \sum_{\substack{b \in \clB_\lm \\ \vep_1(b) = 0}} \A \Gj(b) + \sum_{n \geq 1} f_1^{(n)} (M_\A)_{\lm + n\gamma_1}.
\end{split} \nonumber
\end{align}
Hence, we obtain
\begin{align}
\begin{split}
f_1^{(m)} (M_\A)_\lm & \subset \sum_{\substack{b \in \clB_\lm \\ \vep_1(b) = 0}} \A f_1^{(m)}\Gj(b) + \sum_{n \geq 1} f_1^{(m)}f_1^{(n)} (M_\A)_{\lm + n\gamma_1} \\
&\subset \sum_{\substack{b \in \clB_\lm \\ \vep_1(b) = 0}} \A f_1^{(m)}\Gj(b) + \sum_{n \geq 1} f_1^{(m+n)} (M_\A)_{\lm + n\gamma_1} \\
&= \sum_{\substack{b \in \clB_\lm \\ \vep_1(b) = 0}} \A f_1^{(m)}\Gj(b) + \sum_{\substack{b' \in \clB_\lm \\ \vep_1(b') > m}} \A \Gj(b') \qu (\text{by induction hypothesis}).
\end{split} \nonumber
\end{align}
Also, by the argument above, $f_1^{(m)} \Gj(b)$ with $\vep_1(b) = 0$ is contained in $\sum_{\vep(b') \geq m} \A \Gj(b')$. This completes the proof of part $(1)$.
Next, we turn to prove assertion $(3)$ for $i = 1$ by descending induction on $m := \vep_1(b)$. When $m$ is maximum among $\{ \vep_1(b') \mid b' \in \clB \}$, we have by $(1)$ that
$$
f_1 \Gj(b) \in \sum_{n > m} f_1^{(n)} M_\A = \sum_{\vep_1(b') > m} \A \Gj(b') = 0,
$$
and the equation in $(3)$ holds. Assume that $(3)$ is true for all $m' > m$. As in the proof of $(1)$, let us write
\begin{align}
\begin{split}
\Gj(b) &= f_1^{(m)} \Gj(\etil_1^m b) + \sum_{\substack{b' \in \clB \\ \vep_1(b') > m}} c_{b'} \Gj(b'), \\
\Gj(\ftil_1b) &= f_1^{(m+1)} \Gj(\etil_1^m b) + \sum_{\substack{b'' \in \clB \\ \vep_1(b'') > m+1}} d_{b''} \Gj(b'')
\end{split} \nonumber
\end{align}
for some $c_{b'},d_{b''} \in \A$. Then, we have
\begin{align}
\begin{split}
f_1 \Gj(b) &= [m+1] f_1^{(m+1)} \Gj(\etil_1^m b) + \sum_{\vep_1(b') > m} c_{b'} f_1 \Gj(b') \\
&= [m+1] f_1^{(m+1)} \Gj(\etil_1^m b) + \sum_{\vep_1(b') > m} c_{b'} ([\vep_1(b')+1] \Gj(\ftil_1b') + \sum_{\vep_1(b'') > \vep_1(b')+1} \vphi_{b'',b'}^{(1)} \Gj(b'')) \\
&= [m+1] \Gj(\ftil_1b) + \sum_{\vep_1(b') > m} c_{b'} ([\vep_1(b')+1] \Gj(\ftil_1b') + \sum_{\vep_1(b'') > \vep_1(b')+1} \vphi_{b'',b'}^{(1)} \Gj(b'')) \\
&\qu \qu - \sum_{\substack{b'' \in \clB \\ \vep_1(b'') > m+1}} [m+1] d_{b''} \Gj(b'').
\end{split} \nonumber
\end{align}
Thus, we obtain that $f_i \Gj(b) = [\vep_i(b)+1]\Gj(\ftil_i b) + \sum_{\substack{b' \in \clB \\ \vep_i(b') > \vep_i(b)+1}} \vphi^{(i)}_{b',b} \Gj(b')$ for some $\vphi^{(i)}_{b',b} \in \A$. It remains to prove that $\vphi^{(i)}_{b',b} \in q^{2-\vep_1(b')} \Q[q]$. Let us write
$$
\Gj(b) = \sum_{k \geq m} f_1^{(k)} u_k
$$
for some $u_k \in \clL_{\wtj(b) + k\gamma_1}$ such that $e_1u_k = 0$. Note that $\Gj(b) + q\clL = u_m + q\clL$. Then, we have
$$
f_1 \Gj(b) = [m+1] f_1^{(m+1)}u_m + \sum_{k > m} [k+1] f_1^{(k+1)} u_k,
$$
and that $f_1^{(m+1)} u_m \in \clL$, $f_1^{(m+1)}u_m + q\clL = \ftil_1 b$. Hence, we have $f_1 \Gj(b) = [m+1] \Gj(\ftil_1b) + \sum_{k > m} [k+1] f_1^{(k+1)} u_k$ modulo $q^{2-m}\clL$. Then, rewriting $f_1^{(k+1)}u_k$ as a sum of $\Gj(b')$, $\vep_1(b') \leq k+1$ with coefficients in $q\Ao$, we conclude that the coefficient of $\Gj(b')$ in $f_1 \Gj(b)$ lies in $q^{2-\vep_1(b')}\Ao \cap \A = q^{2-\vep_1(b')} \Q[q]$. This completes the proof.
\end{proof}
For a bipartition $\bflm \in \Pj(M)$, define $I_\bflm(M)$, $W_{\succeq \bflm}(M)$, $W_{\succ \bflm}(M)$, and $W_\bflm(M)$ in a similar way as $I_\lm$, $W_{\succeq \lm}$, $W_{\succ \lm}$, and $W_\lm$, respectively.
\begin{defi}\normalfont
We say that $M$ has the property $(\ast)$ if there exists a poset $(S,\leq)$ and a map $s: \clB \rightarrow S$ satisfying the following:
\begin{enumerate}
\item The abelian group $Q := \sum_{i \in \bbI} \Z \alpha_i$ acts on $S$ freely; the action is written additively.
\item $\sigma \leq \sigma + \lm$ for all $\lm \in Q_+$, $\sigma \in S$.
\item $\sigma + \lm \leq \sigma' + \lm$ for all $\lm \in Q$, $\sigma \leq \sigma' \in S$.
\item $s(b) = s(b')$ only if $\wt(b) = \wt(b')$ for all $b,b' \in \clB$.
\item For $b \in \clB$ and $i \in \Ij \setminus \{1\}$, $s(\Etil_{\ul{i}} b) = s(b) + \alpha_{\ul{i}}$ if $\Etil_{\ul{i}}b \neq 0$.
\item For $i \in \Ij \setminus \{1\}$,
\begin{align}
\begin{split}
e_i \Gj(b) &= [\vphi_{\ul{i}}(b) + 1] \Gj(\Etil_{\ul{i}}b) + \sum_{\substack{b' \in \clB \setminus \{ \Etil_{\ul{i}}b \} \\ \wtj(b') = \wtj(b) + \gamma_i \AND s(b') \leq s(b) + \alpha_{\ul{i}}}} e_{b',b}^{(i)} \Gj(b'), \\
f_i \Gj(b) &= [\vphi_{-\ul{i}}(b) + 1] \Gj(\Etil_{-\ul{i}}b) + \sum_{\substack{b' \in \clB \setminus \{ \Etil_{-\ul{i}}b \} \\ \wtj(b') = \wtj(b) - \gamma_i \AND s(b') \leq s(b) + \alpha_{-\ul{i}}}} f_{b',b}^{(i)} \Gj(b')
\end{split} \nonumber
\end{align}
for some $e_{b',b}^{(i)}, f_{b',b}^{(i)} \in \A$.
\end{enumerate}
\end{defi}
\begin{lem}\label{condition ast}
Let $M \in \Oj$, and $\clL,\clB,\psij_M,M_\A$ as above.
\begin{enumerate}
\item If $r = 1$, then $M$ has the property $(\ast)$.
\item If $M \in \Oint$ and the global $\jmath$-crystal basis is the $\jmath$-canonical basis, then $M$ has the property $(\ast)$.
\end{enumerate}
\end{lem}
\begin{proof}
Setting $S$ and $s$ to be $\Lm$ and $\wt$, respectively, part $(1)$ is obvious, and part $(2)$ follows from Proposition \ref{str const for Gj-basis}.
\end{proof}
The main result in this paper is the following:
\begin{theo}\label{filtration}
Suppose that $M$ has the property $(\ast)$. Then, for each $\bflm \in \Pj(M)$, the following hold:
\begin{enumerate}
\item $W_{\succeq \bflm}(M)$ has a global $\jmath$-crystal basis $W_{\succeq \bflm}(\Gj(\clB)) := \{ \Gj(b) \mid I(b) \succeq \bflm \}$ with the associated balanced triple $(W_{\succeq \bflm}(\clL),W_{\succeq \bflm}(M_{\A}),W_{\succeq \bflm}(\psij_M(\clL)))$, where $W_{\succeq \bflm}(\clL) := W_{\succeq \bflm}(M) \cap \clL$, and so on.
\item $W_{\succ \bflm}(M)$ has a global $\jmath$-crystal basis $W_{\succ \bflm}(\Gj(\clB)) := \{ \Gj(b) \mid I(b) \succ \bflm \}$ with the associated balanced triple $(W_{\succ \bflm}(\clL),W_{\succ \bflm}(M_{\A}),W_{\succ \bflm}(\psij_M(\clL)))$, where $W_{\succ \bflm}(\clL) := W_{\succ \bflm}(M) \cap \clL$, and so on.
\item $W_{\bflm}(M)$ has a global $\jmath$-crystal basis $W_{\bflm}(\Gj(\clB)) := \{ \Gj(b) + W_{\succ \bflm}(M) \mid I(b) = \bflm \}$ with the associated balanced triple $(W_{\bflm}(\clL),W_{\bflm}(M_{\A}),W_{\bflm}(\psij_M(\clL)))$, where $W_{\bflm}(\clL) := W_{\succeq \bflm}(\clL)/W_{\succ \bflm}(\clL)$, and so on.
\item There exists a $\Uj$-module isomorphism $\xi: L(\bflm)^{\oplus m_{\bflm}} \rightarrow W_\bflm(M)$ which induces an isomorphism
$$
(\clL(\bflm)^{\oplus m_\bflm}, (L(\bflm)_{\A})^{\oplus m_\bflm}, \psij_{\bflm}(\clL(\bflm))^{\oplus m_\bflm}) \simeq (W_{\bflm}(\clL),W_{\bflm}(M_{\A}),W_{\bflm}(\psij_M(\clL))),
$$
where $m_\bflm := \dim \Hom_{\Uj}(L(\bflm),M)$ denotes the multiplicity of $L(\bflm)$ in $M$.
\end{enumerate}
\end{theo}
The proof will be given in Section \ref{sec:proof of main theorem}.
\begin{cor}
Let $\bflm \in \Pj$. Then, $\Gjlow(\bflm)$ is a unique global $\jmath$-crystal basis of $L(\bflm)$ satisfying the property $(*)$.
\end{cor}
\subsection{Operators $\tilde{e}_{i^+}$ and $\tilde{f}_{i^+}$}
The definitions of $\etil_{i'}$ and $\ftil_{i'}$ given in \cite{W17} are artificial, namely, they are defined by means of a distinguished basis $\Gjlow(\bflm)$, $\bflm \in \Pj$ (in \cite{W17}, it is denoted by $\{ b_T \mid T \in \clB(\bflm) \}$). Here, we define new operators $\etil_{i^+}$ and $\ftil_{i^+}$ for $i \in \Ij \setminus \{1\}$, and then, explain that the operators $\etil_{i'}$ and $\ftil_{i'}$ on $\jmath$-crystal bases are in fact intrinsic.
\begin{lem}
Let $r \geq 2$, $\bflm \in \Pj$, and consider the irreducible highest weight module $L(\bflm)$. As a $\Uj_{r-1}$-module, $L(\bflm)$ is multiplicity-free.
\end{lem}
\begin{proof}
Let $b \in \clB(\bflm)$ be a $\Uj_{r-1}$-highest weight vector with highest weight, say, $\bfmu \in \Pj_{r-1}$. If we identify $\clB(\bflm)$ with $\SST(\bflm)$, we have $T^\jmath_b \downarrow_{r-1} = T_{\bfmu}$. Since the entries of the boxes of $T^\jmath_{b'}$ corresponding to $\bflm/\bfmu$ are either $-r$ or $r$, it must hold that $\bflm/\bfmu$ is a horizontal strip. Conversely, given $\bfmu \in \Pj_{r-1}$ such that $\bflm/\bfmu$ is a horizontal strip, there exists a unique $b \in \clB(\bflm)$ which is a $\Uj_{r-1}$-highest weight vector with highest weight $\bfmu$. This proves the lemma.
\end{proof}
\begin{lem}\label{e_rprime}
Let $r \geq 2$, $\bflm \in \Pj$. Let $b \in \clB(\bflm)$ be such that $\etil_{r'} b \neq 0$. Then, there exist unique $b' \in \clB(\bflm)$ and $j \in \Ij \setminus \{1\}$ satisfying the following:
\begin{itemize}
\item $b'$ is a $\Uj_{r-1}$-highest weight vector.
\item There exist unique $\vep_i \in \{ \emptyset, \prime \}$ for each $j \leq i \leq r-1$ such that $b = \ftil_{r'} \ftil_{(r-1)^{\vep_{r-1}}} \cdots \ftil_{j^{\vep_j}} b'$.
\end{itemize}
\end{lem}
\begin{proof}
By the definition of $\etil_{r'}$, $b$ is a $\Uj_{r-1}$-highest weight vector with highest weight, say, $\bfmu \in \Pj_{r-1}$ such that $(T^\jmath_b)^- = T_\bflm^-$. Then, $T^\jmath_{\etil_{r'} b} \downarrow_{r-1}$ is obtained from $T_\bfmu$ by adding a box $\fbox{$r-1$}$ to the $(j-1)$-th row for some uniquely determined $j \in \Ij \setminus \{1\}$. Set $b_{r-1} := \etil_{r'} b$. Now, we have exactly one of the following; $\etil_{r-1} b_{r-1} \neq 0$ or $\etil_{(r-1)'} b_{r-1} \neq 0$. Choose a unique $\vep_{r-1} \in \{ \emptyset,\prime \}$ in a way such that $b_{r-2} := \etil_{(r-1)^{\vep_{r-1}}} b_{r-1} \neq 0$. Then, $T^\jmath_{b_{r-2}} \downarrow_{r-1}$ is obtained from $T_\bfmu$ by adding a box $\fbox{$r-2$}$ to the $(j-1)$-th row. Repeating this procedure, we obtain $\vep_{i} \in \{ \emptyset,\prime \}$ and $b_{i-1} \in \clB(\bflm)$ for $j \leq i \leq r-1$. By the construction, $T^\jmath_{b_{j-1}} \downarrow_{r-1}$ is obtained from $T_\bfmu$ by adding a box $\fbox{$j-1$}$ to the $(j-1)$-th row, which turned out to be $T_{\bfmu'}$, where $\bfmu' \in \Pj_{r-1}$ such that $\bfmu'_k = \bfmu_k + \delta_{k,j-1}$, $k \in \{ -(r-1),\ldots,r-1 \}$. Hence, $b_{j-1}$ is a $\Uj_{r-1}$-highest weight vector, and we have $b = \ftil_{r'} \ftil_{(r-1)^{\vep_{r-1}}} \cdots \ftil_{j^{\vep_j}} b_{j-1}$. This proves the assertion.
\end{proof}
Set $E_r(\bflm) := \{ \bfmu \in \Pj_{r-1} \mid \bfmu^- = \bflm^- \downarrow_{r-1} \AND \bflm^+/\bfmu^+ \text{ is a horizontal strip} \}$. Then, the assignment
$$
\{ b \in \clB(\bflm) \mid \etil_{r'}b \neq 0 \} \rightarrow E_r(\bflm);\ b \mapsto I^\jmath_{r-1}(b)
$$
is bijective. To each $\bfmu \in E_r(\bflm)$, we associate $b,b' \in \clB(\bflm)$, $j \in \Ij \setminus \{ 1 \}$, and $\vep_i \in \{ \emptyset,\prime \}$, $j \leq i \leq r-1$ as in Lemma \ref{e_rprime}.
Let $r \geq 2$. We define operators $\etil_{l^+}$ and $\ftil_{l^+}$ on every $\Uj$-modules in $\Oj$ inductively for all $2 \leq l < r$. Let $\bflm \in \Pj$. We define the linear operator $\etil_{r^+}$ on $L(\bflm)$ by
$$
\etil_{r^+} := \bigoplus_{\bfmu \in E_r(\bflm)}p_2(\bfmu) \circ \frac{1}{[\vphi_{\ul{r}}(b_\bfmu)+1]}e_r \circ p_1(\bfmu),
$$
where $b_\bfmu \in \clB(\bflm)$ is the corresponding element to $\bfmu \in E_r(\bflm)$, $p_1(\bfmu)$ is the projection from $L(\bflm)$ to the one-dimensional subspace $L(\bfmu)_{\wtj(\bfmu)}$;
$$
L(\bfmu)_{\wtj(\bfmu)} \subset L(\bfmu) \underset{\text{multiplicity free}}{\hookrightarrow} L(\bflm),
$$
and $p_2(\bfmu)$ is the projection from $L(\bflm)$ to the one-dimensional subspace $\ftil_{(r-1)^{\delta_{r-1}}} \cdots \ftil_{j^{\delta_j}} L(\bfmu')_{\wtj(\bfmu')}$;
$$
\ftil_{(r-1)^{\delta_{r-1}}} \cdots \ftil_{j^{\delta_j}} L(\bfmu')_{\wtj(\bfmu')} \subset L(\bfmu') \underset{\text{multiplicity free}}{\hookrightarrow} L(\bflm),
$$
where $\delta_l = \emptyset$ if $\vep_l = \emptyset$, and $\delta_l = +$ if $\vep_l = \prime$ for $l = j,\ldots,r-1$. Also, we define $\ftil_{r^+}$ by
$$
\ftil_{r^+} = \bigoplus_{\bfmu \in E_r(\bflm)} \etil_{r^+}\inv \circ p_2(\bfmu),
$$
where $\etil_{r^+}\inv$ is the inverse of the linear isomorphism $\etil_{r^+} : L(\bfmu)_{\wtj(\bfmu)} \rightarrow \ftil_{(r-1)^{\delta_{r-1}}} \cdots \ftil_{j^{\delta_j}} L(\bfmu')_{\wtj(\bfmu')}$.
Finally, we extend the definitions of $\etil_{r^+}$ and $\ftil_{r^+}$ to a general $\Uj$-module $M \in \Oj$ by the complete reducibility of $M$.
\begin{prop}
Let $\bflm \in \Pj$ and $v \in L(\bflm)$ a highest weight vector. Then, we have
\begin{align}
\begin{split}
\clL(\bflm) &= \Span_{\Ao} \{ \ftil_{i_1} \cdots \ftil_{i,l}v \mid l \in \Z_{\geq 0},\ i_1,\ldots,i_l \in \Ij \sqcup \{ 2^+,\ldots,r^+ \} \}, \\
\clB(\bflm) &= \{ \ftil_{i_1} \cdots \ftil_{i,l}v + q\clL(\bflm) \mid l \in \Z_{\geq 0},\ i_1,\ldots,i_l \in \Ij \sqcup \{ 2^+,\ldots,r^+ \} \} \setminus \{0\}.
\end{split} \nonumber
\end{align}
Moreover, on $\clB(\bflm)$, we have $\etil_{i'} = \etil_{i^+}$ and $\ftil_{i'} = \ftil_{i^+}$ for all $i \in \Ij \setminus \{1\}$.
\end{prop}
\begin{proof}
We proceed by induction on $r$. Assume that the assertion holds for all $2 \leq l < r$ (we assume nothing when $r=2$). Let $\bfmu \in E_r(\bflm)$ and $b_\bfmu$, $b'$, $\bfmu'$ be as above. By the uniqueness of the $\jmath$-crystal bases for $\Uj_{r-1}$-modules, there exists a unique $v_\bfmu \in \clL(\bflm)$ such that $\Uj_{r-1}v_\bfmu = L(\bfmu)$, $v_\bfmu + q\clL(\bflm) = b_\bfmu$. Then, we can write
$$
v_\bfmu = \Gjlow(b_\bfmu) + \sum_{b' \in \clB(\bflm) \setminus \{ b_\bfmu \}} a_{b'} \Gjlow(b')
$$
for some $a_{b'} \in q\Ao$. Note that this equation implies that $\etil_{r'} (v_\bfmu) \in \Gjlow(\etil_{r'} b_\bfmu) + q\clL(\bflm)$. Also, we have
$$
\frac{1}{[\vphi_{\ul{r}}(b_\bfmu)+1]} e_r v_\bfmu = \Gjlow(\etil_{r'}b_\bfmu) + \sum_{b' \in \clB(\bflm)} c_{b'} \Gjlow(b') \qu (\text{since $\etil_{r'} b_\bfmu = \Etil_{\ul{r}} b_\bfmu$.})
$$
for some $c_{b'} \in \A$. Again, by the complete reducibility of the $\Uj_{r-1}$-crystal bases, there exists a unique $v_{\bfmu'} \in \clL(\bflm)$ such that $\Uj_{r-1} v_{\bfmu'} = L(\bfmu')$, $v_{\bfmu'} + q\clL(\bflm) = b'$. By our induction hypothesis, we have $u := \ftil_{(r-1)^{\delta_{r-1}}} \cdots \ftil_{j^{\delta_j}}(v_{\bfmu'}) \in \clL(\bflm) \cap \Uj_{r-1} v_{\bfmu'}$ and $u + q\clL(\bflm) = \etil_{r'}b_\bfmu$. Then, we can write
$$
u = \Gjlow(\etil_{r'}b_\bfmu) + \sum_{b' \in \clB(\bflm)} d_{b'} \Gjlow(b')
$$
for some $d_{b'} \in q\Ao$. Hence, we have
$$
\etil_{r^+}(v_\bfmu) \in \Gjlow(\etil_{r'}b_\bfmu) + q\clL(\bflm).
$$
Since we took $\bfmu \in E_r(\bflm)$ arbitrarily, this equation ensures that $\etil_{r^+}$ preserves $\clL(\bflm)$ and $\clB(\bflm) \sqcup \{0\}$, and that $\etil_{r^+} = \etil_{r'}$ on $\clB(\bflm)$. By the definition of $\ftil_{r^+}$, it also preserves $\clL(\bflm)$ and $\clB(\bflm) \sqcup \{0\}$, and coincides with $\ftil_{r'}$ on $\clB(\bflm)$. Now, the assertions are clear by the definition of $(\clL(\bflm),\clB(\bflm))$.
\end{proof}
\begin{cor}
Let $M \in \Oj$ be a $\Uj$-module with a $\jmath$-crystal basis $(\clL,\clB)$. Then $\etil_{i'} = \etil_{i^+}$ and $\ftil_{i'} = \ftil_{i^+}$ on $\clB$ for all $i \in \Ij \setminus \{1\}$.
\end{cor}
\section{Representation theory of $\Uj$}\label{RT for Uj}
Let $p$ and $q$ be independent indeterminates.
\subsection{Definition of $\Uj$}
Let $\Lm$ be the free $\Z$-module with a free basis $\{ \epsilon_i \mid i \in I \}$, and with a symmetric bilinear form $(\cdot,\cdot)$ defined by $(\epsilon_i,\epsilon_j) = \delta_{i,j}$. For $i \in \bbI$, set
$$
\alpha_i := \epsilon_{\ul{i}}-\epsilon_{\ul{i+1}}, \qu \qu Q := \sum_{i \in \bbI} \Z \alpha_i, \qu \qu Q_+ := \sum_{i \in \bbI} \Z_{\geq 0} \alpha_i.
$$
For $\lm,\mu \in \Lm$, we write $\mu \leq \lm$ if $\lm - \mu \in Q_+$. This defines a partial order on $\Lm$.
The quantum group $\bfU = \bfU_{2r+1} = U_q(\mathfrak{sl}_{2r+1})$ of type $A_{2r}$ is an associative algebra over $\Q(p,q)$ with generators $E_i,F_i,K_i^{\pm 1}$, $i \in \bbI$ subject to the following relations: For $i,j \in \bbI$,
\begin{align}
\begin{split}
&K_i K_i\inv = K_i\inv K_i = 1, \\
&K_i K_j = K_j K_i, \\
&K_i E_j K_i\inv = q^{(\alpha_i,\alpha_j)} E_i, \\
&K_i F_j K_i\inv = q^{-(\alpha_i,\alpha_j)} F_i, \\
&E_i F_j - F_j E_i = \delta_{i,j} \frac{K_i - K_i\inv}{q - q\inv}, \\
&E_i^2 E_j - (q+q\inv)E_iE_jE_i + E_jE_i^2 = 0 \qu \IF |i-j| = 1, \\
&F_i^2 F_j - (q+q\inv)F_iF_jF_i + F_jF_i^2 = 0 \qu \IF |i-j| = 1, \\
&E_i E_j - E_j E_i = 0 \qu \IF |i-j| > 1, \\
&F_i F_j - F_j F_i = 0 \qu \IF|i-j| > 1.
\end{split} \nonumber
\end{align}
In this paper, we use the comultiplication $\Delta$ of $\bfU$ given by
\begin{align}
\begin{split}
&\Delta(K_i^{\pm1}) = K_i^{\pm1} \otimes K_i^{\pm1}, \qu \Delta(E_i) = 1 \otimes E_i + E_i \otimes K_i\inv, \\
&\Delta(F_i) = F_i \otimes 1 + K_i \otimes F_i \qu i \in \bbI.
\end{split} \nonumber
\end{align}
Let $(\bfU,\Uj)$ denote the quantum symmetric pair over $\Q(p,q)$ of type AIII, that is, $\Uj = \Uj_r$ is the subalgebra of $\bfU$ generated by
\begin{align}
&k_i^{\pm1} := (K_{\ul{i}}K_{-\ul{i}})^{\pm1}, \nonumber\\
&e_i := E_{\ul{i}} + p^{-\delta_{i,1}} F_{-\ul{i}} K_{\ul{i}}\inv, \nonumber\\
&f_i := E_{-\ul{i}} + p^{\delta_{i,1}} K_{-\ul{i}}\inv F_{\ul{i}}, \qu i \in \Ij. \nonumber
\end{align}
The $\Uj$ has the following defining relations (\cite{Le99}, see also \cite{BW13}, \cite{BWW16}): For $i,j \in \Ij$,
\begin{align}
\begin{split}
&k_i k_i\inv = k_i\inv k_i = 1, \\
&k_i k_j = k_j k_i, \\
&k_i e_j k_i\inv = q^{(\alpha_{\ul{i}}-\alpha_{-\ul{i}},\alpha_{\ul{j}})} e_j, \\
&k_i f_j k_i\inv = q^{-(\alpha_{\ul{i}}-\alpha_{-\ul{i}},\alpha_{\ul{j}})} f_j, \\
&e_i f_j - f_j e_i = \delta_{i,j} \frac{k_i - k_i\inv}{q - q\inv} \qu \IF (i,j) \neq (1,1), \\
&e_i^2 e_j - (q+q\inv)e_ie_je_i + e_je_i^2 = 0 \qu \IF |i-j| = 1, \\
&f_i^2 f_j - (q+q\inv)f_if_jf_i + f_jf_i^2 = 0 \qu \IF |i-j| = 1, \\
&e_i e_j - e_j e_i = 0 \qu \IF |i-j| > 1, \\
&f_i f_j - f_j f_i = 0 \qu \IF |i-j| > 1, \\
&e_1^2 f_1 - (q+q\inv)e_1f_1e_1 + f_1e_1^2 = -(q+q\inv) e_1 (pqk_1 + p\inv q\inv k_1\inv), \\
&f_1^2 e_1 - (q+q\inv)f_1e_1f_1 + e_1f_1^2 = -(q+q\inv) (pqk_1 + p\inv q\inv k_1\inv) f_1.
\end{split} \nonumber
\end{align}
\begin{prop}\label{automorphisms}
\begin{enumerate}
\item {\cite[Lemma 6.1 (3)]{BW13}} There exists a unique $\Q$-algebra automorphism $\psij$ of $\Uj$ which maps $e_i,f_i,k_i,p,q$ to $e_i,f_i,k_i\inv,p\inv,q\inv$, respectively.
\item There exists a unique $\Q(p,q)$-algebra anti-automorphism $\sigmaj$ of $\Uj$ which maps $e_i,f_i,k_i$ to $f_i,e_i,k_i$, respectively.
\item {\cite[Proposition 4.6]{BW18}} There exists a unique $\Q(p,q)$-algebra anti-automorphism $\tauj$ of $\Uj$ which maps $e_i,f_i,k_i$ to $p^{-\delta_{i,1}}q\inv k_i\inv f_i$, $p^{\delta_{i,1}}q e_i k_i$, $k_i$, respectively.
\end{enumerate}
\end{prop}
\begin{proof}
It suffices to show that the images of the generators of $\Uj$ satisfy the defining relations of $\Uj$; it is straightforward.
\end{proof}
\begin{rem}\normalfont
We have similar automorphisms on $\U$:
\begin{enumerate}
\item There exists a unique $\Q$-algebra automorphism $\psi$ of $\U$ which maps $E_i,F_i,K_i,p,q$ to $E_i,F_i,K_i\inv,p\inv,q\inv$, respectively.
\item There exists a unique $\Q(p,q)$-algebra anti-automorphism $\sigma$ of $\Uj$ which maps $E_i,F_i,K_i$ to $F_i,E_i,K_i$, respectively.
\item There exists a unique $\Q(p,q)$-algebra anti-automorphism $\tau$ of $\Uj$ which maps $E_i,F_i,K_i$ to $q F_i K_i\inv$, $q\inv K_i E_i$, $K_i$, respectively.
\end{enumerate}
Note that $\tauj$ is the restriction of $\tau$ {\cite[Proposition 4.6]{BW18}}, while the others are not.
\end{rem}
Let $\bfU(\frl)$ denote the subalgebra of $\U$ generated by $E_i,F_i,K_j^{\pm1}$, $i \in \bbI \setminus \{ \ul{1} \}$, $j \in \bbI$. Note that we have $e_i,f_i,k_j \in \bfU(\frl)$ for all $i \in \Ij \setminus \{1\}$, $j \in \Ij$. Note that $\bfU(\frl)$ is the quantum group of type $A_r \times A_{r-1}$ with weight lattice $\Lm$.
\subsection{Category $\Oj$}
Let us extend the bilinear form $(\cdot,\cdot)$ on $\Lm$ to $\Lm_\R := \R \otimes_\Z \Lm$. Set $\beta_i := \alpha_{\ul{i}}-\alpha_{-\ul{i}}$, $i \in \Ij$, and $J := \{ \lm \in \Lm_\R \mid (\beta_i,\lm) = 0 \Forall i \in \Ij \}$. Then, the induced bilinear form $(\cdot, \cdot) :(\sum_{i \in \Ij} \R \beta_i) \times (\Lm_\R/J) \rightarrow \R$ denoted by the same symbol is nondegenerate. Let $\delta_j \in \Lm_\R/J$ be such that $(\beta_i,\delta_j) = \delta_{i,j}$ for all $i,j \in \Ij$. Set $\Lmj := \sum_{i \in \Ij} \Z \delta_i$. Let $\gamma_i := \alpha_{\ul{i}} + J \in \Lmj$, $i \in \Ij$, and $\Qj_+ := \sum_{i \in \Ij} \Z_{\geq 0} \gamma_i \subset \Lmj$. For $\lambda,\mu \in \Lambdaj$, we write $\mu \leqj \lambda$ if $\lambda - \mu \in \Qj_+$. This defines a partial order on $\Lmj$.
For a $\Uj$-module $M$ and $\lambda \in \Lambdaj$, we call $M_\lambda := \{ m \in M \mid k_i m = q^{(\beta_i,\lambda)} m \Forall i \in \Ij \}$ the weight space of $M$ of weight $\lambda$. The category $\Oj$ is the full subcategory of the category of all $\Uj$-modules consisting of $\Uj$-modules $M$ satisfying the following:
\begin{itemize}
\item $M$ has a weight space decomposition, i.e., $M = \bigoplus_{\lambda \in \Lambdaj} M_\lambda$.
\item Each weight space of $M$ is finite-dimensional.
\item There exist $\mu_1,\ldots,\mu_l \in \Lmj$ such that if $M_\lm \neq 0$, then $\lm \leqj \mu_i$ for some $i = 1,\ldots,l$.
\item The $f_i$'s act on $M$ locally nilpotently.
\end{itemize}
\begin{theo}[\cite{W17}]
The following hold:
\begin{enumerate}
\item {\cite[Theorem 4.4.3]{W17}} $\Oj$ is semisimple.
\item {\cite[Corollary 7.6.3, 7.6.4]{W17}} Each irreducible $\Uj$-module in $\Oj$ is isomorphic to the irreducible highest weight module $L(\bflm)$ with highest weight $\bflm$ (in the sense of \cite{W17}) for some $\bflm \in \Pj$.
\item For $\bflm,\bfmu \in \Pj$, we have $L(\bflm) \simeq L(\bfmu)$ if and only if $\bflm_i - \bfmu_i$ is constant as $i$ runs through $-r,\ldots,r$.
\end{enumerate}
\end{theo}
\begin{rem}\normalfont
The last statement follows from the definition of $L(\bflm)$.
\end{rem}
For each $\bflm \in \Pj$, let $\wtj(\bflm) \in \Lmj$ denote the weight of a highest weight vector of $L(\bflm)$, namely,
$$
\wtj(\bflm) := \sum_{i \in \Ij} (\bflm_{i-1}-\bflm_i + \bflm_{-(i-1)}-\bflm_{-i}) \delta_i.
$$
\section{Crystal basis theory}\label{crystal}
\subsection{Crystal bases}\label{sec:crystal bases}
The notion of crystal bases (or local bases at $q = 0$) for integrable modules over quantum groups was introduced independently by Kashiwara and Lusztig in different ways (\cite{K90}, \cite{L90a}). Although we will not review the detail, we formulate here some notations concerning the crystal bases. Let $\Oint$ denote the full subcategory of the BGG-category $\clO$ for $\bfU$ consisting of the integrable modules. Let $\Etil_i, \Ftil_i$, $i \in \bbI$ denote the Kashiwara operators. Let $M \in \Oint$, $(\clL,\clB)$ be its crystal basis. For $b \in \clB$ and $i \in \bbI$, set
$$
\vep_i(b) := \max\{ n \mid \Etil_i^nb \neq 0 \}, \qu \vphi_i(b) := \max\{ n \mid \Ftil_i^nb \neq 0 \}.
$$
Also, $\wt(b) \in \Lm$ denotes the weight of $b$.
Recall that, for each $\lm \in P$, the irreducible module $L(\lm)$ has a unique crystal basis $(\clL(\lm),\clB(\lm))$, which is identical to $\SST(\lm)$. For each $M \in \Oint$ with a crystal basis $(\clL,\clB)$, we have a unique irreducible decomposition $\clB = \bigsqcup_{i=1}^l \clB_i$, where $\clB_i \simeq \clB(\lm_i)$ for some $\lm_i \in P$. By retaking $\lm_i$'s if necessary, we may assume that $|\lm_i| - |\lm_j| < 2r+1$ for all $i,j \in \{1,\ldots,l\}$, and that there exists $i$ such that $(\lm_i)_{2r+1} = 0$. Then, $\lm_i$'s are uniquely determined; we set $P(M) = P_r(M) := \{ \lm_1,\ldots,\lm_l \}$. For $b \in \clB$, we define $I(b) = I_r(b) \in P(M)$ to be $\lm_i$ if $b \in \clB_i$. Also let $C(b) = C_r(b) \subset \clB$ denote the connected component of $\clB$ containing $b$. Furthermore, if we write $b = \Ftil_{i_1} \cdots \Ftil_{i_l}b_0$ for some $i_1,\ldots,i_l \in \bbI$, where $b_0$ denotes the highest weight vector in $C(b)$, then define $T_b \in \SST(I(b))$ by $T_b := \Ftil_{i_1} \cdots \Ftil_{i_l} T_0$, where $T_0 \in \SST(I(b))$ corresponding to $b_0 \in C(b) = \clB(I(b))$.
\subsection{$\jmath$-crystal bases}
In \cite{W17}, the notion of $\jmath$-crystal bases was introduced. Let us recall some properties briefly.
Set $\A := \Q[p,p\inv,q,q\inv]$. We denote by $\A_0$ the subring of $\Q(p,q)$ consisting of all elements of the form $f/g$ with $f,g \in p\Q[p,q,q\inv] \oplus \Q[q]$, $\lim_{q \rightarrow 0} (\lim_{p \rightarrow 0} g) \neq 0$. Let $\IIj := \Ij \sqcup \{ 2',\ldots,r' \}$. The Kashiwara operators are denoted by $\etil_i$ and $\ftil_i$, $i \in \IIj$.
The following are basic results for the crystal basis theory of $\Uj$.
\begin{theo}[{\cite[Theorem 7.7.3]{W17}}]
Let $\bflm \in \Pj$, $v_\bflm \in L(\bflm)$ be a highest weight vector. Set
\begin{align}
\begin{split}
&\clL(\bflm) := \Span_{\Ao} \{ \ftil_{i_1} \cdots \ftil_{i_l} v_\bflm \mid l \in \Z_{\geq 0},\ i_1,\ldots,i_l \in \IIj \}, \\
&\clB(\bflm) := \{ \ftil_{i_1} \cdots \ftil_{i_l} v_\bflm + q\clL(\bflm) \mid l \in \Z_{\geq 0},\ i_1,\ldots,i_l \in \IIj \} \setminus \{ 0 \}.
\end{split} \nonumber
\end{align}
Then, $(\clL(\bflm),\clB(\bflm))$ is a unique $\jmath$-crystal basis of $L(\bflm)$. Moreover, $\clB(\bflm)$ is identical to $\SST(\bflm)$; $v_\bflm + q\clL(\bflm) \in \clB(\bflm)$ corresponds to $\bfT_\bflm \in \SST(\bflm)$.
\end{theo}
\begin{theo}\label{restriction of crystal}
Suppose that $M \in \Oint$ has a crystal basis $(\clL,\clB)$. Then, as a $\Uj$-module, $M$ has a $\jmath$-crystal basis whose underlying sets are equal to $(\clL,\clB)$.
\end{theo}
\begin{proof}
This is an easy consequence of \cite[Corollary 7.7.4]{W17}.
\end{proof}
Let $M \in \Oj$ with a $\jmath$-crystal basis $(\clL,\clB)$. For each $b \in \clB$ and $i \in \IIj$, define $\vep_i(b)$, $\vphi_i(b)$, $\wtj(b) \in \Lmj$, $\Pj(M) = \Pj_r(M) \subset \Pj$, $I^\jmath(b) = I^\jmath_r(b) \in \Pj(M)$, $\Cj(b) = \Cj_r(b) \subset \clB$, and $T^\jmath_b \in \SST(I^\jmath(b))$ in a similar way to Section \ref{sec:crystal bases}.
\section{Global bases}\label{global bases}
\subsection{Balanced triples}
Let $\ol{\ \cdot \ }$ be the $\Q$-linear automorphism of $\Q(p,q)$ sending $p$ and $q$ to $p\inv$ and $q\inv$, respectively. Set $\Ai := \ol{\A_0}$.
\begin{defi}\normalfont
Let $V$ be a $\Q(p,q)$-vector space and $x \in \{ 0,\emptyset,\infty \}$. An $\A_x$-lattice of $V$ is a free $\A_x$-submodule $U_x$ of $V$ of rank $\dim_{\Q(p,q)} V$ such that $\Q(p,q) \otimes_{\A_x} U_x = V$.
\end{defi}
\begin{defi}[{\cite[Definition 2.1.2]{K93}}]\normalfont
Let $V$ be a $\Q(p,q)$-vector space, $U_x$ an $\A_x$-lattice of $V$ for $x \in \{ 0,\emptyset,\infty \}$. The triple $(U_0,U,U_\infty)$ is said to be balanced if the canonical map
\begin{align}
U_0 \cap U \cap U_\infty \rightarrow U_0/qU_0 \nonumber
\end{align}
is an isomorhism of $\Q$-vector spaces.
\end{defi}
Let $V$ be a $\Q(p,q)$-vector space with a balanced triple $(U_0,U,U_\infty)$. Take a $\Q$-basis $\clB$ of $U_0/qU_0$. Since we have an isomorphism $G:U_0/qU_0 \rightarrow U_0 \cap U \cap U_\infty$ of $\Q$-vector spaces, which is the inverse of the canonical map $U_0 \cap U \cap U_\infty \rightarrow U_0 / qU_0$, we obtain an $\A_x$-basis $G(\clB) = \{ G(b) \mid b \in \clB \}$ of $U_x$ for each $x \in \{ 0,\emptyset,\infty \}$. We call $G(\clB)$ the global basis of $V$ associated to the balanced triple $(U_0,U,U_\infty)$ and the basis $\clB$.
\begin{lem}\label{sub and quotient of bt}
Let $V,U_0,U,U_\infty,\clB,G$ be as above. Take a subset $\clB' \subset \clB$ and set $U'_x$ to be the $\A_x$-span of $G(\clB') := \{ G(b) \mid b \in \clB' \}$ for each $x \in \{ 0,\emptyset,\infty \}$. Also, let $V'$ be the $\Q(p,q)$-span of $G(\clB')$. Then, the following hold:
\begin{enumerate}
\item $(U'_0,U',U'_\infty)$ is a balanced triple with the global basis $G(\clB')$.
\item $(U_0/U'_0,U/U',U_\infty/U'_\infty)$ is a balanced triple with the global basis $\{ G(b) + V' \mid b \in \clB \setminus \clB' \}$.
\end{enumerate}
\end{lem}
\subsection{Global crystal bases and global $\jmath$-crystal bases}
Let $\UA$ denote the $\A$-subalgebra of $\U$ generated by $E_i^{(n)}, F_i^{(n)}, K_i^{\pm1}$, $i \in \bbI$, $n \in \Z_{> 0}$. Similaly, define $\UjA$ to be the $\A$-subalgebra of $\Uj$ generated by $e_i^{(n)}, f_i^{(n)}, k_i^{\pm1}$, $i \in \Ij$, $n \in \Z_{> 0}$.
\begin{lem}[{\cite[1.3.5]{L94}}]\label{q-binomial formula}
Let $A$ be a $\Q(q)$-algebra, $x,y \in A$ such that $xy = q^2yx$. Then, for each $n \in \Z_{> 0}$, we have
$$
(x+y)^n = \sum_{t=0}^n q^{t(n-t)} {n \brack t} y^t x^{n-t}.
$$
\end{lem}
\begin{lem}\label{Uja in UA}
We have $\UjA \subset \UA$.
\end{lem}
\begin{proof}
It suffices to show that $e_i^{(n)},f_i^{(n)} \in \UA$ for all $i \in \Ij$, $n \in \Z_{>0}$. We prove $e_i^{(n)} \in \UA$; the proof for $f_i^{(n)} \in \UA$ is similar. Setting $x := E_{\ul{i}}$ and $y := p^{-\delta_{i,1}}F_{-\ul{i}} K_{\ul{i}}\inv$, we see that
$$
e_i = x+y, \qu xy = q^2yx.
$$
Then, we can apply Lemma \ref{q-binomial formula}, and obtain
$$
e_i^{(n)} = \sum_{t=0}^n q^{t(n-t)} y^{(t)} x^{(n-t)}.
$$
It is easy to see that $y^{(t)} = p^{-\delta_{i,1}t} q^{-\delta_{i,1} \frac{t(t-1)}{2}} F_{-\ul{i}}^{(t)} K_{\ul{i}}^t \in \UA$. Hence, the assertion follows.
\end{proof}
Let $V$ be a $\U$-module in $\Oint$ (resp., $\Uj$-module in $\Oj$) with a crystal basis $(\clL,\clB)$ (resp., $\jmath$-crystal basis $(\clL,\clB)$). Assume that $V$ admits a $\Q$-linear involution $\ol{\ \cdot \ }$ satisfying the following:
\begin{align}
\begin{split}
&\ol{xv} = \psi(x) \ol{v}, \qu \Forall x \in \U,\ v \in V \\
(\text{resp., }&\ol{xv} = \psij(x) \ol{v}, \qu \Forall x \in \Uj,\ v \in V).
\end{split} \nonumber
\end{align}
We call such an involution a $\psi$-involution (resp., $\psij$-involution) on $V$. Since $\clL$ is an $\Ao$-lattice of $V$, $\ol{\clL}$ is an $\Ai$-lattice of $V$.
\begin{defi}\normalfont
Let $V,\clL,\clB,\ol{\ \cdot \ }$ be as above. $V$ is said to have a global crystal basis (resp., global $\jmath$-crystal basis) if there exists a $\UA$-submodule (resp., $\UjA$-submodule) $V_\A$ of $V$ which is an $\A$-lattice forming a balanced triple $(\clL,V_\A,\ol{\clL})$. The associated global basis $G(\clB)$ (resp., $\Gj(\clB)$) is called a global crystal basis (resp., global $\jmath$-crystal basis) of $V$.
\end{defi}
\begin{ex}\normalfont
Let $\bflm \in \Pj_1$ and consider the irreducible $\Uj_1$-module $L(\bflm)$. Recall that $L(\bflm)$ is $(\bflm_0-\bflm_{-1}+1)$-dimensional with a basis $\Gj(\bflm) := \{ f_1^{(n)}v \mid 0 \leq n \leq \bflm_0-\bflm_{-1} \}$, where $v$ denotes a highest weight vector. Also, $L(\bflm)$ has a $\jmath$-crystal basis $(\clL(\bflm),\clB(\bflm))$, where $\clL(\bflm)$ is the $\Ao$-span of $\Gj(\bflm)$, and $\clB(\bflm) = \{ f_1^{(n)}v + q\clL(\bflm) \mid 0 \leq n \leq \bflm_0-\bflm_{-1} \}$. Set $L(\bflm)_\A$ to be the $\A$-span of $\Gj(\bflm)$. Note that there exists a unique $\psij$-involution $\psij_{\bflm}$ on $L(\bflm)$ fixing $v$. Then, $(\clL(\bflm), L(\bflm)_\A, \psij_\bflm(\clL(\bflm)))$ is a balanced triple, and $\Gj(\bflm)$ is a global $\jmath$-crystal basis of $L(\bflm)$.
\end{ex}
\begin{prop}
Let $M \in \Oj$ with a global crystal $\jmath$-crystal basis $\Gj(\clB_M)$, and $N \in \Oint$ with a global crystal basis $G(\clB_N)$ Then, $M \otimes N$ has a global $\jmath$-crystal basis of the form
\begin{align}
\begin{split}
&\{ \Gj(b_1) \diamondsuit G(b_2) \mid b_1 \in \clB_M,\ b_2 \in \clB_N \}. \\
&\Gj(b_1) \diamondsuit G(b_2) \in \Gj(b_1) \otimes G(b_2) + \sum_{\substack{b'_1 \in \clB_M,\ b'_2 \in \clB_N \\ \wt(b'_2) < \wt(b_2)}} a_{b'1,b'_2;b_1,b_2} \Gj(b'_1) \otimes G(b'_2), \qu a_{b'1,b'_2;b_1,b_2} \in \A.
\end{split} \nonumber
\end{align}
\end{prop}
\begin{proof}
The fact that $\clB_M \otimes \clB_N$ forms a $\jmath$-crystal basis of $M \otimes N$ is proved in \cite{W17}. Now, one can construct a global $\jmath$-crystal basis of $M \otimes N$ with the desired property in the same way as the proof of \cite[Theorem 4]{BWW18}.
\end{proof}
\subsection{$\jmath$-canonical bases}
In this subsection, we recall the notion of $\jmath$-canonical bases, which was introduced by H. Bao and W. Wang in \cite{BW13}, and explain that $\jmath$-canonical bases are global $\jmath$-crystal bases. One of the key ingredients for a construction of $\jmath$-canonical bases is the intertwiner $\Upsilon$:
\begin{defi}[{\cite[Theorem 6.4]{BW13}}]
Let $\U^-$ denote the subalgebra of $\U$ generated by $F_i$, $i \in \bbI$. For each $\lm \in Q_+$, there exists a unique $\Upsilon_\lm \in \U^-_{-\lm}$ satisfying the following:
\begin{itemize}
\item $\Upsilon_0 = 1$, \\
\item $\Upsilon := \sum_{\lm \in Q_+} \Upsilon_\lm$ satisfies $\psij(x)\Upsilon = \Upsilon \psi(x)$ for all $x \in \Uj$.
\end{itemize}
\end{defi}
\begin{lem}[{\cite[Proposition 6.12]{BW13}}]
Let $M \in \Oint$ with a $\psi$-involution $\psi_M$. Then, the composite $\Upsilon \circ \psi_M$ is a $\psij$-involution of $M$.
\end{lem}
\begin{theo}[{\cite[Theorem 6.24]{BW13}}]\label{j-canonical}
Let $M \in \Oint$ have a global crystal basis $G(\clB)$ with a crystal basis $(\clL,\clB)$, a $\psi$-involution $\psi_M$, and an $\A$-lattice $M_\A$. Set $\psij_M := \Upsilon \circ \psi_M$. Then, for each $b \in \clB$, there exists a unique $\Gj(b) \in M$ satisfying the following.
\begin{enumerate}
\item $\psij_M(\Gj(b)) = \Gj(b)$.
\item $\Gj(b) = G(b) + \sum_{b' \in B} c_{b',b} G(b')$ for some $c_{b',b} \in q\Ao \cap \A$. Moreover, $c_{b',b} = 0$ unless $\wtj(b') = \wtj(b)$ and $\wt(b') < \wt(b)$.
\end{enumerate}
\end{theo}
The new basis $\Gj(\clB) := \{ \Gj(b) \mid b \in \clB \}$ thus constructed is called the $\jmath$-canonical basis of $(M,G(\clB))$.
\begin{prop}
We keep the notation in Theorem \ref{j-canonical}. Then, $(\clL,\clB)$ is a $\jmath$-crystal basis, $(\clL,M_{\A},\psij_M(\clL))$ is a balanced triple, and $\Gj(\clB)$ is the global $\jmath$-crystal basis associated to the balanced triple $(\clL,M_{\A},\psij_M(\clL))$ and the basis $\clB$.
\end{prop}
\begin{proof}
That $(\clL,\clB)$ is a $\jmath$-crystal basis has already been stated in Theorem \ref{restriction of crystal}. Let us prove the rest. By the property $(2)$ of Theorem \ref{j-canonical}, it is clear that $\clL$ (resp., $M_\A$) is spanned by $\Gj(\clB)$ over $\Ao$ (resp., $\A$). Also, by $(1)$ of Theorem \ref{j-canonical}, $\psij_M(\clL)$ is spanned by $\Gj(\clB)$ over $\Ai$. Hence, the canonical homomorphism $\clL \cap M_{\A} \cap \psij_M(\clL) \rightarrow \clL/q\clL$ is an isomorphism, and therefore, $(\clL,M_{\A},\psij_M(\clL))$ is balanced. Finally, by Lemma \ref{Uja in UA}, the $\UA$-module $M_\A$ is also a $\UjA$-module. This proves the proposition.
\end{proof}
\section{Notations}\label{notations}
Throughout this paper, we fix a positive integer $r$. For $n \in \hf \Z$, set $\ul{n} := n - \hf$. Note that $-\ul{n} = -n + \hf \neq \ul{-n}$. We set
$$
I := \{ -r,\ldots,-1,0,1,\ldots,r \}, \qu \bbI := \{ -\ul{r}, \ldots, -\ul{1}, \ul{1}, \ldots, \ul{r} \}, \qu \Ij := \{ 1,\ldots,r \}.
$$
A partition of $n \in \N$ of length $l \in \N$ is a nonincreasing sequence $\lm = (\lm_1,\ldots,\lm_l)$ of nonnegative integers satisfying $\sum_{i=1}^l \lm_i = n$. Let $|\lm| := n$ and $\ell(\lm) := l$, and call them the size and the length of $\lm$, respectively. We denote by $\Par_l(n)$ the set of partitions of $n$ of length $l$.
We often identify a partition with a Young diagram in a usual way. Let $(L,\preceq)$ be a totally ordered set. A semistandard tableau of shape $\lm \in \Par_l(n)$ in letters $L$ is a filling of the Young diagram $\lm$ with elements of $L$, which weakly increases (with respect to the total order $\preceq$) from left to right along the rows, and strictly increases from the top to the bottom along the columns.
A bipartition of $n \in \N$ of length $(l;m) \in \N^2$ is an ordered pair $\bflm := (\bflm^-;\bflm^+)$ of partitions such that $\ell(\bflm^-) = l$, $\ell(\bflm^+) = m$, and $|\bflm| := |\bflm^-|+|\bflm^+| = n$. We denote by $P_{(l;m)}(n)$ the set of bipartitions of $n$ of length $(l;m)$. For totally ordered sets $(L^-,\preceq^-)$ and $(L^+,\preceq^+)$, a semistandard tableau of shape $\bflm \in P_{(l;m)}(n)$ in letters $(L^-;L^+)$ is an ordered pair $\bfT = (T^-;T^+)$, where $T^\pm$ is a semistandard tableau of shape $\bflm^\pm$ in letters $L^\pm$.
For partitions $\mu \subset \lm$, define the skew partition $\lm/\mu$ in a usual way. For bipartitions $\bfmu \subset \bflm$ (i.e., $\bfmu^- \subset \bflm^-$ and $\bfmu^+ \subset \bflm^+$), define the skew bipartition $\bflm/\bfmu$ to be $(\bflm^-/\bfmu^-;\bflm^+/\bfmu^+)$. A skew partition $\lm/\mu$ is said to be a horizontal strip if each column of $\lm/\mu$ contains at most one box. We say that a skew bipartition $\bflm/\bfmu$ is a horizontal strip if $\bflm^\pm/\bfmu^\pm$ are.
Set
\begin{itemize}
\item $P(n) = P_r(n) := \Par_{2r+1}(n)$: the set of partitions of $n$ of length $2r+1$.
\item $P = P_r := \bigsqcup_{n \in \N} P(n)$: the set of partitions of length $2r+1$.
\item $\Par_l := \bigsqcup_{n \in \N} \Par_l(n)$: the set of partitions of length $l$.
\item $\Pj(n) = \Pj_r(n) := P_{(r+1;r)}(n)$: the set of bipartitions of $n$ of length $(r+1;r)$.
\item $\Pj = \Pj_r := \bigsqcup_{n \in \N} \Pj(n)$: the set of bipartitions of length $(r+1;r)$.
\item $\SST(\lm)$: the set of semistandard tableaux of shape $\lm \in P(n)$ in letters $I$.
\item $\SST(\bflm)$: the set of semistandard tableaux of shape $\bflm \in \Pj(n)$ in letters $(I \setminus \Ij;\Ij)$ with total orders $0 \prec^- -1 \prec^- \cdots \prec^- -r$ and $1 \prec^+ \cdots \prec^+ r$.
\end{itemize}
For $\bflm \in \Pj$, we refer the $i$-th row of $\bflm^-$ to as the $-(i-1)$-th row of $\bflm$, and the $j$-th row of $\bflm^+$ to as the $j$-th row of $\bflm$. Also, for $i \in I$, set $\bflm_i$ to be the length of the $i$-th row of $\bflm$, i.e.,
$$
\bflm_i := \begin{cases}
\bflm^-_{-i+1} \qu & \IF i \leq 0, \\
\bflm^+_i \qu & \IF i > 0.
\end{cases}
$$
For $i \in \Ij$, set $\bflm \downarrow_i := (\bflm_0,\bflm_{-1},\ldots,\bflm_{-i};\bflm_1,\ldots,\bflm_i) \in \Pj_i$.
For $\bfT \in \SST(\bflm)$ and $i \in \Ij$, set $\bfT \downarrow_i$ to be the semistandard tableau obtained from $\bfT$ by deleting the boxes whose entries are less than $-i$ or greater than $i$.
For each $\bflm \in \Pj$, let $\bfT_\bflm \in \SST(\bflm)$ be the unique semistandard tableau of shape $\bflm$ whose entries in the $i$-th row are $i$. Note that we have $\bfT_\bflm \downarrow_i = \bfT_{\bflm \downarrow_i}$. For $\bfT \in \SST(\bflm)$ and $i \in I$, set $\bfT(i)$ to be the number of boxes of $\bfT$ whose entries are $i$.
\begin{defi}\normalfont
\
\begin{enumerate}
\item $\preceq$ is a partial order (called the dominance order) on $\Par_l$ defined as follows. For $\lm, \mu \in \Par_l$, we have $\lm \preceq \mu$ if
\begin{enumerate}
\item $|\lm| = |\mu|$ and
\item $\sum_{i=1}^j \lm_i \leq \sum_{i=1}^j \mu_i$ for all $1 \leq j \leq l$.
\end{enumerate}
\item $\preceq$ is a partial order (also called the dominance order) on $\Pj$ defined as follows. For $\bflm, \bfmu \in \Pj$, we have $\bflm \preceq \bfmu$ if
\begin{enumerate}
\item $|\bflm| = |\bfmu|$,
\item $\sum_{i=0}^j \bflm_{-i} \leq \sum_{i=0}^j \bfmu_{-i}$ for all $0 \leq j \leq r$, and
\item $|\bflm^-| + \sum_{i=1}^j \bflm_i \leq |\bfmu^-| + \sum_{i=1}^j \bfmu_i$ for all $1 \leq j \leq r$.
\end{enumerate}
\item $\trilefteq$ is a partial order on $\Pj$ defined as follows. For $\bflm,\bfmu \in \Pj$, we have $\bflm \trilefteq \bfmu$ if $\bflm^- \preceq \bfmu^-$ (dominance order on $\Par_{r+1}$) and $\bflm^+ \preceq \bfmu^+$ (dominance order on $\Par_r$).
\end{enumerate}
\end{defi}
Clearly, $\bflm \trilefteq \bfmu$ implies $\bflm \preceq \bfmu$.
|
\section{Introduction}
A zonotope is the Minkowski sum of several line segments.
Given a set of vectors $V=\{v_1,\dots,v_m\}$ in $\R^n$, we define a zonotope
\[Z(V)=[0,v_1]+\dots + [0,v_m].\]
The combinatorial structure of faces of $Z(V)$ is totally determined by the oriented matroid $\Mt(V)$ of $V$: the face poset of $Z(V)$ is anti-isomorphic with the poset of covectors of $\Mt(V)$ \cite{BLSWZ99}. In this paper we are interested in constructing and understanding zonotopes which are combinatorially different but have the same face vector.
Let $G=(V(G),E(G))$ be a simple (no loops or multiple
edges) connected graph on vertex set $\{1,2,\dots,n\}$. The graphical zonotope $Z_G$ of $G$ is defined by
\[Z_G=\sum_{ij\in E(G)}[e_i, e_j],\]
where $e_1,\dots,e_n$ are the coordinate vectors in $\R^n$. Many properties of graph $G$ are encoded by its graphical zonotope. For example, the volume of $Z_G$ equals the number of spanning trees of $G$ and the number of lattice points in $Z_G$ equals to the number of forests in $G$ \cite[Proposition 2.4]{postnikov09}. The number of vertices of $Z_G$ is equal to the number of acyclic orientations of $G$.
Graphical zonotope of the complete graph is just the permutahedron and all graphical zonotopes are in the class of generalized permutahedron \cite{postnikov09}.
Gruji\'c \cite{Grujic17} showed a relation between the face vector of $Z_G$ and the q-analog of the chromatic symmetric function of $G$. Study of face vectors of graphical zonotopes can also be found in Postnikov et.al. \cite{PRW08}.
For each edge $ij$ of $G$, we denote by $\overrightarrow{{ij}}$ the orientation of $ij$ from $i$ to $j$. A partial orientation $X$ of $G$ is an orientation of a subgraph $H$ of $G$. We can regard $X$ as the directed graph whose underlying graph is $H$. If all the edges of $G$ are oriented, then we say $X$ is a full orientation. The {\em genus} $g(X)$ of $X$ is defined to be the genus $g(H)$ of $H$, which is the edge number of $H$ minus the vertex number of $H$ and then plus the number of connected components of $H$. Let $X^0$ denote the subgraph of $G$ with edge set $E(G)\setminus E(H)$. For two partial orientations $X, Y$ of $G$, we define $X\leq Y$ if $X$ is a sub-digraph of $Y$. Two partial orientations $X$ and $Y$ are {\em orthogonal} if either no edge is simultaneously oriented by $X$ and $Y$ or there exist at least two edges such that one is oriented by $X$ and $Y$ in a same direction and the other is oriented by $X$ and $Y$ in reverse directions.
For a graph $G$, we let $r(G)$ be the edge number of a maximal spanning forest of $G$. In the following, we introduce the oriented matroid of graph $G$ in terms of partial orientations.
A partial orientation $X$ of $G$ is a {\em vector of $G$} if every arc (oriented edge) of $X$ is contained in a directed circuit. Let $\mathcal V(G)$ denote the set of vectors of graph $G$. Then the poset $(\mathcal V(G),\leq)$ is graded and the rank $rank(X)$ of each vector $X$ is given by $g(X)$.
A partial orientation $X$ of $G$ is a {\em covector of $G$} if it is orthogonal with every directed circuit of $G$. Let $\mathcal L(G)$ denote the set of covectors of graph $G$. Then the poset $(\mathcal L(G),\leq)$ is graded and the rank $rank(X)$ of each covector $X$ is given by $r(G)-r(X^0)$ (see \cite[Corollary 4.1.15]{BLSWZ99}).
Given a planar graph $G$, let $G^*$ be the dual planar graph of $G$. Then $(\mathcal V(G^*),\leq)$ is isomorphic with $(\mathcal L(G),\leq)$.
\begin{proposition} \label{prop:anti}
The face poset of $Z_G$ is anti-isomorphic with the poset $(\mathcal L(G), \leq)$.
\end{proposition}
This paper is organized as follows. In Section \ref{sec:flip} we introduce the quadrilateral flip operation on graphs and prove that graphical zonotopes of two flip-equivalent graphs have the same face vector. In Section \ref{sec:gon} we compute the face vector and total face number of graphical zonotopes of triangulations of the $n$-gon.
\section{Flip on graphs}
\label{sec:flip}
Flipping has long been important in the study of
triangulations. Given a triangulation of a point configuration in the plane, a flip means replacing two triangles by two different triangles that cover the same quadrilateral. The flip graph is a graph defined on triangulations such that two triangulations are connected if they are related by a flip. Please refer to \cite{bose09} for a survey of the flip graph of triangulations of planar point set. A generalization of flip operation to higher dimension is called bistellar flip \cite{de10}.
We define the following quadrilateral flip operation on graphs.
\begin{definition}[quadrilateral flip]
Let $G$ be a simple connected graph. Let $v_1,v_2,v_3,v_4$ be four vertices of $G$ such that the subgraph $H$ induced by these four vertices is a $4$-cycle with edges $v_1v_2,v_2v_3,v_3v_4,v_4v_1$ and a diagonal edge $v_2v_4$. Suppose further that we can divide $V(G)\setminus \{v_1,v_2,v_3,v_4\}$ into four disjoint sets $V_1,V_2,V_3,V_4$ such that, for every edge $uv$ of $G$ which is not $v_2v_4$, $u$ and $v$ belong to $V_i\cup\{v_i,v_{i+1}\}$ ($v_5$ is $v_1$) for some $i\in \{1,2,3,4\}$. Then let $G'$ be the new graph obtained from $G$ by deleting edge $v_2v_4$ and then adding edge $v_1v_3$. Such an operation from $G$ to $G'$ is called a {\em quadrilateral flip}, or simply a {\em flip}.
\end{definition}
See figure below for illustration of the flip operation. We say two graphs are flip-equivalent if one can be transformed to another through a sequence of flip operations.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[inner sep=0pt, v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (u) to (a) to (v) to (b) to (u) to (v);
\draw[-,dashed] (a) to [out=90,in=180] (0,1.5) to [out=0,in=90] (u);
\draw[-,dashed] (u) to [out=0,in=90] (1.5,0) to [out=-90,in=0] (b);
\draw[-,dashed] (b) to [out=-90,in=0] (0,-1.5) to [out=-180,in=-90](v);
\draw[-,dashed] (v) to [out=180,in=-90] (-1.5,0)to [out=90,in=180] (a);
\node at (3,0) [label=above:flip]{$\longrightarrow$};
\tikzset{xshift=6cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (a) to (v) to (b) to (u) to (a) to (b);
\draw[-,dashed] (a) to [out=90,in=180] (0,1.5) to [out=0,in=90] (u);
\draw[-,dashed] (u) to [out=0,in=90] (1.5,0) to [out=-90,in=0] (b);
\draw[-,dashed] (b) to [out=-90,in=0] (0,-1.5) to [out=-180,in=-90](v);
\draw[-,dashed] (v) to [out=180,in=-90] (-1.5,0)to [out=90,in=180] (a);
\end{tikzpicture}
\end{figure}
\begin{theorem}\label{thm:flip}
Graphical zonotopes of two flip-equivalent graphs have the same face vector.
\end{theorem}
\begin{proof}
Let $G'$ be a graph obtained from a graph $G$ by a flip that removes edge $v_2v_4$ and adds edge $v_1v_3$. It is obvious that $r(G)=r(G')$. For each $i=1,2,3,4$, let $G_i$ be the induced subgraph of $G$ with vertex set $V_i\cup \{v_i,v_{i+1}\}$. Let $H$ and $H'$ be the induced subgraphs of $G$ and $G'$ with vertex set $\{v_1,v_2,v_3,v_4\}$, respectively. In what follows, we will explicitly construct a bijection from covectors $X$ of $G$ to covectors $Y$ of $G'$ such that $rank(X)=rank(Y)$. Note that $Y$ is a covector of $G'$ if its restrictions on $G_1,\dots,G_4$ and $H'$ are covectors of $G_1,\dots,G_4$ and $H'$, respectively. Let $X'$ and $Y'$ be the restriction of $X$ and $Y$ on $H$ and $H'$, respectively.
\textbf{Case 1.} No edge in $H$ is oriented by $X$. Let $Y$ equals $X$ on edges not in $H'$ and let $Y$ do not orient any edge in $H'$. It is obvious that $rank(X)=rank(Y)$.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[ inner sep=0pt, v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (u) to (a) to (v) to (b) to (u) to (v);
\draw[-,directed] (u) to (a);
\draw[-,directed] (v) to (a);
\node at (1,0) {$\longrightarrow$};
\node at (1,-1.3) {$(1)$};
\tikzset{xshift=2cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (a) to (v) to (b) to (u) to (a) to (b);
\draw[-,directed] (u) to (a);
\draw[-,directed] (v) to (a);
\draw[-,directed] (b) to (a);
\tikzset{xshift=2cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (u) to (a) to (v) to (b) to (u) to (v);
\draw[-,directed] (u) to (a);
\draw[-,directed] (u) to (b);
\draw[-,directed] (u) to (v);
\node at (1,0) {$\longrightarrow$};
\node at (1,-1.3) {$(2)$};
\tikzset{xshift=2cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (a) to (v) to (b) to (u) to (a) to (b);
\draw[-,directed] (u) to (a);
\draw[-,directed] (u) to (b);
\tikzset{xshift=2cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (u) to (a) to (v) to (b) to (u) to (v);
\draw[-,directed] (u) to (a);
\draw[-,directed] (b) to (v);
\draw[-,directed] (u) to (v);
\node at (1,0) {$\longrightarrow$};
\node at (1,-1.3) {$(3)$};
\tikzset{xshift=2cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (a) to (v) to (b) to (u) to (a) to (b);
\draw[-,directed] (u) to (a);
\draw[-,directed] (b) to (v);
\draw[-,directed] (b) to (a);
\end{tikzpicture}
\caption{Case 2.}
\label{fig:casetwo}
\end{figure}
\textbf{Case 2.} $X$ orients two edges in the $4$-cycle of $H$. This case has three subcases: $(1)$, two edges incident with $v_1$ or $v_3$ are oriented; $(2)$, two edges incident with $v_2$ or $v_4$ are oriented, then $v_2v_4$ is also oriented; $(3)$, two non-adjacent edges are oriented, then $v_2v_4$ is also oriented. We have shown one example for each of the three cases in Fig. \ref{fig:casetwo}. For each case, we let $Y$ equals $X$ on all edges of $G'$ other than $v_1v_3$. Then the orientation situation of $Y$ on edge $v_1v_3$ is determined. For example, in figure $(1)$, since $\overrightarrow{v_2v_1},\overrightarrow{v_4v_1}\in Y$, we must have $\overrightarrow{v_3v_1}\in Y$ to make $Y$ a covector. It is easy to see that $Y$ is indeed a covector of $G'$. Moreover, we have $rank(X)=rank(Y)$. Indeed, for example, in case $(1)$, $X^0$ can be obtained from $Y^0$ by adding one more edge between two vertices in a same connected component of $Y^0$. So $r(X^0)=r(Y^0)$, and thus $rank(X)=rank(Y)$.
\textbf{Case 3.} $X$ orients three edges in the $4$-cycle of $H$. See Fig. \ref{fig:casethree} for an example. In this case $u_2u_4$ is also oriented. If not, then there is a triangle with only one edge oriented, contradicting the fact that $X$ is a covector. We let $Y$ equals $X$ on edges of $G'$ other than $u_1u_3$. Note that there is also a triangle of $H'$ containing edge $u_1u_3$ and exactly one of the two boundary edges is oriented. So let $Y$ orient edge $u_1u_3$ in the only allowed direction. Now $X^0=Y^0$, so $rank(X)=rank(Y)$.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[ inner sep=0pt, v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (u) to (a) to (v) to (b) to (u) to (v);
\draw[-,directed] (u) to (a);
\draw[-,directed] (b) to (v);
\draw[-,directed] (u) to (v);
\draw[-,directed] (b) to (u);
\node at (1.5,0) {$\longrightarrow$};
\tikzset{xshift=3cm}
\node[v](v) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](a) [above =of v, label=above left:$v_1$] {};
\node[v](b) [ right =of v, label=below right:$v_3$] {};
\node[v](u) [right =of a, label=above right:$v_2$] {};
\draw[-] (a) to (v) to (b) to (u) to (a) to (b);
\draw[-,directed] (u) to (a);
\draw[-,directed] (b) to (v);
\draw[-,directed] (b) to (u);
\draw[-,directed] (b) to (a);
\end{tikzpicture}
\caption{Case 3.}
\label{fig:casethree}
\end{figure}
\textbf{Case 4.} $X$ orients four edges in the $4$-cycle of $H$. In this case we construct $Y$ in the following way. We first let $Y'$ be isomorphic with $X'$ under the map $v_i\rightarrow v_{i+1}$. So if forgetting about labels, then $Y'$ looks just like the rotation of $X'$ by $90$ degree. For each $i=1,2,3,4$, let $X_i$ and $Y_i$ be the restriction of $X$ and $Y$ on $G_i$ respectively. Then for each $i=1,2,3,4$, if $v_iv_{i+1}$ is oriented in the same direction for $X'$ and $Y'$, set $Y_i=X_i$; otherwise, let $Y_i$ be the reverse of $X_i$, denoted as $-X_i$. See figure below for illustration. Though in example below $v_2v_4$ is oriented in $X$, we also have cases that $v_2v_4$ is not oriented.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[inner sep=0pt, v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\node[v](v4) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](v1) [above =of v4, label=above left:$v_1$] {};
\node[v](v3) [ right =of v4, label=below right:$v_3$] {};
\node[v](v2) [above right =of v4, label=above right:$v_2$] {};
\draw[-,directed] (v1) to (v2);
\draw[-,directed] (v1) to (v4);
\draw[-,directed] (v2) to (v4);
\draw[-,directed] (v2) to (v3);
\draw[-,directed] (v3) to (v4);
\draw[-,dashed] (v1) to [out=90,in=180] (0,1.5) to [out=0,in=90] (v2);
\draw[-,dashed] (v2) to [out=0,in=90] (1.5,0) to [out=-90,in=0] (v3);
\draw[-,dashed] (v3) to [out=-90,in=0] (0,-1.5) to [out=-180,in=-90](v4);
\draw[-,dashed] (v4) to [out=180,in=-90] (-1.5,0)to [out=90,in=180] (v1);
\node at (0,1) {$X_1$};
\node at (1,0) {$X_2$};
\node at (0,-1) {$X_3$};
\node at (-1,0) {$X_4$};
\node at (3,0) {$\longrightarrow$};
\tikzset{xshift=6cm}
\node[v](v4) at (-0.5,-0.5) [label=below left:$v_4$] {};
\node[v](v1) [above =of v4, label=above left:$v_1$] {};
\node[v](v3) [ right =of v4, label=below right:$v_3$] {};
\node[v](v2) [above right =of v4, label=above right:$v_2$] {};
\draw[-,directed] (v4) to (v1);
\draw[-,directed] (v4) to (v3);
\draw[-,directed] (v1) to (v3);
\draw[-,directed] (v1) to (v2);
\draw[-,directed] (v2) to (v3);
\draw[-,dashed] (v1) to [out=90,in=180] (0,1.5) to [out=0,in=90] (v2);
\draw[-,dashed] (v2) to [out=0,in=90] (1.5,0) to [out=-90,in=0] (v3);
\draw[-,dashed] (v3) to [out=-90,in=0] (0,-1.5) to [out=-180,in=-90](v4);
\draw[-,dashed] (v4) to [out=180,in=-90] (-1.5,0)to [out=90,in=180] (v1);
\node at (0,1) {$X_1$};
\node at (1,0) {$X_2$};
\node at (0,-1) {$-X_3$};
\node at (-1,0) {$-X_4$};
\end{tikzpicture}
\end{figure}
It is easy to see that $Y$ is a covector of $G'$. Now we will show that $rank(X)=rank(Y)$. Note that for each $i\in \{1,2,3,4\}$, there is no path from $v_i$ to $v_{i+1}$ in $X_i^0$. If not, together with edge $v_iv_{i+1}$ we have a cycle with only one edge oriented, contradicting the fact that $X$ is a covector. So we see that $r(X^0)=r(Y^0)=r(X_1^0)+\dots+r(X_4^0)$ when $v_2v_4$ is oriented and $r(X^0)=r(Y^0)=r(X_1^0)+\dots+r(X_4^0)+1$ when $v_2v_4$ is not oriented. So $rank(X)=rank(Y)$.
At last, through above cases we have defined a map from covectors $X$ of $G$ to covectors $Y$ of $G'$. We can see that this map is injective. As $G$ and $G'$ are flip-equivalent, we can also define a similar map from $G'$ to $G$. Thus our map is a bijection. We have already shown that this map keeps rank. So the theorem is proved.
\end{proof}
\begin{remark}
In our definition of flip on graphs, the extra condition on $V_1,\dots,V_4$ is essential.
For example, the following two graphs are skeleton graphs of two triangulations of $5$ points on the plane that are related by a flip. But these two graphs are not flip-equivalent in our definition. Their graphical zonotopes do have different face vectors. For the left graph it is $(72,150,102,24)$ and for the right graph it is $(78,168,116,26)$. The computation is done by Polymake \cite{polymake}.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[scale=0.7, inner sep=0pt, v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\draw
(0,0) --++(1,1) --++ (2,-1) --++ (-1,0) --++ (-1,1) --++ (0,-2) --++ (2,1);
\draw[-] (0,0) to (1,-1);
\draw[-] (1,-1) to (2,0);
\tikzset{xshift= 6cm}
\draw
(0,0) --++(1,1) --++ (2,-1) --++ (-3,0) --++ (1,-1) --++ (2,1);
\draw[-] (2,0) to (1,1);
\draw[-] (2,0) to (1,-1);
\end{tikzpicture}
\end{figure}
\end{remark}
\begin{remark}
Whitney \cite{whitney33} characterized when two graphs have isomorphic matroids. The matroids of two flip-equivalent graphs are usually not isomorphic. Thus their graphical zonotopes are usually not combinatorially equivalent.
\end{remark}
\begin{example}
The following kind of graphs admit several flip operations. These graphs are in tree shapes; namely, we can divide such a graph into blocks and then the adjacency relation between these blocks is a tree.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[scale=0.5, inner sep=0pt, v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\def \x{4cm};
\def \y{2cm};
\node[v](v4) at (-0.5,-0.5) {};
\node[v](v1) at (-0.5,0.5) {};
\node[v](v3) at (0.5,-0.5) {};
\node[v](v2) at (0.5,0.5) {};
\draw[-] (v1) to (v2) to (v3) to (v4) to (v1) to (v4) to (v2);
\node[v](m4) at ($(-0.5,-0.5)+(0.5*\x,\y)$) {};
\node[v](m1) at ($(-0.5,0.5)+(0.5*\x,\y)$) {};
\node[v](m3) at ($(0.5,-0.5)+(0.5*\x,\y)$) {};
\node[v](m2) at ($(0.5,0.5)+(0.5*\x,\y)$) {};
\draw[-] (m1) to (m2) to (m3) to (m4) to (m1) to (m4) to (m2);
\node[v](u4) at ($(-0.5,-0.5)+(\x,0)$) {};
\node[v](u1) at ($(-0.5,0.5)+(\x,0)$) {};
\node[v](u3) at ($(0.5,-0.5)+(\x,0)$) {};
\node[v](u2) at ($(0.5,0.5)+(\x,0)$) {};
\draw[-] (u1) to (u2) to (u3) to (u4) to (u1) to (u3);
\node[v](w4) at ($(-0.5,-0.5)+2*(\x,0)$) {};
\node[v](w1) at ($(-0.5,0.5)+2*(\x,0)$) {};
\node[v](w3) at ($(0.5,-0.5)+2*(\x,0)$) {};
\node[v](w2) at ($(0.5,0.5)+2*(\x,0)$) {};
\draw[-] (w1) to (w2) to (w3) to (w4) to (w1) to (w4) to (w2);
\draw[-,dashed]
(m4) to (v2)
(v3) to (u4)
(u1) to (m3)
(u2) to (w1)
(u3) to (w4)
;
\draw[-,dashed] (v1) to [out=90,in=180] ($0.5*(v1)+0.5*(v2)+(0,1)$) to [out=0,in=90] (v2);
\draw[-,dashed] (v3) to [out=-90,in=0] ($0.5*(v3)+0.5*(v4)+(0,-1)$) to [out=-180,in=-90](v4);
\draw[-,dashed] (v4) to [out=180,in=-90] ($0.5*(v1)+0.5*(v4)+(-1,0)$)to [out=90,in=180] (v1);
\draw[-,dashed] (w1) to [out=90,in=180] ($0.5*(w1)+0.5*(w2)+(0,1)$) to [out=0,in=90] (w2);
\draw[-,dashed] (w2) to [out=0,in=90] ($0.5*(w3)+0.5*(w2)+(1,0)$) to [out=-90,in=0] (w3);
\draw[-,dashed] (w3) to [out=-90,in=0] ($0.5*(w3)+0.5*(w4)+(0,-1)$) to [out=-180,in=-90](w4);
\draw[-,dashed] (u1) to [out=90,in=180] ($0.5*(u1)+0.5*(u2)+(0,1)$) to [out=0,in=90] (u2);
\draw[-,dashed] (u3) to [out=-90,in=0] ($0.5*(u3)+0.5*(u4)+(0,-1)$) to [out=-180,in=-90](u4);
\draw[-,dashed] (m1) to [out=90,in=180] ($0.5*(m1)+0.5*(m2)+(0,1)$) to [out=0,in=90] (m2);
\draw[-,dashed] (m2) to [out=0,in=90] ($0.5*(m3)+0.5*(m2)+(1,0)$) to [out=-90,in=0] (m3);
\draw[-,dashed] (m4) to [out=180,in=-90] ($0.5*(m1)+0.5*(m4)+(-1,0)$)to [out=90,in=180] (m1);
\end{tikzpicture}
\end{figure}
\end{example}
\section{Triangulations of the $n$-gon}
\label{sec:gon}
The $n$-gon is a $2$-dimensional polytope with $n$ edges.
In this section we consider special graphs that are skeleton graphs of triangulations of the $n$-gon. From definition, it is clear that if two triangulations of the $n$-gon are related by a flip, then their skeleton graphs are flip-equivalent.
It is well known that the flip graph of triangulations of the $n$-gon can be realized as the skeleton graph of the $n-2$ dimensional associahedron (also called the Stasheff polytope) \cite{Lee89}. The diameter of this flip graph is $2n-8$ when $n> 11$ \cite{pournin14,sleator88}. See Fig. \ref{fig:fg6} for an example of flip graph of triangulations of the $6$-gon.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.3]{fg6}
\caption{Flip graph of triangulations of the $6$-gon \cite{fg6}.}
\label{fig:fg6}
\end{figure}
\begin{proposition}\label{prop:connectivity}
The flip graph of triangulations of the $n$-gon is connected.
\end{proposition}
In the following, we do not distinguish a triangulation of the $n$-gon and its skeleton graph. Combining Theorem \ref{thm:flip} and Proposition \ref{prop:connectivity}, we immediately see that graphical zonotopes of all triangulations of the $n$-gon have a same face vector. We will prove the following results in this section.
\begin{theorem} \label{thm:facevector}
The graphical zonotope of every triangulation of the $n$-gon has the face vector $f=(f_0,f_1,\dots,f_{n-1})$ where for $i=0,\dots,n-1:$
\[f_{n-1-i}=\sum_{m=0}^{\min\{i,n-i\}}2^m3^{i-m}{i-1\choose m-1}{n \choose m+i}.\]
\end{theorem}
\begin{theorem}\label{thm:totalnumber}
The total number of faces of the graphical zonotope of a triangulation of the $(n+2)$-gon is $\left(
\begin{array}{cc}
1 & 1 \\
\end{array}
\right)
\left(
\begin{array}{cc}
1 & 1 \\
2 & 4 \\
\end{array}
\right)^{n}
\left(
\begin{array}{c}
1 \\
2 \\
\end{array}
\right)$, which is also equal to
\noindent $3\sum_{\frac{n}{2}\leq m\leq n}{m\choose n-m}5^{2m-n}(-2)^{n-m}-2\sum_{\frac{n-1}{2}\leq m\leq n-1}{m\choose n-m-1}5^{2m-n+1}(-2)^{n-m-1}$.
\end{theorem}
Let $T_n$ be the triangulation of the $n$-gon on vertex set $\{1,2,\dots,n\}$ such that all its interior edges are incident with vertex $1$. Note that $T_n$ is a planar graph. Its dual planar graph $T_n^*$ can be constructed from the {\em binary caterpillar tree} $C_n$ by gluing all leaves into one vertex, where $C_n$ is a tree such that all its interior vertices have degree three and are on a path. We assume the leaves of $C_n$ are labelled by $1,2,\dots,n$ from left to right. See Fig.~\ref{fig:dual} for an example of $T_5, T_5^*$ and $C_5$.
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[v/.style={circle,draw=black,fill=black, inner sep=0.6pt}]
\coordinate (r) at (-0.5,0);
\foreach \i in {-2,-1,0,1}
{
\coordinate (v\i) at (\i,-1);
\draw[-] (r) to (v\i);
}
\draw[-] (v1) to (v-2);
\coordinate (t) at (-0.5,-1.8);
\coordinate (u1) at (-1.2, -0.7);
\coordinate (u2) at (-0.5,-0.7);
\coordinate (u3) at (0.2,-0.7);
\draw[-,color=red] (t) to (u1) to (u2) to (u3) to (t) to (u2);
\draw[-,color=red] (u1) to [out=150, in=150] (-2.3, -1) to (t);
\draw[-,color=red] (u3) to [out=30, in=30] (1.3, -1) to (t);
\draw[-] (3,-1) to (7,-1);
\draw[-] (5,-1) to (5,-1.7);
\draw[-] (6,-1) to (6,-1.7);
\draw[-] (4,-1) to (4,-1.7);
\end{tikzpicture}
\caption{$T_5, T^*_5$ and $C_5$.}
\label{fig:dual}
\end{figure}
As $T_n^*$ is obtained from $C_n$ by gluing its leaves, we can naturally identify partial orientations of $T_n^*$ with partial orientations of $C_n$. So we let $\mathcal V(C_n)$ be the set of partial orientations of $C_n$ that correspond to vectors of $T_n^*$, and also call elements in $\mathcal V(C_n)$ {\em vectors} of $C_n$. Then we have the following isomorphism
\[ (\mathcal L(T_n),\leq)\cong (\mathcal V(T_n^*),\leq)\cong (\mathcal V(C_n),\leq).\]
The following observation is obvious.
\begin{proposition}
A partial orientation $X$ of $C_n$ is a vector if and only if in $X$ every arc is on a directed path between two leaves of $C_n$.
\end{proposition}
Now we investigate the rank function of poset $(\mathcal V(C_n),\leq)$. For each partial orientation $X$ of $C_n$, let $\Gamma_X$ be the digraph on leaves of $C_n$ such that $\overrightarrow{ij}\in \Gamma_X$ if $i$ can reach $j$ through a directed path in $X$. Let $b_0(\Gamma_X)$ be the number of weakly connected components of $\Gamma_X$.
\begin{lemma}\label{lem:rank}
The rank of an element $X$ in $(\mathcal V(C_n),\leq)$ is $n-b_0(\Gamma_X)$.
\end{lemma}
\begin{proof}
Let $X'$ be the covector of $T_n^*$ that corresponds to covector $X$ of $C_n$. Let $H$ be the underlying graph of $X'$. Then we know that $rank(X)=rank(X')=g(H)$.
Suppose that $\Gamma_X$ have $k$ isolated vertices and $m$ weakly connected components with vertex sizes $c_1,\dots,c_m$ which are at least $2$. Then the underlying graph of $X$ is a disjoint union of $m$ subtrees $T_1,\dots, T_m$ of $C_n$. For each tree $T_i$, let $H_i$ be obtained from $T_i$ by gluing its leaves, then $g(H_i)=c_i-1$. Observe that $H$ can be obtained from $H_1,\dots, H_m$ by taking one vertex from each graph and then gluing these $m$ vertices into one vertex. So $g(H)=g(H_1)+\dots+g(H_m)=c_1+\dots+c_m-m=n-k-m=n-b_0(\Gamma_X)$.
\end{proof}
\begin{lemma} \label{lem:vertex}
The number of full vectors of $C_n$ is $2\cdot 3^{n-2}$ when $n\geq 2$.
\end{lemma}
\begin{proof}
The subgraph obtained from $C_n$ by deleting leaves $n$ and $n-1$ is isomorphic with $C_{n-1}$. The restriction of every full vector of $C_n$ on $C_{n-1}$ gives rise to a full vector of $C_{n-1}$. On the other hand, every full vector of $C_{n-1}$ can be extended to $3$ different full vectors of $C_n$. Also note that $C_2$ has $2$ full vectors. So the number of full vectors of $C_n$ is $2\cdot 3^{n-2}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:facevector}]
In the following, we will compute the vector $f$ basing on $Z_{T_n}$. Let $i$ be an integer in $\{0,1,\dots, n-1\}$. According to Proposition \ref{prop:anti}, $f_{n-1-i}$ is the number of covectors of $T_n$ of rank $i$, which is also the number of vectors of $C_n$ of rank $i$. Then by Lemma \ref{lem:rank}, we conclude that $f_{n-1-i}$ is the number of vectors $X$ of $C_n$ such that $b_0(\Gamma_X)=n-i$.
For each integer $m\leq n-i$, we enumerate the number of vectors $X$ of $C_n$ such that $b_0(\Gamma_X)=n-i$ and exactly $m$ connected components of $\Gamma_X$ are not isolated vertices. Let $L$ be the set of isolated vertices of $\Gamma_X$. Then $|L|=n-i-m$. Let $T$ be the subtree of $C_n$ such that leaves of $T$ are leaves of $C_n$ that are not in $L$, and let $T'$ be obtained from $T$ by contracting all degree $2$ vertices. Then $T'$ is isomorphic with $C_{i+m}$. Note that if $v$ is an interior vertex of $C_n$ that is adjacent with a leaf in $L$, let $uv$ and $wv$ be another two edges incident with $v$, then either $uv, wv\in X^0$ or $\overrightarrow{uv}, \overrightarrow{vw}\in X$ or $\overrightarrow{wv}, \overrightarrow{vu}\in X$. So we let $X'$ be the vector of $T'$ such that each edge $ab$ is oriented from $a$ to $b$ whenever $a$ can reach $b$ in $X$. Now $b_0(\Gamma_{X'})= m$ and $\Gamma_{X'}$ has no isolated vertices. So $X'^0$ must be $m-1$ interior edges of $T'$ such that no two of them are adjacent. Let $T'_1,\dots, T'_m$ be the connected components of the underlying graph of $X'$. We assume their leaf numbers are $c_1,\dots,c_m$ respectively. For each $i=1,\dots,m$, let $X'_i$ be the restriction of $X'$ on $T'_i$, then $X'_i$ is a full vector of $T'_i$.
From above analysis we see that $X$ is totally determined by data $L$, $X'^0$ and $X'_1,\dots,X'_{m}$. We have ${n\choose m+i}$ choice for $L$. To choose $X'^0$ from $T'$ is equivalent with choosing $m-1$ non-adjacent numbers from $\{1,2,\dots,i+m-3\}$, and the number of choices to this classical combinatorial problem is ${(i+m-3)-(m-1)+1 \choose m-1}={i-1\choose m-1}$. After $L$ and $X'^0$ are fixed, then $T'_1,\dots, T'_m$ are also fixed. Their leaf numbers satisfy $c_1+\dots +c_m=m+i$. By Lemma \ref{lem:vertex}, the number of choices of $X'_i$ is $2\cdot 3^{c_i-2}$ for each $i=1,\dots, m$. So the total number of choice of $X'_1,\dots,X'_{m}$ is $\prod_{i=1,\dots, m} 2\cdot 3^{c_i-2}=2^m 3^{c_1+\dots +c_m-2m}=2^m 3^{i-m}$.
To sum up, the number of vectors $X$ of $C_n$ such that $b_0(\Gamma_X)=n-i$ and exactly $m$ connected components of $\Gamma_X$ are not isolated vertices is
\[2^m 3^{i-m}{n\choose m+i}{i-1\choose m-1}.\]
Thus $f_{n-i-1}=\sum_{0\leq m\leq \min\{i,n-i\}}2^m 3^{i-m}{n\choose m+i}{i-1\choose m-1}$.
\end{proof}
To enumerate the total number of faces of the graphical zonotope of a triangulation of the $n$-gon, we find it more convenient to do recursion than applying the formula in Theorem \ref{thm:facevector}.
\begin{proof}[Proof of Theorem \ref{thm:totalnumber}]
Let $\mathcal V_n^0,\mathcal V_n^1,\mathcal V_n^{-1}$ be the sets of vectors of $C_n$ such that the pendent edge of leaf $n$ is not oriented, oriented to $n$, oriented from $n$ respectively. Let $\mathcal V_n$ be the set of all vectors of $C_n$.
Let $v$ be the interior vertex of $C_n$ which is adjacent with leaves $n$ and $n-1$. Let $u$ be another vertex of $C_n$ which is adjacent with $v$. Let $T$ be the subgraph of $C_n$ obtained by deleting leaves $n$ and $n-1$. Note that $T$ is isomorphic with $C_{n-1}$. Every vector $X$ of $C_n$ defines a vector $X'$ of $T$ which is the restriction of $X$ on $T$.
For each vector $X\in \mathcal V_n^0$, if $vn-1,vu\in X^0$ then $X'\in \mathcal V_{n-1}^0$; if $\overrightarrow{uv}, \overrightarrow{vn-1}\in X$ then $X'\in \mathcal V_{n-1}^+$; if $\overrightarrow{vu}, \overrightarrow{n-1v}\in X$ then $X'\in \mathcal V_{n-1}^-$. So we see that
\begin{equation}\label{eq:v0}
|\mathcal V^0_n| = |\mathcal V^0_{n-1}|+ |\mathcal V^+_{n-1}|+ |\mathcal V^-_{n-1}|=|\mathcal V_{n-1}|.
\end{equation}
Similar analysis for $\mathcal V_n^+$ and $\mathcal V_n^-$ shows that
\begin{equation}\label{eq:v+}
|\mathcal V^+_n| = |\mathcal V^0_{n-1}|+ 3|\mathcal V^+_{n-1}|+ |\mathcal V^-_{n-1}|
\end{equation}
and
\begin{equation}\label{eq:v-}
|\mathcal V^-_n| = |\mathcal V^0_{n-1}|+ |\mathcal V^+_{n-1}|+ 3|\mathcal V^-_{n-1}|.
\end{equation}
Denote $\mathcal V^*_n= \mathcal V^+_n\cup \mathcal V^-_n$. Then we have
\begin{equation*}
\left(
\begin{array}{c}
|\mathcal V^0_n| \\
|\mathcal V^*_n| \\
\end{array}
\right)
= \left(
\begin{array}{cc}
1 & 1 \\
2 & 4 \\
\end{array}
\right)
\left(
\begin{array}{c}
|\mathcal V^0_{n-1}| \\
|\mathcal V^*_{n-1}| \\
\end{array}
\right).
\end{equation*}
Note that $(|\mathcal V^0_2|,|\mathcal V^*_2|)=(1,2)$, so we have
\begin{equation*}
\left(
\begin{array}{c}
|\mathcal V^0_{n+2}| \\
|\mathcal V^*_{n+2}| \\
\end{array}
\right)
= \left(
\begin{array}{cc}
1 & 1 \\
2 & 4 \\
\end{array}
\right)^{n}
\left(
\begin{array}{c}
1 \\
2 \\
\end{array}
\right).
\end{equation*}
So $|\mathcal V_{n+2}|=\left(
\begin{array}{cc}
1 & 1 \\
\end{array}
\right)
\left(
\begin{array}{cc}
1 & 1 \\
2 & 4 \\
\end{array}
\right)^{n}
\left(
\begin{array}{c}
1 \\
2 \\
\end{array}
\right).$
If we take the summation of Eqs.\eqref{eq:v0}, \eqref{eq:v+} and \eqref{eq:v-}, we will also obtain the recursive formula
\[|\mathcal V_n|=5|\mathcal V_{n-1}|-2|\mathcal V_{n-2}|.\]
By method of generating function, we then get
\[|\mathcal V_{n+2}|=3\sum_{\frac{n}{2}\leq m\leq n}{m\choose n-m}5^{2m-n}(-2)^{n-m}-2\sum_{\frac{n-1}{2}\leq m\leq n-1}{m\choose n-m-1}5^{2m-n+1}(-2)^{n-m-1}.\]
\end{proof}
\begin{remark}
The sequence of total face number $|\mathcal V_{n}|$ is the sequence A052984 in OEIS \cite{oeis}. They are also the Kekul\'{e} numbers for certain benzenoids \cite[see p. 78]{cyvin13}.
\end{remark}
\begin{figure}[htbp]\label{fig:Lynx}
\centering
\begin{tabular}{|l|llllllll|l|}
\hline
\diagbox[dir=SE]{$n$}{$i$} & 0 & 1& 2& 3& 4& 5 & 6 & 7 & Total\\
\hline
2 & 2 & 1 & & & & & & & 3\\
3 & 6 & 6 & 1 & & & & & & 13\\
4 & 18 & 28 & 12 & 1 & & & & & 59\\
5 & 54 & 114 & 80 & 20 & 1 & & & & 269\\
6 & 162 & 432 & 422 & 180 & 30 & 1 & & & 1227\\
7 & 486 & 1566 & 1962 & 1190 & 350 & 42 & 1 & & 5597\\
8 & 1458 & 5508 & 8424 & 6640 & 2828 & 616 & 56 & 1 & 25531\\
\hline
\end{tabular}
\caption{The face vectors of graphical zonotopes of triangulations of $n$-gon with $n=2, \ldots, 8$.}
\end{figure}
\bibliographystyle{plain}
|
\section{Introduction \& Related Work}
Any social interaction involves an exchange of viewpoints and thoughts. But these views and thoughts can be caustic. Often we see that users resort to verbal abuse to win an argument or overshadow someone's opinion. On Twitter, people from every sphere have experienced online abuse. Be it a famous celebrity with millions of followers or someone representing a marginalized community such as LGBTQ, Women and more. We want to channelize Natural Language Processing (NLP) for social good and aid in the process of flagging abusive tweets and users. Detecting abuse on Twitter can be challenging, particularly because the text is often noisy. Abuse can also have different facets. \cite{waseem2016hateful} released one of the initial data sets from Twitter with the goal of identifying what constitutes racism and sexism.
\cite{davidson} in their work pointed out that hate speech is different from offensive language and released a data set of 25k tweets with the goal of distinguishing hate speech from offensive language.
\begin{table}[H]
\begin{center}
\begin{tabular}{|l|l|}
\hline
\multicolumn{2}{|l|}{\begin{tabular}[c]{@{}l@{}}Stop saying dumb blondes with pretty\\ faces as you need a pretty face to pull \\ them off !!! \#mkr\end{tabular}} \\ \hline
\multicolumn{2}{|l|}{\begin{tabular}[c]{@{}l@{}}In Islam women must be locked in their \\ houses and Muslims claim this is treating\\ them well\end{tabular}} \\ \hline
\end{tabular}
\caption{Tweets from \cite{waseem2016hateful} data set demonstrating online abuse}
\end{center}
\end{table}
They find that racist and homophobic tweets are more likely to be classified as hate speech but sexist tweets are generally classified as offensive. \cite{golbeck} introduced a large, hand-coded corpus of online harassment data for studying the nature of harassing comments and the culture of trolling. Keeping these motivations in mind, we make the following salient contributions:
\begin{itemize}
\item We build a deep context-aware attention-based model for abusive behavior detection on Twitter . To the best of our knowledge ours is the first work that exploits context aware attention for this task.
\item Our model is robust and achieves consistent performance gains in all the three abusive data sets
\item We show how context aware attention helps in focusing on certain abusive keywords when used in specific context and improve the performance of abusive behavior detection .
\end{itemize}
\section{Related Work}
Existing approaches to abusive text detection can be broadly divided into two categories: 1) Feature intensive machine learning algorithms such as Logistic Regression (LR), Multilayer Perceptron (MLP) and etc. 2) Deep Learning models which learn feature representations on their own. \cite{waseem2016hateful} released the popular data set of 16k tweets annotated as belonging to sexism, racism or none class \footnote{http://github.com/zeerakw/hatespeech \label{footnote 1}}, and provided a feature engineered model for detection of abuse in their corpus. \cite{davidson} use a similar handcrafted feature engineered model to identify offensive language and distinguish it from hate speech. \cite{badjatiya2017deep} in their work, experiment with multiple deep learning architectures for the task of hate speech detection on Twitter using the same data set by \cite{waseem2016hateful}. Their best-reported F1-score is achieved using Long Short Term Memory Networks (LSTM) + Gradient Boosting.
On the data set released by \cite{waseem2016hateful}, \cite{park2017one} experiment with a two-step approach of detecting abusive language first and then classifying them into specific types i.e. racist, sexist or none. They achieve best results using a Hybrid Convolution Neural Network (CNN) with the intuition that character level input would counter the purposely or mistakenly misspelled words and made-up vocabularies. \cite{pavlopoulos2017deeper} in their work ran experiments on the Gazetta dataset and the DETOX system (\cite{wulczyn2017ex}) and show that a Recurrent Neural Network (RNN) coupled with deep, classification-specific attention outperforms the previous state of the art in abusive comment moderation. In their more recent work \cite{pavlopoulos2017improved} explored how user embeddings, user-type embeddings, and user type biases can improve their previous RNN based model on the Gazetta dataset. Attentive neural networks have been shown to perform well on a variety of NLP tasks (\cite{yang2016hierarchical}, \cite{yinattention}). \cite{yang2016hierarchical} use hierarchical contextual attention for text classification (i.e attention both at word and sentence level) on six large scale text classification tasks and demonstrate that the proposed architecture outperform previous methods by a substantial margin.We primarily focus on word level attention because most of the tweets are single sentence tweets.
\section{Model}
The best choice for modeling tweets was Long Short Term Memory Networks (LSTMs) because of their ability to capture long-term dependencies by introducing a gating mechanism that ensures the proper gradient propagation through the network. We use bidirectional LSTMs because of their inherent capability of capturing information from both: the past and the future states.
A bidirectional LSTM (BiLSTM) consists of a forward LSTM $\overrightarrow{f}$ that reads the sentence from $x_{1}$ to $x_{T}$ and a backward LSTM $\overleftarrow{f}$ that reads the sentence from $x_{T}$ to $x_{1}$, where T is the number of words in the sentence under consideration and $x_{i}$ is the $i_{th}$ word in the sentence. We obtain the final annotation for a given word $x_{i}$, by concatenating the annotations from both directions (Eq. [1]). \cite{speech} show that LSTMs can benefit from depth in space.Stacking multiple recurrent hidden layers on top of each other, just as feed forward layers are stacked in the conventional deep networks give performance gains .And hence we choose stacked LSTM for our experiments.
\begin{center}
\includegraphics[scale=0.50]{lstm.png}
\captionof{figure}{Architecture for Stacked BiLSTM + Word Level Contextual Attention. Figure is inspired by \cite{yang2016hierarchical}}
\end{center}
\subsection{Word Attention}
The attention mechanism assigns a weight to each word annotation that is obtained from the BiLSTM layer. We compute the fixed representation \textbf{\Large v} of the whole message as a weighted sum of all the word annotations which is then fed to a final fully-connected Softmax layer to obtain the class probabilities.
We first feed the LSTM output $h_{i}$ of each word $x_{i}$ through a Multi Layer Perceptron to get $u_{i}$ as its hidden representation. $u_{c}$ is our word level context vector that is randomly initialized and learned as we train our network. Once $u_{i}$ is obtained we calculate the importance of the word as the similarity of $u_{i}$ with $u_{c}$ and get a normalized importance weight $\alpha_{i}$ through a softmax function. The context vector $u_{c}$ can be seen as a tool which filters which word is more important over all the words like that used in the LSTM. Figure 2 shows the high-level architecture of this model. $W_{h}$ and $b_{h}$ are the attention layer’s weights and biases. More formally,
\begin{equation}
\begin{aligned}
h_{i} = [\overrightarrow{f}, \overleftarrow{f}]
\forall{i = 1,...T}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
u_{i} =tanh(W_{h}.h_{i} + b_{h})
\end{aligned}
\end{equation}
\begin{equation}
a_{i} = \frac{ u_{i}^{T}u_{c} }{ \sum_{j=1}^{T} u_{j}^{T}u_{c} }
\end{equation}
\begin{equation}
v = \sum_{i=1}^{T}a_{i}h_{i}
\end{equation}
\section{Experiments}
In this section we talk about data sets first and then go on to show our results obtained on these three data sets .We also show some examples where our model failed . Finally we show how attention helps us understand the model in a better fashion.
\subsection{Data Sets}
We have used the 3 benchmark data sets for abusive content detection on Twitter.
At the time of the experiment, the \cite{waseem2016hateful} data set had a total of \textbf{15,844} tweets out of which \textbf{1,924} were labelled as belonging to racism, \textbf{3,058} as sexism and \textbf{10,862} as none. The \cite{davidson} data set had a total of \textbf{25,112} tweets out of which \textbf{1498} were labelled as hate speech, \textbf{19,326} as offensive language and \textbf{4,288} as neither. For the \cite{golbeck} data set, there were \textbf{20,362} tweets out of which \textbf{5,235} were positive harassment examples and \textbf{15,127} were negative.
We call \cite{waseem2016hateful} data set as D1 , \cite{davidson} data set as D2 and \cite{golbeck} as D3
\begin{table}
\begin{center}
\begin{tabular}{||c |c ||}
\hline
\textbf{Data Set} & \textbf{Tweets Count} \\ [0.5ex]
\hline\hline
\cite{waseem2016hateful} & 15,844 \\
\hline
\cite{davidson} & 25,112\\
\hline
\cite{golbeck} & 20,362 \\
\hline
\end{tabular}
\caption{Data sets and their total tweets count}
\label{table:1}
\end{center}
\end{table}
For tweet tokenization, we use \textbf{Ekphrasis} which is a text processing tool built specially from social platforms such as Twitter.
\cite{baziotis2017datastories} use a big collection of Twitter messages (330M) to generate word embeddings, with a vocabulary size of 660K words, using GloVe (\cite{pennington2014glove}). We use these pre-trained word embeddings for initializing the first layer (embedding layer) of our neural networks.
\subsection{Results}
The network is trained at a learning rate of 0.001 for 10 epochs, with a dropout of 0.2 to prevent over-fitting. The results are averaged over 10-fold cross-validations for D1 and D3 and 5 fold cross-validations for D2 because \cite{davidson} reported results using 5 fold CV. Because of class imbalance in all our data sets, we report weighted F1 scores.
Table 3 shows our results in detail. We compare our model with the best models reported in each paper. Because \cite{golbeck} is a data set paper, we cannot fill the corresponding row. \textbf{*} denotes the numbers from baseline papers. All the results were reproducible except for the one marked red. For (Waseem and Hovy, 2016) data set, (Badjatiyaet al., 2017) claim that using Gradient Boosting with LSTM embeddings obtained from random word embeddings boosted their performance by 12 F1 from 81.0 to 93.0. When we tried to reproduce the result, we did not find any significant improvement over 81. Results show that our model is robust when it comes to the performance on all of the three data sets.
\begin{table}[H]
\begin{center}
\begin{tabular}{|p{2.2cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|}
\hline
{\textbf{Models}}& {\textbf{D1}} & {\textbf{D2}} & {\textbf{D3}} \\ \hline\hline
{\cite{waseem2016hateful}} & 73.8* & 82.3 & 63.0 \\ \hline
{\cite{davidson}} & 78.0 & 90.0* & 69.0 \\ \hline
{\cite{golbeck}}& - & - & - \\ \hline
{\cite{park2017one}} & 82.7* & 88.0 & 70.6 \\ \hline
{\cite{badjatiya2017deep}} & \textbf{\textcolor{red}{93.1}*} & 88.0 & 65.7 \\ \hline\hline
{\textbf{Our Model}} & 84.2 & \textbf{91.1} & \textbf{72.7} \\ \hline
\end{tabular}
\caption{Data sets and the results of different models. We reproduced the results for each model on three of the data sets.}
\end{center}
\end{table}
We also share some examples from the three data sets in Figure \ref{fig:fig2} which our BiLSTM attention model could not classify correctly. On closer investigation we find that most cases where our model fails are instances where annotation is either noisy or the difference between classes are very blurred and subtle.
\begin{center}
\includegraphics[scale=0.33]{analysis.png}
\captionof{figure}{ The first tweet is a tweet from \cite{waseem2016hateful}, the second tweet is a tweet from from \cite{davidson} data set and the third from the \cite{golbeck} datset }
\label{fig:fig2}
\end{center}
\subsection{Why Contextual Attention?}
Attention mechanism enables our neural network to focus on the relevant parts of the input more than the irrelevant parts while performing a
prediction task. But the relevance is often dependant on the context and so the importance of words is highly context dependent. For example, the word \textbf{islam} may appear in the realm of Racism as well as in any normal conversation.The top tweet in Figure \ref{fig:fig3} belongs to \textbf{None} class while the bottom tweet belongs to \textbf{Racism} class.
\begin{center}
\includegraphics[scale=0.20]{context-attention.png}
\captionof{figure}{An example showing how our model captures
diverse context and assigns context-dependent weights to the same word in two different tweets.}
\label{fig:fig3}
\end{center}
\subsection{Attention Heat Map Visualization}
The color intensity corresponds to the weight given to each word by the contextual attention.
\begin{center}
\includegraphics[scale=0.15]{attention_visualization.png}
\captionof{figure}{ The first tweet is a sexist tweet from \cite{waseem2016hateful} where as the second tweet is an example of racist tweet from the same datset . The third tweet is from \cite{davidson} data set labelled as offensive language.}
\label{fig:fig4}
\end{center}
\section{Conclusion and Future Work}
We successfully built a deep context-aware attention-based model and applied it to the task of abusive tweet detection. We ran experiments on three relevant data sets and empirically showed how our model is robust when it comes to detecting abuse on Twitter. We also show how context-aware attention helps us to interpret the model's performance by visualizing the attention weights and conducting thorough error analysis.\\
As for future work, we want to experiment with a model that learns user embeddings from their historical tweets. We also want to model abusive text classification in Twitter by taking tweets in context because often standalone tweets don't give a clear picture of a tweet's intent.
|
\section*{Introduction} \label{s.intro}
The primary decomposition of ideals in Noetherian rings is a fundamental result in commutative algebra and algebraic geometry.
From a minimal primary decomposition, one can define the set of the associated primes by taking the radical of each ideal in the decomposition. Our goal in this paper is to investigate the associated primes of powers of squarefree monomial ideals.
Squarefree monomial ideals and powers of ideals are central objects in combinatorial and commutative algebra and algebraic geometry for the several connections they encode between these areas, see for instance \cite{MS}.
There are several ways to relate a squarefree monomial ideal to a hypergraph. To serve our intent, we will associate to a hypergraph the squarefree monomial ideal with minimal primes corresponding to the edges of the hypergraph, and vice versa. This ideal is usually called the cover ideal of the hypergraph.
The associated primes of a squarefree monomial ideal are easy to describe, whereas computing the associated primes of a power could be really tricky. Currently, the set of the associated primes of a power of any squarefree monomial ideal is far from being fully understood.
Recently, in \cite{TMR} Kaiser, Stehl\' ik and \u Skrekovski produced an example of a squarefree monomial ideal, precisely the cover ideal of a graph, which fails the persistence property, i.e., the set of the associated primes could ``lose" some elements from a power to the next.
For ideals associated to a combinatorial object, one hopes to explain their behavior in terms of the original object. With this in mind, we define the shadow of a hypergraph, Definition \ref{d.shadow}, as a certain set of smaller hypergraphs related to the original one. We show that, see Theorem \ref{t.oddcycle}, Theorem \ref{t.tildeH}, the shadows preserve information about the associated primes of a power of the cover ideal of the hypergraph.
The paper is organized as follows.
In Section \ref{s.prelims}, we introduce the terminology and the basic results.
In Section \ref{s.intro shadow}, we define the shadows of a hypergraph that are the new tool introduced in this paper. Then we start an investigation of the associated primes of a squarefree monomial ideal in terms of the shadows of the associated hypergraph. In particular, in this section, we deal with the second power.
In Section \ref{s.main case}, under some restrictive conditions, we broaden our investigation to any power.
Finally, in Section \ref{s.application}, we apply the results of Section \ref{s.main case} to the persistence property.
\textbf{Acknowledgments.}
This project started during the PRAGMATIC 2017 Research
School ``Powers of ideals and ideals of powers" held in Catania, Italy. We would like to thank the University of Catania and the organizers of the workshop, Alfio Ragusa, Elena Guardo, Francesco Russo, and Giuseppe Zappal\`a. As well, we are deeply grateful to Adam Van Tuyl for introducing us to this topic and for his guidance and to Huy T\`ai H\`a for the useful comments. We also thank them together with Brian Harbourne and Enrico Carlini for their inspiring lectures. Computations were carried out with CoCoA \cite{COCOA} and Macaulay2 \cite{MACAULAY2}.
\section{Notation and basic facts} \label{s.prelims}
Let $V := \{x_1,\ldots,x_n\}$ and $R=K[V]=K[x_1,\ldots, x_n]$ be the standard polynomial ring in $n$ variables over a field $K$.
A squarefree monomial ideal $I\subseteq R$ always has a unique minimal primary decomposition, $I=\mathfrak p_1\cap \cdots \cap \mathfrak p_t$, as an intersection of squarefree prime ideals $\mathfrak p_i=(x_{i_1}, \ldots, x_{i_s})$. For more details and a full description of the topic we refer to Section 1.3 in \cite{HH}.
This property establishes a one-to-one correspondence between squarefree monomial ideals and finite simple hypergraphs.
First, recall that a (simple) {\bf hypergraph} $H$ is a pair $H=(V,E)$ where $V := \{x_1,\ldots,x_n\}$ is called the set of vertices of $H$ and $E$ is a collection of subsets of $V$. (We will only consider finite simple hypergraphs, these are also called $clutters$ in the literature).
We will denote, for a set $U=\{x_{i_1},\ldots,x_{i_s} \}\subseteq V$, by $$\mathfrak p_{U}:= (x_{i_1},\ldots,x_{i_s}) \subseteq R$$ the prime ideal generated by the variables in $U$ and by
$$x_U:=x_{i_1}\cdots x_{i_s} \in R$$
the monomial given by the product of the variables in U.
Then, a hypergraph $H=(V,E)$ unequivocally corresponds to the squarefree monomial ideal $J(H):=\bigcap\limits_{e\in E} \mathfrak p_e$, called the {\bf cover ideal} of $H$, and vice versa.
Let $H = (V,E)$ be a hypergraph. A subset $T$ of $V$ is a {\bf vertex cover} of $H$ if every edge $e\in E$ contains at least one element of $T$. A vertex cover $T$ is a {\bf minimal vertex cover} if no proper subset of $T$ is a vertex cover.
Minimal vertex covers are related to the minimal generators of $J(H)$. Indeed, $T$ is a minimal vertex cover of $H$ if and only if $x_T\in \mathcal G(J(H)),$ the set of monomials which minimally generates $J(H).$
See \cite{HVT} and \cite{HVT2} for a further investigation on cover ideals of hypergraphs.
In this paper we are interested in the study of the associated prime ideals of the (regular) powers of $J(H)$. Recall the following, classical, definition.
\begin{definition} \label{d.assocprime} Let $R$ be a ring and $I$ an ideal of $R$. A prime ideal $\mathfrak p \subset R$ is called an
{\bf associated prime ideal} of $I$ if there exists some element $m \in R/I$ such that $\mathfrak p=\Ann(m)$, the annihilator of $m$. The set of all associated prime ideals of $I$ is denoted by $\ass(I)$.
\end{definition}
By definition, the hypergraph $H=(V,E)$ easily provides a description of all elements in $\ass(J(H))$. Indeed, $\mathfrak p_U\in \ass (J(H))$ if and only if $U\in E$.
In order to describe the associated primes of the powers of $J(H)$ the next lemma is an essential tool. Recall that for an hypergraph $H = (V,E)$ and $U\subseteq V$ the {\bf induced subhypergraph} of $H$ on $U$ is the hypergraph $H_U=(U,E(U))$ where $E(U)=\{e\in E\ |\ e\subseteq U \}$.
\begin{lemma} [Lemma 2.11 \cite{FHVT}]\label{l.localization} Let $H=(V,E)$ be a hypergraph. Let $U\subseteq V$, then
\[
\mathfrak p_U \in \Ass(R/J(H)^s) \Leftrightarrow \mathfrak p_U \in \Ass(K[U]/J(H_U)^s).
\]
\end{lemma}
Lemma \ref{l.localization} moves the problem of looking for associated prime ideals to maximal ideals. Indeed, $\mathfrak p_U$ is associated to $J(H)^s$, if and only if it is associated to $J(H_U)^s\subseteq K[U]$, and $\mathfrak p_U$ is the maximal ideal in $K[U]$.
An immediate consequence of Lemma \ref{l.localization} is a first, well known, step in the description of the elements in $\ass J(H)^s.$
\begin{lemma}\label{l.edges}
Let $H$ be a hypergraph. Then $\mathfrak p_e\in \ass J(H)^s$ for each integer $s\ge 1$ and for each edge $e$ of $H$.
\end{lemma}
A corollary of Lemma \ref{l.localization} will be useful in Section \ref{s.main case}.
\begin{corollary}\label{c.prime of subhypergraph} Let $H=(V,E)$ be a hypergraph on $V$.
Let $F\subseteq U\subseteq V$, then
\[
\mathfrak p_F \in \Ass(R/J(H)^s) \Leftrightarrow \mathfrak p_F \in \Ass(K[U]/J(H_U)^s).
\]
\end{corollary}
\begin{proof}
It is an immediate consequence of Lemma \ref{l.localization}. Indeed, since $F\subseteq U$ we get $(H_U)_F=H_F.$
\end{proof}
In the literature there are only few other results explicitly describing the elements in $\ass (J(H)^s).$ Most of them deal with the case that $H$ is a graph, i.e., the edges all have cardinality 2. If $H$ is a graph we will often denote it by the letter $G$.
For instance, see proposition below, the authors of \cite{FHVT2} describe the set $\Ass(R/J(G)^2)$. They prove that the new primes match the (minimal) odd cycles of $G.$
Recall that in a graph $G=(V,E)$ a set of distinct vertices $C =\{ x_{i_1}, x_{i_2}, \ldots, x_{i_n}\}\subseteq V$ is called an\textit{ $n$-cycle} (or cycle of length $n$) if $\{x_{i_j}, x_{i_{j+1}}\}\in E$ for each $j \in \{1,\ldots,n\}$ and $x_{i_{n+1}} := x_{i_1}$.
$C$ is called an odd (even) cycle if $n$ is odd (even).
The vertices $x_{i_j}$, $x_{i_{j+1}}$ connected by an edge $\{x_{i_j}, x_{i_{j+1}}\}$ are called {\it adjacent} vertices in $C$. A {\it chord} of $C$ is an edge of $G$ joining two nonadjacent vertices. If $C$ has no chord, we shall call it {\it chordless}.
\begin{proposition}[Corollary 3.4, \cite{FHVT2}]\label{Corollary 3.5 cite{FHVT2}}
Let $G$ be a finite graph.
A prime ideal $\mathfrak p=(x_{i_1},\dots,x_{i_s})$ is in $\ass(J(G)^2)$ if and only if:
\begin{itemize}
\item[(a)] $s=2$ and $\mathfrak p\in \ass(J(G))$; or
\item[(b)] $s$ is odd, and after re-indexing, $\{x_{i_1}, x_{i_2}, \ldots, x_{i_s}\}$ is a chordless cycle of $G$.
\end{itemize}
\end{proposition}
\section{Introducing the shadows}\label{s.intro shadow}
The authors of \cite{FHVT} give a description of the set $\ass(J(H)^s)$ in terms of the coloring properties of the hypergraph $H$. However, their method is not very efficient to list all the elements in $\ass(J(H)^s)$ for any given hypergraph $H$. In this section, we introduce a tool that can be useful for this aim: we define the shadows of a hypergraph. The motivating idea is to take information from some other hypergraphs, smaller than $H$, and to bring it to $H$.
The following is the definition of a shadow of $H.$
\begin{definition}\label{d.shadow}
Let $H=(V,E)$ be a hypergraph. We say that a hypergraph $H'=(V', E')$ is a \textit{shadow} of $H$ if
\begin{itemize}
\item[(a)] $V'\subseteq V$; and
\item[(b)] $|E|=|E'|$ (same cardinalities) and $e\cap V' \in E'$ for each $e\in E.$
\end{itemize}
\end{definition}
We denote by $\mathcal{S}(H)$ the set of all the shadows of $H.$ Note that two different elements in $\mathcal{S}(H)$ have different vertex sets. Thus $H'=(V',E')\in \mathcal{S}(H)$ will be also called the shadow of $H$ on $V'$. By definition,
$H$ is always a shadow of itself on the vertex set $V$; we refer to this as the \textit{trivial} shadow. However, not every subset of $V$ produces a shadow of $H$, as we show in the following example.
\begin{example}\label{e.triangles} Consider the hypergraph $H$ on the vertex set $V= \{x_1,x_2,x_3,x_4,x_5\}$ with the edge set $E=\left\{ \{x_1,x_2,x_3\}, \{x_2,x_3,x_4\},\{x_1,x_4,x_5\} \right\}$.
The set $\mathcal S(H)$ contains non-trivial elements, namely, shadows on the vertex sets $V_1:=\{x_1,x_2,x_4\}$, $V_2:=\{x_1,x_3,x_4\},$ $V_3:=\{x_1,x_2, x_3,x_4\}$, $V_4:=\{x_1,x_3,x_4,x_5\}$ and $V_5:=\{x_1,x_2,x_4,x_5\}.$ Indeed, we have
\begin{align*}
\left(V_1, \{ \{x_1,x_2\}, \{x_2,x_4\},\{x_1,x_4\} \}\right)&\in \mathcal S(H), \text{ and}\\
\left( V_2,\{ \{x_1,x_3\}, \{x_3,x_4\},\{x_1,x_4\} \}\right)&\in \mathcal S(H).
\end{align*}
Both of these shadows are graphs, more precisely they are 3-cycles. Additionally,
\begin{align*}
&\left( V_3, \{ \{x_1,x_2, x_3\}, \{x_2, x_3,x_4\},\{x_1,x_4\} \} \right)\in \mathcal S(H), \\
&\left( V_4, \{ \{x_1,x_3\}, \{x_3,x_4\},\{x_1,x_4,x_5\} \} \right)\in \mathcal S(H) \text{ and }\\
& \left( V_5, \{ \{x_1,x_2\}, \{x_2,x_4\},\{x_1,x_4,x_5\} \} \right)\in \mathcal S(H).
\end{align*}
Furthermore,
for instance, $H$ has no shadow on the set $V_6:=\{x_1,x_2,x_3\}$ since we get
$$\left( V_6, \{ \{x_1,x_2,x_3\}, \{x_2,x_3\},\{x_1\} \}\right),$$
and this fails to be a simple hypergraph.
\end{example}
The hypergraph $H$ and its shadows are showed in the following figures
\begin{center}
\begin{tikzpicture}[scale=0.7]
\node (v1) at (0,0) {};
\node (v2) at (2.5,0) {};
\node (v3) at (1.25,2.17) {};
\node (m1) at (1.88,1.08)[right]{$x_3$};
\node (m2) at (1.25,0) [below]{$x_3$};
\node (m3) at (0.63,1.08) [left]{$x_5$};
\node (f1) at (1.25,-1) [below]{Hypergraph $H$};
\draw [thick] (v1)--(v2);
\draw [thick] (v2)--(v3);
\draw [thick] (v3)--(v1);
\fill (v1) circle (0.1) node [below left] {$x_4$};
\fill (v2) circle (0.1) node [below right] {$x_2$};
\fill (v3) circle (0.1) node [above] {$x_1$};
\node (v4) at (6,0) {};
\node (v5) at (8.5,0) {};
\node (v6) at (7.25,2.17) {};
\node (f2) at (7.25,-1) [below]{Shadow on $V_1$};
\draw [thick] (v4)--(v5);
\draw [thick] (v5)--(v6);
\draw [thick] (v6)--(v4);
\fill (v4) circle (0.1) node [below left] {$x_4$};
\fill (v5) circle (0.1) node [below right] {$x_2$};
\fill (v6) circle (0.1) node [above] {$x_1$};
\node (a1) at (12,0) {};
\node (a2) at (14.5,0) {};
\node (a3) at (13.25,2.17) {};
\node (m1) at (13.88,1.08)[right]{$x_3$};
\node (m2) at (13.25,0) [below]{$x_3$};
\node (f4) at (13.25,-1) [below]{Shadow on $V_3$};
\draw [thick] (a1)--(a2);
\draw [thick] (a2)--(a3);
\draw [thick] (a3)--(a1);
\fill (a1) circle (0.1) node [below left] {$x_4$};
\fill (a2) circle (0.1) node [below right] {$x_2$};
\fill (a3) circle (0.1) node [above] {$x_1$};
\node (v7) at (18,0) {};
\node (v8) at (20.5,0) {};
\node (v9) at (19.25,2.17) {};
\node (m3) at (18.63,1.08) [left]{$x_5$};
\node (f3) at (19.25,-1) [below]{Shadow on $V_5$};
\draw [thick] (v7)--(v8);
\draw [thick] (v8)--(v9);
\draw [thick] (v9)--(v7);
\fill (v7) circle (0.1) node [below left] {$x_4$};
\fill (v8) circle (0.1) node [below right] {$x_2$};
\fill (v9) circle (0.1) node [above] {$x_1$};
\end{tikzpicture}
\end{center}
where the edge $\{a,b,v_1,\ldots, v_m\}$ is depicted as the segment
\begin{tikzpicture}[scale=0.7]
\node (v1) at (0,0) {};
\node (v2) at (2.5,0) {};
\node (m2) at (1.25,0) [above]{$v_1 \ldots v_m$};
\fill (v1) circle (0.1) node [left] {$a$};
\fill (v2) circle (0.1) node [right] {$b$};
\draw [thick] (v1)--(v2);
\end{tikzpicture}. Similarly, we can re-picture the hypergraph $H$ in the following form to better see the shadows on $V_2$ and $V_4$:
\begin{center}
\begin{tikzpicture}[scale=0.7]
\node (v1) at (0,0) {};
\node (v2) at (2.5,0) {};
\node (v3) at (1.25,2.17) {};
\node (m1) at (1.88,1.08)[right]{$x_2$};
\node (m2) at (1.25,0) [below]{$x_2$};
\node (m3) at (0.63,1.08) [left]{$x_5$};
\node (f1) at (1.25,-1) [below]{Hypergraph $H$};
\draw [thick] (v1)--(v2);
\draw [thick] (v2)--(v3);
\draw [thick] (v3)--(v1);
\fill (v1) circle (0.1) node [below left] {$x_4$};
\fill (v2) circle (0.1) node [below right] {$x_3$};
\fill (v3) circle (0.1) node [above] {$x_1$};
\node (v4) at (6,0) {};
\node (v5) at (8.5,0) {};
\node (v6) at (7.25,2.17) {};
\node (f2) at (7.25,-1) [below]{Shadow on $V_2$};
\draw [thick] (v4)--(v5);
\draw [thick] (v5)--(v6);
\draw [thick] (v6)--(v4);
\fill (v4) circle (0.1) node [below left] {$x_4$};
\fill (v5) circle (0.1) node [below right] {$x_3$};
\fill (v6) circle (0.1) node [above] {$x_1$};
\node (v7) at (12,0) {};
\node (v8) at (14.5,0) {};
\node (v9) at (13.25,2.17) {};
\node (m3) at (12.63,1.08) [left]{$x_5$};
\node (f3) at (13.25,-1) [below]{Shadow on $V_4$};
\draw [thick] (v7)--(v8);
\draw [thick] (v8)--(v9);
\draw [thick] (v9)--(v7);
\fill (v7) circle (0.1) node [below left] {$x_4$};
\fill (v8) circle (0.1) node [below right] {$x_3$};
\fill (v9) circle (0.1) node [above] {$x_1$};
\end{tikzpicture}
\end{center}
In the following example, we show a hypergraph which only has trivial shadow.
\begin{example}
Let $H$ be the hypergraph on the vertex set $V= \{x_1,x_2,x_3,x_4,x_5\}$ and the edge set $E=\left\{ \{x_1,x_2,x_3\}, \{x_2,x_3,x_4\},\{x_3,x_4,x_5\},\{x_4,x_5,x_1\},\{x_5,x_1,x_2\} \right\}$. In this case, the set $\mathcal{S}(H)$ has only one element, namely $H$. Indeed, notice that each edge of $H$ contains vertices with ``consecutive" indexes. Since any subset of $V$ with two elements is contained in some edge, then $H$ has no shadow on any set $V'\subsetneq V$. For instance, $H$ has no shadow on the subset $V'$ obtained from $V$ by removing $x_1$ since $\{x_4,x_5\} \subset \{x_3,x_4,x_5\}$.
\end{example}
$J(H')$ is an ideal of $K[V']$ and there is a natural inclusion from $K[V']$ into $K[V]$. The ideal generated by the image of $J(H')$ under this map, i.e., the ideal generated by $\mathcal G(J(H'))\subseteq K[V]$, is called cone ideal of $J(H')$ in $K[V].$
The next lemma provides a connection between the monomial generators of $J(H)$ and $J(H')$ for a shadow $H'$ of $H$.
\begin{lemma} \label{l.gens Shad are gens Hyper}
Let $H=(V,E)$ be a hypergraph and $H'=(V',E')\in \mathcal{S}(H)$ a shadow of $H$. Then $\mathcal G(J(H'))\subseteq \mathcal G(J(H))$.
\end{lemma}
\begin{proof}
The ideal $J(H')$ is generated by monomials $x_U$ where $U$ is a minimal vertex cover of $H'$. By the definition of shadow, $U$ is also a minimal vertex cover of $H$, and $U$ does not involve the variables in $V\setminus V'.$
\end{proof}
\begin{remark}\label{r.J(H') subset J(H)}
From Lemma \ref{l.gens Shad are gens Hyper}, we have $J(H')=K[V']\cap J(H)$. Thus, each element $m$ in $J(H')$ also belongs to $J(H)$.
\end{remark}
As a consequence of Lemma \ref{l.gens Shad are gens Hyper}, we get the following result.
\begin{lemma}\label{l.m not in J(H)^s}
If $(J(H')^s:m)=\mathfrak{p}\neq (1)$ for some prime ideal $\mathfrak{p}$, then $m \notin J(H)^s$.
\end{lemma}
\begin{proof}
Suppose $m \in J(H)^s$, then $m=m_1\cdots m_sM$ where the $m_i$'s are monomial minimal generators of $J(H)$. Since $m$ only contains the variables in $V'$, each $m_i$ will also have this property. That means, $m_i \in J(H')$ for all $i \in \{1,2,\ldots,s\}$. Hence $m \in J(H')^s$, which contradicts $(J(H')^s:m)\neq (1)$.
\end{proof}
The next results show the first evidences that our construction really serves our purpose. We strongly use the classification in Proposition \ref{Corollary 3.5 cite{FHVT2}} and assume the existence of a graph $G\in \mathcal{S}(H)$.
Then, we show that $J(H)^2$ only has associated primes inherited from $J(G)^2$.
The following lemma can be deduced from Corollary 3.4 in \cite{FHVT2}. We also include a proof for the convenience of the reader.
\begin{lemma}\label{l.oddcycle}Let $C_{2n+1}=(V,E)$ be an $(2n+1)$-cycle. Then $(J(C_{2n+1})^2:x_{V})=\mathfrak p_V$.
\end{lemma}
\begin{proof} A minimal cover of $C_{2n+1}$ involves exactly $n+1$ vertices. Then $J(C_{2n+1})^2$ is generated in degree $2n+2$ and $x_V\notin J(C_{2n+1})^2.$ Moreover, $x_1x_V=x_{\{1,2,4,\ldots, 2n \}}\cdot x_{\{1,3,5,\ldots, 2n+1 \}}\in J(C_{2n+1})^2$. Analogously, we get $x_ix_V\in J(C_{2n+1})^2 $ for each $x_i\in V.$
\end{proof}
\begin{theorem}\label{t.oddcycle}
Let $H=(V,E)$ be a hypergraph. If $G\in \mathcal{S}(H)$ is an odd cycle (i.e., $G=C_{2n+1}$ for some positive integer $n$), then $\mathfrak p_V\in \ass J(H)^2$.
\end{theorem}
\begin{proof}
Let $E = (e_1, \ldots, e_k)$. Since $G=(V',E')\in\mathcal S(H)$, the edges of $G$ are given by $\lbrace{e_1', \ldots, e_{k}'}\rbrace$ where $e_i' = e_i\cap V'$. By hypothesis, $G$ is an odd cycle, so $k=2n+1$ for some positive integer $n$.
Without loss of generality, we relabel the vertices of $G$ so that
\[ e_i'= \left\{ \begin{array}{ll}
\lbrace{x_i, x_{i+1}}\rbrace, & \mbox{if $1\leq i \leq 2n$},\\
\lbrace x_{2n+1},x_1\rbrace, & \mbox{if $i=2n+1$}.
\end{array} \right. \]
From Proposition \ref{Corollary 3.5 cite{FHVT2}}, we know that $\mathfrak p_{V'}\in \ass J(G)^2$, and Lemma \ref{l.oddcycle} we have $(J(G)^2:x_{V'})=\mathfrak p_V'$, where $x_{V'}=\prod_{i=1}^{2n+1}x_i$.
Then, we claim that $(J(H)^2:x_{V'})=\mathfrak p_V$. If $x_j\in V'$, $x_j x_{V'} \in J(G)^2\subseteq J(H)^2$. So $x_j\in (J(H)^2:x_{V'})$. Moreover, if $y_j\in V\setminus V'$, then there exists an edge $e_i\in E$ such that $y_j \in e_i$.
Without loss of generality, one can assume that $i=1$.
Thus we have that
\begin{align*}
y_j x_{V'} = y_j x_1x_2\cdots x_{2n+1} = (y_jx_3x_5\cdots x_{2n+1})(x_1x_2x_4\cdots x_{2n}).
\end{align*}
The right hand side of the above equality is in $J(H)^2$ since it is the product of two vertex covers of $H$. Thus, $y_j\in (J(H)^2: x_{V'})$. Finally, $x_{V'}\notin J(H)^2$ since $x_{V'}\notin J(H')^2.$
\end{proof}
\begin{example}
Let $H$ be the hypergraph in Example \ref{e.triangles}. Since, for instance, the shadow of $H$ on $\{x_1,x_2,x_4\}$ is an odd cycle, we can state that
$$\mathfrak p_V=(x_1,x_2,x_3,x_4,x_5)\in \ass(J(H)^2).$$
\end{example}
Now we show that Theorem \ref{t.oddcycle} works in a more general setting. We need some further notation.
Let $H=(V,E)$ be a hypergraph and let $G=(V',E')\in \mathcal S(H)$ be a graph. For a subset $U\subset V'$, we denote by
$$\widehat U:= \bigcup_{e\in E, e'\subseteq U} e \subseteq V.$$
\begin{corollary}\label{c.oddcyle1}
Let $H$ be a hypergraph and $H'$ a shadow of $H$. If $C_{2n+1}$ is an odd cycle that is a subhypergraph of $H'$,
then $\mathfrak p_{\widehat C_{2n+1}}\in \ass (J(H)^2)$.
\end{corollary}
\begin{proof}
Say $H'=(V',E')$. We take the subhypergraph $\tilde H:= H_{\widehat C_{2n+1}}$ of $H$ on the vertex set $\widehat C_{2n+1}.$
Notice that $\tilde H$ has a shadow on $C_{2n+1}$. That is the odd cycle $C_{2n+1}$.
Thus, from Proposition \ref{Corollary 3.5 cite{FHVT2}} and Theorem \ref{t.oddcycle}, $\mathfrak p_{\widehat C_{2n+1}}\in \ass (J(\tilde H)^2)$.
Moreover, from Lemma \ref{l.localization}, we have $\mathfrak p_{\widehat C_{2n+1}}\in \ass (J(H)^2)$.
\end{proof}
\begin{corollary}\label{c.oddcyle2}
Let $H$ be a hypergraph and $\tilde H$ a subhypergraph of $H$. If an odd cycle $C_{2n+1}\in \mathcal S(\tilde H)$, then $\mathfrak p_{\widehat C_{2n+1}}\in \ass ((J(H)^2)$.
\end{corollary}
\begin{example} \label{e.shadow} Let $H=(V,E)$ be the hypergraph with the vertex set $$V=\{x_1,x_2,x_3,x_4,x_5,x_6,x_7,x_8\}$$ and the edge set
$$E=\{ \{x_1,x_2,x_6\}, \{x_2,x_3,x_6\},\{x_3,x_4,x_8\},\{x_4,x_5,x_6\},\{x_1,x_5,x_7\} \}.$$
\begin{center}
\begin{tikzpicture}[scale=0.6]
\node (f1) at (-3.3,1.59) {$H \ =\ $};
\node (v4) at (0,0) {};
\node (v3) at (2,1) {};
\node (v2) at (1.67,3.27) {};
\node (v1) at (-0.54,3.58) {};
\node (v5) at (-1.57,1.59) {};
\node (m1) at (0.56,3.39)[above]{$x_6$};
\node (m2) at (1.83,2.11) [right]{$x_6$};
\node (m3) at (1,0.5) [below right]{$x_8$};
\node (m4) at (-0.78,0.8) [below left]{$x_6$};
\node (m5) at (-1.05,2.59) [above left]{$x_7$};
\draw [thick] (v1)--(v2);
\draw [thick] (v3)--(v2);
\draw [thick] (v3)--(v4);
\draw [thick] (v4)--(v5);
\draw [thick] (v1)--(v5);
\fill (v1) circle (0.1) node [above] {$x_1$};
\fill (v2) circle (0.1) node [right] {$x_2$};
\fill (v3) circle (0.1) node [right] {$x_3$};
\fill (v4) circle (0.1) node [below] {$x_4$};
\fill (v5) circle (0.1) node [left] {$x_5$};
\end{tikzpicture}
\end{center}
The shadow of $H$ on the vertex set $V'=\{x_1,x_2,x_3,x_4,x_5\}$
is
$$H'=(V', \{\{x_1,x_2\}, \{x_2,x_3\},\{x_3,x_4\},\{x_4,x_5\},\{x_1,x_5\}\})\in \mathcal S(H).$$
\begin{center}
\begin{tikzpicture}[scale=0.6]
\node (f1) at (-3.3,1.59) {$H' \ =\ $};
\node (v4) at (0,0) {};
\node (v3) at (2,1) {};
\node (v2) at (1.67,3.27) {};
\node (v1) at (-0.54,3.58) {};
\node (v5) at (-1.57,1.59) {};
\node (m1) at (0.56,3.39)[above]{ };
\node (m2) at (1.83,2.11) [right]{ };
\node (m3) at (1,0.5) [below right]{ };
\node (m4) at (-0.78,0.8) [below left]{ };
\node (m5) at (-1.05,2.59) [above left]{ };
\draw [thick] (v1)--(v2);
\draw [thick] (v3)--(v2);
\draw [thick] (v3)--(v4);
\draw [thick] (v4)--(v5);
\draw [thick] (v1)--(v5);
\fill (v1) circle (0.1) node [above] {$x_1$};
\fill (v2) circle (0.1) node [right] {$x_2$};
\fill (v3) circle (0.1) node [right] {$x_3$};
\fill (v4) circle (0.1) node [below] {$x_4$};
\fill (v5) circle (0.1) node [left] {$x_5$};
\end{tikzpicture}
\end{center}
We see that $H'$ is a graph, precisely it is an odd cycle of length 5.
By Theorem \ref{t.oddcycle}, we have that
$$\mathfrak p_V=(x_1,x_2,x_3,x_4,x_5,x_6,x_7,x_8)\in \ass(J(H)^2).$$
Now, we take the shadow of $H$ on the vertex set $V''=\{x_1,x_3,x_5,x_6,x_8\}$.
The shadow of $H$ on $V''$ is
$$H''=(V'', \left\{ \{x_1,x_6\},\{x_3,x_6\},\{x_3,x_8\},\{x_5,x_6\},\{x_1,x_5\} \right\})\in \mathcal S(H).$$
\begin{center}
\begin{tikzpicture}[scale=0.7]
\node (f1) at (-3.3,1.59) {$H'' \ =\ $};
\node (v4) at (0,0) {};
\node (v3) at (2,1) {};
\node (v2) at (1.67,3.27) {};
\node (v1) at (-0.54,3.58) {};
\node (v5) at (-1.57,1.59) {};
\draw [thick] (v1)--(v2);
\draw [thick] (v3)--(v2);
\draw [thick] (v3)--(v4);
\draw [thick] (v2)--(v5);
\draw [thick] (v1)--(v5);
\fill (v1) circle (0.1) node [above] {$x_1$};
\fill (v2) circle (0.1) node [right] {$x_6$};
\fill (v3) circle (0.1) node [right] {$x_3$};
\fill (v4) circle (0.1) node [below] {$x_8$};
\fill (v5) circle (0.1) node [left] {$x_5$};
\end{tikzpicture}
\end{center}
Note that $H''$ has a subhypergraph that is a cycle of length 3, $C_3=\{\{x_1,x_6\},\{x_5,x_6\},\{x_1,x_5\}\}.$ By Corollary \ref{c.oddcyle1}, this cycle produces an element in $\ass(J(H)^2).$ So, we get
$$\mathfrak p_{\widehat C_3}= (x_1,x_2,x_4,x_5,x_6,x_7)\in \ass (J(H)^2).$$
\end{example}
In the following example we show that condition $(b)$ in Definition \ref{d.shadow} is necessary for Theorem \ref{t.oddcycle}.
\begin{example}
Let $H=(V,E)$ be the hypergraph with the vertex set $$V=\{x_1,x_2,x_3,x_4,x_5,x_6,x_7,x_8,x_9\}$$ and the edge set
\begin{align*}
E=\{&\{x_1, x_2,x_6, x_8\}, \{x_2,x_3,x_8,x_6\},\{x_3,x_4,x_7,x_9\},\\
&\{x_4,x_5,x_6,x_8\},\{x_1,x_5,x_7,x_9\},\{x_8,x_9\}, \{x_6,x_7\} \}.
\end{align*}
A Macaulay2 computation \cite{MACAULAY2} shows that $$\ass (J(H)^2)= \{\mathfrak p_e \ |\ e\in E \}.$$
We claim that ignoring the rule $|E|=|E'|$ in the condition (b) of Definition \ref{d.shadow} Theorem \ref{t.oddcycle} does not hold. Indeed, we get on the vertex set $V'=\{x_1,x_2,x_3,x_4,x_5\}$ the hypergraph
$$H'=(V', \{\{x_1,x_2\}, \{x_2,x_3\},\{x_3,x_4\},\{x_4,x_5\},\{x_1,x_5\}\})\in \mathcal S(H).$$
That is an odd cycle and, see Lemma \ref{l.oddcycle}, we have
$$\mathfrak p_{V'}=(x_1,x_2,x_3,x_4,x_5)\in \ass(J(H')^2).$$
\end{example}
In the last part of this section we prove that, under some suitable hypothesis, all the associated primes of $J(H)^2$ come from some non-trivial shadow (we will see in Proposition \ref{p.associated square}). We need an auxiliary lemma.
\begin{lemma}\label{l.for square}
Let $H=(V,E)$ be a hypergraph, and suppose $(J(H)^{s}: m) = \mathfrak p_{V}$ for some monomial $m$. Let $V'\subsetneq V$ be a proper subset such that $e_i\cap e_j\subseteq V'$ for each $e_i, e_j\in E$, $i\neq j$. Then $y^{s-1}$ does not divide $m$ for each $y\in V\setminus V'$.
\end{lemma}
\begin{proof}
Let $y$ be an element in $V\setminus V'$. We write $m=y^am'$, where, unless to rename, $y\in e_1$ and $y$ does not divide $m'.$
If $a\ge s$ then, since $ym\in J(H)^s,$ we get $ym=m_1 \cdots m_sM$, where $m_j$ corresponds to a minimal vertex cover of $H$ for $j \in \lbrace 1,\ldots, s\rbrace$. Thus
$y$ divides $M$ and $m=m_1 \cdots m_s(M/y).$ This contradicts $m\notin J(H)^s$.
Therefore, we can assume $a=s-1.$
We work by induction on $r=|e_1\setminus V'|$. If $r=1$, i.e., $ e_1=(e_1\cap V')\cup \{y\}$, then from $ym\in J(H)^s,$ we get $ym=(ym_1) \cdots (ym_s)M$, where $ym_j$ are minimal vertex covers of $H$.
Note that, for each $x_j\in e_1\cap V'$, we can see that $x_j$ does not divide $m_1,\ldots, m_s$ (these are minimal vertex covers) and $x_j$ does not divide $M$ (otherwise we can just delete $y$ and get $m\in J(H)^s$). This implies that $m\notin (\mathfrak p_{e_1})^s.$ To get a contradiction, we just take some $z\notin e_1$ and remember that by hypothesis $zm\in J(H)^s$ but $zm\notin (\mathfrak p_{e_1})^s$.
If $r>1$, i.e., $ e_1=(e_1\cap V')\cup \{y_1,\ldots, y_r\}$, then just note that $V''=V'\cup \{y_1,\ldots, \widehat{y_{i}},\ldots, y_{r}\}$ satisfies the hypothesis of the theorem and $ e_1=(e_1\cap V'')\cup \{y_i\}$.
\end{proof}
\begin{proposition}\label{p.associated square} Let $H=(V,E)$ be a hypergraph and $H'=(V',E')\in \mathcal S(H)$ a shadow of $H$. Assume that $e_i\cap e_j\subseteq V'$ for each $e_i, e_j\in E$, where $i\neq j.$ If $\mathfrak p_{V}\in \ass J(H)^{2}$, then $\mathfrak p_{V'}\in \ass J(H')^2.$
\end{proposition}
\begin{proof}
By the definition of associated primes, there exists a monomial $m\in K[V]$ such that $(J(H)^2:m)=\mathfrak p_V.$ Say $V'=\{x_1,\ldots,x_a \}$ and $V\setminus V'=\{y_1,\ldots,y_b \}$.
By Lemma \ref{l.for square} $y_j$ does not divide $m$ for $j=1,\ldots, b.$ Then $m\in K[V']$ and therefore $(J(H')^2:m)=\mathfrak p_{V'}.$
\end{proof}
\section{A first case}\label{s.main case}
In this section we investigate the relations between a hypergraph and its shadows in a particular case of study. Precisely, we consider shadows that only differ from the starting hypergraph by one edge and one vertex.
In this section, we shall use the following notation.
\begin{notation}\label{n.notation}
Let $H=(V,E)$ be a hypergraph and $H'=(X,E')$ a shadow of $H$ such that
\begin{itemize}
\item[(a)] $X=\{x_1,\ldots,x_n\}$ and $V=X\cup\{y\}$; and
\item[(b)] $y$ only belongs to one edge, say $e_y\in E$.
\end{itemize}
After renaming, say $e_y=\{x_1,\ldots,x_t,y\}$. We set $e:=e'_y= \{x_1,\ldots,x_t\}$, then we have $H'=\left\{X,(E\setminus \{e_y\})\cup \{e\} \right\}$.
Moreover, to shorten the notation, $\tilde H$ will denote the subhypergraph of $H$ on $X$.
We denote by $\mathfrak p_e$ and $\mathfrak p_{e_y}$ the prime ideals generated by the variables in $e$ and $e_y$ respectively.
\end{notation}
We remark that, in this setting, the hypergraphs $\tilde H$ and $H'$ share the vertex set $X$. Moreover, they share the same edges except for $e$.
We will abuse notation: given a subset $F\subseteq X\subseteq V$, we will write $\mathfrak{p}_F$ to denote both the ideals in $K[X]$ and in $K[V].$
Here, we anticipate the results of this section.
In the first part of the section, we investigate the relation linking associated primes of $J(\tilde H)^s$ and $J(H')^s$ with the elements in $\ass (J(H)^s)$. We have seen in Corollary \ref{c.prime of subhypergraph} that if $\mathfrak p\in \ass (J(\tilde H)^s)$ then $\mathfrak p\in \ass (J(H)^s).$ What about the associated prime of $J( H')^s$? We will show that if $\mathfrak p\in \ass (J( H')^s)$, then either $\mathfrak p+(y)\in \ass (J(H)^s)$ or $\mathfrak p\in \ass (J(H)^s)$. This depends on a further condition of a monomial $m$ such that $(J(H')^s: m)=\mathfrak p.$ The following diagram summarizes these results.
\begin{center}
\vspace*{-5pt}
\begin{tikzpicture}[scale=0.6]
\matrix [column sep=-5mm, row sep=7mm] {
\node (pHt) [draw, shape=rectangle] {$\mathfrak p\in \ass (J(\tilde H)^s) $}; & & \node (pHs) [draw, shape=rectangle] {$\mathfrak p\in \ass (J(H')^s )$};& \\
& & \node (m) [draw, shape=rectangle] {$\mathfrak p=(J(H')^s: m) $};& \\
& \node (mYes) [draw, shape=circle] {$\mathfrak p= (J(\tilde H)^s: m )$}; & & \node (mNo) [draw, shape=circle] {$\mathfrak p\neq (J(\tilde H)^s: m )$}; \\
\node (pH) [draw, shape=rectangle] {$\mathfrak p\in \ass (J(H)^s) $}; & & & \node (pyH) [draw, shape=rectangle] {$\mathfrak p+(y)\in \ass (J(H)^s )$}; \\
};
\draw[->, thick] (pHt) -- (pH);
\draw[->, thick] (pHs) -- (m);
\draw[->, thick] (m) -- (mYes);
\draw[->, thick] (m) -- (mNo);
\draw[->, thick] (mYes) -- (pHt);
\draw[->, thick] (mNo) -- (pyH);
\end{tikzpicture}
\vspace*{-5pt}
\end{center}
In the second part of the section, we will reverse the investigation. Starting from a prime associated to $J(H)^s$, we will look for which conditions allow us to find a relation with an element in $J(\tilde H)^s$ or $J(H')^s$. Precisely, if $\mathfrak p \in \ass (J(H)^s)$ and $y\notin \mathfrak p$ then $\mathfrak p\in \ass (J(\tilde H)^s)$. Moreover, if $\mathfrak p= (y)+\mathfrak p'$, it seems natural to ask if $\mathfrak p'\in \ass (J(H')^s),$ which we positively answer under an extra (restrictive) condition. In the next section, see \ref{e.persistence does not fail}, we will show that not all the primes $(y)+\mathfrak p'$ associated to $J(H)^s$ come from a prime $\mathfrak p'$ in the shadow.
\begin{center}
\begin{tikzpicture}
\matrix [column sep=-5mm, row sep=7mm] {
& \node (pH) [draw, shape=rectangle] {$\mathfrak p\in \ass J(H)^s $}; & \\
\node (NOy) [draw, shape=circle] {$y\notin \mathfrak p$}; & & \node (YESy) [draw, shape=circle] {$y\in \mathfrak p$}; \\
\node (pHt) [draw, shape=rectangle] {$\mathfrak p\in \ass J(\tilde H)^s $}; & &
\node (pTOq) [draw, shape=rectangle] {$\mathfrak p= \mathfrak p' +(y)$ }; \\
& & \node (qd) [draw, shape=circle] {?}; \\
& & \node (pHs) [draw, shape=rectangle] {$\mathfrak p'\in \ass J(H')^s $}; \\
};
\draw[->, thick] (pH) -- (NOy);
\draw[->, thick] (NOy) -- (pHt);
\draw[->, thick] (YESy)-- (pTOq);
\draw[->, thick] (pH) -- (YESy);
\draw[->, thick] (pTOq) -- (qd);
\draw[->, thick] (qd) -- (pHs);
\end{tikzpicture}
\end{center}
We start with an auxiliary result.
\begin{lemma}
Let $m \in \mathcal G(J(H))$ be a monomial minimal generator of $J(H)$. If $y|m$, then $x_i \not| m$ for all $x_i\in e$.
\end{lemma}
\begin{proof} In our setting, $y$ only belongs to the edge $e_y=\{x_1,\ldots,x_t,y\}$.
Since $m$ is a minimal vertex cover of $H$, if $x_i \in e= \{x_1,\ldots,x_t\}$ divides $m$, then $\dfrac{m}{y}$ is also a vertex cover. This contradicts the minimality of $m$.
\end{proof}
In order to relate the associated primes of $J(H')^s$ to the associated primes of $J(H)^s,$ the following proposition will be crucial.
\begin{proposition}\label{p.on associated to H'}
Let $(J(H')^s:m)=\mathfrak{p}_F$ be a prime ideal, for some $F\subseteq X.$ Then,
$$(J(H)^s:m)= \mathfrak p_F + \mathfrak q$$ where $\mathfrak q \subseteq (y).$ In other words, no monomial only involving the variables in $X\setminus F$ belongs to $(J(H)^s:m)$.
\end{proposition}
\begin{proof}
Say $F:=\{ x_{i_1}, \ldots, x_{i_k}\}$ and $\{x_{\ell_1},\ldots, x_{\ell_r}\}=X\setminus F.$ Recall that $e=\{x_1,\ldots, x_t\}.$
First we show that $\mathfrak{p}_F\subseteq(J(H)^s:m)\subsetneq (1).$
From Lemma \ref{l.m not in J(H)^s} we have $m \notin J(H)^s$ and then $(J(H)^s:m) \neq (1)$. By hypothesis, for each $x_j \in F$ we have $x_jm\in J(H')^s$ i.e. $m=m_1\cdots m_s M$ for some monomials $m_i\in J(H')\subseteq K[X]$. But these monomials, see Remark \ref{r.J(H') subset J(H)} also belongs to $J(H)$. Hence, $x_jm \in J(H)^s$ and $\mathfrak{p}_F \subseteq (J(H)^s:m)\subseteq K[V]$.
In order to conclude the proof, take any monomial $x^{a_1}_{\ell_1} \cdots x^{a_t}_{\ell_t}$ in variables in $X\setminus F$. Suppose that $x^{a_1}_{\ell_1} \cdots x^{a_t}_{\ell_t}m=m_1\cdots m_sM \in J(H)^s$, where the $m_j$'s are minimal generators of $J(H)$ in the variables in $X$. The monomials $m_j \in J(H')$ and then $x^{a_1}_{\ell_1} \cdots x^{a_t}_{\ell_t} \in (J(H')^s:m)=\mathfrak{p}_F$, which is a contradiction.
\end{proof}
\begin{lemma} \label{p.tildeH}
Let $(J(H)^s:m)= \mathfrak p$ and $y\notin \mathfrak p$. Then $(J(\tilde H)^s:m)= \mathfrak p$.
\end{lemma}
\begin{proof} Say $\mathfrak p=\mathfrak p_F$ for some $F\subseteq X.$ First note that $m\notin J(\tilde H)^s.$ Indeed, if $m=m_1\cdots m_s \cdot M\in J(\tilde H)^s$ with $m_1,\ldots, m_s$ minimal vertex covers of $\tilde H$, then $y^sm\in J(H)^s$. This contradicts $(J(H)^s:m)= \mathfrak p$. We claim that $(J(\tilde H)^s:m)\supseteq \mathfrak p.$ Indeed, if $x_j\in F$, then $x_jm\in J(H)^s\subseteq J(\tilde H)^s.$
In order to obtain the assertion, we take a monomial $T\notin \mathfrak p_F$ and assume that $Tm\in J(\tilde H)^s$. Again from $Tm=m_1\cdots m_s \cdot M\in J(\tilde H)^s$ with $m_1,\ldots, m_s$ minimal vertex covers of $\tilde H$, we get $Ty^s\in (J(H)^s:m)$ which contradicts the hypothesis.
\end{proof}
\begin{theorem}\label{t.tildeH} Let $(J(H')^s:m)= \mathfrak p$. Then, we have
\begin{itemize}
\item[(a)] $(J(H)^s:m) = \mathfrak p$ if and only if $(J(\tilde H)^s:m)= \mathfrak p$;
\item[(b)] $(J(H)^s:m\cdot m_0) = \mathfrak p+(y)$, for some monomial $m_0\notin \mathfrak p,$ if and only if $(J(\tilde H)^s:m)\neq \mathfrak p$
\end{itemize}
\end{theorem}
\begin{proof}
Note that $y\notin \mathfrak p$, so one implication in (a) follows from Lemma \ref{p.tildeH}.
Set $\mathfrak p_F:=\mathfrak p=(J(\tilde H)^s:m)$ and say $X\setminus F=\{x_{\ell_1},\ldots, x_{\ell_r}\}.$
By Proposition \ref{p.on associated to H'}, we have $(J(H)^s:m)=\mathfrak p+ \mathfrak q$ where either $\mathfrak q=(0)$ or $\mathfrak q$ is minimally generated by monomials $y^a\cdot x_{\ell_1}^{a_1}x_{\ell_2}^{a_2}\cdots x_{\ell_r}^{a_r}$ for some $a>0$ and $a_1,\ldots, a_r\ge 0$.
We claim that $\mathfrak q=(0)$. Indeed, if $T:=y^a\cdot x_{\ell_1}^{a_1}x_{\ell_2}^{a_2}\cdots x_{\ell_r}^{a_r}\in \mathfrak q$, we get $\dfrac{T}{y^a}\in (J(\tilde H)^s:m)=\mathfrak p_F$ which contradicts the hypothesis.
Now we prove item (b). With the notation as above, we have $(J(H)^s:m)=\mathfrak p+ \mathfrak q$.
First assume $(J(\tilde H)^s:m)\neq \mathfrak p$. Then $\mathfrak q$ is not the zero ideal. Consider the non-empty set $$\{b\in \mathbb{N} \mid y^b\ \text{ divides }\ M\ \text{ for some }\ M\in \mathfrak q\},$$
and let $a$ be its minimum element.
Let $T:=y^a\cdot x_{\ell_1}^{a_1}x_{\ell_2}^{a_2}\cdots x_{\ell_r}^{a_r}\in \mathfrak q$ be a monomial minimal generator in $\mathfrak q$.
We collect some relevant facts:
\begin{itemize}
\item $a>0,$ by Proposition \ref{p.on associated to H'};
\item $m\dfrac{T}{y}\notin J(H)^s$, by the minimality of $T;$
\item $x_{\ell_1}^{a_1}x_{\ell_2}^{a_2}\cdots x_{\ell_r}^{a_r}\cdot m\dfrac{T}{y}\notin J(H)^s$, by the minimality of $a$;
\item $y \cdot m\dfrac{T}{y}=mT \in J(H)^s$.
\end{itemize}
Then, we get
$\left(J(H)^s: m\dfrac{T}{y}\right)=\mathfrak p+ (y),$ and $\mathfrak p+(y)\in \ass (J(H)^s).$
Vice versa, assume $(J(H)^s:m\cdot m_0) = \mathfrak p+(y)$, for some monomial $m_0\notin \mathfrak p.$ So, we have $ymm_0\in J(H)^s$ and say $ymm_0=ym_1\cdot m_2\cdots m_s\cdot M\in J(H)^s$ with $ym_1,\ldots, m_s$ corresponding to minimal vertex covers of $H$.
Then, we get $mm_0=m_1\cdot m_2\cdots m_s\cdot M\in J(\tilde H)^s$, i.e., $m_0 \in (J(\tilde H)^s:m)$. Since $m_0$ does not involve the variables in $\mathfrak p$, we get a contradiction.
\end{proof}
In particular, the next result shows that item (a) in Theorem \ref{t.tildeH} is always satisfied if $\mathfrak p_e\not\subseteq \mathfrak p$.
\begin{proposition}\label{p.shadow1} Let $(J(H')^s:m)= \mathfrak p$. If $\mathfrak p_e \not\subseteq \mathfrak p$, then $(J(H)^s:m)= \mathfrak p$.
\end{proposition}
\begin{proof}
Say $\mathfrak p= \mathfrak p_F$ with $F:=\{ x_{i_1}, \ldots, x_{i_k}\}$ and $\{x_{\ell_1},\ldots, x_{\ell_r}\}=X\setminus F.$
By Proposition \ref{p.on associated to H'} we have $(J(H):m)=\mathfrak p+ \mathfrak q$ where $\mathfrak q$ is an ideal minimally generated by monomials which are not only in variables $\{x_{\ell_1},\ldots, x_{\ell_r}\}= X\setminus F$; i.e., a minimal generator of $\mathfrak q$ is a monomial $y^bx_{\ell_1}^{a_1}\cdots x_{\ell_r}^{a_r}$ for some $a_1, \ldots, a_r\ge 0$ and $b>0$.
Assume on the contrary that $\mathfrak q\neq 0$.
Take any minimal generator in $\mathfrak q$, say $T:=y^bx_{\ell_1}^{a_1}\cdots x_{\ell_r}^{a_r}$.
Then $m\cdot T= m_1\cdots m_s\cdot M\in J(H)^s $ where the $m_i$'s are minimal vertex covers of $H.$ Note that $y$ does not divide $M$. Otherwise, we get $x_{\ell_1}^{a_1}x_{\ell_2}^{a_2}\cdots x_{\ell_r}^{a_r}y^{b-1}\in (J(H)^s:m)$, contradicting the minimality of $T.$
Then we can write, after relabeling, $m_{i}=ym_{i}'$ for $i=1, \ldots b,$ $m\cdot T= (ym_1')\cdots(ym_b')\cdot m_{b+1} \cdots m_s\cdot M\in J(H)^s.$
Say $x_1\in \mathfrak p_e$ and $x_1\notin \mathfrak p$, then we get $$m\cdot T\dfrac{x_1^b}{y^b}= (x_1m_1')\cdots(x_1m_b')\cdot m_{b+1} \cdots m_s\cdot M\in J(H)^s.$$
But $m\cdot T\dfrac{x_1^b}{y^b}$ only contains variables of $X$. Then $T\dfrac{x_1^b}{y^b}\in (J(H'):m)= \mathfrak p.$ By Proposition \ref{p.on associated to H'}, this is a contradiction since $T\dfrac{x_1^b}{y^b}$ only contains variables not in $\mathfrak p.$
\end{proof}
Recall that by Corollary \ref{c.prime of subhypergraph}, a prime associated to $J(H)^s$ either belongs to $\ass (J(\tilde H)^s)$ or it contains the variable $y$. This is summarized in the following statement.
\begin{corollary}\label{c.dec ass J(H)^s} We have
$$\ass (J(H)^s) = \ass (J(\tilde H)^s) \cup \mathcal A,$$ where if $\mathfrak p\in \mathcal A$, then $y\in \mathfrak p.$
\end{corollary}
\begin{question}\label{q.shadow}
Do the elements in $\mathcal A$, mentioned in Corollary \ref{c.dec ass J(H)^s}, all come from the shadow? More precisely, if $\mathfrak p=\mathfrak p'+(y)\in \ass J(H)^s$, then is $\mathfrak p'\in \ass J(H')^s$?
\end{question}
We will show in the next section, see Example \ref{e.persistence does not fail}, that such question has in general a negative answer.
By the way, in the next theorem, we positively answer this question under a suitable condition.
\begin{theorem}\label{t.from H to H'}
Let $\mathfrak p=\mathfrak p'+(y)\in \ass (J(H)^s)$. If $\mathfrak p\notin \ass (J(H)^s:y)$, then $\mathfrak p'\in \ass (J(H')^s)$.
\end{theorem}
\begin{proof}
Take the short exact sequence
$$ 0\to \dfrac{K[V]}{J(H)^s:y} \to \dfrac{K[V]}{J(H)^s} \to \dfrac{K[V]}{J(H)^s+(y)}\to 0.$$
From theorem 6.3 in \cite{Matsumura} we have that
$$\Ass (K[V]/J(H)^s)\subseteq \Ass (K[V]/J(H)^s:y) \bigcup \Ass (K[V]/J(H)^s+(y)).$$
Denoted by $J'$ the cone ideal of $J(H')^s$ in the ring $K[V]$, note that $K[V]/J(H)^s+(y)=K[V]/J'+(y).$
Since, by hypothesis $\mathfrak p\in \Ass (K[V]/J(H)^s+(y))$, then $\mathfrak p \in \ass (J' +(y))$, i.e. $\mathfrak p' \in \ass (J(H')^s)$.
\end{proof}
\begin{remark}
Question \ref{q.shadow} has a positive answer if $(J(H)^s:m)=\mathfrak p+(y)$ for some $m\in K[X].$
\end{remark}
In the next examples we show how to describe all the associated prime ideals of $J(H)^s$ from $\ass (J(\tilde H)^s)$ and $\ass (J(H')^s).$
\begin{example}
Let $H$ be the hypergraph on the vertex set $V=\{x_1,x_2,x_3,x_4,x_5, y\}$ and the edge set
$$E=\{\{x_1,x_2\}, \{x_2,x_3\}, \{x_3,x_4\}, \{x_4,x_5\}, \{x_1,x_5\}, \{x_1,x_3,y\} \}.$$
Set $X:=\{x_1,x_2,x_3,x_4,x_5\}.$
Then the shadow of $H$ on $X:=\{x_1,x_2,x_3,x_4,x_5\}$ is
$$H' =( X, \{\{x_1,x_2\}, \{x_2,x_3\}, \{x_3,x_4\}, \{x_4,x_5\}, \{x_1,x_5\}, \{x_1,x_3\} \} ).$$
Moreover, the subhypergraph of $H$ on $X$ is
$$\tilde H =(X, \{\{x_1,x_2\}, \{x_2,x_3\}, \{x_3,x_4\}, \{x_4,x_5\}, \{x_1,x_5\} \}).$$
A Macaulay2 computation shows that
$$\ass J(H')^3=\{(x_1,x_2), (x_2,x_3), (x_3,x_4), (x_4,x_5), (x_1,x_5), (x_1,x_3) \}\cup \{(x_1,x_2,x_3)\}$$
and
$$\ass J(\tilde H)^3=\{(x_1,x_2), (x_2,x_3), (x_3,x_4), (x_4,x_5), (x_1,x_5)\}\cup \{(x_1,x_2,x_3,x_4,x_5)\}. $$
From Theorem \ref{t.tildeH}, we know that $(x_1,x_2,x_3,y)\in \ass (J(H)^3)$.
Moreover, one can check that
$$\ass (J(H)^3)= \ass (J(\tilde H)^3)\cup \{(x_1,x_3,y),(x_1,x_2,x_3,y)\}.$$
\end{example}
\begin{example}
Let $H$ be the hypergraph on the vertex set $V=\{x_1,x_2,x_3,x_4,x_5,y\}$ given by
$$H=(V,\{\{x_1,x_2,x_3\}, \{x_1,x_4\}, \{x_2,x_4\}, \{x_2,x_5\}, \{x_3,x_5\}, \{x_4,x_5,y\} \}).$$
Set $X:=\{x_1,x_2,x_3,x_4,x_5\}$; then the shadow of $H$ on $X$ is
$$H' =\big( X, \{\{x_1,x_2,x_3\}, \{x_1,x_4\}, \{x_2,x_4\}, \{x_2,x_5\}, \{x_3,x_5\}, \{x_4,x_5\} \} \big).$$
Moreover, the subhypergraph of $H$ on $X$ is
$$\tilde H =\big(X, \{\{x_1,x_2,x_3\}, \{x_1,x_4\}, \{x_2,x_4\}, \{x_2,x_5\}, \{x_3,x_5\} \}\big ).$$
Using Macaulay2, we compute that
$$\ass J(H')^2=\{(x_1,x_2,x_3), (x_1,x_4), (x_2,x_4), (x_2,x_5), (x_3,x_5), (x_4,x_5) \}\cup$$
$$\cup \{(x_1,x_2,x_3,x_4),(x_1,x_2,x_3,x_5),(x_2,x_4,x_5)\}$$
and $$\ass J(H')^3= \{(x_1,x_2,x_3), (x_1,x_4), (x_2,x_4), (x_2,x_5), (x_3,x_5), (x_4,x_5) \}\cup$$
$$\cup \{(x_1,x_2,x_3,x_4),(x_1,x_2,x_3,x_5),(x_2,x_4,x_5)\}\cup$$
$$\cup \{(x_1,x_2,x_3,x_4,x_5)\}.$$
We also know that $\ass (J(\tilde H)^3)$ and $\ass (J(\tilde H)^2)$ share the same elements, precisely
$$\{(x_1,x_2,x_3), (x_1,x_4), (x_2,x_4), (x_2,x_5), (x_3,x_5) \}$$
$$\cup \{(x_1,x_2,x_3,x_4),(x_1,x_2,x_3,x_5),(x_1,x_2,x_3,x_4,x_5)\}.$$
Then, from Proposition \ref{p.associated square} and Theorem \ref{t.tildeH}, we have
$$\ass (J(H)^2)=\ass (J(\tilde H)^2) \cup \{(x_4,x_5,y), (x_2,x_4,x_5,y)\}.$$
What about $\ass (J(H)^3)$?
The element $(x_1,x_2,x_3,x_4,x_5)$ appears both in $\ass (J(\tilde H)^3)$ and $\ass (J(H')^3)$ and it contains $(x_4,x_5)$. One can check that
$$\ass (J(H')^3:m) =(x_1,x_2,x_3,x_4,x_5) $$
and
$$\ass (J(\tilde H)^3:m) =(x_1,x_2,x_3,x_4,x_5)$$
where $m:=x_1x_2^2x_3x_4^2x_5^2.$
Thus, by Theorem \ref{t.tildeH}, $(x_1,x_2,x_3,x_4,x_5,y)\notin \ass (J(H)^3)$ and one can check that
$$\ass (J(H)^3)= \ass (J(\tilde H)^3)\cup \{(x_4,x_5,y), (x_2,x_4,x_5,y)\}.$$
\end{example}
\begin{example}\label{e.p and p+y}
Let $H$ be the hypergraph on the vertex set $V:=\{x_1,x_2,x_3,x_4,x_5,y\},$ given by
$$H=\left(V, \{\{x_1,x_2\}, \{x_1,x_3\}, \{x_1,x_4\}, \{x_1,x_5,y\}, \{x_2,x_3,x_4,x_5\}\}\right).$$
Let $H'$ be the shadow of $H$ on the vertex set $X:=\{x_1,x_2,x_3,x_4,x_5\}$ and $\tilde H$ the subhypergraph of $H$ on $X.$ Then
$$H'=\left(X, \{\{x_1,x_2\}, \{x_1,x_3\}, \{x_1,x_4\}, \{x_1,x_5\}, \{x_2,x_3,x_4,x_5\}\}\right)$$
and
$$\tilde H=\left(X, \{\{x_1,x_2\}, \{x_1,x_3\}, \{x_1,x_4\}, \{x_2,x_3,x_4,x_5\}\}\right).$$
A Macaulay2 computation shows that
$$\ass J(H')^2=\{(x_1,x_2), (x_1,x_3), (x_1,x_4), (x_1,x_5), (x_2,x_3,x_4,x_5)\}\cup \{(x_1,x_2,x_3,x_4,x_5)\} $$
and
$$\ass J(\tilde H)^2= \{(x_1,x_2), (x_1,x_3), (x_1,x_4), (x_2,x_3,x_4,x_5)\}\cup \{(x_1,x_2,x_3,x_4,x_5)\}.$$
The element $(x_1,x_2,x_3,x_4,x_5)$ appears both in $\ass (J(\tilde H)^3)$ and $\ass (J(H')^3)$ and it contains $(x_1,x_5)$. One can check that
$$(J(H')^2:x_1x_2x_3x_4x_5) = (x_1,x_2,x_3,x_4,x_5)$$
but $x_1x_2x_3x_4x_5\in J(\tilde H)^2.$ Thus
$$\ass (J(H)^2)= \ass (J(\tilde H)^2)\cup \{(x_1,x_5,y), (x_1,x_2,x_3,x_4,x_5, y)\}.$$
\end{example}
\section{An application to the persistence property}\label{s.application}
In this section, we apply the results of Section \ref{s.main case} to the persistence problem.
A squarefree monomial ideal $I$ is said to have the persistence property if $\ass(I^s)\subseteq\ass(I^{s+1})$ for any integer $s>0.$
The authors of \cite{TMR} describe an example of a cover ideal of a graph failing the persistence property.
We show how to construct, starting from a hypergraph whose cover ideal fails the persistence property, a new hypergraph whose cover ideal fail such property.
We use the notation introduced in Section \ref{s.main case}.
\begin{theorem}\label{t.persistence}
Let $H=(V,E)$ be a hypergraph where $V=X\cup \{y\}$ such that
\begin{itemize}
\item[1.] there exists only one edge $e_y\in E$ containing $y$;
\item[2.] $H$ has a shadow on $X$, say $H'=(X, E')\in \mathcal S(H).$
\end{itemize}
Suppose $J(H')$ fails the persistence property and let $\mathfrak p'\in \ass (J(H')^s)$ and $\mathfrak p'\notin \ass (J(H')^{s+1})$ for some $s>0$. Set $\tilde H:=H_X$ the subhypergraph of $H$ on $X$.
If the following conditions hold,
\begin{itemize}
\item[3.] $\mathfrak p'\notin ass(J(\tilde H)^s)$; and
\item[4.] $\mathfrak p'+(y)\notin ass(J(H)^{s+1}:y)$,
\end{itemize}
then $J(H)$ fails the persistence property.
\end{theorem}
\begin{proof}
By hypothesis we have $\mathfrak p'\in \ass (J(H')^s)$ and $\mathfrak p'\notin ass(J(\tilde H)^s)$. So by Theorem
\ref{t.tildeH}, one gets that
$$ \mathfrak p'+(y)\in \ass (J(H)^s).$$
Moreover, the hypothesis also ensures that
$\mathfrak p'\in \ass (J(H')^{s+1})$ and $\mathfrak p'+(y)\notin \ass(J(H)^{s+1}:y)$. Hence, from Theorem
\ref{t.from H to H'}, we have $ \mathfrak p'+(y)\notin \ass (J(H)^{s+1}).$
\end{proof}
\begin{example}\label{e.persistence fails}
In \cite{TMR}, Theorem 11 provides an example of a graph failing the persistence property.
The graph, denoted by $H_4$, has the vertex set on $X:=\{x_1,\ldots, x_{12}\}$ and the edge set
\begin{align*}
E:=\{&\{x_1,x_2\}, \{x_1,x_5\}, \{x_1,x_9\}, \{x_1,x_{12}\},\{x_2,x_3\}, \{x_2,x_6\},\{x_2,x_{10}\}, \{x_3,x_4\},\\
&\{x_3,x_7\},\{x_3,x_{11}\},\{x_4,x_8\}, \{x_4,x_9\}, \{x_4,x_{12}\}, \{x_5,x_6\},\{x_5,x_8\},\{x_5,x_9\}, \\
&\{x_6,x_7\},\{x_6,x_{10}\},\{x_7,x_8\},\{x_7,x_{11}\},\{x_8,x_{12}\},\{x_9,x_{10}\},\{x_{10},x_{11}\},\{x_{11},x_{12}\}\}.
\end{align*}
The persistence property fails since $\ass (J(H_4)^3)\not\subseteq \ass (J(H_4)^4)$.
In particular, $\mathfrak p_X\in \ass (J(H_4)^3)\setminus \ass (J(H_4)^4)$
We consider now the hypergraph $H$ on vertex set $V:=X\cup \{y\}$, constructed from $H_4$ by adding the variable $``y"$ only to the edge $\{x_1, x_2\}:$ $$H=(X\cup\{y\}, (E\setminus \left\{\{x_1,x_2\}\right\})\cup \left\{\{x_1,x_2, y\}\right\}).$$
\begin{center}
\begin{tikzpicture}[scale=0.7]
\node (v1) at (0,0) {};
\node (v2) at (3,0) {};
\node (v3) at (6,0) {};
\node (v4) at (9,0) {};
\node (v5) at (0,-2) {};
\node (v6) at (3,-2) {};
\node (v7) at (6,-2) {};
\node (v8) at (9,-2) {};
\node (v9) at (0,-4) {};
\node (v10) at (3,-4) {};
\node (v11) at (6,-4) {};
\node (v12) at (9,-4) {};
\node (y) at (1.5,0)[above] {$y$};
\draw [thick] (v1)--(v2);
\draw [thick] (v3)--(v2);
\draw [thick] (v3)--(v4);
\draw [thick] (v1)--(v5);
\draw [thick] (v5)--(v9);
\draw [thick] (v5)--(v6);
\draw [thick] (v6)--(v7);
\draw [thick] (v7)--(v8);
\draw [thick] (v4)--(v8);
\draw [thick] (v8)--(v12);
\draw [thick] (v9)--(v10);
\draw [thick] (v10)--(v11);
\draw [thick] (v11)--(v12);
\draw [thick] (v2)--(v6);
\draw [thick] (v10)--(v6);
\draw [thick] (v3)--(v7);
\draw [thick] (v7)--(v11);
\draw [thick] (v1) .. controls (-1.5,-1.5) and (-1.5,-2.5) .. (v9);
\draw [thick] (v2) .. controls (1.5,-1.5) and (1.5,-2.5) .. (v10);
\draw [thick] (v3) .. controls (4.5,-1.5) and (4.5,-2.5) .. (v11);
\draw [thick] (v4) .. controls (7.5,-1.5) and (7.5,-2.5) .. (v12);
\draw [thick] (v5) .. controls (3,-0.25) and (6,-0.25) .. (v8);
\draw [thick] (v9) .. controls (11,-8) and (12,-3) .. (v4);
\draw [thick] (v1) .. controls (-5.5,-4.5) and (3,-7.5).. (v12);
\fill (v1) circle (0.1) node [above] {$x_1$};
\fill (v2) circle (0.1) node [above] {$x_2$};
\fill (v3) circle (0.1) node [above] {$x_3$};
\fill (v4) circle (0.1) node [above] {$x_4$};
\fill (v5) circle (0.1) node [left] {$x_5$};
\fill (v6) circle (0.1) node [above right] {$x_6$};
\fill (v7) circle (0.1) node [above right] {$x_7$};
\fill (v8) circle (0.1) node [right] {$x_8$};
\fill (v9) circle (0.1) node [left] {$x_9$};
\fill (v10) circle (0.1) node [below] {$x_{10}$};
\fill (v11) circle (0.1) node [below] {$x_{11}$};
\fill (v12) circle (0.1) node [right] {$x_{12}$};
\end{tikzpicture}
\vspace*{-40pt}
\end{center}
With this construction, $H_4$ is the shadow of $H$ on the set $X$.
Moreover, the subhypergraph of $H$ on $X$ is $$\tilde H=(X, E\setminus \left\{\{x_1,x_2\}\right\}).$$
By Theorem \ref{t.persistence}, $H$ fails the persistence property and
$$\mathfrak p_V\in \ass (J(H)^3)\setminus \ass (J(H)^4).$$
One can check, for instance using Macaulay2, that actually $\mathfrak p_X\notin ass(J(\tilde H)^3)$ and $\mathfrak p_V\notin ass(J(H)^4:y).$
\end{example}
\begin{example}\label{e.persistence does not fail}
Take the hypergraph $H'$ on the vertex set $X=\{x_1,\ldots, x_{12}, x_{13}\}$ and the edge set
\begin{align*}
E=\{&\{x_1,x_2,x_{13}\}, \{x_1,x_5\}, \{x_1,x_9,x_{13}\}, \{x_1,x_{12},x_{13}\}, \{x_2,x_3,x_{13}\}, \{x_2,x_6,x_{13}\},\\
&\{x_2,x_{10},x_{13}\},\{x_3,x_4,x_{13}\}, \{x_3,x_7,x_{13}\},\{x_3,x_{11},x_{13}\},\{x_4,x_8,x_{13}\}, \{x_4,x_9,x_{13}\},\\
& \{x_4,x_{12},x_{13}\},\{x_5,x_6,x_{13}\}, \{x_5,x_8,x_{13}\},\{x_5,x_9,x_{13}\},\{x_6,x_7,x_{13}\},\{x_6,x_{10},x_{13}\},\\
&\{x_7,x_8,x_{13}\},\{x_7,x_{11},x_{13}\},\{x_8,x_{12},x_{13}\},\{x_9,x_{10},x_{13}\},\{x_{10},x_{11},x_{13}\},\{x_{11},x_{12},x_{13}\}\}.
\end{align*}
It was constructed from $H_4$, see example \ref{e.persistence fails}, by adding a new variable $``x_{13}"$ to all the edges but $\{x_1,x_5\}.$
Consider now the hypergraph $H$ on vertex set $V:=X\cup \{y\}$, constructed from $H'$ by adding the variable $``y"$ only to the edge $\{x_1, x_5\}$: $$H=(X\cup\{y\}, (E\setminus \left\{\{x_1,x_5\}\right\})\cup \left\{\{x_1,x_5, y\}\right\}).$$
With this construction, $H'$ is the shadow of $H$ on the set $X$.
A computation with Macaulay2 shows that $\mathfrak p_V\in J(H)^4$, but $\mathfrak p_X\notin J(H')^4$.
Indeed, we found two minimal monomials $m_1$, $m_2$ such that $\mathfrak p_V= (J(H)^4:m_1)=(J(H)^4:m_2)$ that are
$$m_1:=x_1^2x_2^3x_3^3x_4^2x_5^2 x_6^3 x_7^2 x_8^3 x_9^3 x_{10}^2 x_{11}^3 x_{12}^3y,\ \ m_2:=x_1^2 x_2^2 x_3^2 x_4^2 x_5^2 x_6^2 x_7^2 x_8^2 x_9^2 x_{10}^2 x_{11}^2 x_{12}^2x_{13}y.$$
Both are divisible by $y.$
Then the conditions in the statement of Theorem \ref{t.persistence} are not satisfied. By Theorem \ref{t.from H to H'}, we get $\mathfrak p_V\in \ass (J(H)^4:y)$.
Using Macaulay2, one can check that even if the hypergraph $H'$ fails the persistence property,
and in particular $\mathfrak p_X\in \ass (J(H')^3)\setminus \ass (J(H')^4)$,
we have $\ass (J(H)^3)\subseteq \ass (J(H)^4).$
\end{example}
|
\section{Introduction}
The problem considered in this note is to determine generators and relations for the graded rings $M_*(\Gamma_K)$ of Hilbert modular forms over real-quadratic fields $K$. There is some history to this. The structure has been worked out in the cases $\mathbb{Q}(\sqrt{5})$ (Gundlach, \cite{Gu}), $\mathbb{Q}(\sqrt{2})$ (Hammond, \cite{H}), $\mathbb{Q}(\sqrt{13})$ (van der Geer and Zagier, \cite{GZ}), $\mathbb{Q}(\sqrt{17})$ (Hermann, \cite{He} determined the structure of the subring of symmetric forms; the full ring appears in section 6.5 of \cite{M}), and $\mathbb{Q}(\sqrt{6})$ (van der Geer, \cite{G}). Section 6.2 of \cite{M} gives a more detailed overview. Simpler derivations for the fields $\mathbb{Q}(\sqrt{5}), \mathbb{Q}(\sqrt{13}), \mathbb{Q}(\sqrt{17})$ were given by Mayer \cite{M} using a reduction process against Borcherds products with simple zeros on the diagonal, and the work here extends this idea. \\
We only consider fields $K = \mathbb{Q}(\sqrt{p})$ of prime discriminant $p \equiv 1 \, (4)$ and class number one as this makes a number of aspects of the theory simpler (but there are ultimately workarounds for both of these assumptions). The first few such primes are $p=5,13,17,$ and the corresponding graded rings have the form
\begin{align*} M_*(\Gamma_{\mathbb{Q}(\sqrt{5})}) &= \mathbb{C}[X_2,X_5,X_6,X_{15}]/(R_{30}), \\ M_*(\Gamma_{\mathbb{Q}(\sqrt{13})}) &= \mathbb{C}[X_2,X_3,X_4,X_5,X_6,X_8,X_9] / (R_9,R_{10},R_{12},R_{18}), \\ M_*(\Gamma_{\mathbb{Q}(\sqrt{17})}) &= \mathbb{C}[X_2,X_3,X_4^{(1)},X_4^{(2)},X_5^{(1)},X_5^{(2)},X_6^{(1)},X_6^{(2)},X_7,X_8,X_9] / \mathcal{I}, \\ &\quad\quad\quad \mathcal{I} = (R_9,R_{10},R_{11},R_{12}^{(1)},R_{12}^{(2)},R_{13},R_{14},R_{18}),\end{align*} where $X_k,X_k^{(i)}$ denote modular forms of weight $k$ and where $R_k,R_k^{(i)}$ denote relations that are homogeneous of weight $k$, see \cite{M}. \\
No results seem to have been written down for larger primes. We could expect this to be more difficult for primes $p >17$, $p \equiv 1\, (4)$ because the Weil representation associated to the finite quadratic module $(\mathcal{O}_K^{\#}/ \mathcal{O}_K, N_{K/\mathbb{Q}})$ for the lattice of integers $\mathcal{O}_K$ admits nonzero cusp forms of weight $2$. More precisely, there are $\lceil \frac{p-5}{24} \rceil$ linearly independent cusp forms of weight $2$ as is well-known (and originally due to Hecke). For one thing, these cusp forms lift injectively under the Doi-Naganuma map to holomorphic differentials on the Hilbert modular surfaces such that those surfaces are never rational (see also \cite{HV}, chapter 3 for the classification of Hilbert modular surfaces) so one roughly expects the rings in question to be more complicated. More immediately, these cusp forms act as obstructions to the possible divisors of Borcherds products in the sense of section 3.2 of \cite{H1} so it is generally not possible to find modular forms with simple divisors to reduce against. \\
In this note we are able to solve this problem for the smallest two cases $\mathbb{Q}(\sqrt{29})$ and $\mathbb{Q}(\sqrt{37})$ where obstructions occur by considering the restrictions of Hilbert modular forms both to the diagonal and to other Hirzebruch-Zagier curves and by matching these simultaneously to polynomials in Eisenstein series and Borcherds products. The graded rings have the form \begin{align*} M_*(\Gamma_{\mathbb{Q}(\sqrt{29})}) &= \mathbb{C}[X_2^{(1)},X_2^{(2)},X_3^{(1)},X_3^{(2)},X_4,X_5,X_6^{(1)},X_6^{(2)},X_7,X_8,X_9] / \mathcal{I}_{29}, \\ M_*(\Gamma_{\mathbb{Q}(\sqrt{37})}) &= \mathbb{C}[X_2^{(1)},X_2^{(2)},X_3^{(1)},X_3^{(2)},X_3^{(3)},X_4^{(1)},X_4^{(2)},X_5^{(1)},X_5^{(2)},X_6^{(1)},X_6^{(2)},X_6^{(3)},X_7,X_8,X_9] / \mathcal{I}_{37}, \end{align*} where both $\mathcal{I}_{29}$ and $\mathcal{I}_{37}$ are homogeneous ideals of relations of weight between $6$ and $18$. We found minimal systems of $35$ generators of $\mathcal{I}_{29}$ and $77$ generators of $\mathcal{I}_{37}$. These are given explicitly in the ancillary material on arXiv and the author's website. \\
The methods used here likely apply to some other real-quadratic fields. Unfortunately the construction of generators is rather ad hoc, and we are unable to answer some natural questions. For example, are the graded rings always generated by Eisenstein series and Borcherds products? (Some preliminary calculations with the field $\mathbb{Q}(\sqrt{41})$ suggest that they may not be enough to produce all modular forms of weight 5.) Also, if $K \ne \mathbb{Q}(\sqrt{5})$, then are the graded rings always presented by generators of weight at most 9 and by relations of weight at most $18$? (The last question is similar to Main Theorem 1.4 of \cite{VZ}, which implies a statement of this type for all modular curves but which does not seem to apply to the Hilbert modular surfaces in an obvious way.) \\
Computing the Fourier expansion of a Borcherds product is not entirely trivial to do, even in the relatively simple case of Hilbert modular forms. (The paper \cite{GKR} discusses that problem in general and offers some improvements over the naive method of computing.) The ideals of relations would probably be difficult to work out directly. In this note we are able to avoid that problem entirely by using Koecher's principle and the existence of certain Borcherds products with convenient divisors. In particular we never need to compute the Fourier expansion of any Hilbert modular form explicitly. We only use the interpretation of restrictions along various curves on the Hilbert modular surface as elliptic modular forms, which are easier to compute. \\
\textbf{Acknowledgments.} A large part of what is presented below is the result of computations in SAGE and Macaulay2. Some easier computations were carried out on a CoCalc server. The Magma implementation of Hilbert modular forms as described in \cite{DV} was useful for checking correctness. I thank Martin Raum for providing resources with which the more time-consuming computations were performed. I am also grateful to Jan Hendrik Bruinier and John Voight for helpful discussions. This work was supported by the LOEWE research unit Uniformized Structures in Arithmetic and Geometry.
\section{Hilbert modular forms}
\textbf{2.1 Hilbert modular forms.} We quickly recall some of the basic theory of Hilbert modular forms attached to quadratic fields. The book \cite{H1} introduces Hilbert modular forms with a view toward Borcherds products and is therefore a useful reference for more details. Let $\mathbb{H}$ be the usual upper half-plane and let $K$ be a real-quadratic field of discriminant $d_K$ with ring of integers $\mathcal{O}_K$. For simplicity we assume that $K$ has class number one. A \textbf{Hilbert modular form} of (parallel) weight $k \in \mathbb{N}_0$ is a holomorphic function of two variables $f : \mathbb{H} \times \mathbb{H} \rightarrow \mathbb{C}$ which satisfies $$f\Big( \frac{a \tau_1 + b}{c \tau_1 + d}, \frac{a' \tau_2 + b'}{c' \tau_2 + d'} \Big) = (c \tau_1 + d)^k (c' \tau_2 + d')^k f(\tau_1,\tau_2)$$ for all $M = \begin{psmallmatrix} a & b \\ c & d \end{psmallmatrix} \in \Gamma_K = SL_2(\mathcal{O}_K)$. Here $a'$ denotes the conjugate of $a \in K$. \\
The invariance of $f$ under all translations $T_b = \begin{psmallmatrix} 1 & b \\ 0 & 1 \end{psmallmatrix}$, $b \in \mathcal{O}_K$ implies that $f$ has a Fourier series which we write in the form $$f(\tau_1, \tau_2) = \sum_{\nu \in \mathcal{O}_K^{\#}} a(\nu) q_1^{\nu} q_2^{\nu'}, \; \; q_i = e^{2\pi i \tau_i}, \; a(\nu) \in \mathbb{C}.$$ Here $\mathcal{O}_K^{\#}$ is the codifferent $$\mathcal{O}_K^{\#} = \{\nu \in K: \; \mathrm{Tr}_{K/\mathbb{Q}}(\nu x) \in \mathbb{Z} \; \text{for all} \; x \in \mathcal{O}_K\} = \sqrt{1 / d_K} \mathcal{O}_K.$$ The G\"otzky-Koecher principle (\cite{H1}, theorem 1.20) states that $a(\nu) = 0$ whenever either $\nu < 0$ or $\nu' < 0$ so the usual boundedness condition at cusps for automorphic forms is automatically satisfied. \\
Some of the simplest Hilbert modular forms are the \textbf{Hecke Eisenstein series} $$\mathbf{E}_k(\tau_1,\tau_2) = 1 + \frac{4}{\zeta_K(1-k)} \sum_{\substack{\nu \in \mathcal{O}_K^{\#} \\ \nu,\nu' > 0}} \sigma_{k-1}\Big( \nu \sqrt{d_K} \mathcal{O}_K \Big) q_1^{\nu} q_2^{\nu'}, \; \; k \ge 2 \; \text{even},$$ where $\zeta_K(s)$ is the Dedekind zeta function of $K$ and where $\sigma_{k-1}(\mathfrak{a}) = \sum_{\mathfrak{c} | \mathfrak{a}} N_{K/\mathbb{Q}}(\mathfrak{c})^{k-1}$ is the ideal divisor sum. When $k \ge 4$ these arise out of an averaging process similar to the Eisenstein series for $SL_2(\mathbb{Z})$, and the case $k=2$ can be recovered using Hecke's convergence trick. See \cite{H1}, section 1.5 for details. \\
Cusp forms are Hilbert modular forms that also vanish at the cusps: $(\infty,\infty)$ and $(a,a')$ for all $a \in K$. Up to equivalence by $\Gamma_K$ the number of cusps is exactly the class number of $K$. We have assumed that $K$ has class number one and therefore a form $f(\tau_1,\tau_2) = \sum_{\nu} c(\nu) q_1^{\nu} q_2^{\nu'}$ is a cusp form if and only if $c(0) = 0$. \\
One can generalize Hilbert modular forms to include a character of $SL_2(\mathcal{O}_K)$. The character group is always finite and depends up to isomorphism only on the ramification behavior of $2$ and $3$ in $\mathcal{O}_K$ (see \cite{BS}). We will only consider \textbf{symmetric characters} $\chi$: those for which $\chi(M) = \chi(M')$ for all $M \in SL_2(\mathcal{O}_K)$. This is equivalent to $f(\tau_1,\tau_2)$ and $f^{\sigma}(\tau_1,\tau_2) = f(\tau_2,\tau_1)$ having the same character. Any Hilbert modular form $f$ for a symmetric character can be decomposed into its symmetric and antisymmetric parts as $$f = f^+ + f^-, \;\; f^+ = \frac{f + f^{\sigma}}{2}, \; f^- = \frac{f - f^{\sigma}}{2}.$$ (Here $f$ is symmetric or antisymmetric if $f = f^{\sigma}$ or $f = -f^{\sigma}$, respectively.) \\
\textbf{2.2 Graded rings of Hilbert modular forms.} Let $K$ be a real-quadratic number field and set $\Gamma_K = SL_2(\mathcal{O}_K)$. The set $M_{*,char}(\Gamma_K)$ of all Hilbert modular forms with all characters is naturally a graded ring where forms are graded by both their weights and characters. We denote by $M_{*,sym}(\Gamma_K)$ and $M_*(\Gamma_K)$ the subrings of Hilbert modular forms with symmetric and trivial characters, respectively. \\
The graded rings $M_*(\Gamma_K)$ are always finitely generated. In fact the Baily-Borel theorem implies that their projective spectra $\mathrm{Proj} \, M_{*}(\Gamma_K)$ are complex surfaces. Together with the Noether normalization lemma this implies that we can always find a set of three algebraically independent forms $f_1,f_2,f_3$ such that $M_*(\Gamma_K)$ is finite over $\mathbb{C}[f_1,f_2,f_3]$ (i.e. finitely generated as a module). If $f_1,f_2,f_3$ have weights $k_1,k_2,k_3$ respectively then the theory of Hilbert series implies that there is a polynomial $p(t)$ for which $$\sum_{k=0}^{\infty} \mathrm{dim} \, M_k(\Gamma_K) t^k = \frac{p(t)}{(1 - t^{k_1})(1 - t^{k_2})(1- t^{k_3})}.$$ Of course neither $\{f_1,f_2,f_3\}$ nor their weights $\{k_1,k_2,k_3\}$ are unique. Since $M_{*,char}(\Gamma_K)$ is finite over $M_*(\Gamma_K)$ (as some power of any Hilbert modular form will have trivial character), we get the same result for the larger ring. In particular, for every character $\chi$ of $\Gamma_K$ there is a polynomial $p_{\chi}(t)$ such that, with the same exponents $k_1,k_2,k_3$, $$\sum_{k=0}^{\infty} \mathrm{dim} \, M_{k,\chi}(\Gamma_K) t^k = \frac{p_{\chi}(t)}{(1 - t^{k_1})(1 - t^{k_2})(1- t^{k_3})}.$$
\begin{rem} It was proved in \cite{TV} (theorem 3.4) that the Hilbert series of even-weight modular forms for every real quadratic field except $\mathbb{Q}(\sqrt{5})$ has the form $$\sum_{k=0}^{\infty} \mathrm{dim}\, M_{2k}(\Gamma_K) t^k = \frac{p(t)}{(1 - t)^2 (1 - t^3)}$$ for an explicit polynomial $p(t)$ of degree exactly $6$. This is not a special case of the paragraph above, and it does not hold for the full ring of modular forms.
\end{rem}
\textbf{2.3 Restrictions to Hirzebruch-Zagier curves.} Hilbert modular forms can be restricted to the Hirzebruch-Zagier cycles of \cite{HZ} to produce elliptic modular forms for $\Gamma_0(n)$. Recall that the Hirzebruch-Zagier cycle of discriminant $n$ is the set $T_n \subseteq \mathbb{H} \times \mathbb{H}$ of all points $(\tau_1,\tau_2)$ that satisfy an equation of the form $$a \tau_1 \tau_2 + \lambda \tau_1 + \lambda' \tau_2 + b = 0$$ for some $a,b \in \mathbb{Z}$ and $\lambda \in \mathcal{O}_K^{\#}$ with $ab - \lambda \lambda' = n$. These cycles are $\Gamma_K$-invariant and there are inclusions $T_n \subseteq T_{nd^2}$ for all $n,d \in \mathbb{N}$ which are strict if $d > 1$. \\
Suppose $\lambda \in \mathcal{O}_K$ is a totally positive integer of norm $\ell = \lambda \lambda'$. Then the function $$g(\tau) = \mathrm{Res}_{\lambda} f(\tau) = f(\lambda \tau, \lambda' \tau)$$ satisfies \begin{align*} g(M \cdot \tau) &= f\Big( \lambda \frac{a \tau + b}{\ell c \tau + d}, \lambda' \frac{a \tau + b}{\ell c \tau + d} \Big) \\ &= f\left( \begin{psmallmatrix} a & \lambda b \\ \lambda' c & d \end{psmallmatrix} \cdot (\lambda \tau), \begin{psmallmatrix} a & \lambda' b \\ \lambda c & d \end{psmallmatrix} \cdot (\lambda' \tau) \right) \\ &= \chi \left( \begin{psmallmatrix} a & \lambda b \\ \lambda' c & d \end{psmallmatrix} \right) (\ell c \tau + d)^{2k} g(\tau) \end{align*} for all $M = \begin{psmallmatrix} a & b \\ \ell c & d \end{psmallmatrix} \in \Gamma_0(\ell).$ Moreover the growth condition at cusps follows from that of $f$ and therefore $g(\tau)$ is a modular form of weight $2k$ and level $\Gamma_0(\ell)$ for the character $$\chi_g \left( \begin{psmallmatrix} a & b \\ \ell c & d \end{psmallmatrix} \right) = \chi \left( \begin{psmallmatrix} a & \lambda b \\ \lambda' c & d \end{psmallmatrix} \right).$$ In addition, if $\ell$ is prime and $f$ is symmetric, then we find \begin{align*} g(-1/\ell \tau) &= f\Big(-\frac{\lambda}{\ell \tau}, -\frac{\lambda'}{\ell \tau} \Big) \\ &= f\Big( -\frac{1}{\lambda' \tau}, -\frac{1}{\lambda \tau}\Big) \\ &= f\Big(-\frac{1}{\lambda \tau}, -\frac{1}{\lambda' \tau} \Big) \\ &= (\lambda \tau)^k (\lambda' \tau)^k f(\lambda \tau, \lambda' \tau) \\ &= \ell^k \tau^{2k} g(\tau), \end{align*} i.e. $g$ is an eigenform of the Atkin-Lehner involution $W_{\ell} g(\tau) = \ell^{-k} \tau^{-2k} g(-1/\ell \tau)$ with eigenvalue $+1$. Similarly, if $f$ is antisymmetric, then $g$ will have eigenvalue $-1$ under $W_{\ell}$. \\
\begin{rem} If $\ell$ is a prime and $\lambda$ is a fixed, totally positive integer of norm $\ell$, then the Hirzebruch-Zagier cycle $T_{\ell}$ is exactly the orbit of $$\{(\lambda \tau, \lambda' \tau): \; \tau \in \mathbb{H}\} \cup \{(\lambda' \tau, \lambda \tau): \; \tau \in \mathbb{H}\}$$ under $SL_2(\mathcal{O}_K)$. In that sense $f(\lambda \tau, \lambda' \tau)$ is a restriction of $f$ to $T_{\ell}$. In particular, if $f$ is either symmetric or antisymmetric (which includes Borcherds products), then it vanishes along the entire divisor $T_{\ell}$ whenever $f(\lambda \tau, \lambda' \tau) \equiv 0$ for any such element $\lambda$. Also note that $T_1$ is simply the orbit of the diagonal.
\end{rem}
\begin{rem} Even if the Hilbert modular form $f(\tau_1,\tau_2)$ vanishes along the diagonal, we can take its Taylor expansion in the variable $\tau_1$ about the point $\tau_2$: $$f(\tau_1,\tau_2) = g(\tau_2) (\tau_1 - \tau_2)^N + O((\tau_1 - \tau_2)^{N+1}), \; \; |\tau_1 - \tau_2| < y_2,$$ for some function $g$. For any matrix $M = \begin{psmallmatrix} a & b \\ c & d \end{psmallmatrix} \in SL_2(\mathbb{Z})$ compare $$f(M \cdot \tau_1, M \cdot \tau_2) = (c \tau_1 + d)^{-N} (c \tau_2 + d)^{-N} g(M \cdot \tau_2) (\tau_1 - \tau_2)^N + O((\tau_1 - \tau_2)^{N+1})$$ with $$(c \tau_1 + d)^k (c \tau_2 + d)^k f(M \cdot \tau_1, M \cdot \tau_2) = (c \tau_1 + d)^k (c \tau_2 + d)^k g(\tau_2) (\tau_1 - \tau_2)^N + O((\tau_1 - \tau_2)^{N+1})$$ to see that $g$ is a modular form of weight $2k+2N$. If $N \ge 1$, then $(\tau_1 - \tau_2)^N$ is not bounded, and therefore the growth condition on $f$ forces $g$ to be a cusp form. Similar expansions can be written out for Hilbert modular forms that vanish along the other Hirzebruch-Zagier curves, but we will not need any of these except for the diagonal.
\end{rem}
\textbf{2.4 Forced zeros.} Modular forms for $SL_2(\mathbb{Z})$ of certain weights have forced zeros at elliptic CM points. Namely, if the weight of a level one modular form $f$ is not divisible by $4$, then $f(i) = 0$; and if the weight is not divisible by $3$, then $f(\rho) = 0$ where $\rho =e^{2\pi i / 3}$. Some statements for Hilbert modular forms that are roughly in this spirit will be important later. We assume that $K = \mathbb{Q}(\sqrt{p})$ has prime discriminant $p$ and fundamental unit $\varepsilon$.
\begin{lem} (i) Every antisymmetric Hilbert modular form vanishes along the diagonal. \\ (ii) Every symmetric Hilbert modular form of weight $k$ and character $\chi$ for which $\chi(\begin{psmallmatrix} \varepsilon & 0 \\ 0 & \varepsilon^{-1} \end{psmallmatrix}) \ne (-1)^k$ vanishes on the Hirzebruch-Zagier divisor $T_p$ of discriminant $p$. \\ (iii) Every antisymmetric Hilbert modular form of weight $k$ and character $\chi$ with $\chi(\begin{psmallmatrix} \varepsilon & 0 \\ 0 & \varepsilon^{-1} \end{psmallmatrix}) \ne (-1)^{k+1}$ vanishes on the Hirzebruch-Zagier divisor $T_p$ of discriminant $p$.
\end{lem}
\begin{proof} (i) $f(\tau_1,\tau_2) = -f(\tau_2,\tau_1)$ implies $f(\tau,\tau) = 0$ identically. \\ (ii) Since $K$ has prime discriminant, the fundamental unit $\varepsilon$ has negative norm. In particular $\lambda = \varepsilon \sqrt{p}$ is totally positive and $T_p$ is the $\Gamma_K$-orbit of $\{(\lambda \tau, \lambda'\tau) : \; \tau \in \mathbb{H}\}$. We find $$f(\lambda \tau, \lambda' \tau) = \chi \left( \begin{psmallmatrix} \varepsilon & 0 \\ 0 & \varepsilon^{-1} \end{psmallmatrix} \right) (-1)^k f(\lambda' \tau, \lambda \tau) = \chi\left( \begin{psmallmatrix} \varepsilon & 0 \\ 0 & \varepsilon^{-1} \end{psmallmatrix} \right) (-1)^k f(\lambda \tau, \lambda' \tau),$$ using the transformation under $\begin{psmallmatrix} \varepsilon & 0 \\ 0 & \varepsilon^{-1} \end{psmallmatrix}$ which sends $(\lambda' \tau, \lambda \tau)$ to $(\lambda \tau, \lambda' \tau)$ and using the symmetry of $f$. This implies the claim. \\ (iii) This uses nearly the same argument as (ii).
\end{proof}
\section{Two kinds of Borcherds products}
\textbf{3.1 Borcherds products for $\Gamma_0(p)$.} Let $p$ be a prime. We recall a generalization of theorem 14.1 of \cite{B1} which produces modular products of level $\Gamma_0(p)$. (This is itself a special case of theorem 13.3 of \cite{B2}.) Let $A_p$ denote the Kohnen plus space of nearly-holomorphic modular forms of weight $1/2$ and level $(\Gamma_0(4p),\chi_{\vartheta})$ with rational Fourier coefficients for the multiplier system $\chi_{\vartheta}$ of the classical theta function $\vartheta(\tau) = 1 + 2q + 2q^4 + 2q^9 + ...$ In other words $A_p$ consists of Laurent series $$f(q) = \sum_{n \gg -\infty} c(n) q^n, \; q = e^{2\pi i \tau}, \; c(n) \in \mathbb{Q}$$ for which $c(n) = 0$ if $n$ is a quadratic nonresidue mod $4p$, and for which $f \cdot \vartheta$ is a modular form of weight $1$ and level $\Gamma_0(4p)$ for the Dirichlet character $$\chi(n) = \begin{cases} \left( \frac{-1}{n}\right) : & n \; \text{coprime to} \; 4p \\ 0: & \text{otherwise}. \end{cases}$$
Define $$\delta_p(n) = \begin{cases} 1: & n \equiv 0 \, (p) \\ 1/2: & n \not \equiv 0 \, (p), \end{cases}$$ and for $f(\tau) = \sum_n c(n) q^n$ let $\tilde c(n) = c(n) \delta_p(n)$. For any modular form $f \in A_p$ as above, Borcherds' theorem states that there is a rational number $h \in \mathbb{Q}$ (the ``Weyl vector") such that $$\Psi_f(\tau) = q^h \cdot \prod_{n=1}^{\infty} (1 - q^n)^{\tilde c(n^2)}$$ converges for large enough $\mathrm{im}(\tau)$ and extends to a multivalued modular form of weight $c(0)$ and level $\Gamma_0(p)$ and some character. All zeros and singularities $w$ of $\Psi_f$ on $\mathbb{H}$ satisfy a quadratic equation of the form $aw^2 + bw + c = 0$ with $a,b,c \in \mathbb{Z}$ and $a \equiv 0 \, (p)$, and the order of $\Psi_f$ at such a point is a sum of coefficients for negative exponents in $f$: $$\mathrm{ord}(\Psi_f;w) = \sum_{m=1}^{\infty} \tilde c(m^2(b^2 - 4ac)).$$ In particular, $\Psi_f$ has no branch cuts if these orders are all integral, and it is a holomorphic away from the cusps if these orders are also nonnegative. The Weyl vector $h$ can be calculated as the constant term of the product of $f$ with a mock Eisenstein series of weight $3/2$ and level $4p$ which is closely related to Zagier's Eisenstein series \cite{Z}. Of course one can also expand the product above and simply guess the (unique) value of $h$.
\begin{ex} Let $p = 3$; then $A_3$ contains the input function $$f(\tau) = q^{-3} + 1 - 12q + 42q^4 - 76q^9 + 168q^{12} - 378q^{13} + 690q^{16} - 897q^{21} + 1456q^{24} - 3468q^{25} \pm ...$$ which can be computed in SAGE to any desired precision by multiplying the weight four form $$1 - 2q + q^3 - 12q^4 + 24q^5 + 44q^7 - 108q^8 - 2q^9 + 84q^{11} - 84q^{12} + O(q^{13}) \in M_4(\Gamma_0(12))$$ by the eta product $\frac{\eta(2\tau)}{\eta(\tau)^2 \eta(12\tau)^6}$. Its lift $\Psi_f$ is a modular form of weight $1$ and level $(\Gamma_0(3),\chi)$ which has simple zeros exactly on the $\Gamma_0(3)$-orbit of the CM point $w = \frac{1}{2} + \frac{\sqrt{3}}{6}i$ (i.e. the solution $w \in \mathbb{H}$ of $3w^2 - 3w + 1 = 0$). The Weyl vector is $h=0$. From the Fourier expansion $$\Psi_f(\tau) = (1 - q)^{-12/2} (1 - q^2)^{42/2} (1-q^3)^{-76} (1 - q^4)^{690/2} (1 - q^5)^{-3468/2}...= 1 + 6q + 6q^3 + 6q^4 + O(q^6)$$ we see that $\Psi_f$ is the theta function of $x^2 + xy + y^2$ and that $\chi$ is the nontrivial Nebentypus mod $3$.
\end{ex}
\begin{rem} Borcherds products are more naturally orthogonal modular forms; in the special case above, one should really consider $\Gamma_0(p)$ acting by conjugation on a certain lattice of symmetric matrices and preserving the determinant (a quadratic form of signature $(1,2)$). The Atkin-Lehner involution $W_p f(\tau) = p^{-k/2}\tau^{-k} f(-\frac{1}{p\tau})$ also preserves the determinant, and therefore all Borcherds products $\Psi_f$ are eigenforms of $W_p$. We should also mention that passing to vector-valued input functions makes the condition that $p$ is prime or even squarefree unnecessary. \\
\end{rem}
\textbf{3.2 Borcherds products for $\mathbb{Q}(\sqrt{p})$.} We recall the construction of Hilbert forms by Borcherds products (using \cite{B2},\cite{BB}; see also section 3.2 of \cite{H1} for an introduction). Let $p \equiv 1 \, (4)$ be a prime discriminant and let $K = \mathbb{Q}(\sqrt{p})$ with ring of integers $\mathcal{O}_K = \mathbb{Z}[\frac{1 + \sqrt{p}}{2}]$. One can essentially identify Hilbert modular forms of level $\mathcal{O}_K$ with orthogonal modular forms for the signature $(2,2)$ lattice $\mathcal{O}_K \oplus II_{1,1}$, where $\mathcal{O}_K$ carries the norm-form $N_{K/\mathbb{Q}}$. \\
Let $\chi$ denote the quadratic Dirichlet character attached to $K$: $\chi(n) = \left(\frac{n}{p}\right)$, and let $M_0^!(\Gamma_0(p),\chi)$ denote the space of nearly-holomorphic modular forms of weight zero, level $\Gamma_0(p)$ and Nebentypus $\chi$. Denote by $B_p$ the plus-space $$B_p = \Big\{ F = \sum_{n \gg -\infty} c(n) q^n \in M_0^!(\Gamma_0(p),\chi): \; c(n) \in \mathbb{Q}, \; c(n) = 0 \; \text{if} \; \chi(n) = -1\Big\}.$$ As before we define $\delta_p(n) = 1$ if $n \equiv 0 \, (p)$ and $\delta_p(n) = 1/2$ otherwise, and set $\tilde c(n) = c(n) \delta_p(n)$ for any series $F = \sum_n c(n) q^n$. \\
Let $f \in B_p$. As in section 3.2 of \cite{H1}, the principal part of $F$ determines a splitting of $\mathbb{R}_{>0} \times \mathbb{R}_{>0}$ into \textbf{Weyl chambers} $W$. These are the connected components that are left after removing the curves $\{(\lambda y, \lambda' y): \, y \in \mathbb{R}_{>0}\}$ for all totally positive $\lambda \in \mathcal{O}_K$ with $c(-\lambda \lambda') \ne 0$. For any Weyl chamber $W$, Borcherds' theorem states that there is a number $h_W \in K$ (the ``Weyl vector") such that $$\Psi_F(\tau_1,\tau_2) = q_1^{h_W} q_2^{h_W'} \prod_{\substack{\nu \in \mathcal{O}_K^{\#} \\ \langle \nu, W \rangle > 0}} (1 - q_1^{\nu} q_2^{\nu'})^{\tilde c(p \cdot N_{K/\mathbb{Q}}\nu)}$$ converges for all $(\tau_1,\tau_2) \in \mathbb{H} \times \mathbb{H}$ with $\mathrm{im}(\tau_1\tau_2)$ sufficiently large and that $\Psi_F$ extends to a multivalued Hilbert modular form of weight $c(0)/2$ for some character. Here $\langle \nu, W \rangle > 0$ means that $\mathrm{Tr}_{K/\mathbb{Q}}(\nu w) > 0$ for all $w \in K$ with $(w,w') \in W$. The divisor of $\Psi_F$ is a sum of Hirzebruch-Zagier cycles: $$\mathrm{div} \, \Psi_F = \sum_{n=1}^{\infty} \tilde c(-n) T_n.$$ In particular, $\Psi_F$ is holomorphic if and only if all orders $\sum_{m=1}^{\infty} \tilde c(-nm^2)$ are nonnegative integral. (Note that this does not imply that all $c(-n)$ are nonnegative.)
\begin{rem} The Borcherds products attached to the various Weyl chambers and a single input function all have the same weight and divisor so the quotient between any two is a nonzero constant (by the G\"otzky-Koecher principle). That constant is not generally $1$. From this it follows that the Borcherds products attached to the different Weyl chambers all have the same character as well.
\end{rem}
\begin{rem} A holomorphic Borcherds product $\Psi_F$ is a cusp form if and only if the compactification of its divisor intersects all cusps. This too can be read off the input function. When $K$ has class number one, it is sufficient for the input function $f$ to have a nonzero coefficient in any exponent $q^{-n}$ with $n \in N_{K/\mathbb{Q}} \mathcal{O}_K$. Therefore Hilbert modular form Borcherds products of small weight tend to be cusp forms. Indeed all Borcherds products we construct in this note for $\mathbb{Q}(\sqrt{29})$ and $\mathbb{Q}(\sqrt{37})$ are cusp forms. \\
\end{rem}
\textbf{3.3 Restricting Borcherds products.} Let $K = \mathbb{Q}(\sqrt{p})$ for a prime $p \equiv 1 \, (4)$. Suppose $\Psi_F(\tau_1,\tau_2) = q_1^h q_2^{h'} \prod_{\langle \nu, W \rangle > 0} (1 - q_1^{\nu} q_2^{\nu'})^{\tilde c(p \cdot N_{K/\mathbb{Q}} \nu)}$ is the expansion of the Borcherds lift of $$F(\tau) = \sum_{n \in \mathbb{Z}} c(n) q^n$$ within some Weyl chamber $W$, and set $\tilde c(n) = c(n)$ or $c(n)/2$ as in section 3.2. Also suppose that $\lambda \in \mathcal{O}_K$ satisfies $(\lambda,\lambda') \in W$ (in particular, $\lambda$ is totally positive) and that $N_{K/\mathbb{Q}}(\lambda) = \ell$ is $1$ or a prime. Restricting gives
$$\Psi_F(\lambda \tau, \lambda' \tau) = q^{\mathrm{Tr}(\lambda h)} \prod_{n=1}^{\infty} (1 - q^n)^{\tilde b(n^2)}$$ for the coefficients $$\tilde b(n^2) = \sum_{\substack{ \langle \nu, W \rangle > 0 \\ \mathrm{Tr}(\lambda \nu) = n}} \tilde c(p \cdot N_{K/\mathbb{Q}} \nu) = \sum_{\mathrm{Tr}(\lambda \nu) = n} \tilde c(p \cdot N_{K/\mathbb{Q}} \nu),$$ where we use the fact that $\mathrm{Tr}(\lambda \nu) > 0$ implies $\langle \nu, W \rangle > 0$, by lemma 3.2 of \cite{Br}. This suggests that $\mathrm{Res}_{\lambda} \Psi_F(\tau) = \Psi_F(\lambda \tau, \lambda' \tau)$ itself is likely a Borcherds product as in section 2.1. This is true and for $\ell \ne p$ the input function can be produced from $F$ as follows. Divide the coefficients of all $q^n$, $n \not \equiv 0 \, (p)$ in $F$ by two (to get the modified coefficients $\tilde c(n)$), change variables $q \mapsto q^{4\ell/p}$, multiply the result by $\vartheta(\tau/p)$ where $\vartheta(\tau) = 1 + 2q + 2q^4 + 2q^9 + ...$ is the theta function, and restrict to integer exponents. In other words we take the input function $$f(\tau) = \sum_{n=0}^{\infty} \sum_{\substack{r \in \mathbb{Z} \\ 4\ell n + r^2 \equiv 0 \, (p)}} \tilde c(n) q^{(4\ell n + r^2)/p} = \sum_{m=0}^{\infty} b(m) q^m, \; \; b(m) = \sum_{\substack{r \in \mathbb{Z} \\ mp - r^2 \equiv 0 \, (4\ell)}} \tilde c\left( \frac{mp-r^2}{4\ell}\right).$$
The procedure for constructing $f$ (and in particular the proof that $f \in A_{\ell}$ is a valid input function in the sense of section 3.1) is the same as remark 10 and example 11 of \cite{W} with minor changes. That is, the vector-valued modular form corresponding to $F$ is transformed into a weak Jacobi form $\varphi(\tau,z)$ of fractional index $\frac{p}{4\ell}$ through the theta decomposition and then $f(\tau)$ is the Kohnen plus form corresponding to the vector-valued modular form $\varphi(\tau,0)$. This is ultimately the same procedure as the theta-contraction of \cite{Ma}.\\
The square coefficients are $$b(n^2) = \sum_{\substack{r \in \mathbb{Z} \\ pn^2 - r^2 \equiv 0 \, (4\ell)}} \tilde c\left( p \cdot N_{K/\mathbb{Q}}\nu\right), \; \; \nu = \frac{n - r / \sqrt{p}}{2 \lambda}.$$ Note that if $n \not \equiv 0 \, (\ell)$ and $pn^2 - r^2 \equiv 0 \, (4\ell)$, then exactly one of $\sqrt{p}n \pm r$ is divisible by $2\lambda$ in $\mathcal{O}_K$ (and the other is divisible by $2\lambda'$) because $\lambda$ is prime and $\lambda \nmid \mathrm{gcd}(\sqrt{p}n+r,\sqrt{p}n-r)$. (Remember that we assume $\mathcal{O}_K$ has class number one.) If $n \equiv 0 \, (\ell)$ and therefore also $r \equiv 0 \, (\ell)$, then both $\sqrt{p}n \pm r$ are divisible by $2\ell$ and therefore by $2\lambda$. Altogether it follows that the $\tilde b(n^2) = b(n^2) \delta_p(n)$ satisfy
$$\tilde b(n^2) = \sum_{\substack{\nu \in \mathcal{O}_K^{\#} \\ \mathrm{Tr}_{K/\mathbb{Q}}(\lambda \nu) = n}} \tilde c(p \cdot N_{K/\mathbb{Q}}\nu) = \begin{cases} b(n^2)/2: \; n \not \equiv 0 \, (p) \\ b(n^2): \; n \equiv 0 \, (p), \end{cases}$$ and therefore that $\mathrm{Res}_{\lambda} \Psi_F(\tau) = \Psi_f(\tau)$ holds. (That the Weyl vectors match up follows from the fact that $\mathrm{Res}_{\lambda} \Psi_F(\tau)$ and $\Psi_f(\tau)$ are meromorphic modular forms of the same weight and level whose Fourier expansions are equal up to a shift; so their quotient is a power of $q = e^{2\pi i \tau}$ which is invariant under $SL_2(\mathbb{Z})$ and therefore equals $1$. It does not seem to be as easy to show that the Weyl vectors match directly from their definitions.) \\
When $\ell = p$ the procedure of theta contraction is similar. The only difference in the result is that we do not divide any of the coefficients of $F$ by 2; instead we multiply $F(4\tau) \vartheta(\tau/p)$ directly and restrict to integer exponents.
\begin{ex} Let $p = 5$. Then $B_5$ contains the input function $$F(\tau) = 2q^{-1} + 10 + 22q - 108q^4 + 110q^5 + 88q^6 - 790q^9 + 680q^{10} \pm ...$$ which lifts to Gundlach's weight 5 cusp form $s_5(\tau_1,\tau_2)$ for $\mathbb{Q}(\sqrt{5})$ that vanishes exactly on the $\Gamma_K$-orbit of the diagonal \cite{Gu}. There is a Weyl chamber of $F$ with Weyl vector $h = \frac{1}{2} - \frac{\sqrt{5}}{10}$. We compute the restriction to the Hirzebruch-Zagier curve $T_5$ as follows. Take the totally positive element $\lambda_5 = \frac{5 + \sqrt{5}}{2}$ of norm $5$, which satisfies $\mathrm{Tr}_{K/\mathbb{Q}}(\lambda_5 h) > 0$. We multiply \begin{align*} &\quad F(4\tau) \vartheta(\tau/5) \\ &=\Big( 2q^{-4} + 10 + 22q^4 - 108q^{16} + 110q^{20} + 88q^{24} \pm ... \Big) \Big( 1 + 2q^{1/5} + 2q^{4/5} + 2q^{9/5} + ... \Big) \\ &= 2q^{-4} + 4q^{-19/5} + 4q^{-16/5} + 4q^{-11/5} + 4q^{-4/5} + 10 + 20q^{1/5} + 20q^{4/5} + 4q + ...\end{align*} and restrict to integer exponents to find the input function $$f(\tau) = 2q^{-4} + 10 + 4q + 22q^4 + 20q^5 + 44q^9 - 104q^{16} + 130q^{20} \pm ... \in A_5.$$ Its lift is \begin{align*} s_5(\lambda_5 \tau, \lambda_5' \tau) &= q^{\mathrm{Tr}_{K/\mathbb{Q}}(\lambda_5 h)} (1 - q)^2(1-q^2)^{11} (1 - q^3)^{22} (1 - q^4)^{-52}... \\ &= q^2 (1 - 2q - 10q^2 + 140q^4 + ...) \\ &= \frac{1}{4} \eta(\tau)^8 \eta(5\tau)^8 (5 E_2(5\tau) - E_2(\tau)), \end{align*} where $E_2(\tau) = 1 - 24 \sum_{n=1}^{\infty} \sigma_1(n) q^n$ with $\sigma_1(n) = \sum_{d | n} d$, and $\eta(\tau) = q^{1/24} \prod_{n=1}^{\infty} (1 - q^n)$. In particular, the case of Borcherds' theorem in section 3.1 shows that this has simple zeros exactly on the $\Gamma_0(5)$-orbit of the discriminant $4$ CM point $\tau = \frac{2+i}{5}$, so $(\lambda_5 \tau,\lambda_5' \tau)$ represents the intersection $T_1 \cap T_5$. \\
To compute the restriction onto the curve $T_{11}$ with $\lambda_{11} = \frac{7 + \sqrt{5}}{2}$, we have to pass from the coefficients $c(n)$ to $\tilde c(n)$. Therefore we multiply \begin{align*} &\Big( q^{-44/5} + 10 + 11q^{44/5} - 54q^{176/5} + 110q^{44} + 44q^{264/5} \pm ... \Big) \Big(1 + 2q^{1/5} + 2q^{4/5} + 2q^{9/5} + ... \Big) \\ &= q^{-44/5} + 2q^{-43/5} + 2q^{-8} + 2q^{-7} + 2q^{-28/5} + 2q^{-19/5} + 2q^{-8/5} + 10 + 20q^{1/5} + 20q^{4/5} + 2q + ... \end{align*} and restrict to integer exponents to find the input form $$f(\tau) = 2q^{-8} + 2q^{-7} + 10 + 2q + 2q^4 + 20q^5 + 22q^9 + 22q^{12} + ... \in A_{11}$$ whose Borcherds lift is \begin{align*} s_5(\lambda_{11} \tau, \lambda_{11}' \tau) &= q^{\mathrm{Tr}_{K/\mathbb{Q}}(\lambda_{11} h)} (1 - q)(1-q^2)(1-q^3)^{11} ... \\ &= q^3 (1 - q - q^2 - 10q^3 + 10q^5 + 121q^6 \pm ...) \in S_{10}(\Gamma_0(11)). \end{align*} Note that $\lambda_5$ and $\lambda_{11}$ lie in the same Weyl chamber of $F$ so there really is a single Hilbert modular form with these restrictions. In general this would only hold up to a constant multiple as in remark 7.
\end{ex}
\section{The graded ring of Hilbert modular forms for $\mathbb{Q}(\sqrt{29})$}
Let $K = \mathbb{Q}(\sqrt{29})$ with ring of integers $\mathcal{O}_K = \mathbb{Z}[\frac{1 + \sqrt{29}}{2}]$. Since $2$ and $3$ both remain inert in $\mathcal{O}_K$, there are no nontrivial characters of $SL_2(\mathcal{O}_K)$. \\
The principal parts that extend to nearly-holomorphic forms in $M_0^!(\Gamma_0(29),\chi)$ can be determined by theorem 3.49 of \cite{H1}; essentially the only obstruction is that the product with the unique cusp form in the plus space $$q - 3q^4 - 3q^5 + 5q^6 + 2q^7 - 2q^9 - q^{13} - q^{16} \pm ... \in S_2(\Gamma_0(29),\chi)$$ has constant term zero. In this way one can prove that there exist nearly-holomorphic modular forms $F_2,F_3,G_3,F_4,F_6,G_6$ of weight $0$ and level $\Gamma_0(29)$ and Nebentypus $\chi(d) = \left(\frac{29}{d}\right)$ whose Fourier expansions begin as follows: \begin{align*} F_2(\tau) &= 2q^{-4} + 6q^{-1} + 4 + 2q - 2q^4 + 8q^5 - 2q^6 + 14q^7 + 4q^9 - 10q^{13} \pm ... \\ F_3(\tau) &= 2q^{-6} + 4q^{-4} + 2q^{-1} + 6 - 6q - 2q^4 + 12q^5 + 16q^6 + 18q^7 + 28q^9 - 10q^{13} \pm ... \\ G_3(\tau) &= 2q^{-5} + 6q^{-1} + 6 + 10q + 4q^4 + 12q^5 + 0q^6 + 0q^7 - 6q^9 + 20q^{13} \pm ... \\ F_4(\tau) &= 2q^{-6} + 2q^{-5} + 2q^{-4} + 2q^{-1} + 8 + 2q + 4q^4 + 16q^5 + 18q^6 + 4q^7 + 18q^9 + 20q^{13} \pm ... \\ F_6(\tau) &= q^{-29} + 6q^{-1} + 12 + 30q + 124q^4 + 162q^5 + 252q^6 + 336q^7 + 648q^9 + 2050q^{13} + ... \\ G_6(\tau) &= 2q^{-16} + 2q^{-6} + 2q^{-5} - 2q^{-1} + 12 -10q - 34q^4 - 48q^5 - 70q^6 + 80q^7 + 186q^9 + 336q^{13} \pm ... \end{align*}
The Borcherds lifts $\phi_i = \Psi_{F_i}$, $\psi_i = \Psi_{G_i}$ have weight $i$, and their divisors can be read off the principal parts of the input functions:
$$\mathrm{div}\, \phi_2 = 3T_1 + T_4, \; \; \mathrm{div} \, \phi_3 = T_1 + 2T_4 + T_6, \; \; \mathrm{div} \, \psi_3 = 3T_1 + T_5,$$ $$\mathrm{div} \, \phi_4 = T_1 + T_4 + T_5 + T_6, \; \; \mathrm{div} \, \phi_6 = 3T_1 + T_{29}, \; \; \mathrm{div} \, \psi_6 = -T_1 + T_5 + T_6 + T_{16}.$$
We will also need the following (holomorphic) quotients of the above forms as generators:
\begin{align*} \phi_5 &= \frac{\psi_3 \phi_4}{\phi_2}, \; \; \mathrm{div} \, \phi_5 = T_1 + 2T_5 + T_6; \\ \phi_7 &= \frac{\phi_3 \phi_6}{\phi_2}, \; \; \mathrm{div} \, \phi_7 = T_1 + T_4 + T_6 + T_{29}; \\ \phi_8 &= \frac{\phi_3 \psi_3 \phi_6}{\phi_2^2}, \; \; \mathrm{div} \, \phi_8 = T_1 + T_5 + T_6 + T_{29}; \\ \phi_9 &= \frac{\phi_3^2 \psi_3 \phi_6}{\phi_2^3}, \; \; \mathrm{div} \, \phi_9 = -T_1 + T_4 + T_5 + 2T_6 + T_{29}.\end{align*}
All are cusp forms. We fix $\lambda_1 = 1$ and $\lambda_5 = \frac{7 + \sqrt{29}}{2}$ and compute restrictions up to a constant multiple (which does not matter here and is omitted by abuse of notation) using the procedure in section 3.3:
\begin{align*} \mathrm{Res}_{\lambda_1} \psi_6(\tau) &= \Delta(\tau) = q - 24q^2 + 252q^3 - 1472q^4 + 4830q^5 \pm ... \\ \mathrm{Res}_{\lambda_1} \phi_9(\tau) &= \Delta(\tau) E_6(\tau) = q - 528q^2 - 4284q^3 + 147712q^4 \pm ... \\ \mathrm{Res}_{\lambda_1} \mathbf{E}_2(\tau) &= E_4(\tau) = 1 + 240q + 2160q^2 + 6720q^3 + 17520q^4 + ...\end{align*} (none of which actually require computation because the respective spaces of modular forms or cusp forms are one-dimensional) and
\begin{align*} \mathrm{Res}_{\lambda_5} \phi_2(\tau) &= s_4(\tau) = q - 4q^2 + 2q^3 + 8q^4 - 5q^5 \pm ... \\ \mathrm{Res}_{\lambda_5} \phi_3(\tau) &= s_4(\tau) e_2(\tau) = q + 2q^2 - 4q^3 - 28q^4 + 25q^5 \pm ... \\ \mathrm{Res}_{\lambda_5} \phi_4(\tau) &= s_4(\tau)^2 = q^2 - 8q^3 + 20q^4 - 70q^6 \pm ... \\ \mathrm{Res}_{\lambda_5} \phi_6(\tau) &= s_4(\tau)^2 e_4(\tau) = q^2 - 18q^3 + 10q^4 + 240q^5 \pm ... \\ \mathrm{Res}_{\lambda_5} \phi_7(\tau) &= e_2(\tau) e_4(\tau) s_4(\tau)^2 = q^2 - 12q^3 - 80q^4 + 870q^6 \pm ... \\ \mathrm{Res}_{\lambda_5} \mathbf{E}_2(\tau) &= e_2(\tau)^2 - 4s_4(\tau) = 1 + 8q + 88q^2 + 256q^3 + 664q^4 + 1400q^5 + ...\end{align*} with the omitted restrictions above being zero, where we have fixed the following generators of the ring of modular forms for $\Gamma_0(5)$:
\begin{align*} e_2(\tau) &= \frac{5 E_2(5 \tau) - E_2(\tau)}{4} = 1 + 6q + 18q^2 + 24q^3 + 42q^4 + 6q^5 + ... \\ e_4(\tau) &= \frac{25 E_4(5 \tau) - E_5(\tau)}{24} = 1 - 10q - 90q^2 - 280q^3 - 730q^4 - 1010q^5 - ... \\ s_4(\tau) &= \eta(\tau)^4 \eta(5\tau)^4 = q - 4q^2 + 2q^3 + 8q^4 - 5q^5 \pm ... \end{align*} which satisfy a single relation $$e_4^2 = e_2^4 - 44 s_4 e_2^2 - 16s_4^2$$ in weight $8$. Note that $e_2,e_4$ are eigenforms of the Atkin-Lehner involution $W_5$ with eigenvalue $-1$ while $s_4$ has eigenvalue $+1$ under $W_5$.
\begin{lem} Let $f$ be a Hilbert modular form of parallel weight for $\mathbb{Q}(\sqrt{29})$ and let $\lambda_5 = \frac{7 + \sqrt{29}}{2}$. Then the restriction $f(\lambda_5 \tau, \lambda_5' \tau)$ of $f$ to $T_5$ coincides with the restriction of some polynomial in $\mathbf{E}_2,\phi_2,\phi_3,\phi_6,\phi_7$.
\end{lem}
\begin{proof} Any form $f$ can be split into its symmetric and antisymmetric parts, and it is enough to argue for each of those parts separately. We also argue differently depending on whether the weight is even or odd. \\
Case 1: suppose $f \in M_k^{sym}(\Gamma_K)$ is symmetric and has even weight. Then $\mathrm{Res}_{\lambda_5} f$ is an eigenform of $W_5$ with eigenvalue $+1$ and has weight divisible by $4$ and is therefore a polynomial expression $P$ in $e_2^2,s_4$. Since $$e_2^2 = \mathrm{Res}_{\lambda_5}(\mathbf{E}_2 + 4\phi_2), \; \; s_4 = \mathrm{Res}_{\lambda_5} \phi_2$$ we find $$\mathrm{Res}_{\lambda_5} f = \mathrm{Res}_{\lambda_5} P(\mathbf{E}_2 + 4\phi_2, \phi_2) = \mathrm{Res}_{\lambda_5} \tilde P(\mathbf{E}_2,\phi_2)$$ for some polynomial $\tilde P$. \\
Case 2: suppose $f \in M_k^{anti}(\Gamma_K)$ is antisymmetric and has odd weight. In particular $f$ vanishes on the diagonal and is therefore a cusp form; so $\mathrm{Res}_{\lambda_5} f$ is also a cusp form (and is therefore divisible by $s_4$) and has weight $2$ mod $4$ (and is therefore divisible by $e_2$). The remainder $\frac{\mathrm{Res}_{\lambda_5} f}{e_2 s_4}$ has eigenvalue $+1$ under the involution $W_5$ and is therefore a polynomial in $e_2^2, s_4$. Since $$e_2 s_4 = \mathrm{Res}_{\lambda_5} \phi_3$$ it follows from case 1 that $\mathrm{Res}_{\lambda_5} f$ coincides with the restriction of some polynomial in $\mathbf{E}_2,\phi_2,\phi_3$. \\
Case 3: suppose $f \in M_k^{sym}(\Gamma_K)$ is symmetric and has odd weight. In particular it has a forced zero on $T_{29}$. After multiplying by $\phi_2$, we obtain a form $\phi_2 f$ which vanishes on $T_{29}$ and vanishes to order at least three along the diagonal and is therefore divisible by $\phi_6$. The remainder $\frac{f \phi_2}{\phi_6}$ is antisymmetric of odd weight so by the argument in case $2$ we find $$\mathrm{Res}_{\lambda_5} (f \phi_2 \phi_6^{-1}) = e_2 s_4 P(e_2^2,s_4)$$ for some polynomial $P$, and therefore $$\mathrm{Res}_{\lambda_5} f = \frac{\mathrm{Res}_{\lambda_5} \phi_6}{\mathrm{Res}_{\lambda_5} \phi_2} \cdot e_2 s_4 P(e_2^2,s_4) = e_2 e_4 s_4^2 P(e_2^2,s_4).$$ Since $e_2 e_4 s_4^2 = \mathrm{Res}_{\lambda_5} \phi_7$ and $e_2^2, s_4$ were accounted for in case 1, we see that $\mathrm{Res}_{\lambda_5} f$ is the restriction of some polynomial in $\mathbf{E}_2,\phi_2,\phi_7$. \\
Case 4: suppose $f \in M_k^{anti}(\Gamma_K)$ is antisymmetric and has even weight; in this case there are forced zeros on both $T_1$ and $T_{29}$. After multiplying by $\phi_3$ we see that $\phi_3 f$ is divisible by $\phi_6$, and the remainder is antisymmetric of odd weight. By the argument in case $2$ again we find $$\mathrm{Res}_{\lambda_5} (f \phi_3 \phi_6^{-1}) = e_2 s_4 P(e_2^2,s_4)$$ for some polynomial $P$, and therefore $$\mathrm{Res}_{\lambda_5} f = \frac{\mathrm{Res}_{\lambda_5} \phi_6}{\mathrm{Res}_{\lambda_5} \phi_3} \cdot e_2 s_4 P(e_2^2,s_4) = e_4 s_4^2 P(e_2^2,s_4).$$ Since $e_4 s_4^2 = \mathrm{Res}_{\lambda_5} \phi_6$ we see that $\mathrm{Res}_{\lambda_5} f$ is the restriction of some polynomial in $\mathbf{E}_2,\phi_2,\phi_6$.
\end{proof}
\begin{prop} Every Hilbert modular form of parallel weight for $\mathbb{Q}(\sqrt{29})$ is an isobaric polynomial in the generators $$\mathbf{E}_2,\phi_2,\phi_3,\psi_3,\phi_4,\phi_5,\phi_6,\psi_6,\phi_7,\phi_8,\phi_9.$$
\end{prop}
\begin{proof} We use induction on the weight $k$ of $f$. If $k \le 0$, then $f$ is constant. \\
In general, we can assume by the previous lemma that $f$ vanishes on $T_5$. In particular, $f$ is a cusp form. As before, we split $f$ into its symmetric and antisymmetric parts and obtain four cases to distinguish. \\
Case 1: $k$ is odd and $f$ is symmetric. In particular $f$ has a forced zero on $T_{29}$. Then the restriction of $f$ to the diagonal is a cusp form of weight $2k \equiv 2 \, (4)$ and therefore has the form $$f(\tau,\tau) = \Delta(\tau) E_6(\tau) P(E_4,\Delta)$$ for some polynomial $P$; so $g := f - \phi_9 P(\mathbf{E}_2,\psi_6)$ vanishes on both $T_1$ and $T_{29}$. Actually $g$ must have at least a double zero along $T_1$ because $g \cdot \phi_4$ is divisible by $\phi_6$ and the quotient $\frac{g \phi_4}{\phi_6}$ is antisymmetric and therefore vanishes on the diagonal. Therefore, $g$ expands about the diagonal in the form $$g(\tau_1,\tau_2) = h(\tau_2) (\tau_1 - \tau_2)^2 + O((\tau_1 - \tau_2)^3),$$ where $h$ is a cusp form of weight $2k + 4 \equiv 2 \, (4)$ and therefore has the form $$h(\tau) = \Delta(\tau) E_6(\tau) P_2(E_4, \Delta)$$ for some polynomial $P_2$. On the other hand, the form $\phi_7$ has a double zero on $T_1$ and the expansion $$\phi_7(\tau_1,\tau_2) = \Delta(\tau_2) E_6(\tau_2) (\tau_1 - \tau_2)^2 + O((\tau_1 - \tau_2)^3)$$ up to a constant multiple (since $\Delta E_6$ is the unique normalized cusp form of weight $18$); for the argument here there is no loss of generality in assuming that constant is $1$. Therefore $$g - \phi_7 P_2(\mathbf{E}_2,\psi_6) = f - \phi_9 P(\mathbf{E}_2,\psi_6) - \phi_7 P_2(\mathbf{E}_2,\psi_6)$$ has a zero on $T_{29}$ and at least a triple zero on $T_1$ and is therefore divisible by $\phi_6$. The quotient $$\frac{f - \phi_9 P(\mathbf{E}_2,\psi_6) - \phi_7 P_2(\mathbf{E}_2,\psi_6)}{\phi_6}$$ has smaller weight than $f$, and the claim follows by induction. \\
Case 2: $k$ is odd and $f$ is antisymmetric. Expand $f$ about the diagonal in the form $$f(\tau_1,\tau_2) = h(\tau_2) (\tau_1 - \tau_2) + O((\tau_1 - \tau_2)^2);$$ then $h(\tau_2)$ is a cusp form of weight $2k+2 \equiv 0 \, (4)$ and therefore of the form $\Delta(\tau) P(E_4,\Delta)$ for some polynomial $P$. Since $\phi_5$ has the expansion $$\phi_5(\tau_1,\tau_2) = \Delta(\tau_2) (\tau_1 - \tau_2) + O((\tau_1 - \tau_2)^2)$$ (up to a possible constant multiple, $\Delta(\tau_2)$ is the unique cusp form of weight $12$), it follows that $$g = f - \phi_5 P(\mathbf{E}_2,\psi_6)$$ has at least a double zero along $T_1$ and still vanishes along $T_5$ (because $\phi_5$ does). Actually $g$ has at least a triple zero along $T_1$ because $g \phi_5$ is divisible by $\psi_3$ with the result being antisymmetric and therefore still vanishing along $T_1$. It follows that $g$ itself is divisble by $\psi_3$. The quotient $\frac{g}{\psi_3}$ has smaller weight than $f$, and the claim follows by induction. \\
Case 3: $k$ is even and $f$ is symmetric. The restriction $\mathrm{Res}_{\lambda_1} f$ has the form $\Delta(\tau) P(E_4,\Delta)$ for some polynomial $P$ and therefore $g = f - \psi_6 P(\mathbf{E}_2,\psi_6)$ vanishes along $T_1$ (and it still vanishes along $T_5$, because both $f$ and $\psi_6$ do). Actually it must have at least a double zero along $T_1$, because $g\phi_4$ is divisible by $\psi_3$ with the result being antisymmetric and therefore still vanishing along $T_1$. We expand $g$ along the diagonal as $$g(\tau_1,\tau_2) = h(\tau_2) (\tau_1 - \tau_2)^2 + O((\tau_1 - \tau_2)^3);$$ then $h$ is a cusp form of weight divisble by $4$ so it is $h = \Delta \cdot P_2(E_4,\Delta)$ for another polynomial $P_2$. With the expansion $$\phi_4(\tau_1,\tau_2) = \Delta(\tau_2) (\tau_1 - \tau_2)^2 + O((\tau_1 - \tau_2)^3)$$ (up to a constant multiple), it follows that $g - \phi_4 P_2(\mathbf{E}_2,\psi_6)$ has at least a triple zero along $T_1$ and continues to vanish along $T_5$ (since both $g$ and $\phi_4$ do). Therefore $g - \phi_4 P_2(\mathbf{E}_2,\psi_6)$ is divisible by $\psi_3$ with the quotient having smaller weight, and the claim follows by induction. \\
Case 4: $k$ is even and $f$ is antisymmetric (and therefore $f$ vanishes on both $T_1$ and $T_{29}$). Expand $f$ about the diagonal: $$f(\tau_1,\tau_2) = h(\tau_2) (\tau_1 - \tau_2) + O((\tau_1 - \tau_2)^2).$$ Then $h$ is a cusp form of weight $2$ mod $4$ and therefore $h = \Delta E_6 \cdot P(E_4,\Delta)$ for some polynomial $P$. Since $$\phi_8(\tau_1,\tau_2) = \Delta(\tau_2) E_6(\tau_2) (\tau_1 - \tau_2) + O((\tau_1 - \tau_2)^2)$$ (up to a constant multiple), it follows that $f - \phi_8 P(\mathbf{E}_2,\psi_6)$ has at least a double zero along $T_1$ (and continues to vanish on $T_5$, because $\phi_8$ does). As in the previous cases it follows that $f - \phi_8 P(\mathbf{E}_2,\psi_6)$ actually has at least a triple zero along $T_1$. It is therefore divisible by $\psi_3$ with the result having smaller weight, and the claim follows by induction.
\end{proof}
\begin{prop} The graded ring $M_*(\Gamma_{\mathbb{Q}(\sqrt{29})})$ is defined by the 11 generators $$\mathbf{E}_2,\phi_2,\phi_3,\psi_3,\phi_4,\phi_5,\phi_6,\psi_6,\phi_7,\phi_8,\phi_9$$ in weights $2,2,3,3,4,5,6,6,7,8,9$ and by 35 relations in weights 6 through 18.
\end{prop}
\begin{proof} The strategy we used to compute relations is as follows. There are totally positive elements $$\lambda_{23} = \frac{11 + \sqrt{29}}{2}, \; \; \lambda_{35} = \frac{13 + \sqrt{29}}{2}$$ of norms $23$ and $35$, respectively, which lie in the same Weyl chamber with respect to all $10$ of the Borcherds products considered above. Moreover the obstruction principle shows that there is a nearly-holomorphic modular form whose Fourier expansion begins $2q^{-35} + 2q^{-23} + 48 + ...$ and which lifts to a Borcherds product $\phi_{24}$ of weight $24$ with simple zeros along $T_{23}$ and $T_{35}$. Any nonzero Hilbert modular form $f$ whose restrictions to $T_{23}$ and $T_{35}$ both vanish is divisible by $\phi_{24}$; the quotient $\frac{f}{\phi_{24}}$ is also holomorphic in the cusps by the G\"otzky-Koecher principle and therefore has nonnegative weight so $f$ must have had weight at least $24$. \\
In this way we are able to compute all relations up to weight $24$ by computing relations only among the restrictions to $T_{23}$ and $T_{35}$ which is somewhat easier. The exact relations we found are tedious to write out and are left to the ancillary material. We mention here that $\{\mathbf{E}_2,\phi_3,\psi_3\}$ is algebraically independent (because the restrictions of $\mathbf{E}_2$ and $\phi_3$ to $T_5$ are already algebraically independent, and $\psi_3$ vanishes there) and that any relation can be reduced against the relations we found up to weight $18$ to eliminate the variables other than $\mathbf{E}_2,\phi_3,\psi_3$. It follows that the 35 relations we found up to weight $18$ generate the entire ideal. The Hilbert series is \[ \sum_{k=0}^{\infty} \mathrm{dim}\, M_k(\Gamma_{\mathbb{Q}(\sqrt{29})}) t^k = \frac{(1 + t + t^2)(1 - t + t^2 + t^4 + t^6 - t^7 + t^8)}{(1 - t^2)(1 - t^3)^2}. \qedhere \]
\end{proof}
\section{The graded ring of Hilbert modular forms for $\mathbb{Q}(\sqrt{37})$}
Let $K = \mathbb{Q}(\sqrt{37})$ with ring of integers $\mathcal{O}_K = \mathbb{Z}[\omega]$, $\omega = \frac{1 + \sqrt{37}}{2}$. Since $2$ is inert and $3$ splits in $\mathcal{O}_K$, theorem 1 of \cite{BS} implies that the full character group of $SL_2(\mathcal{O}_K)$ is generated by two characters $\chi_1, \chi_{\omega}$ which are characterized by their effect on the translations $T_1,T_{\omega}$ by $1$ and $\omega$: $$\chi_1(T_1) = e^{2\pi i / 3}, \; \chi_1(T_{\omega}) = 1$$ and $$\chi_{\omega}(T_1) = 1, \; \chi_{\omega}(T_{\omega}) = e^{2\pi i / 3}.$$ The condition for a character $\chi$ being symmetric is $\chi(T_{\omega}) = \chi(T_{\omega'}) = \chi(T_1) \chi(T_{\omega}^{-1})$, i.e. $\chi(T_1) = \chi(T_{\omega})^2$. In particular the group of symmetric characters of $SL_2(\mathcal{O}_K)$ is cyclic of order three and generated by $\chi = \chi_1 \chi_{\omega}^{-1}$. \\
We compute input functions using the obstruction principle as in the last section; there is again a single cusp form that acts as the obstruction to extending arbitrary principal parts to nearly-holomorphic modular forms and it is $$q - q^3 - 2q^4 + 3q^7 - 2q^9 + 4q^{10} - 3q^{11} + 2q^{12} \pm ... \in S_2(\Gamma_0(37),\chi).$$ Therefore there exist nearly-holomorphic modular forms $F_1,G_1,F_2,G_2,F_4,G_7$ of weight $0$ and level $\Gamma_0(37)$ and Nebentypus $\chi(d) = \left(\frac{37}{d}\right)$ whose Fourier expansions begin as follows:
\begin{align*} F_1(\tau) &= 2q^{-3} + 2q^{-1} + 2 + 2q + 4q^3 + 2q^7 +0q^9 - 4q^{10} + 2q^{11} +4q^{12} \pm ... \\ G_1(\tau) &= 2q^{-4} + 4q^{-1} + 2 - 2q + 2q^3 - 2q^4 + 6q^7 + 8q^9 + 2q^{10} + 2q^{11} - 4q^{12} \pm ... \\ F_2(\tau) &= 2q^{-10} + 2q^{-4} + 2q^{-3} - 2q^{-1} + 4 - 6q - 8q^3 + 8q^4 + 18q^7 - 4q^9 + 6q^{10} + 34q^{11} + 30q^{12} \pm ... \\ G_2(\tau) &= 2q^{-10} + 4q^{-4} + 4 - 10q - 10q^3 + 6q^4 + 22q^7 + 4q^9 + 12q^{10} + 34q^{11} + 22q^{12} \pm ... \\ F_4(\tau) &= q^{-37} + 2q^{-1} + 8 + 24q + 42q^3 + 80q^4 + 170q^7 + 300q^9 + 416q^{10} + 504q^{11} + 664q^{12} + ... \\ G_7(\tau) &= q^{-37} + 2q^{-12} + 2q^{-3} + 14 + 28q + 60q^3 + 68q^4 + 150q^7 + 308q^9 + 450q^{10} + 520q^{11} + 700q^{12} + ... \end{align*}
The Borcherds lifts $\phi_i = \Psi_{F_i}$, $\psi_i = \Psi_{G_i}$ have weight $i$ and their divisors can be read off the principal parts of the input functions: $$\mathrm{div} \, \phi_1 = T_1 + T_3, \; \; \mathrm{div} \, \psi_1 = 2T_1 + T_4, \; \mathrm{div} \, \phi_2 = -T_1 + T_3 + T_4 + T_{10},$$ $$\mathrm{div} \, \psi_2 = 2T_4 + T_{10}, \; \mathrm{div} \, \phi_4 = T_1 + T_{37}, \; \mathrm{div} \, \psi_7 = T_3 + T_{12} + T_{37}.$$ In particular $\phi_1,\psi_1,\psi_2,\phi_4$ vanish on the diagonal and $\phi_2,\psi_7$ do not. Moreover $\phi_1,\psi_1,\phi_4$ are antisymmetric and $\phi_2,\psi_2,\psi_7$ are symmetric, and all of the products above have a nontrivial character of order $3$; the notation is chosen such that $\phi_1,\phi_2,\phi_4$ have the same character $\chi$ and that $\psi_1,\psi_2,\psi_7$ have the same character $\chi^{-1}$. All of this can be proved directly, but it is easier to read it off of the restrictions to the diagonal and to the curve $T_3$ which are worked out in the next paragraph. Note that by construction $G_2 = F_2 + G_1 - F_1$ and therefore $\phi_1 \psi_2 = \psi_1 \phi_2$. \\
We fix the totally positive integers $\lambda_1 = 1, \; \lambda_3 = \frac{7 + \sqrt{37}}{2}$. Then one can compute the restrictions $$\mathrm{Res}_{\lambda_1} \phi_2(\tau) = \eta(\tau)^8, \; \; \mathrm{Res}_{\lambda_1} \psi_7(\tau) = E_6(\tau) \eta(\tau)^{16}, \; \; \mathrm{Res}_{\lambda_1} \mathbf{E}_2(\tau) = E_4(\tau)$$ and \begin{align*} \mathrm{Res}_{\lambda_3} \psi_1(\tau) &= \eta(\tau)^2 \eta(3\tau)^2 = q^{1/3} (1 - 2q - q^2 + 5q^4 + 4q^5 - 7q^6 \pm ...) \\ \mathrm{Res}_{\lambda_3} \psi_2(\tau) &= \eta(\tau)^2 \eta(3\tau)^2 e_2(\tau) = q^{1/3} (1 + 10q + 11q^2 - 72q^3 + 29q^4 - 44q^5 + 29q^6 \pm ...) \\ \mathrm{Res}_{\lambda_3} \phi_4(\tau) &= \eta(\tau)^4 \eta(3 \tau)^4 e_4(\tau) = q^{2/3} (1 - 34q - 148q^2 + 454q^3 - 559q^4 + 2418q^5 + 680q^6 \pm ...) \\ \mathrm{Res}_{\lambda_3} \mathbf{E}_2(\tau) &= e_2(\tau)^2 = 1 + 24q + 216q^2 + 888q^3 + 1752q^4 + 3024q^5 + 7992q^6 \pm ...\end{align*} with the omitted restrictions above being zero. Here we use the notation $$e_2(\tau) = \frac{3 E_2(3\tau) - E_2(\tau)}{2} = 1 + 12q + 36q^2 + ..., \; \; e_4(\tau) = \frac{9E_4(3\tau) - E_4(\tau)}{8} = 1 - 30q - 270q^2 -...$$ Using these computations we can prove:
\begin{lem} Let $f \in M_*(\Gamma_K)$ be a Hilbert modular form of parallel weight and trivial character. Then there is a polynomial $P$ such that $f - P(\psi_1,\psi_2,\phi_4,\mathbf{E}_2)$ vanishes on the Hirzebruch-Zagier divisor $T_3$. If $f$ is symmetric or antisymmetric, then one can also choose $P(\psi_1,\psi_2,\phi_4,\mathbf{E}_2)$ to be symmetric or antisymmetric, respectively.
\end{lem}
\begin{proof} The ring of modular forms of level $\Gamma_0(3)$ and trivial character is generated by the forms \begin{align*} e_2(\tau) &= \frac{3 E_2(3\tau) - E_2(\tau)}{2} = 1 + 12q + 36q^2 + 12q^3 + ... \\ e_4(\tau) &= \frac{9 E_4(3\tau) - E_4(\tau)}{8} = 1 - 30q - 270q^2 - 570q^3 - ... \\ s_6(\tau) &= \eta(\tau)^6 \eta(3\tau)^6 = q - 6q^2 + 9q^3 + 4q^4 \pm ... \end{align*} where $e_2$ has weight $2$, $e_4$ has weight $4$, $s_6$ is a cusp form of weight $6$, together with a single relation $$e_4^2 = e_2^4 - 108 e_2 s_6$$ in weight $8$. All three generators are eigenforms of the Atkin-Lehner operator $W_3$ with eigenvalue $-1$. We split $f$ into its symmetric and antisymmetric parts and argue for each part separately. \\
Case 1: suppose $f \in M_k^{sym}(\Gamma_K)$ is symmetric and has even weight. Then $\mathrm{Res}_{\lambda_3}f$ is an eigenform of $W_3$ with eigenvalue $+1$ and has weight divisible by $4$; so it is a polynomial expression in $e_2^2, e_2 s_6, s_6^2$. Since $$e_2^2 = \mathrm{Res}_{\lambda_3} \mathbf{E}_2, \; e_2 s_6 = \mathrm{Res}_{\lambda_3}(\psi_1^2 \psi_2), \; s_6^2 = \mathrm{Res}_{\lambda_3} \psi_1^6,$$ we can find a polynomial in $\psi_1^6,\psi_1^2\psi_2,\mathbf{E}_2$ whose restriction to $T_3$ equals $\mathrm{Res}_{\lambda_3} f$. This will also be symmetric because $\psi_1^2, \psi_2, \mathbf{E}_2$ are symmetric. \\
Case 2: suppose $f \in M_k^{sym}(\Gamma_K)$ is symmetric and has odd weight. Then $f$ is a cusp form and therefore $\mathrm{Res}_{\lambda_3} f$ is also a cusp form; in particular it is divisble by $s_6$, and $s_6^{-1} \mathrm{Res}_{\lambda_3}$ has eigenvalue $-1$ under $W_3$. Therefore all monomials in $e_2,e_4,s_6$ that occur in $s_6^{-1} \mathrm{Res}_{\lambda_3} f$ must contain $e_4$ (as all expressions involving only $e_2,s_6$ in weights $0$ mod $4$ will have eigenvalue $+1$) so $\mathrm{Res}_{\lambda_3} f$ is $e_4 s_6$ multiplied by some polynomial in $e_2^2, e_2 s_6, s_6^2$. The claim follows from the previous case together with $$e_4 s_6 = \mathrm{Res}_{\lambda_3}(\phi_4 \psi_1),$$ and the resulting polynomial expression will again be symmetric because $\phi_4 \psi_1$ is symmetric. \\
Case 3: suppose $f \in M_k^{anti}(\Gamma_K)$ is antisymmetric. In particular $f$ vanishes along the diagonal and therefore also at the cusps; so $\mathrm{Res}_{\lambda_3} f$ is a cusp form that is an eigenform of $W_3$ with eigenvalue $-1$. After dividing by $s_6$, we reduce to case $1$; since $s_6 = \mathrm{Res}_{\lambda_3} \psi_1^3$ and $\psi_1^3$ is antisymmetric, the resulting polynomial expression is antisymmetric.
\end{proof}
This quickly implies the same result for all of the symmetric characters:
\begin{lem} Let $f \in M_*(\Gamma_K,\chi)$ be a Hilbert modular form of any weight with symmetric character $\chi$. Then there is a polynomial $P$ as above for which $f - P(\psi_1,\psi_2,\phi_4,\mathbf{E}_2)$ vanishes on the divisor $T_3$.
\end{lem}
\begin{proof} The nontrivial symmetric characters restrict on $T_3$ to the characters of $s_2(\tau) = \eta(\tau)^2 \eta(3\tau)^2$ and its square. Note that any modular form of level $\Gamma_0(3)$ that transforms under one of those characters is a cusp form (because the order at all cusps is nonintegral and in particular nonzero) and is therefore divisible by $s_2$ or $s_2^2$, respectively. Since $\mathrm{Res}_{\lambda_3} \psi_1 = s_2$, we can use the previous argument.
\end{proof}
By further restricting to the diagonal we can compute generators:
\begin{prop} Every Hilbert modular form $f \in M_*(\Gamma_K,\chi)$ for a symmetric character $\chi$ is an isobaric polynomial in $\phi_1,\psi_1,\phi_2,\psi_2,\phi_4,\psi_7,\mathbf{E}_2$.
\end{prop}
\begin{proof} We induct on the weight of $f$. If $f$ has weight zero, then it is constant. Otherwise, we again split $f$ into its symmetric and antisymmetric parts and argue for each part separately. \\
Case 1: suppose $f$ is symmetric and has odd weight. Then $f$ has a forced zero on $T_{37}$ so its restriction $g = \mathrm{Res}_{\lambda_1} f$ to the diagonal is a cusp form for $SL_2(\mathbb{Z})$ of weight $2$ mod $4$ for some cubic character, and is therefore some polynomial expression $P$ in $E_4, \eta^8, E_6 \eta^{16}$. More precisely, if $g$ has the character of $\eta^8$ then $\eta^{-8}g$ has weight $0$ mod $4$ and no character, and is therefore a polynomial in $E_4$ and $\Delta = (\eta^8)^3$. If $g$ has the character of $\eta^{16}$ then $\eta^{-16} g$ has weight $2$ mod $4$ and no character, and therefore is also divisible by $E_6$, and $E_6^{-1} \eta^{-16} g$ is holomorphic of weight $0$ mod $4$ so the previous sentence applies. Finally, if $g$ has the trivial character then it vanishes at $i$ and is therefore divisible by $E_6$. As $g$ is a cusp form, we can divide by $\Delta E_6 = \eta^8 \cdot E_6 \eta^{16}$ with the remainder having weight $0$ mod $4$ again. Altogether, we can write $$\mathrm{Res}_{\lambda_1} f = P(E_4,\eta^8,E_6 \eta^{16}) = \mathrm{Res}_{\lambda_1} P(\mathbf{E}_2,\phi_2,\psi_7).$$ In particular, $f - P(\mathbf{E}_2,\phi_2,\psi_7)$ vanishes on $T_1$ and is still symmetric and of odd weight so it continues to vanish on $T_{37}$. Therefore we can divide by $\phi_4$ (which has simple zeros only on $T_1$ and $T_{37}$) to obtain a holomorphic form $\frac{f - P(\mathbf{E}_2,\phi_2,\psi_7)}{\phi_4}$ of lower weight which is a polynomial as in the claim by induction. This implies the claim for $f$ itself. \\
Case 2: suppose $f$ is antisymmetric and has odd weight. We use the previous lemmas and assume without loss of generality that $f$ already vanishes on $T_3$. As an antisymmetric form, $f$ has a forced zero on $T_1$. In particular it is divisible by $\phi_1$ and $\frac{f}{\phi_1}$ has smaller weight than $f$ so the claim follows by induction. \\
Case 3: Now suppose that $f$ has even weight; and again by the previous lemmas, assume without loss of generality that $f$ already vanishes on $T_3$. Then $f$ is a cusp form, and its restriction to the diagonal is a cusp form of level $1$ whose weight is divisible by $4$ and which is therefore some polynomial in $E_4, \eta^8$ in which all monomials contain $\eta^8$. In particular, $\mathrm{Res}_{\lambda_1} f$ is the restriction of some polynomial $Q(\mathbf{E}_2,\phi_2)$ in which all monomials contain $\phi_2$. Since $\phi_2$ vanishes on $T_3$, we find that $f - Q(\mathbf{E}_2,\phi_2)$ vanishes on both $T_1$ and $T_3$, so it is divisible by $\phi_1$; as $\frac{f - Q(\mathbf{E}_2,\phi_2)}{\phi_1}$ has weight strictly less than that of $f$, the claim follows by induction.
\end{proof}
\begin{prop} The graded ring $M_{*,sym}(\Gamma_K)$ of modular forms with symmetric characters is presented by seven generators $$\phi_1,\psi_1,\mathbf{E}_2,\phi_2,\psi_2,\phi_4,\psi_7$$ of weights $1,1,2,2,2,4,7$ and by 9 relations $R_{3,1},R_{4,1},R_{4,\chi},R_{8,1},R_{8,\chi},R_{8,\chi^2},R_{9,\chi},R_{11,1},R_{14,\chi}$ in weights $3$ through $14$.
\end{prop}
\begin{proof} We take the positive-definite integers $$\lambda_{21} = \frac{11 + \sqrt{37}}{2}, \; \lambda_{33} = \frac{13 + \sqrt{37}}{2}$$ of norm $21$ and $33$ which lie in a common Weyl chamber for each of the Borcherds products above. The obstruction principle shows that there is a nearly-holomorphic modular form in $B_{37}$ whose Fourier expansion begins $2q^{-33} + 2q^{-21} + 32 + ...$ and which therefore lifts to a Hilbert modular form $f_{16}$ of weight $16$ with simple zeros exactly along $T_{21}$ and $T_{33}$. By the G\"otzky-Koecher principle (arguing as in the previous section) every nonzero modular form which vanishes along both $T_{21}$ and $T_{33}$ must have weight at least $16$ so we are able to find all relations up to weight $16$ by only computing relations among the restrictions to $T_{21},T_{33}$. \\
In this way we compute the relations
\begin{align*} R_{3,1} &= \psi_1 \phi_2 - \phi_1 \psi_2; \\ R_{4,1} &= 5\phi_2 \psi_2 - 5\phi_1 \psi_1 \mathbf{E}_2 - 4 \phi_1^2 \psi_1^2 - 20\phi_1^2 \phi_2; \\ R_{4,\chi} &= 5 \psi_2^2 - 5 \psi_1^2 \mathbf{E}_2 -4 \phi_1 \psi_1^3 - 20 \phi_1^2 \psi_2; \\ R_{8,1} &= 5 \phi_2^2 \phi_4 - 5 \phi_1^2 \mathbf{E}_2 \phi_4 - 20\phi_1 \psi_7 + 36 \phi_1^3 \psi_1 \phi_4; \\ R_{8,\chi} &= \psi_1 \psi_7 - \phi_1^2 \phi_2 \phi_4 - 2\phi_1^2 \psi_1^2 \phi_4; \\ R_{8,\chi^2} &= 125 \phi_4^2 - 125 \psi_1^4 \mathbf{E}_2^2 + 13500 \psi_1^6 \psi_2 - 250\phi_1 \psi_1 \psi_2 \mathbf{E}_2^2 + 38050\phi_1 \psi_1^5 \mathbf{E}_2 - 125 \phi_1^2 \mathbf{E}_2^3 + 27000\phi_1^2 \phi_2^3 \\ &\quad \quad + 35850\phi_1^2 \psi_1^2 \psi_2 \mathbf{E}_2 + 33895\phi_1^2 \psi_1^6 + 11700\phi_1^3 \psi_1 \mathbf{E}_2^2 + 163340 \phi_1^3 \psi_1^3 \psi_2 + 42750 \phi_1^4 \phi_2 \mathbf{E}_2 \\ &\quad\quad + 119710\phi_1^4 \psi_1^2 \mathbf{E}_2 + 82466 \phi_1^5 \psi_1^3 + 447700 \phi_1^6 \psi_2 + 3375 \phi_1^8; \\ R_{9,\chi} &= 5 \psi_2 \psi_7 - 5\psi_1^3 \mathbf{E}_2 \phi_4 - 20 \phi_1^2 \psi_7 - 10 \phi_1^2 \psi_1 \psi_2 \phi_4 + 36 \phi_1^4 \psi_1 \phi_4; \\ R_{11,1} &= 500\phi_4 \psi_7 -125\phi_1\phi_2^2 \mathbf{E}_2^3 + 27000\phi_1\phi_2^5 -1000\phi_1^2\psi_1^5 \mathbf{E}_2^2 + 108000\phi_1^2 \psi_1^7 \psi_2 + 125\phi_1^3 \mathbf{E}_2^4 \\ &\quad\quad + 15750\phi_1^3 \phi_2^3 \mathbf{E}_2 -2500\phi_1^3 \psi_1^2 \psi_2 \mathbf{E}_2^2 + 358400\phi_1^3 \psi_1^6 \mathbf{E}_2 -1900\phi_1^4 \psi_1 \mathbf{E}_2^3 + 439000\phi_1^4 \psi_1^3 \psi_2 \mathbf{E}_2 \\ &\quad\quad + 314360\phi_1^4 \psi_1^7 + 50\phi_1^5 \phi_2 \mathbf{E}_2^2 + 236200\phi_1^5 \psi_1^2 \mathbf{E}_2^2 + 1658300\phi_1^5 \psi_1^4 \psi_2 + 1725760\phi_1^6 \psi_1^3 \mathbf{E}_2 \\ &\quad\quad + 1554640\phi1^7 \psi_2 \mathbf{E}_2 + 2571775\phi_1^7 \phi_2^2 + 1182416\phi_1^7 \psi_1^4 + 6680424\phi_1^8 \psi_1 \psi_2 -3375\phi_1^9 \mathbf{E}_2 + 24300\phi_1^{10} \psi_1; \end{align*} \begin{align*} R_{14,\chi} &= 10000\psi_7^2 -625\phi_2^4 \mathbf{E}_2^3 + 135000\phi_2^7 + 1250\phi_1^2 \phi_2^2 \mathbf{E}_2^4 -56250\phi_1^2 \phi_2^5 \mathbf{E}_2 -625\phi_1^4 \mathbf{E}_2^5 -78500\phi_1^4 \phi_2^3 \mathbf{E}_2^2 \\ &\quad\quad -40000\phi_1^4 \psi_1^6 \mathbf{E}_2^2 + 4320000\phi_1^4 \psi_1^8 \psi_2 -120000\phi_1^5 \psi_1^3 \psi_2 \mathbf{E}_2^2 + 16496000\phi_1^5 \psi_1^7 \mathbf{E}_2 -56250\phi_1^6 \phi_2 \mathbf{E}_2^3 \\ &\quad\quad + 16746875\phi_1^6 \phi_2^4 -129600\phi_1^6 \psi_1^2 \mathbf{E}_2^3 + 24728000\phi_1^6 \psi_1^4 \psi_2 \mathbf{E}_2 + 14302400\phi_1^6 \psi_1^8 + 18188000\phi_1^7 \psi_1^3 \mathbf{E}_2^2 \\ &\quad\quad + 81259200\phi_1^7 \psi_1^5 \psi_2 + 6064800\phi_1^8 \psi_2 \mathbf{E}_2^2 + 24373050\phi_1^8 \phi_2^2 \mathbf{E}_2 + 109220400\phi_1^8 \psi_1^4 \mathbf{E}_2 \\ &\quad\quad + 133829600\phi_1^9 \psi_1 \psi_2 \mathbf{E}_2 + 73829440\phi_1^9 \psi_1^5 + 16875\phi_1^{10} \mathbf{E}_2^2 + 424151360\phi_1^{10} \psi_1^2 \psi_2 \\ &\quad \quad + 229181280\phi_1^{11} \psi_1 \mathbf{E}_2 + 917697120\phi_1^{12} \phi_2 + 184414224\phi_1^{12} \psi_1^2.\end{align*}
Any relation of weight greater than $16$ can be reduced against the relations above to eliminate the variables $\phi_2,\psi_2,\phi_4,\psi_7$, and the remaining generators $\phi_1,\psi_1,\mathbf{E}_2$ are algebraically independent (because the restrictions of $\psi_1, \mathbf{E}_2$ to $T_3$ are already algebraically independent, and $\phi_1$ vanishes along $T_3$). Therefore the relations above are enough. Let $\mathfrak{e}_1,\mathfrak{e}_{\chi},\mathfrak{e}_{\chi^2}$ be multiplicative symbols which represent the symmetric characters of $SL_2(\mathcal{O}_K)$; then we get the Hilbert series $$\sum_{k=0}^{\infty} \sum_{i \in \mathbb{Z}/3\mathbb{Z}} \mathrm{dim}\, M_{k,\chi^i}(\Gamma_K) t^k \mathfrak{e}_{\chi^i} = \frac{1 + t^2 (\mathfrak{e}_{\chi} + \mathfrak{e}_{\chi^2}) - t^3 + t^4 (\mathfrak{e}_{\chi} + \mathfrak{e}_{\chi^2}) - t^5 \mathfrak{e}_{\chi} + 2t^6 + t^6 \mathfrak{e}_{\chi^2} - t^7 \mathfrak{e}_{\chi}}{(1-t \mathfrak{e}_{\chi}) (1 - t \mathfrak{e}_{\chi^2})(1-t^2)}.$$ (This can be computed in Macaulay2, for example; the exponents and characters in the denominator come from $\phi_1,\psi_1,\mathbf{E}_2$.) With a bit of algebra (e.g. replacing $\frac{1}{1 - t \mathfrak{e}_{\chi}}$ by $\frac{1 +t \mathfrak{e}_{\chi} + t^2 \mathfrak{e}_{\chi^2}}{1 - t^3}$) it follows that the dimensions of Hilbert modular forms with trivial character have the generating series $$\sum_{k=0}^{\infty} \mathrm{dim}\, M_k(\Gamma_K) t^k = \frac{1 - t + t^2 + t^3 + t^4 + t^5 + t^6 - t^7 + t^8}{(1 - t)(1-t^2)(1-t^3)}.$$
\end{proof}
\begin{prop} The graded ring of modular forms $M_*(\Gamma_K)$ is presented by the $15$ generators $$\mathbf{E}_2, \phi_1 \psi_1, \phi_1^3, \phi_1 \psi_2, \psi_1^3, \phi_2 \psi_2, \psi_1^2 \psi_2, \psi_1 \phi_4, \phi_1 \phi_2^2, \phi_2^3, \psi_2 \phi_4, \phi_1^2 \phi_4, \phi_1 \phi_2 \phi_4, \phi_2^2 \phi_4, \phi_2 \psi_7$$ of weights $2,2,3,3,3,4,4,5,5,6,6,6,7,8,9$ and by 77 relations in weights 6 through 18.
\end{prop}
A number of the 77 defining relations are those that are implied by the notation, e.g. $(\phi_1 \psi_1)^3 = \phi_1^3 \cdot \psi_1^3$.
\begin{proof} As generators we can take the monomials in $\mathbf{E}_2,\phi_1,\psi_1,\phi_2,\psi_2,\phi_4,\psi_7$ that have trivial character (and that cannot be further split into monomials with trivial character). Many of these monomials turn out not to be necessary due to relations such as $\psi_1 \phi_2 = \phi_1 \psi_2$. The 15 given above are minimal in that sense. \\
The ideal of relations is given by intersecting the ideal of relations from the previous lemma with the subring of Hilbert modular forms with trivial character. Finding this intersection and minimal generators for it is a straightforward Gr\"obner basis computation which was done in Macaulay2. We leave the explicit relations to the supplementary material.
\end{proof}
\bibliographystyle{plainnat}
|
\section{Introduction}
The ultrastrong-coupling (USC) between an effective two-level system (TLS) and the electromagnetic field has been realized in several solid state systems \cite{Kockum2018,Forn-Diaz2018}. In this regime of quantum light-matter interaction, going beyond weak and strong coupling, the coupling strength becomes comparable to the transition frequencies in the system. Recently, circuit QED experiments involving a single LC-oscillator mode
coupled to a flux qubit superconducting quantum circuit have realized coupling strengths larger then the system transition frequencies \cite{Yoshihara2017,Yoshihara2017b}. This extreme interaction regime has been denoted deep strong coupling (DSC).
In these interaction regimes, several properties of coupled light-matter
systems change drastically, opening the way to a wealth of new intriguing physical effects (see, e.g. \cite{Ashhab2010,Casanova2010,Niemczyk2010,Ridolfo2012,Ridolfo2013,Stassi2013,Cirio2016, DeLiberato2009, DeLiberato2014,Garziano2015,Garziano2016, Garziano2017,DiStefano2017,Kockum2017a,Felicetti2018}) which may offer opportunities for the development of quantum technologies \cite{Nataf2010,Nataf2011,Romero2012,Kyaw2015,Wang2016,Stassi2017,Kockum2017, Armata2017,Stassi2018}.
The form of the electron-photon interaction
is gauge dependent (see, e.g., \cite{Babiker1983}). However, all physical
results must be independent of this choice.
Gauge invariance is a general guiding principle in
building the theory of fundamental interactions (see, e.g., \cite{Aitchison1989, Maggiore2005}).
Let us consider, for example, a particle field whose action is invariant under a global phase change [$U(1)$ invariance]. If this phase is allowed to depend on the space-time coordinate $x$, its action is not invariant because the coordinate-dependent phase $\phi(x)$ does not commute with the derivatives (corresponding to the four-momentum) in the action. The symmetry can be restored replacing these derivatives with covariant derivatives: $D_\mu = (\partial_\mu + i q A_\mu)$, where $q$ is the charge parameter and $A_\mu$ is the gauge potential. If the transformation of the particle field $\psi \to e^{i \phi(x)}$ is associated to a gauge transformation of the four-vector potential $A_\mu \to A_\mu - \partial_\mu \theta$, the local $U(1)$ symmetry is restored.
The coupling to the field, obtained performing the replacement $\partial_\mu \to D_\mu$ in the free Lagrangian, is called the ``minimal coupling''.
It has been shown by several authors \cite{Lamb1952,Lamb1987,Starace1971, Girlanda1981,Ismail-Beigi2001} that approximate models for light-matter interactions
derived in different gauges may lead to different predictions. It has also been shown that the convergence of the two-photon transition rate depends significantly on the choice of the gauge \cite{Bassani1977}.
When the light-matter interaction becomes very strong, different gauges can lead to drastically different predictions, leading to controversies \cite{Hepp1973, Wang1973, Rza1975,Lambert2004, Keeling2007,Nataf2010a,Vukics2014,Grie2016,Bamba2016b,DeBern2018}.
For example, in the case of several TLSs interacting with a single mode of an optical resonator \cite{Dicke}, different gauges may even lead to
very different predictions, such as the presence or the absence of a
quantum phase transition.
One important conclusion that can be drawn
from these controversies is that, once the light-matter coupling becomes non-perturbative, the validity of the two-level
approximation for the atomic dipoles depends explicitly
on the choice of gauge. In particular, it has been shown \cite{DeBernardis2018} that
in the electric dipole gauge (the multipolar gauge in the dipole approximation) the two-level approximation can be performed as long as the Rabi
frequency remains much smaller than the energies of the off-resonant higher-energy levels. However, it can dramatically fail
in the Coulomb gauge, even for systems with an extremely anharmonic spectrum \cite{DeBernardis2018}. This unexpected result is particularly unsatisfactory, since the general procedure to derive the multipolar gauge consists of using the minimal coupling replacement first, and then applying the Coulomb gauge \cite{Babiker1983}.
In a recent work \cite{Stokes2018}, considering arbitrary gauges for the interaction between a TLS and the electromagnetic field, it has been shown that physical predictions in the USC and DSC regimes strongly depend on the gauge choice. The gauge which provides the more accurate results seems to depend on the detuning between the cavity mode resonance and the atomic transition frequency.
From all of these previous studies, it is clear that approximations in the description of the matter system, as e.g., a finite-level truncation, ruin the gauge invariance of the theory. In 1970, Ref.~\cite{Starace1971} pointed out that gauge ambiguities in the calculation of atomic oscillator strengths, originate from the occurrence of nonlocal potentials determined by the approximation procedures.
Since a {\em nonlocal} potential in the coordinate representation is an integral operator, it does not commute with the coordinate operator. Indeed, it is easy to show that it can be expressed as a {\em local} momentum-dependent operator $V (\hat {\bf r}, \hat {\bf p})$. This affects the interaction of light with quantum systems described by approximate Hamiltonians. Specifically, in order to introduce the coupling of the matter system with the electromagnetic field, the minimal replacement rule $\hat {\bf p} \to \hat {\bf \Pi} = \hat {\bf p} - \hat {\bf A}( {\bf r},t)$ ($\hat {\bf A}$ is the vector potential) has to be applied not only to the kinetic energy terms, but also to the {\em nonlocal potentials} in the effective Hamiltonian of the particles in the system.
By applying such a procedure, the ambiguities in the calculation of approximate matrix elements for electric dipole transitions can be removed \cite{Starace1971}.
Moreover, taking into account the nonlocality of the approximate potential, two-photon transition rates involving Wannier excitons in semiconductors become gauge invariant \cite{Girlanda1981}.
Also the microscopic quantum theory of excitonic polaritons is affected by the presence of nonlocal potentials. The application of the standard minimal coupling replacement leads to a total Hamiltonian which, besides the usual $\hat {\bf A} \cdot \hat {\bf p}$ term, displays an additional diamagnetic term $ e^2 \hat{\bf A}^2/ 2m$, where $e$ and $m$ are the electron charge and mass respectively.
If a limited number of exciton levels are explicitly included in the model, the long wavelength solution of the polariton dispersion cannot be recovered without introducing {\em ad hoc} the Thomas-Reiche-Kuhn sum rule and truncating the summation consistently \cite{Combescot1991}. This procedure presents some ambiguity and appears to be rather artificial. However, if the nonlocality of the approximate potential is taken into account and the resulting additional terms are included, up to second order in the vector potential, the correct dispersion relation is automatically recovered \cite{Savasta1995,Savasta1996}.
In summary, the concept of approximation-induced nonlocal potentials, together with an expansion of the interaction Hamiltonian up to second order in the vector potential, was able to overcome a number of gauge ambiguities.
Here we investigate if this strategy can work in the maximally truncated Hilbert space provided by a TLS, and in the nonperturbative regimes of cavity QED. This investigation is relevant not only in order to remove gauge ambiguities in quantum optical systems which are attracting great interest, but also provides a general insight on gauge invariance in extreme interaction regimes.
We find that {\em the usual strategy}, which consists of taking into account the nonlocality of the atomic potential, performing the minimal coupling replacement and developing the resulting interaction Hamiltonian up to second order in the vector potential, {\em fails when the coupling strength reaches a significant fraction of the resonance frequencies of the system}.
We demonstrate that these gauge ambiguities can be eliminated for arbitrary coupling strengths only by taking into account the approximation-induced nonlocality and keeping the resulting interaction Hamiltonian to all orders in the vector potential $\hat {\bf A}$. The results presented here solve all the long-lasting controversies arising from gauge ambiguities in the quantum Rabi and Dicke models.
\section{Results}
\subsection{Nonlocal potentials}
In order to understand why local potentials become nonlocal when the Hilbert space is truncated, let us consider a one-dimensional potential $\hat V$. In the position basis, it can written as
\begin{equation}
\hat V = \int dx dx'\, \langle x | \hat V | x' \rangle\, \ketbra{x}{x'}
\end{equation}
If the potential is local, its matrix elements can be written as $\langle x | \hat V | x' \rangle \equiv V(x,x') = W(x) \delta (x-x')$. If ${|n \rangle}$ is a complete orthonormal basis, the matrix elements can be expressed as $V(x,x') = W(x) \delta (x-x')=\sum_{n,n'} W_{n,n'} \psi^*_{n'}(x') \psi_n(x)$, where we defined $\psi_n (x) \equiv \langle x|n \rangle$. Notice that the Dirac delta function can be reconstructed only by keeping all the infinite vectors of the basis. Hence, any truncation of the complete basis can transform a local potential into a nonlocal one. If only two states are included, e.g., the two lowest energy levels, we obtain
\begin{equation}\label{nonlocality}
V(x,x') = W_{1,0}\left[ \psi_0^*(x') \psi_1(x) + \psi_1^*(x') \psi_0(x) \right]
\end{equation}
where, for simplicity, we assumed parity symmetry and real matrix elements.
It is evident that adding the two terms in Eq.(\ref{nonlocality}), which are products of two smooth wavefunctions it is not possible to reproduce the Dirac-delta function and this will result into a potential with an high degree of spatial nonlocality.
It has been shown by several authors \cite{Starace1971,Girlanda1981,Ismail-Beigi2001} that a nonlocal potential can be expressed as a momentum-dependent operator $ V( \hat{\bf r}, \hat {\bf p})$. Indeed, by using the translation operator $\psi(x') = \exp{[i(x'-x) \hat p ]} \psi (x)$, where $\hat p$ is the momentum operator, we obtain
\begin{equation}
\int V(x,x') \psi(x') dx' =V(x, \hat p) \psi(x)\, .
\end{equation}
\subsection{The dipole gauge}
We consider a single electric dipole with charge $q$ coupled to a single mode of the electromagnetic field. The field mode is described by a harmonic oscillator with
bare frequency $\omega_c$ and annihilation and creation operators
$\hat a$ and $\hat a^\dag$. In one dimension, the dipole can be
modeled as an effective particle of mass $m$ in a potential $V(x)$. Its unperturbed Hamiltonian is $\hat H_0 = \hat p^2/({2 m}) + \hat V(x)$.
Applying the dipole approximation, the total Hamiltonian in the dipole gauge is
\begin{equation}\label{fulldipole}
\hat H_D = \hbar \omega_c \hat a^\dag \hat a + \hat H_0
+ \frac{q^2 A_0^2 \omega_c}{\hbar} \hat x^2 +i q \omega_c \hat x A_0 (\hat a^\dag - \hat a)\, ,
\end{equation}
where $A_0$ is the zero point fluctuation amplitude of the vector potential $\hat A = A_0 (\hat a + \hat a^\dag)$.
If the two lowest-energy levels of the effective quantum particle are well separated from the higher energy levels, as in the case of flux qubits \cite{Yoshihara2017}, and if the detuning $\delta = \omega_{10} - \omega_c$ ($\omega_{10}$ is the transition frequency of the two lowest energy levels) is much smaller than the detunings with respect to other transitions,
it is possible to truncate the Hilbert space considering only the two lowest-energy levels. In this case, the unperturbed particle Hamiltonian reduces to $\hat {\cal H}_0 = \hbar \omega_{10} \hat \sigma_z /2$, and the total Hamiltonian becomes
\begin{equation}\label{RabiD}
\hat {\cal H}_D = \hbar \omega_c \hat a^\dag \hat a + \frac{\hbar \omega_{10}}{2} \hat \sigma_z + i \hbar g_D (\hat a^\dag - \hat a) \hat \sigma_x \,
\end{equation}
where $g_D = \omega_c A_0 d_{10}/ \hbar$ and $d_{10} \equiv q \langle 1|\hat x | 0\rangle$ is the dipole matrix element. In Eq.~(\ref{RabiD}) we dropped the constant term ${\cal C} = {d^2_{10} \omega_c A_0^2}/{\hbar}$, which results from the expectation value $\langle i|\hat x^2 | j \rangle$ by using the TLS identity operator $\hat I_2 = |0\rangle \langle 0| + |1\rangle \langle 1|$. This term could be calculated more accurately by including it in the particle potential, before the diagonalization. However, we made the choice of considering the interaction terms only after the Hilbert space truncation. De Bernardis et al.~\cite{DeBernardis2018} have shown that, for an effective particle displaying the two lowest-energy levels well separated from the higher energy levels ($\omega_{21} \gg g_D$), the two level approximation provides results which are in agreement with those obtained using the full Hamiltonian in Eq.~(\ref{fulldipole}), even for extreme coupling strengths ($g_D \gg \omega_{10}$).
\subsection{Derivation of the quantum Rabi model in the Coulomb gauge}
In the Coulomb gauge, the Hamiltonian describing the interaction of an effective quantum particle with the electromagnetic field can be obtained applying the substitution $\hat p \to \hat p - q \hat A$ to the particle Hamiltonian $\hat H_0$. The well-known result is
\begin{equation}\label{fullRabi}
\hat H_C = \hbar \omega_c \hat a^\dag \hat a + \hat H_0
- i q \hat p A_0 (\hat a^\dag + \hat a) + \frac{q^2 A_0^2}{2 m} (\hat a^\dag + \hat a)^2\, .
\end{equation}
Projecting $\hat H_C$ in a two-level space, we obtain
\begin{equation}\label{RabiC}
\hat {\cal H}'_C = \hbar \omega_c \hat a^\dag \hat a + \frac{\hbar \omega_{10}}{2} \hat \sigma_z + \hbar g_C \sigma_y (\hat a^\dag + \hat a)
+ \frac{q^2 A_0^2}{2 m} (\hat a^\dag + \hat a)^2\, ,
\end{equation}
where $g_C = g_D\, \omega_{10}/\omega_c$.
\begin{figure}
\centering
\includegraphics[width= 0.95 \linewidth]{fig1.pdf}
\caption{Comparison of the energy
spectra as a function of the normalized coupling $\eta= g_D /\omega_{c}$, obtained from the quantum Rabi Hamiltonians in the dipole gauge ($\hat {\cal H}_D$), in the Coulomb gauge (standard derivation, $\hat {\cal H}'_C$), and in the Coulomb gauge taking into account the presence of nonlocal potentials ($\hat {\cal H}_C$). The plots in (a) have been obtained using a detuning $\delta=0$; the plots in (b) using $\delta= 0.5 \omega_c$.
\label{fig:1}}
\end{figure}
The diamagnetic term $q^2 \hat A^2/(2m)$ can be absorbed by using a Bogoliubov transformation involving only the photon operators \cite{Yoshihara2017, DeBernardis2018}.
In Fig.~\ref{fig:1} we plot the energy differences ($E_n- E_0$) for the lowest eigenstates
of $\hat {\cal H}_D$, $\hat {\cal H}'_C$, and $\hat {\cal H}_C$ [see Eq.~(\ref{SSC2})], as a function of the normalized coupling $\eta= g_D /\omega_{c}$. Figure~1a has been obtained at zero detuning ($\delta=0$), while we used $\delta = 0.5\, \omega_c$ for Fig.~1b. The comparison in Fig.~\ref{fig:1} shows that, for very small values of the coupling, the eigenvalues of the different Hamiltonians reproduce the expected behaviour.
However, already at moderate coupling strengths, $\eta \sim 0.1$,
there are significant deviations in the predicted energies, in agreement with the results obtained in Ref.~\cite{DeBernardis2018}. For values of $\eta \gtrsim 0.5$ the differences become dramatic. The analysis carried out in Ref.~\cite{DeBernardis2018} clarifies that the disagreement between the eigenvalues of $\hat {\cal H}_D$ and $\hat {\cal H}'_C$ is due to the failure of the Rabi model to provide correct results beyond small coupling values in the Coulomb gauge ( $\hat {\cal H}'_C$).
However, this derivation of the quantum Rabi model in the Coulomb gauge does not take into account that, in the presence of a truncated Hilbert space, the particle potential can loose its locality $\hat V(\hat x) \to \hat V'(\hat x, \hat p)$. In this case, in order to preserve gauge invariance, one has to apply the substitution $\hat p \to \hat p - q \hat A$ {\em also} to the potential. In principle, this procedure can give rise to additional terms in the interaction Hamiltonian to all orders in the vector potential. These higher order terms are expected to be negligible for small normalized couplings $\eta$, but can become relevant at higher coupling strengths.
As shown in detail in the Supplementary Note~\textcolor{blue}{1}, using some general operator theorems \cite{Louisell1973}, it is possible to apply the minimal coupling replacement to both the kinetic energy and the nonlocal potential of the effective Hamiltonian of a quantum particle by employing a unitary transformation \cite{Savasta1995}:
\begin{equation}\label{SSC}
\hat H_C = \hat U \hat H_0 \hat U^{\dag} + \hat H_{\rm ph}\, ,
\end{equation}
where $\hat H_{\rm ph} = \hbar \omega_c \hat a^\dag \hat a$.
In the dipole approximation, the unitary operator is
\begin{equation}\label{U}
\hat U = \exp \left[ i \frac{q \hat x \hat A}{\hbar}\right]\, .
\end{equation}
\begin{figure}
\centering
\includegraphics[width= 0.95 \linewidth]{fig2.pdf}
\caption{Comparison of the energy
spectra as a function of the normalized coupling $\eta= g_D /\omega_{c}$, obtained from the quantum Rabi Hamiltonians in the dipole gauge ($\hat {\cal H}_D$), in the Coulomb gauge (standard derivation, $\hat {\cal H}'_C$), and in the Coulomb gauge taking into account the presence of nonlocal potentials ($\hat {\cal H}_C$). The plots in (a) have been obtained using a detuning $\delta=0$; the plots in (b) using $\delta= 0.5 \omega_c$.
\label{fig:2}}
\end{figure}
If the two-level approximation is used, in order to include the interaction, one has to consistently project Eq.~(\ref{SSC}) in the two-level space before performing the unitary transformation of $\hat {\cal H}_0$:
\begin{equation}\label{SSC2}
{red}{
\hat {\cal H}_C = \hat {\cal U} \hat {\cal H}_0 \hat {\cal U}^{ \dag} + \hat H_{\rm ph}\, ,}
\end{equation}
where the projected unitary operator is \begin{equation}
\hat {\cal U} = \exp \left[ i \eta \hat \sigma_x (\hat a + \hat a^\dag)\right]\, .
\end{equation}
{Note that the standard procedure consists of applying the minimal coupling replacement to the exact matter Hamiltonian $\hat H_0$ as, e.g., in Eq.~(\ref{SSC}), and then to project the resulting Hamiltonian in the truncated Hilbert space:
\begin{equation}\label{Rabl}
\hat {\cal H}'_C = \hat P \hat U \hat H_0 \hat U^{\dag}\hat P + \hat H_{\rm ph}
= \hat P \left[ \hat U \frac{\hat p^2}{2m} \hat U^{\dag}
+ \hat V \right] \hat P + \hat H_{\rm ph}
\, ,
\end{equation}
where
$\hat P$ is the projection operator for the truncated Hilbert space, and we used the relation $\hat U \hat V \hat U^\dag = \hat V$, valid when the potential is local. Note that, in contrast to Eq.~(\ref{SSC2}), Eq.~(\ref{Rabl}) contains a nonlocal potential $\hat P \hat V \hat P$ to which the gauge principle has not been applied.
This analysis explains why the standard approach fails when applied to a matter system described in a truncated Hilbert space.}
Of course, this procedure can be easily generalized to any truncated Hilbert space. It is remarkable that it takes automatically into account the presence of nonlocal potentials, even without knowing them explicitly.
Noticing that the operator $\hat{\cal U}$ corresponds to a spin rotation along the $x$ axis, we obtain
\begin{equation}\label{CRabi}
{red}{\begin{aligned}
\hat {\cal H}_C &= \hbar \omega_c \hat a^\dag \hat a + \frac{\hbar \omega_{10}}{2} \\ &\left\{ \hat \sigma_z \cos{\left[ 2 \eta (\hat a + \hat a^\dag)\right]} +
\hat \sigma_y \sin{\left[ 2 \eta (\hat a + \hat a^\dag)\right]} \right\}
\, .
\end{aligned}}
\end{equation}
This is the correct Hamiltonian in the Coulomb gauge of the quantum Rabi model. The price one has to pay for preserving the gauge principle in such a truncated space is that the resulting Hamiltonian will contain field operators at all orders. The generalization to multimode fields \cite{Male2017,Gely2017,Munoz2018} is straightforward. When the normalized interaction strength is sufficiently weak, it is possible to expand the trigonometric functions up to the linear or quadratic terms. However, when $g_D$ becomes comparable with $\omega_c$ (i.e., the USC regime), such expansion is not sufficient. When approaching, or reaching, the deep strong coupling regime ($\eta \gtrsim 1$) the field terms have to be included at all orders.
Figure~(\ref{fig:1}) shows that the numerically calculated energy spectrum of $\hat {\cal H}_C$ coincides with that of $\hat {\cal H}_D$. Hence we can conclude that, {\em when the presence of nonlocal potentials, owing to the two-level truncation, is properly taken into account, it is possible to obtain correct results also in the Coulomb gauge}.
Equation~(\ref{SSC2}) can be easily generalized beyond the single-mode approximation. Expanding Eq.~(\ref{SSC2}) up to the second order in the potential, we obtain
\begin{equation}\label{SSC22}
\hat {\cal H}^{(2)}_C = \hbar \omega_c \hat a^\dag \hat a
+ \frac{\hbar \omega_{10}}{2} \hat \sigma_z + \hbar g_C \hat \sigma_y (\hat a + \hat a^\dag)+
\frac{\hbar g^2_C}{\omega_{10}} \hat \sigma_z (\hat a + \hat a^\dag)^2\, .
\end{equation}
This equation differs from the standard quantum Rabi Hamiltonian in the Coulomb gauge $\hat {\cal H}'_C$ for the diamagnetic term, which now displays a different coefficient and depends on $\hat \sigma_z$. By introducing the Thomas-Reiche-Kuhn sum rule in the diamagnetic term of $\hat {\cal H}'_C$ and applying the two-level truncation \cite{Stokes2018}, $\hat {\cal H}'_C$ turns into $\hat {\cal H}^{(2)}_C$. However, this procedure is quite arbitrary. Figure~(\ref{fig:2}) compares the energy spectra of $\hat {\cal H}'_C$, $\hat {\cal H}^{(2)}_C$, and $\hat {\cal H}_C$.
In particular, we observe that for values of the normalized coupling below $0.1$ the spectra of both $\hat {\cal H}^{(2)}_C$ (red lines) and $\hat {\cal H}'_C$ (blue dashed lines) are in quite good agreement with the exact results (expecially those of $\hat {\cal H}^{(2)}_C$) given by $\hat {\cal H}_C$ (black dotted lines). Increasing the coupling, $\hat {\cal H}'_C$ gives spectra which rapidly become very different from the exact energy levels. The spectra of $\hat {\cal H}^{(2)}_C$ provide an improved approximation, which, for the lowest few levels, is acceptable for $\eta \lesssim 0.2$. The range of values of $\eta$ where the approximate models provide acceptable spectra reduces at nonzero detuning [Fig.~(\ref{fig:2}) (b)].
In order to understand how many powers of the photon operators have to be included in $\hat {\cal H}_C$ to obtain correct spectra at higher $\eta$, we compared in Fig.~3 the approximate spectra, calculated from different $n$-order Taylor expansions $\hat {\cal H}^{(n)}_C$ of $\hat {\cal H}_C$, with the exact spectra (the eigenvalues of $\hat {\cal H}_C$ ). The results are interesting: for $n=3$ there is already a significant improvement (with respect to $n=2$), up to $\eta \lesssim 0.25$. However, the spectra become completely wrong already at $\eta \sim 0.3$. Accuracy improves for $n = 10$, but only up to $\eta \lesssim 0.25$. For $n = 200$, there is an excellent agreement, but only for $\eta \lesssim 1.3$. These results show that for values of $\eta$ larger than $1$ (DSC), a very large $n$ is needed to get the correct spectra. However, further increasing $\eta$ requires the inclusion of more and more terms in the expansion. This confirms the {\em nonperturbative spatial nonlocality} which occurs when heavily truncating the particle's Hilbert space.
Specific results for circuit-QED systems can be obtained analogously. For example, the full Hamiltonian of a fluxonium capacitively coupled to an LC oscillator circuit \cite{Manucharyan2009} (corresponding to the charge gauge) can be obtained through an analogous minimal coupling replacement. As shown in the Supplementary Note~\textcolor{blue}{2}, the resulting total Hamiltonian for the two-level model in the charge gauge is very similar to Eq.~(\ref{SSC2}).
\subsection{Resolution of gauge ambiguities}
The Hamiltonians $\hat H_D$ and $\hat H_C$ are related by a gauge transformation \cite{Babiker1983}, which can be expressed by the unitary transformation $\hat H_C = \hat {G} \hat H_D \hat {G}^\dag$. In the dipole approximation $\hat {G} = \hat U$ [see Eq.~(\ref{U})].
Applying this gauge transformation, it is easy to obtain the quantum Rabi Hamiltonian in the Coulomb gauge Eq.~(\ref{fullRabi}) from
Eq.~(\ref{fulldipole}). Of course, the inverse relationship also holds: $\hat H_D = \hat {G}^\dag \hat H_C \hat {G}$.
As discussed above, it has been shown that $\hat {\cal H}'_C$ (the two-level approximation of $\hat {H}_C$) gives rise to wrong spectra, thus ruining gauge invariance.
In the previous subsection, we have demonstrated that the correct quantum Rabi Hamiltonian in the Coulomb gauge, based on the complete introduction of the gauge-preserving replacement $\hat {\bf p} \to \hat {\bf \Pi} = \hat {\bf p} - \hat {\bf A}(\hat {\bf r},t)$, is $\hat {\cal H}_C$ in Eq.~(\ref{CRabi}).
Numerical results in Fig.~\ref{fig:1} show that
$\hat {\cal H}_C$ gives correct spectra, thus
providing clear indications that the procedure developed here restores gauge invariance in TLSs.
We now present an analytical demonstration of the gauge invariance of a TLS coupled to the electromagnetic field.
We show that gauge invariance in a truncated Hilbert space can be fully preserved if the gauge transformation is consistently performed. We start from ${\cal H}_D$, which was shown to be a very good approximation of the full Hamiltonain $H_D$, and apply the gauge transformation projecting the unitary operator $\hat G = \hat U$ in the two-level space:
$
\hat {\cal G} \hat {\cal H}_D \hat {\cal G}^{\dag}\, ,
$
where
$\hat {\cal G} = \hat {\cal U}$. The result of this unitary transformation should be $\hat {\cal H}_D \to \hat {\cal H}_C$.
Noticing that the operator $\hat {\cal U}$ corresponds to a spin rotation along the $x$ axis, and using the Baker-Campbell-Hausdorff lemma, it is easy to obtain
\begin{equation}\label{TLgauge}
{red}{\begin{aligned}
\hat {\cal G} \hat {\cal H}_D \hat {\cal G}^{\dag} &= \hbar \omega_c \hat a^\dag \hat a + \frac{\hbar \omega_0}{2} \left\{ \hat \sigma_z \cos{\left[ 2 \eta (\hat a + \hat a^\dag)\right]} \right.
\\ &+ \left.
\hat \sigma_y \sin{\left[2 \eta (\hat a + \hat a^\dag)\right]} \right\} = \hat {\cal H}_C\, .
\end{aligned}}
\end{equation}
This equation shows that indeed $\hat {\cal G} \hat {\cal H}_D \hat {\cal G}^{\dag} = \hat {\cal H}_C $. This result demonstrates that gauge invariance is preserved in a two-level truncated space, if the gauge transformation is applied consistently, and if we use $\hat {\cal H}_C$ instead of $\hat {\cal H}'_C$.
\begin{figure*}
\centering
\includegraphics[width= \linewidth]{fig3.pdf}
\caption{Energy spectra as a function of the normalized coupling $\eta$, obtained from the quantum Rabi Hamiltonians in the Coulomb gauge taking into account the presence of nonlocal potentials ($\hat {\cal H}_C$).
The panel shows the comparison between the exact (containing all terms) (black continuous curve, see also the bottom-right panel) and approximated energy levels (red curves). The calculations are performed at different order of approximation. The value $n$ indicates the order of the Taylor expansion used in the approximated Hamiltonian.
\label{fig:3}}
\end{figure*}
Following Ref.~\cite{Stokes2018}, it is possible to employ a formulation in which the gauge freedom is contained within a single real continuous parameter $\alpha$, which determines the gauge through a function $X_\alpha$.
The general gauge transformation in the dipole approximation is generated by a unitary transformation determined by the unitary operators
$
\hat U_\alpha = \exp \left[ i \hat X_\alpha\right]$,
where $\hat X_\alpha=\alpha q \hat x \hat A/{\hbar}$.
The values $\alpha= 0,1$ specify the dipole and the Coulomb gauge, respectively.
The two-level projected unitary operator is $\hat {\cal U}_\alpha = \exp \left[ i \hat \chi_\alpha\right]$, where $\hat \chi_\alpha=\alpha ({g_D}/{\omega_c}) \hat \sigma_x (\hat a + \hat a^\dag)$.
The correct $\alpha$-gauge Hamiltonian for a TLS is thus
\begin{equation}\label{Ualpha}
\hat {\cal H}^{(\alpha)} = \hat {\cal U}_\alpha \hat {\cal H}_D \hat {\cal U
}^{\dag}_\alpha\, .
\end{equation}
We obtain
\begin{eqnarray}\label{Halpha}
&&\hat {\cal H}^{(\alpha)} = \hbar \omega_c \hat a^\dag \hat a -i(1-\alpha) g_D (\hat a - \hat a^\dag)\sigma_x\\ &&+ \frac{\hbar \omega_0}{2} \left\{ \hat \sigma_z \cos{\left[2 \alpha \eta (\hat a + \hat a^\dag)\right]} +
\hat \sigma_y \sin{\left[2 \alpha \eta (\hat a + \hat a^\dag)\right]} \right\} \nonumber .
\nonumber
\end{eqnarray}
Since, the transformation in Eq.~(\ref{Ualpha}) is unitary, the Hamiltonian (\ref{Halpha}) will have the same energy spectra of $\hat {\cal H}_D = \hat {\cal H}^{(\alpha=0)}$ and of $\hat {\cal H}_C = \hat {\cal H}^{(\alpha=1)}$ for any value of $\alpha$. {\em This eliminates any gauge ambiguity of the quantum Rabi model}.
\subsection{Derivation of the Dicke model in the Coulomb gauge}
We now extend the results obtained above for a single two-level dipole (Rabi) to the multi-dipole case (Dicke) \cite{Hepp1973, Shammah2017, DeBern2018}. The standard Dicke Hamiltonian in the Coulomb gauge can be written as
\begin{eqnarray} \label{Heff}
\begin{split}
\hat {\cal H}^{\prime N}_{C} & =\hbar \omega_{c} \hat a^{\dag}\hat a +
\hbar \omega_{10} \hat J_z\\
& + 2 \hbar g_c \left(\hat a^{\dag}+\hat a \right) \hat J_y
+ j\frac{q^{2} A_0^{2}}{m} \left(\hat a^{\dag}+\hat a \right)^{2}\, ,
\end{split}
\end{eqnarray}
where the number $N$ of dipoles determines the effective main angular momentum quantum number $j = N/2$, and $2 \hat J_i=\sum_{k=1}^N \hat \sigma_{i}^{(k)}$.
Similarly to the quantum Rabi Hamiltonian, it is possible to introduce the coupling between the matter Hamiltonian (in this case ${\cal H}^N_0 = \hbar \omega_{10} \hat J_z$) and the electromagnetic field, by the minimal coupling replacement. In the presence of nonlocal potentials, this can be done by applying the unitary transformation
\begin{equation}\label{DickeU}
\hat {\cal H}^{N}_{C} = \hat {\cal U}_N \hat {\cal H}_0^N {\cal U}^\dag_N
+ \hat H_{\rm ph}\, ,
\end{equation}
where
\begin{equation}
\hat {\cal U}_N =\exp \! \left [i 2 \eta (\hat a^{\dag}+\hat a) \hat J_x \right ]
\end{equation}
is the projection of
$\hat U_N = \exp \! \left[ i {q \sum_k \hat x_k \hat A}/{\hbar}\right]$ in a reduced Hilbert space, which is the tensor product of $N$ two-dimensional Hilbert spaces. We obtain
\begin{equation} \label{DickeC}
{red}{\begin{aligned}
\hat {\cal H}^{N}_{C} = \hbar \omega_{c} \hat a^{\dag}\hat a
&+ \hbar \omega_{10} \left \{\hat J_z \cos \left[ 4 \eta (\hat a^{\dag}+\hat a)\right] \right. \\&+ \left. \hat J_y \sin \left[ 4 \eta (\hat a^{\dag}+\hat a)\right]\right\}\, .
\end{aligned}}
\end{equation}
This is the correct Hamiltonian in the Coulomb gauge for the Dicke model. Its gauge invariance can be easily proved following the same procedure shown in the previous subsection.
{\em This result eliminates any gauge ambiguity} \cite{Keeling2007, Nataf2010a, DeBern2018} {\em of the Dicke model}.
{The Hamiltonian in Eq.~(\ref{DickeC}) is equivalent to the one in Eq.~(45) of Ref.~\cite{DeBern2018} (if direct dipole-dipole interactions are not included). The latter has been obtained applying a polaron transformation to the Dicke Hamiltonian in the dipole gauge. This unitary transformation, as observed by the authors ~\cite{DeBern2018}, is equivalent to a gauge transformation and hence gives rise to a Dicke Hamiltonian in the Coulomb gauge. However, as pointed out there, Ref.~\cite{DeBern2018} does not solve the gauge ambiguities of the Rabi and Dicke models (see also Ref.~\cite{DeBern2018,DeBernardis2018}). Indeed, their direct derivation of the Dicke Hamiltonian in the Coulomb gauge was obtained employing the standard minimal coupling substitution, instead of the correct generalized one proposed here [Eq.~(\ref{DickeU}]}.
\section{Discussion}
Approximate models for light-matter interactions derived in different gauges have produced drastically different predictions when the light-matter interaction strength reaches high values, such as the presence or absence of a quantum phase transition. {\em We succeeded to identify the origin of these gauge ambiguities recently pointed out}.
Specifically, we found that these ambiguities originate from the presence of potentials depending on both position and momentum (nonlocal potentials) in the effective Hamltonians describing the matter systems. This is an unavoidable consequence of the usual approximations (like Hilbert space truncation). In these cases, gauge invariance of the light-matter Hamiltonian can be preserved, even in the presence of extreme light-matter interactions, only by applying the minimal coupling replacement to {\em both} the kinetic and the potential terms of the effective Hamiltonian.
A careful application of the gauge principle generates a quantum Rabi Hamiltonian in the Coulomb gauge, which is significantly different from the standard quantum Rabi model. It contains field operators to all orders that cannot be neglected when the coupling strength is high. The derivation presented here is based on the fundamental minimal coupling replacement, and thus is not limited to the Coulomb gauge. We have also derived
arbitrary gauge descriptions (depending on a real parameter $\alpha$)
of cavity-QED Hamiltonians for TLSs. {\em These results resolve all the controversies arising from gauge ambiguities in the quantum Rabi and Dicke models.}
This analysis bring a deeper understanding of the
subtleties of the gauge principle, which cannot be ignored in the nonperturbative and extreme interaction regimes, which are now experimentally accessible. The lesson that can be learned is very general: when introducing interactions according to the gauge principle, any approximation in the description of a quantum system requires a {\em generalization of the minimal coupling replacement} in order to avoid inconsistencies and completely wrong predictions.
{We presented a general method for the derivation of gauge-invariant light-matter Hamiltonians in truncated Hilbert spaces.}
This approach can be extended to study ultrastrong and deepstrong light-matter interactions beyond the dipole approximation \cite{Ismail-Beigi2001}, where the multipolar gauge \cite{Babiker1983} is also affected by the presence of nonlocal potentials. A further interesting development could be the application of this analysis to many-body quantum systems strongly interacting with the electromagnetic field \cite{Cortese2017}.
\vspace{0.4 cm}
\acknowledgments{We acknowledge conversations with S. De Liberato and P. Rabl. FN is supported in part by the: MURI Center for Dynamic Magneto-Optics via the Air Force Office of Scientific Research (AFOSR) (FA9550-14-1-0040), Army Research Office (ARO) (Grant No. 73315PH), Asian Office of Aerospace Research and Development (AOARD) (Grant No. FA2386-18-1-4045), Japan Science and Technology Agency (JST) (the ImPACT program and CREST Grant No. JPMJCR1676), Japan Society for the Promotion of Science (JSPS) (JSPS-RFBR Grant No. 17-52-50023, and JSPS-FWO Grant No. VS.059.18N), RIKEN-AIST Challenge Research Fund, and the John Templeton Foundation.}
The authors declare no competing interests
\section*{Supplementary Note 1: Generalized minimal coupling replacement}
In this section, by using some general operator theorems \cite{Louisell1973}, we show how to implement the minimal coupling replacement on a generic operator $O(\hat x, \hat p)$ by a unitary transformation \cite{Savasta1995}.
Given two noncommuting operators $\hat \alpha$ and $\hat \beta$ and a parameter $\mu$ we want to calculate $e^{\mu \hat \beta} O(\hat \alpha) e^{-\mu \hat \beta}$.
The function $O(\hat \alpha)$ can be expanded in a power series:
\begin{equation}\label{expan}
O(\hat \alpha)=\sum_n c_n \hat \alpha^n\, .
\end{equation}
Using \eqref{expan}, we have
\begin{equation}
e^{\mu \hat \beta} O(\hat \alpha) e^{-\mu \hat \beta}=
\sum_n c_n e^{\mu \hat \beta} \hat \alpha^n e^{-\mu \hat \beta}\,.
\end{equation}
Observing that
\begin{eqnarray}
e^{\mu \hat \beta} \hat \alpha^n e^{-\mu \hat \beta}=\nonumber
&&e^{\mu \hat \beta} \hat \alpha e^{-\mu \hat \beta}
e^{\mu \hat \beta} \hat \alpha e^{-\mu \hat \beta}\\ \nonumber\ldots\,
&&e^{\mu \hat \beta} \hat \alpha e^{-\mu \hat \beta}=
\left(e^{\mu \hat \beta} \hat \alpha e^{-\mu \hat \beta}\right)^n\, .
\end{eqnarray}
We have:
\begin{equation}\label{opeq}
e^{\mu \hat \beta} O(\hat \alpha) e^{-\mu \hat \beta}=
\sum_n c_n \left(e^{\mu \hat \beta} \hat \alpha e^{-\mu \hat \beta}\right)^n=
O(e^{\mu \hat \beta} \hat \alpha e^{-\mu \hat \beta})\,.
\end{equation}
We now apply \eqref{opeq} to $ e^{i \chi(\hat{x})/\hbar} O(\hat x,\hat p)e^{-i \chi(\hat{x})/\hbar}$. For the sake of simplicity, we consider here the $\rm 1D$ case. Generalization to $\rm 3D$ is straightforward. We obtain
\begin{equation} \label{wowp}
e^{i \hat\chi( x)/\hbar} O(x,\hat p)e^{-i \hat\chi( x)/\hbar} = O( x,e^{i \hat\chi(x)/\hbar}\hat pe^{-i \chi(x)/\hbar})\, .
\end{equation}
Then, by using the Baker-Campbell-Hausdorff Lemma, we obtain:
\begin{eqnarray}\label{epe}
e^{i \hat \chi(x)/\hbar} \hat pe^{-i \hat\chi(x)/\hbar}&=& \hat p+ \frac{i}{\hbar} [ \hat\chi( x), \hat p]+ \nonumber \frac{1}{2}\left(\frac{i}{\hbar}\right)^2[\chi(\hat x),[ \chi(\hat x), \hat p]]+\\ &&\ldots= \hat p- \partial_x \hat\chi(x)\, ,
\end{eqnarray}
where we have used the result $[\hat \chi( x), \hat p]=i\hbar\partial_x \hat\chi(x)$.
In conclusion, using \eqref{epe}, \eqref{wowp} becomes
\begin{equation}\label{wow}
e^{i \hat\chi( x)/\hbar} O( x,\hat p)e^{-i \hat\chi( x)/\hbar}= O( x,\hat p-\partial_ { x} \hat\chi( x))\,.
\end{equation}
Considering now the special function
\begin{equation} \label{chi}
\hat\chi( x)= x \hat A_0\, ,
\end{equation}
where $\hat A_0 \equiv \hat A(x_0)$ is the field potential calculated at the atom position $x_0$, we obtain
\begin{equation} \label{chi}
\partial_ { x} \hat \chi( x)= \hat A_0\, .
\end{equation}
If we plug this result into Eq.~(\ref{wow}), we get
\begin{equation}\label{wow2}
e^{i \hat \chi( x)/\hbar} O( x,\hat p)e^{-i \hat \chi( x)/\hbar}= O( x,\hat p- \hat A_0)\, ,
\end{equation}
demonstrating that the unitary transformation in Eq.~(\ref{wow}) corresponds to the application of the minimal coupling replacement in the dipole approximation.
\section*{Supplementary Note 2: Fluxonium {\em LC} oscillator Hamiltonian in the charge gauge}
Here we derive the total Hamiltonian describing a
fluxonium qubit (TLS) interacting with a superconducting {\em LC} oscillator \cite{Stokes2018, Manucharyan2009}.
Considering for simplicity the case of a zero-flux offset, the fluxonium Hamiltonian is:
\begin{equation}\label{hfluxon}
\hat H_{\rm flux}= 4\tilde E_C \hat N^2 +\frac{\tilde E_L}{2}\hat\phi^2-E_J \cos{\hat\phi} \, .
\end{equation}
where $\hat \phi=\hat\Phi/\hat\Phi_0=2e/\hbar\,\hat\Phi $ is the reduced flux operator with conjugate momentum $\hat N=\hat Q/2e$ (the reduced charge), such that $[\hat\phi, \hat N] = i$. In Eq.~(\ref{hfluxon}), $\tilde E_C$, $\tilde E_L$, and $E_J$ are, respectively, the capacitive, inductive, and Josephson energies.
The {\em LC} oscillator is characterized by a capacitance $C$ and an inductance $L$. Its resonance frequency is $ \omega_{c}=1/\sqrt{L C}$ and its characteristic impedance is $Z=\sqrt{L/C}$.
The Hamiltonian of the {\em LC} oscillator is
\begin{equation}\label{hlc}
\hat H_{\rm osc}=
\frac{\hat Q^2}{2C}+\frac{\hat \Phi^2}{2L}\, ,
\end{equation}
where $\hat Q$ and $\hat \Phi$ are the charge and the flux operator, respectively.
We now introduce the reduced flux operator $\hat \varphi= 2 \pi \Phi /\Phi_0$ and the reduced charge operator $\hat \chi = \hat Q/ (2e)$, whose commutator is $[\hat \varphi, \hat \chi]=i$. These can be expanded in terms of the creation and annihilation
operators:
$$\hat \varphi=\varphi_{0}(\hat a +\hat a^\dag)\, ,$$
$$\hat \chi=-i \chi_{0} (\hat a -\hat a^\dag)\, ,$$
where $\varphi_{0} =\sqrt{2e^2{L}\hbar\Omega}=\sqrt{2\hbar e^2Z}$,
and $\chi_{0} =\sqrt{{C}\hbar\Omega/(8e^2)}=\sqrt{\hbar /8e^2Z}$.
The Hamiltonian of the LC oscillator can be expressed as
\begin{equation}\label{hlc}
\hat H_{\rm osc}=4E_{C}\, \hat\zeta^2+ E_{L}\frac{\hat\varphi^2}{2}\, ,
\end{equation}
where $E_{C} = e^2/(2 C)$ and $E_{L} = (\phi_0/2 \pi)^2 /L$.
The capacitive coupling (charge gauge) between the fluxonium and the $LC$ oscillator
can be described by making the substitution
$\hat N \to \hat N+\hat \chi$ in the fluxonium Hamiltonian in \eqref{hfluxon}. The total Hamiltonian then becomes:
\begin{equation}\label{HC}
\hat H_{C}=4\tilde E_C (\hat N+\hat \chi)^2 -E_J\cos( \hat \phi) +\frac{\tilde E_L}{2}\hat \phi^2+ 4E_{C} \hat\chi^2+ E_L\frac{\hat\varphi^2}{2}\, .
\end{equation}
This replacement is different from the standard {\em minimal coupling replacement}, it involves two conjugate momenta, instead of a conjugate momentum and a field coordinate. This minimal coupling replacement can also be obtained by applying a unitary transformation to the fluxonium Hamiltonian $\hat H_{\rm flux}$:
\begin{equation}
\hat H_{C} = \hat H_{\rm osc} + \hat {R}\, \hat { H}_{\rm flux} \hat { R}^{\dag}\, ,
\end{equation}
where $\hat { R}=\exp[{i\hat\phi\hat\chi}]$. This can be proved by following the procedure described in the next section.
Diagonalizing $\hat H_{\rm flux}$ and then projecting Eq.~(\ref{HC}) in a two-level space, spanned by the eigenstates $\ket{0}$ (ground state) and $\ket{1}$ (excited state), we obtain
\begin{equation}\label{Hpcharge}
\hat {\cal H}'_{C}= \hat {\cal H}_{\rm flux} + \hbar \omega_c \hat a^\dag \hat a +i\hbar g_C \hat \sigma_y(\hat a -\hat a^\dag) -
4 \tilde E_C \chi_{0}^2
(\hat a -\hat a^\dag)^2\, ,
\end{equation}
where in the two-level approximation the reduced flux operator becomes
$\hat\phi= \Phi_{10}\hat \sigma_x$ ($\Phi_{10}\equiv \langle e| \hat \phi |g \rangle$ is assumed real), and $g_C=\omega_{10}\phi_{10}\chi_{0}$. We also have:
\begin{equation}
\hat {\cal H}_{\rm flux}= \frac{\hbar\omega_{10}}{2} \hat\sigma_z \,,
\end{equation}
with $\hat\sigma_z=\ketbra{1}{1}-\ketbra{0}{0}$.
It can be shown that the total Hamiltonian in Eq.~(\ref{Hpcharge}) is not gauge invariant because its derivation does not take into account the presence of an effective nonlocal potential which originates from the two-level truncation. In analogy with the procedure employed in the main text, {\em the correct gauge invariant total Hamiltonian} can be obtained applying the following unitary transformation to $\hat {\cal H}_{\rm flux}$:
\begin{equation}\label{mcrS}
\hat {\cal H}_{C} = \hbar \omega_{c}\, \hat a^\dag \hat a + \hat {\cal R}\, \hat {\cal H}_{\rm flux} \hat {\cal R}^{\dag} \, ,
\end{equation}
where
\begin{equation}
\hat {\cal R}= \exp\left[{\frac{g_C}{\omega_{10}}\hat\sigma_x(\hat a -\hat a^\dag) }\right]
\end{equation}
is the unitary operator resulting from the truncation in the two-level space of the unitary operator $\hat R$. The minimal coupling replacement in the truncated space \eqref{mcrS} gives
\begin{eqnarray}\label{CoulombS}
{red}{\begin{aligned}
\hat {\cal H}_C &=\hbar \omega_c \hat a^{\dag} \hat a
+\frac{\hbar\omega_{10}}{2}
\left\{\cosh\left[{\frac{2g_C}{\omega_{10}}}(\hat a -\hat a^\dag)\right]\hat\sigma_z+ \right. \\&+ \left.i\sinh\left[{\frac{2g_C}{\omega_{10}}}(\hat a -\hat a^\dag)\hat\sigma_y\right]
\right\} \nonumber\,.
\end{aligned}}
\end{eqnarray}
This equation describes the correct (gauge invariant) total Hamiltonian for a fluxonium qubit interacting with an LC circuit in the charge gauge.
|
\section{Introduction}
\label{sec:intro}
Inflationary cosmology has come a long way, starting from being the theory to explain the shortcomings of Big Bang model
\cite{starbsky, guth, linde} to the most accepted model with
the current observations. Success of $NASA's$ first CMB satellite COBE (Cosmic Background Explorer satellite), paved way for future projects in CMB.
These results gave an accurate thermal spectrum of CMB, the first detection of primordial CMB fluctuations with normalization at 10$^{-5}$.
It also gave cosmic variance limit of large angular scale power spectrum \cite{cobe}.
Among the CMB missions which followed COBE, BICEP mission was designed to measure the B-mode polarization
at degree angular
scales and can be considered as the first mission to constrain the tensor-to-scalar ratio $r$ directly from CMB B-mode polarization.
The mission gave the $r$ constraints as
\begin{equation*}
r=0.02^{+0.31}_{-0.26} \quad or \quad r<0.72 \quad at \quad 95\%\quad C.L.
\end{equation*}
The direct measurement of $r$ from the BB spectrum has advantage, as the measurements of TT are limited by cosmic variance at large angular scales
and the $r$ constraints from TT have strong parameter degeneracies such as with the
scalar spectral index $n_s$, while the BB amplitude depends primarily on \textit{r}
\cite{bicep}.
The SPT-SZ survey helped to remove the degeneracy between $n_s$ and $r$ by measurements of the power spectrum from angular scales corresponding
to multipole range 650 $<l<$ 3000, that is from the third acoustic peak through the CMB damping tail. The constrain on scalar spectral index tightens
to $n_s<1.0$ at $3.9\sigma$ by using SPT and WMAP7 data \cite{spt}.
The B and E mode polarization data gives constraints on the recombination (last-scattering) and reionization epochs. The cross-correlation between
temperature anisotropies and polarization signal is a consistency check for inflationary cosmology and helps in breaking degeneracies among the cosmological
parameters \cite{cmbpol}.
Analysis of the combined BICEP2 and Keck Array data in combination with the 2015 Planck data put an upper limit $r_{0.05} < 0.12 $ at 95$\%$
confidence \cite{comb}.
2018 release of the Planck temperature, polarization and lensing
data determine the spectral index of scalar perturbations to be
\begin{equation*}
n_s= 0.965\pm0.004
\end{equation*}
at $68\%$ CL and find no evidence for a scale dependence
of $n_s$, either as a running or as a running of the running.
The latest Planck $95\%$ CL upper limit on the tensor-to-scalar ratio,
$r_{0.002} < 0.10 $, is further tightened by combining with the BICEP2/Keck
Array BK14 data to obtain $r_{0.002} < 0.064 $ \cite{planck2018t, planck2018}.
In the present work we study the constraints on the tensor-to-scalar ratio \textit{r} using an alternative approximation technique, instead of
slow-roll approximation, with the Power law inflationary model. We first enumerate the impact of \textit{r} on different inflationary models and cosmological parameters,
followed by a brief review on some alternative approximation techniques. Scalar and tensor power spectra
using the WKB approximation technique are obtained. Hence we try to constrain tensor to scalar ratio with WKB approximation and check the consistency
with latest Planck observations.
\section{Tensor to scalar ratio}
The study of tensor to scalar ratio \textit{r} gives insight into large working of the universe :-
\begin{itemize}
\item
Primordial Gravitational waves - The tensor perturbations in the FLRW background metric during inflation are called the primordial gravitational waves and CMB is
the only observation to validate it. The CMB sky is measured with two observables, the temperature anisotropy and the polarisation. Primordial gravitational
waves imprints very slight variations on these variables and are sensitive to \textit{r} which is defined as \cite{cmbpol}
\begin{equation*}
r=\frac{\mathcal{P}_t}{\mathcal{P}_s}
\end{equation*}
It contributes to the temperature anisotropies as a local quadruple since the gravitational wave metric is traceless. The primordial tensor power spectrum \cite{star}
is scale
invariant and the shear is impulsive at horizon entry \cite{wang}. \\
\item Energy scale of universe - \textit{r} depends on the time-evolution of the inflaton field and the relation to inflaton potential \cite{liddle} is given as
\begin{equation*}
r=\frac{8}{M_{Pl}^2}\bigg(\frac{\dot\phi}{H}\bigg)^2
\end{equation*}
\begin{equation*}
V^{1/4}=1.06\times10^{16} \bigg(\frac{r_{\phi}}{0.01}\bigg)^{1/4} GeV
\end{equation*} where $r_{\phi}$ is the tensor to scalar ratio in CMB scales \cite{cmbpol}.
\item Consistency relation - \textit{r} gives a test for consistency for single-field slow roll inflation as
\begin{equation*}
r=-8n_{t}sin^2\Delta
\end{equation*}
where $sin^2\Delta$ parameterizes the ratio between the adiabatic power spectrum at horizon-exit during inflation and the observed power spectrum giving information about presence of multiple fields. For non-slow roll inflation, with a non-trivial $c_s$, it is given as
\begin{equation*}
r=-n_{t}c_s
\end{equation*}
where it is seen that the non-slow roll evolution of the inflaton field is driven by a non-canonical kinetic term.
\item Inflaton Field Excursion - The inflaton field $\Delta \phi$ is related to \textit{r} giving the effective number of e-folds as
\begin{equation*}
N_{eff}=\int_{0}^{N_{\star}}dN\bigg(\frac{r(N)}{r_{\star}}\bigg)^{\frac{1}{2}}
\end{equation*}
where $N_{\star}$ is the number of e-folds between the end of inflation and the horizon exit of the CMB pivot scale. By determining,
if the inflaton-field excursion was super-planckian or not we can probe aspects of scalar field space and the UV completion of gravity.
\item Inflationary models - The shape of the inflaton potential is controlled by the slow-roll parameters and hence measurements of
\textit{r}, $n_s$, $n_t$, $\alpha_{s}$, $\alpha_t$ help in constraining different inflationary models and also removes the degeneracies between the slow-roll parameters.
\end{itemize}
\section{Alternatives for Standard Slow-roll approximation}
There are many models proposed which give the similar observations, which doesn't explicitly consider slow roll approximation for inflation but gives very
similar results without the approximation \cite{martin}. For example,
\subsection{Hamilton-Jacobi formulation of inflation}
Hamilton-Jacobi formulation \cite{mus, bond} is a method to rewrite the equations of motion. It considers the scalar inflaton field itself to be
time-varying but it breaks down during the oscillatory epoch that ends inflation \cite{liddlelyth, minu2}.
We know the equations of motion for a spatially flat universe as
\begin{equation}
H^2=\frac{1}{3M_{pl}^{2}}(V(\phi)+\frac{1}{2}\dot{\phi^{2}})
\end{equation}
and
\begin{equation}
\ddot{\phi}+3H\dot{\phi}=-\frac{dV}{d\phi} ,
\end{equation} where $V(\phi)$ is the potential of the scalar field.
By differentiating the first equation of motion with respect to time and by substituting in the second equation of motion, we get
\begin{equation}
2\dot{H}=-\frac{\dot{\phi}^{2}}{M_{pl}^{2}}
\end{equation}
Dividing both sides by $\dot{\phi}$ gives,
\begin{equation}
\dot{\phi}=-2M_{pl}^{2}H'(\phi)
\end{equation}
which gives the Freidmann-equations in the first order form,
\begin{equation}
H'(\phi)^{2}-\frac{3}{2M_{pl}^{2}}H^{2}(\phi)=-\frac{1}{2M_{pl}^{4}}V(\phi) .
\end{equation}
When $H(\phi)$ is specified it gives the corresponding potential and all the other inflationary solutions can be derived from it.
For example,
\begin{equation}
H(\phi)\propto exp\Big(-\sqrt{\frac{1}{2p}}\frac{\phi}{M_{pl}}\Big)
\end{equation}
gives the Power-law inflation.
The slow-roll parameters can be written as,
\begin{equation}
\epsilon_{H}=3\frac{\dot{\phi}^2}{2}/\Big(V+\frac{\dot{\phi}^2}{2}\Big)=-\frac{dlnH}{dlna}
\end{equation}
\begin{equation}
\eta_{H}=-\frac{\ddot{\phi}}{H\dot{\phi}}=-\frac{dln\dot{\phi}}{dlna}
\end{equation}
and inflation is precisely given as,
\begin{equation}
\ddot{a}>0 \Longleftrightarrow\epsilon_{H}<1 .
\end{equation}
\subsection{Perturbed Power law inflation}
With the standard Power law model the scale factor is given as, $a(t)=t^p$
where $p>1$ and the slow-roll parameters are given as $\epsilon=-\eta=\frac{1}{p}$.
The Power-law also states that \cite{lucchin} - \cite{ts_thesis},
\begin{itemize}
\item[(a)]Expansion is uniformly accelerated \textit{ie }$\epsilon=constant$
\item[(b)]The perturbations in the inflaton field is similar to that for a massless scalar field $m^2_{eff}=0$.
\end{itemize}
The power spectrum is given as $\mathcal{P}_{k}=Ak^{n_s -1}$ and spectral indexes are $n_s$ and $n_t=n_s-1$ which are constant values.
The perturbed Power law \cite{minu1, ppl2} can be given as a slight deviation from the scale invariant spectra. Here, the inflaton potential is estimated from the `reduced Hamilton -Jacobi equation'.
The corresponding slow-roll parameters are given as,
$\epsilon=\frac{M_{pl}^2}{4\pi}\Big(\frac{H'}{H}\Big)^2$ and $\eta=-\frac{M_{pl}^2}{4\pi}\Big(\frac{H''}{H}\Big)$. The spectral indices are then,
\begin{equation}
n_t=n_s-1=\frac{-2\epsilon}{1-\epsilon}
\end{equation}where $\epsilon=(\frac{\partial ln H}{\partial \phi})^2$.
The Perturbed power-law is taken for small deviations from the uniform acceleration in terms of $\frac{\partial^2 lnH}{\partial\phi^2}$. The scalar perturbation spectrum
is then obtained as,
\begin{equation}
\mathcal{P}_{s}(k)=\frac{A(\mu_T , \mu_S)}{2\pi}\Big(\frac{H}{2\pi}\Big)^2 \frac{1}{\epsilon(\eta)}
\end{equation} and the corresponding tensor power-spectrum is,
\begin{equation}
\mathcal{P}_t(k)=\frac{8A(\mu_T , \mu_S)}{2\pi} \Big(\frac{H}{2\pi}\Big)^2
\end{equation}
where $A(x,y)=\frac{4^y\Gamma^2(y)}{(x-1/2)^{2y-1}}$, $\Gamma(y)$ is the Euler gamma function where $\mu_T\rightarrow \mu $ and \\
$\mu_s$ = $\Big(\mu^2 - (\mu-1/2)^2\frac{m^2_{eff}}{H^2}\Big)^{1/4}$.
\section{WKB approximation}
The WKB approximation is a method used to get approximate solutions to linear differential equations. Here it is applied on the Radial wave equations.
It was proved that the Balmer formula for Hydrogen atom could be derived from correct application of the WKB approximation. Hence the method was used for
cosmological perturbations which had an analogy with the Hydrogen atom formalism. The use of this method also gives the advantage that it can be used for
both subhorizon and the superhorizon scales and it is also an alternate method than the standard slow-roll approximation.
The non-singular solutions with initial dense de-sitter phase, first mentioned by Starobinsky gave
the tensor perturbation formulation \cite{star}. Mukhanov and Chibisov proposed that the spectrum of quantum fluctuation in the non-singular solution could evolve
into the present universe and derived the
scalar perturbation equation \cite{mukh}, followed by works of Hawking \cite{hawk} and Starobinsky \cite{star2}.
These scalar and tensor perturbations formulations are reduced to a single variable \cite{guthpi} denoted by $\mu_S$ or $\mu_T$, and the equation is similar to
the Schr$\ddot{o}$dinger equation \cite{Schro}.
\begin{equation}
\frac{d^2}{d\eta^2}\mu +\big(k^2-U(\eta)\big)\mu = \frac{d^2\mu}{d\eta^2}+\omega^2(\eta)\mu =0
\end{equation}where $U(\eta)$ is the potential. Then the corresponding power spectra and the spectral indices are given as,
\begin{equation}
\mathcal{P}_s=\frac{k^3}{8\pi^2}\Big|\frac{\mu_s}{z_s}\Big|^2, \quad \mathcal{P}_T=\frac{2k^3}{\pi^2}\Big|\frac{\mu_T}{z_T}\Big|^2
\end{equation}
where $z_s=a\sqrt{\frac{-aa''}{a'}}$ and $z_T=a$. The spectral indices being,
\begin{equation}
n_s-1=\Big|\frac{dln \mathcal{P}_s}{dlnk}\Big|_{k=k_*},\quad n_T=\Big|\frac{dln\mathcal{P}_T}{dlnk}\Big|_{k=k_*} .
\end{equation}
The WKB spectra is taken as the solution for the equation,
\begin{equation}
\mu''_{WKB}(k,\eta)+[\omega^2(k,\eta)-Q(k,\eta)]\mu_{WKB}=0
\end{equation} where
\begin{equation}
Q(k,\eta)=\frac{3}{4}\frac{(\omega')^2}{\omega}-\frac{\omega''}{2\omega} .
\end{equation}
The mode function $\mu_{WKB}$ is an approximation to the actual mode function in the limit,
\begin{equation}
\Big|\frac{Q}{\omega^2}\Big|<1
\end{equation}
For application of the WKB approximation, the variables are transformed to
\begin{equation}
x=ln\Big(\frac{Ha}{k}\Big),\quad u=(1-\epsilon_1)^{\frac{1}{2}}e^{\frac{x}{2}}\mu
\end{equation}
which in turn gives the Power spectra to be,
\begin{equation}
\mathcal{P}_s=\frac{H^2}{\pi\epsilon_1M_{pl}^2}\Big(\frac{k}{aH}\Big)^3\frac{e^{2\psi_s}}{(1-\epsilon_1)|\omega_s|}
\end{equation}
\begin{equation}
\mathcal{P}_T=\frac{16H^2}{\pi M_{pl}^2}\Big(\frac{k}{aH}\Big)^3 \frac{e^{2\psi_T}}{(1-\epsilon_1)|\omega_T|}
\end{equation}
Here we have seen a brief on the general cosmological perturbation equations and the application of WKB approximation on them as was given in the work
by J. Martin and D. Schwarz (2003) \cite{mainref} and the derivations of the power spectra for the scalar and the tensor perturbations. Further we
use the above method for Power law inflation as the model is exactly solvable and hence comparable with the observation data.
\section{Application of WKB approximation on Power law inflationary model}
For Power law inflation, the scale factor is
\begin{equation}
a(\eta)= l_{o}|\eta|^{1+\beta}
\end{equation}
where $\beta\leq-2$ and $l_{o}$ is the Hubble's radius which is a constant when $\beta=-2$. The corresponding Horizon flow functions are given as,
\begin{equation}
\epsilon_{n+1}= \frac{d ln |\epsilon_{n}|}{dN}
\end{equation} for $n\geq0$.
Inflation takes place when $\epsilon_{1}<1$.
The exact power spectrum is given as
\begin{equation}
\mathcal{P}_{\zeta}=\frac{l^{2}_{pl}}{l^{2}_{0}}\frac{f(\beta)k^{2\beta+4}}{\pi\epsilon_{1}}
\end{equation}
\begin{equation}
\mathcal{P}_{h}=\frac{l^{2}_{pl}}{l^{2}_{0}}\frac{16f(\beta)k^{2\beta+4}}{\pi}
\end{equation}
where $l_{pl}=m_{pl}^{-1}$ is the Planck Length and the function $f(\beta)$ is given by \begin{equation}
f(\beta)=\frac{1}{\pi}\Big[\frac{\Gamma(-\beta-\frac{1}{2})}{2^{\beta+1}}\Big]^{2}
\end{equation} where $\Gamma$ is the Euler integral of the second kind.
When $f(\beta=-2)=1$, this becomes as a special case because when $\beta=-2$ the expression of the scalar spectrum blows up.
By applying the WKB approximation the effective frequency is given as,
\begin{equation}
\omega^2(\eta)=k^2-\beta(\beta+1)/\eta^2
\label{ome}
\end{equation}
and $|\frac{Q}{\omega^2}|=\frac{1}{4\beta(\beta+1)}$.
As for inflation $\frac{Q}{\omega^2}\ll1$, when $\beta=-2$ the ratio goes to 1/8 hence satisfying the condition.
By applying the WKB approximation we get the power spectra to be
\begin{equation}
\mathcal{P}_{\zeta}=\frac{l^{2}_{pl}}{l^{2}_{0}}\frac{1}{\pi\epsilon_{1}}g(\beta)k^{2\beta+4}
\label{P1}
\end{equation}
\begin{equation}
\mathcal{P}_{h}=\frac{l^{2}_{pl}}{l^{2}_{0}}\frac{16}{\pi}g(\beta)k^{2\beta+4}
\label{P2}
\end{equation}
and the function is given as,
\begin{equation}
g(\beta)=\frac{2e^{2\beta+1}}{(2\beta+1)^{2\beta+2}} .
\end{equation}
The corresponding spectral indices are then computed to be
\begin{equation}
n_s-1=n_T=2\beta+4 .
\end{equation}
The Constant-roll inflation or the Ultra-slow-roll inflationary model is also a viable alternative to the standard slow-roll approximation.
In this model, the inflationary scenario is investigated with constant rate of roll,
$\ddot{\Phi}/ H\dot{\Phi}$ = $-3-\alpha$ (remains a constant) \cite{moho}.
The model gives experimentally viable results for $\alpha \lesssim -3$. The parameter constrains on the constant
roll model and comparison with observations is explored by Motohashi and Starobinsky \cite{mohoo} which gives a permitted range
of $r$ values as $r \lesssim 0.07 $ to $r \ll 10^{-3} $.
\section{WKB Power Spectra and Tensor to scalar ratio(\textit{r})}
By using the WKB approximation, we first see that the effective frequencies in Eq. (\ref{ome}) are given as,
\begin{equation}
\omega_s^2=k^2\eta^2 -\Big(\frac{9}{4}+3\epsilon_1+\frac{3}{2}\epsilon
_2\Big)
\end{equation}
\begin{equation}
\omega_T^2=k^2\eta^2-\Big(\frac{9}{4}+3\epsilon_1\Big)
\end{equation}
Then corresponding WKB spectra are derived from solving Eq. (\ref{P1}) $\& $ (\ref{P2}) with the above frequencies and we get the scalar power spectrum as,
\begin{equation}
\mathcal{P}_{\zeta}= \frac{H^{2}}{\pi\epsilon_{1}m^{2}_{pl}}(18e^{-3})[1-2(D+1)\epsilon_{1}-D\epsilon_{2}-(2\epsilon_{1}+\epsilon_{2})ln\Big(\frac{k}{k_{*}}\Big)+\textit{O}(\epsilon_{n}^{2})]
\label{pks}
\end{equation}
and the corresponding tensor power spectrum as,
\begin{equation}
\mathcal{P}_{h}=\frac{16H^{2}}{\pi m^{2}_{pl}}(18e^{-3})[1-2(D+1)\epsilon_{1}-2\epsilon_{1}ln\Big(\frac{k}{k_{*}}\Big)+\textit{O}(\epsilon_{n}^{2})]
\label{pkt}
\end{equation}
with $D\equiv\frac{1}{3}-ln3\approx-0.765$.
The Spectral indices calculated from above equations are in accordance with the spectral indices given by slow-roll inflation but it gives the next higher order
terms which increases it's accuracy. The spectral indices are obtained as,
\begin{equation}
n_{s}-1=\frac{-2\epsilon_{1}}{1-\epsilon_{1}}-\frac{\epsilon_{2}}{1-\epsilon_{1}}+
\frac{-2\epsilon_{1}-\epsilon_{2}}{[1-2(D+1)\epsilon_{1}-D\epsilon_{2}-(2\epsilon_{1}+\epsilon_{2})ln(\frac{k}{k_{*}})]}+\mathcal{O}(\dot{\epsilon_{n}})
\end{equation}
\begin{equation}
n_{T}=\frac{-2\epsilon_{1}}{1-\epsilon_{1}}+\frac{-2\epsilon_1}{[1-2(D+1)\epsilon_{1}-(2\epsilon_{1})ln(\frac{k}{k_{*}})]} + \mathcal{O}(\dot{\epsilon_{n}})
\end{equation}
With Slow-roll approximation the power spectra are \cite{stewart}
\begin{equation}
\mathcal{P}_{\zeta}=\frac{H^{2}}{\pi\epsilon_{1}m^{2}_{pl}}[1-2(C+1)\epsilon_{1}-C\epsilon_{2}-(2\epsilon_{1}+\epsilon_{2})ln\Big(\frac{k}{k_{*}}\Big)+\textit{O}(\epsilon_{n}^{2})]
\end{equation}
and the tensor powerspectrum,
\begin{equation}
\mathcal{P}_{h}=\frac{16H^{2}}{\pi m^{2}_{pl}}[1-2(C+1)\epsilon_{1}-2\epsilon_{1}ln\Big(\frac{k}{k_{*}}\Big)+\textit{O}(\epsilon_{n}^{2})]
\end{equation}
where C=$\gamma_{E}+ln2-2\simeq$ -0.7296 , $\gamma_{E}\simeq$0.5772 which is the Euler constant. All the quantities are calculated at $\eta_*$ which is when $k_*=aH(N_*)$. This scale $k_*$ is called the Pivot scale.
The spectral indices with the slow-roll approximation are given as,
\begin{equation}
n_s-1=-2\epsilon_1-\epsilon_2, \quad n_T=-2\epsilon_1
\end{equation}
From this we could readily see that the spectral indices of the WKB approximation reduces to the spectral indices of the slow-roll approximation.
The scalar and tensor power spectra are represented by using the Power law form. For which the expressions for Hubble's constant is
\begin{equation}
H=-\frac{\dot{H}}{H^2}=\frac{H'}{aH^2}=(1+\beta)\eta^{-(\beta+2)}/l_o
\end{equation}
and $\epsilon_2=0$.
The power-spectra are analytically represented and by setting the pivot scale at $k_*=0.05 Mpc^{-1}$.
The plots show the scale-invariant scalar and tensor power spectra with the slow-roll approximation and with the WKB approximation.
The power-spectra are calculated for the $\beta=-2.005$.
\begin{figure}[!hbt]
\centering
\includegraphics[width=4.5in]{s12_beta_2005.png}
\label{fig:1}
\caption{\small Scalar power-spectra (1)with slow-roll approximation and (2)with WKB approximation}
\end{figure}
\begin{figure}[!hbt]
\centering
\includegraphics[width=4.5in]{t12_beta_2005.png}
\caption{Tensor power spectra (1)with slow roll approximation and (2)with WKB approximation}
\label{fig:2}
\end{figure}
It is clear that the power spectra are scale invariant with both approximation techniques and the WKB approximation
spectra shows a slight variation in the amplitude which can be compensated by changing the amplitude terms in Eq. (\ref{pks}) and (\ref{pkt}).
The corresponding spectral indices are calculated for these spectra. With the WKB approximation it is $n_s-1=n_T=3.12177276e-04 $ and with the slow roll approximation
$ n_s-1=n_T=3.08732253e-04 $.
The application of WKB approximation helps in determining the power spectra even when the Slow-roll approximation is invalid and hence gives us a new frontier
for viewing inflation. The added advantage is that it gives solutions with higher accuracy and higher order terms which is a necessity to match with the very
precise observations of WMAP and Planck.
We use CosmoMC \cite{cosmomc}, to establish the constrains on the parameters and to identify the maximum likelihood region in the parameter space.
The primordial power-spectrum in CAMB of CosmoMC is changed to the WKB approximation power spectrum.
As even a slight change can produce large deviations in the CMB sky, we have constrained the WKB approximation parameter
to $\beta=-2.005$. The modified power spectrum is used to then generate the CMB angular power spectrum with flat priors and
using the whole base-set of parameters with power law $\Lambda CDM$ model.
We stay with the minimal six parameter spatially flat $\Lambda CDM$ cosmological model as our base model with base
parameters as baryon density today $\Omega_{b}h^2$, cold dark matter density today $\Omega_{c}h^2$, amplitude of scalar
power spectrum $A_{s}$, spectral index of scalar power spectrum $n_s$, tensor-to-scalar ratio \textit{r} and the tilt of
tensor power spectrum $n_t$ but with WKB primordial power spectrum. This model has four free non-primordial cosmological parameters same as the power-law
baseline \textit{i.e} $\omega_b$, $\omega_c$, $\theta_{MC}$ and $\tau$. Here we try to estimate the impact of primordial power spectrum with higher-order terms of Hubble-flow
functions, which are analogous to slow-roll parameters with emphasis on the deviations in tensor-to-scalar ratio. We also take to consideration three neutrinos species,
with two mass-less states and a single massive neutrino of mass $m_{\nu}$ = 0.06eV.
The Planck likelihood code$^\dag$ {\footnotetext[1]{2018 Planck CMB likelihoods (except for lensing) are not published yet.
The present discussion refer to the 2015 likelihoods.}} (PLC/clik) and parameter chains are available from the Planck Legacy Archive.
The Planck lensing and Bicep-Keck-Planck likelihoods are included with the cosmomc.
\begin{figure}[H]
\centering
\includegraphics[width=6.1in]{WKBCP.pdf}
\caption{\small 1D marginalized posterior distribution of the inflationary parameters and cosmological parameters with WKB approximation}
\label{fig:3}
\end{figure}
The limits and marginalized densities given in Fig. \ref{fig:3} are generated using the getdist package which is part of CosmoMC, it
uses the baseline likelihood which is a hybrid,
by connecting together a low-multipole likelihood with the Gaussian likelihood constructed from the higher multipoles. We study the Bayesian factors
of the model
with respect to the power-law base model and the $\chi^2_{eff}$ values with power-law model.
As it is clear the WKB power spectrum is scale invariant and the parameters are in the $95\%$ CL of Planck results.
\begin{figure}[H]
\centering
\includegraphics[width=5.1in]{WKBrT.pdf}
\caption{\small 1D marginalized posterior distribution of the Tensor to Scalar ratio \textit{r} with WKB approximation}
\label{fig:4}
\end{figure}
By combining Planck temperature, low- polarization and lensing 2018 Planck results give $r_{0.002} < 0.10 $ ($95\%$ CL, Planck TT+lowE+lensing)
\cite{planck2018}.
This constraint slightly improves on the corresponding Planck 2015, $95\%$ CL bound, i.e., $r_{0.002} < 0.11 $ \cite{planck2015} which is
consistent with the B-mode polarization constraint $r < 0.12$ (95 $\%$ CL), obtained from a joint analysis of the BICEP2/Keck Array and Planck data.
With the WKB approximation, the upper limit on \textit{r} is obtained
as $\textit{r} < 0.1109$ and $\textit{r}_{0.002} < 0.1055 $ (BKPlanck$^\dag${\footnotetext[2]{BICEP2/Keck Array and 2015 Planck joint analysis}}+ lensing).
This gives us a window into higher order corrections
and better fit to the Planck results. It is seen that the WKB best fit value of the spectral index is $n_s=0.9657$ with $H_{o}=67.64$ which are
in the $95\%$ CL of Planck observations.
\section{Conclusion}
Present work studied the scalar and tensor power spectra with the WKB approximation and constrained the value of the WKB parameter $\beta$.
The WKB approximation helps in determing the power spectra even when the Slow-roll approximation is invalid, gives the higher order corrections and
the precise details of Power law model which will give insights into paramount questions
in the Inflationary model of Universe. The spectral index values and spectra generated with WKB approximation are compared with the observation data and we see that
the WKB approximated power spectra could become a viable alternative for Slow-roll approximation. Running Monte-Carlo chains with CosmoMC and generating the plots with
Getdist, the tensor to scalar ratio(\textit{r}) is also analysed by
constraining the cosmological parameters with WKB approximation, setting an upper limit at $ \textit{r} < 0.1109 $. Here \textit{r}
is defined at the pivot scale $k_* = 0.05Mpc^{-1}$. We also report the bounds on $\textit{r}_{0.002}$, the tensor to scalar ratio at $k_* = 0.002 Mpc^{-1}$
as $\textit{r}_{0.002} < 0.1055 $ which is consistent within $1\sigma$ with the Planck results.
\section{Acknowledgments}
We acknowledge the use of high performance computing system at IUCAA. MJ acknowledges the support from U.G.C. Major
research project grant, $F. No.-43-523/2014(SR)$ and the Associateship of
IUCAA. AA acknowledges the Junior Research Fellowship from University Grants Commission (UGC), India. We thank the referee for the valuable suggestions and
comments.
\section*{References}
|
\section{Introduction and Motivation}
Interactions do not exist in a vacuum, but are surrounded by many social and physical constructs that shape and constraint that interaction \cite{Argyle}. Understanding and modeling the context of interactions is essential in the design of systems that are both realistic and computationally feasible. In fact, if context is not properly considered \emph{``the patterns and the outcomes of the interactions are inherently unpredictable, and predicting the behavior of the overall system based on its constituent components is extremely difficult (sometimes impossible) because of the high likelihood of emergent (and unwanted) behavior"}~\cite{jennings:00}. In human societies, we use social practices as means to cope with uncertainty of outcome of interaction. Interactions are embedded in a broad network of societal and institutional contexts, and social practices provide means to deal with this complexity. Endowing agents with means to represent and reason with social practices, will enable smooth, flexible, context-aware human-agent interaction.
Applied to human-agent interactions, such as companion robots, human-agent-robot teamwork, or persuasive applications, social practices can simplify the deliberation of the agent in complex contexts. In these domains, the agent is expected to support the user but there are no predefined protocols or fixed objectives, so agent actions must be based on a close observation of the situation and evaluation of possible user aims, and reaction to user actions and environment changes. This is exactly the reason for social practices in human interactions. Social practices refer to everyday practices and the way these are typically performed by (most of) the members in the society. Even though the subject of emergence and evolution of social practices is an important one, for simplicity in this paper we assume social practices to be given and fixed. I.e. we will not elaborate on how the practice come to be, or how they are maintained and shared. Our conceptual agent architecture therefore takes a set of social practices as given, and the focus is on how the agent uses practices in its deliberation and planning.
In this paper, we present a formalization of the theory of social practices as proposed in \cite{shove2012,Reckwitz} to be used as basis for a cognitive architecture. Our
approach can be seen as a middle way between using fixed interaction scripts and full free deliberation reactive to user and environment sensing. We build on the concept of landmark \cite{Dignum2007} in the sense, that social practices enable the specification of possible (abstract) plans without needing to represent every single component of the interaction. An agent will then adopt a social practice, based on its evaluation of the context, and \textit{`fill-in the gaps'}, using its own capabilities and according to its own intentions. We provide a logic semantics for social practices, based on modalities to represent expectation, capability, roles, norms and contexts.
\section{Background}
Social practices are accepted ways of doing things, contextual and materially mediated, that are shared between actors and routinized over time \cite{Reckwitz}. They can be seen as patterns which can be filled in by a multitude of single and often unique actions. Through (joint) performance, the patterns provided by the practice are filled out and reproduced.
Each time it is used, elements of the practice, including know-how, meanings and purposes,
are reconfigured and adapted \cite{shove2012}. Therefore the use of social practices includes a constant learning of the individuals using the social practice in ever changing contexts. In this way social practices guide the learning process of agents in a natural way. In \cite{shove2012} the social aspect of social practices is emphasized by giving the social practice center stage in interactions and letting individuals be supporters of the social practice. It shows that social practices are shared (social) concepts. The mere fact that they are shared and jointly created and maintained means that individuals playing a role in a social practice will expect certain behavior and reactions of the other participants in the social practice. Thus it is this aspect that makes the social practices so suitable for use in individual planning in social situations. Because, in this paper, we concentrate on the individual planning, we will not see much of the particular social aspects of the social
practices.
The formal model proposed in this paper is based on modal logics, extending work on agent organizations and landmarks, in particular the Logic of Agent Organizations (LAO) \cite{Dignumigpl} that provides a formal, provable representation of organizations based on the
notions of capability, `bring it about' -- or stit -- \cite{Porn1974}, attempt and responsibility. Landmarks
represent points in the plans of an agent that must always be achieve. This concept has been extensively used in the planning community \cite{}. We use the ADICO grammar proposed by Ostrom to specify rules, norms, and strategies \cite{crawford1995grammar}. ADICO statements
are formed using the following five components: Attribute (or Acting entity), Deontic, aIm (or Intention), Condition , and Or else (or sanction). In ADICO Strategies are statements including only the acting entity, intention, and condition (AIC); Norms
include the acting entity, deontic, intention, and condition (ADIC); and Rules include all 5 components (ADICO). The use of landmarks in the specification of agent plans
and interactions has been proposed by \cite{kumar02,Dignum2007}. Formally, landmarks are conjunctions of logical expressions that are true in a state, representing families of protocols. The level of landmark detail determines the degree of actors freedom.
Related research on work practices and cognitive architectures are the closest to our proposal. Work practice research recognizes the inherent difference between the work flows as described and prescribed by the organization and employee behavior. The work practice model Brahams enables to define the behavior of entities by means of activities and workframes, amongst others~\cite{sierhuis09} but lacks learning capabilities to adjust priorities, and means to distinguish between context and action preconditions. Cognitive architectures \cite{sun2009motivational} use drives as basis to dynamically derive goals during agent interactions. As such, it can capture the motivational complexity of the human mind~\cite{newell1994unified}, but it takes an agent perspective rather than a societal one, such as the one we propose.
\section{Characteristics of Social Practices}\label{sec:spelements}
In Social Sciences, \emph{social practices} (SP) are defined on the basis of materials, meanings and competences~\cite{holtz2014}. In the following we adapt an initial conceptualization of these concepts, as introduced in \cite{Dignum2015}:\\
\textit{\textbf{Context}}~\cite{Zimmermann:2007}
\begin{itemize}
\item \textit{Roles} describe the competencies and expectations about a certain type of actors~\cite{sunstein96}. Thus a lecturer is expected to deliver the presentation.
\item \textit{Actors} are all people and autonomous systems involved, that have capability to reason and (inter)act. This indicates the other agents that are expected to fulfill a part in the practice.
\item \textit{Resources} are objects that are used by the actions in the practice such as seats, projector, screen, etc. So, they are assumed to be available both for standard actions and for the planning within the practice.
\item \textit{Affordances} are the properties of the context that permit social actions and depend on the match between context conditions and actor characteristics~\cite{gaver1996situating}.
\item \textit{Places} indicates where all objects and actors are usually located relatively to each other, in space or time: Seats in a lecture theater all face the front of the room, etc.
\end{itemize}
\textit{\textbf{Meaning}}
\begin{itemize}
\item \textit{Purpose} determines the social interpretation of actions and of certain physical situations.
\item \textit{Promotes} indicates the values that are promoted (or demoted, by promoting the opposite) by the social practice.
\item \textit{Counts-as} are rules of the type "X counts as Y in C" linking brute facts (X) and institutional facts (Y) in the context (C) \cite{searle1995construction}. E.g., in a voting place, filling out a ballot counts as a vote.
\end{itemize}
\textit{\textbf{Expectations}}
\begin{itemize}
\item \textit{Plan patterns} describe usual patterns of actions~\cite{bresciani2004tropos} defined by the landmarks that are expected to occur.
\item \textit{Norms} describe the rules of (expected) behavior within the practice, as statements of the form ADIC or ADICO.
\item \textit{Strategies} indicate the possible activities that are expected within the practice. Not all activities need to be performed! They are meant as potential courses of action. Strategies are specified as AIC statements
\item \textit{Start condition}, or trigger, indicating how SP starts
\item \textit{Duration}, or End condition, indicating how SP ends
\end{itemize}
\textbf{\textit{Activities}}
\begin{itemize}
\item \textit{Possible actions} describes the expected actions by actors in the social practice
\item \textit{Requirements} indicate the type of capabilities or competences that the agent is expected to have in order to perform the activities within this practice.
\end{itemize}
Social Practices show some resemblance to agent organizations (see e.g. \cite{OperA}), in the sense that both provide structure to the interactions between agents. However, organizations provide an imposed (top-down) structure, while the social practices form a structure that arises from the bottom up. Thus instead of compliance, interaction patterns in a social practice indicate expectations on the behavior of its actors. Therefore no guarantee is given that behavior will occur (exactly) as expected. This has a large influence on the way the social practices are formally specified, focusing on possibilities and priorities rather than prescribing behavior. Cf. Table \ref{table:sp}.A for an informal description of how to use social practices.
Expectations are given in the form of roles that determine possible and expected behavior and also indicate who is supposed to take initiative at certain points. E.g. the lecturer is supposed to take the initiative to start the lecture. Norms determine the normal patterns of behavior and thus also determine a certain type of expectation, namely the patterns of actions that are allowed or prohibited. The plan patterns also determine partly the sequences of actions that are expected. The different scenes are temporally ordered and have a specific starting and end situation. Thus the whole practice should be fulfilled by following a trace that fits through that plan pattern.
In the next section we will start defining a formal representation for social practices based on the above observations.
\section{Formal model}\label{sec:logic}
Because the description of actions and action sequences is important as well as epistemic/doxastic states of the agents we will use a combination of dynamic logic with epistemic logic as the basis for our formalization.\footnote{Given that we only need to represent epistemic and dynamic operators, and not their dynamics, Dynamic epistemic logic \cite{Ditmarsch} is not needed here.}
\subsection{Activities}
We start by formalizing social practice activities in terms of actions and capabilities.
Let $Act=\{\alpha_1,...\alpha_n\}$ be the set of basic actions and $Ag=\{a_1,...a_m\}$ be the set of actors. We define a function $Capability: Agx2^{Act}$ that indicates for each actor the set of (basic) actions it is capable to perform. We have $Cap(a,\alpha) {\bf iff} \alpha\in Capability(a)$.\\
We use $[\alpha(a)]\phi$ to denote that if actor $a$ performs action $\alpha$ the result $\phi$ will become true afterwards. We assume the axiom: $([\alpha(a)]{\bf true}) \rightarrow Cap(a,\alpha)$ such that we do not have to specify the capability as a precondition every time we use the action description with an actor.\\
We can generalise these definitions for unions of actions, parallel actions, sequences of actions and repetitions:\\
The set ${\sl CA}$ of complex action expressions is given as the smallest set closed under:
\begin{itemize}
\item[(i).] $Act \subseteq {\sl CA}$
\item[(ii).] $ \alpha_1 , \alpha_2 \in {\sl CA} \Longrightarrow \alpha_1
+ \alpha_2 \in {\sl CA}$
\item[(iii).] $ \alpha_1 , \alpha_2 \in {\sl CA} \Longrightarrow \alpha_1
\& \alpha_2 \in {\sl CA}$
\item[(iv).] $ \alpha_1 , \alpha_2 \in {\sl CA} \Longrightarrow \alpha_1 ; \alpha_2 \in {\sl CA}$
\item[(v).] $ \alpha \in {\sl CA} \Longrightarrow \alpha * \in {\sl CA}$
\end{itemize}
Let $\gamma \in \sl CA$ and $A \subseteq Ag$ then $[\gamma(A)]\phi$ denotes that $\phi$ will be true after the group of actors $A$ has executed the complex action $\gamma$. Here we use $\gamma(A)$ as an abbreviation for the following:\\
Let $\gamma\equiv \gamma_1 \circ \gamma_2$ with $"\circ"$ is $"+", "\&"$ or $";"$ or $\gamma\equiv \gamma_1 ; \gamma_2 \wedge \gamma_2\equiv \gamma_1 *$ then
\[\gamma(A) \equiv \exists A',A'': A'\cup A''=A \wedge \gamma_1(A')\circ\gamma_2(A'')\]
Thus if we indicate that a group $A$ performs an action we explicitly do not differentiate who performs what part of that action. If we want to be more specific we can indicate the specific group performing a sub-action with that sub-action directly. Thus:
\[ (\gamma_1 \circ \gamma_2)(A) \not\equiv \gamma_1(A)\circ\gamma_2(A)\]
We use $act(\gamma)$ to denote the set of basic actions that are part of the complex action $\gamma$.
º
It will be handy to refer to abstract actions such as "an action that achieves $\phi$" or "an action performed by $A$ that achieves $\phi$". For this we introduce the following definitions:\\
Let $\alpha_i \in \sl CA$
\[ \alpha\phi \equiv \cup\alpha_i : [\alpha_i]\phi \]
\[ \alpha(A)\phi \equiv \cup\alpha_i(A) : [\alpha_i(A)]\phi \]
Note that if $\phi=true$ the abstract action refers to any action that can be performed.
We use $DO(a,\alpha)$ (resp. $DONE(a,\alpha)$) to denote that actor $a$ performs action $\alpha$ next (resp. actor $a$ performed action $\alpha$ as the last action). These action modalities can be represented in the semantics of dynamic logic by introducing a selection function on the links that indicate all possible actions in the current state. (see e.g. \cite{Dignumigpl} for details). Again we assume $DO(a,\alpha) \rightarrow Cap(a,\alpha)$ and $DONE(a,\alpha) \rightarrow Cap(a,\alpha)$.\\
We also extend this notation to groups of actors and use $DO(A,\alpha)$ and $DONE(A,\alpha)$ to denote a group $A$ doing an action together. Finally we use $DO(a,\alpha(A))$ and $DONE(a,\alpha(A))$ to denote that $a$ performs actions as part of the group $A$ in order to perform $\alpha$ together.
Besides actions we will also refer to contexts within which an action can be performed. This will be especially important when effects of an action depend on the context. E.g. the effect of raising your hand can be different in an auction and in a classroom. Intuitively a context stands for a temporal and/or spatially defined interval within which an action takes place, such as "the lunch break" or "the board room of university".
There is a vast amount of literature on reasoning in context. We will not tap on this literature, because we mainly use contexts (in this paper) as reference structures needed to specialize interaction expectations and beliefs. Every social practice can also function as a context by using only those parts that give temporal and spatial restrictions on the actions and agents.\\
We assume a set $C=\{c_1,...,c_n\}$ of special context constants (or names). Let $SP$ be the set of social practice identifiers then: $SP \subset C$. (Thus every social practice is a context, but not every context is a social practice). We explicitly do not have closeness of this set under combinations of contexts in any way! We also do not assume that any combination of specifications of temporal and/or spatial intervals defines a context. However, we do have that with every element $c_i \in C$ there is associated a formula $\Psi_i$ written as $con(c_i,\Psi_i)$ such that:
\[ active(a,\gamma, c_i) \equiv \Psi_i(a,\gamma)\]
meaning that context $c_i$ is active for actor $a$ while performing action $\gamma$ if the conditions associated with that context ($\Psi$) hold for $a$ performing $\gamma$.\footnote{This is a very simplified version of contexts such as is developed in \cite{GrossiDignum}}.If $c_i$ is a social practice then:
\[\Psi_i \rightarrow Sc_{c_i} \wedge \neg D_{c_i}\]
where $Sc_{c_i}$ is the start condition of $c_i$ and $D_{c_i}$ is the duration or end condition.\\
In general it can be the case that more than one context is active when an actor performs an action. This is not always a problem, but especially for concepts like expectations we want to have a unique context to narrow down (disambiguate) which expectations are relevant. Therefore we also assume there is a function $salient: CxC \rightarrow C$ that indicates for any pair of contexts which is the most salient. The following two restrictions hold for this function:
\[active(a,\gamma,c_1) \wedge \neg active(a,\gamma,c_2) \rightarrow salient(c_1,c_2)=c_1\]
\begin{gather*}
active(a,\gamma,c_1) \wedge active(a,\gamma,c_2) \wedge con(c_1,\Psi_1) \wedge con(c_2,\Psi_2) \\
\wedge (\Psi_1(a,\gamma)\rightarrow \Psi_2(a,\gamma))\rightarrow salient(c_1,c_2)=c_1
\end{gather*}
the first restriction states that only active contexts can be salient. The second states that more specific contexts are salient.\\
Finally we define $Salient(a,\gamma,c_i)$ as:
\[active(a,\gamma,c_i) \wedge \neg\exists c_j: salient(c_i,c_j)=c_j\]
which gives the most salient context that is active for actor $a$ when performing $\gamma$.
\subsection{Beliefs and Assumptions}
From epistemic logic we will use especially the operators for "everyone in the set of actors $A$ believes" ($EB_A$) and "it is common belief between the actors in $A$" ($CB_A$), They are defined in the usual way:
\[ EB_A\phi \equiv \wedge_{a\in A} B_a\phi\]
\[ CB_A\phi \equiv \wedge_{i=0}^{\infty} EB^i_A\phi\]
It is exactly the property that several parts of the social practice are common beliefs that makes them work efficiently.
Given the epistemic operators we can now also define things like the purpose of an action (and a social practice) and expectations within a social practice. The intuitive meaning of the purpose of an action is the reason for which the action is performed. It is part of the intended result of the action and thus indicates that the action is performed with the \emph{goal} of achieving that result. Thus we define the purpose of an action $\alpha$ performed by $a \in A_c$ in a context $c$ as follows:
\begin{gather*} purpose(a,\alpha,c)=\phi \equiv
CB_{A_c} ((Salient(a,\alpha,c) \\
\wedge DO(a,\alpha))\rightarrow Goal(a,\phi) \wedge B_a([\alpha(a)]\phi)
\end{gather*}
If we talk about the general purpose of an action $\alpha$ in context $c$ we mean that whenever any actor performs that action we infer that it was done to achieve that purpose.
\begin{gather*} purpose(\alpha,c)=\phi \equiv \\
CB_{A_c} (\forall a\in A: (Salient(a,\alpha,c) \wedge DO(a,\alpha))\rightarrow \\
Goal(a,\phi) \wedge B_a([\alpha(a)]\phi)
\end{gather*}
We can make one more abstraction if we let $\alpha$ not be a concrete action but the representation of an abstract action. I.e. $\gamma$=$\alpha\psi$ or in general let $\gamma$=$\cup_{i=1...n}\alpha_i$ then
\[ purpose(\gamma,c)=\phi \equiv \forall \alpha_i: purpose(\alpha_i,c)=\phi\]
\begin{table*}[!th]
\centering
\caption{Social Practice application (A: informal; B: formal)}\label{table:sp}
\begin{tabular}{|M{2.4cm}|M{0.6cm}|M{6.4cm}|M{7.2cm}|N} \hline
\textbf{Social Practice}&&A: Lecture SP (informal)&B: Lecture SP (formal)&\\ \hline
\textbf{Context} && &\\
\multicolumn{1}{|r|}{Roles}&$R_{sp}$ &lecturer, student & \textit{l}, \textit{s}\\
\multicolumn{1}{|r|}{Actors}&$A_{sp}$&people in the room & $a_1, a_2, ..., a_n$\\
& & $play(a_1,l)$, $\forall a_i, i>1: play(a_i, s)$&\\
\multicolumn{1}{|r|}{Resources} &$O_{sp}$&seats, projectors, whiteboard, markers & $o_1,...o_m$\\
\multicolumn{1}{|r|}{Affordances} &$\text{Af}_{sp}$ &sit on, present, write with & $\forall o_i$ \emph{afford}$(c_i, \text{\textit{aff}}_j)$, $\text{\textit{aff}}_j \in \{sit, present, write\}$ \\
\multicolumn{1}{|r|}{Places} &$\text{Pl}_{sp}$& lecture room & \textit{room}\\ \hline
\textbf{Meaning} && &\\
\multicolumn{1}{|r|}{Purpose} &$P_{sp}$&All students learn topic & $purpose(attend,sp)= \newline \forall a\in: plays(a,s), learntopic(a)$\\
\multicolumn{1}{|r|}{Promoted Values} &$\text{Pv}_{sp}$& attending lecture \textit{promotes} Self Enhancement, & $\forall i>1: promotes(sp, attend(a_i), SelfEnhance)$\\
&& noisy students \textit{demotes} Respect& $\forall i>1: promotes(sp, noisy(a_i),\neg Respect)$\\
\multicolumn{1}{|r|}{Counts-as} &$\text{Co}_{sp}$& student raising hand counts as has a question, & $countsas(sp,raise(s,hand),question(s))$\\
&& lecturer raising hand counts as start lecture & $countsas(sp,raise(l,hand),start(lecture))$ \\ \hline
\textbf{Expectations} && &\\
\multicolumn{1}{|r|}{Plan Patterns}&$\text{PP}_{sp}$& See Figure \ref{fig:pp}& $\alpha_1(lstarted);$\\
&&& $(\alpha_2(presented)+\alpha_3(q\&a'ed));\alpha_4(lended)$\\
\multicolumn{1}{|r|}{Norms} &$N_{sp}$& when lecturer talks students are quiet, & $F(s,talk(l),\neg talk(s))$\\
&& student must raise hand when has a question & $O(s,talk(s),raise(s,hand))$\\
\multicolumn{1}{|r|}{Strategies} &$S_{sp}$& students sit, & $strategy(\top, DO(s, sit))$\\
&& lecturer stops talk when student has question&$strategy(raise(s,hand), DO(lec,stop(talk)))$\\
\multicolumn{1}{|r|}{Start Condition} &$\text{Sc}_{sp}$& 9am, students present, lecturer declares start&$strategy(startcondition, DO(all, lecture))$\\
\multicolumn{1}{|r|}{Duration} &$D_{sp}$& 11am or lecturer finishes or students gone& $time(11) \vee finish(l) \vee \forall a: play(a,s), leave(a)$\\ \hline
\textbf{Activities} && &\\
\multicolumn{1}{|r|}{Possible Actions} &$\text{Ac}_{sp}$& sit, talk, raise hand, ...& $sit(X),talk(X), raise(X,hand),...$\\
\multicolumn{1}{|r|}{Requirements} &$\text{Re}_{sp}$& topic of lecture is known& $CB_A(sp,topic(x))$\\
&& lecturer: knows topic, can present& $\forall a_i, play(a_i,l)$: \\
&& & $cap(a_i, talk)$, $cap(a_i,raise(a_i,hand))$ \\
&& student: has prereq. knowledge for lecture & $\forall a_i, play(a_i,s)$: \\
&& & $cap(a_i, sit)$, $cap(a_i,raise(a_i,hand))$\\ \hline
\end{tabular}
\end{table*}
We have similar definitions for the purpose of complex actions. Let $\gamma \in \sl CA$ then
\begin{gather*}
purpose(a,\gamma,c)=\phi \equiv \\
CB_{A_c} (Salient(a,\gamma,c) \wedge \forall \alpha_i\in act(\gamma):\\
DO(a,\alpha_i)\rightarrow Goal(a,\phi) \wedge B_a([\gamma(a)]\phi))
\end{gather*}
Note that this definition only indicates that the whole sequence $\gamma$ leads to $\phi$ and does not restrict exactly which actions are part of that sequence. However, every action is believed to be done because $a$ has the goal to achieve $\phi$. In this way we represent the intuition that each action of the complex action $\gamma$ contributes to the goal $\phi$.\\
We can generalize this definition first by allowing the different actions in $\gamma$ to be done by different actors $a_j$ from a set of actors $A_c$. And we can also abstract altogether from the agency and give the purpose of the complex action $\gamma$. E.g. checking the minutes of the previous meeting at the start of a formal meeting is done for the purpose of creating common ground and agreement about what happened in last meeting.
Let $act(\gamma)=\{\alpha_1;...;\alpha_n\}$ and $A_c=\{a_1,...,a_m\}$ then
\begin{gather*}
purpose(A_c,\gamma,c)=\phi \equiv \\
CB_{A_c} (Salient(A_c,\gamma,c) \wedge \forall \alpha_i\exists a_j:\\
DO(a_j,\alpha_i)\rightarrow Goal(a_j,\phi) \wedge B_{a_j}([\gamma]\phi)
\end{gather*}
and of course we can talk about the purpose of an abstract sequence of actions $\delta$=$\cup_{i=1...n}\beta_i$:
\[ purpose(\delta,c)=\phi \equiv \forall \beta_i purpose(\beta_i,c)=\phi\]
Using the above definitions we can now give a definition of plan patterns in a social practice $sp$.
Let PP be the smallest set closed under:
\begin{gather*}
\gamma\phi \in PP\\
\gamma_1\phi_1,\gamma_2\phi_2 \in PP \Rightarrow \gamma_1\phi_1 +\gamma_2\phi_2 \in PP \\
\gamma_1\phi_1,\gamma_2\phi_2 \in PP \Rightarrow \gamma_1\phi_1 \&\gamma_2\phi_2 \in PP \\
\gamma_1\phi_1,\gamma_2\phi_2 \in PP \Rightarrow \gamma_1\phi_1 ;\gamma_2\phi_2 \in PP\\
\gamma\phi \in PP \Rightarrow (\gamma\phi)* \in PP
\end{gather*}
Let $\gamma\phi,\gamma_1\phi_1 \in PP$ then we use $\gamma_1\phi_1 \in \gamma\phi$ if $\gamma_1\phi_1$ occurs in $\gamma\phi$.\\
This leads us to the following definition of a plan pattern of a social practice:
\begin{gather*}
planpattern(\gamma\phi,sp) \rightarrow purpose(\gamma\phi,sp)=\phi \wedge \\
\forall\gamma_i\phi_i \in \gamma\phi: purpose(\gamma_i\phi_i,sp)=\phi_i \wedge
\forall\gamma_1\phi_1;\gamma_2\phi_2 \in \gamma\phi:\\ strategy(DONE(A',\gamma_1\phi_1),DO(A''\gamma_2\phi_2),sp)
\end{gather*}
where $A',A'' \subset A_{sp}$.
This states that plan patterns of a social practice are abstract (sequences of) actions for which the purpose in the context of the social practice is to reach the formula $\phi$ associated with that abstract action. And if the plan pattern specifies two abstract actions in a sequence there is an expectation that when the first part is done (by a subset $B$ of the agents) the second abstract action will be done. (See section 4.3 for more explanation on the strategy relation).
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{LectureSPpp.pdf}
\caption{Plan Pattern for SP Lecture}
\label{fig:pp}
\end{figure}
\subsection{Meanings}
Besides the above elements that are used in several parts of the formal representation of social practices we also have some more ontological elements that are used to describe meanings and resources. Due to space limitations we will not give complete formal definitions of these elements but refer to other work that gives the (intricate) logical formalization. First of all we define the counts-as relation
\[countsas(sp,\alpha_1,\alpha_2)\]
to mean that performing action $\alpha_1$ in context $sp$ is seen as performing action $\alpha_2$. E.g. the lecturer raising her hand counts as starting the lecture. A further formal definition of this concept of counts-as can be taken from \cite{Grossiphd}.
The second concept that we need is the fact that an action is promoting or demoting a certain value. E.g. students talking during a lecture demotes the value of respect. This concept is used to indicate some abstract social effects of actions. Again this relation is context dependent. We follow \cite{Dignumigpl} and use $play(a,r)$ to indicate that actor $a$ enacts role $r$. Let $sp$ be a social practice $\alpha \in \sl CA$ an (abstract) action, $r$ a role and $v \in V$ a value,
\[play(a,r) \rightarrow promotes(sp,\alpha(a),v)\]
indicates that if an actor $a$ plays the role $r$ and performs the action $\alpha$ she promotes the value $v$. For a more formal characterization of the promotes relation see \cite{Weidephd}.
Finally we also introduce the concept of affordances here. Affordances are used to describe the type of actions that are normally expected to be performed with an object and can also be used to describe types of objects in an abstract way. E.g. an object that affords to sit (which can be a chair or a bank or ...). The affordance relation is denoted as:
\[\emph{afford}(o,\alpha)\]
where $\alpha \in \sl CA$ and $o$ denotes an object. One could make the affordance relation also context dependent, but as argued in~\cite{oijen2014} a context independent approach suffices.
\subsection{Expectations}
Many elements of the social practice can be seen as expressing a kind of expectation. As stated before expectations with social practices come in different forms. The plan patterns can be seen as expectations because they indicate the general patterns of behavior that are expected. However, we also have more specific expectations. Although not all parts of a social practice are fixed there are some points where specific types of actions are expected. E.g. if in a greeting practice one person extends his hand it is expected that the other person shakes the hand.
Let $\gamma \in \sl CA$ and let $B\subset A_{sp}$ be a set of actors (possibly one), then we define strategies (as in ADICO) as follows:
\begin{gather*}
strategy(\phi, DO(B,\gamma),sp) \equiv\\
CB_{A_{sp}}(CB_B\phi) \rightarrow \forall a\in B:CB_{A_{sp}}(DO(a,\gamma(B))
\end{gather*}
Thus if all actors (involved in the social practice $sp$) believe that $B$ believes the condition $\phi$ then they all believe that all actors in $B$ will perform their part of $\gamma$ next. The condition $\phi$ can be of the form $DO(C,\gamma_1)$ or $DONE(C,\gamma_1)$ or a description of other facts. The first two indicate a synchronization and sequencing of actions, while the last represents that it is expected that $B$ takes the initiative to perform some action when a certain state is reached.\\
Note that the above definition entails that if actor $a \in B$ then she believes that if she does her own part of $\gamma$ the rest of $B$ will do their part of $\gamma$.
As stated before norms are also a kind of expectations. However, instead of a belief that an action will be taken they lead to a normative statement that an action \emph{should} be taken. In the context of social practices, the norms are connected to the roles of the social practice. Thus any actor that enacts a role in the social practice should follow the norms specified for that role.
Norms are now described as follows:
\[ O(r,\phi,\gamma) \equiv \forall a: play(a,r) \wedge B_a(\phi) \rightarrow O(\gamma(a))\]
\[ F(r,\phi,\gamma) \equiv \forall a: play(a,r) \wedge B_a(\phi) \rightarrow F(\gamma(a))\]
where $O(\gamma)$ and $F(\gamma)$ have the standard dynamic deontic logic semantics.
\section{Representing and Using Social Practices}
In the previous sections we have described all ingredients for representing social practices, which allowed us to fill table \ref{table:sp} completely. Formally a social practice can now be defined as an operator with as arguments all the elements from the table.
\begin{gather*} socialpractice(sp,R_{sp},A_{sp},O_{sp},Af_{sp},Pl_{sp},P_{sp},\\
Pv_{sp},Co_{sp},PP_{sp},N_{sp},S_{sp},Sc_{sp},D_{sp},Ac_{sp},Re_{sp})
\end{gather*}
However in order to achieve the goal of using social practices we should have social practices that can guarantee their fulfillment and give enough expectations to facilitate efficient planning and deliberation.
A social practice $sp$ is useful ($\emph{useful}(sp)$) iff:
\begin{enumerate}
\item for all actions necessary to create at least one trace $t$ that fulfills the plan pattern, norms and all relevant strategies of $sp$ there are objects and actors with capabilities that make those actions possible
\item if $t$ contains a sequence of actions $\alpha_1(A);\alpha_2(B)$ with $A\not=B$ then $sp$ contains a strategy $strategy(DONE(A,\alpha_1),DO(B,\alpha_2),sp)$
\end{enumerate}
The first item indicates that there should be at least one way to fulfill the social practice without violating expectations. The second item indicates that interactions between actors in a social practice are all regulated by strategies and thus do not require explicit synchronization.
In the remainder of this section, we will briefly describe how an actor will use social practices to deliberate about its context and plan its actions. We consider an extremely simplified scenario of attending a lecture on \textit{social practices}. Table \ref{table:sp} describes this social practice along the elements specified in Section \ref{sec:spelements}, providing an informal description and the formal specification using the language introduced in Section \ref{sec:logic}.
We now consider an actor $a$ intending to learn about SP. Assuming this actor to follow a BDI-like architecture, the goal ``know about SP'' would be part of the Intentions of $a$. We further assume that both the social practice described in table \ref{table:sp} and the prerequisite knowledge are part of $a$'s beliefs. Actor $a$ also has the capability to sit leading $a$ to belief that it could fulfill the role of \textit{student}. By using a social practice-based deliberation process, once actor $a$ beliefs to be itself in the context of the \textit{Lecture SP}, it can then form a plan for its participation in the lecture, based on the expectations, meaning and activities the social practice describes for actors of role `student'. This plan is very simple, $a$ sits in the lecture room and leaves at the end of the lecture. While $a$ is sitting it can ask questions when it does not understand the presentation. This is a reactive behavior that is governed by a strategy and a norm. Thus all parts of the plan for the student are given. In fact, based on the plan pattern of sp $a$ can conclude that by fulfilling its part, the desired result will be achieved, given that it expects all other agents to fulfill their own parts.
For the lecturer there is a bit more planning left. The lecturer has to plan the sequence of actions that will instantiate the abstract "present" action.\\
Note also that the current social practice does not require the students to actually pay attention. In fact, the (expected) social effects of actions is an important issue but of the scope of this paper. Social effects are not visible but can be assumed given the social practice. E.g. if no questions being asked can be taken to mean that students understand the subject so far, and therefore no explicit checking is needed.
\section{Conclusions}
In this paper we gave a first formalization of social practices that aims at support actors planning social interactions. We have briefly shown that {\emph useful} social practices can simplify the amount of actions that have to be planned by an actor and that it can make assumptions about interactions with other actors in the social practice without explicitly having to coordinate.
Due to space limitations we cannot present the full logic nor discuss the formal properties of social practices that follow from its formalization. We leave these for future work. Another important area for future work is the development of learning mechanisms that enable efficient matching of sensing information and beliefs to the components of a social practice description.
\clearpage
\newpage
\bibliographystyle{named}
|
\section{Introduction}
\renewcommand*{\thefootnote}{\fnsymbol{footnote}}
\footnotetext[1]{Work done during an internship at ETH Z{\"u}rich.}
\renewcommand*{\thefootnote}{\arabic{footnote}}
In the word embeddings literature, it has previously been of interest to find interpretable representations:
individual dimensions should capture a distinct semantic meaning, such that humans are able to understand why a word is
encoded in a particular vector.
With a cognitive plausibility argument from~\citet{nnse}, the interpretability can be linked to sparse representations:
they argue that the representation should model a wide range of features in the
data and that every sample should be characterized by the presence of a small number of key features.
\citet{arora}~use this idea to recover and disentangle the different meanings of polysemous words.
The above\-/named approaches, as well as those by~\citet{spine,spowv}, recover an interpretable sparse representation
in a separate, post-processing step on top of the uninterpretable, dense embeddings of the original model (often
word2vec or GloVe).
This is commonly done using sparse coding or a downstream model.
Additionally to understanding a model's intermediate representation, there has been work
on constructing models that inherently use a sparse embedded representation by learning it during the training
process~\citep{sun2016sparse,chen2017learning}.
This is motivated by the idea that the model should include the prior that each word is a sparse combination of
disentangled features from the very beginning.
In contrast, when computing dense embeddings first, it is less likely that this representation will be easily
disentanglable in the post-processing step.
\citet{thoughtvectors} argues that sparse representations can be used to explain image and sentence embeddings as well.
To be precise, the author focuses on encoder-decoder neural networks and uses sparse coding to recover interpretable
features in the latent spaces of a variational autoencoder~\citep{vae} and an image captioning system based
on~\citep{showandtell}.
In this paper, we aim to use sparse methods to disentangle sentence embeddings' dimensions.
We focus on a simple sentence autoencoder model, and apply both a sparse-coding-based post-processing
technique, as well as model constraints during training time, to obtain sparse vector representations of sentences.
We aim to increase the understanding of the latent space, which helps us gain insight into how the inference and
learning process works by identifying the patterns in the data that the model learns to recognize and
encode in this representation.
To compare our different approaches, as well as measure the improvement compared to the baseline of a dense autoencoder model,
we introduce a novel, quantitative and automated metric of the mentioned interpretability properties.
It is based on the notion of topic coherence and further develops it for the case of sentences.
We observe that the new measure reflects our manual judgment on the interpretability of the embeddings.
Additionally, we track reconstruction quality and performance in downstream tasks, showing that sparse approaches
can obtain a remarkable increase in interpretability at a moderate cost in quality.
\section{Models}
\label{sec:models}
Our models are based on a standard recurrent neural network autoencoder following the \emph{Sequence-to-Sequence}
architecture~\citecomment{seq2seq;}{seq2seq}{}.
This architecture is based on the encoder-decoder scheme, where an encoder network maps the input to a dense, embedded
representation $\mathbf{z}$, and a decoder net reconstructs the input from $\mathbf{z}$.
In Section~\ref{subsec:experiments}, we give a more detailed description of our experimental setup.
\subsection{Enforcing Sparsity by Post-Processing Dense Embeddings}
Consider a dataset $x_1, \ldots, x_N$ of $N$ sentences.
We train a dense autoencoder net with a hidden state size $D'=500$ to convergence, and compute
$\mathbf{Z} = [\mathbf{z}_1, \ldots, \mathbf{z}_N]^\mathsf{T} \in \mathbb{R}^{N \times D'}$, the vector representations of the data arranged as
the rows of a matrix.
We follow the approaches by~\citet{arora,thoughtvectors} and compute a sparse representation of size $D=2000$
on top of $\mathbf{Z}$, where all but $k$ values have to be zero.
We do this by solving the following sparse dictionary learning problem:
\begin{equation}
\begin{split}
\mathbf{E}, \mathbf{U} = \arg \min_{\mathbf{E}, \mathbf{U}} ||\mathbf{E} \mathbf{U} - \mathbf{Z} ||_F^2, \\
s.t. \; ||\mathbf{e}_{i}||_0 \leq k, ||\mathbf{u}_{j}||_2 = 1, \; \forall i, j,
\end{split}
\end{equation}
whereby we obtain $\mathbf{E} = [\mathbf{e}_1, \ldots, \mathbf{e}_N]^\mathsf{T} \in \mathbb{R}^{N \times D}$, a set of new, sparse vector representations,
and $\mathbf{U} = [\mathbf{u}_1, \ldots, \mathbf{u}_D]^\mathsf{T} \in \mathbb{R}^{D \times D'}$, a dictionary of \emph{atoms} found in $\mathbf{Z}$.
We solve this problem with the $k$-SVD algorithm~\citecomment{}{ksvd}{%
we use an open source implementation called \emph{pyksvd}\footnote{\url{https://github.com/hoytak/pyksvd}}}.
The intuition behind this sparse coding approach is as follows.
The atoms $\mathbf{U}$ are intended to represent a wide range of the 2000 most important features that explain the data in the
dense latent space of the model.
By solving this problem we decompose the intermediate representation $\mathbf{z}_i$ of a sample $x_i$ into a linear combination
$\mathbf{e}_i$ of atoms.
By constraining $\mathbf{e}_i$ to a fixed and low sparsity level $k$ we aim to disentangle this representation and therefore
increase interpretability.
We refer to this representation as the $k$-SVD model.
\subsection{Enforcing Sparsity during Embedding Learning}
The $k$-SVD model proposed in the previous section obtains sparse representations through solving two independent
problems: finding a fixed-size vector representation for sentences with a neural model and, in a separate step, mapping
it to an interpretable, sparse representation.
As we mention in the introduction, we conjecture that interpretability can be further increased with an end-to-end
approach.
In this section, we introduce modifications to the model architecture that will force the neural nets to encode and
understand sparse representations of the data during training time.
We propose an additional layer that is inserted between the encoder and decoder net.
We map the vector representation $\mathbf{z}$ to a vector $\mathbf{e}$ of the same size in a \emph{sparsity transformation}.
The only requirements for this mapping are that the output $\mathbf{e}$ is sparse and differentiable (or that we can
define a custom gradient) to allow backpropagation through it.
We then feed $\mathbf{e}$ through the decoder net instead of $\mathbf{z}$ and train the whole net end-to-end.
In the rest of this section, we propose two mappings for such a sparsity transformation.
\subsubsection*{$k$-Sparse.}
For this model, we draw inspiration from the \emph{$k$-Sparse Autoencoder} by~\citet{ksparseae}.
We again introduce a hyperparameter $k$ and define $\mathbf{e}$ by keeping the $k$ largest activations in $\mathbf{z}$,
the \emph{support set}, and setting all other units to zero.
We backpropagate only through the support set.
This is a simple way of enforcing a hard specification for the sparsity level as an integral part of the model.
\subsubsection*{Sparsemax.}
The $k$-Sparse approach has the drawback of requiring a fixed sparsity level for all samples.
To allow for a variable, per-sample sparsity level, we use the \emph{Sparsemax} layer, introduced by \citet{sparsemax}.
Sparsemax is an alternative to Softmax---however, unlike Softmax, it is able to return sparse probability distributions.
It is defined as:
\begin{equation}
\text{Sparsemax}(\mathbf{z}) = \arg\min\limits_{\mathbf{p} \in \Delta^{D - 1}} ||\mathbf{z} - \mathbf{p}||_2^2,
\end{equation}
where $D$ is the dimensionality of $\mathbf{z}$ and $\Delta^{D - 1}$ is the $(D-1)$-dimensional simplex
$= \cbraces{\mathbf{p} \in \mathbb{R}^D \, | \, \mathbf{1}^\mathsf{T} \mathbf{p} = 1, \mathbf{p} \geq 0}$.
Similar to Softmax, Sparsemax supports a temperature mechanism, where a hyperparameter
$\tau$ trades off the ``confidence'' in the output probability of the largest input unit.
To be precise, as $\tau$ approaches $0$, the probability distribution $\text{Sparsemax}\left(\frac{\mathbf{z}}{\tau}\right)$
approaches the distribution peaked on the maximum components of $\mathbf{z}$.
Additionally to Softmax, Sparsemax has the property that this output distribution becomes increasingly sparse.
Putting this together, we introduce a hyperparameter $\tau$ and define a sparsity transformation by defining
$\mathbf{e} = \text{Sparsemax}\left(\frac{\mathbf{z}}{\tau}\right)$.
\section{Experiments}
\label{subsec:experiments}
\subsection{Training Details}
In our experiments, we use a vocabulary size of 20,000, with the symbolic words \texttt{<person>}, \texttt{<unk>}, and
\texttt{<eos>} for names and out-of-vocabulary (OOV) words in the dataset, and the end-of-sentence marker, respectively.
We convert words to 100-dimensional word embeddings by looking them up in a trainable matrix $\mathbf{V}$ (note that, in
general, this matrix is not sparse---sparsity is only imposed on the latent space of the sentences\footnote{%
On a side note, sparsity can be imposed on the word embeddings by adding an L1\-/regularizer to
$\mathbf{V}$~\citep{sun2016sparse,chen2017learning}.}).
Our encoder and decoder nets are recurrent neural networks that use a single GRU~\citep{gru} layer.
They have the same hidden dimensionality but share no parameters.
We obtain the model predictions as the Softmax of a learned, affine transformation to 20,000-dimensional space at every
time step of the decoder net.
We minimize the mean cross-entropy loss over all timesteps.
We use a batch size of 64 and the Adam optimization algorithm.
\subsection{Data}
We train our models on the Cornell Movie-Dialogs Corpus and MS Common Objects in Context
datasets~\citecomment{respectively}{cornell,coco}{}.
The Movie-Dialogs Corpus is a collection of movie lines, therefore it contains a wide variety of different utterences
and allows us to explore general-purpose sentence embeddings.
We preprocess this data by splitting the movie lines into separate sentences, thereby obtaining more than 500,000
samples.
This dataset has no predefined split;
we define a validation and test set by setting aside 50,000 samples each.
The MS COCO dataset contains images showing scenes with objects in numerous configurations.
Every image contains 5 human-annotated variations of a caption that describe the scene.
In our experiments, we use only these captions and refer to this as the COCO Captions data.
They total over 600,000 samples and allow us to explore sentence embeddings of a more narrow language:
since they merely describe objects and scenes, they tend to follow the same, simple sentence structure.
The dataset comes with a predefined training\slash validation split.
For tokenizing and splitting movie lines into sentences we use the NLP library SpaCy\footnote{\url{https://spacy.io}}.
All our models are implemented in TensorFlow~\citep{tensorflow}.
\input{reconstructed.tex}
\section{A Quantitative and Automated Evaluation Metric}
\label{sec:evaluation}
The most common quantitative interpretability measure for embeddings (in particular word embeddings) is
the intrusion test, first introduced in \citep{wordintrusion}.
This test involves generating 5-tuples of samples, where according to the embeddings model four are related and one
stands out.
The better human judges identify the intruder, the more interpretable the model is considered.
This evaluation method has the drawback of requiring human attention, thereby it is expensive and slow to evaluate.
For our evaluation, we introduce an automated interpretability test, based on topic coherence, that does not require
human attention.
We describe our method in this section.
A \emph{topic model} defines a set of topics in a corpus of documents and allows us to find the top $n$ most likely words
that belong to each topic.
\emph{Topic coherence} is an automated evaluation method of the interpretability of topic models, which has been shown to
correlate well with human assessments~\citep{topiccoherence,mimno2011optimizing}.
Given a symmetric similarity measure of two words (e.g.\ pointwise mutual information),
the coherence of a topic is defined as the mean pairwise similarity of all pairs of words.
The total topic coherence of the model is the mean coherence over all topics.
We devise an evaluation scheme based on topic coherence.
Instead of looking at words in topics, we consider the highest-ranked sentences in the dimensions of our embeddings and replace
the word similarity measure with a sentence similarity measure.
Let $x_d^{(p)}$ be the sample that has rank $p$ in the order given by the $d$-th dimension in the embedding.
For a similarity measure $sim_*$, the coherence of a single dimension $d$ is defined as:
\begin{equation}
\begin{multlined}
coh_*(d) = \frac{2}{n \cdot (n-1)} \sum_{p=1}^{n-1} \sum_{q=p+1}^{n} sim_*(x_d^{(p)},x_d^{(q)}).
\label{eq:tcformula}
\end{multlined}
\end{equation}
The coherence of the model is defined as the mean coherence over all dimensions:
\begin{equation}
coh_*(1, \ldots, D) = \frac{1}{D} \sum_d coh_*(d).
\label{eq:tcformulaall}
\end{equation}
In addition, to determine how much the coherence deteriorates when looking beyond the top ranks, we consider all non-zero samples
of a dimension and we evaluate Equation~\ref{eq:tcformula} on $n$ sentences sampled at random and without replacement from
$\cbraces{x_i \mid e_{i,d} \neq 0}$ instead of $x_d^{(1)}, \ldots, x_d^{(n)}$.
We compute this on the validation set of our data.
We strip all stop words from all sentences.
We consider $n=10$ sentences per dimension, unless a dimension has a non-zero value for less than $n$ samples, in which
case we compute Equation~\ref{eq:tcformula} on all pairs of sentences.
In the following, we define three choices for a sentence similarity measure $sim_*$.
\paragraph{Jaccard Similarity.}
We regard the sentences as sets of words and compute the Jaccard similarity:
\begin{equation}
sim_\text{J}(x_i, x_j) = \frac{|x_i \cap x_j|}{|x_i \cup x_j|}.
\end{equation}
\paragraph{BoW Similarity.}
We consider the Bag-of-Words (BoW) vectors $\mathbf{b}_i, \mathbf{b}_j$ of the two sentences, i.e.\ the vectors with the number of
occurrences of each vocabulary word in $x_i, x_j$, respectively.
The similarity is defined as the cosine of the angle between these vectors:
\begin{equation}
sim_{\text{BoW}}(x_i, x_j) = \frac{\mathbf{b}_i^\mathsf{T} \mathbf{b}_j}{||\mathbf{b}_i||_2 \, ||\mathbf{b}_j||_2}.
\end{equation}
\paragraph{WMD Similarity.}
The Jaccard and BoW similarity measures have a drawback in that they do not take semantic relatedness of different
words into account.
The \emph{Word Mover's Distance} \citecomment{WMD;}{wmd}{} remedies this problem:
the authors define a document distance measure that relies on the word2vec latent space to make a better assessment
of the semantic distance of sentences, based on the distance of the words they consist of.
We use the negative WMD to obtain a similarity measure:
\begin{equation}
sim_{\text{WMD}}(x_i, x_j) = -\text{WMD}(x_i, x_j).
\label{eq:wmdsimilarity}
\end{equation}
\section{Results}
\label{sec:results}
\subsection{Reconstruction Quality}
\input{topsamples.tex}
We start off by looking at the amount of information lost by our models due to sparsity constraints.
In general, we observe that as the sparsity level is decreased, the reconstructions start to deteriorate.
At low values of $k$, our sparse models often fail to restore the exact meaning or phrasing, but still generate
sentences with correct grammar and related topics.
For example, they turn ``\texttt{waiting at a bench}'' into ``\texttt{sit at a park}'', ``\texttt{sidewalk}'' into
``\texttt{curb}'', ``\texttt{two cars}'' into ``\texttt{two buses}'', and similar.
The $k$-SVD model generally does this less than the other models but in some cases it fails as well.
See examples of typical reconstruction errors by our $k$-Sparse, $k=15$ model in Table~\ref{tab:samplesentences}.
\subsection{Highest-Ranked Samples}
\label{subsec:highestrankedsamples}
We examine the top samples in the dimensions of our embedding models and observe that sparse models often group
sentences s.t.\ they have a common syntactic element or talk about a common concept.
For example, in our $k$-Sparse, $k=15$ model trained on the COCO Captions dataset, we identify dimensions that
represent sentences about objects in \texttt{water}, people \texttt{holding} things, \texttt{horse} (and occasionally
\texttt{bicycle}) \texttt{riders}, sentences starting with common prefixes such as \texttt{there is a [\ldots]}, etc\@.
We give examples in Table~\ref{tab:rnntopsamples} and in the appendix in Table~\ref{tab:appendixtopsamples}.
For some dimensions, this pattern is not only recognizable in the top ranks but for all samples $x_i$ with
$e_{i,d} \neq 0$.
We are able to find such patterns in all sparse models, but the lower the sparsity level, the more apparent these
patterns become.
$k$-SVD based models exhibit these properties to a lesser extent.
\input{resultstable.tex}
\subsection{Quantitative Evaluation}
In Table~\ref{tab:rnntopsamples} we additionally report the coherence $coh_\text{WMD}(d)$ of the presented dimensions $d$
(see Equations~\ref{eq:tcformula},~\ref{eq:wmdsimilarity}).
We observe that this score correlates with our empirical assessment of the interpretability of the dimension.
For example, we observe on the COCO dataset that, while unrelated groups of sentences usually have a coherence score
of $< -3$, sentences with common or semantically related subjects and objects have higher coherence scores (usually
between $-2.8$ and $-2.2$).
Groups of sentences with very close semantic meaning or large common prefixes have coherence scores around $-2$ or
higher.
We report the topic coherence of our models (Equation~\ref{eq:tcformulaall}) in
Table~\ref{tab:tcevaluation}.
As rough reference values for the metrics, we include the mean similarity of pairs of random sentences from the
dataset (estimated on 500 randomly sampled pairs), and the topic coherence of a 500-dimensional dense
autoencoder model.
In accordance with our empirical observations, we see an increase in interpretability in the sparse models.
For example, on the COCO Captions data, a random pair of sentences has a WMD-based similarity of -3.12,
and the WMD-based coherence score of a dense autoencoder model is -3.
With the additional sparse coding step on top of that, we can increase the coherence to -2.86.
\subsection{Downstream Tasks}
\label{sec:downstream}
\input{senteval.tex}
Additionally to the interpretability properties of sparse sentence embeddings, it is of interest whether
sparsity decreases their usefulness in downstream tasks. To evaluate this, we use the \emph{SentEval}
framework~\citep{senteval}, which learns downstream models on top of the provided sentence embeddings
to solve a variety of transfer tasks.
We report the accuracy on the standard classification problems the framework provides, namely
binary sentiment of movie reviews (MR), movie lines (SST2) and product reviews (CR), five-class sentiment of movie lines
(SST5), subjectivity/objectivity (SUBJ), binary opinion polarity (MPQA), and six-class question type (TREC) classification.
To look at semantic entailment/similarity of pairs of sentences, we report Spearman correlation with human-labelled
ground truth on the five-class semantic relatedness (STS14, SICK-R), and three-class semantic entailment (SICK-E) tasks,
and accuracy on the binary paraphrase detection (MRPC) task.
We configure the framework to use Logistic Regression for downstream models.
More details on the tasks and evaluation methods can be found in the SentEval paper.
We evaluate these tasks on the Movie-Dialogs models only, because COCO is unsuited for general-purpose sentence
embeddings.
We show the results of this evaluation in Table~\ref{tab:senteval}.
We again observe that sparse representations perform, in many cases, worse than their dense equivalent, therefore,
trading quality for interpretability\footnote{%
On a side note, we address the noticeable fact that the transfer tasks are solved with low accuracy in general.
For numbers comparable to the state of the art literature, more powerful sentence embedding models
\citecomment{such as self-attentive networks, InferSent, SkipThought etc., see}{infersent,skipthought}{}
with a higher latent dimensionality, more layers, and a larger and more diverse dataset are required.
Further, SentEval provides slower but more powerful MLP downstream models instead of Logistic Regression.}.
However, this does not occur on all tasks: for example, SST2 and SST5 clearly benefit from a sparse representation.
\subsection{Discussion}
The results of our quantitative evaluation method confirm the tendencies we observed in our empirical evaluation.
It appears that embedding dimensions generated by sparse models are coherent to a higher extent---in particular, the
lower the sparsity level, the more apparent topics can be found in the embedding dimensions.
The price of good interpretability is a higher reconstruction error.
As we impose more sparsity in the representation, the model is forced to ``cut corners'' and single slots in the embedding
are designated broader collections of traits in the data.
This results in more coherent topics, however, the narrow information bandwidth reduces the decoder net's
ability to reconstruct the exact sentence.
The fact that sparse representations carry less information may also explain the lower utility in some of the downstream tasks.
Other tasks (e.g.\ sentiment classification) can be solved with greater accuracy, which suggests that a sparse and
interpretable representation discovers more useful features for a simple downstream model like Logistic Regression.
As we force a model to deal with a sparse representation already during the training phase, finding sensible atoms
gets incorporated into the encoding and decoding mechanism.
We found that, in comparison to extracting this information from a dense model's intermediate
representation, this results in an observable and measurable boost in interpretability.
We can explain this by the fact that this architecture makes it \emph{part of the model's task} to find a sparse and
accurate representation of the data, whereas in a post-processing approach the model focuses on reconstruction only.
On the other side of the coin, this modification interferes with the training process.
To be more precise, the model follows a more complex objective, and the sparse layer is limited in the amount of
information that can be forward and backpropagated through it at a time---hence we observe convergence at a higher
loss value and bigger reconstruction errors.
We note that our Sparsemax-based approach does not perform particularly well in our evaluation, although in some
cases it outperforms other approaches when we consider samples beyond the 10 highest-ranked.
This can be explained by the fact that the sparsity level is not fixed, and that due to the Sparsemax layer,
the embeddings $\mathbf{e}_i$ are valid probability distributions.
A high value $e_{i,d}$ does, therefore, not necessarily indicate a strong presence of feature $d$ in sample $i$, but also
a lack of other features.
On the other hand, Sparsemax is better at determining feature presence\slash absence in general, due to not being
constrained to find an exact number of features.
\section{Related Work}
\label{sec:relatedwork}
As aforesaid, there has been work in the NLP literature on the interpretability of word embeddings.
\citet{nnse} suggest that sparse embeddings can be linked to a disentangled, and thus interpretable, representations.
This idea is also applied in~\citep{arora,spowv,spine}, commonly by solving a sparse dictionary learning problem on top
of dense embeddings.
In the papers \citep{sun2016sparse,chen2017learning,oiwe}, the authors learn sparse word embeddings during the training
phase.
\citet{thoughtvectors} applies above-named approaches to image embeddings, and the intermediate representation of an
image captioning model.
\citet{ksparseae} define the $k$-Sparse autoencoder. They use a $k$-Sparse layer in a shallow autoencoder
trained on the MNIST and NORB datasets, focusing on unsupervised feature learning, improvement in classification
accuracy, and a fast alternative to sparse coding.
\citet{sparsemax} develop Sparsemax as an alternative to the Softmax layer that is able to output exactly zero
probabilities, their work is focused on classification problems and attention mechanisms.
Interpretability metrics are usually of interest for word embeddings, where the predominant evaluation method is the
word intrusion test~\citep{wordintrusion}.
Our interpretability metric is based on topic coherence~\citep{topiccoherence}, a comparison of different variants of
this method can be found in~\citep{exploringtc}.
\section{Conclusion}
\label{sec:conclusion}
Being able to understand the intermediate representation of a neural net increases our model understanding.
In this paper we have taken a step towards this goal by introducing several sparse methods for a sentence autoencoder,
inspired by previous work on word embeddings.
The evaluation of our proposed models supports our hypothesis that sparse methods benefit the interpretability of the
embedding.
It is intuitive that a vector restricted to many zero values inevitably carries less information,
and indeed we have found that this increase in interpretability comes at a cost in reconstruction quality and, in some
cases, utility in downstream tasks.
It is, however, possible to strike a balance and achieve good interpretability without a large penalty.
We have devised a novel, automated method of quantifying said interpretability, based on topic coherence.
In our experiments, we observe that this evaluation corresponds to our manual assessment of interpretability.
It is fully automated, and therefore cheap and fast to run.
It can easily be extended by using different sentence similarity metrics or other topic coherence variants.
An interpretable sentence representation has further applications beyond model understanding: for example, it allows us
to develop a sentence similarity measure, that can justify \emph{why} two sentences are similar.
It can also help us understand downstream models on top of sentence embeddings.
For example, consider the case of a linear classification model:
we can inspect the largest positive and negative weights and understand which features in a source sentence influence
the model's decision.
For future work, it suggests itself to apply sparsity constraints to more sophisticated sentence embedding models
such as \emph{SkipThought} or \emph{InferSent}~\citecomment{respectively}{skipthought,infersent}{}.
Our methods can also be used to construct sparse encoder-decoder models for further tasks, such as image captioning,
machine translation, or recommender systems.
\bibliographystyle{acl_natbib_nourl}
|
\section{\label{sec:level1}INTRODUCTION}
According to the traditional Landau-Ginzburg paradigm, states of matter are defined by the symmetries broken in thermal equilibrium that are preserved by the underlying Hamiltonian, and phase transitions acquire universal features that only depend on the symmetries involved and the spatial dimension. However, this definition proves inadequate for topological phases, in which the ground state wavefunction of the bulk system is characterized by a global, topological quantum number which distinguishes it from a conventional phase with the same symmetries \cite{moore2010birth, schnyder2008classification, hasan2010colloquium}. Naturally, critical points separating these phases fall outside the tradition paradigm as well. The most striking consequence of the non-trivial bulk topology is the presence of robust surface states where the bulk terminates. One of the most celebrated topological phases in condensed matter systems is the time-reversal symmetric strong topological insulator (TI) in three dimensions, which is characterized by a $\mathbb{Z}_2$ topological invariant $\nu=\text{odd/even}$ \cite{fu2007topological, chen2009experimental}. The surface manifestation of the bulk topology in this phase is the presence of an odd number of pseudo-relativistic, helical surface states that are robust against non-magnetic perturbations. Numerous materials have been predicted to be in this phase, and many of them have been experimentally confirmed. Additionally, several TIs can be tuned into trivial insulators with doping, thus allowing experimental access to the quantum critical point separating them.
A close cousin of the topological insulator is the time-reversal symmetric topological superconductor (TSC) in 3D \cite{schnyder2008classification}. (See Fig. 1). Here, the superconducting gap plays the role of the insulating gap of the insulator, the topological invariant is $\nu\in\mathbb{Z}=0,1,2\dots$, and the surface hosts $\nu$ helical Majorana fermions instead of electrons. These states have been shown to serve as a condensed matter realization of spontaneous supersymmetry breaking \cite{qi2009time,grover2014emergent}, while a related particle -- the Majorana zero mode, shown to exist in various heterostructures of superconductors and spin-orbit coupled systems including topological insulators \cite{mourik2012signatures, fu2008superconducting} -- has been extensively proposed as the building block for a fault-tolerant quantum computer \cite{kitaev2003fault}. Despite such remarkable theoretical predictions, however, 3D topological superconductors have been elusive experimentally. The main material platform that has been studied experimentally is the prototypical TI Bi$_2$Se$_3$ doped with Cu, in addition to theoretical predictions on a handful of rare earth compounds \cite{chadov2010tunable}. Unfortunately, undoped Bi$_2$Se$_3$ does not display superconductivity at ambient pressures, so a topological-to-trivial superconductor phase transition does not occur in this system.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.7\textwidth,,height=7cm]{Fig_one.pdf}
\caption{ \textbf{3D topological superconducting phase transition. } Upper-left: topological metal: Fermi level E$_f$ is placed above (or below) the topological surface state SS. Upper-right: Doping drains the bulk bands and lowers (raises) the Fermi level to expose the SS. Lower-left: Inducing even-parity superconductivity in the bulk of a topological metal gaps out the bulk and the Dirac SS, while (lower-right) odd-parity paring turns the bulk system into a 3D topological superconductor, which hosts in-gap, helical Majorana fermion surface states. While signatures of Majarana zero mode (MZM) have been demonstrated in several experiments (in superconducting TI and proximity induced superconductivity heterostructure), the 2D Majorana fluid unique to 3D TSC has been elusive. Importantly, this system could exhibit a topological quantum phase transition in the superconducting state. }
\end{figure*}
In this work, we investigate the transport properties of $\beta$-PdBi$_2$ single crystals with and without K doping. Our normal state magneto-transport experiments on K-doped $\beta$-PdBi$_2$ show that the longitudinal magnetoresistivity (LMR) exhibits weak anti-localization (WAL) at low temperatures, which is a signature of the presence of spin-momentum locked Dirac fermions, while the undoped system lacks any such features in LMR. Next, we probe superconductivity in these systems using various bulk and surface transport measurements. We find the undoped material to be a topologically trivial superconductor; in contrast, we find evidence that the doped system realizes a TSC. Specifically, point-contact spectroscopy (PCS) in the superconducting state shows signatures of Majorana surface states protected by time-reversal symmetry as theoretically predicted for time-reversal-invariant 3D topological superconductors. On studying the superconductivity in the bulk, we find that (i) the upper critical field exceeds the prediction by the Werthemer-Helfand-Hohenberg (WHH) orbital model for conventional $s$-wave pairing, but is consistent with the prediction for polar $p$-wave pairing, and (ii) the magnetization shows an anomalous behavior, characteristic of spin-triplet superconductivity. Fu and Berg showed that time-reversal symmetric odd-parity superconductors are topological if their normal state Fermi surface encloses an odd number of time-reversal invariant momenta \cite{fu2010odd}. Existing band structure calculations have found that this is indeed the case in $\beta$-PdBi$_2$ (See Fig. 2a) \cite{sakano2015topologically}. Thus, K-doped $\beta$-PdBi$_2$ is likely to be a topological superconductor, and could undergo an unprecedented 3D topological superconducting phase transition between a trivial and a topological superconductor. If there is an intermediate magnetic phase, the TSC-magnetism critical point would be a condensed matter realization of supersymmetry \cite{grover2014emergent}. In addition, the transport properties across the across the topological quantum phase transition can be useful for technological applications.
$\beta$-PdBi$_2$ has a layered, tetragonal crystal structure, belonging to the I4/mmm space group. Band structure calculations \cite{shein2013electronic, sakano2015topologically} show that the compound is a multi-band metal, with two hole-like and two electron-like pockets formed by Bi-6p and Pd-4d bands at the Fermi level ($E_f$). The Bi-6p and Pd-4d bands are already inverted without spin-orbit coupling (SOC) \cite{shein2013electronic, sakano2015topologically}, and the role of SOC is to drive the opening of a bulk band gap below $E_f$, i.e., to create a metallic band structure that can be transformed into that of an insulator by smoothly deforming the bands and lowering $E_f$. In addition, (spin-) angle-resolved photoemission spectroscopy (spin-ARPES) and quasiparticle interference imaging have revealed the presence of spin-polarized topological surface states around $E_f$ \cite{sakano2015topologically, iwaya2017full}. Intrinsic superconductivity robust to different dopants \cite{dopant, zhao2015chemical} and a relatively high T$_c$ compared to other systems make this material an attractive candidate for realizing topological superconductivity. Until now, however, transport properties of the topological surface states in single crystals of $\beta$-PdBi$_2$ have been difficult to measure due to the coexistence of trivial bands with non-trivial surface bands around $E_f$.
\section{\label{sec:level1} NORMAL STATE TRANSPORT PROPERTIES}
\begin{figure*}[]
\centering
\includegraphics[width=0.7\textwidth,,height=14cm]{fig1}
\caption{ \textbf{Transport properties of K doped and pristine $\beta$-PdBi$_{2}$ above $T_c$ } (a) Fermi surfaces of $\beta$-PdBi$_{2}$ as calculated in ref. \cite{sakano2015topologically} showing that the Fermi surfaces enclosed an odd number of time-reversal invariant momentum points. (b) Crystal structure of $\beta$-PdBi$_{2}$. Note that K atom is bigger than Pd and Bi atoms. (c) Symmetrized Hall resistivity of K doped $\beta$-PdBi$_{2}$ at 6 K showing the non-linear behavior characteristic of two channel systems (see Supplementary Information S2 Ref. \cite{Supplemental}). The pristine $\beta$-PdBi$_{2}$ data in c is from \cite{zhao2015chemical}. (d) Electrical resistivity along the a-b plane of pristine and K doped $\beta$-PdBi$_{2}$ as a function of temperature (T). (e) LMR in $\beta$-PdBi$_{2}$: There is no clear sign of WAL as the LMR scales with B$^2$. (g) LMR in K doped $\beta$-PdBi$_{2}$: WAL can be seen clearly from 15 K till the lowest temperature (6 K) in our experiment. At higher temperatures, the phase coherence length becomes shorter and quantum interference effects are lost, resulting in an LMR that scales with $B^2$ at 20 K and above in the low field range, and (f) is temperature dependence of the phase coherence length.}
\end{figure*}
We begin by briefly studying the transport properties above the superconducting transition temperature of pristine ($T_c$ = 5.3 K) and potassium doped (0.3 \%) $\beta$-PdBi$_2$ ($T_c$ = 4.4 K). In Fig. 2, we explore the temperature and magnetic field dependence of the longitudinal and Hall resistivity. Potassium can be an ambipolar dopant in this system. It can either enter interstitial sites and serve as an electron donor, or replace the elements and become an acceptor. We find evidence in the Hall resistance curves (Fig. 2c) that K acts as an acceptor. At above 20 K, it changes transport from electron- to hole-dominated upon doping, while the low temperature data exhibits a non-linearity characteristic of two-channel conduction -- a highly mobility (4280 cm$^2$/Vs) electronic channel and relatively lower mobility (32 cm$^2$/Vs) hole channel. (See Supplementary Information S2 Ref. \cite{Supplemental}). Following Ref. \cite{ando2013topological}, we attribute the high mobility to surface Dirac electrons and the low mobility to holes in the bulk. At higher temperatures, the quantum effect of the surface Dirac electrons is diminished and the bulk carriers dominate transport in the doped system, resulting in a classical, hole-dominated linear Hall resistance.
The normal state LMR $\rho_{xx}(B)$, of doped $\beta$-PdBi$_2$ (Fig. 2g) is compared with the pristine sample (Fig. 2e). The LMR of $\beta$-PdBi$_2$ lacks clear evidence of WAL as it scales with $B^2$ in the lower B field range. The doped crystal, in contrast, shows a sharp magnetoresistance cusp characteristic of WAL. WAL is a striking manifestation of the $\pi$ Berry phase of spin-momentum locked states on the surface of a TI, which protects them against localization by non-magnetic impurities \cite{ando2013topological}. This effect is quickly suppressed in the presence of magnetic field which breaks the time-reversal symmetry and results in a sharp increase in resistance, reflected by the cusp in the magnetoresistivity. The conductivity for small $B$ can be described by Hikami-Larkin-Nagaoka (HLN) formula \cite{hikami1980spin}. At 6 K, the phase coherent length $l_\phi$ = 41 nm, and it rapidly decreases with temperature. For 2D transport, the temperature dependence is expected to follow \cite{altshuler1982effects} $l_\phi(T)$ $\propto$ $T^{-0.5}$; in Fig. 2f the temperature dependence can be fitted with $l_\phi(T) \propto T^{-0.49}$, suggesting that the WAL occurs in the 2D surface states.
The above discussion provides evidence that doping depletes the trivial bands and exposes the spin-polarized topologically non-trivial surface bands which have been already measured by spin-resolved ARPES \cite{sakano2015topologically, iwaya2017full}, thus unveiling its topological nature in the normal phase. We are now well-positioned to explore the effects of topological state on superconductivity, especially, on Cooper pairing in the bulk and its interplay with the band structure to produce non-trivial topology in the superconducting phase.
\section{\label{sec:level1} SUPERCONDUCTING TRANSPORT PROPERTIES}
\subsubsection{Point-Contact Spectroscopy}
To investigate the surface of the superconductor, we performed 'soft' point-contact spectroscopy \cite{ daghero2010probing} (See Supplementary Information S3 Ref. \cite{Supplemental} ) on K-doped $\beta$-PdBi$_2$ cooled down to 300 mK, studying the magnetic field and temperature dependence of the differential conductance, d$I$/d$V$. We present the magnetic field dependence of d$I$/d$V$ with current along the $ab$ plane and the magnetic field along the $c$ axis in Fig. 3a. z is representative of the barrier strength: z = 0, Andreev spectroscopy; while z = $\infty$ (z $\sim$ 5 in experiments), tunneling spectroscopy. Here is z = 0.4. We note that the spectrum at zero magnetic field looks remarkably different from the rest. In particular, in Fig. 3b (where we have normalized d$I$/d$V$ to 1 as V $\rightarrow$ $\infty$), we see conductance dips at $\pm$ 1 meV and peaks exceeding the value predicted by the Blonder-Tinkham-Klapwijk (BTK) formalism of Andreev reflection \cite{blonder1982transition} for a conventional SC-insulator-normal metal interface (with z = 0.4). In fact, the peaks at 0.3 K exceed the theoretical maximum value of 2 (required by Andreev process) for a gapped superconductor, and might thus be indicative of gapless states.
Let us recall the known causes of conductance dip in PCS d$I$/d$V$ spectra. (i) \textit{Critical current or heating effect}\textemdash dips at positions larger than the superconducting energy gap are often found in the spectrum when the contacts on the sample are in the thermal limit \cite{sheet2004role}. At the critical current the superconductor turns into a normal metal, and when measurements are carried out in the thermal limit, the resistance of the bulk sample is measured in the d$I$/d$V$ spectrum. Since the critical current required to limit superconductivity reduces with increasing magnetic field and temperature, these dips are found to occur at position of decreasing lower bias voltage. In our experiments, the dip position does not reduce with increase in temperature and magnetic field (in fact it was only observed at zero magnetic field), so the critical current effect is ruled out (See Supplementary Information S4 Ref. \cite{Supplemental}). (ii) \textit{ 1D, 2D and 3D TSC} \textemdash topological superconductors feature dips at $\pm \Delta$. A simple physical explanation for this is the transfer of spectra weight from the states near the gap to make up for the in-gap states. In addition to the dips, 1D and 2D TSCs feature zero-bias conductance peak (ZBCP) due to Andreev bound state (ABS), while TRI 3D TSCs do not \cite{ mourik2012signatures, kashiwaya2014tunneling, sasaki2011topological, yamakage2012theory, sato2017topological}.
For a finite potential barrier between the contact and an ideal 3D topological superconductor, dI/dV $\propto$ surface density of states, and differential conductance spectrum should produce a double peak structure \cite{sato2017topological, sasaki2011topological, yamakage2012theory} \textemdash that is, ZBCP is \textit{not} expected for fully gapped TRI 3D topological superconductors. It should be pointing out, however, that the tunneling conductance can feature a ZBCP due to various effects. In studies for superconducting 3D TI, remnant Dirac fermions from the normal state are found to modify in the superconducting state in two ways: one, it enhances the pair potential resulting in a larger gap for the surface superconductivity \cite{mizushima2014dirac}. Two, if the Dirac surface state is well separated from the bulk, it can twist the surface Majorana cone, resulting in the ZBCP \cite{yamakage2012theory}. Furthermore, if the bulk superconductivity is not fully-gapped as is the case for Cu$_x$Bi$_2$Se$_3$, for example, the tunneling conductance features a ZBCP \cite{sasaki2011topological}. Otherwise in the ideal case, the differential conductance features a double peak. Here, the presence of dI/dV$_{norm}$ $>$ 2 (for z = 0.4) and the non-trivial conductance dips at zero magnetic field are in good agreement with the theoretical prediction for 3D TRI TSCs.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.7\textwidth,height=12cm]{fig3x}
\caption{ \textbf{Point-contact spectroscopy.} (a) Magnetic field dependence of the d$I$/d$V$ vs bias voltage for K doped $\beta$-PdBi$_{2}$ at 0.3 K. (b, c) BTK fitting of the d$I$/d$V$ spectrum at 0 T for 0.3 K and 1.1 K. The fit is poor. In comparison, the fit is good at 0.1T in (d). (e) Field evolution of the quasi-particle lifetime broadening parameter $\Gamma$. (f) close up view of d$I$/d$V$ vs bias voltage at 0.7 T and 0.8 T. (g) Attempts to fit the gap with the BCS magnetic field dependence equation.}
\end{figure*}
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth,height=8cm]{fig3}
\caption{\textbf{Upper-critical field analysis.} (a) and (b), Variation of the upper critical field $B_{c2}$ as a function of temperature in pristine and K doped $\beta$-PdBi$_{2}$. (c) the Ginzburg-Landau fit is shown in red and blue for pristine and K doped $\beta$-PdBi$_2$ respectively. (d) Plot of the reduced upper critical field, $b^* = B_{c2}/|dB_{c2}/dt|_{t = 1}$ as a function of the reduced temperature $t = T/T_c$. The red dash is the upper limit for s-wave superconductivity according to the WHH model. A conventional superconductor with finite SOC and Maki parameter will be below the universal WHH model curve. K doped $\beta$-PdBi$_2$ lies above the upper-limit of WHH model and closer to the polar p-wave model, pointing to K doped $\beta$-PdBi$_2$ as an odd-parity superconductor.}
\end{figure}
\begin{figure}[]
\begin{center}
\includegraphics[width=0.5\textwidth,height=12cm]{fig4}
\end{center}
\caption{\textbf{Spin-triplet Vortex State.} (a) $B_{c1}<B<B_{c2}$ is the vortex state: flux vortices starts to penetrate the superconductor at $B_{C1}$ and form a vortex lattice leads to irreversibility of the magnetization loop due to flux pinning effects. At the irreversibility field, the vortex lattice melts, restoring the reversibility. Doping increases the the number of pinning sites and consequently, irrersibility. However, a spin-triplet superconductor can demonstrate the opposite behavior, dubbed `type 1.5'. (b) Comparison of the normalized magnetization along the $B||c$ plane of both the undoped and doped sample, at 2.5 K and 1.8 K respectively. Beyond $B_{c1}$ of K doped $\beta$-PdBi$_{2}$ the magnetization shows an anomalous increase in magnitude which is at odds with the conventional type-II "alloy" superconductors but consistent with the expectation for a spin-triplet pairing. (c)-(d) magnetization for the pristine and doped sample respectively.}\label{lines}
\end{figure}
Looking beyond the presence or absence of the ZBCP as a precise evidence of non-trivial superconductivity, we study the effect of a time-reversal breaking perturbation, that is, magnetic field on the surface states. Analogous to the time-reversed Dirac fermions on the surface of a TI which are protected from backscattering, the superconducting state is expected host helical pairs of Majorana fermions which are robust against non-magnetic disturbances \cite{xu2014momentum, qi2009time}. This physics is captured here: applying a magnetic field breaks time-reversal symmetry\textemdash and protection from scattering is lifted. The helical surface states is destroyed and the underlying superconductivity described by the usual BTK-like spectrum is uncovered. We see in Fig. 3b-d that the spectrum under 0.1 T at 0.3 K fits the BTK model for conventional superconductivity \cite{blonder1982transition, dynes1978direct} while that under zero magnetic field does not. Note that 0.1 T $\ll B_{c2}\sim$ 0.8 T (as estimated below), so modifications to d$I$/d$V$ data due to vortices and due to suppression of the bulk gap are expected to be negligible.
We extracted the superconducting gap at different magnetic fields using the BTK formalism and attempted to fit its evolution with the prediction for a conventional Bardeen-Cooper-Schrieffer (BCS) superconductor: $\Delta(B) = \Delta_0 (1 - B/B_{c2} )^{1/2}$. Experimentally, $B_{c2}$ is found to be somewhere between 0.7 and 0.8 T according to the B field dependence of Andreev reflection as shown in Fig. 3f. However, the gap could not be described by BCS using 0.7 T $< B_{c2} <$ 0.8 T ; the misfit for 0.8 T is shown in 3g. By making both the $\Delta_0$ and $B_{c2}$ free parameters, we got the best fit with 0.63 T. Clearly, as shown in Fig. 3f the crystal is still superconducting up till at least 0.7 T. This proposes that the bulk superconducting state might not be entirely described by BCS theory. Odd-parity, unconventional bulk superconductivity is a necessary condition for 3D topological superconductivity -- in which the just demonstrated helical surface states are a signature. The next step, therefore, is to study the bulk superconductivity.
\subsubsection{Upper-Critical Field Limiting Effect}
To better understand the nature of the bulk superconducting state from which the surface state arises, we proceed to study the upper critical field limiting effect. In Fig. 4, the upper critical field $B_{c2}$ at different temperatures below $T_c$ is plotted and extrapolated to T = 0 by using the Ginzburg-Landau form, $B_{c2}(t) = B_{c2}(0)(1- t^2)/(1 + t^2)$, where $t = T/T_c$. $B_{c2}(0)$ is found to be 0.89 T for K doped crystal. This value is clearly higher than the value of 0.63 T obtained by fitting the d$I$/d$V$ data to the BTK formula for a conventional BCS SC, confirming that the superconductivity is unconventional.
It is remarkable that doped $\beta$-PdBi$_2$ has lower $T_c$ but higher $B_{c2}(0)$, as this is suggestive of spin-triplet pairing. Indeed, the mean free path $l$ is greater than the coherent length $\xi$, and the WHH and Pauli limiting effects are absent(See supplementary Information S5 Ref. \cite{Supplemental}). We can gain more insight by comparing the $B_{c2}(T)$ data with the well known theoretical model for s-wave and polar p-wave \cite{werthamer1966temperature, scharnberg1980p}. Fig. 4 is the plot of the reduced upper critical field, $b^* = B_{c2}/|dB_{c2}/dt|_{t = 1}$ versus the reduced temperature $t = T/T_c$ compared to the theoretical models for s-wave and polar p-wave SCs. For the doped crystal, we note that the experimental data fits better to the p-wave than to the s-wave model. The pristine $\beta$-PdBi$_2$ in comparison lies below the upper limit of the WHH s-wave theoretical prediction. The s-wave WHH model presented in Fig. 4 is the upper limit, derived for $\alpha$ = $\lambda_{so}$ = 0, where $\alpha$ and $\lambda_{so}$ are the Maki parameter \cite{maki1966effect} and the spin-orbit strength, respectively; non-zero $\alpha$ and $\lambda_{so}$ moves the curves lower. With $\alpha$ = $0.53 |dB_{c2}/dT|_{T_c}$ = 0.11 and a finite $\lambda_{so}$ \cite{shein2013electronic} in $\beta$-PdBi$_2$, the pristine crystal can be well described by the WHH model. This is in contrast to the doped crystal where b$^*$ lies above the s-wave WHH upper limit. For further evidence of unconventional superconductivity, we carried out magnetization experiments to study the vortex state.
\subsubsection{Magnetization: Vortex State}
The magnetization for a spin-triplet superconductor, as demonstrated for Cu$_x$Bi$_2$Se$_3$ \cite{das2011spin}, exhibits the so-called Type 1.5 like behavior (see Fig. 5). Consider the effect of the magnetic field induced by the persistent vortex current on the spins of the spin-triplet pairs. The induced magnetic field polarizes the Cooper pairs and an additional spin magnetization arises. The total magnetic flux in the vortex now consists of the current and spin magnetization contributions \textemdash and is quantized. The quantization of magnetic flux causes current inversion in parts of the vortex, favoring their formation just above $B_{c1}$ by driving an attractive interaction between the vortices. This explains the anomalous increase of the magnetization past the $B_{c1}$ threshold and the low irreversibility often observed in the magnetization vs magnetic field loop of Cu$_x$Bi$_2$Se$_3$. We also observe the sharp increase in magnetization beyond $B_{c1}$ and subsequent low irreversibility in doped PdBi$_2$ (which is absent in the pristine sample) and propose that this effect can be explained by spin-triplet pairing.
In the Fig. 5, we compare the magnetization of the pristine crystal with the doped one. Below $B_{c1}$, labeled region I, diamagnetization occurs at the same rate for both samples; above $B_{c1}$ the effect of current inversion is observed in the doped crystal. As the vortices becomes attractive, flux vortices are forced in the doped sample and the magnetization increases sharply in region II and continues to do so till the beginning of region III. In the third region, the magnetization rate saturates because the repulsive force between vortices due to high density counterbalances the attractive force effect. Also the vortex-lattice state melts into vortex-liquid state at a faster rate; thus the low irreversibility in spite of the doping.
\section{\label{sec:level1} DISCUSSION}
Having studied various bulk and surface transport properties in the normal and superconducting phases, we address the question of \emph{why} the doped system behaves so differently from the undoped one.
\textit{Effects on surface states in the normal state}: Firstly, the nontrivial surface states around E$_f$ in the normal phase coexist with trivial surface and bulk bands \cite{sakano2015topologically}. While a direct probe of the band structure such as spin-ARPES can detect the topological surface states, it is difficult to probe them in a transport experiment. One route for reducing the bulk contributions is using thin films. Alternately, appropriate dopants can deplete the trivial bands, uncovering the topological ones, and we find that K doping indeed accomplishes this goal.
\textit{Effects on bulk superconductivity}: Secondly, we recall that sufficient conditions for topological superconductivity in a 3D TRI material are that the normal state Fermi surfaces enclose odd number of TRIM \emph{and} the fullu-gapped bulk superconductivity pairing be odd under inversion. In pristine $\beta$-PdBi$_2$ only the former condition met. Bulk superconductivity is s-wave \cite{2016single,herrera2015magnetic,che2016absence}, so topological superconductivity is not expected (see lower-left corner of Fig. 1); instead, one gets a Fu-Kane-like superconductor with surface Majorana zero modes in vortex cores \cite{fu2008superconducting, lv2017experimental}, which are distinct from 2D helical Majorana surface states. On doping with K, we find that the c-lattice parameter increases (see Supplementary Information S1 Ref. \cite{Supplemental}) and transport changes from electron to hole dominated, indicating that larger K$^+$ ions have replaced the smaller Pd and Bi ions. This would lead to local inversion symmetry breaking without breaking the centrosymmetry of the bulk crystal, which is enough to induce spin polarizations in the bulk of layered, centrosymmetric systems \cite{zhang2014hidden, riley2014direct}. It is now well known that in the presence of spin-orbit-coupling, electron-phonon interactions can give rise to even- as well as odd-parity pairing \cite{brydon2014odd, kozii2015odd, wang2016topological}. The s-wave, even-parity state invariably onsets at a higher $T_c$ \cite{brydon2014odd}, driving the odd-parity state to T = 0. Suppressing the s-wave channel, for example, by adding minute local spin-polarization could promote the odd-parity channel. In centrosymmetric Bi$_2$Se$_3$, superconductivity is induced by intercalating A = Cu, Nb, Sr into the non-superconducting parent compound. Recent spin, transport and thermodynamics experimental studies on A$_x$Bi$_2$Se$_3$ have shown the evidence of 2-fold pairing symmetry consistent with odd-parity, nematic superconductivity as opposed to the 6-fold symmetry of the hexagonal Bi$_2$Se$_3$ structure \cite{yonezawa2017thermodynamic, pan2016rotational, asaba2017rotational}.
We emphasize that the bulk superconductivity intrinsic to $\beta$-PdBi$_2$ opens the possibility of a topological phase transition in the superconducting state from a trivial to a topological superconductor, which prior to this work has not been realized because undoped Bi$_2$Se$_3$ does not host intrinsic superconductivity. In case the topological superconductor first transitions into a magnetic phase, the accompanying critical point will present a unique laboratory realization of emergent supersymmetry \cite{grover2014emergent}. Thus, $\beta$-PdBi$_2$ with intrinsic superconductivity and a centrosymmetric layered structure presents a unique and robust platform for exploring various aspects of 2D helical Majorana fermions, just as 3D topological insulators enabled the study of 2D helical electrons on their surface.
\section{\label{sec:level1} MATERIALS AND METHODS}
$\textbf{Crystal growth and characterization.} $ Single crystals of $\beta$-PdBi$_2$ were successfully synthesized by solid-state melt method. Stoichiometric amounts of Pd grains and Bi powder, mixed at a molar ratio of 1:2, were sealed in an evacuated quartz tube and ran through the thermal profile as follows: the mixture was heated up to 900$^o$C in 4 hours and left to melt and mix for 20 hours. It was then cooled down at 3$^o$C/ hour to allow for crystal formation and swiftly quenched in iced water at 480$^o$C. The quenching is best done above 450$^o$C to prevent the formation of the $\alpha$-phase. The potassium doped sample was synthesized by nominally replacing Bi with the dopant element. The composition was characterized by wavelength dispersive spectroscopy; chemical analysis puts the actual potassium content much lower than the starting composition. To confirm good crystallization, we carried out x-ray diffraction experiments-- all the peaks were index and no traces of the $\alpha$-phase was detected. Furthermore, the doped crystal was checked for single crystalline domain orientation by Laue back-reflection diffraction method. Our Laue analysis on the pattern of dots indicates the formation of single crystalline domain orientation.
$\textbf{Transport measurements.}$ The samples were first cut into rectangular shape, then cleaved into the desired thickness by gently peeling of the layered crystal layers by layers using scotch tape. The freshly cleaved surface which is mirror-like and shiny is then gold-plated at the contact points to achieve better ohmic contact. The four-probe technique was used for the longitudinal resistance, $R_{xx}$, and the Hall resistance, $R_{xy}$, was acquired by the standard method. The magneto-resistance and magnetization measurements were carried out using Quantum Design Inc.'s Physical Properties Measurement System (PPMS) and Magnetic Property Measurement System (MPMS) respectively. The point-contact spectroscopy was performed in an He-3 refrigerator.
\begin{acknowledgements}
The authors wish to acknowledge R. Forrest, K. Dahal, Y. Lyu, S. Huyan and U. Saparamadu for technical assistance. We thank W.-P. Su, C. Ting, A. Guloy, D. Stokes, Lei Hao, Hai Li and R. Du for helpful discussions. The work in Houston is supported in part by the State of Texas through the Texas Center for Superconductivity at University of Houston (AK and JHM); and the U.S. Air Force Office of Scientific Research Grant No. FA9550-15-1-0236, and the T.L.L. Temple Foundation, the John J. and Rebecca Moores Endowment (AK); and by the Division of Research, Department of Physics and the College of Natural Sciences and Mathematics at the University of Houston (PH). TL is supported by NSF grant number DMR-1508644.
\end{acknowledgements}
\putbib
\end{bibunit}
\onecolumngrid
\begin{center}
\textbf{\large Supplementary material for: Transport evidence for three-dimensional topological superconductivity in doped $\beta$-PdBi$_2$}\\[.2cm]
Ayo Kolapo,$^{1,*}$ Tingxin Li,$^{2}$, Pavan Hosur$^{1,*}$ and John H. Miller$^1$ \\[.1cm]
{\itshape ${}^1$Texas Center for Superconductivity and Department of Physics, University of Houston, 3201 Cullen Boulevard, Houston, Texas 77204, USA,\\ Department of Physics and Astronomy, Rice University, Houston, Texas 77251, USA\\}
${}^*$email: <EMAIL>, phosur@uh.edu\\
(Dated: \today)\\[1cm]
\end{center}
\setcounter{equation}{0}
\setcounter{figure}{0}
\setcounter{table}{0}
\setcounter{page}{1}
\renewcommand{\theequation}{S\arabic{equation}}
\renewcommand{\thefigure}{S\arabic{figure}}
\begin{bibunit}
Here we present the x-ray diffraction (XRD), Hall resistivity analysis, 'soft' point-contact spectroscopy experimental methods, and analysis of WHH orbital and Pauli limiting effect on potassium doped $\beta$-PdBi$_{2}$.
\section*{S1. X-ray Diffraction}
Single crystalline samples of pristine and 0.3 \% K doped $\beta$-PdBi$_2$ were characterized by XRD using the Rigaku DMAX IIIB diffractometer at room temperature with a Cu K ($\lambda$ $=$ 0.154056 nm) radiation source. The $hkl$ ($00l$) reflection patterns for undoped and K doped $\beta$-PdBi$_2$ are shown in FIG. S1. The inset shows the 2$\theta$ shift in the doped crystal revealing a 0.3 \% increase in the c lattice parameter.
\begin{figure}[!htbp]
\centering
\includegraphics[width=13cm]{xrd}
\caption{Single crystal x-ray diffraction for pristine (black line) and K-doped $\beta$-PdBi$_{2}$ crystal (red line)}
\end{figure}
\pagebreak
\section*{S2. Hall Resistivity: Two-Channel Analysis}
The Hall resistivity data show non-linear behavior below 20 K, while it display linear behavior from above 20 K up to room temperature. In topological materials, this non-linearity is due to parallel contribution from the surface and bulk states \cite{ando2013topological, qu2010quantum}.
Both the surface sheet and the bulk contributions to the Hall conductivity can be analyzed using \cite{ashcroft1976nd}:
\begin{equation}
\rho_{xy} = \frac{ ( R_s \rho^2_b + R_b \rho^2_s)B + R_sR_b (R_s + R_b) B^3}
{ (\rho_b + \rho_s)^2 + (R_s + R_b)^2 B^2 }
\end{equation}
where $\rho_s $ and $\rho_b$ is the surface sheet and bulk resistivity respectively; and $R_s$ and $R_b$ is the surface sheet and bulk Hall coefficient respectively. Here $R_b= \frac{1}{en_b}$, $R_s = \frac{t}{en_s}$, and $\rho_s = \frac {1}{\sigma_s} = \frac {t}{G_s} $.
Using the value the longitudinal resistivity at zero field as a constraint, that is, $ \rho_{xx}^{-1}$ = $\rho_{s} ^{-1} + \rho_{b} ^{-1}$, we fit the Hall resistivity curves as shown in Fig. S2. It can be fitted well with the 2-channel model.
From the fitting value of $R_s$, we estimate $n_s$ = $- 7.7 \times 10^{16}$ \ cm$^{-2}$ for the surface carrier density and $n_b$ = $ 3.4 \times 10^{22}$ \ cm$^{-3}$ for the bulk. From these derived fitting parameters of $\rho_{s}$ and $\rho_{b}$, the percentage of surface contribution is calculated as:
\[ \frac{G_s}{G_s + \sigma_b t } = 0.06\]
estimating that $\sim$ 6 \% of the total conductance is due to the surface states at 6 K.
The mobility for the surface electron is $\mu_s = \frac{R_s}{\rho_s}$ = 4280 \ cm$^2/Vs$ which is 134 times larger than the mobility of the bulk carriers $\mu_b = \frac{R_b}{\rho_b} =$ 32 \ cm$^2/Vs$. With this high mobility, the surface states dominate the transport at low temperatures.
\begin{figure}[b!]
\centering
\includegraphics[width=0.5\textwidth,,height=6.5cm]{Hall}
\caption{Symmetrized Hall resistivity of K doped $\beta$-PdBi$_{2}$ at 6 K showing the non-linear behavior characteristic of two channel systems. }
\end{figure}
\pagebreak
\section*{S3. Point-contact Spectroscopy}
The point-contact spectroscopy set up is shown in Fig. S3. Here, ballistic transport is achieved through several channels of nanometer sized silver particles contained in the silver epoxy \cite{daghero2010probing, sasaki2011topological}. This method is in contrast to the hard metallic etched tip used in Scanning tunneling spectroscopy (STS), hence, the 'soft' appellation. The d$I$/d$V$ measurements were obtained by superimposing a constant 5 $\mu$A AC current with a sweeping DC current. The corresponding AC voltage can be picked up with great accuracy by the lock-in amplifier. The experiments were then carried out in a He-3 refrigerator.
\begin{figure}[h]
\begin{center}
\includegraphics[width=8cm]{2}
\end{center}
\caption{'soft' point-contact spectroscopy set up. The current is past through the the thin Au wire to the sample through a 30 $\mu$m tiny drop of Ag nano-particle epoxy paint.}
\end{figure}
\vspace{20mm}
\section*{S4. Differential Conductance}
Here we comment and highlight on the feature in the d$I$/d$V$ spectrum presented in the main text.
The BTK formalism for the Andreev reflection can be written as:
\begin{equation}
\frac{dI}{dV}_{Norm} \propto \int ^{\infty} _{-\infty} dE \frac{df(E+eV}{d(eV)} [1 + A(E) - B(E)]
\end{equation}
where $ { \frac{dI}{dV} }_{Norm}$ = $\frac{{\frac{dI}{dV}}_{NS}}{{\frac{dI}{dV}}_N} $, and $A(E)$ is the probability amplitude for Andreev reflection, and $B(E)$ is the amplitude for specular reflection. At $T$ = 0 and in the limit z $\rightarrow$ 0, then for a particle with $E < \Delta$ undergoing a complete Andreev reflection, we have $\frac{dI}{dV}$ = 2, because B(E) = 0. When $T$ and $B(E)$ is finite, $\frac{dI}{dV} <$ 2.
In our experiments, the spectroscopy is consistent with the BTK formalism when the magnetic field is turned (that is, when the magnetic field localizes the helical surface states). In the absence of the magnetic, we see the normalized differential conductance suddenly rising to $>$ 2. We also see that the conductance peak around $\pm \Delta$ is not due to critical current. In this paper, we have shown that K-doped $\beta$-PdBi$_{2}$ satisfies the necessary condition for bulk topological superconductivity, therefore an helical surface surface is guaranteed to exist (whether its detected in transport measurements or not). Then, it is reasonable to suggest that the anomalous zero magnetic field Andreev spectrum here is due to helical in-gap states, which in this situation are expected to be 2D surface Majorana fluids (distinct from the Majorana zero mode that has been detected in heterostructures).
\begin{figure}[]
\begin{center}
\includegraphics[width=16cm]{S4}
\end{center}
\caption{(a) the simulation for the following parameter, similar to actual experimental values: superconducting gap, $\Delta$ = 0.7 meV, barrier strength, z = 0.4, temperature, T = 0.3 K, and quasiparticle interference parameter $\Gamma$ = 0.05 meV (b) d$I$/d$V$ spectrum at 300 mK (c) 3D plot of the d$I$/d$V$ spectrum depicting the dramatic increase of d$I$/d$V$ at zero magnetic field.}
\end{figure}
\clearpage
\section*{S5. Absence of WHH orbital and Pauli Limiting Effect}
Backscattering by impurities, even non-magnetic ones, suppresses odd-parity superconductivity as Anderson theorem does not hold \cite{fay1980coexistence, foulkes1977p}. Thus, we first check if the mean free path is greater than the coherent length. As shown in Fig. 4 in the main text, the upper critical field $B_{c2}$ values at different temperatures below $T_c$ is plotted and extrapolated to T = 0 by using the Ginzburg Landau (G-L) theory. $B_{c2}(0)$ is 0.69 T for $\beta$-PdBi$_2$ and a higher value of 0.89 T for K-doped crystal even with lower $T_c$. Using $B_{c2} = \Phi_0/2\pi\xi_{c2}^2$, where $\xi_{c2}$ is the Ginzburg-Landau coherence length and $\Phi_0$ is the flux quantum, we obtain $\xi$ = 19 nm for K-doped $\beta$-PdBi$_2$. Assuming a spherical Fermi surface, we have wavenumber $k_F = (3\pi^2n)^{1/3}$ and with n = 4.81 x 10$^{27}$ m$^{-3}$ derived from the linear part of $\rho_{xy}$ for K doped $\beta$-PdBi$_2$, we find $l$ = 75 nm. This $l>>\xi$ combines contribution from both the surface and bulk state. If we use the bulk density, n = 3.4 x 10$^{28}$ m$^{-3}$, derived from S2, we have $l$ = 22 nm, which is still greater than $\xi$ = 19 nm. The doped crystal is sufficiently pure for odd-parity superconductivity.
Under the BCS theory, superconductivity can be limited by orbital and spin effect of magnetic field. The orbital depairing effect is described by the WHH theory while the spin limiting effect is described by the Pauli paramagnetism formalism by equating the paramagnetic polarization energy to the SC condensation energy $\chi_n B_p^2 = N[0]\Delta^2$, where N[0] is the density of state, $\Delta$ is the SC gap, and from which the polarization field, $B_{c2}^{p}(0)$ = 1.86 T$_c$, is obtained. Under WHH theory in the clean limit, $B_{c2}^{orb}(0) = 0.72 T_c|dB_{c2}/dT|_{T_c}$ = 0.75 T for the doped sample. This is below the experimental $B_{c2}$ value, suggesting that superconductivity is not orbital-limited. The spin limiting effect is described by the Pauli paramagnetism, $B_{c2}^{p}(0)$ = 1.86 T$_c$ = 8.184 T, which is way above the experimental B$_c$$_2$. So we have the relation: $B_{c2}^{orb}(0) < B_{c2}(0) << B_{c2}^{p}(0)$, a relation which is also observed in Cu$_x$Bi$_2$Se$_3$ \cite{bay2012superconductivity}. Now, when both the orbital and Pauli limiting effects are present, then $B_{c2} =B_{c2}^{orb}(0)/ \sqrt[]{1 + \alpha^2} $ = 0.74 T. $\alpha$ here is the Maki parameter \cite{maki1966effect}; $\alpha$= $\sqrt[]{2}B_{c2}^{orb}(0)/ B_{c2}^{p}(0)$ = 0.13. The expected theoretical $B_{c2}$ in the presence of both the orbital and spin limiting effects is lower than the experimental value 0.89 T. We can possibly conclude that the Pauli limiting effect is also absent.
\putbib
\end{bibunit}
\end{document}
|
\section{Introduction}
Tuning of spin-orbit coupling (SOC) in graphene \cite{spintronics} is one of the fundamental steps to engineer graphene-based spintronics devices. One promising route to achieve this goal is the so-called spin-orbit proximity effect, recently extensively studied from both theoretical and experimental side.\cite{proximity, fluorine, weeks, tmds, njp, smallcells, copper, carbon, soc_prb, voloshina,krivenkov} This mechanism of inducing SOC extrinsically relies on the proximity between graphene (G) and a metal;
the SOC of the heavy atoms might be \textit{transferred} to the G when both materials are brought
sufficiently close to each other. Experimental realizations of spin-orbit proximity have revealed
several important phenomena, such as spin Hall effect at room temperature shown by Avsar
\textit{et al.}~\cite{proximity} or even a more intriguing electron confinement associated to
multiple topologically non-trivial gaps observed by Calleja \textit{et al.} in graphene
on Ir intercalated by Pb nanoislands (Pb/Ir).\cite{calleja}
Recently, we have reported that the mechanism of inducing SOC in G when adsorbed on heavy metal
surfaces is far more complex than it had been predicted before.\cite{soc_prb} DFT calculations of
graphene on Pt(111) and on Au/Ni(111) showed that the induced spin texture is a result of
spin-dependent hybridization between the Dirac cones (DCs) and the surface $d$-bands of the metal.
The spin vector of graphene is determined by that of the substrate bands, and undertakes
rotations wherever hybridization with any of the spin-orbit splitted metal bands occur.
Consequently, the reported non-trivial spin textures, although intriguing from the
fundamental point of view, seem difficult to control in any practical device.
Furthermore, although hybridizations locally open mini-gaps around which the SOC-derived spin
splitting may reach giant values above 100~meV, in the quasi-linear regions, where
the G transport properties are most relevant, the splittings are typically of the order of
just 10~meV.\cite{soc_prb, voloshina}
The main purpose of this study is to theoretically explore alternative routes to increase the
SOC derived splittings in the G by incorporating single metal adatoms at the graphene/metal
interface. We consider two types of adsorption which should lead to two very different
interaction scenarios:
(i) decoration defined as the adsorption of the adatom on top of graphene and,
(ii) intercalation of the adatom between the G and the metallic surface.
The first case should induce changes mainly in the G's properties already perturbed by the metal
surface, while the latter might
significantly alter the graphene-substrate proximity, as graphene will now interact with the
metal mainly via the intercalated adatom. Importantly, both decoration and intercalation can be realized experimentally\cite{decoration1, decoration2, decoration3, decoration4, intercalation1, intercalation2, intercalation3, calleja, misha, jorge, 2dmaterials, krivenkov} and are known to provide several interesting options for engineering of graphene's properties, in addition to any possible enhancement of SOC.\cite{krajl} Here, we will focus on two previously studied models, G/Pt(111) and G/Au/Ni(111) which present markedly different electronic and magnetic properties, and consider the adsorption of one species for each system, namely, a Pt adatom for G/Pt(111) and
an Au adatom for G/Au/Ni(111).
The paper is organized as follows. In Sec II we provide a brief description of DFT calculations. Section III reports the electronic properties and spin textures of the G/Pt(111) calculated defect-free case and both types of adsorption. In Sec. IV we present a similiar analysis for G/Au/Ni(111) structures. The conclusions are summarized in Sec. V.
\section{Methods}
Our large-scale DFT calculations have been performed with the {\sc SIESTA} code~\cite{siesta} as implemented within the {\sc GREEN} package.\cite{green,loit} The exchange-correlation (XC) potential has been treated using the generalized gradient approximation (GGA) in the Perdew, Burke, and Ernzerhof formalism.\cite{pbe} Dispersion forces were included via the
semi-empirical scheme of Ortmann and Bechstedt.\cite{vdw_ortmann} Spin-orbit coupling has been self-consistently taken into account as implemented in Ref. \onlinecite{soc}. Core electrons have been simulated employing norm-conserving pseudopotentials of the Troulliers-Martin type, including core corrections for the metal atoms.
The atomic orbital (AO) basis set based on double-zeta polarized strictly localized numerical
orbitals has been generated employing a confinement energy of 100~meV.
Real space three-center integrals have been computed over 3D-grids with a resolution equivalent to
500~Rydbergs mesh cut-off, while the Brillouin zone integrations have been performed over
$k$-supercells of around ($18\times18$) with respect to the G-($1\times1$) unit cell. The
temperature $kT$ in the Fermi-Dirac distribution has been set to 10 meV in all cases.
We have employed realistically large supercells to properly account for the the moir\'e patterns
and reconstructions known for these systems as well as to minimize the direct interaction
between the adatoms (Fig.\ref{geom}). In the case of the G/Pt(111) we
considered a thick Pt(111) slab (6 layers) with graphene adsorbed on top assuming a
G-($3\times3$)/Pt-$(\sqrt{7}\times\sqrt{7})$R$19.1^{\circ}$\ supercell which corresponds to a minimal lattice mismatch.\cite{martingago}
In order to reduce the interaction between defects among neighbouring supercells we have enlarged
the $(3\times3)$ supercell to a $(6\times6)$ and placed a Pt adatom either on top of the G
in an $atop$ configuration (Pt$_{\mathrm{ad}}$), or between the G and the Pt surface at an $fcc$ site and
below a C atom (Pt$_{\mathrm{in}}$). On the other hand, we modeled the G/Au/Ni(111) system assuming a
($9\times9$)/($8\times8$)/($9\times9$) commensurability between the G, Au and Ni lattices, respectively, with the Au layer intercalated between the G and the four Ni layers thick slab.
The Au adatoms have been incorporated either on top of the graphene at an $atop$ site (Au$_{\mathrm{ad}}$),
or in between the G and the Au layer below a C atom and at an $hcp$ site (Au$_{\mathrm{in}}$). The final adsorption structures have been
obtained after relaxing the graphene, the adatom, and the first two metal layers
until forces were smaller than 0.04 eV/\AA. In all calculations including SOC for the
G/Au/Ni(111) systems the spin quantization axis was set along the $z$ direction (out-of-plane).
Although estimates of the magnetic anisotropy employing the force theorem indicate that the
in-plane magnetization is more favourable, we have chosen the out-of-plane orientation in order
to preserve the $p3m$ symmetry and thus facilitate the interpretation of the spin textures. The
effect of choosing a different spin quantization axis will be briefly discussed in section IV.
Finally, the electronic structures have been evaluated in the form of projected density of states PDOS($\vec{k}, E$)\, calculated for the semi-infinite surfaces constructed after replacing the bottom layers of
the slabs by a semi-inifinte bulk following the Green's functions based prescription detailed in
Refs.~\onlinecite{ysi2,loit}. Unfortunately, unfolding the G-projected band structure into its
primitive BZ is not possible in the adatom configurations since the strong interaction induce
large distortions which break the translation symmetry within the G layer. Hence, all projections
are presented folded into the supercell's BZ.
\begin{figure*}
\includegraphics[width=\textwidth]{Fig1.eps}
\caption{\label{geom}
(a) Top and side view of the G/Pt(111). The ($3\times3$) supercell has been enlarged to ($6\times6$) to avoid interactions between adatoms in configurations (b) and (c). (b) Side view of G/Pt(111) with Pt adatom intercalated between graphene and the first Pt layer. (c) Same as (b), but with Pt$_{\mathrm{ad}}$ placed above (on top) of a C atom in graphene. (d) Relaxed geometry of the G/Au/Ni(111) structure. (e) Same as (b) for G/Au/Ni(111). (f) Same as (c) for G/Au/Ni(111). Graphene is represented either by red balls or sticks, while Pt, Au and Ni atoms by blue, yellow and grey balls, respectively. The black parallelograms in (a) and (d) mark the G($3\times3$)/Pt($\sqrt{7}\times\sqrt{7})R19.1^{\circ}$ and G($9\times9$)/Au($8\times8$)/Ni($9\times9$) supercells, respectively. All distances are given in angstroms.}
\end{figure*}
\section{G/Pt(111): intercalation vs decoration with Pt adatoms}
Figure~\ref{geom}(a-c) shows the relaxed geometries of all considered G/Pt(111) structures, that is;
the defect-free case in (a), the intercalated adatom
between G and the Pt(111) surface in (b) and the atop adatom adsorption in (c).
Figure~\ref{pt} presents all the corresponding
electronic structures and spin textures along the high-symmetry lines of the shrinked
$(6\times6)$ BZ.
Let us first briefly summarize the main results obtained for the defect-free configuration as
a detailed study for this case has already been presented in Ref.~\onlinecite{soc_prb}.
Given the weak interaction indicated by the large G-metal distance of 3.37~\AA\ (physisorption
regime\cite{fingerprints}) the DCs can still be clearly resolved in the PDOS map in Fig.~\ref{pt}(a), where the G
(red) and surface Pt (light blue) projections have been superimposed --recall that the G's $K$ and
$K'$ points are backfolded into the supercell's $\Gamma$ point.
In Figs.~\ref{pt}(b)-(c) we present the spin textures projected on the G and the Pt surface,
respectively, where we have simultaneously plotted the three spatial components of the spin
polarization employing a different color scheme for each of them: $s_\parallel$\ green, $s_\perp$\ red and
$s_{z}$ blue tones, where $s_\parallel$\ and $s_\perp$\ correspond to the in-plane spin components
projected along the $k$-line and perpendicular to it, respectively, and $s_z$ to the out-of-plane
component. Contrary to the PDOS case, the G-Pt interaction can be clearly appreciated
in these maps via the rich spin texture induced in the DCs by the hybridization with the $d$-bands,
involving multiple spin reorientations in all the occupied states region and up to around 0.8~eV
above the Fermi level (E$_f$). Furthermore, and as shown in Ref.~\onlinecite{soc_prb}, the splitting
of the G bands is by no means uniform, attaining {\it giant} values larger than 100~meV at mini-gaps, but only a few tens of meV at most in the quasi-linear regions.
\begin{figure*}
\includegraphics[width=\textwidth]{pt_defects.eps}
\caption{\label{pt}
Electronic and spin structure of G/Pt(111) intercalated/decorated with single Pt atoms. (a) Density of states of G/Pt(111) calculated in ($6\times6$) supercell and projected on graphene (red) and Pt (light-blue). (b) Corresponding spin texture projected on graphene. The color scheme is defined as follows: green/red shades refer to the direction of spin parallel/perpendicular to the momentum, while blue corresponds to the out-of-plane component; light/dark tones denotes positive/negative values of each component, see also the inset summarizing the legends in the bottom of panel (d). (c) Same as (b) projected on Pt substrate. (d) Brillouin zones of the ($6\times6$) supercell (small black hexagons), G-($1\times1$) primitive cell (red hexagon), and Pt-($1\times1$) primitive cell (blue hexagon). The selected $k$-lines are marked within the yellow hexagon in the center. (a'-c') Same as (a-c) for the configuration with intercalated Pt atom; its PDOS in (a') is colored in yellow, and its spin texture is displayed in (d'). (a''-d'') Same as (a'-d') for configuration with single Pt atoms adsorbed on top of G.
}
\end{figure*}
\subsection{Intercalation between graphene and Pt surface}
Intercalation of the Pt adatom (Pt$_{\mathrm{in}}$) between the G and the substrate induces a
strong buckling in the former with a corrugation as large as 0.8~\AA, with short bond lengths of
2.1~\AA\ between Pt$_{\mathrm{in}}$ and the closest carbon atoms. At the same time, the G layer is displaced
upwards so that the lowest C atoms lie 3.7~\AA\ above the Pt surface.
In such geometry, we expect a weakening of the overall interaction of the G with the Pt surface
at the expense of a stronger one with the intercalated defect.
In the PDOS map presented in
Fig.~\ref{pt}(a'), consisting of superimposed bands of G (red), Pt adatom (yellow) and the Pt
surface (light blue), the adatom contribution appears as a rather faint smudge (yellowish tones)
indicating, as expected, a strong hybridization with the continuum of Pt bulk states.
Close proximity of the C atoms with the Pt$_{\mathrm{in}}$ leads to important changes in the
DCs with respect to the defect-free case; one of the cones vanishes almost entirely below E$_f$
while the other remains well-preserved but strongly broadened in the whole considered region.
The G's spin structure, shown in panel (b'), also reveals strong differences with respect to the
defect-free case. As can be inferred from the substrate's and adatom's spin textures shown in
(c') and (d'), it now follows more closely the spin of the latter. In fact, due
to the strong G-Pt$_{\mathrm{in}}$ interaction, the Dirac point can be clearly resolved in
panel (d') as well as the strong hybridization with one of the DCs.
Surprisingly, the Pt$_{\mathrm{in}}$ spin texture is markedly different from that at
the Pt surface, which closely resembles the defect-free case (panel (c)), implying that the
SOC at the surface is hardly affected by the presence of the adatom.
On the other hand, at energies above $\sim 1$~eV, there are hardly any Pt$_{\mathrm{in}}$ states
and the DCs appear at first sight very similar as in the defect free case, allowing a direct
comparison between their respective SOC induced splittings.
Figure ~\ref{peaks_pt}(b) presents single spectra corresponding to spin vector versus energy curves
$\vec{s}$(E)
extracted from panel (b') for two selected $k$-points in the empty states region (indicated by
white line segments). They are compared versus analogous data calculated for the defect-free model.
The spin-splittings are clearly larger by at least a factor of two in the case of the intercalated
model, although the PDOS (gray lines) is significantly broadened as a result of the strong
G-Pt$_{\mathrm{in}}$ interaction. Thus, Pt$_{\mathrm{in}}$
intercalation appears as a quite efficient way to enhance the spin-orbit proximity effect.
\begin{figure}
\includegraphics[width=\columnwidth]{peaks_pt.eps}
\caption{\label{peaks_pt}
PDOS(E) and $\vec{s}$(E) single spectra extracted from the maps in Fig.\ref{pt} (a)-(a') at two different k-points (left-hand and right-hand panels) marked with white lines in Fig.\ref{pt}. Panel (a) corresponds to the defect free case and (b) to the intercalated model. Only unoccupied DC branches are shown. The numbers shown in the plots refer to the values of spin-orbit derived spin-splitting of the bands corresponding to each peak in PDOS(E). Grey, red, green and blue lines represent the PDOS, and $s_\perp$, $s_\parallel$\, and $s_z$ components, respectively.
}
\end{figure}
\subsection{Pt adsorption on top of G/Pt(111)}
Contrary to the intercalation case, the adsorption of a Pt adatom on top of G/Pt(111) leads to
hardly any buckling of the G with a corrugation below 0.1~\AA\ (see Fig. \ref{geom}).
However, a very short distance between the adatom and the G (2.14~\AA) induces a strong
interaction and important changes in the G's electronic structure, as can be noticed
in Fig.~\ref{pt} (a'') where the PDOS($\vec{k}, E$)\, of G, Pt$_{\mathrm{ad}}$ and Pt(111) are superimposed
following the same color scheme as in (a'). The most striking feature is the bunch of intense
localized bands belonging to the adatom (yellow) which completely tear the lower DCs and notably
alter the upper ones.
Such picture is consistent with a simpler model where the Pt surface has
been removed. Indeed, the PDOS of a pure G+Pt$_{\mathrm{ad}}$ configuration, shown in
Fig.~\ref{s1}(a) in the Appendix A, strongly resembles the one in panel (a''), indicating
that the G-Pt$_{\mathrm{ad}}$ interaction overrules that with the Pt substrate as expected from
their close proximity.
An orbital analysis of the adatom's states reveals that below E$_f$ all of them are mainly
of 5$d$ character, while only the band at approximately +400 meV, which crosses the Dirac point,
has an $sp$ origin.
The same applies to the spin textures shown in Figs.~\ref{pt}(b''-d''). The G and Pt$_{\mathrm{ad}}$\
projections (panels (b'') and (d''), respectively) are highly reminiscent of their substrate-free
counterparts in Fig.~\ref{s1}(b) and (c). The quasi-atomic states at energies around $-0.1$, $-0.4$
and $-0.6$~eV can be clearly seen in the G-projected $(\vec k,E)$ map throughout the entire BZ due to their
strong hybridization; they present a spin splitting of $\sim 200$~meV and as they tear the DCs,
the $\pi$-bands are endowed with similar splittings. At each anti-crossing region their
magnetization aligns with that of the Pt$_{\mathrm{ad}}$\ state (of intrinsic character) and mantain
this orientation (mainly out-of-plane, $\pm s_z$) until the next anti-crossing. The Pt surface,
nevertheless, still influences the G's spin texture, specially at energies where the Pt$_{\mathrm{ad}}$\ bands
are absent: below $-1$~eV and above $+1$~eV, where the in-plane $s_\perp$ and $s_\parallel$ spin
components become patent.
\subsection{Dirac point analysis}
\begin{figure}[ht!]
\includegraphics[width=0.98\columnwidth]{pt_dc.eps}
\caption{\label{pt_dc}
G-projected electronic and spin structure around the $\Gamma$ point for the G/Pt(111) systems:
(a) defect-free, (b) intercalated
Pt$_{\mathrm{in}}$\ defect and (c) decorated Pt$_{\mathrm{ad}}$\ defect. First and second columns show the PDOS
obtained without and with SOC, respectively, while right-hand columns correspond to the in-plane
spin components, $s_\parallel$ and $s_\perp$.
}
\end{figure}
In Figure~\ref{pt_dc} we present high resolution graphene projected PDOS and $s_\parallel$ and
$s_\perp$ ($\vec k,E$) maps around the Dirac point (DP) for the three configurations considered; the $s_z$
component has been omitted since it is significanly less intense than the in-plane ones in
all cases. Additionally, and in order to visualize the role played by the SOC, in the
leftmost column we present the graphene's PDOS calculated under the scalar-relativistic
approximation. For the defect free case, panel (a), and in the absence of SOC we obtain
sharp linear $\pi$-bands and a gapless DC
consistent with the quasi-free standing character of the G. When the spin-orbit coupling
is turned on, the intrinsic SOC opens a small gap (below 10~meV) which, however, is hindered
by the broadening of the $\pi$-bands due to their hybridization with the Pt substrate. Hence, no quantum spin Hall phase is expected. On the other hand, Rashba SOC is patent in the
$s_{\parallel/\perp}$ maps with splittings of the order of 10~meV (30~meV) in the upper (lower) cones.
Furthermore, the spin texture is far from helical, having a larger $s_\parallel$ component
than $s_\perp$.
The quasi-free standing picture changes drastically for the two defected
configurations. In the intercalated case, panel (b), sublattice symmetry is broken since
the Pt adatom resides below a C atom (sublattice A), opening a large gap ($\approx 130$~meV)
between its associated DCs, while the other DP (sublattice B) remains gapless, although the
bands loose their linear behaviour. Furthermore, the G's PDOS intensity is significanly
smaller than in the defect-free case due to the reduced C-Pt distance.
The main effect of the SOC here is an
increase in the gap for DP-A and of the Rashba splitting of all cones. This is particularly clear in the
lower DC-A, where the splittings attain values close to 40~meV.
When the Pt adatom is adsorbed on top of a
C atom, panel (c), the sublattice symmetry is again broken and a gap larger than 150~meV opens at the DP-A. On the other hand, the lower DC associated to the sublattice B is destroyed
due to the presence of the Pt$_{\mathrm{ad}}$\ $sp$ atomic level at around 0.4~eV. Apart from the Rashba
splitting of the lower DC-A (larger than 40~meV) and, to a less extent, of the upper DC-B, SOC
induces a splitting of the adatom's $sp$ state, so that one component remains flat
(at around 0.34~eV) while the other bends as it anti-crosses the DC.
\section{Adsorption of single Au adatoms in G/Au/Ni(111)}
\begin{figure*}
\includegraphics[width=\textwidth]{au_defects.eps}
\caption{\label{au}
Electronic and spin structure of G/Au/Ni(111) intercalated/decorated with single Au atoms. (a) Band structure of G/Au/Ni(111) along $\Gamma-K-M$ in folded $(9\times9)$ BZ represented as PDOS($\vec{k}, E$) projected on graphene (red), gold (light blue) and Ni surface (dark blue) superimposed at one map. (b) Corresponding graphene's spin texture after superimposing the $x/y/z$ components, each color coded as explained in Fig.\ref{pt}. (c) Same as (b), but projected on intercalated Au layer. (d) Brillouin zones of the ($9\times9$) supercell (small black hexagons), and G-($1\times1$) primitive cell (red hexagon); the considered $k$-lines are labeled within the yellow hexagon. (a'-c') Same as (a-c) for the configuration with additional Au atom intercalated below the G; yellow shades in (a') denote its PDOS, while panel (d') shows its spin texture. (a''-d'') Same as (a'-d') for the configuration of G/Au/Ni(111) with Au atoms adsorbed on top of the G. The spin textures projected on Ni(111) are neglected in all cases.
}
\end{figure*}
\begin{figure}
\includegraphics[width=\columnwidth]{peaks_au.eps}
\caption{\label{peaks_au}
PDOS(E) and $\vec{s}$(E) single spectra extracted from the maps in Fig.\ref{au} (a)-(a') at specific k-point marked with white lines. Panel (a) corresponds to the defect-free case and (b) to the model containing additional intercalated Au atom. Left-hand (right-hand) panel shows occupied (unoccupied) DC branches. The numbers shown in the plots refer to the values of spin-orbit derived spin-splitting of the bands corresponding to each peak in PDOS(E). Grey, red, green and blue lines represent the PDOS, and $s_\perp$, $s_\parallel$\, and $s_z$ components, respectively.
}
\end{figure}
The relaxed geometries for the G/Au/Ni(111) system are shown in Figures~\ref{geom}(d), (e) and (f)
for the defect-free case, the intercalated Au$_{\mathrm{in}}$\ adatom and the adatom Au$_{\mathrm{ad}}$
on top of the G,
respectively. In the former, the weak G-Au interaction~\cite{soc_prb,gold-prl,klusek,gold-dft,gold-domains,gold-rpa}
leaves an uncorrugated graphene layer lying 3.4~\AA\ above the metal surface.
Figures~\ref{au}(a-c) summarize its associated electronic and spin structure along the
high-symmetry lines of the supercell's BZ.\cite{soc_prb}.
Overall, the hybridization between graphene and the underlying Au/Ni(111) is weaker than in
G/Pt(111) case.
In the combined PDOS($\vec{k}, E$)\, map (a), the G (red), Au (light blue) and Ni surface
(dark blue) projections have been superimposed. The quasi-freestanding character of the G
manifests in almost undoped and well-preserved DCs down to binding energies of around $-1$~eV,
in agreement with previous experimental works.\cite{first13meV, main_nat}
The most intense Ni related features are located at approximately $-0.6$ and $+0.1$~eV,
corresponding to the top of the majority and minority $d$-bands, respectively. Several
gold $sp$ bands (the most prominent of them the Shockley-type surface state~\cite{soc,soc_exp}
(SS) emerging from $\Gamma$ at $-0.33$~eV) cross the BZ whereas fingerprints of the Au
5$d$-bands (light blue) appear below $-1$~eV distorting the DCs.
In spite of the fact that this system is magnetic and, hence, there exists an interplay between
SOC and exchange interactions, the G's spin texture shown in Fig.~\ref{au}(b) appears far
less complex than in the G/Pt(111) case. Indeed, in the [-1,+1]~eV range where the DCs appear
almost intact, their spin vector has only two components both perpendicular to the momentum~:\cite{soc_prb}
an in-plane helical component arising solely from the SOC, $s_\perp$, and an
out-of-plane one, $s_z$, mainly induced by the Ni magnetic order.
It is also noteworthy the different broadenings of the spin-splitted branches,
particularly around the DP at $\Gamma$ where the minority (dark blue) component is much broader than
the minority one (light blue) whereas along $K-M$ and at around $-0.9$~eV the opposite behavior holds.
The $\pi$-band splittings in this energy window are only of the order of 10 meV, in
agreement with previous experimental data\cite{first13meV} and several theoretical results\cite{main_nat,soc_prb, voloshina}.
However, the helical spin texture should not hold anymore when domains with
different in-plane magnetizations are present at the Ni surface, as expected in real
samples\cite{main_nat}. In such case, one may still expect that the values of the splittings will
remain small since their magnitude is mainly related to the magnetic coupling between the G and the
Au/Ni(111) surface rather than to the SOC.
The spin projected on the intercalated Au layer (panel (c)), on the other hand, mainly reflects the
hybridization with the Ni(111) spin-polarized bands again displaying light (majority) and dark
blue (minority) regions.
SOC manisfests most notably in the lower energy region (below $-1.0$~eV), where large in-plane
components (red) can be clearly seen at several energies.
\subsection{Intercalation with single gold atoms}
We again explored the role of adatoms either adsorbed above the graphene or intercalated between
the graphene and the Au monolayer. The relaxed structures, shown in Figs.~\ref{geom}(e) and (f),
follow analogous trends as in the G/Pt system. The intercalated adatom
induces a significant buckling in the G (0.8~\AA) while its average distance to the top Au layer
is significantly increased from 3.4~\AA\ to 3.9~\AA.
The associated PDOS and spin ($\vec{k}, E$)\ maps are presented in the middle
panels in Fig.~\ref{au}.
As shown in (a') where the additional Au$_{\mathrm{in}}$\ projection is colored in yellow, and
in contrast to the G/Pt(111) case, the intercalated adatom
introduces only subtle changes in the band structure
(e.g. removal of the Au's SS) leaving the graphene's DCs hardly affected. The contribution
of the highly delocalized Au$_{\mathrm{in}}$\ $sp$ states covers most of the map as can be seen by the change
in the blue tones compared to the defect-free configuration in panel (a), while intense
$d$-states appear below $-1$~eV showing little dispersion. We also note that Au$_{\mathrm{in}}$\ shows
no significant spin-polarization (below 0.01~$\mu$B) when intercalated.
The spin textures projected on the G and the gold surface layer, panels (b') and (c'), respectively,
are very similar to their defect-free counterparts ((b) and (c)), implying that the adatom has
little impact on them. The main difference is a reduction of the $\pi$-band
broadening due to the enlarged G-Au average distance. In Fig.~\ref{peaks_au} we compare G-projected DOS$(E)$ and $\vec s(E)$ curves between the
defect-free (a) and the intercalated (b) cases for both the lower and upper DCs
at a representative $k$-point (marked by the white segments in Figs.~\ref{au}a-a').
There are only very small changes (a few meV) in the splittings between both systems, with values
of $\sim 10$~meV in the upper cones and $\sim 20$~meV in the lower ones.
Therefore, intercalation of an Au adatom hardly enhances the SOC derived spin splitting in
the G/Au/Ni(111) system, in contrast to the G/Pt(111) case. We assign this difference to the
absence of Au$_{\mathrm{in}}$-$d$ states close to E$_f$.
\subsection{Decoration with single gold atoms}
When Au$_{ad}$ is adsorbed on top of the G, the latter remains hardly
corrugated (0.15 \AA), while the C-Au$_{\mathrm{ad}}$\ bond distance becomes very short (2.46 \AA).
Below $-1$~eV, the atomic-like Au$_{\mathrm{ad}}$\ $d$-states (intense yellow in (a'')) strongly hybridize
with the DCs opening multiple gaps. Moreover, the most relevant
feature is the pair of flat bands that run above and below $E_f$ and which strongly perturb and
tear the $\pi$-bands close to the DP.
As can be clearly seen in the G and Au$_{\mathrm{ad}}$\ spin projections of panels (b'') and
(d''), each band holds opposite spins with only $s_z$ component. An orbital
analysis reveals that they correspond to the 6$s$ state of Au$_{\mathrm{ad}}$\ which is
exchange splitted by $\sim 0.4$~eV and, in analogy with an Au isolated atom,
is responsible for the adatom's spin polarization (the total Au$_{\mathrm{ad}}$'s magnetic
moment is 0.56~$\mu$B).
The analogous calculation for the simpler G+Au$_{\mathrm{ad}}$ model
(that is, after removing the Au/Ni(111) surface), shown in Figure~\ref{s2}, yields a
very similar band and spin structure, indicating that the spin polarization of
the Au$_{\mathrm{ad}}$\ atom is unrelated to that of the substrate. This is further
corroborated by the fact that the spin texture projected on the Au layer (c'')
is almost identical to that of the defect-free case (c).
\subsection{Dirac point analysis}
\begin{figure}[ht!]
\includegraphics[width=0.99\columnwidth]{au_dc.eps}
\caption{\label{au_dc}
G-projected electronic and spin structure around the $\Gamma$ point for the G/Au/Ni(111) systems:
(a) defect-free, (b) intercalated
Au$_{\mathrm{in}}$\ defect and (c) decorated Au$_{\mathrm{ad}}$\ defect. Left, center and right columns show the PDOS,
as well as the $s_\perp$ and $s_z$ spin components, respectively.
}
\end{figure}
High-resolution graphene projected PDOS and $s_\perp$ and $s_z$ ($\vec k,E$) maps are displayed
in Figure~\ref{au_dc} for the three G/Au/Ni(111) configurations (this time the $s_\parallel$
component is negligible in all cases). The equivalent PDOS and $s_z$ maps calculated neglecting
SOC have been omitted since they are visually identical to those shown in the figure. Therefore,
as stated above, the role of SOC in this system is mainly to introduce an $s_\perp$ spin
component (helical spin texture). Siimilar to the G/Pt case, the defect-free configuration
presents quasi-perfect DCs while the possible presence of a small gap due to intrinsic SOC is
masked by the broadening of the $\pi$-bands. The broadening, in fact, is much larger for the
$-s_z$ bands (dark), as could be expected from the fact that the Au/Ni(111) PDOS
around the Fermi level is mainly occupied by the minority Ni bands.
As shown in panel (b), intercalation of Au$_{\mathrm{in}}$\ hardly alters the DP or its
spin texture due to the low Au PDOS around the $E_F$ (the adatom's $d$-states all lie
at binding energies below $-1$~eV). However, the situation is drastically different when
the G is decorated by the adatom (panel (c)). The interaction between the spin-splitted
Au$_{\mathrm{ad}}$\ $s$-levels induces a large gap in the DCs associated to sublattice A (the C atom below
Au$_{\mathrm{ad}}$) which are also spin-splitted (bright and dark parabolas in the $s_z$ map). In contrast,
the DCs of the sublattice B remain fairly linear except for a small gap at $\Gamma$.
\section{Summary and conclusions}
We have investigated the spin-orbit proximity effect in graphene on metallic substrates decorated
or intercalated by metallic adatoms focusing on two specific graphene/metal systems, non-magnetic
G/Pt(111) and magnetic G/Au/Ni(111) previously studied experimentally. \cite{platinum,platinum2,first13meV,main_nat,krivenkov}
Depending on the location of the adatom, two very different scenarios are reached; adsorption on top
leaves the graphene essentially uncorrugated but hybridizations with the atomic-like $d$-states
leads to densely teared $\pi$ bands resembling freestanding graphene decorated by adatoms. It turns
out that in the two systems considered the Pt$_{\mathrm{ad}}$\ and Au$_{\mathrm{ad}}$\ adatoms present states close to
E$_f$, thus the quasi-linear parts of the DCs close to the DP are largely distorted and the
electronic structure of G loses its linear character.
On the other hand, when intercalated between the graphene and the metal surface the former
becomes highly corrugated making short bonds with the adatom but with an average distance to the
surface larger (by $\sim 0.4$~\AA) than in the defect-free case. In this geometry, the
adatom's states strongly hybridize with the substrate's continuum of bands largely losing
their atomic-like character and therefore, their effect on the $\pi$-bands is less intense than
for adsorption on top. In G/Pt(111) the upper DCs remain almost unaltered exhibiting
a SOC-induced complex spin texture similar to the defect-free case. Interestingly, the close
proximity of the G to the Pt$_{\mathrm{in}}$\ leads to an
increase in the $\pi$-band splittings in the empty states region by up to a factor of three.
This is not the case, however, for G/Au/Ni(111) which presents similar splittings as in the
defect-free case since the Au$_{\mathrm{in}}$\ $d$-states lie at higher
binding energies and their impact on the upper DC is less significant.
A detailed analysis of the G's Dirac point shows
that the role of intrinsic SOC is minimal in all configurations, inducing gaps smaller than the
broadening of the $\pi$-bands; this is an expected result since in all the defected configurations
the G's sublattice symmetry is broken.\cite{weeks} Therefore, the proximity effect
in the systems under consideration relies mainly on the Rashba-type SOC transfer.
Finally, we recall that the G/Au/Ni(111) system is the most puzzling one, since two
very different spin-splittings for the $\pi$-bands have been reported: around 10~meV~\cite{first13meV} and
giant values close to 100~meV~\cite{main_nat}. A subsequent STM study~\cite{krivenkov}, including simplified
theoretical models, tentatively assigned the small splittings to a full gold monolayer, while the
giant values would correspond to sub-monolayer phases where small Au clusters or even individual atoms
lie intercalated between the Ni(111) surface and the G. Furthermore, the Ni surfacemost layer was shown
to be reconstructed presenting a misfit dislocation loop structure.~\cite{interface} All our models considered,
based on a full gold monolayer (plus an adatom), and even taking into account the reconstructed
Ni(111) surface~\cite{soc_prb}, always yield small splittings of the order of 10~meV which are
driven by the substrate's spin polarization.
Therefore, even if the giant splittings come from
gold sub-monolayer phases or a different type of Ni-Au surface alloying (which we have not considered), we
believe their magnitudes are determined by the magnetic coupling with the metal surface and not by the SOC.
\begin{acknowledgments}
J.S. acknowledges Polish Ministry of Science and Higher Education for the Mobility Plus Fellowship (Grant No. 910/MOB/2012/0). J.I.C acknowledges support from the Spanish Ministry of Economy and
Competitiveness under contract Nos. MAT2015-66888-C3-1R and RTI2018-097895-C41. Part of the
calculations have been done using the supercomputer facilities at the Barcelona Supercomputing Center under the activity
ID QCM-2015-2-0008.
\end{acknowledgments}
|
\section{Introduction}
Resistive cooling is a well-established technique for slowing down the motion and reducing the amplitude of oscillation of charged particles in Penning traps \cite{rs1,rs2,rs3}. It is based on the dissipation of particle kinetic energy into resistive elements that represent a heat bath at some low target temperature. To this end, a resistive circuit is set up to connect geometrically suited trap electrodes \cite{dfa}. The particle oscillation in the trap causes oscillations of the image charges induced in these electrodes (Shockley-Ramo theorem \cite{sr1,sr2}) and leads to an image current through the resistive circuit that is hence extenuated to the point where the confined charged particles are in thermal equilibrium with the circuit. Usually, a resonant (RLC) circuit is used, based on the high achievable resonance resistances which lead to efficient cooling \cite{cir1,cir2,cir3}. For highly charged ions, resistive cooling is particularly suited due to the scaling of the cooling rate with the square of the particle charge $q$ \cite{rs3}. Highly charged ions are inapt for laser cooling due to their lack of broad optical transitions \cite{nolas}, and have large charge exchange cross sections that make buffer gas cooling or electron cooling unfavourable. They can be sympathetically \cite{sym1,sym2} cooled to a crystalline state well below the ambience temperature by interaction with singly charged laser-cooled ions \cite{sym3} when storage of both species with sufficient proximity and laser access is possible.
In the present ARTEMIS experiment \cite{art0,art1}, cooling of the trapped ions is used to suppress Doppler effects in optical and microwave spectroscopy which will be used to determine the magnetic moments of the ionic nucleus and the bound electron in highly charged ions. The work presented has been performed in the framework of the HITRAP facility at GSI, Germany \cite{kluge,gfak}.
In the following, we will give a short account of our setup and experimental procedures, particularly for in-trap production of highly charged ions, their non-destructive detection and resistive cooling.
Note that presently, detection and cooling are both aspects of the same interaction of the confined ions with the resistive circuit.
We will briefly review the theory of resistive cooling and of mass spectrometry of ion ensembles in a Penning trap, discuss their thermalization behavior, and how they are described in terms of a non-neutral plasma. The results section then shows spectrometric data, allowing us to determine the energy, spectral width, and spectral position of all species present in the ion ensembles during resistive cooling. While at low ion number densities, purely exponential cooling behavior in agreement with naive theory is observed, dense ion ensembles show spectral discontinuities that agree with predictions of a transition to a fluid-like state. As we present measurements with ion numbers of the order of several thousands and above, these ensembles are distinct from both the well-known single-ion case and the small ensembles studied in \cite{rs3}, and show collective behavior, depending mainly on the rate of thermalization. We compare the cooling dynamics to the non-exponential behavior reported earlier \cite{rs3} to show that fast thermalization within the ensemble is critical for efficient cooling of ion ensembles, rather than achieving high effective resistances with high cooling rates, but limited to a narrow energy distribution. This points to a potential limitation for resistive cooling in traps with highly tuned potentials or narrow-band cooling circuits.
\section{Setup and Procedures}
The experimental setup has been described in detail in \cite{art0,art1}. Briefly, an arrangement of stacked cylindrical Penning traps is located in the homogeneous field of a superconducting magnet and is cooled to liquid helium temperature (Fig. \ref{one}). This arrangement consists of a creation trap with multiple potential wells for ion confinement, and of a trap optimized for optical and microwave spectroscopy of confined ions under large solid angles \cite{ito}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig1}
\caption{\small (Color online) Left: Stacked configuration of Penning traps for ion creation, selection, and cooling (lower part) and spectroscopy (upper part). Right: Overview of setup showing the cryogenic trap chamber in the field centre of the superconducting magnet.}
\label{one}
\end{center}
\end{figure}
In a Penning trap, the homogeneous magnetic field $B_0$ along the $z$-direction forces the charged particles on orbits, hence confining them radially, while the electrode voltages of the trap are chosen to create a well at $z=0$ in the axial direction, thereby confining the ions to a small volume around the trap center $\rho=z=0$. In the absence of ion-ion interactions, each individual ion performs an oscillatory motion consisting of three eigenmotions, two in the radial plane perpendicular to the magnetic field, and one axial oscillation parallel to the trap axis \cite{bro86}.
In a cloud of interacting ions, the overall motion is more complicated on account of ion-ion interactions. Nevertheless, the axial and radial oscillatory motions in arbitrary ion clouds can be detected non-destructively, which is used to perform charge-to-mass spectrometry of the stored particles. At present, we are concerned with the detection and cooling of the axial motion only.
In the creation trap, ensembles of highly charged ions are produced by impact ionization of gas by an electron beam from a field emission tip, in close similarity to the charge-breeding process in electron-beam ion sources \cite{art1,joseba}.
Figure \ref{two} shows the part of the creation trap relevant for ion detection and the cooling measurements: Three electrodes A, B, and C form a geometrically compensated trap \cite{gab84} with B acting as the ring electrode and A and C being the endcaps. Electrode A is connected to a resonant (RLC) circuit for detection of the axial ion motion. The trap voltage $U_0$ is symmetrically split between ring and endcap electrodes as indicated.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig2}
\caption{\small (Color online) Schematic of the part of the creation trap used for the cooling measurements (cut for presentation of the inner structure): electrode A is connected to a resonant (RLC) circuit used for both detection and cooling of the axial ion motion and is read out by a spectrum analyzer via two amplifiers. All elements within the dotted box are cooled to liquid-helium temperature.}
\label{two}
\end{center}
\end{figure}
Non-destructive spectrometry of the confined ions is performed by recording the signal induced in the axial resonator circuit used for cooling as a function of the trap voltage (Fig. \ref{three}). Presently, the detection system is resonant at a fixed frequency of $\omega_R=2\pi\times 741.23$\,kHz and has a quality factor of about $Q=900$, such that the spectral width of the circuit is $\delta \omega_R=\omega_R/Q \approx 2 \pi \times 820$\,Hz.
When the trap voltage $U_0$ is scanned, ion species with different charge-to-mass ratios $q/m$ subsequently come into resonance when their axial oscillation frequency $\omega_z(U_0)$ (see Eq. (\ref{z})) matches the resonance frequency $\omega_R$ of the detection system, i.e. when for their specific value of $q/m$ we have $\omega_z(U_0)=\omega_R$. Since the detection system records a spectral power density within its frequency
band for a certain scan time, each peak area is proportional to an energy. To be more
precise, it is proportional to the individual particle kinetic energy times the number of particles contributing to that peak. This means that at constant particle number, the area measures the individual particle energy.
In our experiment, the particle number can indeed be regarded as constant, since ion loss due to collision or charge exchange with neutral species has been observed to be negligible on the time scale of the cooling experiments. As expected from kinetic gas theory, the cryogenic temperature of about 4\,K ensures efficient cryo-pumping of residual gases. From a measurement of the ion signal as a function of time over the course of 48 hours, a charge state lifetime (half-life) for Ar$^{13+}$ of 22 days is extracted, which at a cross section for charge exchange with helium gas of $(3.25 \pm 0.25) \times 10^{-15}$\,cm$^2$ \cite{mann} indicates a residual gas pressure of about $3 \times 10^{-16}$\,hPa.
Figure \ref{three} shows a spectrum of ions produced from supplied argon gas which also contains tungsten ions sputtered from the field emission tip. The signals in-between could not be assigned to an ion species. The spectrum is cleaned from all ions except argon prior to the following measurements by application of the SWIFT technique \cite{li} to the axial motion, i.e. by simultaneous resonant ejection of all unwanted species from the trap. This selective excitation increases the axial ion kinetic energy of unwanted species until it exceeds the trap depth and the ions are axially lost from confinement. The procedure takes place on the milli-second time scale and is presently combined with a temporal lowering of the trap potential to ease the ejection. Depending on the details, the number of unwanted ions is lowered to undetectable levels, while spurious excitation reduces the number of wanted ions by up to about 50\,\%. The result of this procedure (upon further cooling) is a spectrum as shown in Fig. \ref{peaks}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig3}
\caption{\small Ion spectrum upon ion creation. It shows charge states 9+ to 16+ of the desired argon ions and a number of charge states of tungsten from the field emission tip.}
\label{three}
\end{center}
\end{figure}
The argon ions are confined in a specific well of the creation trap for the present experiments, see Fig. \ref{two}.
In full similarity to the cleaning procedure above, pure ion clouds containing only a single ion species even in a selected $q/m$-state can be produced by resonant ejection of unwanted ions from the trap. This is done by resonant dipole excitation of the axial oscillation again using the SWIFT technique \cite{li}.
\section{Theory of Resistive Cooling}
In an ideal cylindrical Penning trap, a single confined ion
obeys the axial equation of motion \cite{gho}
\begin{equation}
\label{eins}
\frac{\mbox{d}^2}{\mbox{d}t^2} z + \omega_z^2 z =0,
\end{equation}
where the axial oscillation frequency $\omega_z$ follows from the axial trapping potential
\begin{equation}
\label{harm}
V(z)=\frac{U_0C_2}{2d^2} z^2
\end{equation}
according to
\begin{equation}
\label{pott}
\omega_z^2 z = \frac{q}{m} \frac{\mbox{d}V}{\mbox{d} z},
\end{equation}
such that for the present geometry the frequency of axial oscillation $\omega_z$ is given by \cite{gho}
\begin{equation}
\label{z}
\omega_z=\sqrt{\frac{qU_0C_2}{md^2}} \;\;\; \mbox{with} \;\;\, d^2=\frac{z_0^2}{2}+\frac{\rho_0^2}{4}.
\end{equation}
Here,
$C_2U_0$ constitutes the depth of the potential well for axial confinement, $z_0$ and $\rho_0$ are the axial and radial dimensions of the trap leading to $d=6.73$\,mm, and $C_2=0.563$ is a coefficient that accounts for our cylindrical trap geometry \cite{bro86,gab89}.
Within this description, resistive cooling may be modelled as a friction force which depends on the axial ion velocity {d$z$/d$t$}. The equation of motion then reads \cite{gho}
\begin{equation}
\label{vier}
\frac{\mbox{d}^2}{\mbox{d}t^2} z - \gamma \frac{\mbox{d}}{\mbox{d}t}z +\omega_z^2 z =0,
\end{equation}
where $\gamma$ denotes the cooling rate.
It is given by \cite{book}
\begin{equation}
\label{cool}
\gamma=\frac{R}{D^2} \frac{q^2}{m},
\end{equation}
where $R$ is the resonance resistance of the circuit used for cooling and $D$ is the so-called effective electrode distance which will be detailed below. The evolution of the ion kinetic energy in the presence of resistive cooling is given by
an exponential decay of the kind \cite{gho}
\begin{equation}
\label{dec}
E=E_0 \exp\left(-\gamma t \right).
\end{equation}
Note, that the cooling rate $\gamma$ and the exponential shape of the cooling curve (energy as a function of time) according to equations (\ref{cool}) and (\ref{dec}) is predicted to be independent of time and of the absolute value of the energy $E_0$ at the beginning of the cooling.
Further, it is an important prediction of a detailed calculation that for a thermal ensemble of identical particles, the cooling rate of the total kinetic energy is expected to be identical to the single-particle case \cite{rs3,gho,book}.
The effective electrode distance $D$ in equation (\ref{cool}) contains the information about the location and the geometry of the electrodes connected by the circuit with respect to the center of the ion oscillation \cite{book}. For the present geometry (see figure \ref{two}) with electrode (A) used for pick-up of the signal while the trap center is in electrode (B), $D$ can be calculated by \cite{book}
\begin{equation}\label{effd}
D^{-1}(\rho,z)=\frac{\partial}{\partial z} \Xi(\rho,z).
\end{equation}
where $\Xi(\rho,z)$ reads
\begin{equation}\nonumber
\Xi(\rho,z)=\frac{-1}{\rho_{0}\pi}\int_{0 }^{\infty}\frac{\textrm{I}_0~(x\rho/\rho_0)}{\textrm{I}_0~(x)} G \mbox{d}x,
\end{equation}
in which the geometry function $G$ is given by
\begin{equation}\nonumber
G=\textrm{sinc}(x\frac{z-z_F}{\pi \rho_0})(z-z_F)- \textrm{sinc}(x\frac{z-z_N}{\pi \rho_0})(z-z_N)
\end{equation}
and $\rho_0$ is the inner radius of the cylindrical electrodes, $z_N$ and $z_F$ are the distances from the center of the trap to the nearest and farthest edge of the
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig4}
\caption{\small (Color online) Plot of the effective electrode distance $D$ according to Eq. (\ref{effd}) as a function of the axial position $z$ for different values of the radial coordinate $\rho$ for the geometry shown in Fig. \ref{two}.}
\label{dee}
\end{center}
\end{figure}
electrode used for cooling (against ground), and $\textrm{I}_0(x)$ is the modified Bessel function of first kind. In our case, we have geometry values of $\rho_0 = 8.70$\,mm, $z_N= 7.46$\,mm and $z_F = 21.57$\,mm. Figure \ref{dee} shows the values of $D$ resulting from equation (\ref{effd}) as a function of the axial position $z$ for different radial positions between the trap centre $\rho=0$ and the inner electrode surface $\rho=\rho_0$. Obviously, the value of $D$ depends on both the axial and radial position in the trap, such that the expected cooling rate $\gamma$ (and likewise the cooling time constant $\tau=1/\gamma$) also depends on the ion location. Hence, for a large and/or hot ion ensemble that occupies a significant volume around the trap centre, this effect needs to be taken into account and the cooling rate has to determined via integration over the full ion distribution.
\section{Ion Ensemble Properties}
\subsection{Ion Number and Space-Charge Effects}
\label{spacecharge}
The ion number density in the trap can be deduced from the observed space-charge shift of the axial oscillation frequency of a given ion species upon production and initial cooling.
The presence of the other charges effectively reduces the trap potential experienced by an ion and thus reduces its axial frequency to a value \cite{yu}
\begin{equation}
\omega_z'=\omega_z \sqrt{1-\frac{\omega_p^2}{3 \omega_z^2}},
\end{equation}
where $\omega_p^2=q^2n/(\epsilon_0m)$ is the plasma frequency that depends mainly on the ion charge $q$ and ion number density $n$.
In spectrometry, the space-charge shift is seen as a shift of the observed trap voltage $U'_0$ that brings a given ion species into resonance with the detection circuit relative to the single-ion value $U_0$. From this shift, the ion number density can be calculated as
\begin{equation}
n= \frac{3\epsilon_0m}{q^2} \omega_R^2 \left[1-\left( \frac{U'_0}{U_0} \right) \right].
\end{equation}
In the present measurements, $n$ is between a few $10^3$ cm$^{-3}$ and a few $10^6$ cm$^{-3}$, depending on the values of the creation parameters (trap depth, charge breeding duration, electron current and gas density).
Reproduction of a specific ion number density is difficult, a given value can be measured by the voltage shift to a factor of less than two.
At the present trap parameters, the density is hence up to 10\,\% of the electric trap limit (from the stability criterion $\omega_c^2>2\omega_z^2$ with $\omega_c=qB_0/m$ \cite{book}) and up to 1\,\% of the Brillouin limit \cite{plas,book} at the magnetic field strength of 7\,T. The observed corresponding shifts are non-negligible and care is required for a proper identification of the ion species in the spectrum.
\subsection{Spectral Properties as a Function of Energy}
\label{axtrans}
Imperfections of the confining electric or magnetic fields make the axial oscillation frequency of any given ion dependent on its kinetic energy.
The most relevant imperfections of the confining fields are deviations of the electric field from the quadrupolar case. These effects have been carefully discussed in \cite{bro86,gab89,sens}. In traps like the present one, the dominant
contributions to an energy-dependent shift of the axial frequency come from higher-order dependences of the axial trapping potential on the coordinates and are quantified by coefficients $C_4$ and $C_6$ as defined in \cite{bro86,gab89}. The relative shift of the axial oscillation frequency of any given ion with its axial energy $E_z$ is given by \cite{gab89}
\begin{equation}
\label{shift}
\frac{\Delta \omega_z}{\omega_z} = \frac{3}{2} \frac{C_4}{C_2^2} \frac{E_z}{qU_0} + \frac{15}{4} \frac{C_6}{C_2^3} \left( \frac{E_z}{qU_0} \right)^2 .
\end{equation}
Hence, for a thermal distribution of ion energies we expect a corresponding distribution of axial oscillation frequencies, shifting and broadening the signal of each ion species in the spectrum. For a thermalized ion cloud at temperature $T$, the distribution of axial and radial energies is Boltzmann-like, with the expectation value $\langle E_z \rangle = k_BT/2$ and a typical width of roughly $2k_BT$ \cite{rs3}.
Hence, we expect the average relative shift of the axial oscillation frequency distribution of each ion species (with respect to $T$=0) to be given by
\begin{equation}
\label{shift1}
\langle \frac{\Delta \omega_z}{\omega_z} \rangle = \frac{3}{4} \frac{C_4}{C_2^2} \frac{k_BT}{qU_0} + \frac{15}{16} \frac{C_6}{C_2^3} \left( \frac{k_BT}{qU_0} \right)^2 ,
\end{equation}
and the relative width of the axial oscillation frequency distribution of each ion species to be given by
\begin{equation}
\label{wid}
\frac{\delta \omega_z}{\omega_z} \approx \frac{3C_4}{C_2^2} \frac{k_BT}{qU_0} + \frac{15C_6}{C_2^3} \left( \frac{k_BT}{qU_0} \right)^2 .
\end{equation}
In principle, the ensemble temperature $T$ can be determined either from an observed signal shift or width when the coefficients $C_i$ are known, however the determination via the shift requires zero-point information and also can be obstructed by space-charge effects. Thus, a determination via the observed signal width is preferred. In the present experiment, we have $C_2 = 0.563$, $C_4 = 0.001$, and $C_6 = 0.05$. These coefficients have been calculated from the measured trap geometry according to the formalism in \cite{gab89}.
The corresponding dependence of the relative signal width on the ensemble temperature is shown in Fig. \ref{width} for a trap potential chosen such that the ion species are in resonance with the circuit at $\omega_R=2\pi\times 741.23$\,kHz. Note that in a thermal ensemble, the relative widths of the signals as a function of temperature are identical for all ion species, i.e. identical for all ion charge states present in a spectrum.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig5}
\caption{\small (Color online) Relative frequency width $\delta \omega_z/\omega_z$ of the axial ion oscillation as a function of the ion ensemble temperature $T$, according to Eq. (\ref{wid}).}
\label{width}
\end{center}
\end{figure}
\subsection{Collisional Thermalization}
\label{rad}
Assuming an ion cloud with an arbitrary initial energy distribution and in the absence of external forces, ion-ion interactions (Coulomb collisions) thermalize the ions, eventually leading to the same Boltzmann distribution of energies within each degree of freedom.
To quantify the time scale for this, we use the thermalization time constant ('Spitzer self-collision time'), estimated by \cite{plas}
\begin{equation}
\label{dav}
\tau_T \approx (4\pi\epsilon_0)^2\, \frac{3\sqrt{m}\, (k_BT)^{3/2}}{4 \sqrt{\pi}\, n\,q^4 \ln \Lambda},
\end{equation}
where $\ln \Lambda$ is the so-called 'Coulomb logarithm' given by \cite{plas}
\begin{equation}
\ln \Lambda = 23- \ln \left(\frac{2nq^6}{e^6T^3} \right)^{1/2},
\end{equation}
where the ion number density $n$ is given in units of cm$^{-3}$ and $T$ is given in units of eV.
Assuming argon ions of an average charge state of 11 (which is the average squared charge state of the ion distribution), even at high ion energies upon production of 10 eV and low densities of order $10^3$ cm$^{-3}$, we have $\ln \Lambda \approx 15$, and the thermalization time constant is below one second. This is much smaller than the scan time of 550\,s for recording of an ion spectrum such as Fig. \ref{three} and also smaller than the time of each voltage step applied during the scan. Additionally, during cooling, as the density increases, the thermalization time constant becomes successively smaller. From this, we deduce that in all present measurements, the ion ensembles are thermalized during cooling and spectrometry.
\subsection{Ion Ensemble as a Plasma}
An ensemble of ions confined in a Penning trap can be characterized in terms of a non-neutral plasma \cite{dub}. Depending on the strength of the ion-ion interaction via the Coulomb force in relation to the ion kinetic energy, the collective behavior of the ion ensemble can be gas-like, fluid-like, or crystal-like. The strength of the motional correlation is usually expressed by the so-called `plasma parameter' \cite{book}
\begin{equation}
\label{gam}
\Gamma=\frac{q^2}{4\pi\epsilon_0ak_BT},
\end{equation}
where $a^3=3/(4\pi n)$. We have a weakly correlated plasma (gas-like state, independent-ion picture) for $\Gamma \ll 1$, an intermediate correlation (fluid-like state) between $\Gamma \approx 2$ and $\Gamma \approx 170$, and strong correlation (ions crystals) for about $\Gamma > 170$ \cite{dub99}. Due to the factor $q^2$ in the numerator, highly charged ions have much higher degrees of correlation at the same given temperature when compared to singly charged ions. In turn, they reach fluid-like or solid-like states already at higher temperatures.
An ion fluid can be seen as a distinct state between the largely uncorrelated ion gas and the rigid ion crystal.
In an ion crystal, a geometric ordering results from the mutual repulsion of the ions within the common potential well of the trap, and the dominance of the Coulomb interaction energy between ions over their kinetic energy leads to a high degree of motional correlation, with each ion being bound to a fixed location within the geometric structure \cite{drew}.
Ions in a fluid perform larger excursions within a certain volume around these locations on account of their higher kinetic ion energy, and may thermally `hop' from one such volume to another one. Correspondingly, in an ion fluid, the amplitude of motion of each individual ion is comparable to the average ion-ion distance, whereas in an ion crystal, it is at least one order of magnitude smaller. Still, the degree of motional correlation is too high for independent-particle pictures to be applicable, as the Coulomb interaction energy is at least comparable to the ion kinetic energy. Possible scientific applications of such highly correlated ion systems have been discussed in \cite{fifteen}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig6}
\caption{\small (Color online) Plasma parameter $\Gamma$ as a function of the ion number density according to equation (\ref{gam}) for Ar$^{13+}$ at three different values of the ensemble temperature $T$. For comparison, the same is shown for Pb$^{81+}$ ions.}
\label{gamma}
\end{center}
\end{figure}
Figure \ref{gamma} shows the plasma parameter $\Gamma$ according to Eq. (\ref{gam}) for Ar$^{13+}$ at temperatures of 4.2\,K, 7.5\,K, and 14\,K as a function of the ion number density $n$. The dotted lines indicate $\Gamma=2$ and $\Gamma=170$ above which a fluid-like and crystal-like behavior of the ion ensemble is expected, respectively. Accordingly, for our conditions with ion number densities in the range of a few $10^3$ cm$^{-3}$ to a few $10^6$ cm$^{-3}$ and temperatures below 15\,K the regime of intermediate (fluid-like) correlations should be accessible.
For comparison, $\Gamma$ is shown also for Pb$^{81+}$ ions. Here, the same plasma parameter $\Gamma$ is reached already at higher temperatures $T$ and even the crystal-like state might be achieved solely by resistive cooling.
\section{Results}
\subsection{Cooling Effect on the Ion Energy}
Argon ions have been created, confined, and cooled in the creation trap segment of the stacked Penning trap setup shown in Fig. \ref{one}.
Cooling is achieved by scanning the trap voltage and hence subsequently bringing each part of the ion spectrum into resonance with the cooling circuit, detecting the spectrum at the same time, see also Fig. \ref{two}. Each individual scan ('cooling cycle') takes 550\,s and is repeated to monitor the temporal evolution of the spectrum. So each data point of the resulting cooling curve is the outcome of one such scan.
Figure \ref{peaks} shows a typical spectrum.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig7}
\caption{\small (Color online) Ion signal as a function of trap voltage. This is a typical spectrum obtained from a voltage scan (`cooling cycle').}
\label{peaks}
\end{center}
\end{figure}
The full width at half maximum of a peak is indicated by $\delta U_0$, see for example the peak of Ar$^{11+}$ in Fig.\,\ref{peaks}. The value of $\delta U_0$ yields the width $\delta \omega_z$ of the corresponding axial frequency distribution according to
\begin{equation}
\delta \omega_z=\left( \frac{q}{md^2} \right)^{1/2} U_0^{-1/2} \, \delta U_0.
\end{equation}
For the present measurements, we have chosen the timing conditions such that thermalization amongst all ion species is much faster than resistive cooling, so that cooling affects all ion species similarly at all times. The time constant for thermalization $\tau_T$ (see Eq. \ref{dav}) is of order seconds and below, which is much smaller than the time of 550\,s used for each individual voltage scan. This is done to investigate whether the non-exponential cooling behaviour of ensembles of highly charged ions reported earlier \cite{rs3} can be attributed to insufficient thermalization.
In order to bring the present measurements of cooling by subsequently putting all species in the ion spectrum in resonance with the fixed circuit in accordance to the model of cooling underlying Eq. (\ref{cool}) (which assumes that any ion is in resonance with the circuit all of the time), we introduce two correction factors: One accounts for the fraction of the resonator bandwidth with respect to the total frequency span of the spectrum.
The frequency span $\delta$ of the spectrum is roughly $2 \pi \times 330$\,kHz, of which the resonator covers $\delta \omega_R \approx 2 \pi \times 820$\,Hz at a time. Hence, the ratio $\delta \omega_R/\delta \approx 1/400$ accounts for how much of the complete spectrum is covered by the resonator at any time.
The second correction factor concerns the shape of the spectrum, in specific the fact that the spectrum consists of discrete resonance lines. Cooling is only effective within a resonance line but not between the lines.
To compare this to a situation in which all ions are in resonance all of the time, the summed widths of the resonance lines with respect to the full span of the spectrum has to be accounted for. When we use the sum $p$ of all peak widths and the total spectral span $\delta$, the ratio $p/\delta$ is a reasonable correction factor accounting for the spectral shape.
The equivalent duration $t_{eq}$ of a cooling cycle is then given by the product of the cycle time $t_{cyc}$ and the two correction factors,
\begin{equation}
t_{eq}= \frac{p}{\delta} \frac{\delta \omega_R}{\delta} t_{cyc} =\frac{p \delta \omega_R}{\delta^2} t_{cyc}.
\end{equation}
and the effective time $t$ after a certain number of cycles is just the sum of all $t_{eq}$ summed up to this point. From here on, the time $t$ used in the discussions is the effective time obtained with this procedure.
Figure \ref{cycle} shows
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig8}
\caption{\small (Color online) Ion kinetic energy as a function of time when the ensemble is resistively cooled and re-heated by white noise (indicated by red arrows) several times. The parts marked `A' and `B' are identical to the ones in Fig. \ref{cool1} (top).}
\label{cycle}
\end{center}
\end{figure}
the ion energy as a function of time when the ensemble is cooled and re-heated by white noise several times. In all four cycles which represent different initial energies and trapping times, the observed cooling is exponential and occurs at the same rate. Thus, 'resistive cooling has no memory', as known for the single-particle case, and also as expected from theory for large ensembles. From the production parameters and the space-charge shift, the ion number density is estimated to be around 3 x $10^3$ cm$^{-3}$ in these measurements \cite{sad}.
The two indicated parts `A' and `B' of the cooling curve from Fig. \ref{cycle} have been evaluated separately.
In Fig. \ref{cool1} (top), the total ion kinetic energy of measurements A and B is given as a function of the effective time $t$. A single exponential decay of the form
\begin{equation}
E(t)=E_\infty + E_0 e^{-t/\tau}
\end{equation}
fits the data well, yielding a cooling time constant of $\tau=1/\gamma=4.65(79)$\,s (measurement `A'). Here, $E_\infty$ is the energy offset, i.e. the energy at the end point of cooling. A second measurement starting at lower initial energy yields a time constant of $\tau=4.76(83)$\,s (measurement `B'), in agreement with the first one.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig9}
\caption{\small (Color online) Effects of resistive cooling as a function of time: on the total ion kinetic energy (top), on the individual thermal signal width of Ar$^{13+}$ (middle), and on the individual center frequency of Ar$^{13+}$.}
\label{cool1}
\end{center}
\end{figure}
A similar measurement has been performed with an ensemble of pure Ar$^{13+}$ ions. To that end, the same creation procedure as before has been used, and all unwanted ion species have been removed from the spectrum by selective excitation of their axial motion beyond the confining well of the trap. This has been achieved with the so-called SWIFT technique applied to axial dipole excitation. Fig. \ref{ar13} shows the ion energy as a function of time for Ar$^{13+}$ and the signal for Ar$^{12+}$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig10}
\caption{\small (Color online) Ion energy of a pure Ar$^{13+}$ ensemble as a function of time and the corresponding Ar$^{12}$ signal.}
\label{ar13}
\end{center}
\end{figure}
The Ar$^{13+}$ signal shows an exponential decay like observed in the cooling of mixed ion ensembles, and the cooling time constant is in fair agreement with the previous values. The fact that the signal of Ar$^{12+}$ remains at zero for the complete measurement time confirms the efficiency of the SWIFT process and corroborates the above-stated absence of charge exchange which would produce Ar$^{12+}$ from Ar$^{13+}$.
Let us compare the observed cooling times with expected values. The resonance circuit used for cooling and detection of the axial ion motion is resonant at a frequency of 2$\pi \times 741.23$\,kHz and has a Q-value of about 900. Its inductance is about 3\,mH, and correspondingly it represents a resonance resistance of
\begin{equation}
R=QL\omega_z \approx 12.5 \mbox{M$\Omega$}.
\end{equation}
Inserting this value into Eq. (\ref{cool}) and averaging the effective electrode distance $D$ over the relevant trap volume we have $\langle D^2 \rangle ^{1/2} \approx 40$\,mm, and using the average squared ion charge state $\langle q^2 \rangle$ for q=8 to q=15, we have an expected cooling time constant of $\tau =1/\gamma \approx 4$\,s, which is in agreement with the measured values, both for the mixed and pure ion ensembles.
The observed cooling behavior is different from the one reported in \cite{rs3} as it does not display two largely different cooling time constants with a plateau-like structure between the two domains of the cooling curve belonging to different cooling time constants. Such a feature can apparently only be expected when the axial and radial degrees of freedom are cooled with significantly different time constants, as is the case for the dilute ion cloud in the highly tuned trap used in \cite{rs3}, leading to a complicated time dependence of the observed ion energy \cite{rs3}. In the present work, thermalization amongst all degrees of freedom is made to be faster than resistive cooling at any time during the process, leading to the energies in all degrees of freedom being reduced at the same rate, and hence leading to a single-exponential cooling behaviour.
\subsection{Cooling Effect on Signal Width and Shift}
The middle and lower panels of Fig. \ref{cool1} show the observed signal peak width of Ar$^{13+}$ and the corresponding center frequency of that peak as a function of time, respectively. From our discussion in section \ref{axtrans} we expect the thermal width of each peak to decrease with decreasing energy according to Eq. (\ref{wid}), i.e. to decrease with ongoing cooling time. This is in fact what is observed. The two measured time constants of $\tau=8.34(95)$\,s (for measurement `A') and $\tau=6.71(89)$\,s for measurement `B' agree within their uncertainties. They differ from the time constants observed for the decrease of the ion kinetic energy, because of the higher-order contributions of the temperature in Eq. (\ref{wid}). The presence of higher-order contributions makes the effect of the temperature decrease during cooling non-proportional to the temperature itself, and leads to a different observed time constant.
The same is true for the shift of the center frequency as a function of time. Also here, the two measurements `A' and `B' agree with each other, but do not necessarily give the same time constant as the ion energy. The observed shift of the center frequency is of the order of several hundreds of hertz which is a relative value of about $10^{-3}$ and can be attributed to the thermal change of the distribution in the presence of the given field imperfections.
\subsection{Final Temperature and Plasma Parameter}
From the known coefficients $C_2$, $C_4$ and $C_6$ of the trap potential and the measured width of the signal peaks, the ensemble temperature can be derived by inversion of Eq. (\ref{wid}), see also Fig. \ref{width}. The final width of the Ar$^{13+}$ signal at the endpoint of cooling is $\delta \omega_z \approx 2 \pi \times 7.3$\,kHz (Fig. \ref{cool1}, middle), which corresponds to a relative width of $\delta \omega_z/\omega_z$ of just below 0.01. According to Eq. (\ref{wid}) and Fig. \ref{width}, this corresponds to a temperature of about 7.5\,K, just above the ambient temperature of 4.2\,K.
According to Fig. \ref{gamma}, resistive cooling to a temperature of 7.5\,K in combination with an ion number density of several 10$^4$ cm$^{-3}$ brings the ion ensembles into the regime of non-negligible correlation with values of $\Gamma$ above 2, where fluid-like behavior can be expected. With our upper limit of ion number densities of several 10$^6$ cm$^{-3}$, values of $\Gamma$ up to 10 are accessible for this temperature. In addition, at these ion number densities, temperatures as high as about 15\,K should still allow us to reach $\Gamma$=2. Note, that the final temperature of resistive cooling is the electronic noise temperature of the circuit, not the physical ambient temperature, as has experimentally been shown in \cite{dje}. Often, the noise temperature is slightly above the ambient temperature, see for example \cite{haff}. Presently, the lowest observed temperature, irrespective of the ion number density and species, is around 7.5\,K.
\subsection{Indication of a Phase Transition}
We have performed cooling measurements like the one shown in Fig. \ref{cool1} at a significantly higher ion number density of a few times $10^6$ cm$^{-3}$.
Note, that the values of the axial frequencies are different from the previous measurements (Fig. \ref{cool1}) due to different choices of the trapping voltage and the RLC-circuit resonance frequency.
Figure \ref{cool2} shows the effect of resistive cooling on the ion energy, the signal peak width, and the signal center frequency for two selected ion species from the spectrum.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{fig11}
\caption{\small (Color online) Measurement as the one shown in Fig. \ref{cool1}, but at much higher ion number density: Effects of resistive cooling on the total ion energy (top), the peak widths (middle), and the axial oscillation frequency (bottom) of Ar$^{11+}$ and Ar$^{12+}$.}
\label{cool2}
\end{center}
\end{figure}
The axial energy of the ions follows an exponential decay like in all cases discussed above with similar time constants.
The same is initially true for the peak width.
There is, however, a significant discontinuity in the peak width as well as the center frequency of the ion resonances. After that discontinuity, they both remain nearly constant at a low level. The initial behavior of the ion frequency is qualitatively different from the case shown in Fig. \ref{cool1} (bottom): Instead of an exponential dependence with slightly increasing frequency, the frequency here decreases non-exponentially by some kHz. This variation is on the percent scale and hence more pronounced than the thermal shift shown in Fig. \ref{cool1} by more than one order of magnitude.
The same behavior is also present for the remaining ions in the spectrum (Ar$^{8+}$ to Ar$^{14+}$) and is independent of the step size of the voltage scan, i.e. the discontinuity occurs at the same effective cooling time for all ion species and step sizes.
The fact that the discontinuity is observed for all ions at the same time during cooling corroborates the assumption of a well-thermalized mixed ion ensemble.
From the observed relative spectral width $\delta \omega_z/\omega_z \approx 0.035$ at the discontinuity (dotted line in Fig. \ref{cool2}), Eq. (\ref{wid}) and Fig. \ref{width} yield an ensemble temperature of about 14\,K.
Although the phase transition around 14\,K
collapses the line width to an apparently constant value and therefore it cannot be used to extract the plasma temperature anymore, we believe that the ensemble keeps cooling to a final temperature of around 7.5 K, similar
to the cases where the number density is not sufficient to observe the
phase transition.
While for the measurement in Fig. \ref{cool1} the plasma parameter according to Eq. (\ref{wid}) remains well below $\Gamma = 2$ all of the time, the estimated value at the present discontinuity is $\Gamma \approx 5$. Hence, during cooling, the present ensemble crosses the expected transition from gas-like to fluid-like behavior.
Based on this, the observed discontinuity can be attributed to the onset of a significant level of correlations of the axial ion motions, leading to a sudden narrowing of the distribution of axial oscillation frequencies (see Fig. \ref{cool2} (middle)) and a sudden downward shift of the axial oscillation frequency by about 25\,kHz out of about 1\,MHz (see Fig. \ref{cool2} (bottom)).
This can be explained by a higher ion number density in the axial direction leading to a corresponding space-charge shift towards lower axial frequencies. So while the axial ion energies and thus axial oscillation amplitudes of all ions decrease continuously, at the critical temperature the motional correlation changes the oscillation frequency distribution non-continuously.
Based on the observed discontinuities in the ion spectra and the quantitative determination of values of the plasma parameter $\Gamma \approx 5$, we infer that a fluid-like state of the confined ion ensemble has been reached by pure resistive cooling.
\section{Conclusion}
We have measured the resistive cooling behavior of large ensembles of highly charged argon ions in a Penning trap under conditions ensuring efficient thermalization of the ensemble during the cooling process. Under these conditions, the cooling of an ion ensemble is observed as a decrease of the ion kinetic energy that leads to a narrowing and shift of the ion signal peaks. We have shown that all of these temporal evolutions can be described by a simple exponential decay, in contrast to earlier observations made under conditions which lead to inefficient thermalization and consequently to the occurrence of more involved cooling dynamics.
We have also shown that the exponential nature of resistive cooling of large ion ensembles and their cooling time constant $\tau$ and final temperature are independent of the initial ion kinetic energy and of the number of previous cooling or heating cycles (`resistive cooling has no memory'), as expected from the case of a single ion. Our observations further corroborate the theoretical expectation that the cooling time constant and final temperature of a thermalized ensemble is identical to the single-ion case, which at present is shown to be true for both mixed and pure ion ensembles.
The final temperatures reached in the present experiments are slightly above the ambient temperature of liquid helium, for which plasma parameters $\Gamma$ up to about 10 are accessible, depending on the actual ion number density.
While for low ion number densities, the observed cooling behavior is continuous down to the final temperature, for ion ensembles of high density, we have observed a significantly discontinuous behavior of spectral features during cooling, indicating an evolution of the ion ensemble across the $\Gamma \approx 2$ phase transition. From this, we infer that the present ion ensembles reach a fluid-like state at temperatures of around 14\,K without the need to apply laser cooling.
As a consequence, candidate ions for the envisaged optical spectroscopy measurements \cite{art0} such as Pb$^{81+}$, Bi$^{82+}$, or U$^{91+}$ are expected to form ion crystals by mere resistive cooling, which facilitates highly precise optical and microwave spectroscopy.
|
\section{Introduction}
Today's industries in various domains are becoming more and more driven by software as innovator. They range from web applications powering our increasingly digitalized daily lives to deeply embedded systems driving complex and safety-critical cyber-physical systems (CPS) as in, for example, self-driving vehicles. Companies need to continuously rejuvenate their product portfolio for adopting new ideas to remain competitive. A recent idea that is permeating from its original application domain of financial use cases are blockchains, where researchers and companies try to apply key ideas behind them to other domains.
CPS are found in various domains like automotive, smart energy systems, and home automation. Many systems therein are strictly regulated as they provide safety-critical functionality to their users. Such domains are usually more conservative in adopting fundamentally new ideas as many hurdles need to be overcome to unlock an idea's full potential.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{NewScoping.pdf}
\caption{Key ideas behind blockchains might be applicable to safety-critical system as found in cyber-physical systems where liability is of importance.
}
\label{fig:IntersectionSet}
\end{figure}
Figure \ref{fig:IntersectionSet} provides an overview of topic areas addressed in our research, which are currently potentially in conflict in industry: Today's companies that are developing products containing safety-critical components typically deal with the intersection set ``safety-critical systems'' and ``liability'' (for example, ISO-26262 in the automotive domain or insurance regulations in buildings and homes) to document what principles and state-of-practice were followed during the development and how the system was designed to reduce the risk of exposing its users to potentially harmful consequences. Such application scenarios, their benefits as well as their potential conflicts are analyzed and discussed in this paper.
Our research is motivated by the aforementioned intersection areas, where key ideas behind blockchains might be applied to development and usage scenarios of safety-critical systems (intersection set ``A''). Furthermore, the design of blockchains might support certain areas for liability aspects (intersection set ``B''). Finally, key ideas behind blockchain are apparently of interest for the intersection of all three (labeled ``C''). Our goal with this paper is to explore the question: \textit{What research and engineering challenges discussed in literature could be addressed with key ideas behind blockchains?}
The main contributions of our work comprise the analysis of challenges for CPS in the three (a) domains automotive, (b) smart energy systems, and (c) home automation to identify benefits and challenges of adopting key ideas behind blockchains for deriving a research agenda focusing on CPS. While we limit our analysis to examples from the aforementioned areas, our research roadmap is not necessarily limited to them, but will contain topics that similarly affect other domains like aerospace. Furthermore, our study does not look into prototypical implementations of concrete blockchain technologies but discusses the benefits and advantages primarily from a conceptual perspective.
The rest of this article is structured as follows: Section \ref{sec:RelatedWork} presents and discusses related work. Section \ref{sec:Challenges} outlines three application domains and discusses specific challenges therein to motivate open research challenges to be tackled as part of a research agenda outlined in Section \ref{sec:ResearchAgenda}; the research agenda is provided in Section \ref{sec:Discussion} before the article is concluded in Section \ref{sec:Conclusion}.
\section{Related Work}
\label{sec:RelatedWork}
This section outlines the related work in blockchain technology, self-driving vehicles, smart energy systems, and home automation.
\subsection{Blockchain Technology}
Swan describes one of the uses of blockchain as ``Github for memory''~\cite[p.46]{7360255}. Blockchain technology is the secure decentralized computing ledger that enables Bitcoin and crypto-currencies. More profoundly, it is a next-generation global-scale decentralized infrastructure and mechanism for securely updating distributed computing nodes with ongoing consensus truth states~\cite[p.42]{7360255} -- recently also named as ``Internet-of-Values''. Blockchain 3.0 comprises applications beyond currency, finance, and markets -- particularly in the areas of government, health, science, literacy, culture, and art~\cite{swan2015blockchain}.
Also the ICSE community has acknowledged the research potential to enable blockchain application development and Porru et al.~pointed out that the transactionality of blockchain requires major research efforts in software engineering~\cite{PorruPMT17} to fully understand and adopt its potential.
Today's challenges for blockchains in general are their limited throughput, latency, size and bandwidth, security, inefficient resources usage, usability, and versioning, and hard forks resulting in multiple chains~\cite[p.82/83]{swan2015blockchain}. Furthermore, Caes et al.~\cite{caes2016robust} add scalability, compatibility, legal uncertainty, lack of industry standards, and cost efficiency.
\subsection{Self-Driving Vehicles}
The technology around self-driving vehicles has been demonstrated at large-scale for the first time during the 2004 \& 2005 DARPA Grand Challenges and with a focus on urban-like environments in the 2007 DARPA Urban Challenge (cf.~Rauskolb et al.~\cite{RBL+10}) ten years ago. Since then, technology start-ups and all major automotive original equipment manufacturers (OEMs) have started to make the technology more reliable and to prepare their product portfolio to provide autonomous driving for their customers\footnote{The California DMV currently lists more than 40 companies that have a permission to test their self-driving vehicles on public roads: \url{https://www.dmv.ca.gov/portal/dmv/detail/vehindustry/ol/auton\_veh\_tester}}. Recent examples from a US manufacturer of premium electric vehicles also demonstrate that newcomers in this domain started to adopt accelerated development and deployment at large scale even for safety-critical systems to better understand how the customers use their products in the field (cf.~Giaimo and Berger \cite{GB17}). Such over-the-air software updates for certified systems would require end-to-end documentation from a change request to its implementation and tests, up to its deployment and usage in the field.
However, following the ISO-26262 standard for vehicle functions from the area of autonomous driving bears many more challenges to overcome as mentioned by Martin et al.~\cite{MTBW17} and Spanfelner et al.~\cite{SRS+12}; for example, the rising area of machine-learning seems to be highly controversial when safety standards like ISO-26262 shall be followed. Salay et al.~\cite{SQC17} argue for instance that functionality, which is only partially specifiable like, for instance, algorithms for perceiving the surroundings, should be treated differently from fully specifiable ones. Similarly, V\"{o}st and Wagner \cite{VW17} argue that current literature about combining accelerated development (``continuous delivery'') with safety critical domains is scarce.
In our work, we explore opportunities and limitations from key principles behind blockchain to enable, for instance, accelerated development including continuous experimentation and thus, monitoring a product in the field and systematically modifying it after its initial deployment as the next step.
\subsection{Smart Energy Systems}
Current energy systems are transitioning towards smart grids that integrate an increasing list of energy sources, including wind, solar, water, gas, nuclear, and coal. \textit{Smart} means intelligent, neat, trim, stylish, or operating in automation; a grid is a network of electrical conductors that deliver electricity to certain points~\cite{tuballa2016review}. According to the Strategic Deployment Document for Europe's Electricity Networks of the Future, a Smart Grid is an electricity network that can intelligently integrate the actions of all users connected thereto; such actors include generators, consumers, and those that do both in order to efficiently deliver sustainable, economic, and secure electricity supplies~\cite{smartgrids2010smartgrids}.
Smart grids are an important component of a smart city~\cite{bulkeley} and potentially an application area for the sharing economy~\cite{kalathil}. The problem is that smart grids have not yet been proven in the context of the utility providers' desired specifications. ``Utilities are in critical need of a near-real-world environment, with real loads, distribution gear, and diverse consumption profiles, to develop, test, and validate their required smart grid solutions.''~\cite{farhangi2010path} Such an environment can be provided by a smart microgrid that would then be extended over time, to extensively test function bidding~\cite{kamyab2016demand} and to solve communication~\cite{tsampasis2016communication} as well as other challenges~\cite{tuballa2016review}.
\subsection{Home Automation}
Another example for cyber-physical systems relates to Home Automation. The European Union identified that buildings consume 40\% of the energy and emit 36\% of CO2, being the largest end-user sector followed by transport (32\%), industry (24\%), and agriculture (2\%). The SMARTer 2030 report~\cite{smarter2030} by the Global eSustainability Initiative identified that the installation of building management systems (``smart home systems'') by occupants could offer a major opportunity to reduce the global CO2 emissions of buildings by a ratio of 15-30\%.
Although the possibilities arising from smart home technologies are promising, the market is highly fractionated and no set of standards for cross-vendor integration has emerged yet~\cite{icomp2016}. Therefore, we can see several open source projects emerging to offer the freedom to integrate different vendor technologies in a single smart home system. Despite the fact that the approach can lead to a fully integrated system, continuous integration of new vendor-specific technologies
are giving open source developer teams a hard time to keep the systems consistent and reliable.
As insurers have already identified smart home technologies as a valid tool for preventing damages and insurance cases, they report the heterogeneity of available systems and the difficult integration of software versions as the major hampering reason for a solid business case~\cite{MITTechReview10-2016}.
Similar to the application of blockchain technology in smart homes for security and privacy~\cite{DBLP:journals/corr/DorriKJ16}, our research explores possibilities using blockchain technology to give house owners as well as insurers an on-demand overview of installed smart home components, their configuration, as well as their used software version. This would enable insurers, for example, to accept smart home systems when different vendors' products and their transactions would be transparently documented in an end-to-end manner.
\section{Challenges from Cyber-Physical Systems}
\label{sec:Challenges}
Cyber-physical systems denote systems where computation meets physical processes in various contexts ranging from classical control loops to, for example, monitor chemical processes, up to complex system-of-systems where several agents collaborate towards a joint goal, like, for instance harbor logistics. In the recent years, different research challenges have been suggested to be addressed from various perspectives.
Lee~\cite{Lee08} primarily highlights in his position paper limitations originating from the level of abstraction that is introduced with the computational view for CPS. In addition, he also points out that reliability and especially interfaces between CPS are critical to be designed and realized properly. In this regard, we see here a primary application domain for blockchain as transactions between different CPS or a CPS and its back-end in the cloud, for example, can be transparently documented in an end-to-end way.
Cheng et al.~\cite{CLG+09} present an extensive research roadmap primarily addressing self-adaptive systems. Several core issues that the authors identify for self-adaptive systems also concern CPS in general as both have to cope with uncertainty after deployment. The authors name, for instance, safe mechanisms to achieve adaptation at runtime. While they mainly look at self-adaptation, such type of modification at runtime can be found with continuous integration and continuous deployment as well as it is on purpose aiming at system modification. The authors suggest, for example, proof-carrying code (PCC) as an instrument to ensure that parts of a software unit adhere to certain safety rules. When using blockchains, for example, such safety rules could be stored thereon to document what types of rules were in effect at a certain point in time to cope with liability claims in case of system failures for example.
Furthermore, Cheng et al.~\cite{CLG+09} have identified challenges for self-adaptive systems, where we also see application areas for blockchains that hold for CPS in general. For example, the authors name traceability from requirements to implementation for documentation; here, we clearly see a benefit of using key ideas from the technology behind blockchain to capture all transactions that turn requirements into code to also address liability concerns again to cope with auditing.
Sanislav and Miclea~\cite{SM12} are concerned in their work with challenges and research areas for CPS in general. They look into the domains of aviation, automotive, energy, and healthcare. They have identified reliability and security as a prevailing topic, which is very important among all of them. Here, blockchain can serve as an instrument to address security when, for instance, interacting agents communicate with each other and need to log their interacting transactions.
The research agenda CPS by Giaimo and Berger \cite{GB15} also lists relevant challenges from CPS in general, where we foresee application areas for blockchains. The authors name, amongst others, safety and dependability of systems at runtime; as elaborated before, the auditing feature of blockchain could come into play here as well. Similarly, evolution of CPS can be supported by blockchain due to transparent documentation of transactions that model a system's evolution.
The possible application cases that we have described so far address, in general, documentation topics. Thereby, audits or certification of safety functionality can be tackled from an end-to-end perspective assuming an appropriate tool support. Today's rise of machine learning approaches, though, requires also a seamless documentation of the training and validation stages during the development, especially after a system's deployment to the field when more data and situations could be collected to modify a neural network (NN) retroactively: In cases where a system would expose unwanted side-effects or a system failure, the complete documentation of any transactions around a NN can address liability concerns.
\subsection{Selected Challenges from Self-Driving Vehicles}
\label{sec:selfdrivingvehicles}
Compared to the aforementioned challenges, the development of self-driving vehicular functionality is specifically influenced from the ISO-26262 standard concerning safety-critical functionality for road vehicles to, for instance, address liability concerns. Apparent application scenarios for blockchain would typically address end-to-end documentation of and about development artifacts. For example, changing one software unit on a self-driving vehicle could require a modification to the blockchain describing all software units present in the distributed system. Other software units could only accept messages sent from the updated software unit if these messages are signed with the most recent hash value corresponding to the head of the blockchain; otherwise, such messages could simply be ignored. Thereby, after-market modifications to a complex system-of-system could also be regulated.
The use of vehicle-to-infrastructure or vehicle-to-vehicle (V2X) communication could also benefit from using ideas from blockchain to improve the trustworthiness among communication parties: Using a V2X infrastructure could be combined with a blockchain that documents credibility of its communication participants, where several observers transparently document whether communicated intentions of a participating unit match with a participant's observed intentions; as soon as observed actions deviate from intended behavior of a participant, its credibility score would decrease in the blockchain-based documentation. The consequences of this unmotivated deviating behavior, for example, could be penalized from an Intelligent Transportation System (ITS) to reject any requests from accessing a shared resource like a fast-lane on highways or with an increased taxation.
\subsection{Selected Challenges from Smart Energy Systems}
There is a number of challenges in the smart energy systems domain, the two most important aspects being privacy~\cite{bekara2014security,rawat2015cyber} and security issues~\cite{crispim2014smart,xenias2015uk}.
The community needs to design strong data encryption schemes, and security solutions vary among domains~\cite{wang2013cyber}.
Becker et al.~\cite{becker2016requirements} advocate for the importance of requirements engineering systems that support the sustainability of human lifestyle on Earth. They point out that one of the challenges is to have the buy-in and trust from all involved stakeholders. A similar trust issue comes up when talking about emission trading in the form of smart contracts~\cite{alharby2017blockchain}. One potential solution could involve small-scale emission certificate trading using blockchain technology (similar to micro credits in Africa and India, cf.~\url{http://www.kiva.org}).
\subsection{Selected Challenges from Home Automation}
Smart Home systems are typically a system-of-systems and therefore consist of a multitude of different software artifacts. In the aforementioned case, where insurers are willing to support Smart Home installations for better home insurances, keeping track of the correct use and maintenance of the insured valuables (e.g., by smoke detectors, heating control systems, or similar), the used code in the different components, their configurations as well as their combination is part of liability concerns. In case of damage, the insurance company needs to be sure that smart home system itself has not caused or supported the damage, not been tampered with, that all measures previously agreed on have been taken to prevent the damage, and that the extent of the damage has been limited as good as possible. In order to ensure these aspects, several application scenarios for blockchains can be derived like keeping track of software-versions and configurations of the different artifacts. In the previous example an end-to-end documentation of the deployed software artifacts can be ensured by the blockchain technology as well as their individual configuration in the specific use case.
But beyond keeping track of installed and modified software versions, blockchains can also be used to monitor the transactions and messages exchanged between the different smart home components. This way, it can be tracked if users have been warned about critical states or have been informed to take specific actions in order to avoid severe damage. A challenge within this scenario is for sure the limited computational power of the different nodes within a smart home system but at least an integrating smart home server could perform more advanced blockchain operations.
Furthermore Dorri, Kanhere, and Jurdak reported about the use of blockchain to tackle privacy and security issues in the Internet of Things, especially in the domain of interconnected Smart Homes \cite{DBLP:journals/corr/DorriKJ16}.
In order to foster the uptake of Smart Home Technology, blockchains can be used to keep track of the energy consumption of each home to reliably document fair use of resources by the inhabitants. Utilizing this, energy providers could either offer cheaper energy prices for more ``green'' families or a game-based competition between households could be started in order to motivate a more careful use of resources (cf.~\cite{gnauk2012leveraging,grossberg2015gamified}). This could ultimately lead to an approach of emission trading on a micro-scale within a street, a city, or a region.
\section{Research Agenda for Applying blockchain Ideas with Safety-Critical Systems}
\label{sec:ResearchAgenda}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{ResearchAgenda.pdf}
\caption{Research agenda for development and usage scenarios for key ideas behind blockchain in Cyber-Physical Systems.}\label{fig:RA}
\end{figure*}
Based on the challenges explored in the previous section, we outline several research topics originating from the overlapping circles in Figure~\ref{fig:IntersectionSet} describing potential application areas for key principles behind the blockchain technology. We suggest potential application scenarios; limitations of the current blockchain technology that might prevent a larger adoption, though, are further discussed in Section~\ref{sec:Discussion}.
\ul{Using Blockchain -- General Case}
\textit{Replacing central agencies using the distributed and synchronized ledger idea.} One of the core ideas behind bitcoins is to make a regular bank account obsolete as the distributed blockchain behind bitcoins transparently contains all transactions between involved parties and a bank account balance is simply the summation of all transactions. The same principle can be used to replace existing centralized registrations like taxi license concessions or driver's licenses.
\textit{Integral and decentralized index to existing data sources to verify authenticity and data integrity.} As the key idea behind blockchain is its immutability, this aspect can be exploited to establish and verify data integrity even for complex data sets. For example, sensor-readings over time from a test-drive of a self-driving vehicle could be stored by computing a hash-value per message that in turn is stored on the blockchain as the capacity on the blockchain is limited. A recipient of the complete dataset can then simply verify the integrity of the data recording by computing an arbitrarily chosen sequence of messages and verify their hash-values on the blockchain. Thereby, recent threats to pure hash-values applied to the entire dataset at large could be addressed\footnote{cf.~\url{https://goo.gl/lRCdTZ}}.
\textit{Regulating the access to systems.} The technology behind blockchain can also be used to replace current ``man-in-the-middle''-marketplaces like ride-sharing providers\footnote{cf.~\url{https://goo.gl/WfePp6}}. The blockchain would act as instrument to document supply and demand in a decentralized and distributed way to build the basis for service contracts for intelligent mobility. This would enable a cross-manufacturer marketplace to access intelligent mobility solutions in urban environments.
\ul{Using Blockchain for Safety-Critical Systems}
\textit{Management of certified software assets.} The development of a safety-critical system requires as of today strict documentation, for example, following the ISO-26262 in the automotive domain. Using blockchains, complete and full traceability of software assets that are used to create a software artifact including license management can be achieved to address auditing needs.
\textit{Smart Contracts.} Recent blockchain-based technology like Ethereum embed a virtual machine that allows the execution of scripts. Such scripts could codify Smart Contracts as extension of the aforementioned topic, which can be verified at run-time (cf.~\cite{alharby2017blockchain}). Such Smart Contracts could be used, for example, for safety-critical software to verify certain properties of a blockchain-documented runtime environment before it is enabled to provide a contractually agreed level of service.
\textit{Trustworthy behavior.} Blockchains can be used to describe intended future behavior in a trustworthy way: An entity X can document its intended behavior on the blockchain and X's collaborators can verify if the actual observed behavior of X matches with the communicated behavior beforehand. Any deviation therefrom might result in a context-dependent penalty like exclusion of communication or rejected access to resources. This aspect, for example, can be used to enable a blockchain-based watchdog for safety-critical systems.
\ul{Using Blockchain to Address Liability Aspects}
\textit{Auditing.} Blockchains can be used to document requirements to implementation in a transparent, traceable, and non-falsifiable way. Thereby, the issues mentioned by V\"{o}st and Wagner~\cite{VW17} regarding how to handle artifacts around functional safety can be stored on a blockchain. In combination with the Smart Contract feature, automated auditing thereof is enabled to also pave the road towards true continuous integration even for safety-critical systems.
\textit{Certification.} Similarly, blockchains can be used to document a certified system structure by documenting all involved subsystems and software units. Thereby, unintended or uncertified system modifications are simply rejected and a system would simply refuse to get activated. Furthermore, due to the transparency in a distributed ledger, verification of a system with what has been certified before is enabled. Thereby, the remote removal of software not fulfilling a given certification after deployment would be transparent as hash-codes from individual software units would be document on an immutable blockchain and thus, any modification after the certification time point would be visible.
\textit{Trustworthy software evolution.} Today's software engineering is significantly powered by the version tracking software Git\footnote{cf. \url{https://www.openhub.net/repositories/compare}}. Compared to blockchain, both approaches implement Merkle trees. However, in contrast to blockchains, anyone can simply add a new hashed block to Git's Merkle tree. Blockchains when used for Bitcoins, for example, require a Proof-of-Work block to be added before any further transactions are considered verified and transparently documented across the distributed ledger. As such, a blockchain-based Merkle tree combined with Proof-of-Work blocks can help to realize auditable and transparent software evolution.
\ul{Using Blockchain for Safety-Critical Systems to Address Liability Aspects}
\textit{Online transaction documentation.} Blockchains can be used to document online transactions of a distributed system; exchanged messages (or a subset sufficient enough to recreate a certain system state) are stored on a blockchain. Transparently documenting (sub-)system interactions even, for example, for V2X scenarios can help to resolve liability questions in cases of system failure like a traffic accident.
\textit{Regulating access to a critical system resource.} As an extension of the aforementioned scenario, blockchains can be used to regulate and document the access to a critical system resource like V2X-supported traffic management at intersections. The consensus-principle behind blockchain could be exploited to prioritize who is scheduled to pass next through an intersection.
\textit{Certification for regulated (safety-critical) systems.} Publicly accessible and verifiable documentation of how safety-critical systems performed in standardized tests (e.g., vehicles' safety system in standardized New-Car-Assessment-Programs such as EuroNCAP) could be enabled through the blockchain technology.
\ul{Using Blockchains to Enable Continuous-X for Safety-Critical Systems to Address Liability Aspects}
\textit{Pre-announcing changes.} Similar to the aforementioned example of documented intended future behavior of interacting agents, blockchains can be used to prepare and document what is about to change as part of Continuous-X: Intended modifications to a system can be codified as Smart Contracts and verified before activation after deployment. Thereby, violating behavior can be spotted and documented in a transparent way withstanding auditing.
\textit{Online Documentation for Continuous Experimentation.} Using blockchains, an experimental algorithm that is part of a Continuous Experiment (like, for example, an A/B test) and running in inert mode in the background is monitored by storing its incoming and outgoing data onto a blockchain to document its transaction history. Thereby, a system integrator can provide a safety-regulated instrument for its suppliers to change software running on a supplier's components after initial system deployment.
\textit{Software evolution documentation.} Blockchain can be used to document in a transparent way when suppliers or third-party entities change components in a system-of-systems after deployment. Thereby, certification violations and root-cause analysis can be facilitated in case of unwanted system behavior, malfunction, or system failure.
\section{Discussion}
\label{sec:Discussion}
The aforementioned research agenda focuses primarily on some core key principles behind the blockchain idea, namely documentation of transactions between interacting peers, distributed and decentralized storage, and immutability. These properties already enable critical application areas such as certification, smart contracting, or data integrity. However, cyber-physical systems might in many use cases require further properties that are either not or only partially addressed by blockchains as of today. In the following, we are discussing several challenges that we foresee when aiming for applying key principles behind blockchain to the domain of embedded and cyber-physical systems in particular.
The most critical limitation for embedded and cyber-physical systems in use cases that aim for online (or on-board) usage is the \textit{limited realtime capability}; the extension of the blockchain powering bitcoins is based on blocks that have been found using an increasingly hard computational task as Proof-of-Work. Embedded systems have only very limited power to act as Proof-of-Work providers. In addition, the computational effort requires a substantial amount of energy that might not be available on Internet-of-Things (IoT) devices that operate on battery for example.
Furthermore, due to the decentralized nature of blockchain, \textit{trust among nodes} is established using consensus about the longest valid blockchain. Therefore, blocks or the entire blockchain needs to be replicated among participants, which poses issues when either only limited storage capacity is available on a cyber-physical system, or when the network bandwidth is not sufficient or not present. The challenge, hence, is how a node could provide its service when a live network connection is a pre-condition therefor.
Especially for complex systems such as self-driving vehicles as presented in Section \ref{sec:selfdrivingvehicles}, the \textit{amount of data to be processed} can easily reach several hundreds of megabytes per second. Thus, storing such data on a blockchain is challenging as, for example, the block size in use for bitcoins is limited to 1MB. While this could be modified to any arbitrarily chosen size, the replication feature as described before must be preserved. Therefore, small block sizes are rather favorable over large storage capacity. A potential approach, though, could be to simply store digests or hash values from a real data snippet instead of the raw data itself.
The \textit{block size limitation}, however, is also a limiting factor for the number of transactions that can be handled per block. Current estimations for Ethereum are, for example, around 20 transactions per second depending on the maximum available Ethereum's gas limit. This might seem enough for an application to realize a decentralized intersection management system to replace traffic lights, but this limit would be valid globally and hence, not useful in practice.
The hurdles to apply key ideas behind blockchains to cyber-physical systems as we foresee them originate mainly from the \textit{trade-off between decentralized design to achieve trust and consensus and number of participating nodes to benefit from the use of blockchains}. The more nodes need to replicate (parts of) a blockchain as well as use or contribute to a blockchain, the more important are solutions and strategies to increase realtime-capability and replication/storage efficiency.
\section{Conclusion}
\label{sec:Conclusion}
In this work, we analyzed challenges for safety-critical embedded and cyber-physical systems to address the research question, which of these challenges could be overcome when adopting key ideas behind blockchains. This technology's principles, such as transactionality, immutable evolution based on Merkle trees, and distributed and decentralized storage offer apparent applicability to the intersection set between safety-critical systems and liability concerns. We outlined various examples for possible application areas looking specifically at the borders within the intersection sets safety-critical systems, liability, and blockchains: For example, software evolution even initiated by sub-system suppliers could be made transparent and auditable using blockchains as documentation instrument; similarly, certification of safety-critical embedded systems could be made more easily accessible using blockchains to enable verification of a system's components; in combination with Smart Contracts, such system evolution and certification could be combined and even enforced at runtime. Furthermore, an accelerated development and deployment process (e.g., continuous integration and continuous deployment) for such systems might benefit from the transactionality of blockchains, as an accelerated development and release cycle requires a much more stringent way of documenting changes in an immutable and auditable way.
Despite potential application areas for key principles behind blockchains, current challenges that also present issues in the operation for Bitcoins need to be overcome. Especially for cyber-physical systems, real-time capability by finding better or complementary instruments for the proof-of-work idea is needed. Furthermore, we need a solution to address the storage limitations per block to avoid expensive and lengthy replication but yet ensure transparency and consensus.
|
\section{Introduction}
Let $G$ be a simple graph. A signed graph $G_{\sigma}$ is obtained from $G$ by a sign function $\sigma: E(G)\rightarrow \{+1,-1\}$. The adjacency matrix $A(G_{\sigma})$ is a $(0,\pm 1)$-matrix where
\[
A(G_{\sigma})_{ij}=\begin{cases}
\sigma(ij) & ij\in E(G),\\
0 & \text{othewise.}\\
\end{cases}
\]
For a positive integer $n$, we denote by $[n]$, the set $\{1,\ldots,n\}$. The Laplacian matrix of a simple graph $G$, denoted by $L_{G}$, is $D(G)-A(G)$, where $D(G)$ is a diagonal matrix of the degree sequence of $G$ and $A(G)$ is the adjacency matrix of $G$.
Similar to the $L_{G}$, the Laplacian matrix of a signed graph $G_{\sigma}$ is defined as $D(G)-A(G_{\sigma})$ and is denoted by $L_{G_{\sigma}}$. The following well-known conjecture states an upper bounds for the sum of the largest eigenvalues of $L_{G}$.
\begin{conjecture}[Brouwer's Conjecture]\cite{bh}
Let $k,n\in \mathbb{N}$ and $k\in[n]$. For any simple graph $G$ of order $n$, if $\mu_1(L_G)\geq\cdots\geq\mu_n(L_G)$ are the Laplacian eigenvalues of $G$, then we have $\sum_{i=1}^{k}\mu_i(L_G)\leq e(G)+\binom{k+1}{2}$.
\end{conjecture}
In the recent paper of Wang and Hou, "Wang D., Hou Y., On the sum of Laplacian eigenvalues of a signed graph, Linear Algebra and its Applications, 555(2018), 39-52", they conjectured an upper bound for the sum of the $k$ largest Laplacian eigenvalues of signed graphs.
\begin{conjecture}\label{whc}\cite{WH}
Let $k,n\in \mathbb{N}$ and $k\in[n]$. For any signed graph $G_{\sigma}$, if $\mu_1(L_{G_{\sigma}})\geq\cdots\geq\mu_n(L_{G_{\sigma}})$ are the Laplacian eigenvalues of $G_{\sigma}$, then we have $\sum_{i=1}^{k}\mu_i(L_{G_{\sigma}})\leq e(G)+\binom{k+1}{2}+1$.
\end{conjecture}
We considered Conjecture \ref{whc} for graphs of order $7$ and $8$ and by using computer-assisted computations to signing the complete graphs of order $7$ and $8$, we found some counterexamples for $k=4$ and $k=5$.
\begin{example}\label{74e}
For the graph $G_{\sigma}$ as shown below, Figure \ref{74}, we have
\[L_{G_{\sigma}}=
\begin{pmatrix}
6& 1& 1& -1& 1& -1& 1\\
1& 6& 1& 1& -1& 1& 1\\
1& 1& 6& 1& 1& -1& -1\\
-1& 1& 1& 6& 1& 1& -1\\
1& -1& 1& 1& 6& 1& 1\\
-1& 1& -1& 1& 1& 6& 1\\
1& 1& -1& -1& 1& 1& 6
\end{pmatrix} \,\, \text{and}\]
\[
(\mu_1(L_{G_{\sigma}}),\cdots,\mu_7(L_{G_{\sigma}}))\simeq (8.7015,8.2360,8.2360,7.,3.7639,3.7639,2.2984).\] So, $\sum_{i=1}^{4}\mu_i(L_{G_{\sigma}})\simeq 32.1735\nleq e(G)+\binom{5}{2}+1=32$.
\end{example}
\begin{figure}[H]\label{74}
\centering
\includegraphics[scale=0.4]{74.png}
\caption{The graph $G_{\sigma}$ in Example \ref{74e}.}
\end{figure}
\begin{example}\label{85e}
For the graph $H_{\sigma}$ as shown below, Figure \ref{85}, we have
\[L_{H_{\sigma}}=
\begin{pmatrix}
7& 1& 1& 1& -1& 1& 1& 1\\
1& 7& 1& 1& 1& -1& -1& 1\\
1& 1& 7& 1& 1& -1& 1& -1\\
1& 1& 1& 7& 1& 1& -1& -1\\
-1& 1& 1& 1& 7& 1& 1& 1\\
1& -1& -1& 1& 1& 7& 1& 1\\
1& -1& 1& -1& 1& 1& 7& 1\\
1& 1& -1 &-1& 1& 1& 1& 7
\end{pmatrix}\,\, \text{and} \]
\[
(\mu_1(L_{H_{\sigma}}),\cdots,\mu_8(L_{H_{\sigma}}))\simeq(10.6056,10.,8.,8.,8.,4.,4.,3.3944).\]
So, $\sum_{i=1}^{5}\mu_i(L_{H_{\sigma}})\simeq 44.6056\nleq e(H)+\binom{6}{2}+1=44$.
\end{example}
\begin{figure}[H]\label{85}
\centering
\includegraphics[scale=.4]{85.png}
\caption{The graph $H_{\sigma}$ in Example \ref{85e}.}
\end{figure}
\section*{Acknowledgments}
The author is indebted to Iran National Science Foundation (INSF) for supporting this research under grant number 95005902.
|
\section{Introduction}
\label{sec:introduction}
Wealth inequality is an important matter for economic theory and policy.
Current debates have been discussing recent rise in wealth inequality in connection
with recent development of active financial markets around the world
(\cite{Beck et al. 2007, Piketty 2013,Krugman 2013,Stiglitz 2012,Solow 2014}).
While some argued for the role of financial development in reducing income inequality and benefiting the poorer, others criticize the increasing concentration of financial capital and related income as the main source of raising inequality. Moreover, issues of wealth distribution were raised by the 99\% movement in US in the aftermath of the Global Financial Crisis of 2007-08, claiming that the increased financialisation of economy and society in recent decades has involved an increased appropriation of wealth by the richest 1\%
of the population at the detriment of the remaining 99\%,
implying a more unequal and allegedly unfair wealth distribution.
Existing literature on wealth distribution connects the origins of wealth inequality with a variety of drivers
(\cite{Bertola et al. 2006}, \cite{Snowdon and Vane 2005}).
Already at the beginning of the XX century, pointing to the concentration of wealth in economy and society,
the economist and sociologist V. Pareto suggested the so-called Pareto wealth distribution as an empirical regularity
(\cite{Pareto 1897}), while the economic statistician C. Gini developed ingenious statistical techniques to summarise
wealth inequality through the so-called Gini Index (\cite{Gini 1912}).
In this context, a relevant literature stream draws upon \cite{Champernowne 1953} and \cite{Rutherford 1955}
to develop an elegant formal modelling that explains wealth concentration and the Pareto wealth distribution
under conditions of financial market efficiency and the stochastic distribution of financial returns across individuals
that invest in that market (\cite{Levy 2005,Levy and Levy 2003}).
This modelling strategy considers financial investment as a multiplicative process
closely related to a Kesten process (\cite{Kesten 1973,Redner 1990}).
Existing literature and the current debate on wealth inequality foreshadow some connection between wealth concentration and the financial investment process through active financial markets. Our approach consists in disentangling the minimal components of this process in order to formalise its working and infer general results on its impact on economy and society. In particular, we identify two minimal institutions that surround individual investment strategies: compound return structure, and active financial markets.
On the one hand, compound return structure stands at the core of financial investment process.
It implies that individual investors keep reinvesting financial proceeds together with the previous capital stock
through time and circumstances.
On the other hand, active financial markets assure that this financial investment process
is played as a fair game, involving some market order over return-seeking individual strategies.
Under efficient financial markets, individual investors extract their realised returns from
the same
distribution, while individual results remain {independent through space and time},
preventing arbitrage opportunities
(\cite{Samuelson 1965,Samuelson 1973,Fama 1995,Biondi and Righi 2017}
providing further literature review).
Our mathematical modelling shall formalise this financial investment process in line with both minimal institutions, in order to study its relationship with relative wealth concentration across individuals in economy and society. In particular, we shall prove that, let alone without countering forces, financial accumulation implies ever-increasing relative wealth inequality through time. We define financial accumulation as the peculiar financial investment process in which financial proceeds are systematically reinvested together with the previous capital stock through time and circumstances.
The rest of the article is organised as follows. The next section summarises our modelling strategy.
Section \ref{sec:model-proof-under} and \autoref{sec:model-proof} provide mathematical proof of
the relationship between wealth concentration and the financial investment process respectively under compound
and simple return structures. Section~\ref{sec:limits-implications} develops some implications
of and perspectives on our results. Section \ref{sec:concluding-remarks} concludes.
\section{Modelling Strategy}
\label{sec:modelling-strategy}
Existing models of economic process generally combine a multiplicative factor called capital with an additive factor called labour, in order to explain economic growth and study the aggregate distribution of wealth
(\cite{Nirei and Souma 2007} providing further references).
Our modelling strategy focuses only on the capital factor,
in view to disentangle featuring financial investment mechanisms and their impact on economy and society.
Generally speaking, financial investment process under efficient capital markets implies that all and every individual investment strategies generate independent and independently distributed returns extracted from the same distribution for all investors and at every point of time, altough real financial markets factually show some correlation through time and cross-sectionally. Our modelling strategy relaxes this reference framework while featuring cumulative effects through space and time. Our results do not require financial returns to be independent and independently distributed over time and across agents. Only some decorrelation through time is required. Although some previous studies noticed the non-stationarity of wealth distribution driven by financial investment (\cite{Kalecki 1945, Biondi and Righi 2019}), its cumulative effects and implications were generally neglected.
Cumulative effects relate to the financial accumulation process that is embedded in compound return structure. Accordingly, every investor starts investing an initial capital value through time by realising one given return at each point of time. Individual wealth investment is then repeatedly submitted to a series of realised returns that singles out each and every individual investment portfolio. Compound return structure further assures that financial investment proceeds sum-up with the past-period capital stock to be reinvested in the next period, generating individual financial accumulation through time.
Our modelling strategy points to this cumulative effect by featuring its impact on both individual financial accumulation through time, and the relative wealth concentration across individuals at some point of time. Our minimalist modelling strategy enables disentangling key features of financial accumulation process that characterises a large set of real world mechanisms such as saving deposits and mutual investment funds. In this context, it must be stressed that compound return is generally taken as benchmarking reference for assessing financial investment performance.
\section{Model and proof under compound interest structure }
\label{sec:model-proof-under}
Consider a population of $n$ individual investors $i = 1, \dots, n$, holding financial wealth available
for investment $W_{i}(t)$ at some {time $t$. We consider here discrete time dynamics, so $t$ is integer valued.}
We formalise the financial investment process according to the familiar structure of compound return
where the wealth $W_{i}(t+1)$ of investor $i$ at time $t+1$ depends on the past wealth stock $W_{i}(t)$
invested at some return $r_{i,t+1}$ as follows:
\begin{equation}
W_{i}(t+1)= W_{i}(t) (1 + r_{i,t+1})
\label{eq:1}
\end{equation}
where $\{\mathbf r_t = (r_{1,t}, \dots, r_{n,t})\}_{t\in \mathbb N}$
is an infinite sequence of identically distributed random vectors in $\mathbb R^n$.
We do not require that the components of these random vectors are independent.
Without loss of generality, we may impose $r_{i,t}\ge -1$.
In fact, we do only impose conditions that are required for mathematical proof.
For sake of clarity, if $r_{i,t} \ge 0$,
the investor $i$ maintains or increases its past wealth stock that was invested at time $t-1$,
while that investor loses it at least partly if $r_{i,t} < 0$.
This setting generates the following cumulative process through time from initial time $t=1$ to time $t=T$:
\begin{equation}
W_{i}(T) = W_{i}(0) \prod_{t=1}^T \{ 1+ r_{i,t}\}= W_{i}(0)\exp \left\{ \sum_{t=1}^T Z_{i,t} \right\},
\label{eq:2}
\end{equation}
where $Z_{i,t} = \log[1+r_{i,t}]$.
This logarithmic transformation defines, for any $i=1,\dots,n$, the individual series of realised financial returns
$\{ Z_{i,t} ; \ t=1,\dots,T \}$ as a sequence of identically distributed random variables,
and we require the following conditions:
\begin{enumerate}[(i)]
\item strictly positive finite second moments $0< \mathbb E Z_{i,t}^2 < +\infty$ \label{var1},
\item $|\mathbb E \left(Z_{i,t} Z_{j,t}\right)|^2 < \mathbb E Z_{i,t}^2 \mathbb E Z_{j,t}^2$ for $i\neq j$ and any $t\ge 1$.
\label{var2}
\end{enumerate}
Defining $\{Y_{i,j}(t) = Z_{i,t} - Z_{j,t}, t\in \mathbb N\}$, for $i\neq j$, by (\ref{var2})
we have that $\mathbb E (Y_{i,j}(t)^2)>0$.
Furthermore we assume the following condition on the temporal dependence of $Y_{i,j}(t)$:
\begin{equation}
\label{eq:3}
\inf_{i\neq j} \left|\sum_{t=1}^T Y_{i,j}(t)\right| \mathop{\longrightarrow}_{T\to\infty} + \infty
\qquad \text{in probability},
\end{equation}
which means that for any $0<M<+\infty$:
\begin{equation*}
\lim_{T\to\infty}\mathbb P\left( \inf_{i\neq j} \left|\sum_{t=1}^T Y_{i,j}(t)\right| \ge M\right) = 1.
\end{equation*}
There are many explicit conditions that imply \eqref{eq:3} for different reasons:
\begin{itemize}
\item If $\mathbb E(Z_{i,t}) \neq \mathbb E(Z_{j,t})$ for all $i\neq j$.
\item If $\{Y_{i,j}(t)\}_t$ are \emph{martingale increments}, i.e. $\mathbb E(Y_{i,j}(t)| Y_{i,j}(s), s<t) = 0$.
\item If $\{Y_{i,j}(t)\}_t$
is strongly mixing satisfying a condition for a limit theorem (cf. \cite{Ibragimov and Linnik 1971}).
\end{itemize}
Define $X_i(T) = \sum_{t=1}^T Z_{i,t}$, $W_i(T) = W_{i}(0)\; e^{X_i(T)}$ and
\begin{equation*}
\theta_{i}(T) = \frac{W_{i}(T)}{\sum_j W_j(T)}, \qquad i=1, \dots, n,
\end{equation*}
and $\tilde \theta^T_{i}(T)$ the corresponding ranked sequence at time $T$:
$\tilde\theta_{i}(T) \le \tilde\theta_{i+1}(T)$.
This indicator ranks all the relative individual wealth $W_i(T)$.
In particular $\tilde W^T_i(0)$ is the initial position for the sequence ranked at time $T$.
\begin{thm} Under the above conditions, the following limits holds in probability:
\begin{equation}
\label{eq:1}
\lim_{T\to\infty} \tilde\theta_{n}(T) = 1, \qquad \lim_{T\to\infty} \tilde\theta_{i}(T) = 0 \quad \forall i\neq n
\end{equation}
\end{thm}
\begin{proof}
Let $\tilde X_i(T), \tilde W_i(T)$ be the corresponding ranked sequences at time $T$,
\begin{equation*}
\begin{split}
\tilde\theta_{n}(T) = \left(1 + \sum_{j\neq n}\tilde W_j(T) \tilde W_n(T)^{-1}\right)^{-1}
= \left(1 + \sum_{j\neq n} \frac{\tilde W^T_j(0)}{\tilde W_n^T(0)} e^{\tilde X_j(T) - \tilde X_n(T)}\right)^{-1} \\
\ge \left(1 + n \left(\sup_{j\neq i} \frac{ W_j(0)}{ W_i(0)} \right)
e^{-\inf_{j\neq n}(\tilde X_n(T) - \tilde X_j(T))}\right)^{-1}.
\end{split}
\end{equation*}
Since $W_j(0)>0$ for all $j$ we have that $\sup_{j\neq i} \frac{ W_j(0)}{ W_i(0)} < \infty$. Furthermore
\begin{equation*}
\inf_{j\neq n} \left(\tilde X_n(T) - \tilde X_j(T)\right) \ge \inf_{i\neq j} |X_i(T) - X_j(T)| =
\inf_{i\neq j} |\sum_{t=1}^T Y_{i,j}(t)|,
\end{equation*}
that implies $ \mathbb P\left(\tilde\theta_{n}(T) \ge 1-\epsilon\right)
\mathop{\longrightarrow}_{T\to\infty} 1$ for any $\epsilon>0$.
\end{proof}
\begin{rem}
Notice that the convergence is only in probability, and not almost sure. This means that
a single realization of a trajectory of $(\tilde\theta_{n}(t), t\ge 0)$ oscillates almost surely,
but the probability to find $\tilde\theta_{n}(t)$ strictly smaller than one converges to 0 as $t\to \infty$.
More precisely, a more accurate analysis shows that
\begin{equation}
\label{eq:5}
\limsup_{T\to\infty} \tilde\theta_{n}(t) = 1, \qquad \liminf_{T\to\infty} \tilde\theta_{n}(t) = 0, \quad
\text{almost surely.}
\end{equation}
\end{rem}
This result implies that financial accumulation through time by individual investors leads
to ever-increasing wealth concentration and inequality across them most of the time.
This result holds for whatever initial wealth distribution and whatever mean return for financial investments over time.
The result does not imply that one single individual will own all the wealth indefinitely, but in a large time scale it can move from one investor to another. The result says that most the time wealth remains
concentrated in the hand of one individual.
\section{Model and proof under simple return structure}
\label{sec:model-proof}
The previous section proved the link between financial accumulation through time and ever-increasing wealth concentration across individual investors. This section provides a complementary proof by studying simple return structure. The latter excludes financial accumulation from the financial investment process. Financial proceeds are then consumed or held for precaution (if positive), and they are not reinvested together with the previous capital stock of wealth. Simple return structure corresponds to the historical cost accounting rule of financial performance, which assumes an additive process through time between invested capital stock and generated income flow. Therefore the capital stock is reproduced more than accumulated. Financial capital is remunerated as a productive factor but not financially accumulated over time. We formalise this additive process as follows:
\begin{equation}
W_i(T) = W_i(0) \left( 1+ \sum_{t=1}^T r_{i,t}\right)
\label{eq:4}
\end{equation}
assuming that $r_{i,t} >= -1$.
For sake of simplicity, we do not impose further restrictions to make this investment process more realistic.
Note that individual wealth may become negative at some point of time.
We include only conditions that are required for mathematical proof.
This formalisation defines $\{ \mathbf r_t = (r_{1,t}, \dots, r_{n,t}), t = 1,\dots,T\}$
as {an ergodic sequence of positive random vectors} with finite expectation:
$E r_{i,t} = \mu_i < +\infty$.
\begin{thm}
With probability one, we have
\begin{equation}
\label{eq:2}
\lim_{T\to\infty} \theta_{i}(T) = \ \frac{W_i(0)\mu_i}{\sum_{ j=1}^n W_j(0) \mu_j} .
\end{equation}
\end{thm}
\begin{proof}
This is a direct consequence of the ergodic theorem. In fact with probability one
\begin{equation*}
\theta_{i}(T) \ =\ \frac{W_{i}(T)/T}{\sum_j W_j(T)/T} \ \mathop{\longrightarrow}_{T\to\infty}
\ \frac{W_i(0)\mu_i}{\sum_ j W_j(0) \mu_j} .
\end{equation*}
\end{proof}
\begin{cor}
If $\mu_j = \mu$ for any $j$, then $ \theta_{i}(T) \mathop{\longrightarrow}_{T\to\infty}\theta_i(0)$.
\end{cor}
\begin{rem}
No condition is required on the correlations between the components of
the random vector $\mathbf r_t = (r_{1,t}, \dots, r_{n,t})$.
\end{rem}
\begin{rem} Notice that
the convergence in \eqref{eq:2} is with probability one, stronger than the convergence in probability
in \eqref{eq:1}.
\end{rem}
\section{ Limits and implications}
\label{sec:limits-implications}
Our result concerning wealth concentration are twofold. In the simple interest return scenario, with probability one wealth concentration converges to certain values determined by the initial conditions. In constrast, in the compound interest return scenario, the probability that maximal wealth concentration does not occur tends to zero.
More precisely, most of the time, there exists one individual who owns all the wealth,
though not always the same individual through time: in a very large time scale
wealth can move from one lucky individual to another, passing through a fast redistribution.
Our result implies therefore that financial accumulation through time implies ever-increasing wealth concentration and inequality across individuals.
Wealth distribution tends to degenerate so that, at a certain point of time,
a single individual eventually owns virtually all wealth, relatively speaking.
This result holds for whatever initial wealth distribution, and it critically depends
on some variance in financial returns and some return decorrelation through time.
Moreover, this result does not depend on a specific distribution of returns as long as
it involves some finite positive variance of them.
Therefore, even efficient financial markets -- implying a fair investment game
that individual investors do repeatedly play through time -- involve ever-increasing wealth inequality.
{Since this result depends only on some temporal decorrelation through time,}
it may be extended to include mutual investment opportunities that deliver
a joint return to subsets of individual investors at one point of time.
Furthermore, our analysis may be extended by studying the likely transition
time patterns that generate wealth concentration and inequality over time.
Our analysis identifies the financial accumulation process as an important driver of increasing wealth inequality in economy and society. This effect does not depend on the condition that capital investment is remunerated as a productive factor, but on the peculiar opportunity that its financial proceeds are systematically reinvested, compounding financial returns. Ever-increasing inequality depends on some temporal decorrelation between individual return trajectories.
Further mechanisms may be introduced which counter this wealth inequality effect. Countering forces may include aggregate mechanisms that redistribute wealth $W_{i,t}$ across individuals at some point of time, compensating for increasing wealth concentration and inequality over time. An obvious candidate mechanism may be taxation.
Another countering force may emerge by reaching some limit to economic process. Although generally assumed by economic models, it may be unrealistic to assume compound returns that last indefinitely over time and circumstances
(\cite{Voinov and Farley 2007,Biondi 2011}). This implicitly maintains constant returns to scale for whatever involved size, while some decreasing returns may occur after some threshold due to natural and artificial limits to growth. Furthermore, recent studies suggest that higher inequality may reduce growth, limiting financial investment opportunities
(\cite{Berg and Ostry 2013,Ostry and al. 2014}). However, according to our analysis, as long as compound returns apply, financial accumulation at those returns involves ever-increasing wealth concentration and inequality.
\section{Concluding remarks}
\label{sec:concluding-remarks}
Our approach focalises only on one dimension of economic process, namely its capital factor and related financial investment through time and circumstances. Through formal modelling, we disentangle its featuring mechanisms and study financial investment process respectively under compound and simple return structures. By construction, our modelling strategy neglects interaction with other dimensions that may affect its impact on inequality. Mathematical proofs identify the set of conditions under which financial accumulation generates ever-increasing wealth concentration and inequality. Let alone without countering forces, compound return structure constitutes then an important driver for inequality in economy and society.
Recent evolution in economy and society has been featured by increasing financial investment opportunities through active capital markets around the world. Our result corroborates that this phenomenon may have contributed to increase wealth concentration and inequality across individuals in recent decades.
From an epistemological viewpoint, our approach points to the interaction between microscopic and macroscopic dimensions of the economic system under investigation. One cumulative effect concerns compound financial returns through time at the level of every individual investment portfolio. Another cumulative effect concerns wealth concentration across population of individual investors at every point of time.
Our approach contributes therefore to include space heterogeneity and time evolution into economic modelling. Concerning time, financial accumulation has been somewhat neglected because modelling strategies were considering one-period or two-period horizons, while the compound interest structure plays its featuring role when at least three successive periods are considered. Concerning space, our approach introduces a population of heterogeneous individuals. Individual financial investment trajectories constitute a heterogeneous set of financial accumulation processes performed under active financial markets through time and circumstances.
|
\section{First application: Detecting a changepoint in across-epochs correlation over a frequency band, for a single channel}\label{changepoint}
To illustrate the pertinence of the dependence issues related to brain signals for certain frequency bands, here we use experimental data from \citet{wann2017} on local field potentials measured on rats' cortex. To summarize that experiment, local field potentials were recorded from $32$ microelectrodes placed on $4$ cortical layers (each with $8$ electrodes). This setup is illustrated in Figure \ref{RatBrain}. On these $32$ microelectrodes (channels), using insulated stainless steel wire electrodes, data have been recorded for 5 minutes where each second represents a single epoch which consists of $T=1000$ time points. After these five minutes, a stroke have been mechanically induced using an hemostat clamp on the brain artery located on the second column of electrodes (from the left) recording microelectrodes $2,10,18$ and $26$. Then, for five minutes, data, divided in the same way as the previous five minutes, has been recorded.
\begin{figure}
\begin{center}
\includegraphics[scale=0.28 ]{RatBrain8.jpg}
\caption{Placement of the $32$ electrodes on the cortex of the rat. There are $4$ layers (having a different depth in the cortex: respectively $300\mu m$, $700\mu m$, $1100\mu m$ and $1500\mu m$) and each layer has $8$ electrodes. For details, see \citet{wann2017}. }
\label{RatBrain}
\end{center}
\end{figure}
Our interest in this section is to identify the epoch $r^{*}$ where the dependence between successive epochs $r^{*}-1$ and $r^{*}$ differ from the dependence between epochs $r^{*}$ and $r^{*}+1$. Hence, we are interested to identity that epoch $r^{*}$ for each microelectrode for all the $\Omega_\kappa \in \{ \Delta, \theta, \alpha, \beta, \gamma \}$ frequency bands. With LFP data, the dependence between $\delta_{\ell,\Omega_\kappa}^{(r)}$ and $\delta_{\ell,\Omega_\kappa}^{(r+1)}$ (no matters if these epochs are considered as a changepoint or not) exhibit frequently complex structure. For example, on Figure \ref{scatterplotCh17}, one sees (for rat id $141020$) for microelectrode (channel) $17$ some structures where the magnitudes of the Fourier coefficients for two successive epochs are highly dependent in their lower tail, and become more and more independent as they one moves toward their higher tail. This particular structure is easily representable through a copula function (Clayton copula will be considered in this case), but is not through any linear correlation (specifically coherence in the spectral domain) structure. For channel $1$ (Figure \ref{scatterplotCh1}) , one notices, still for the same rat, that the dependence structure change with stroke: for example, one notices the difference between the upper tails. This difference gets more and more obvious as we are looking for dependence between these magnitudes for epochs which get closer to the temporal interval: epochs $375$ to $380$. One notices that data in the first row are the one used in our algorithm. However, for a visualization purpose, data are log scaled in the second row in order to respond to skewness of large magnitude values.
\begin{figure}
\begin{center}
\includegraphics[scale=0.48]{ch1_dep_log.jpeg}
\caption{Changes in between-epoch dependence for microelectrode (channel) $1$, for rat id $141020$. \textbf{First row: }Plot of the rescaled magnitudes into unit interval (as used in the changepoint detection algorithm) for pre-stroke \textsl{(left)}, early post-stroke \textsl{(middle)} and late post-stroke \textsl{(right)}. \textbf{Second row: }Plot of the log-scaled magnitudes (for visualization purpose) for pre-stroke \textsl{(left)}, early post-stroke \textsl{(middle)} and late post-stroke \textsl{(right)} .}
\label{scatterplotCh1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.35]{Fig4_17.jpeg}
\caption{Changes in between-epoch dependence for microelectrode $17$, for rat id $141020$. Plot of the rescaled magnitudes into unit interval for pre-stroke \textsl{(left)}, early post-stroke \textsl{(middle)} and late post-stroke \textsl{(right)} .}
\label{scatterplotCh17}
\end{center}
\end{figure}
For this data, the expected major changepoint is $r^{*}=301$ which is the stroke onset. It is likely too that other changepoints would be observed after the stroke. However, for some biological issues, the peak of this observation might be delayed between the $375$$-th$ and the $380$$-th$ epoch (from $75$ to $80$ seconds after the stroke) for a majority of the frequency channels, on most of their frequency bands. We observe that this $5-$seconds activity window is subject to change in function of the rat on which experiment is conducted. We note that the way we segregated epochs (changepoint vs stable epoch) is based on the empirical setting presented in \ref{empiricalSetting}.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[scale=0.60 ]{ExampleBeta.jpeg}
\caption{Three different patterns in the regime of $\delta_{1, \beta}^{(1:600)},\delta_{9, \beta}^{(1:600)}$ and $\delta_{17, \beta}^{(1:600)}$. Red dotted lines represent the moment when the stroke is artificially induced.}
\label{threePat}
\end{center}
\end{figure}
We remark that, without regard to the frequency band, mainly three patterns are present in the regime of $\delta_{\ell, \Omega_\kappa}^{(1:600)}=[\delta_{\ell, \Omega_\kappa}^{(1)},...,\delta_{\ell, \Omega_\kappa}^{(600)}]$ with rat id $141020$. An interesting fact is that even if the location of the clamped artery is on column $2$, these three patterns are observed on column $1$. They are respectively microelectrodes (channels) $1$, $9$ and $17$. Figure \ref{threePat} exhibits the averaged amplitude (per epoch) for each one of these microelectrodes. The results for these three microelectrodes (for the five frequency bands) are presented as these are representative of our methodology.
\vspace{0.5em}
\hrule\vspace{0.3em}
{ \noindent \textbf{\underline{ALGORITHM 1:} Detection of a changepoint over many epochs, for a particular microelectrode (channel) and a given frequency band}}\hrule \vspace{0.5em}
{\noindent \textbf{for} (epochs $r=1$ to $r=600$) \newline
1: \hspace{0.3in} Standardize (scale data into $[0,1]$ interval) such that \newline \hspace*{1in}
$\tilde{\delta}_{\ell, \Omega_\kappa}^{(r)}=(\delta_{{\ell, \Omega_\kappa}^{(r)}}-\min(\delta^{(1:600)}_{\ell, \Omega_\kappa}))/(\max(\delta^{(1:600)}_{\ell, \Omega_k})-\min(\delta^{(1:600)}_{\ell, \Omega_\kappa}))$.
\newline
2: \hspace{0.3in} Apply the moving block bootstrap (to conserve the temporal structure inside \newline
\hspace*{0.5in} data, see Section \ref{section2}) by sampling on $X_\ell^{(r)}$ to obtain robust\newline
\hspace*{0.5in} estimations of the shape $\nu$ and the rate $\upsilon$ values of a Gamma \newline
\hspace*{0.5in} distribution and fit its cdf $\Gamma^{(r)}_{\ell,\Omega_\kappa}$ with parameters $(\nu^{(r)}_{\ell,\Omega_\kappa}, \upsilon^{(r)}_{\ell,\Omega_\kappa})$ \newline
\textbf{end } \newline
\textbf{for $r=1,...,599$} \newline
3: \hspace{0.3in} Compute Kendall's tau between $\delta_{\ell, \Omega_\kappa}^{(r)}$ and $\delta_{\ell, \Omega_\kappa}^{(r+1)}$. \newline
4: \hspace{0.3in} Among a predefined panel of parametric copulas, select using AIC the \newline
\hspace*{0.5in}most suitable copula model and using inverse Kendall's\newline
\hspace*{0.5in} tau method, estimate the corresponding copula dependence parameter. \newline
\hspace*{0.5in}This copula is noted $C^{(r,r+1)}_{\ell, \Omega_\kappa}\left( \Gamma^{(r)}_{\ell,\Omega_\kappa}(u),\Gamma^{(r+1)}_{\ell,\Omega_\kappa}(v) \right) \forall (u,v) \in [0,1]\times [0,1]$ \newline
\textbf{end } \newline
\textbf{for $r=2,...,599$} \newline
5: \hspace{0.3in} Compute all the bivariate Kolmogorov-Smirnov statistic $D(u,v)$ between \newline
\hspace*{0.5in} copulas $C^{(r-1,r)}_{\ell, \Omega_\kappa}\left( \Gamma^{(r-1)}_{\ell,\Omega_\kappa}(u),\Gamma^{(r)}_{\ell,\Omega_\kappa}(v) \right)$ and $C^{(r,r+1)}_{\ell, \Omega_\kappa}\left( \Gamma^{(r)}_{\ell,\Omega_\kappa}(u),\Gamma^{(r+1)}_{\ell,\Omega_\kappa}(v) \right)$ \newline
\textbf{end }\newline
\textbf{Output:} Kolmogorov-Smirnov statistics above a threshold determined for a desired significance level are said to be related to changepoint epochs.
\vspace{0.5em} \hrule}
Algorithm 1 presents the proposed methodology to detect one or many changepoint(s) in the regime of a channel for a given frequency band. Some remarks are as follows. Firstly, in this algorithm, we adopt the tilde notation (e.g., $\tilde{\delta}_{\ell, \Omega_\kappa}^{(r)}$) to differentiate standardized variables from regular variables, but in this paper we assume all variables to be standardized after this step. Therefore, we will no longer use this notation in order to simplify the text. Also, still in the same step, note that we use unique minimum and maximum over all the $600$ epochs range. It is crucial to use these standardization values in order to get comparable Kolmogorov-Smirnov statistics. Secondly, our panel of copulas consists only of Archimedean models in order to avoid potential problems of comparing widely different models. Finally, the thresholds for the Kolmogorov-Smornov statistics are determined empirically and the procedure is described in Section \ref{empiricalSetting}.
\input{sectionThresh2.tex}
\subsection{Changepoints observed on LFP}
To verify the validity of our method which has been applied to LFP data (results for rat id $141020$ are presented here), we decide to use also an estimator to detect changepoint by pruned objectives (see \citet{james2015}), and to verify if changepoint(s) detected by both methods concord. Figure \ref{changepoints} presents for two channels the detected changepoints, which are stated on table \ref{changeEpochs}, for a significance level of $\bar{\alpha}=1\%$. We remark that most of the time, epochs $374$ to $379$ are significant changepoints.
\begin{figure}
\begin{center}
\includegraphics[scale=0.65 ]{changepoints.jpeg}
\caption{Examples (for two microelectrodes, for rat id $141020$) of detection of the epochs related to a changepoint in the dependence structure (for $\bar{alpha}-1\%$) using the copula-based method, represented by the vertical red dashed lines. \textbf{Left:} For channel 1, alpha band. \textbf{Right:} For channel 17, gamma band.}
\label{changepoints}
\end{center}
\end{figure}
\begin{table}{\footnotesize
\begin{tabular}{|l||*{1}{c|c|}}
\hline
\textbf{Channel and} & Copula-based & Condordance with \\
\textbf{Frequency band} & algorithm & James algorithm \\
\hline \hline
Channel 1, $\Delta$ &$374,375,376,377,378,399,425$ & Yes\\
Channel 1, $\theta$ &$\*$(\textit{25 changepoints have been detected}) & Yes \\
Channel 1, $\alpha$ &$375,376,377,378$ & Yes \\
Channel 1, $\beta$ &$374,375,376,378,379,380,387$ & Yes \\
Channel 1, $\gamma$ &$374,375,376,378$ & Yes \\
\hline
Channel 9, $\Delta$ &$59,60,61,62,179,180,181,182,239,240$ & Yes \\
Channel 9, $\theta$ &$121,372,373$ & Yes \\
Channel 9, $\alpha$ & $214, 309, 372, 373, 374, 375, 376, 480$ & No \\
Channel 9, $\beta$ &$56,372,373,374,375,376$ & Yes\\
Channel 9, $\gamma$ & $60, 180, 182, 372, 373, 374, 375, 376$ & No \\
\hline
Channel 17, $\Delta$ &$179, 180, 181, 344, 345, 374, 375, 376, 377, 378, 399, 424, 425, 426, 524, 525$ & No \\
Channel 17, $\theta$ & $18, 88, 191, 215, 339, 373, 378, 390, 446, 467, 473, 563$ & Yes \\
Channel 17, $\alpha$ &$\*$(\textit{30 changepoints have been detected}) & Yes \\
Channel 17, $\beta$ & $\*$(\textit{91 changepoints have been detected}) & Yes \\
Channel 17, $\gamma$ &$375, 376, 377, 378$ & No \\
\hline
\end{tabular}}
\caption{Epochs considered as changepoints using the copula-based algorithm (with a threshold of $\bar{\alpha}=1\%$), for rat id $141020$. }
\label{changeEpochs}
\end{table}
\section{Conclusion}
This paper related some approaches to assess both the dependence and the information we can learn from that dependence (i.e. changepoint, change in a regime, etc.). By considering more complex structures than simply linear relations (e.g., linear correlation, coherence) to assess dependence between brain signals, we modeled the relations between these signals. We also proposed algorithms with which we determined if one can presume a change or not on these complex structures of dependence. Such a methodology aims to show his utility in the future because research about specific types of strokes like CVA gets more and more funded in order to do prevention in the society.
In closing, we address two potential criticisms of the proposed work. Firstly, we used only parametric copula models when it is true that in general, non-parametric models are more flexible to data. But in our context, the dimension of some frequency bands is not large enough to ensure the robustness of a nonparametric model as a parametric model might be. Secondly, the analysis was conducted only on four rats. It is true that data from many more rats will increase the power of the neurological conclusions. However, the work done here was an explanatory study of a copula-based approach for such data, and having to analyze data from many more rats will complexity the computational work. Many future research avenues are possible from what we did. One of these avenues is to study the impact of taking copulas on more than $2$ epochs while processing the iterative algorithm when studying a changepoint for a single brain channel. Indeed, it will allow to detect changes that occur for small windows of time instead that punctually. Another one is to study our copula-based approach the possible lagged dependence(s) between two different frequency bands.
\section{Statistical prologue and notation}\label{notation}
To facilitate ease of reading of the paper, we include the notations in Appendix \ref{appendixNotation}. Let ${\bf{X}}=\left[ X_1,X_2,...,X_d\right]$ a three-dimensional matrix of dimension $T \times d \times R$ ($d$ brain channels divided into $R$ possibly over-lapping epochs (i.e., equal segmentation of $1$ second into $T$ time points)). Thus, one represents a single element of that matrix by $X_\ell^{(r)}(t)$, $\ell=1,...,d$; $ t=1,...,T$; $ r=1,...,R$, which might be seen as any recorded measure on channel $\ell$ during epoch $r$. Therefore, the 3-dimensional matrix is composed of $R$ matrices of size $T \times d$ denoted by ${\bf X}^{(r)} = [X_1^{(r)}, \ldots, X_{d}^{(r)}]$.
This paper focuses on dependence among brain channels in the frequency domain. We remark that from the experimental perspective, the channels are defined by microelectrodes different parts of the brain. The Fourier coefficient for the channel $\ell=1,...,d$, at epoch $r=1,...,R$ and at fundamental Fourier frequency $\omega_k = k/T$ is defined to be
$$f_{\ell \omega_k}^{(r)} = \frac{1}{\sqrt{T}} \sum_{t=1}^{T} X_\ell^{(r)}(t)
\exp(-i 2 \pi \omega_k t).$$ Because this transformation outputs single frequencies and in our context we are interested in frequency bands, we have to segregate frequencies according to the bands' ranges and to smooth their magnitudes by averaging.
In this paper, we will study the dependence of magnitudes of the Fourier coefficients (or square roots of periodograms) between the different pairs of channels $\ell$ and $ \ell'$; $\ell,\ell'=1,...,d$ for the same epoch $r$. In addition, we will investigate the dependence between successive pairs of epochs $r$ and $r+1$. We denote the frequency bands by $\Omega_\kappa$ where $\kappa=1,...,Q$ represents the frequency band, and where $Q$ is the number of frequency bands considered in this paper (see Section \ref{section2} for details about the chosen bands). Thus, $\Omega_\kappa$ is a collection of fundamental Fourier frequencies $\omega_j$, $j=\text{index}(\Omega_\kappa^{(\text{min})}),...,\text{index}(\Omega_\kappa^{(\text{max})})$ where $\text{index}(\Omega_\kappa^{(\text{min})})$ is the index of the first value constituting the frequency band among the $k$ fundamental frequencies, and $\text{index}(\Omega_\kappa^{(\text{max})})$ is the one of the last frequency constituting that band.
We now define ${\bf{\delta}}^{(r)}_{\Omega_\kappa} = [\delta_{1,\Omega_\kappa}^{(r)},...,\delta_{d,\Omega_\kappa}^{(r)}]$ which is the matrix of dimension $\text{card}(\Omega_\kappa) \times d$ where any column is a different channel $\ell=1,...,d$. Therefore, each column is represented by $\delta_{\ell,\Omega_\kappa}^{(r)}=[| f_{\ell,\Omega_\kappa^{(\text{(min)})}}^{(r)}|,...,| f_{d,\Omega_\kappa^{(\text{(max)})}}^{(r)}|]'
$, a vector of length $\text{card}(\Omega_\kappa)$ containing the magnitude for each Fourier fundamental frequency constituting the frequency band $\Omega_\kappa$ at epoch $r$. Hence, in the rest of this paper, we will consider $\delta_{\ell, \Omega_\kappa}^{(r)}, \delta_{\ell', \Omega_\kappa}^{(r)}$ as the random vectors on which our methodology is applied.
\subsection*{Copula function}
Let the brain channels be indexed by $\ell,\ell' \in \{1,2,...,d\}$, let the epochs be indexed by $r=1,...,R$ and let the frequency bands of interest to be $\Omega_{\kappa}$ and $\Omega_{\kappa'}$. Our goal is to assess the dependence between $\delta_{\ell,\Omega_\kappa}^{(r)},\delta_{\ell ',\Omega_{\kappa '}}^{(r')}$ in the cases where $\ell=\ell', r\neq r', \Omega_\kappa=\Omega_{\kappa'}$, where $\ell \neq\ell', r=r', \Omega_\kappa=\Omega_{\kappa'}$ and where $\ell =\ell', r=r', \Omega_\kappa \neq \Omega_{\kappa'}$, we will express the dependence between these two quantities by expressing their joint cumulative distribution function. To this end, one writes $H_{(\ell,\Omega_\kappa)(\ell ', \Omega_{\kappa '})}^{(r,r')}(\delta_{\ell,\Omega_\kappa}^{(r)},\delta_{\ell ',\Omega_{\kappa '}}^{(r')})$ as the bivariate joint distribution for the random variables $\delta_{\ell,\Omega_\kappa}^{(r)}$ and $\delta_{\ell ',\Omega_{\kappa '}}^{(r')}$. We denote its marginal distributions by $H_{\ell,\Omega_\kappa}^{(r)}(\delta_{\ell,\Omega_\kappa}^{(r)})$ and $H_{\ell',\Omega_{\kappa'}}^{(r')}(\delta_{\ell ',\Omega_{\kappa'}}^{(r')})$. Using Sklar's theorem\cite*{sklar1959}, this joint distribution can be rewritten in terms of a unique copula:
$$
H_{(\ell,\Omega_\kappa)(\ell ', \Omega_{\kappa '})}^{(r,r')}(\delta_{\ell,\Omega_\kappa}^{(r)},\delta_{\ell ',\Omega_{\kappa '}}^{(r')})=C_{(\ell,\Omega_\kappa)(\ell ', \Omega_{\kappa '})}^{(r,r')}\left( H_{\ell,\Omega_\kappa}^{(r)}(\delta_{\ell,\Omega_\kappa}^{(r)}),H_{\ell',\Omega_{\kappa'}}^{(r')} (\delta_{\ell ',\Omega_{\kappa '}}^{(r')}) \right)
$$ where $C$ is the exact copula linking $\delta_{\ell,\Omega_\kappa}^{(r)}$ to $\delta_{\ell',\Omega_{\kappa'}}^{(r')}$. \citet{fontaine2018} provided an inferential framework for such a joint model in the spectral domain. For the rest of this paper, in the case of $\ell=\ell'$ and $\kappa=\kappa'$, we reduce this notation to $C_{\ell,\Omega_\kappa}^{(r,r')}$. We also assume that the copulas are fully parametric meaning that either the copula structures than the marginal distributions are parametric. Furthermore, we assume the true copula parameter $\breve{\theta}$ to be inferred in two possible ways (depending on the clinical question we are trying to answer): by a maximum likelihood estimation denoted $\hat{\theta}_{{\bf K}}$ or by the inversion of Kendall's tau method, denoted $\hat{\theta}_\tau$. We remark that although many parametric families of copulas have been studied in the literature (see \citet{genest1986} or \citet{nelsen2007}), selecting a suitable copula model may be tricky. Therefore, in Section \ref{copulaSel}, we discuss the selection of a copula model and the impact of misspecification.
\section{Theoretical framework related to copulas and distributions}\label{section2}
Prior to any statistical modeling, we applied a Fourier transform to the multichannel multiple-epoch brain signals, considering each microelectrode $\ell$ at each epoch $r$ as a single data vector. This transform was in order to obtain the values of the periodograms and then compute the magnitudes of these Fourier coefficients. As mentioned earlier, we decided to use frequency bands rather than single frequencies. To determine the range of the considered bands, we based our choice on a classical text in the literature: \citet{buzsaki2006} about the rhythms of the brain. Hence, we adopted the following bands: $\Delta \in (0,4)$Hertz, $\theta \in (4,8)$Hertz, $\alpha \in (8,12)$Hertz, $\beta \in (12,30)$Hertz and $\gamma \geq 30$ Hertz. We note that in our applications, we truncated $\gamma$ at $300$ Hertz and applied a notch filter to remove the $60$ Hertz activity.
Before describing the specific cases where we assess the dependence among brain channels for particular frequency bands, we first discuss the application of the Kolmogorov-Smirnov statistic to the multivariate setting. Kolmogorov-Smirnov might be used to compare two cdf together. In the univariate case, if $A(x)$ and $B(x)$ are two cdf, to test $H_0: A=B$ versus $H_1:A \neq B$, we use the statistic $$ D = \sup_{x \in \mathbb{R}} \|A(x)-B(x) \|,$$ which is known to converge almost surely to 0 under $H_0$ due to Donsker's theorem\cite*{donsker1952}.
In this paper, we are interested in the empirical value of that statistic in a multivariate context. We use it in different ways according what we study. However, the way that we deal with the statistics of test remains the same. In a bivariate situation, let ${\bf{X}}=(X_1,X_2)'$ and ${\bf{Y}}=(Y_1,Y_2)'$ be two random variables with respective joint cdf $A$ and $B$. Also, let $u,v$ be two finite partitions in any closed subset of $\mathbb{R}^2$, large enough to contain the supports of $\bf{X}$ and $\bf{Y}$. Hence, we define our computational approach of the bivariate Kolmogorov-Smirnov statistic as
\begin{eqnarray*}
D(u,v)&=& \sup_{(u,v)} \left| A(u,v)-B(u,v) \right| \\
&=& \sup_{(u,v)} \left| C_{\bf{X}}(A_{X_1}(u),A_{X_2}(v))-C_{\bf{Y}}(B_{Y_1}(u),B_{Y_2}(v)) \right|
\end{eqnarray*}
where $C_{\bf{X}},C_{\bf{Y}}$ are respectively the unique copulas equal to $A$ and $B$ according to Sklar. In practice, variables are on different supports (e.g., the amplitude of signals for $\delta$-frequency band versus the one for $\beta$-frequency band) and finding a finite grid of values $u$ and $v$ containing the support of both $\bf{X}$ and $\bf{Y}$ might be a tricky task. That is the reason why we standardize data into the $[0,1]$ interval (see how in Section \ref{changepoint}).
Under a real equality in distribution for $C_{\bf{X}}$ and $C_{\bf{Y}}$, for $\tilde{u}\in [0,1]$ and $\tilde{v}\in [0,1]$ standardized versions of $u$ and $v$ being vectors of sufficiently large dimension, the statistic $D_{\bf{\tilde{X},\tilde{Y}}}$, where ${\bf{\tilde{X},\tilde{Y}}}$ are standardized versions of $\bf{X},\bf{Y}$, is nothing more than the bivariate version of the usual Kolmogorov-Smirnov statistic.
A remaining issue with Kolmogorov-Smirnov is that the validity of this statistic relies on the robustness of the distributions. However, due to the cardinality of the low-frequencies bands, estimating any parameter directly on these bands will lead to non-robust distributions. That is the reason why one has to use resampling techniques in order to obtain some distributions and then derive their parameters (e.g., mean and standard deviation).
\subsection{Block bootstrap for small frequency bands}
Due to the small cardinality of some frequency bands (i.e., those composed of a small quantity of single frequencies) such that $\Delta , \theta , \alpha$ or $\beta$ (e.g., the actual frequencies considered in the $\Delta$ band are $\{ 1,2,3,4 \}$ Hertz), any standard parametric inference methodology applied on the magnitude of the different Fourier frequencies within them, for a fixed epoch, will suffer from a lack of robustness. Indeed, with such small populations, any standard estimation (e.g., estimation of the parameters of the distribution) will lead to a statistic for which its variance with likely suffer from a lack of robustness. It is the reason why one has to use resampling methods while inferring distribution parameters in order to obtain a gain in robustness the variance of the estimators.
Let $X^{(r)}_\ell$ be the time-domain valued vector, of dimension $T$, for channel $\ell$ at epoch $r$. Computing straightforwardly the modulus of the Fourier transform, one obtains $\delta^{(r)}_{\Omega_\kappa, \ell}$, a vector whose cardinality might not be sufficiently large. We applied resampling techniques in order to obtain an empirial distribution of $\delta^{(r)}_{\Omega_\kappa, \ell}$. However, any naive use of bootstrap methods (\citet{efron1994}) will destroy the temporal structure among the $T$ observations of $X^{(r)}_\ell$. For this reason we apply the moving block bootstrap (see \citet{politis1994} or \citet{radovanov2014}) which preserves the temporal structure of the time series within an epoch. Here, we define $M$ to be the number of blocks, each with $T/M$ observations. Thus, one gets the bootstrapped variables $X_\ell^{b,(r)}$ for $b=1,...,B$ the number of iterations. One remarks that in this work, bootstrapped observations are only used to estimate the parameters of the distributions of $\delta_{\Omega_\kappa}^{(r)}$, they are not directly used on any measure of the strength of the dependence between variables represented through Kendall's tau or coherence measure.
\subsection{Estimation of the distributions}
Still for a reason of data size of $\delta_{\Omega_\kappa, \ell}^{(r)}$, we decided to avoid any empirical or non-parametric estimation of the distribution of $\delta_{\Omega_\kappa, \ell}^{(r)}$. As shown in \citet{brockwell2013}, the asymptotic distribution of the periodogram of a time series follows an exponential distribution with mean $\lambda$ equals to the spectrum. By some algebraic manipulations, we show in Appendix \ref{appendixRay} that the square root of an exponential distribution follows a Rayleigh distribution of parameter $1/\sqrt{2\lambda}$. Note that Rayleigh is a special case of the generalized Gamma distribution. Since the generalized gamma distribution is a model with three parameters (which allows room for computational bias in their estimation due to the idiosyncrasies of data for some frequency bands), we decided to use two-parameter models of that family to infer the distribution of $\delta_{\Omega_\kappa, \ell}^{(r)}$ in order to reduce inferential bias due to the small data size as well as to increase computational speed in the inferential process. Thus, we compared the likelihood of fitting a gamma distribution versus the one of fitting a two-parameters Weibull distribution to the LFP data, on all channel. Hence, with the help of an information criterion ($BIC$ - see Section \ref{copulaSel}), we decided to use the gamma distribution to model $\delta_{\Omega_\kappa, \ell}^{(r)}$. In the rest of this paper, we adopt the notation $\Gamma^{(r)}_{\Omega_\kappa, \ell}$ to denote the estimated distribution of the variable $\delta_{\Omega_\kappa, \ell}^{(r)}$ where the parameters $(\nu, \iota)$ are estimated by the maximum likelihood estimators $(\hat{\nu}, \hat{\iota})$.
\subsection{Selection of a copula model} \label{copulaSel}
The copula-based algorithms to detect changes in brain signals, which are presented in this paper, can be fit using various types of copula functions. Among the most common ways of model selection from a wide set of possible types of copulas, we find those based on an information criterion. For instance, Akaike Information Criterion (AIC, \cite{akaike1998}), Bayesian Information Criterion (BIC, \cite{schwarz1978}) or Copula Information Criterion (CIC, \citet{gronneberg2014}) are some of these possibilities. In this paper, to reduce the computational aspect of the algorithms and because only some slightly differences has been shown to exist between AIC and CIC (\citet{jordanger2014}), we use AIC to select all the copula models.
The range of copula models to consider for such a methodology is arbitrary. In this paper, in an attempt to avoid any numerical issues/misscomputations while computing the differences between some copula models (e.g., the difference between a normal copula and a Gumbel copula might be very high for border values due to their divergent behavior in these areas), we confined our choice only to the Archimedean family of copulas. We made this choice based on the flexibility of that family: elliptical copulas exhibit always a radial symmetry, which is not the case of the Archimedean copulas; furthermore Archimean copulas allow easily to model skewed distributions with non-symmetric tails. Thus, the panel of considered copulas was restricted to: independent, Clayton, Gumbel, Frank, Joe and rotated Joe ($180$ degrees) copulas (see \citet{cech2006} for more about rotated copulas).
\subsubsection{Effect of misspecification}
In this work, we suggest to limit the panel of available copula models to $6$ types of copulas from the Archimedean family. Let $\breve{C}^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \breve{\theta})$, $r=1,...,R$ be the true copula (with its true parameter $\breve{\theta}$) which is maybe or not in our selection panel, and $C^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \bar{\theta})$ be the one selected using AIC (or any other method based on likelihood information) with its parameter. Then, we express the Kullback-Leibler Information Criterion (KLIC) by
$$\text{KLIC}
=\int_0^1 \int_0^1 \text{log} \left( \frac{\breve{c}^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \breve{\theta})}{c^{(r,r')}_{\ell, \Omega_\kappa}(u,v;\bar{\theta})} \right) d\breve{C}^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \breve{\theta}). $$
If $\breve{C}^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \breve{\theta})$ is in the panel of considered copulas choose based on a likelihood-based criterion, then indeed $\breve{C}^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \breve{\theta})=C^{(r,r')}_{\ell, \Omega_\kappa}(u,v; \bar{\theta})$ and the KLIC will equal $0$. Otherwise, concerning the copula structure, as long as the real copula function $\breve{C}^{(r,r')}_{\ell, \Omega_\kappa}$ is unknown, it is not realistic to give a value of KLIC. However, we can minimize this criterion using a panel of flexible and rich possible families of copulas. Nevertheless, under a misspecified model, it is possible that the equivalent of $\breve{\theta}$ for the selected model does not exist. However, a pseudo-true parameter $\bar{\theta}$ exists. From \citet{white1982}, under conditions of continuity and measurability (which are met in this paper by LFP data) an estimator $\hat{\theta}_k$ of $\bar{\theta}$ obtained by a maximum likelihood estimator computed from the misspecified model will be, as $n=\text{card}(\Omega_\kappa) \to \infty$ (which is the case for Gamma band in the experimental setting described below, but might be the case for all frequency bands when there are enough time points within each epoch), consistent.
\section*{Acknowledgements}
Ron D. Frostig was supported by the Leducq Foundation (grand 15CVD02).
\bibliographystyle{chicago}
\begin{flushleft}
\end{flushleft}
{\setstretch{1.5}
\section{Introduction}
Brain stroke occurs when blood circulation in one of the cerebral blood vessels is abnormally weak, and in such case, leads to death of the cells.. Brain stroke has been studied for years by biologists and neurologists. Studying this disorder from the perspective of the changes in the brain’s electrical activities among different regions has yielded many clinically important results: these changes are so important that often they do irreversible damages to patients and incur extravagant costs to society (\textit{e.g., high medical expenses and low quality of patients’ lives}). In order to reduce these societal costs, neuroscientists study the behavior of the cortex activity by inducing stroke in rats. Due to ethical considerations, stroke experiments are conducted mostly only on rats. This paper is based on an experimental setup designed to induce stroke in a rat and to study the electrical oscillations among different regions in his brain. Using the copula information, we developed methods for assessing and analyzing dependence between the rat’s brain regions. Our work is in collaboration with neuroscientists from University of California at Irvine (co-author Frostig and student \citet{wann2017}) who mechanically induced brain stroke in the rats by clamping a brain artery and recorded the brain activity on $32$ microelectrodes (or channels) before and after the stroke. Figure \ref{channel1Intro} shows, for one of the analyzed rats, how the data act differently in the pre-stroke phase (first $5$ minutes or first $300$ epochs) versus in the post-stroke phase (last $5$ minutes or last $300$ epochs). The detailed setup is described in Section \ref{changepoint}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.60 ]{channel1Intro.jpeg}
\caption{Microelectrode (channel) $1$ for rat id $141020$. Red dotted line represents the onset of the induced stroke. \textbf{Left:} Local field potential recording on channel $1$. \textbf{Middle:} Plot of the magnitude of Fourier coefficients in the delta band across all epochs. \textbf{Right:} Plot of the magnitude of Fourier coefficients in the beta band across all epochs.}
\label{channel1Intro}
\end{center}
\end{figure}
One goal is to analyze the changes in the dependence between some channels for all frequency bands by using flexible models. Most analyses use coherence or correlation which are simple to implement but they are severely limited because they capture only linear dependence structures. Thus, we present an innovative methodology based on the notion of copula function to capture the complexity of the dependence and by comparing two (or more) copulas. Moreover, we assess whether or not the dependence between pairs of epochs changes across ($1$-second) epochs of the entire recording period. From the recorded electrical activity during the laboratory experiment, the impact of brain stroke is observable for most microelectrodes on almost all frequency bands; and the effect of the stroke on brain signals appear to last throughout the entire post-stroke recording (see Appendix \ref{appChannels}). Our contributions in this paper are the following. First, we present an algorithm to help recognize which epoch(s) exhibit changes in the dependence structure of the brain signals. This recognition of a changepoint is key to understanding the biological mechanisms occurring in the time window between the onset of the stroke and the moment where significant changes occur. Second, we present a method to assess if the dependence structure during pre-stroke differs from the one from post-stroke. This assessment is crucial to understand if this channel will be impacted by the stroke. This method is also used later to compare the dependence structure among two different channels for a given frequency band.
In the literature, many studies investigated changes in dependence for brain channels (in electroencephalograms) defined in the spectral domain. Among them, we highlight \citet{ombao2016}, \citet{fiecas2016}, \citet{long2004}, \citet{purdon2001}, \citet{nunez2006} and \citet{gotman1982}. However, the primary limitation of these studies is that they look only into the linear dependence between signals. Thus, they could miss potential complex (or non-linear) dependence structure between signals. Most methods reported explored the problem of detecting one (or many) changepoint moment(s) (e.g., within an epoch). A major approach is based on segmentation of the series in order to assess a possible discrepancy between these segments: on either a change in mean or a difference in the correlation structure. Many authors considered the segmentation: e.g., \citet{adak1998} with binary trees and windowed spectra to adaptively partition data; \citet{ombao2001} derive a segmentation by selecting the best localized basis from the SLEX (smooth local exponential) library. Another example is the estimation of a penalized minimum contrast (\citet{lavielle1999}). Its principle has two steps. In the first step, a contrast function is computed over a segments of a time period (or a sequence defined in the frequency domain - see \citet{lavielle2000}). The changepoints are then selected to be a solution to the minimization problem. Another example of that segmentation is based on probabilistic pruning methods. The principle of pruning is to predict the probability that a segment belongs to a stationary process rather than its likelihood. This method has been well studied by \citet{james2015} and \citet{kifer2004}. Another approach presented by \citet{davis2006} is the Auto-PARM: it consists in fitting multiple auto-regressive (AR) functions to segments of time. But fitting the AR model could be subject to model misspecification. The third kind of methods for detecting changepoints is based on hypothesis tests. \citet{dette2009} and \citet{dette2012} proposed an approach to test the equality of spectrum between two successive segments. This idea iis interesting but it does not take into account the nature and the structure of the dependence between these successive segments.
The use of the joint cumulative distribution functions with brain signals has also been explored to study dependence between random variables in general. These functions, namely \textit{copula} models, have the main advantage to represent the dependence as functions that provide the information of both ''strength'' and ''structure'' of the relation between two variables. For example, in Figure \ref{channel1Intro}, for the three cases, it is obvious that the dependence pattern between succeeding epochs during pre-stroke (first $300$ epochs) and the one during post-stroke (last $300$ epochs) are different and that the dependence structure from epoch $300$ to epoch $400$ is not the same than the one between epoch $500$ and epoch $600$. These particularities in dependence structure will be fully detected with a copula under a right specification. \citet{iyengar2010} used it to quantify synchronicity between multiples electroencephalographic (EEG) signals. \citet{dauwels2012} used copulas in their attempt to design brain network. \citet{ince2017} presented a framework to assess dependence for neuroimaging data based on the gaussian copula. Even if all of these approaches presented a copula-based framework for brain signals data, none of them was interested in a detection of a change (or of a changepoint) between successive epochs.
\input{exampleMotivation.tex}
We present in this article a copula-based framework to analyze changes between brain signals on given frequency bands for three different contexts. Firstly, we are interested in the detection of one (or many) changepoint(s) in the regime of a brain channel for a given frequency band. Secondly, we compare, within a single channel (microelectrode), if there is a difference in the dependence between successive epochs across the pre-stroke and post-stroke epochs. Thirdly, we compare the dependence structure of two different microelectrodes still on a given frequency band, on the entire recording time of $5$ minutes priot to and $5$ minutes post to the induced stroke.
The remainder of this paper is organized as follows. In Section \ref{notation}, we present present briefly the transformation of time data to spectral data as well as the copula function in order to introduce our notation. In Section \ref{section2}, we present the necessary theoretical background to introduce our models and algorithms. Then, in Sections \ref{changepoint}, \ref{prepost} and \ref{twoChan}, we present analyses of the local field potential data recording during a span of $10$ minutes ($5$ minutes pre-stroke and $5$ minutes post-stroke). Our methodology directly applied to these data shows its performance by, at first, assessing the statistically significant changepoints in dependence between successive epochs for some specific channels. Secondly, it shows that the whole dependence structure between pre-stroke epochs is not significantly different, for all the channels, than the one for post-stroke epochs.
\section{Second application: Comparing dependence prior to and post induced stroke}\label{prepost}
Our goal here is to compare the spectral dependence of the magnitude of Fourier coefficients pre-stroke versus the one post-stroke (i.e., to compare for a fixed microelectrode - understand ''a fixed channel''- and a fixed frequency band if there is a change in the entire structure of dependence among the $300$ epochs before the stroke versus after).
To do so, consider a given frequency band $\Omega_\kappa$, $\kappa=1,...,Q$. We define the multivariate matrices $ \delta_{\ell,\Omega_\kappa}^{(1:300)}=[\delta_{\ell,\Omega_\kappa}^{(1)},...,\delta_{\ell,\Omega_\kappa}^{(300)} ]$ and $ \delta_{\ell,\Omega_\kappa}^{(301:600)}=[\delta_{\ell,\Omega_\kappa}^{(301)},...,\delta_{\ell,\Omega_\kappa}^{(600)} ]$ (two matrices of dimension $\text{card}(\Omega_\kappa) \times 300$) as two single structures of the dependence. Using straightforwardly a single parametric copula in each case would be an enormous mistake. In fact, the parameter(s) of any Archimedean copula is too general to represent at the same time both the dependence measure between $\delta_{\ell,\Omega_\kappa}^{(1)}$ and $\delta_{\ell,\Omega_\kappa}^{(2)}$ and the dependence measure between $\delta_{\ell,\Omega_\kappa}^{(r)}$ and $\delta_{\ell,\Omega_\kappa}^{(r+1)}; r=2,...,299$. That's the reason why we propose here to use \textit{vine copulas} (for information, see see \citet{bedford2002} and \citet{aas2009}) to represent the dependence between these sets of variables.
The principle of vine copulas is the representation of a multivariate copula as a nested network of bivariate copulas where each single copula is named a \textit{node} and each link between two nodes (defining the order of the copulas and their relations among themselves) is named an \textit{edge}. Each level of dependence in this nested network is named a \textit{tree}. For all type of vines, the first tree is always the set of copulas between the univariate nodes (variables) and for the following trees, the nodes are always conditionals to at least one variable.
In this paper, due to the temporal relation between the consecutive $\delta_{\ell, \Omega_\kappa}^{(r)}, r=1,...,600$, we assume the structures of dependence for the multivariate sets $ \delta_{\ell,\Omega_\kappa}^{(1:300)}$ and $ \delta_{\ell,\Omega_\kappa}^{(301:600)}$ being represented by drawable vine (D-Vine) copulas. In other words, we assume that the edge between any node in the first tree only link the consecutive variables $\delta_{\ell,\Omega_\kappa}^{(r)},\delta_{\ell,\Omega_\kappa}^{(r+1)}; r=1,...,599$.
Obviously, we cannot use a 300-variate version of the Kolmogorov-Smirnov statistic to compare $ \delta_{\ell,\Omega_\kappa}^{(1:300)}$ to $ \delta_{\ell,\Omega_\kappa}^{(301:600)}$. It is still computationally unfeasible. That's the reason why, in our work, we propose to adapt a test comparing two vine copulas models (see \citet{clarke2007})to our context consisting in determining any difference in the structure of dependence between them. We remark that this test is mainly used in the literature in a goodness of fit perspective of a vine structure given a set of data. Since $ \delta_{\ell,\Omega_\kappa}^{(1:300)}$ and $ \delta_{\ell,\Omega_\kappa}^{(301:600)}$ are two different set of data having the same dimensionality which does not need to be independent, it is appropriate to use it.
The principle of that test is as follow. Let the ratios of the log-likelihhood for each Fourier frequency in the band under consideration $m_{i;\ell,\Omega_\kappa}=\text{log }\left\{ \frac{c^{(1:300)}_{\ell,\Omega_\kappa}({\bf{u}}_i|\theta^{(1:300)}_{\ell,\Omega_\kappa})}{c^{(301:600)}_{\ell,\Omega_\kappa}({\bf{u}}_i|\theta^{(301:600)}_{\ell,\Omega_\kappa})} \right\} $, $i=1,...,\text{card}(\Omega_\kappa)$ where $c$ stands for the density of a D-Vine copula function, $\theta^{(1:300)}_{\ell,\Omega_\kappa}, \theta^{(301:600)}_{\ell,\Omega_\kappa}$ the vectors of the copula parameters for each vine structure and ${\bf{u}}_i$ the vector of observations. Thus, if there is no difference between the vine copulas of the sets of variables $ \delta_{\ell,\Omega_\kappa}^{(1:300)}$ and $ \delta_{\ell,\Omega_\kappa}^{(301:600)}$, the ratios of the log-likelihhood $m_{i;\ell,\Omega_\kappa}$ should be uniformly distributed around zero and $50\%$ of them should be greater than 0 (for details and proof, see \citet{vuong1989}). Thus, we are testing for all $i=1,...,\text{card}(\Omega_\kappa)$: $$ \begin{cases} H_0:\mathbb{P}\left( m_{i;\ell,\Omega_\kappa} > 0 \right)=0.5, \\ H_1:\mathbb{P}\left( m_{i;\ell,\Omega_\kappa} > 0 \right)\neq 0.5. \end{cases}$$ Therefore, the statistic of test is $$\xi_{\ell,\Omega_\kappa}=\sum_{i=1}^{\text{card}(\Omega_\kappa)}\mathds{1}_{(0,\infty)}\left( m_{i;\ell,\Omega_\kappa} \right)$$ where $\mathds{1}$ stands for the indicator function. Then, under the null hypothesis, $\xi_{\ell,\Omega_\kappa} \sim \mathcal{B}\text{\textit{in}}(\text{card}(\Omega_\kappa),0.5)$ and we can interpreted the statistic of test such that the vine copula pre-stroke is statistically equivalent to the one post stroke if $\xi$ is not statistically different from $\mathbb{P}(m_{i;\ell,\Omega_\kappa} \leq 0) \times \text{card}(\Omega_\kappa) = 0.5\text{card}(\Omega_\kappa)$ for a given significance level. This test is known as Clarke's test \cite{clarke2007} and has been considered in most of the literature comparing two vine structures (see \citet{joe2011}).
We performed our version of that test on the gamma band for the whole LFP data set, as it is the one where we can visually observe on some channels aspects of change and on some other channels aspect of stability. We note that, in order to reduce noise, we truncated gamma band to $300$ $Hz$ such that $\gamma \in \{ 30, 300 \} Hz$. Figure \ref{ratbrainPrePost} shows, for rat id $141020$, in the way the electrodes are placed in the rat's brain the p-value obtained for each channel. In Appendix \ref{appendixPrePost}, one observes the statistics of test we obtained for each channel (second row) as well as these p-values for each one of the four rats. Microelectrodes in black suggest to accept $H_0$ for a significance level of $2.5\%$. Thus, under that significance level, we can say that for $\gamma$-band, there are no strong evidences of a change in the brain activity of the rat after the induced stroke for channels $1,2,13,14, 15,16 \text{ and }23$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.28]{RatBrain7.jpg}
\caption{Schema showing, for rat id $141020$, the p-values found on the $32$ channels for Gamma band.}
\label{ratbrainPrePost}
\end{center}
\end{figure}
\subsection{Empirical thresholds for Kolmogorov-Smirnov statistics}\label{empiricalSetting}
The goal of this subsection is to determine empirical threshold(s) for the bivariate Kolmogorov-Smirnov statistics, which will be used in order to test for a change in the auto-correlation between succeeding epochs. The determination of a theoretical threshold under the conditions on data used in this paper is a work in progress. Thus, as it is an explanatory work where we want to illustrate the potential of our methodology, we establish from some simulations these thresholds through two main scenarios of data generating processes (DGPs).
The overall idea in all these DGPs is to simulate two or more time series (with moderate noises) in a given DGP, from a latent signal derived from an autoregressive process. The reason justifying to simulate many different series in each simulation is to explore the effect of various latent signals with our copula-based algorithm.
We mention that the hypothesis we are considering to establish a significant threshold are the equivalence of $C^{(r-1,r)}_{\ell, \Omega_\kappa} \equiv C^{(r,r+1)}_{\ell, \Omega_\kappa}$, $r=2,...,R-1$ under the null hypothesis against the hypothesis of non-equivalence under the alternative one. It can be rewritten as:
$$
\begin{cases}
H_0: \left| C^{(r-1,r)}_{\ell, \Omega_\kappa}(u,v) - C^{(r,r+1)}_{\ell, \Omega_\kappa}(u,v) \right| = 0
& \forall (u,v) \in [0,1] \times [0,1];
\\[4pt]
H_1: \left| C^{(r-1,r)}_{\ell, \Omega_\kappa}(u,v) - C^{(r,r+1)}_{\ell, \Omega_\kappa}(u,v) \right| > 0
& \text{for some } (u,v) \in [0,1] \times [0,1].
\end{cases}
$$
Thus, setting up an experimental-based threshold that provides a critical value to test these hypotheses at a significance level $\bar{\alpha}$ (in order to avoid confusion with $\alpha$, a frequency band) is our challenge here. We remark that we fixed our risk of type 1 error to $\bar{\alpha}=1\%$.
\subsubsection{Deriving the empirical thresholds under the null hypothesis}
\paragraph{DGP 1}
In this DGP, we simulated two different scenarios where, for each scenario, we simulated $R=100$ epochs with $T=1000$ timepoints per epoch. The first scenario follows a stationary $AR(1)$, then the second one follows a similar $AR(1)$ process where we added a constant. The simulations setting is, for $t=1,...,T$:
\begin{itemize}
\item $Z_{t,A}^{(r)}= 0.9X_t^{(r)}+\epsilon_t^{(r)}$ where $X_t^{(r)} \sim \text{AR}(1) \text{ of parameter }\phi=0.9$, $\epsilon_t^{(r)} \sim \mathcal{N}(0,0.1)$ for $r=1,...,R$
\item $Z_{t,B}^{(r)}= 1+0.9X_t^{(r)}+\epsilon_t^{(r)}$ where $X_t^{(r)} \sim \text{AR}(1) \text{ of parameter }\phi=0.9$, $\epsilon_t^{(r)} \sim \mathcal{N}(0,0.1)$ for $r=1,...,R$
\end{itemize}
We computed in both cases the bivariate Kolmogorov-Smirnov statistics, $D(u,v)$, between each consecutive pairs of copulas $C^{(r-1,r)}$ and $C^{(r,r+1)}$, $r=2,...,199$. These statistics are plotted for each of the three frequency bands on Figure \ref{KSdgp1}. We remark that this DGP is considered being a basic simulation model. The goal here is to establish an empirical distribution of the Kolmogorov-Smirnov statistic under the null hypotheses and to identify the $99-th$ percentile which will serve as the threshold that satisfies $\mathbb{P}$(Type I error)$ = \bar{\alpha} = 0.01$.
\begin{table}[h]
\begin{center}
\begin{tabular}{|l||*{1}{c|}}
\hline
\textbf{Frequency band} & Threshold \\
& for $\bar{\alpha}=1\%$ \\
\hline \hline
Delta band $(\Delta)$ & $D(u,v) > 0.0102$ \\ \hline
Theta band $(\theta)$ & $D(u,v) > 0.0451$ \\ \hline
Alpha band $(\alpha)$ & $D(u,v) > 0.0090$ \\ \hline
Beta band $(\beta)$ & $D(u,v) > 0.0038$ \\ \hline
Gamma band $(\gamma)$ & $D(u,v) > 0.0048$ \\ \hline
\end{tabular}
\caption{DGP 1: threshold on the Kolmogorov-Smirnov statistics for a significance levels of $\bar{\alpha}=1\%$. }
\label{tableDGP1}
\end{center}
\end{table}
\paragraph{DGP 2}
The second DGP is based on some AR(2) processes. There main idea for this DGP is to analyze time series generated from multiple latent signals, where the time series used is tributary of the frequency band on which is performed the analysis. We notice the stationarity here across epochs (i.e., dependence between successive epochs does not change).
The principle is that latent signals from six AR(2) processes are observed for $100$ epochs. Thus, the six latent signals are: $X_{t,i} \sim AR(2), i=1,...,6$ with polynomial function whose roots are complex-valued with respectively, for each latent signal, phases $p_1= \pm 4/T \cdot 2\pi$, $p_2=\pm 6/T \cdot 2\pi$, $p_3= \pm 9/T \cdot 2\pi$, $p_4= \pm 13/T \cdot 2\pi$, $p_5=\pm 15/T \cdot 2\pi$, and $p_6= \pm 150/T \cdot 2\pi$; for $t=1,...,1000$. Thus, the spectra of these latent signals are concentrated on the phases of each one of the bands of interest. Our simulation setup, for $t=1,...,T$ and for $r=1,...,100$, is:
\begin{itemize}
\item $Z_{t,i}^{(r)}=X_{t,i}^{(r)}+\epsilon_t^{(r)},$ with noise $\epsilon_t^{(r)} \sim \mathcal{N}(0, 0.1 \sigma_{X_{t,i}^{(r)}}); \text{ for }i=1,...,6.$
\end{itemize}
Table \ref{tableDGP2a} presents the thresholds obtained for a significance value of $\bar{\alpha}=1\%$. Hence, for Delta band for example, based on these simulations, assuming the null hypothesis true, the epochs related any Kolmogorov Smirnov statistic valued greater than $0.0149$ will be considered as a changepoint in the dependence structure.
\begin{table}[h]
\begin{center}
\begin{tabular}{|l||*{1}{c|}}
\hline
\textbf{Frequency band} & Threshold for $\bar{\alpha}=1\%$ \\
\hline \hline
Delta band $(\Delta)$ & $D(u,v) > 0.0149$ \\ \hline
Theta band $(\theta)$ & $D(u,v) > 0.0625$ \\ \hline
Alpha band $(\alpha)$ & $D(u,v) > 0.0101$ \\ \hline
Beta band $(\beta)$ & $D(u,v) > 0.0050$ \\ \hline
Gamma band $(\gamma)$ & $D(u,v) > 0.0103$ \\ \hline
\end{tabular}
\caption{DGP 2: Threshold on the Kolmogorov-Smirnov statistics for a significance value of $\bar{\alpha}=1\%$.}
\label{tableDGP2a}
\end{center}
\end{table}
\subsubsection{Empirical threshold}
We conclude that each frequency band has its own threshold. We note that these critical values are not based on a theoretical development but they are empirically based on an explanatory work. Thus, they are tributary to the way we infer the copulas in our code as well as the way that we compute Kolmogorov-Smirnov statistics over a bidimensional grid of evaluation points. However, as our methodology and our code remain the same to analyze LFP data, these threshold are a reliable way to determine changepoint(s) in the rats brain activity. Hence, as all the thresholds determined in DGP 2 are more conservative than the one in DGP 1, we will consider the latter (see table \ref{tableDGP2a}) in our local field potential of a rat study. \
\subsubsection{Illustration of the power of the test, under $H_1$}
To assess the power of the test (under the alternative hypothesis), we decided to retake both DGPs from the last section and to combine the simulations settings in the same scenario. For example, for DGP $1$, the simulation setting now becomes a scenario of $200$ epochs such that we observe $Z_{t,A}^{(r)}$ for $r=1,...,100$ and then $Z_{t,B}^{(r)}$ for $r=101,...,200$. Thus, we observe two consecutive stationary series where the dependence between epochs does not change in the first half ($r=1,...,100$), the one between epochs does not change too in the second half ($r=101,...,200$), and a changepoint is expected between both stationary blocks.
For each DGP, we collect all the Kolmogorov-Smirnov values and verify that the known changepoint(s) (location between two consecutive series) are above the threshold. For example, as the series $Z_{t,A}^{(r)}$ for epochs $r=1,...100$ and the series $Z_{t,B}^{(r)}$ for epochs $r=101,...,200$ are different, we will expect to detect a changepoint; which means that $|C^{(99,100)}-C^{(100,101)}|\geq \bar{\alpha}$ as well as $|C^{(100,101)}-C^{(101,102)}|\geq \bar{\alpha}$ for a significance level $\bar{\alpha}$ determined empirically; which leads to $2$ values above the thresholds.
We present, in appendix \ref{appendixThresh}, for three frequency bands, the Kolmogorov-Smirnov statistics for DGP $1$ when, for a scenario of $200$ epochs, for $t=1,...,1000$, the setup becomes $Z_{t,A}^{(r)}$ for $r=1,...,100$ and $Z_{t,B}^{(r)}$ for $r=101,...,200$. We remark on Figure \ref{KSdgp1} that the Kolmogorov-Smirnov statistics related to epochs $99$ and $101$ are above the threshold line.
Concerning DGP $2$, we decided to combine three of the six series: $Z_{t,2}^{(r)} \text{ for }r=1,...,100, Z_{t,5}^{(r)} \text{ for }r=101,...,200 \text{ and }Z_{t,6}^{(r)} \text{ for }r=201,...,300$. We show the results in Figure \ref{DGP2powerTheta} for theta band.
\begin{figure}
\begin{center}
\includegraphics[scale=0.55 ]{DGP2powerTheta.jpeg}
\caption{DGP 2: Illustration of the power of the test, under the alternative hypothesis, for three observed series and for theta band. \textbf{Left:} Plot of the magnitude of Fourier coefficients (averaged per epoch) for the three consecutive series. \textbf{Right:} Kolmogorov-Smirnov statistics obtained and the empirically derived threshold for $\bar{\alpha}=1\%.$}
\label{DGP2powerTheta}
\end{center}
\end{figure}
\section{Third application: Comparing the dependence behavior of two different channels for a given frequency band}\label{twoChan}
This section is in fact a brief note to show that one can apply the methodology from Section \ref{prepost} to compare, based on the dependence structure, if two different microelectrodes (brain channels), for a given frequency band, act similarly during all the regime of the experiment (i.e., during the $600$ epochs). Thus, we test exactly the same hypothesis but this time, $m_{i;\ell,\ell';\Omega_\kappa}$ is defined differently. For epochs $1$ to $600$, this log-likelihood ratio is defined by:
$$
m_{i;\ell,\ell';\Omega_\kappa}=\text{log }\left\{ \frac{c^{(1:600)}_{\ell,\Omega_\kappa}({\bf{u}}_i|\theta^{(1:600)}_{\ell,\Omega_\kappa})}{c^{(1:600)}_{\ell',\Omega_\kappa}({\bf{u}}_i|\theta^{(1:600)}_{\ell',\Omega_\kappa})} \right\}
$$
where $ i=1,...,\text{card}(\Omega_\kappa)$, and $\ell$ and $\ell'$ are obviously two different channels. We applied this test to the channels defined on the two first columns of microelectrodes in the rat brain (i.e., channels linked to microelectrodes $1,2,9,10,17,18,25$ and $26$; which means a total of $28$ possible combinations). We show the results for the gamma band, for the four experimental rats, in Appendix \ref{appendixPrePost} (Table \ref{changeChannels}). We remind that to not reject $H_0$, the Clarke's statistic of test must not be significantly different to $\text{card}(\Omega_\kappa)/2$. As the results are for gamma band, they should not be significantly different to $(300\text{Hz}-31{Hz})/2=135$ under $H_0$. That said, even if the Clarke's statistics are valued on a wide range from $0$ to $270$, one observes that these $8$ channels are considered being completely different on their whole regime for a significance level: the p-values are always lower than $0.0001$ for all the $28$ possible combinations.
|
\section*{Supplementary Material}
The supplementary video, code, trained models, appendix and a preprocessed dataset will be made available at \url{http://prg.cs.umd.edu/ECN.html}.
\section{INTRODUCTION}
With recent advances in the field of autonomous driving, autonomous platforms are no longer restricted to research laboratories and testing facilities - they are designed to operate in an open world, where reliability and safety are key factors. Modern self-driving cars are often fitted with a sophisticated sensor rig, featuring a number of LiDARs, cameras and radars, but even those undoubtedly expensive setups are prone to misperform in difficult conditions - snow, fog, rain or at night.
Recently, there has been much progress in imaging sensor technology, offering alternative solutions to scene perception. A neuromorphic imaging device, called Dynamic Vision Sensor (DVS)~\cite{lichtsteiner_latency_2008}, inspired by the transient pathway of mammalian vision, can offer exciting alternatives for visual motion perception. The DVS does not record image frames, but instead - the changes of lighting occurring independently at every camera pixel. Each of these changes is transmitted asynchronously and is called an \textit{event}.
By its design, this sensor accommodates a large dynamic range, provides high temporal resolution and low latency -- ideal properties for applications where high quality motion estimation and tolerance towards challenging lighting conditions are desirable.
The DVS offers new opportunities for robust visual perception so much needed in autonomous robotics, but challenges associated with the sensor output, such as high noise, relatively low spatial resolution and sparsity, ask for different visual processing approaches.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\columnwidth]{img/teaser.png}
\end{center}
\vspace{-0.5\baselineskip}
\caption{\textit{\small{Optical flow and depth inference on sparse event data in night scene: event camera output (left), ground truth (middle column), network output (right) (top row - flow, bottom row - depth). The event data is overlaid on the ground truth and inference images in blue. Note, how our network is able to `fill in' the sparse regions and reconstruct the car on the right.}}}
\vspace{-1.0\baselineskip}
\label{fig:teaser}
\end{figure}
In this work we introduce a novel lightweight encoding-decoding neural network architecture - the Evenly-Cascaded convolutional Network (ECN) to address the problems of event data sparsity for depth, optical flow and egomotion estimation in a self driving car setting. Despite having just 150k parameters, our network is able to generalize well to
different types of sequences.
The simple nature of our pipeline allows it to run at more than 250 inferences per second on a single NVIDIA 1080 Ti GPU. We perform ablation studies using the SfMlearner architecture \cite{zhou2017unsupervised} as a baseline and evaluate different normalization techniques (including our novel \textit{feature decorrelation}) to show that our model is well suited for event data.
We demonstrate superior generalization to low-light scenes. Fig. \ref{fig:teaser} shows an example featuring night driving - the network trained on a day light scene was able to predict both depth and flow even with a low event rate and abundance of noise. This is facilitated by our event-image representation: instead of the latest event timestamps, we use the average timestamp of the events generated at a given pixel. The averaging helps to reduce the noise without losing the timestamp information. Moreover, we use multiple slices as input to our model to better preserve the 3D structure of the event cloud and more robustly estimate egomotion. The main contributions of our work can be summarized as:
\begin{itemize}
\item The first unsupervised learning-based approach to structure from motion recovery using monocular DVS input.
\item Demonstrating that dense, meaningful scene and motion information can be reliably recovered from sparse event data.
\item A new lightweight high-performance
network architecture -- $ECN$.
\item A new normalization technique, called \textit{feature decorrelation}, which significantly improves training time and inference quality.
\item Quantitative evaluation on the MVSEC dataset \cite{MVSEC} of dense and sparse depth, optical flow and egomotion.
\item A pre-processesed MVSEC \cite{MVSEC} dataset facilitating further research on event-based SfM.
\end{itemize}
\section{Related Work}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\columnwidth]{img/representation.png}
\end{center}
\vspace{-0.5\baselineskip}
\caption{\textit{A three-channel DVS data representation. The first channel represents the time image described in \cite{iROSBetterFlow}. The second and third channels represent the per-pixel positive and negative event counts. Best viewed in color.}}
\vspace{-1.0\baselineskip}
\label{fig:representation}
\end{figure}
\subsection{Event-based Depth Estimation}
The majority of event-based depth estimation methods
\cite{zhou2018semi,UpennSlices} use two or more event cameras. As our proposed approach uses only one event camera, we focus our discussion on monocular depth estimation methods. The first works on event-based monocular depth estimation were presented in \cite{BMVC2016_63} and \cite{10.1007/978-3-319-46466-4_21}. Rebecq \textit{et al}. \cite{BMVC2016_63} used a space-sweep voting mechanism and maximization strategy to estimate semi-dense depth maps where the trajectory is known. Kim \textit{et al}. \cite{10.1007/978-3-319-46466-4_21} used probabilistic filters to jointly estimate the motion of the event camera, a 3D map of the scene, and the intensity image. More recently, Gallego \textit{et al}. \cite{Gallego2018AUC} proposed a unified framework for joint estimation of depth, motion and optical flow. So far there has been no deep learning framework to predict depths from a monocular event camera.
\subsection{Event-based Optical Flow}
Previous approaches to image motion estimation used local information in event-space. The different methods adapt in smart ways one of the three principles known from frame-based vision, namely correlation \cite{delbruck_frame_2008,liu2017block}, gradient \cite{benosman_asynchronous_2012} and local frequency estimation \cite{tschechne2014event}
The most popular approaches are gradient based - namely, to fit local planes to event clouds \cite{benosman_event_2014}
As discussed in \cite{barranco2014contour}, local event information is inherently ambiguous. To resolve the ambiguity Barranco et al. \cite{barranco2014contour} proposed to collect events over a longer time intervals and compute the motion from the trace of events at contours.
Recently, neural network approaches have shown promising results in various estimation problems without explicit feature engineering. Orchard and Etienne-Cummings \cite{6891162} used a spiking neural network
to estimate flow. Most recently, Zhu \textit{et al}. \cite{zhu2018ev} released the MVSEC dataset \cite{MVSEC} and proposed self-supervised learning algorithm to estimate optical flow. Unlike \cite{zhu2018ev}, which uses grayscale information as a supervision signal, our proposed framework uses only events and thus can work in challenging lighting conditions.
\subsection{Self-supervised Structure from Motion}
The unsupervised learning framework for 3D scene understanding has recently gained popularity in frame-based vision research. Zhou et. al~\cite{zhou2017unsupervised} pioneered this line of work. The authors followed a traditional geometric modeling approach and built two neural networks, one for learning pose from single image frames, and one for pose from consecutive frames, which were self-supervised by aligning the frames via the flow. Follow-up works~
(e.g. \cite{ DBLP:journals/corr/abs-1802-05522}
have used similar formulations with better loss functions and networks, and recently
\cite{UpennSlices} proposed SfM learning from stereo DVS data.
\section{Methods}
\subsection{Ego-motion Model}
We assume that the camera is moving with rigid motion with translational velocity $v = (v_x,v_y,v_z)^T$ and rotational velocity $\omega = (\omega_x, \omega_y,\omega_z)$, and that the camera intrinsic parameters are provided. Let $\mathbf{X}=(X,Y,Z)^T$ be the world coordinates of a point, and $\mathbf{x}= (x,y)^T$ be the corresponding pixel coordinates in the calibrated camera. Under the assumption of rigid motion, the image velocity $\mathbf{u} = (u,v)^T$ at $(x,y)^T$ is given as:
\begin{equation}
\resizebox{.9\hsize}{!}{$
\begin{pmatrix}
u\\v
\end{pmatrix}\\
=\frac{1}{Z}
\begin{pmatrix}
-1 &0 &x\\
0 &-1 &y
\end{pmatrix}
\begin{pmatrix}
v_x\\
v_y\\
v_z
\end{pmatrix}+
\begin{pmatrix}
xy &-1-x^2 &y\\
1+y^2 &-xy &-x
\end{pmatrix}
\begin{pmatrix}
\omega_x\\
\omega_y\\
\omega_z
\end{pmatrix}\\
= A \mathbf{p}
$}
\label{uv2}
\end{equation}
In words, given the inverse depth and the ego-motion velocities $\mathbf{v},\omega$, we can calculate the optical flow or pixel velocity at a point using a simple matrix multiplication (Equation~\ref{uv2}) Here $\mathbf{p}$ is used to denote the pose vector $(\mathbf{v},\omega)^T$, and $A$ is a $2\times6$ matrix. Due to scaling ambiguity in this formulation, depth $Z$ and translation $(v_x, v_y, v_z)$ are computed up to a scaling factor.
\subsection {Input Data Representation}
The raw data from the DVS sensor is a stream of events, which we treat as data of 3 dimensions. Each event encodes the pixel coordinate $(x, y)$ and the \textit{timestamp} $t$.
In addition, it also carries information about its \textit{polarity} - a binary value that disambiguates events generated on rising light intensity (positive polarity) and events generated on falling light intensity (negative polarity).
The 3D $(x, y, t)$ event cloud (within a small time slice), called event slice, is projected onto a plane and converted to a 3-channel image. An example of such image can be seen in Fig. \ref{fig:representation}. The two channels of the image are the per-pixel counts of positive and negative events. The third channel is the \textit{time image} as described in \cite{iROSBetterFlow} - each pixel consists of the average timestamp of the events generated on this pixel,
because the averaging of timestamps provides better noise tolerance.
The neural network input consists of up to 5 such consecutive slice images to better preserve the timestamp information of the event cloud.
\subsection {The Pipeline}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.95\columnwidth]{img/pipeline.png}
\end{center}
\vspace{-0.5\baselineskip}
\caption{\textit{\small{The depth network (left) with an encoder-decoder architecture is used to estimate scene depth. The pose network (right) takes consecutive frames to estimate the translational velocity and rotational velocity with respect to the middle frame. Given the poses of neighboring frames and the depth of the middle frame, we calculate the optical flow. The neighboring frames are inversely warped to the middle frame and the warping difference provides the supervision loss.}}}
\vspace{-1.0\baselineskip}
\label{fig:pipeline}
\end{figure}
The global network structure is similar to the one proposed in \cite{zhou2017unsupervised}. It consists of a depth prediction component and a parallel pose prediction component, which both feed into a optical flow component to warp successive event slices. The loss from the warping error is backpropagated to train flow, inverse depth, and pose.
Our network components, instead of the standard CNNs, are based on our ECN network structure.
An (ECN based) encoding-decoding architecture is used to estimate scaled inverse depth $\frac{1}{Z}$ from a single slice of events.
To address the data sparsity, we use bilinear interpolation, which propagates local information and fills in the gaps between events. A second network, which takes consecutive slices of signals, is used to derive $v$ and $\omega$.
Then, using the rigid motion and inverse depth to predict the optical flow, neighboring slices at neighboring time instances $T+1, T+2$ and $T-1, T-2$ are warped to the slice at $T$ (Fig.~\ref{fig:pipeline}). The $l^1$ loss is used to measure the difference between the warped events and the middle slice as the supervision signal.
\subsection {Evenly Cascaded Network Architecture}
We use an encoder network
to estimate pose from consecutive frames and an encoder-decoder network~\cite{DBLP:conf/miccai/RonnebergerFB15} (Fig.~\ref{fig:ecn}) to estimate the scaled depth. Next, we describe its main novelties.
\iffalse
\begin{figure}
\centering
\includegraphics[width=.8\columnwidth]{img/ecn_encoder.png}
\caption{\textit{\small{The evenly-cascaded encoder structure. Each encoding layer has two streams of feature maps. The feedback signal is added to the current features themselves to improve them (the improving times is braced in the superscript). A set of high-level features are generated and concatenated with existing feature channels. We make pose predictions using all levels of features.}}}
\vspace{-1.0\baselineskip}
\label{fig:ecn_encoder}
\end{figure}
\fi
\begin{figure}
\centering
\includegraphics[width=1.\columnwidth]{img/ecn.png}
\caption{\textit{\small{The encoder-decoder structure. Each encoding layer has two streams of feature maps. The feedback signal is added to the downsampled current features themselves to improve them (with improving times braced in the superscript). A set of high-level features are generated and concatenated with existing feature channels. The decoding stage performs reversely as the encoding stage, in which the feature maps expand in size. The highest-level features and the corresponding encoding layer (via skip link) are merged back into the lower-level features as modulation. We make predictions of depth at multiple scales using different features. A coarse, backbone depth is predicted then upscaled and refined using modulated lower-level features.}}}
\vspace{-1.0\baselineskip}
\label{fig:ecn}
\end{figure}
First, our transform of features takes biological inspiration from multi-stage information distillation, and incorporates feedback.
Our encoding layers~\cite{DBLP:conf/bigdataconf/YeDMFA18} split the (layer) input into two streams of features (Fig.~\ref{fig:ecn}): one incorporates `feedback' via residual learning~\cite{DBLP:journals/corr/HeZRS15} (convolution outputs are added to \textit{downsampled} existing features as modulation signals); the other directly generates a set of higher level features (convolution outputs are directly utilized). At the end of the encoding stage, the network has a multi-scale feature representation. This representation is used in our pose prediction.
In each decoding layer, the highest level features, together with the corresponding features in the encoding layers are convolved and added back to the \textit{upsampled} lower level features as modulation. At the end of the decoding stage, the network acquires a set of modulated high resolution low-level features.
Our evenly-cascaded (EC) structure facilitates training by providing easily-trainable shortcuts in the architecture. The mapping $f$ from network input to output is decomposed into a series of progressively deeper, and therefore harder to train functions: $f=f_1+f_2+...+f_N$. The leftmost pathway (the green blocks) in Fig.~\ref{fig:ecn} contains the easiest to train, lowest-level features. A backbone pathway remains \textit{\textbf{unblocked}} by only going through downsampling and upsampling, promises the final gradient can be propagated to the first network layer. This construction alleviates the vanishing gradient problem in neural network training. More difficult-to-learn, modulation signals are added to this backbone pathway, allowing the network to selectively enhance and utilize multiple levels of features.
Second,
to tackle the challenges raised by sparse event data and evenly resize the features, we use bilinear interpolation. In the encoding layers, our network evenly downscales the feature maps by a scaling factor of $(s < 1)$ to get coarser and coarser features until the feature size falls below a predefined threshold. In the decoding layers, the feature maps are reversely upscaled by a scaling factor of $1/s$. The network construction is automatic and is controlled by the scaling factor. Bilinear interpolation propagates the sparse data spatially, facilitating dense prediction of depth and optical flow.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.475\textwidth]{img/big_pic.png}
\end{center}
\vspace{-0.5\baselineskip}
\caption{\textit{\small{Qualitative results from our evaluation. The table entries from left to right: DVS input, ground truth optical flow, network output for flow, ground truth for depth, network output for depth. The event counts are overlaid in blue for better visualization. Examples were collected from sequences of the MVSEC \cite{MVSEC} dataset: (top to bottom) outdoor day 1, outdoor day 1, indoor flying 1, indoor flying 2, outdoor night 1, outdoor night 2, outdoor night 3. It can be seen that on the `night' sequences the ground truth is occasionally missing due to Lidar limitations but the pipeline performs well. Best viewed in color.}}}
\vspace{-1.0\baselineskip}
\label{fig:big_pic}
\end{figure}
\subsection {Depth Predictions}
\label{sec:depth_prediction}
In the decoding stage, we make predictions from features at different resolutions and levels (Fig.~\ref{fig:ecn}). Initially, both high and low-level coarse features are used to predict a backbone depth map. The depth map is then upsampled with bilinear interpolation for refinement. Then the enhanced lower level features are used to estimate the prediction residue, which are usually also low-level structures. The residue is added to the backbone estimation to refine it. The final prediction map is therefore obtained through successive upsamplings and refinements. We also apply a smoothness constraint on the depth prediction.
Our pipeline is based on monocular vision and predicts depth up to a scale.
In real world driving scenes, the mean depth value stays relatively stable. Taking advantage of this observation, we use batch normalization before making the depth prediction, and this way the predicted depths have similar range. Optionally, we also apply depth normalization~\cite{DBLP:journals/corr/abs-1712-00175} to strictly enforce that the mean depth remains a constant.
\subsection {Feature Decorrelation}
Gradient descent training of neural networks can be challenging if the features are not properly conditioned, and features are correlated.
Researchers have proposed normalization strategies~\cite{DBLP:journals/corr/IoffeS15,DBLP:journals/corr/abs-1803-08494} to account for the scale inconsistency problem. We proceed one step further with a decorrelation algorithm to combat the feature collinearity problem.
We apply Denman-Beavers iterations~\cite{Denman:1976:MSF:2612816.2613024} to decorrelate the feature channels in a simple and forward fashion. Given a symmetric positive definite covariance matrix $C$, Denman-Beavers iterations start with initial values $Y_0=C$, $Z_0=I$. Then we iteratively compute: $Y_{k+1}=\frac{1}{2}Y_{k}(3I-Z_{k}Y_{k}),Z_{k+1}=\frac{1}{2}(3I-Z_{k}Y_{k})Z_{k}$. We then have $Z_{k}\xrightarrow{} C^{-\frac{1}{2}}$~\cite{DBLP:journals/corr/LinM17}. In our implementation, we evenly divide the features into 16 groups as proposed in group normalization~\cite{DBLP:journals/corr/abs-1803-08494}, and reduce the correlation between the groups by performing a few (1-10) Denman-Beavers iterations. We notice that a few iterations lead to significantly faster convergence and better results.
\renewcommand{\arraystretch}{1.3}
\begin{table*}[t]
\caption{\small{Evaluation of the optical flow pipeline}}
\vspace{-0.5\baselineskip}
\begin{center}
\resizebox{1.0\textwidth}{!}{\begin{tabular}{@{\extracolsep{0pt}}lcccccccccc}
\hline
& \multicolumn{2}{l}{outdoor day 1} & \multicolumn{2}{l}{outdoor night 1} & \multicolumn{2}{l}{outdoor night 2} & \multicolumn{2}{l}{outdoor night 3} & \multicolumn{2}{l}{indoor flying} \\
\cline{2-3}\cline{4-5}\cline{6-7}\cline{8-9}\cline{10-11}
& AEE & \% Outlier & AEE & \% Outlier & AEE & \% Outlier & AEE & \% Outlier & AEE & \% Outlier \\
\hline\hline
\textbf{$ECN$} & 0.35 & 0.04 & \textbf{0.49} & \textbf{0.82} & \textbf{0.43} & \textbf{0.79} & \textbf{0.48} & \textbf{0.80} & 0.21 & \textbf{0.01} \\
\textbf{$ECN_{masked}$} & \textbf{0.30} & 0.02 & 0.53 & 1.1 & 0.49 & 0.98 & 0.49 & 1.1 & \textbf{0.20} & \textbf{0.01} \\
\textbf{$ECN_{erate}$} & 0.28 & 0.02 & 0.46 & 0.67 & 0.40 & 0.53 & 0.43 & 0.67 & 0.20 & 0.01 \\
\textbf{$Zhu18$ \cite{UpennSlices}} & 0.32 & \textbf{0.0} & - & - & - & - & - & - & 0.84 & 2.50 \\
\textbf{$EV{\mhyphen}FlowNet_{best}$ \cite{zhu2018ev}} & 0.49 & 0.20 & - & - & - & - & - & - & 1.45 & 10.26 \\
\textbf{$SfMlearner$} & 0.58 & 0.89 & 0.59 & 1.01 & 0.78 & 1.32 & 0.59 & 1.38 & 0.55 & 0.29 \\
\hline
\end{tabular}}
\vspace{-2.0\baselineskip}
\end{center}
\label{table:main_table}
\end{table*}
\renewcommand{\arraystretch}{1}
\section{Experimental Evaluation}
Our our self-supervised learning framework can infer both dense optical flow and depth from sparse event data. We evaluate our work on the MVSEC \cite{MVSEC} event camera dataset which, given a ground truth frequency of 20 Hz, contains over 40000 ground truth images.
The MVSEC dataset, inspired by KITTI \cite{KITTI1},
features 5 sequences of a car on the street (2 during the day and 3 during the night), as well as 4 short indoor sequences shot from a flying quadrotor. MVSEC was shot in a variety of lighting conditions and features low-light and high dynamic range frames which are often challenging for an analysis with classical cameras.
\begin{figure}[b]
\begin{center}
\includegraphics[width=1.0\columnwidth]{img/norm_plot.png}
\end{center}
\vspace{-1.0\baselineskip}
\caption{\textit{\small{Comparison of Abs Rel Errors using different normalization methods on outdoor day 1 sequence (less is better).}}}
\vspace{-0.5\baselineskip}
\label{fig:normalizations}
\end{figure}
\subsection{Implementation Details}
Our standard network architecture has scaling rates of $0.5$ and $2.0$ respectively for the encoding and decoding layers, and results in $5$ encoding/decoding layers. Our depth network has $8$ initial hidden channels and expands with a growth rate of $8$. We halve these settings to $4$ for our pose network. The pose network takes $5$ consecutive event slices and predicts $6d$ pose vectors.
We use $3\times 3$ convolutions, and the combined network has $150k$ parameters. We train the network with a batch size of $32$ and use the Adam optimizer with a learning rate of $0.01$. Interestingly, we notice that compared to the standard architecture of the SfMlearner, the learning rate is higher. Thus, the new design allows us to learn at a faster rate. The learning rate is annealed using cosine scheduling, and we let the training run for $30$ epochs. Our training takes $7$-minutes for each epoch using a single Nvidia GTX 1080Ti GPU. Our model using batch normalization runs at 250 FPS at inference as it has been heavily GPU optimized. The model using feature decorrelation runs at 40 FPS. The slow down is mainly due to matrix multiplications in our customized layer which are less optimized for the GPU.
In the experiments with the outdoor sequences, we trained the network using only the \textit{outdoor day 2} sequence with the hood of the car cropped. Our experiments demonstrate that our training generalizes well to the notably different \textit{outdoor day 1} sequence, as well as to the \textit{night} sequences. For the \textit{indoor} sequences, since they were too short to create a representative training set, we used $80\%$ of each sequence for training and evaluated on the remaining $20\%$. We would like to note that the \textit{outdoor night} sequences have occasional errors in the ground truth (see for example Fig. \ref{fig:big_pic} last three rows, or Fig. \ref{fig:failure_1}). All incorrect frames had to be manually removed for the evaluation. In our experiments, we use fixed-width time slices of $1/40$-th of a second.
\subsection{Ablation Studies}
\subsubsection{Testing on the \textit{SfMlearner}}
As baseline we use the state-of-the-art \textit{SfMlearner} \cite{zhou2017unsupervised} on our data (event images). \textit{SfMlearner} has a fixed structure of 7 encoding and 7 decoding layers. It has 32 initial hidden channels and expands to 512 channels. Overall the model contains 33M parameters. SfMlearner is trained using Adam optimizer with a learning rate of $2e-4$ and a batch size of $4$. We replace the inputs with our event slices, and we include the evaluation results for flow and egomotion in tables \ref{table:main_table} and \ref{table:main_table_egomotion}.
\subsubsection{Normalization Methods}
We compare two normalization methods and our decorrelation method on the validation set portion of the outdoor day 2 sequence. We apply $5$ Denman-Beavers iterations in the decorrelation procedure. Compared with normalization methods, decorrelation leads to more thorough data whitening, and we have noticed this layer-wise whitening lead to faster convergence and lower evaluation loss (Fig.~\ref{fig:normalizations}).
\subsubsection{Visualizing a Shallow and Tiny Network}
Our lightweight multi-level, multi-resolution design allows us to construct networks of any depth and size. As a preliminary attempt, we set the scaling rate to $1/3$ and $3.0$ for encoding and decoding layers respectively, so the network has only 3 encoding/decoding layers. As the network is small, we can directly visualize all its internal feature maps. A deeper and wider network would produce a higher quality output but also more feature maps, which we do not list here. In Fig.~\ref{fig:depthnet_maps} we have listed all the feature maps in the small depth network. The row number corresponds to the level number of the features for each figure. We notice the encoder seems to play a feature extraction role in the network and the decoder starts to produce semantically meaningful representation corresponding to the desired output (depth). By scrutinizing the pose network outputs, we notice the network is intelligent enough to aggregate information corresponding to different time periods of the events in the first layer (Fig.~\ref{fig:posenet_maps}). Otherwise, mixing up the events at different time period would make the pose estimation harder.
\begin{figure}[h]
\centering
\subfigure[]{\includegraphics[width=.45\columnwidth]{viz/depth_encoder_maps_layer0.png}}
\subfigure[]{\includegraphics[width=.85\columnwidth]{viz/depth_encoder_maps_layer1.png}}
\subfigure[]{\includegraphics[width=.5\columnwidth]{viz/depth_encoder_maps_layer2.png}}\\
\subfigure[]{\includegraphics[width=.25\columnwidth]{viz/depth_encoder_maps_layer3.png}}\\
\subfigure[]{\includegraphics[width=.5\columnwidth]{viz/depth_decoder_maps_layer2.png}}
\subfigure[]{\includegraphics[width=.85\columnwidth]{viz/depth_decoder_maps_layer1.png}}
\subfigure[]{\includegraphics[width=.45\columnwidth]{viz/depth_decoder_maps_layer0.png}}\\
\subfigure[]{\includegraphics[width=.08\columnwidth]{viz/depth_predict3.png}}
\subfigure[]{\includegraphics[width=.11\columnwidth]{viz/depth_predict2.png}}
\subfigure[]{\includegraphics[width=.16\columnwidth]{viz/depth_predict1.png}}
\caption{Visualization of feature maps of the depth network. (a) Input channels. (b-g) Feature maps of the encoder-decoder network. (h-j) Multiscale predictions by layers (e-g).}
\label{fig:depthnet_maps}
\end{figure}
\begin{figure}[h]
\centering
\subfigure[]{\includegraphics[width=.95\columnwidth]{viz/pose_encoder_maps_layer0.png}}
\subfigure[]{\includegraphics[width=.495\columnwidth]{viz/pose_encoder_maps_layer1.png}}
\subfigure[]{\includegraphics[width=.22\columnwidth]{viz/pose_encoder_maps_layer2.png}}
\subfigure[]{\includegraphics[width=.11\columnwidth]{viz/pose_encoder_maps_layer3.png}}
\caption{Visualization of the pose network. (a) Input channels. (b-d) Feature maps of the pose network.}
\label{fig:posenet_maps}
\end{figure}
\subsubsection{Performance Versus Event Rate}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\columnwidth]{img/plot_aee_ecount.png}
\end{center}
\vspace{-1.0\baselineskip}
\caption{\textit{\small{The Average Endpoint Error (blue) and the number of pixels with at least one event (red) for the first 1500 frames of `outdoor_day1' sequence of the MVSEC \cite{MVSEC} dataset. Both plots are normalized so that the mean value is 0.5 for easier comparison.}}}
\vspace{-0.5\baselineskip}
\label{fig:plot_aee_ecount}
\end{figure}
Since the event data is inherently sparse, we investigate the performance of the pipeline in relation to the data sparsity.
We measure the data sparsity as a percentage of the pixels on the input images with at least one event. Fig.~\ref{fig:plot_aee_ecount} demonstrates how the event rate is inversely proportional to the average endpoint error for the optical flow (we have observed similar behavior for depth and egomotion). The \textit{outdoor day 1} sequence is used to minimize the influence of the noise.
We find the inverse correlation between event rate and inference quality to be a useful observation, as this property could be efficiently used in sensor fusion in a robotic system. Motivated by that, we provide an additional row to the Table \ref{table:main_table}: $ECN_{erate}$, and report our error metrics once again only for the frames with \textit{higher than average} number of event pixels across the datasets.
\subsection{Qualitative Results}
In addition to the quantitative evaluation, we present a number of samples for qualitative analysis in Fig. \ref{fig:big_pic}. The last three rows of the table show the night sequences, and how the pipeline performs well even when only a few events are available. The third and the fourth rows show indoor scenes. The indoor sequences were relatively small and it is highly possible that the quality of the output would increase given a larger dataset.
\subsection{Optical Flow Evaluation}
We evaluate our optical flow results in terms of Average Endpoint Error ($AEE = \frac{1}{n}\sum \norm{\vec{y} - \vec{y}^*}_2$ with $y^*$ and $y$ the estimated and ground truth value, and $n$ the number of events) and compare our results against two state-of-the-art optical flow methods for event-based cameras: $EV-FlowNet$ \cite{zhu2018ev} and a recent stereo method \cite{UpennSlices} (in the tables - $Zhu18$).
Because our network produces flow and depth values for every image pixel, our evaluation is not constrained by pixels which did not trigger a DVS event. Still, for consistency reasons, we report both numbers for each of our experiments (for example, $ECN$ and $ECN_{masked}$, where the latter has errors computed only on the pixels with at least one event). Similar to KITTI and EV-FlowNet, we report the percentage of outliers - values with error more than 3 pixels or 5\% of the flow vector magnitude.
To compare against \cite{zhu2018ev} and \cite{UpennSlices}, we account for the difference in the frame rates (for example, EV-FlowNet uses the frame rate of the DAVIS classical frames) by scaling our optical flow. We also provide aggregated results for the indoor scenes (split on a train and a test set 80/20 as described above), although these are not the main focus of our study. Our main results are presented in the Table \ref{table:main_table}.
We show that our optical flow performs well during both day and night, all on unseen sequences. The results are typically better for the experiments with event masks except for the \textit{outdoor night}. One possible explanation for that is that this sequence is much noisier with events being generated not only on the edges, which leads to suboptimal masking.
\subsection{Depth Evaluation}
Since there are currently no monocular event-based methods for the depth estimation based on unsupervised learning, we provide the classical scale-invariant depth metrics, used in many works such as \cite{DEPTH_COOLGUY}, \cite{zhou2017unsupervised}, \cite{MetricUser}:
\begin{equation}
Accuracy: \% of y_i\ s.t.\ max(\frac{y_i}{y_i^*}, \frac{y_i^*}{y_i}) = \delta < th
\label{eq:delta}
\vspace{-0.5\baselineskip}
\end{equation}
\begin{equation}
SILog: \frac{1}{n}\sum d_i^2 - \frac{1}{n^2}(\sum d_i)^2, d_i = \log y_i - \log y_i^*
\label{eq:SILog}
\vspace{-0.5\baselineskip}
\end{equation}
\begin{equation}
Absolute Relative Difference: \frac{1}{n}\sum \frac{|y - y^*|}{y^*}
\label{eq:AbsRelDiff}
\vspace{-0.5\baselineskip}
\end{equation}
\begin{equation}
Logarithmic RMSE: \sqrt{\frac{1}{n}\sum \norm{\log y - \log y^*}^2}
\label{eq:RMSElog}
\end{equation}
Our results are presented in Table \ref{table:main_table_depth} for both event count-masked depth values and full, dense depth. Since the night driving scenes have similar depth geometries, we aggregate all 3 sequences in one table entry.
Applying an event mask during the evaluation increases accuracy for all scenes - this is expected, as the inference is indeed more accurate on the pixels with event data. On the contrary, the error rate increases on the outdoor scenes and decreases on the indoor scenes. This is probably due to higher variation of the outdoor scenes and also faster motion of the car.
\renewcommand{\arraystretch}{1.3}
\begin{table}[]
\caption{\small{Evaluation of the depth estimation pipeline (results on masked, sparce depth are in braces)}}
\begin{tabular}{cllll}
\hline
\multicolumn{1}{l}{} & & outdoor day 1 & outdoor night & indoor flying \\ \hline \hline
\multicolumn{1}{c|}{\multirow{3}{*}{\rotatebox{90}{Error}}} & \multicolumn{1}{l|}{Abs Rel} & 0.29 (0.33) & 0.34 (0.39) & 0.28 (0.22) \\
\multicolumn{1}{c|}{} & \multicolumn{1}{l|}{RMSE log} & 0.29 (0.33) & 0.38 (0.42) & 0.29 (0.25) \\
\multicolumn{1}{c|}{} & \multicolumn{1}{l|}{SILog} & 0.12 (0.14) & 0.15 (0.18) & 0.11 (0.11) \\
\multicolumn{1}{c|}{\multirow{3}{*}{\rotatebox{90}{Accuracy}}} & \multicolumn{1}{l|}{$\delta < 1.25$} & 0.80 (0.97) & 0.67 (0.95) & 0.75 (0.98) \\
\multicolumn{1}{c|}{} & \multicolumn{1}{l|}{$\delta < 1.25^2$} & 0.91 (0.98) & 0.85 (0.98) & 0.91 (0.99) \\
\multicolumn{1}{c|}{} & \multicolumn{1}{l|}{$\delta < 1.25^3$} & 0.96 (0.99) & 0.93 (0.99) & 0.96 (1.0)
\end{tabular}
\label{table:main_table_depth}
\end{table}
\renewcommand{\arraystretch}{1}
\subsection{Egomotion Estimation}
\label{sec:egomotion}
Our pipeline is capable of inferring egomotion on both day and night sequences, and transfers well from \textit{outdoor day 2} onto \textit{outdoor day 1} and \textit{outdoor night 1,2,3}. Since our pipeline is monocular, we predict the translational component of the velocity up to a scaling factor, while the rotational velocity does not need scaling. Despite our network outputs full 6 degree of freedom velocity, we did not achieve high quality on indoor sequences. This is likely due to highly more complicated motion types and a small size of the indoor dataset. We further discuss this in sec. \ref{sec:discussion}.
For the driving scenarios we can \textit{make an important observation} - the mean distance of the road in respect to the camera is often a constant. We crop the lower middle part of the inferred depth image and apply a scaling factor such that the mean depth value (corresponding to the road location) is constant. Consequently, only a single extrinsic value (camera height on the car) is needed to reconstruct the scaled motion. In our experiments, we report egomotion with translational scales taken both from ground truth ($AEE_{tr}^{gt}$) and using the depth constancy constraint ($AEE_{tr}^{depth}$), with a global scale taken from ground truth. The qualitative results can be seen in Fig. \ref{fig:trajectories}.
Unlike \cite{UpennSlices}, we train \textit{SfMlearner} on the event images, and not on the classical frames to allow for evaluation on the night sequences. We provide comparison to the work in \cite{UpennSlices}, although it uses a stereo pipeline and reports results only on the \textit{outdoor day 1} sequence.
To be consistent with \cite{UpennSlices}, we report our trajectory estimation relative pose and relative rotation errors as $RPE = arccos(\frac{t_{pred} \cdot t_{gt}}{\|t_{pred}\|_2 \cdot \|t_{gt}\|_2})$ and $RRE = \|logm(R_{pred}^TR_{gt})\|_2$. Here $logm$ is matrix logarithm and $R$ are Euler rotation matrices. The $RPE$ essentially amounts to the angular error between translational vectors (ignoring the scale), and $RRE$ amounts to the total 3-dimensional angular rotation error. To account for translational scale errors, we report classical Endpoint Errors - computed as a magnitude of the differences in translational component of the velocities. Our quantitative results are presented in Table \ref{table:main_table_egomotion}.
\begin{figure}
\centering
\subfigure[\textit{outdoor day 1} sequence]{\includegraphics[width=.95\columnwidth]{img/tj_od1.png}} \\
\subfigure[\textit{outdoor night 2} sequence]{\includegraphics[width=.95\columnwidth]{img/tj_on2.png}}
\caption{\textit{\small{Estimated trajectories on `outdoor day 1' and `outdoor night 2' sequences, acquired by integrating the egomotion predictions. The network was trained only on `outdoor day 2'. Black: ground truth. Red: network prediction with translational scale applied from ground truth. Cyan: network prediction by assuming the mean depth is fixed throughout the sequence (see sec. \ref{sec:egomotion}) and by applying a single global scaling to the translational pose.}}}
\label{fig:trajectories}
\end{figure}
\renewcommand{\arraystretch}{1.6}
\begin{table}[h!]
\caption{Egomotion Estimation Results}
\label{table:main_table_egomotion}
\begin{tabular}{cccccc}
\hline
& & \multicolumn{1}{c}{$ARPE$} & \multicolumn{1}{c}{$ARRE$} & \multicolumn{1}{c}{$AEE_{tr}^{gt}$} & \multicolumn{1}{c}{$AEE_{tr}^{depth}$} \\ \hline
\hline
\multicolumn{1}{l|}{$ECN$} & \multicolumn{1}{l|}{\multirow{3}{*}{\rotatebox{90}{outdoor} \rotatebox{90}{ day 1}}} & 3.98 & 0.00267 & 0.70 & 1.29 \\
\multicolumn{1}{l|}{$Zhu18$ \cite{UpennSlices}} & \multicolumn{1}{l|}{} & 7.74 & 0.00867 & - & - \\
\multicolumn{1}{l|}{$SfMlearner$} & \multicolumn{1}{l|}{} & 16.99 & 0.00916 & 1.59 & 2.39 \\
\hline
\multicolumn{6}{l}{outdoor night \rule{0pt}{4ex} } \\
\hline\hline
\multicolumn{1}{l|}{$ECN$} & \multicolumn{1}{c|}{\multirow{2}{*}{1}} & 3.90 & 0.00139 & 0.42 & 1.26 \\
\multicolumn{1}{l|}{$SfMlearner$} & \multicolumn{1}{c|}{} & 9.95 & 0.00433 & 1.04 & 2.47 \\
\multicolumn{1}{l|}{$ECN$} & \multicolumn{1}{c|}{\multirow{2}{*}{2}} & 3.44 & 0.00202 & 0.80 & 1.34 \\
\multicolumn{1}{l|}{$SfMlearner$} & \multicolumn{1}{c|}{} & 13.51 & 0.00499 & 1.66 & 3.15 \\
\multicolumn{1}{l|}{$ECN$} & \multicolumn{1}{c|}{\multirow{2}{*}{3}} & 3.28 & 0.00202 & 0.49 & 1.03 \\
\multicolumn{1}{l|}{$SfMlearner$} & \multicolumn{1}{c|}{} & 1.053 & 0.00482 & 1.42 & 2.74
\end{tabular}
\caption*{\textit{\small{Table 3: Egomotion estimation results on driving sequences - `outdoor day 1' and `outdoor night 1,2,3'. The ARPE and ARRE are reported in degrees and radians respectively \cite{UpennSlices}, AEE is in m/s. $AEE_{tr}^{gt}$ - translational endpoint error with ground truth normalization. $AEE_{tr}^{depth}$ - normalized using depth prediction and a global scaling factor (see sec. \ref{sec:egomotion}).}}}
\end{table}
\renewcommand{\arraystretch}{1.0}
\subsection{Discussion and Failure Cases}
\label{sec:discussion}
A monocular pipeline tends infer more information from the shape of the contours on depth estimation and hence would transfer poorly on completely different scenarios. Nevertheless, we were able to achieve good generalization on night sequences and demonstrate promising results for depth and flow for indoor scenes (trained separately on parts of indoor sequences).
We observe a relatively small angular drift on trajectory estimation (Fig. \ref{fig:trajectories}). Despite our model predicting a full 6 degree of freedom
motion we admit that in the car scenario only 2 motion parameters play meaningful role and the network may tend to overfit. For this reason, training on the indoor scenes, featuring more complicated motion would require a notably larger dataset than presented in MVSEC. We still achieve results superior to \textit{SfMlearner} and the stereo method \cite{UpennSlices}, while for the comparison with the latter we must attribute some of our success to the fact that our translational velocity prediction is only up to scale.
A common shortcoming of event-based sensors in the lack of data when the relative motion is not present. Fig. \ref{fig:failure_1} shows such an example. This issue (although it does not affect egomotion) can be solved only by fusing data from other visual sensors or by moving the event-based sensor continuously. Because of the smoothness constraint used to combat data sparsity, the network tends to blur object boundaries. Still, for the driving environment the contours of obstacles, people and cars are clearly visible, as can be seen in Fig. \ref{fig:big_pic}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{img/failure_1.png}
\end{center}
\vspace{-1.0\baselineskip}
\caption{\textit{\small{A typical failure case and a dataset artifact: A non-moving car (visible in the middle ground truth inverse depth image) is not visible on the DAVIS camera (left image) which prevents ECN from inferring optical flow or depth correctly (right image is the inference inverse depth image). On the contrary, the moving car on the left side of the road is clearly visible in the event space and its depth inference is correct, but due to the Lidar limitations the depth ground truth is completely missing. This frame is taken from the `outdoor\_night 1' MVSEC sequence.}}}
\vspace{-0.5\baselineskip}
\label{fig:failure_1}
\end{figure}
\iffalse
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{img/fast_motion.png}
\end{center}
\vspace{-1.0\baselineskip}
\caption{\textit{\small{A case of fast motion: even though the recent events overwrite older pixels, the network is still capable of predicting accurate flow. Ground truth is in the middle, the inference image is on the right. This frame is taken from the `indoor\_flight 2' MVSEC sequence.}}}
\vspace{-0.5\baselineskip}
\label{fig:fast_motion}
\end{figure}
\fi
\section{CONCLUSION}
We have presented a novel low-parameter pipeline for generating dense optical flow, depth and egomotion from sparse event camera data.
We also have shown experimentally that our new neural network architecture using multi-level features improves upon existing work. Future work will investigate the estimation of moving objects as part of the pipeline, using event cloud representations instead of accumulated events, and the use of space-time frequency representations in the learning.
|
\section{Introduction}
\label{sec:intro}
To collect a large statistics neutrino experiments typically use massive nuclear targets,
which are particularly relevant in long-baseline oscillation experiments due to the
reduced flux at the far site~\cite{Acciarri:2015uup,Acciarri:2016ooe}. An understanding
of the structure and interactions of hadrons within the nuclear
targets is, therefore, crucial to achieve accurate measurements of neutrino interactions.
The existing uncertainties in the modeling of the nuclear effects, which modify the
neutrino cross-sections, as well as the final state interactions within the nucleus
are not adequate to meet the precisions required by next-generation neutrino experiments~\cite{Alvarez-Ruso:2017oui}.
Modern (anti)neutrino beams can deliver high intensity fluxes alleviating
one of the primary limitations of past experiments.
This progress allows the use of high resolution detectors with a relatively small fiducial mass
of a few tons to obtain a more accurate reconstruction of (anti)neutrino interactions.
However, all of these improvements are voided without a control of the
configuration, chemical composition, size, and mass of the neutrino
targets comparable to the one traditionally achieved in electron scattering experiments.
Furthermore, available measurements of various (anti)neutrino exclusive processes
and cross-sections on nuclei indicate many outstanding discrepancies among different
experiments, as well as with existing theoretical models~\cite{Alvarez-Ruso:2017oui}.
The limited resolution and acceptance of some of those measurements, and the complexity
of the weak current (compared to the electromagnetic one) and of the associated
nuclear effects in heavy nuclei are typically responsible for the observed discrepancies.
To this end, in addition to a high experimental resolution, a large acceptance,
and a complete control of the targets, a necessary condition for any program
of precision measurements with (anti)neutrino interactions is to have a
complementary control target free from nuclear effects:
hydrogen (free proton)~\footnote{Elastic scattering off electrons (mainly NC) can
also provide complementary information free from nuclear effects.
However, the limited statistics and the additional smearing associated
to the outgoing neutrino and the beam divergence
limit the physics sensitivity of this channel.}.
The only available data from $\nu(\bar \nu)$-H interactions were collected by the early
bubble chamber experiments ANL~\cite{Barish:1978pj}, BNL~\cite{Baker:1982ty},
BEBC~\cite{Jones:1991gd,Jones:1990au}, and FNAL~\cite{Bell:1978ta,Derrick:1981zb}.
In spite of the excellent experimental resolution of those measurements,
the overall statistics is limited to about 16,000 events
and is insufficient for current needs. Since then safety requirements related to the
underground operation and practical considerations favoring modern electronic detectors
have prevented new experimental measurements of high statistics $\nu(\bar \nu)$-H samples.
In this paper we propose a novel approach for precision measurements of
$\nu(\bar \nu)$-H Charged Current (CC) interactions, which is both simple and safe to implement.
Interactions on hydrogen are extracted by subtracting measurements on a dedicated
graphite (pure C) target from those on a dedicated CH$_2$ plastic target, within a
low-density high resolution detector. This concept appears to be a viable and realistic
alternative to the use of liquid H$_2$ detectors to obtain high statistics samples
of (anti)neutrino interactions on hydrogen.
This paper is organized as follows. In Sec.~\ref{sec:targets} we discuss how we can
achieve a control of (anti)neutrino targets comparable to electron scattering experiments,
as well as the basic concept of the subtraction between CH$_2$ and C targets.
In Sec.~\ref{sec:Hsele} we present a detailed kinematic selection of
$\nu(\bar \nu)$-H interactions from the CH$_2$ plastic target for various event
topologies. In Sec.~\ref{sec:disc} we discuss our results and in Sec.~\ref{sec:sum} we
summarize our findings.
\section{Control of Neutrino Targets}
\label{sec:targets}
All detectors for neutrino interactions suffer from a tension between the opposing requirements
of a large target mass versus the detector resolution necessary to achieve the physics goals.
Furthermore, in most fine-grained neutrino detectors the active detector, consisting of
several different materials and components, also provides the target mass.
A drawback of this configuration is the difficulty to precisely
control the target chemical composition and mass, limiting the ultimate precision
achievable in the measurements.
A detector technology designed to offer a solution to the outstanding issues above
is a Straw Tube Tracker (STT) in which the neutrino targets are physically
separated from the actual tracking system (of negligible mass).
To achieve high resolution measurements,
the target mass is spread out uniformly throughout the entire tracking volume,
by keeping the average density low enough
-- similar to that of liquid deuterium $\rho \sim 0.17$ g/cm$^3$ --
to have a detector transparent (size comparable to one radiation length)
to final state particles produced in neutrino interactions.
This detector concept is rather flexible and its main operating parameters can
be adapted to different detector configurations and geometries.
Examples of different implementations of the STT technology can be found in the
CDR of the DUNE experiment~\cite{Acciarri:2016ooe} and in a recently proposed
near detector enhancement for the Long-Baseline Neutrino Facility (LBNF)~\cite{ESGprop}.
The detector considered here is based upon a central STT inserted in a
0.6 T magnetic field and surrounded by a $4\pi$ electromagnetic
calorimeter (ECAL)~\cite{Adinolfi:2002zx}.
The base tracking technology is provided by
low-mass straws similar to the ones used in many modern experiments
for precision physics or the search for rare
processes~\cite{Sergi:2012twa,Anelli:2015pba,Nishiguchi:2017gei,Lee20162530,Gianotti:2013ipa}.
Thin layers -- typically 1-2\% of radiation length -- of various target
materials (100\% chemical purity) are alternated with straw layers so that they
represent more than 95\% of the total detector mass
(the mass of the straws being $<5\%$). This feature, combined
with the excellent vertex, angular, momentum, and timing
resolutions are key factors to correctly associate
neutrino interactions to each target material, as well as for an
accurate measurement of the four-momenta of the final state particles.
As a result, we can achieve a control of the configuration, chemical composition,
size, and mass of the (anti)neutrino targets comparable to electron scattering
experiments.
In this paper we will focus on two specific materials to be used as
(anti)neutrino targets within the detector described above:
polypropylene (CH$_2$) and graphite (pure C). The former is one of the
plastic materials with the highest hydrogen content and can be easily manufactured
in thin foils of arbitrary size.
Both targets have comparable thickness in terms of radiation length and
are alternated throughout the entire tracking volume to ensure that
they result in the same detector acceptances for final state particles.
We further assume that they are exposed simultaneously to the same
(anti)neutrino beam. By performing a statistical subtraction between the data collected
from the CH$_2$ plastic and the ones collected from the graphite, we can in principle
obtain (anti)neutrino interactions on a hydrogen target, i.e. on free protons.
Given the purity (100\%) of the CH$_2$ and C targets and the
accuracy in associating the interactions to each target, the
normalization of the H signal and the C background is based upon the
corresponding relative abundances in CH$_2$.
This technique can be safely implemented and can offer
a relatively large ($\sim 0.7$ ton with 5 tons of CH$_2$) target mass of hydrogen.
The main issues to address are the physics sensitivity achievable
and whether the uncertainties associated to the subtraction procedure allow
competitive $\nu(\bar \nu)$-H measurements. In the following we will present
detailed studies of the corresponding event selection with realistic assumptions
for the detector smearing and the physics modeling.
\section{Selection of \boldmath $\nu(\bar \nu)$-H Interactions}
\label{sec:Hsele}
The detection technique described in Sec.~\ref{sec:targets} allows a complete control of the
(anti)neutrino targets in both size and material(s), as well as an accurate location
of the interaction vertices within each target element.
However, the H content by weight in the CH$_2$ target is only 14.4\% of the total and
most of the interactions are still originated from the C nucleus.
We can improve the signal/background ratio in the selection of H interactions by exploiting
the event kinematics. Since the H target is at rest, CC events are expected
to be perfectly balanced in a plane transverse to the beam direction (up to the tiny beam divergence)
and the muon and hadron vectors are back-to-back in the same plane. Instead, events
from nuclear targets are affected by the smearing with the energy-momentum
distribution of bound nucleons (Fermi motion and binding), the off-shell modifications of bound nucleons, meson exchange currents and nuclear
shadowing~\cite{Alvarez-Ruso:2017oui,Kulagin:2004ie,Kulagin:2007ju,Kulagin:2014vsa},
as well as by final state interactions (FSI)~\cite{Alvarez-Ruso:2017oui}.
These nuclear effects result in a significant missing transverse momentum
and a smearing of the transverse plane kinematics.
The use of transverse plane variables and event kinematics to select various Neutral
Current (NC) and CC (anti)neutrino topologies was pioneered by the NOMAD experiment~\cite{Altegoer:1997gv}.
The analysis described in this paper is largely based upon the variables and techniques
developed and validated with NOMAD data~\cite{Astier:2001yj,Naumov:2004wa,Astier:2000ng}.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.00\textwidth]{Hsel-Rmh}
\end{center}
\caption{Comparison of the reconstructed $p_T$ asymmetry
$R_{mH} = (p_T^m - p_T^H)/(p_T^m + p_T^H)$ in
$\nu_\mu p \to \mu^- p \pi^+$ and $\bar \nu_\mu p \to \mu^+ p \pi^-$ processes
on H (signal) and in the corresponding background topologies from the C nucleus.
Results for both neutrino (left panel) and antineutrino (right panel) interactions are shown. All distributions are normalized to unit area.
}
\label{fig:Rmh}
\end{figure}
\subsection{Analysis framework}
\label{sec:framework}
We simulate (anti)neutrino interactions on CH$_2$, H, and C targets with three different event
generators: NuWro~\cite{Juszczak:2005zs}, GiBUU~\cite{Buss:2011mx}, and GENIE~\cite{Andreopoulos:2009rq}. While we use NuWro as our default generator,
we compare our results with both GiBUU and GENIE to check the sensitivity of our
analysis to the details of the input modeling. These generators are based upon
different assumptions and modeling of nuclear effects and final state interactions,
with GiBUU using a conceptually different approach from heavy ion physics,
based upon the Boltzmann-Uehling-Uhlenbeck equation for the
particle propagation through the nuclear medium. For a detailed review and
comparison of the generators used we refer to Ref.~\cite{Mosel:2019vhx}.
The studies of nuclear FSI by the MINER$\nu$A and T2K
experiments~\cite{Lu:2018stk,Abe:2018pwo} have found an unphysical excess
of $hA$ elastic scattering processes in the FSI simulated by GENIE, which is in
disagreement with (anti)neutrino data. We follow the corresponding prescriptions by
MINER$\nu$A and T2K and discard such elastic $hA$ FSI processes in GENIE.
We generate inclusive CC interactions including all
processes available in the event generators -- quasi-elastic (QE), $\Delta(1232)$
and higher resonances (RES), non-resonant processes and deep inelastic scattering (DIS) --
with defaults settings and the input (anti)neutrino spectra
expected in the Long-Baseline Neutrino Facility
(LBNF)~\cite{Petti2018,ESGprop}.
We use the GEANT4~\cite{Agostinelli:2002hh} program to evaluate
detector effects and apply to all final state particles a parameterized reconstruction
smearing consistent with the NOMAD data~\cite{Altegoer:1997gv}.
The detector smearing has also been independently checked
with the FLUKA~\cite{fluka2} simulation program in Ref.~\cite{ESGprop}.
The acceptance for individual final state particles ($p,n,\pi^\pm, \pi^0,\mu$) takes into
account the detector geometry, the event topology, and the material traversed by the
particles and is folded into the analysis.
We emphasize that the STT detector has been explicitly designed to offer the same
acceptance for particles produced in both the CH$_2$ and graphite targets, as discussed
in Sec.~\ref{sec:targets}.
On average, the momentum resolution is
about 5\% and the angular resolution is about 2 mrad for charged tracks.
The reconstruction of protons is typically worse due to the shorter track length,
with average momentum resolution of 6.5\% for H events and 8.4\% for C events.
The difference between H and C events is related to the different momentum and
angular distributions introduced by nuclear effects.
The vertex resolution can vary from 100 $\mu m$ to about
500 $\mu m$ for multi-track events, depending upon the geometry and the event topology~\cite{Altegoer:1997gv}. We note that this parameter is not critical in STT
-- even for single track events (Sec.~\ref{sec:numubarqe}) --
since the actual targets have 100\% chemical purity and represent more than 95\% of the
total detector mass (Sec.~\ref{sec:targets}). Particle identification is provided by
various methods including dE/dx, range, and transition radiation in STT, as well as
the energy deposition, timing and track segments in the surrounding ECAL and
muon identifier. For more details about the detector performance we refer to
Refs.~\cite{Acciarri:2016ooe,ESGprop,Altegoer:1997gv}.
We analyze $\nu_\mu$ and $\bar \nu_\mu$ CC interactions originated
from the CH$_2$ and graphite targets described in Sec.~\ref{sec:targets}. We determine
the momentum vectors of charged particles from the track
curvature in the B filed, while for neutral particles we use either the
energy deposited in the various sub-detectors or, whenever available, the
secondary charged tracks originated by the neutral particles in STT.
The momentum vector of the total hadron system,
$\vec{p}_H$, is obtained from the sum of the momenta of all the reconstructed
final state hadrons~\cite{Astier:2001yj}.
In the following section we discuss a unified approach to the kinematic selection
of all the event topologies available in $\nu_\mu$-H and $\bar \nu_\mu$-H CC
interactions, as well as the corresponding inclusive samples. For more details about the
analysis technique used we refer to Ref.~\cite{Astier:2001yj}.
\subsection{Kinematic analysis}
\label{sec:kine}
\subsubsection{Selection of $\nu_\mu p \to \mu^- p \pi^+$
and $\bar \nu_\mu p \to \mu^+ p \pi^-$}
\label{sec:3trk}
In order to illustrate the potential of the technique we propose for the measurement
of H interactions, we start from an analysis based upon simple kinematic cuts
and the cleanest topologies $\nu_\mu p \to \mu^- p \pi^+$ and
$\bar \nu_\mu p \to \mu^+ p \pi^-$, mainly originated from resonance production.
Detailed GEANT4 simulations indicate that the average proton reconstruction
efficiency~\footnote{To be reconstructed the proton track must have enough hits to
allow a momentum determination from the bending in B field. In principle some of
the events failing this requirement can still be reconstructed from the energy
deposition in the available hits. We do not make any attempt to recover such protons.}
is about 93\% for H events and 67\% in C events, since in this latter case nuclear
effects result, on average, in smaller kinetic energies and larger angles for the protons.
As discussed in Sec.~\ref{sec:framework}, the final state particles from these processes
on H can be accurately reconstructed in the detector described in Sec.~\ref{sec:targets},
thus resulting in an excellent measurement of all the relevant kinematic variables.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.00\textwidth]{Hsel-Hvars}
\end{center}
\caption{Some of the correlations between kinematic variables used to construct $\ln \lambda^H$
and $\ln \lambda_4^H$ for H signal (red color) and C background (blue color)
exclusive $\mu^- p \pi^+$ CC topologies. The dotted line on the bottom right plots corresponds to a
cut $R_{mH}<-0.7$ (Fig.~\ref{fig:Rmh}).
}
\label{fig:Hsel-Hvar}
\end{figure}
The most powerful kinematic variable to separate H interactions from the ones
originated in nuclear targets is $R_{mH} \equiv (p_T^m - p_T^H)/(p_T^m + p_T^H)$,
the asymmetry between the missing transverse momentum, $p_T^m$, and the
transverse momentum of the hadron vector, $p_T^H$. For H interactions $p_T^m$ is
consistent with zero up to reconstruction effects and hence we expect $R_{mH} \sim -1$.
Instead, if the interactions occur inside a nuclear target we expect, on average,
a substantial $p_T^m$ together with smaller values of $p_T^H$, due to the nuclear smearing.
Furthermore, the missing transverse momentum is mainly generated inside the hadron
system and it is expected to be correlated with the latter.
All these nuclear effects result in much larger values of $R_{mH}$.
As shown in Fig.~\ref{fig:Rmh}, this variable can be efficiently used to separate H
interactions, as well as to probe various aspects of the nuclear smearing.
Another useful variable is the magnitude of the component of the hadron transverse momentum
perpendicular to the transverse momentum of the lepton, $p_{T \perp}^H$.
In H interactions the transverse momenta of the lepton and of the hadron system are back-to-back,
thus resulting in a a sharp peak around zero in $p_{T \perp}^H$.
Interactions from nuclear targets have a much broader distribution originating from the
nuclear smearing. The use of this variable to study H interactions within composite targets
was suggested in Ref.~\cite{Lu:2015hea}. Since $p_{T \perp}^H$ is selecting topologies in which
the transverse momenta of the lepton and the hadron system are back-to-back, the effect
of this variable is similar to the use of the angle between those transverse vectors,
$\Phi_{lH}$. Used as a single variable in the H selection, $p_{T \perp}^H$ has less discriminating power
than both $R_{mH}$ and the other kinematic variables discussed in this section.
However, it provides information complementary to $R_{mH}$, so that the combined effect
of both these variables improves the overall selection efficiency.
The use of simple cuts $R_{mH}<-0.6$ and $p_{T \perp}^H <0.03$ GeV/c
provides a selection of $\nu_\mu p \to \mu^- p \pi^+$ H interactions
from the CH$_2$ plastic with an efficiency of 93\% and a purity
of 86\%, including non-resonant backgrounds as well as higher
order resonances above $\Delta(1232)$.
Similarly, we can select the $\bar \nu_\mu p \to \mu^+ p \pi^-$ topology
with a 89\% efficiency and a 84\% purity.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.00\textwidth]{Hsel-Hvars2}
\end{center}
\caption{Some of the kinematic variables and correlations used to
construct $\ln \lambda^H_{\rm IN}$ for H signal (red color) and C background (blue color)
exclusive $\mu^- p \pi^+$ CC topologies. All histograms in the right plots are normalized to
unit area.}
\label{fig:varin}
\end{figure}
We can improve the selection of H interactions by using multivariate techniques exploiting
the complete event kinematics~\cite{Astier:2001yj,Astier:2000ng,Naumov:2004wa}.
Assuming the two momentum vectors of the lepton and hadron system, we have in total
3 transverse and 2 longitudinal degrees of freedom in the event selection, due to the
invariance for an arbitrary rotation in the transverse plane. Since we want to separate
the same CC events with and without nuclear effects, we can further assume that
the overall reconstructed energy spectra are similar,
thus somewhat reducing the rejection power of one of the longitudinal variables
compared to the transverse ones. As a result, we can define a complete kinematic set as
3 transverse plus one longitudinal variables.
We select the angle between the total visible momentum vector and
the incident neutrino direction ($z$ axis), $\theta_{\nu T}$, as the variable including
longitudinal information. This variable is expected to be close to zero in H interactions,
up to the tiny beam divergence, while it is
much larger in interactions originated from nuclear targets.
We use likelihood functions incorporating multi-dimensional correlations among
kinematic variables. A study of the kinematic selection suggests the following
function using only three-dimensional correlations:
\begin{equation}
{\cal L}^{\rm H} \equiv \left[ [\;R_{mH}, \; p_{T \perp}^H, \; \theta_{\nu T} \;],
\; p_T^m, \; \Phi_{lH} \; \right]
\label{eq:lnH}
\end{equation}
where the square brackets denote correlations (Fig.~\ref{fig:Hsel-Hvar}).
The ${\cal L}^{\rm H}$ function is over-constrained in the transverse plane to compensate for the
reduced phase space. An optimal function strictly based upon a complete set of
kinematic variables is the four-dimensional:
\begin{equation}
{\cal L}^{\rm H}_4=[\; p_T^H, \; p_T^m, \; \Phi_{lH} , \; \theta_{\nu T} \; ]
\label{eq:lnH4}
\end{equation}
which incorporates the $R_{mH}$ variable through its underlying
correlation $[\; p_T^H, \; p_T^m \;]$ (Fig.~\ref{fig:Hsel-Hvar}).
Although the use of the ${\cal L}^{\rm H}_4$ function requires a larger statistics to
achieve a sensible binning, we used it for our kinematic selection and obtained
results similar to ${\cal L}^{\rm H}$.
We build the ${\cal L}^{\rm H}$ and ${\cal L}^{\rm H}_4$ probability density functions (pdf) for
the two test hypotheses of H interactions (signal) and C interactions (background).
The individual pdf are properly smoothed and are built with samples independent
from the test ones to avoid large statistical biases. As it is common practice, the logarithm
of the final likelihood ratio between signal and background hypotheses,
$\ln \lambda^{\rm H}$ or $\ln \lambda_4^{\rm H}$, is used as discriminant~\cite{Astier:2001yj}.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.00\textwidth]{Hsel-Hlike2}
\end{center}
\caption{Left plot: Distributions of $\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$
for the H signal, the C background, and the CH$_2$ plastic (sum) for the
exclusive $\mu^-p\pi^+$ CC topologies. The multiple peaks are the effect of the
binning used to build ${\cal L}^{\rm H}$ and ${\cal L}^{\rm H}_{\rm IN}$.
The C distribution is normalized to unit area while the H distribution is normalized to
the expected relative abundance in CH$_2$.
Right plot: Efficiency (red color) and purity (blue color) as a function of the cut on
$\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$ for the kinematic
selection of the exclusive processes $\nu_\mu p \to \mu^-p\pi^+$
(solid lines) and $\bar \nu_\mu p \to \mu^+p\pi^-$ (dashed-dotted lines)
on hydrogen from the CH$_2$ target.
}
\label{fig:Hlike}
\end{figure}
We can also exploit the information related to the individual particles
inside the hadron system. For $\mu^\mp p \pi^\pm$ topologies we have
a total of 3 additional degrees of freedom,
since the total hadron momentum vector is constrained by the global event kinematics
in ${\cal L}^{\rm H}$ and ${\cal L}_4^{\rm H}$. We select one of such variables as the
difference $\Delta E$ between the energy of the neutrino calculated from the
$\mu$ and $\pi$ momenta by applying energy-momentum conservation
and the one reconstructed from the sum of the momentum vector of all 3
particles $\mu^\mp p \pi^\pm$:
\begin{equation}
\Delta E_p = \frac{m_\mu^2-m_{\pi^\pm}^2+2M_p\left( E_\mu + E_{\pi^\pm} \right) -
2p_\mu \cdot p_{\pi^\pm}}
{2\left( M_p - E_\mu - E_{\pi^\pm} + \mid \vec{p}_\mu \mid \cos \theta_\mu +
\mid \vec{p}_{\pi^\pm} \mid \cos \theta_{\pi^\pm} \right)} -
\mid \vec{p}_\mu + \vec{p}_{\pi^{\pm}} + \vec{p}_p \mid
\label{eq:RES1}
\end{equation}
where $M_p, M_n, m_\mu, m_{\pi^\pm}$ are the masses of the proton, neutron, muon, and
pion, respectively, $p_{\mu (\pi^\pm)}, \vec{p}_{\mu (\pi^\pm)}, E_{\mu (\pi^\pm)}$
and $\theta_{\mu (\pi^\pm)}$ are the 4-momentum, momentum vector, energy
and angle of the outgoing muon (pion), and $\vec{p}_p$ is the proton momentum vector.
The variable $\Delta E$ is close to zero up to reconstruction effects in hydrogen, while
it is significantly larger in carbon events, due to the nuclear smearing.
Another useful variable is the invariant mass of the reconstructed neutrino, calculated as
$m_0 = \; \mid p_\mu + p_{\pi^\pm} + p_p - p_N \mid$
where $p_p$ and $p_N$ are the
4-momenta of the outgoing proton and of the target proton assumed at rest, respectively.
We use the following likelihood function using information from
the internal $p \pi$ structure:
\begin{equation}
{\cal L}^{\rm H}_{\rm IN}=[\; \Delta E_p, \; p_L^p, \; m_0 \;]
\label{eq:lnHin}
\end{equation}
where $p_L^p$ is the longitudinal component of the proton
momentum vector along the beam direction.
Figure~\ref{fig:varin} shows the main variables and
correlations included in ${\cal L}^{\rm H}_{\rm IN}$. Since ${\cal L}^{\rm H}_{\rm IN}$
is essentially independent from ${\cal L}^{\rm H}$ and ${\cal L}^{\rm H}_{4}$ we
multiply the corresponding density functions and use the sum
$\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$ or
$\ln \lambda^{\rm H}_4$+$\ln \lambda^{\rm H}_{\rm IN}$ as the final discriminant
for our analysis.
\begin{table}[tb]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|c|} \hline
& \multicolumn{4}{|c|}{ $\nu_\mu$-H CC}
& \multicolumn{6}{|c|}{$\bar \nu_\mu$-H CC} \\
Process & $\mu^- p \pi^+$ & $\mu^- p \pi^+ X$ & $\mu^- n \pi^+ \pi^+ X$ & Inclusive &
$\mu^+ p \pi^-$ & $\mu^+ n \pi^0$ & $\mu^+ n$ & $\mu^+ p \pi^- X$ & $\mu^+ n \pi \pi X$ & Inclusive \\
\hline\hline
Eff. $\varepsilon$ & 96\% & 89\% & 75\% & 93\% & 94\% & 84\% & 75\%& 85\% & 82\% & 80\% \\
Purity & 95\% & 93\% & 70\%& 93\% & 95\% & 84\% & 80\% & 94\% & 84\% & 84\% \\
\hline
\end{tabular}
\caption{Efficiency $\varepsilon$ and purity for the kinematic selection of H interactions from the CH$_2$ plastic target using the likelihood ratio
$\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$
or $\ln \lambda^{\rm H}_4$+$\ln \lambda^{\rm H}_{\rm IN}$. For the $\mu^+ n$ QE
topologies $\ln \lambda^{\rm H}_{\rm QE}$ is used instead.
The cuts applied for each channel are chosen to maximize the sensitivity defined
as $S/\sqrt{S+B}$, where $S$ is the H signal and $B$ the C background. The CC inclusive samples
are obtained from the combination of the corresponding exclusive channels.
}
\label{tab:effpur}
\end{center}
\end{table}
The distributions of $\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$
for the H signal and the C background in
$\mu^- p \pi^+$ topologies are shown in Fig.~\ref{fig:Hlike} (left plot).
The corresponding purity and efficiency achievable as a function of the
$\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$ cut
are given in the right plot of Fig.~\ref{fig:Hlike}, for both the
$\nu_\mu p \to \mu^- p \pi^+$ and $\bar \nu_\mu p \to \mu^+ p \pi^-$ samples.
Both the efficiency and the purity appear relatively uniform as a function of the
neutrino energy.
Table~\ref{tab:effpur} summarizes the results obtained by applying the cut on
$\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$
maximizing the sensitivity $S/\sqrt{S+B}$, where $S$ is the H signal and $B$ the C background
The fact that the maximum sensitivity corresponds to regions with high purity for the
selected H signal indicates that the kinematic selection is optimal.
The use of a multi-variate selection further reduces the background levels with respect
to the simple cut analysis, without dramatically changing the overall results.
A key advantage of this approach is that the likelihood function allows to assign,
on an event-by-event basis, the probability that a given (anti)neutrino interaction
originated from either the hydrogen or the carbon nucleus. Furthermore, it provides a better
control of the selection procedure by easily varying the efficiency/purity and by
offering relatively clean control samples.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.00\textwidth]{Hsel-Hlike3}
\end{center}
\caption{Left plot: Distributions of $\ln \lambda^{\rm H}_{\rm QE}$ for
the H signal, the C background, and the CH$_2$ plastic (sum) for
the selection of $\bar \nu_\mu p \to \mu^+ n$ QE events.
Right plot: Distributions of $\ln \lambda^{\rm H}_4$
for the selection of $\bar \nu_\mu p \to \mu^+ n \pi^0$ events.
The multiple peaks are the effect of the binning used to build the likelihood functions.
The C distributions are normalized to unit area while the H distributions are normalized to
the expected relative abundance in CH$_2$. See the text for details.
}
\label{fig:Hsel-likeHqepi0}
\end{figure}
\subsubsection{Selection of $\bar \nu_\mu p \to \mu^+ n$ QE}
\label{sec:numubarqe}
An important exclusive process available in $\bar \nu_\mu$ CC
interactions is the quasi-elastic $\bar \nu_\mu p \to \mu^+ n$ on
hydrogen. The reconstruction of this topology is more complex because of the
presence of a neutron in the final state and a single charged track.
We performed detailed GEANT4 simulations of the detector to study the
reconstruction efficiency of the neutrons. The results were independently
checked using FLUKA simulations.
About 25\% of the neutrons produced in
$\bar \nu_\mu p \to \mu^+ n$ QE events on H interact inside the STT and
can be detected. An additional fraction of about 55\% of the neutrons
can be detected in the ECAL~\cite{Adinolfi:2002zx} surrounding the tracker,
thus allowing a combined reconstruction efficiency around 80\%.
Conversely, the combined detection efficiency for neutrons originated
from the C background events is only about 65\%, since
they are typically affected/produced by nuclear effects, resulting, on average,
in a smaller kinetic energy.
For events with a single charged track the resolution on the position of the primary vertex
is worse than for multi-track events, and is essentially defined by the thickness of a single
CH$_2$ (or C) target plane by noting the absence of straw tube hits preceding the presumed target.
However, events can still be efficiency associated to the correct target material,
due to the lightness of the tracking straws and the purity of the target itself. The corresponding
uncertainty is given by the ratio between the thickness of the straw walls and the thickness of
a single CH$_2$ target, resulting in an efficiency $>99\%$. From the positions of the
primary vertex and that of the neutron interaction within the detector we can
reconstruct the neutron line of flight. Assuming that the target proton is at rest,
we calculate the energy of the incoming antineutrino as:
\begin{equation}
E_\nu = \frac{M_n^2-m_\mu^2+2M_pE_\mu -M_p^2}{2\left( M_p- E_\mu +\mid \vec{p}_\mu \mid \cos \theta_\mu\right)}
\label{eq:QE}
\end{equation}
with the same notations as in \eq{eq:RES1}.
The energy of the neutron is $E_n=E_\nu+M_p-E_\mu$ and the momentum
vector of the neutron is obtained from the measured neutron direction and the
calculated energy $E_n$~\footnote{It is possible to calculate the complete
momentum vector of the neutron from energy-momentum conservation.
However, the direct use of transverse plane kinematics would void the kinematic
selection against the C background. For this reason we combine the calculated energy $E_n$
with the measured neutron direction.}.
We note that \eq{eq:QE} is correct only for interactions on hydrogen and not for
the ones on carbon, due to nuclear effects.
We use a realistic smearing on the measured direction of the neutron obtained
from a detailed FLUKA simulation of the detector and including the non gaussian
tails, mainly related to elastic scattering of the neutrons before interacting.
We can then reconstruct the complete event kinematics for $\bar \nu_\mu p \to \mu^+ n$
interactions on hydrogen and apply the same kinematic selection described in
Sec.~\ref{sec:kine}. The resolution achievable on many of the kinematic variables
is dominated by the reconstruction of the neutron direction rather than that of
the $\mu^+$ momentum vector, given the excellent momentum and angular
resolution of the STT.
An additional background source to consider for the selection of
the process $\bar \nu_\mu p \to \mu^+ n$ on H is given by the uncorrelated neutrons
originated from (anti)neutrino interactions occurring in the large amount of
materials surrounding the STT, including the ECAL, the magnet elements,
and the external rocks. We estimate this background by randomly overlaying
a genuine $\bar \nu_\mu p \to \mu^+ n$ event on H with a neutron extracted
from either a different CH$_2$ event or from an event in the surrounding materials.
We then reconstruct the event kinematics using the measured direction of the
random neutron and the calculated energy $E_n$, as described above.
Since the random neutron is uncorrelated with the primary vertex from the
$\bar \nu_\mu p \to \mu^+ n$ interaction on H, the resulting kinematic
distributions are inconsistent with those expected from a genuine H signal.
In addition,
we can require that the measured time difference between the interaction
point of the random neutron and the primary vertex is consistent with the expected time of flight.
We have less kinematic information available for the selection of
$\bar \nu_\mu p \to \mu^+ n$ QE on H with respect to other channels.
We use the optimal function ${\cal L}^{\rm H}_4$ from \eq{eq:lnH4} as the basis
of our kinematic analysis.
The likelihood function ${\cal L}^{\rm H}_{\rm IN}$, as defined in \eq{eq:lnHin},
is not strictly applicable, since the hadron system is composed of a single neutron
without internal structure.
However, we can still exploit one additional longitudinal degree of freedom
with respect to \eq{eq:lnHin}.
In analogy of \eq{eq:RES1}, we calculate $\Delta E_n = E_\nu \; -
\mid \vec{p}_\mu + \vec{p}_n \mid $, where $E_\nu$ is given by \eq{eq:QE} and
$\vec{p}_n$ is the neutron momentum calculated as described above.
We use the following likelihood function for the selection of
$\bar \nu_\mu p \to \mu^+ n$ QE on H:
\begin{equation}
{\cal L}^{\rm H}_{\rm QE}= \left[ [\; p_T^H, \; p_T^m, \; \Phi_{lH} , \; \theta_{\nu T} \; ],
\; \Delta E_n, \; p_L^n \; \right]
\label{eq:QE2}
\end{equation}
which is over-constrained to compensate for the reduced phase space.
The variable $p_L^n$ represents the longitudinal component of the neutron
momentum vector along the beam direction.
We note that, although the two-dimensional correlation
$ [\; \Delta E_n, \; p_L^n \;]$
appears similar to the corresponding one in \eq{eq:lnHin},
it has a different role in the QE channel, as a part of the global event kinematics.
We therefore include it into a single likelihood function, since it is expected to be
correlated with the other kinematic variables.
Figure~\ref{fig:Hsel-likeHqepi0} shows the distributions of
$\ln \lambda_{\rm QE}^{\rm H}$ for the H signal and the
C background. Table~\ref{tab:effpur} summarizes the efficiency and purity for a cut on
$\ln \lambda_{\rm QE}^{\rm H}$ maximizing the sensitivity of the analysis.
\subsubsection{Selection of $\bar \nu_\mu p \to \mu^+ n \pi^0$}
\label{sec:numubarpi0}
Another exclusive process available only in $\bar \nu_\mu$ CC interactions is
$\bar \nu_\mu p \to \mu^+ n \pi^0$ on hydrogen, mainly originated from
resonance production. The kinematic selection of this channel is similar to the
one discussed in Sec.~\ref{sec:3trk} for the complementary process
$\bar \nu_\mu p \to \mu^+ p \pi^-$,
with the exception of a few specific features related to the reconstruction of
the neutron and the $\pi^0$.
The relative fractions of these two channels are about comparable in
$\bar \nu_\mu$ CC interactions on hydrogen.
For the reconstruction of the $\pi^0$ we distinguish the case in which $\gamma$s
are converted into $e^+e^-$ pairs within the STT volume and the one in which $\gamma$s are
detected only in the surrounding ECAL. In the former case we use the momentum and angle
smearing for $e^\pm$ tracks in STT, while in the latter case we apply the corresponding
smearing obtained from a detailed ECAL simulation. On average, about 50\% of all
$\pi^0$ have at least one converted $\gamma$ within the STT volume, providing a
more accurate reconstruction of the $\pi^0$ direction.
We calculate the energy of the neutrons interacting in the detector as:
\begin{equation}
E_n = \frac{M_n^2-m_\mu^2-m_{\pi^0}^2+2M_p\left( E_\mu + E_{\pi^0} \right) - 2p_\mu \cdot p_{\pi^0}-M_p^2}
{2\left( M_p - E_\mu - E_{\pi^0} + \mid \vec{p}_\mu \mid \cos \theta_\mu +
\mid \vec{p}_{\pi^0} \mid \cos \theta_{\pi^0} \right)} + M_p - E_\mu - E_{\pi^0}
\label{eq:RES2}
\end{equation}
where $p_{\pi^0}, \vec{p}_{\pi^0}, E_{\pi^0)}$ and $\theta_{\pi^0}$
are the 4-momentum, momentum vector, energy and angle of the outgoing $\pi^0$, respectively.
The first term on the r.h.s. of \eq{eq:RES2} represents the neutrino energy
calculated from the muon and $\pi^0$ momenta using energy-momentum conservation.
We note that \eq{eq:RES2} is correct only for interactions on hydrogen since it
assumes a target proton at rest. We reconstruct the neutron momentum vector $\vec{p}_n$
following the same procedure described in Sec.~\ref{sec:numubarqe} for the QE process,
combining the measured (smeared) direction of the neutron
with the energy $E_n$ calculated from \eq{eq:RES2}.
We use the likelihood function ${\cal L}^{\rm H}_4$ from \eq{eq:lnH4} to describe the
global event kinematics.
In analogy to the $\bar \nu_\mu p \to \mu^+ p \pi^-$ case, it is possible to exploit the
additional information related to the individual particles within the hadron system
with the function ${\cal L}^{\rm H}_{\rm IN}= [\; \Delta E_{\pi^0}, \; p_L^{\pi^0}, \; m_0 \;]$,
where $\Delta E_{\pi^0} = E_\nu - \mid \vec{p}_\mu + \vec{p}_{\pi^0} + \vec{p}_n \mid$,
$E_\nu$ is the neutrino energy calculated from the muon and neutron momenta using
energy-momentum conservation, and $p_L^{\pi^0}$ is the longitudinal
component of the momentum vector of the $\pi^0$.
The use of $\Delta E_{\pi^0}$ is preferable with respect to
the equivalent quantity for the neutron, $\Delta E_n$, since this latter is partially
biased by the calculation of the neutron energy from \eq{eq:RES2}.
Since the average reconstruction smearing for the neutron and $\pi^0$ is larger than for
most charged particles, the $\ln \lambda^{\rm H}_{\rm IN}$ improves only marginally
the selection of $\bar \nu_\mu p \to \mu^+ n \pi^0$ on H and can be dropped for this
channel. Figure~\ref{fig:Hsel-likeHqepi0} shows the distributions of
$\ln \lambda_4^{\rm H}$ for the H signal and the
C background. Table~\ref{tab:effpur} summarizes the efficiency and purity for a cut on
$\ln \lambda_4^{\rm H}$ maximizing the sensitivity of the analysis.
Similar results are obtained with $\ln \lambda^{\rm H}$.
\subsubsection{Selection of $\nu_\mu p \to \mu^- p \pi^+ X$
and $\bar \nu_\mu p \to \mu^+ p \pi^- X$}
\label{sec:disp}
In this section and in the following one
we consider the collective selection of all the inelastic topologies produced
in $\nu_\mu p$ and $\bar \nu_\mu p$ CC interactions on H and
different from the $\mu^- p \pi^+$, $\mu^+ p \pi^-$,
and $\mu^+ n \pi^0$ topologies discussed in
Sec.~\ref{sec:3trk} and Sec.~\ref{sec:numubarpi0}. These samples are
dominated by the Deep Inelastic Scattering (DIS): about 63\%
of the $\mu^- \pi^+ X$ events
and 7\% of the $\mu^- p \pi^+$
sample have $W>1.8$ GeV with the default low energy beam option considered.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=1.00\textwidth]{Hsel-Hlike4}
\end{center}
\caption{Distributions of $\ln \lambda^{\rm H}_4$+$\ln \lambda^{\rm H}_{\rm IN}$
for the H signal, the C background, and the CH$_2$ plastic (sum) for
$\nu_\mu p \to \mu^- \pi^+ X$ (left plot) and $\bar \nu_\mu p \to \mu^+ \pi X$ (right plot),
including the sum of the subsamples with a proton and a neutron.
The multiple peaks are the effect of the binning used to build
${\cal L}_4^{\rm H}$ and ${\cal L}^{\rm H}_{\rm IN}$.
The C distributions are normalized to unit area while the H distributions are normalized to
the expected relative abundance in CH$_2$. See the text for details.
}
\label{fig:Hsel-likeHdis}
\end{figure}
The inelastic $\nu_\mu p \to \mu^- \pi^+ X$ and $\bar \nu_\mu p \to \mu^+ \pi X$
samples on H are characterized, in general, by a higher multiplicity and a worse
reconstruction of some of the events, compared to the $\mu^- p \pi^+$, $\mu^+ p \pi^-$,
and $\mu^+ n \pi^0$ topologies.
The larger detector smearing directly affects the reconstruction
of the kinematic variables, thus somewhat reducing their discriminating power.
However, these effects are even larger for the DIS interactions
originated in nuclear targets, primarily because of the final state interactions.
We can therefore still achieve an adequate separation of the interactions on hydrogen
from the carbon background, which largely dominates the statistics of the
$\mu^- \pi^+ X$ and $\mu^+ \pi X$ samples from the CH$_2$ target.
Since we want to select interactions on protons, we require that the total charge measured
at the primary vertex is $C=+1$. This cut rejects a large fraction of the interactions on
neutrons in background events and has an efficiency of about 23\%(33\%) for
$\nu_\mu (\bar \nu_\mu)$ CC interactions on C.
We also require the presence of at least one $\pi^+$ in $\nu_\mu p$ and one $\pi^-$
or $\bar \nu_\mu p$ events to select inelastic interactions.
In $\nu_\mu p$ and $\bar \nu_\mu p$ CC interactions on H we expect a single nucleon --
either a proton or a neutron -- in the final state, while for interactions on C nuclear effects
including final state interactions can result in higher multiplicities.
We perform a separate analysis of
the two subsamples with a proton and a neutron.
In this section we discuss the selection of the subsamples with a reconstructed proton.
Detailed GEANT4 simulations indicate that the average proton reconstruction
efficiency is about 96\% for H events and 74\% in C events.
The kinematic selection of $\nu_\mu p \to \mu^- p \pi^+ X$
and $\bar \nu_\mu p \to \mu^+ p \pi^- X$ on H can exploit additional degrees of freedom
with respect to the $\mu^- p \pi^+$, $\mu^+ p \pi^-$, and $\mu^+ n \pi^0$ topologies,
due to the higher number of particles in the hadron system. We use the likelihood
function ${\cal L}^{\rm H}_4$ from \eq{eq:lnH4} to describe the
global event kinematics. We then tag the single particle within the hadron system
potentially affected by the largest nuclear effects by maximizing the magnitude
$\mid \Delta E_{h_i} \mid$, defined as:
\begin{equation}
\Delta E_{h_i} =
\frac{m_{h_i}^2-m_\mu^2-(W^{\prime})^2+2M_p\left( E_\mu + E_{H^\prime} \right) - 2p_\mu \cdot p_{H^\prime}-M_p^2}
{2\left( M_p - E_\mu - E_{H^\prime} + \mid \vec{p}_\mu \mid \cos \theta_\mu +
\mid \vec{p}_{H^\prime} \mid \cos \theta_{H^\prime} \right)}
- \mid \vec{p}_\mu + \sum_k \vec{p}_{h_k} \mid
\label{eq:DIS}
\end{equation}
which is calculated for each hadron particle $h_i$ of mass $m_{h_i}$ and momentum
vector $\vec{p}_{h_i}$. The first term on the r.h.s. of \eq{eq:DIS} is similar to the equivalent
one in \eq{eq:RES2}, with the $\pi^0$ replaced by a reduced hadron system
$H^\prime$ excluding the single particle $h_i$ being considered. The quantity
$(W^\prime)^2=E_{H^\prime}^2-\mid \vec{p}_{H^\prime} \mid^2$ is the invariant mass of
the reduced hadron system, and $E_{H^\prime}=\sum_{k\neq i} E_{h_k}$ and
$\vec{p}_{H^\prime} = \sum_{k \neq i} \vec{p}_{h_k}$ are the corresponding energy
and momentum vector. We select the hadron particle maximizing the magnitude
$\mid \Delta E_{h_i} \mid$ and use the corresponding value $\Delta E_{h_i}^{\rm max}$ (with sign)
and the longitudinal momentum of such a particle, $p_L^{h_i}$, as input for the likelihood
function~\footnote{In principle the same approach can be used for
$\mu^- p \pi^+$, $\mu^+ p \pi^-$, and $\mu^+ n \pi^0$ topologies, further improving
the selection of these channels with respect to the results presented
in Sec.~\ref{sec:3trk} and Tab.~\ref{tab:effpur}.}
based upon information internal to the hadron system:
\begin{equation}
{\cal L}^{\rm H}_{\rm IN}=[\; \Delta E_{h_i}^{\rm max}, \; p_L^{h_i}, \; m_0, \; ]
\label{eq:DIS2}
\end{equation}
where we use similar notations as in \eq{eq:lnHin}.
Figure~\ref{fig:Hsel-likeHdis} shows the distributions of
$\ln \lambda_4^{\rm H}$+$\ln \lambda_{\rm IN}^{\rm H}$ for the H signal and the C background
in $\nu_\mu p \to \mu^- \pi^+ X$ and $\bar \nu_\mu p \to \mu^+ \pi X$ topologies.
Table~\ref{tab:effpur} summarizes the
efficiency and purity in the selection of both $\nu_\mu p \to \mu^- p \pi^+ X$
and $\bar \nu_\mu p \to \mu^+ p \pi^- X$ processes on H with
a cut on $\ln \lambda_4^{\rm H}$+$\ln \lambda_{\rm IN}^{\rm H}$
maximizing the sensitivity of the analysis.
Similar results are obtained with $\ln \lambda^{\rm H}$+$\ln \lambda_{\rm IN}^{\rm H}$.
\subsubsection{Selection of $\nu_\mu p \to \mu^- n \pi^+ \pi^+ X$
and $\bar \nu_\mu p \to \mu^+ n \pi \pi X$}
\label{sec:disn}
In this section we discuss the selection of the subsamples of the inelastic
$\nu_\mu p$ and $\bar \nu_\mu p$ CC interactions on H with a detected neutron in
the final state and different from the $\mu^+ n \pi^0$ topologies (Sec.~\ref{sec:numubarpi0}).
These samples are complementary with respect to the similar ones with a reconstructed
proton in the final state described in Sec.~\ref{sec:disp}.
For $\nu_\mu p$ CC on H the presence of a neutron in the final state is
tagged by requiring that no proton and two $\pi^+$ are present.
These criteria correctly tag the neutron in 99.8\% of the cases,
including events in which the neutron is not detected.
Events with a neutron in $\bar \nu_\mu p$ CC on H are tagged with an efficiency of
88\% by requiring that no proton and a number of pions ($\pi^-$+$\pi^+$+$\pi^0$)
$\geq 2$ with total charge equal to zero are present in the final state.
The analysis follows closely the one described in Sec.~\ref{sec:disp} for the
subsamples with a proton, the main difference being the treatment of the neutron.
We calculate the neutron energy using energy-momentum conservation and \eq{eq:RES2} with the
$\pi^0$ replaced by a reduced hadron system $H^\prime$ excluding the neutron:
\begin{equation}
E_n = \frac{M_n^2-m_\mu^2-(W^{\prime})^2+2M_p\left( E_\mu + E_{H^\prime} \right) - 2p_\mu \cdot p_{H^\prime}-M_p^2}
{2\left( M_p - E_\mu - E_{H^\prime} + \mid \vec{p}_\mu \mid \cos \theta_\mu +
\mid \vec{p}_{H^\prime} \mid \cos \theta_{H^\prime} \right)} + M_p - E_\mu - E_{H^\prime}
\label{eq:DIS3}
\end{equation}
where we use the same notations as in \eq{eq:DIS} with $h_i\equiv n$.
We reconstruct the neutron momentum vector $\vec{p}_n$
following the same procedure described in Sec.~\ref{sec:numubarqe} for the QE process,
combining the measured (smeared) direction of the neutron
with the energy $E_n$ calculated from \eq{eq:DIS3}.
For events with more than one neutron detected the calculation above is not applicable
and we ignore the neutrons.
We use ${\cal L}^{\rm H}_4$ from \eq{eq:lnH4} to describe the
global event kinematics and ${\cal L}^{\rm H}_{\rm IN}$ from \eq{eq:DIS2} for the
information related to the individual particles inside the hadron system.
Since the angular smearing for the detected neutrons (Sec.~\ref{sec:numubarqe})
is typically larger than for other particles, we use the track with the largest
angle with respect to the beam direction to calculate $ \Delta E_{h_i}^{\rm max}$
in \eq{eq:DIS2}, rather than explicitly maximizing $ \mid \Delta E_{h_i} \mid$.
Figure~\ref{fig:Hsel-likeHdis} shows the distributions of
$\ln \lambda_4^{\rm H}$+$\ln \lambda_{\rm IN}^{\rm H}$ for the H signal and the C background
in $\nu_\mu p \to \mu^- \pi^+ X$ and $\bar \nu_\mu p \to \mu^+ \pi X$ topologies.
Table~\ref{tab:effpur} summarizes the
efficiency and purity in the selection of both $\nu_\mu p \to \mu^- n \pi^+ \pi^+ X$
and $\bar \nu_\mu p \to \mu^+ n \pi \pi X$ processes on H with
a cut on $\ln \lambda_4^{\rm H}$+$\ln \lambda_{\rm IN}^{\rm H}$
maximizing the sensitivity of the analysis.
Similar results are obtained with $\ln \lambda^{\rm H}$+$\ln \lambda_{\rm IN}^{\rm H}$.
\begin{table}[tb]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|c|} \hline
& \multicolumn{4}{|c|}{ $\nu_\mu$-H CC, $\varepsilon\equiv 75\%$}
& \multicolumn{6}{|c|}{$\bar \nu_\mu$-H CC, $\varepsilon\equiv 75\%$} \\
Process & $\mu^- p \pi^+$ & $\mu^- p \pi^+ X$ & $\mu^- n \pi^+ \pi^+ X$ & Inclusive &
$\mu^+ p \pi^-$ & $\mu^+ n \pi^0$ & $\mu^+ n$ & $\mu^+ p \pi^- X$ & $\mu^+ n \pi \pi X$ & Inclusive \\
\hline\hline
Purity & 99\% & 99\% & 70\% & 98\% & 99\% & 90\% & 80\% & 98\% & 90\% & 86\% \\
\hline
\end{tabular}
\caption{Purity achieved with the kinematic selection of H interactions from the
CH$_2$ plastic target using a cut on the likelihood ratio
$\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$ or
$\ln \lambda^{\rm H}_4$+$\ln \lambda^{\rm H}_{\rm IN}$ resulting in the fixed H signal
efficiency $\varepsilon$ specified. For the $\mu^+ n$ QE topologies
$\ln \lambda^{\rm H}_{\rm QE}$ is used instead.
For illustration purpose, the value of the efficiency is chosen as the lowest
among the ones listed in Tab.~\ref{tab:effpur} for individual topologies.
The CC inclusive samples are obtained from the combination
of the corresponding exclusive channels.
}
\label{tab:inclusive}
\end{center}
\end{table}
\subsubsection{Selection of $\nu_\mu p$ and $\bar \nu_\mu p$ CC inclusive}
\label{sec:inclusive}
In the previous sections we optimized the selection of the various exclusive
topologies available in $\nu_\mu p$ and $\bar \nu_\mu p$ CC interactions on H
by maximizing independently the corresponding sensitivities.
The results summarized in Tab.~\ref{tab:effpur} are characterized by varying
efficiencies and purities across different channels. For measurements requiring the
inclusive CC samples we can combine the individual exclusive topologies with their
corresponding relative fractions in $\nu_\mu p$ and $\bar \nu_\mu p$ CC
interactions on H. The average efficiency and purity of the resulting inclusive
CC samples on H are listed in Tab.~\ref{tab:effpur}.
An alternative approach to obtain inclusive $\nu_\mu p$ and $\bar \nu_\mu p$ CC
samples is to tune the kinematic selection to achieve a constant fixed efficiency
across all individual channels. We can then directly sum all the selected exclusive
topologies independently from the corresponding relative fractions.
The purities of the individual exclusive processes will still be different, but the use
of the dedicated graphite target for the background subtraction automatically
corrects for such differences in a model-independent way. The use of a
multi-variate selection allows an easy variation of the fixed efficiency, as
shown in Fig.~\ref{fig:Hlike}. As an example, Tab.~\ref{tab:inclusive} illustrates the
results obtained by imposing the lowest efficiency achieved among individual channels
in Tab.~\ref{tab:effpur}.
\subsection{Achievable statistics}
\label{sec:stat}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.70\textwidth]{Hsel-Hstat}
\end{center}
\caption{Ratio between the relative statistical uncertainty $\Delta_{\rm stat}$ on the H sample
obtained after the C background subtraction and the corresponding ideal one from a pure
H$_2$ sample of the same size, as a function of the ratio between the fiducial masses of the
graphite and CH$_2$ targets, $M_C/M_{CH_2}$.
The curves for a H purity of 90\% (solid line) and 60\% (dashed-dotted line) and 90\% efficiency,
as well as for a 99\% purity and a 75\% efficiency (dashed line)
are shown to illustrate the impact of the kinematic selection. See text for details.
}
\label{fig:Hsel-Hstat}
\end{figure}
In the following we will assume an overall fiducial mass of 5 tons for the CH$_2$ targets.
This value is realistically achievable with the detector technology discussed in
Sec.~\ref{sec:targets} and a relatively compact tracking volume around 45 m$^3$
(without fiducial cuts), depending upon the specific configuration of the main STT parameters.
The distributions of the generic kinematic variables
$\vec x \equiv (x_1, x_2, ....., x_n)$ in $\nu(\bar \nu)$-H CC interactions are obtained as:
\begin{equation}
N_H(\vec x) \equiv N_{CH_2} (\vec x) - N_{C} (\vec x) \times \frac{M_{C/CH_2}}{M_C}
\label{eq:Hevt}
\end{equation}
where $N_{CH_2}$ and $N_C$ are the data from the CH$_2$ plastic and graphite (C)
targets. The interactions from this latter are normalized by the ratio between the total fiducial
masses of C within the graphite and CH$_2$ targets, $M_{C/CH_2}/M_C$.
The subtraction in \eq{eq:Hevt} is performed after all the selection cuts, including the
kinematic analysis described in Sec.~\ref{sec:Hsele} and
resulting in the purities and efficiencies summarized
in Tab.~\ref{tab:effpur}. Practical considerations require the graphite targets to be smaller
than the actual amount of C inside the CH$_2$ plastic, thus resulting in a statistical penalty
associated with the subtraction procedure. Figure~\ref{fig:Hsel-Hstat} illustrates how the
total statistical uncertainty on $N_H$ from \eq{eq:Hevt} compares to the ideal one expected
from a pure H$_2$ sample equivalent to the statistics of H interactions within CH$_2$.
For a given efficiency, the purity of the H samples achievable by the kinematic selection
is crucial for the feasibility of this technique. Our analysis suggests that a fiducial mass
for the graphite targets in the range 500-600 kg (i.e. $M_C/M_{CH_2}\sim0.1$-0.12) provides a
reasonable compromise with a statistical penalty around 30\%.
We note that this statistical penalty can be further reduced by analytically
smoothing the measured distributions from the graphite target
and/or by using a tighter kinematic selection,
as illustrated in Tab.~\ref{tab:inclusive} and Fig.~\ref{fig:Hsel-Hstat}.
\begin{table}[tb]
\begin{center}
\begin{tabular}{|l|r|r|r|r|r|} \hline
CC process & ~~CH$_2$ target~~ & ~~H target~~ & ~~CH$_2$ selected~~ & ~~C bkgnd~~ & ~~H selected~~\\ \hline\hline
$\nu_\mu p \to \mu^- p \pi^+$ & 5,615,000 & 2,453,000 & 2,305,000 & 115,000 & 2,190,000 \\
$\nu_\mu p \to \mu^- p \pi^+ X$~~~ & 11,444,000 & 955,000 & 877,000 & 61,000 & 816,000 \\
$\nu_\mu p \to \mu^- n \pi^+ \pi^+ X$~~~ & 3,533,000 & 183,000 & 158,000 & 48,000 & 110,000 \\
\hline
$\nu_\mu$ CC inclusive~~ & 34,900,000 & 3,591,000 & 3,340,000 & 224,000 & 3,116,000 \\
\hline\hline
$\bar \nu_\mu p \to \mu^+ n$ & 4,450,000 & 1,688,000 & 1,274,000 & 255,000 & 1,019,000 \\
$\bar \nu_\mu p \to \mu^+ p \pi^-$ & 827,000 & 372,000 & 342,000 & 17,000 & 325,000 \\
$\bar \nu_\mu p \to \mu^+ n \pi^0$ & 791,000 & 366,000 & 295,000 & 48,000 & 247,000 \\
$\bar \nu_\mu p \to \mu^+ p \pi^- X$ & 2,270,000 & 176,000 & 153,000 & 9,000 & 144,000 \\
$\bar \nu_\mu p \to \mu^+ n \pi \pi X$ & 2,324,000 & 280,000 & 220,000 & 35,000 & 185,000 \\
\hline
$\bar \nu_\mu$ CC inclusive~~ & 13,000,000 & 2,882,000 & 2,284,000 & 364,000 & 1,920,000 \\
\hline
\end{tabular}
\caption{Number of events expected in the selection of all the various processes
on H with the default low energy (anti)neutrino beams available at the
LBNF~\cite{Acciarri:2015uup,Acciarri:2016ooe}, assuming 5+5 years of data taking
with the neutrino and antineutrino beams. The first two columns (CH$_2$ and H targets)
refer to the initial statistics, while the last three include all selection cuts
described in this paper (Sec.~\ref{sec:Hsele} and Tab.~\ref{tab:effpur}).
For the CH$_2$ and C targets the numbers refer to the given final state topologies
originated from either $p$ or $n$ interactions.
The fifth column shows the total residual C background to be subtracted from the
corresponding CH$_2$ selected samples. We use a ratio $M_C/M_{C/CH_2} =0.12$
to measure the C backgrounds from the graphite targets.
See the text for details.
}
\label{tab:events}
\end{center}
\end{table}
As an example of application of our technique, we consider the neutrino and antineutrino
beam spectra expected in the LBNF project. To this end, we assume a
nominal beam power of 1.2 MW with $1.1 \times 10^{21}$ pot/year and a
data taking time of 5+5 years with the default low-energy neutrino and antineutrino beams~\cite{Petti2018,ESGprop}.
Table~\ref{tab:events} summarizes the total number of events expected for the various
topologies and targets. An interesting option available at LBNF is a
high-energy beam optimized to detect the $\nu_\tau$ appearance from neutrino
oscillations in the far detector, which would result in an increase by a factor 2.4 of the
$\nu(\bar \nu)$-H rates, combined with a much harder spectrum.
It is conceivable to have a dedicated 2 year run with this high energy
beam after the completion of the nominal data taking~\cite{Petti2018,ESGprop}.
By that time the planned upgrades of the beam intensity to a nominal power
of 2.4 MW would more than double the available statistics.
\section{Discussion}
\label{sec:disc}
\subsection{Systematic uncertainties}
\label{sec:syst}
The kinematic analysis described in Sec.~\ref{sec:kine} allows the identification
of all the various $\nu(\bar \nu)$-H CC topologies within the CH$_2$ target
with little residual backgrounds 5-20\% from interactions on the carbon
nucleus. This selection dramatically reduces not only the statistical uncertainty from the
background subtraction procedure (Sec.~\ref{sec:stat}), but also the impact of systematic
uncertainties on the modeling of nuclear effects in carbon~\cite{Alvarez-Ruso:2017oui}.
These latter are further suppressed by a model-independent background
subtraction using the data obtained from the dedicated graphite target,
which also provides a pure background sample to build the likelihood functions
used in the selection.
The CH$_2$ plastic can be considered as an effective tank filled with hydrogen,
while the graphite represents the empty tank. This technique is similar to what
is done in electron scattering experiments, in which a cryogenic tank
is filled with a liquid H$_2$ target and special runs with the empty tank are taken
for background subtraction~\footnote{One of the dominant systematic uncertainties in
electron scattering experiments on hydrogen is typically given by the knowledge of the
density/mass of the cryogenic target. Our technique for measuring $\nu(\bar \nu)$-H
interactions allows a more precise knowledge of the actual target mass.}.
The detector technology discussed in Sec.~\ref{sec:targets} is essential, since
the CH$_2$ and C targets are configured as thin layers spread out uniformly
over the tracking volume, ensuring that the corresponding acceptance
corrections are small and, most importantly, the same on average for both targets.
We verified this latter condition with detailed detector simulations using the
GEANT4 program~\cite{Agostinelli:2002hh}.
We emphasize that the data from the graphite target automatically
include all types of interactions, as well as reconstruction effects, relevant for our analysis.
The impact of possible model dependencies through the acceptance corrections is
therefore negligible, since they would appear as third order effects on
the data-driven subtraction of small backgrounds.
\begin{table}[tb]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|} \hline
& & \multicolumn{2}{|c|}{ NuWro}
& \multicolumn{2}{|c|}{GiBUU} & \multicolumn{2}{|c|}{GENIE} \\
Process & Selection & ~~Efficiency~~ & ~~Purity~~ & ~~Efficiency~~ & ~~Purity~~ & ~~Efficiency~~ & ~~Purity~~ \\ \hline\hline
$\nu_\mu p \to \mu^- p \pi^+$ & $\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$ & 96\% & 95\% & & & 96\% & 96\% \\
& $R_{mH}$ \& $p_{T \perp}^H$ & 93\% & 86\% & 93\% & 84\% & 93\% & 91\% \\
$\bar \nu_\mu p \to \mu^+ p \pi^-$ & $\ln \lambda^{\rm H}$+$\ln \lambda^{\rm H}_{\rm IN}$ & 94\% & 95\% & & & 94\% & 98\% \\
& $R_{mH}$ \& $p_{T \perp}^H$ & 89\% & 84\% & 89\% & 87\% & 89\% & 89\% \\
\hline
\end{tabular}
\caption{Comparison of the efficiency and purity for the kinematic selection of H
interactions from the CH$_2$ target
(Sec.~\ref{sec:kine}) with the NuWro~\cite{Juszczak:2005zs}, GiBUU~\cite{Buss:2011mx},
and GENIE~\cite{Andreopoulos:2009rq} event generators.
The same cuts on the likelihood ratios as in Tab.~\ref{tab:effpur} are used.
In all cases the likelihood functions are built from the default NuWro generator and
are not recalculated when different generators are tested. See text for details.
} \label{tab:effpurMC}
\end{center}
\end{table}
The kinematic selection of the signal from $\nu_\mu$-H and $\bar \nu_\mu$-H CC
interactions described in Sec.~\ref{sec:kine} relies upon the absence of
nuclear effects -- i.e. target proton at rest and no final state interactions --
rather than on the specific kinematics of the interactions with the free nucleon.
Uncertainties on these latter, including form factors and structure functions,
therefore can affect the resulting efficiencies only as second order
effects through the small acceptance corrections for the individual exclusive
topologies. Furthermore, uncertainties
related to the structure of the free nucleon would affect both signal and backgrounds
and largely cancel in a selection entirely based upon the differences introduced
by nuclear effects. We note that the relevant proton form factors and structure functions
can be determined in a model-independent way from the measured
$Q^2$ and $x$ distributions, using the procedure described
in details in Ref.~\cite{Duyang:2019prb}.
Although the technique we propose is designed to be model-independent,
in the absence of actual data from STT the efficiencies and purities
listed in Tab.~\ref{tab:effpur} are sensitive to the details of the interaction modeling
implemented in the simulations. In order to estimate the impact of
such effects, we repeat the event selection
described in Sec.~\ref{sec:kine}
with three event generators: NuWro~\cite{Juszczak:2005zs},
GiBUU~\cite{Buss:2011mx}, and GENIE~\cite{Andreopoulos:2009rq}.
These generators use rather different assumptions for the modeling of
(anti)neutrino-nucleus interactions -- including both initial and final state nuclear effects --
as outlined in Ref.~\cite{Mosel:2019vhx}.
To this end, we do not recalculate the likelihood functions but rather use the ones
obtained from the default NuWro generator throughout.
Given the differences among generators~\cite{Mosel:2019vhx,Duyang:2019prb},
this assumption helps to understand the impact of potential discrepancies
between data and simulations on the H selection.
As shown in Tab.~\ref{tab:effpurMC}, our kinematic selection of H interactions from the
CH$_2$ targets is rather stable and we obtain comparable efficiencies and purities with
all the three event generators. We emphasize that this test is only meant to
estimate the degree to which we can rely upon the numbers in Tab.~\ref{tab:effpur},
since the actual technique is entirely data-driven and all backgrounds and
efficiencies will be directly determined from the measured interactions.
Reconstruction effects on the four-momenta of the final state particles can in principle
degrade the kinematic selection. For this reason in our studies we used a realistic
detector smearing and checked its consistency with GEANT4 and FLUKA
simulations (Sec.~\ref{sec:framework}). Furthermore, we validated the effects of
the detector acceptance, smearing, and track reconstruction
with NOMAD data~\cite{Altegoer:1997gv}, although the NOMAD
detector had worse acceptance and granularity than the STT.
We note that similar kinematic selections were successfully
demonstrated by NOMAD~\footnote{The higher neutrino energy in NOMAD
implies higher backgrounds and more difficult kinematic selections for low multiplicity
processes like RES and QE compared to our case study. Furthermore, the lowest
usable energy in NOMAD was about 5 GeV, which is reasonably close to the
spectra expected at LBNF.}
in more severe background conditions (rejections up to $10^5$) in various published
analyses~\cite{Astier:2001yj,Naumov:2004wa,Astier:2000ng}, as well as in single track
measurements of $\bar \nu_\mu$ QE and inverse muon decay~\cite{Lyubushkin:2008pe}.
The momentum scale of charged particles is calibrated
with the mass peak of the large samples of reconstructed
$K_0 \to \pi^+\pi^-$ decays~\cite{Duyang:2019prb}.
In our study we assume the same energy scale uncertainty of 0.2\% achieved
by the NOMAD experiment using this technique~\cite{Wu:2007ab}.
We note that the STT would provide 25 time higher granularity
than NOMAD and about 40 times higher $K_0$
statistics~\cite{Duyang:2019prb}. Similarly, the proton identification and reconstruction
efficiency can be accurately calibrated with the large samples
of $\Lambda \to p \pi^-$ decays available~\cite{Duyang:2019prb}.
Both $\Lambda$ and $K_0$ decays can be
used to constrain the systematic uncertainty on the reconstruction of the track angles.
The use of a likelihood function in the kinematic analysis (Sec.~\ref{sec:kine}) provides
a simple way to vary the purity and efficiency of the selected samples (Fig.~\ref{fig:Hlike})
to validate the background subtraction and the selection efficiencies through appropriate
control samples.
\subsection{Unique option for \boldmath $\nu(\bar \nu)$-H CC interactions}
\label{sec:options}
The technique we propose does not have many practical alternatives for the
experimental detection of high statistics samples of $\nu(\bar \nu)$-H CC interactions.
The only technology ideally able to deliver more accurate measurements
(considering both statistics and resolution) would be
a dedicated magnetized liquid H$_2$ detector -- either a bubble chamber or a
time-projection chamber (TPC), surrounded by an electromagnetic calorimeter and a
muon identifier -- with a fiducial volume $> 10$ m$^3$.
However, modern safety requirements for underground experimental
halls and related costs make this option extremely challenging.
We note that even such liquid H$_2$ detectors cannot
provide the unique capability of STT to incorporate a wide range of nuclear targets to be
compared with H measurements within the same detector for the study of nuclear effects.
Our technique represents an acceptable approximation to pure H$_2$ targets,
since it can achieve relatively small dilution factors (Tab.~\ref{tab:effpur} and Tab.~\ref{tab:inclusive})
using a stringent kinematic selection.
Filling a large TPC with high pressure H$_2$ gas, besides
safety/cost concerns, would result in an insufficient statistics for any sensible physics
measurement~\footnote{A total TPC volume of about 2,500 m$^3$ filled with H$_2$ gas at
10 atm would be required to match the statistics achievable with our technique.}.
The use of a plastic scintillator target cannot deliver the precision required for the
detection of $\nu(\bar \nu)$-H interactions via subtraction with a graphite target.
In addition to the smaller hydrogen content (CH vs. CH$_2$), the coarser vertex, angular,
and momentum resolution achievable introduce larger systematic uncertainties
in the location of $\nu (\bar \nu)$-H interactions and higher backgrounds from the
kinematic selection (Fig.~\ref{fig:Hsel-Hstat}).
Furthermore, the many impurities from glue, fibers, coatings, etc.
make the subtraction of nuclear components problematic in relation to the small
H content of about 7\%.
\subsection{Physics Measurements}
\label{sec:apps}
The availability of high statistics samples of $\nu(\bar \nu)$-H CC interactions can
represent a significant advancement for neutrino scattering physics, as well as for
long-baseline oscillation experiments. In this section we outline some of the key physics
measurements~\cite{Duyang:2019prb,Petti2018,ESGprop}.
The limited knowledge of the (anti)neutrino flux has always been
a bottleneck for accelerator based neutrino experiments.
The exclusive $\nu_\mu p \to \mu^- p \pi^+$, $\bar \nu_\mu p \to \mu^+ p \pi^-$,
and $\bar \nu_\mu p \to \mu^+ n$ processes on hydrogen
with small energy transfer $\nu$ offer a measurement of relative fluxes as a
function of energy with little hadronic uncertainties.
To this end, the use of the samples described in Sec.~\ref{sec:stat} allows a determination of the
$\nu_\mu$ and $\bar \nu_\mu$ relative flux to a precision better than 1\%~\cite{Duyang:2019prb},
not achievable with other techniques using nuclear targets.
The measurement of the $\bar \nu_\mu p \to \mu^+ n$ QE interactions on H at small
momentum transfer $Q$ also provides an absolute measurement of the $\bar \nu_\mu$ flux,
since the corresponding cross-section in the limit $Q\to 0$ is a constant known
to high accuracy from neutron $\beta$ decay~\cite{Duyang:2019prb}.
A comparison of $\nu(\bar \nu)$-H CC interactions
with the corresponding ones from nuclear targets within the same detector provides a direct
measurement/constraint of nuclear effects, which typically result in a substantial
smearing of the observed interactions. This study can be performed with both inclusive
CC events and with various exclusive topologies. Constraining the nuclear
smearing from initial and final state interactions is required to reduce the
uncertainties in the unfolding of data collected from heavy targets and to calibrate
the reconstructed neutrino energy scale.
The possibility to control the (anti)neutrino target and the input fluxes to an
unprecedented accuracy~\cite{Duyang:2019prb} would open up a sensible
program of precision tests of
fundamental interactions~\cite{Petti2018,ESGprop}. An example is given by the
Adler sum rule~\cite{Adler:1964yx}, $S_A=\int^1_0 dx/x \left( F_2^{\bar \nu p} -
F_2^{\nu p} \right) = 2$, which is based upon current algebra and
was tested only by BEBC~\cite{Allasia:1985hw} with a few thousand events.
Similarly, by exploiting the isospin symmetry $F_2^{\nu n} = F_2^{\bar \nu p}$, we
can obtain a model-independent measurement of the free neutron structure
functions, as well as a measurement of the large $x$ behavior of the $d/u$ quark
ratio~\cite{Alekhin:2017fpf}. These measurements can also be used for
precision tests of the isospin (charge) symmetry and would help to elucidate the
flavor structure of the nucleon~\cite{Alekhin:2018dbs}.
\section{Summary}
\label{sec:sum}
We presented a novel approach to achieve precision measurements of
$\nu(\bar \nu)$-hydrogen CC interactions via subtraction between dedicated
CH$_2$ plastic and graphite (pure C) targets, embedded within a low
density tracker allowing a control of the configuration, chemical composition,
size, and mass of the targets similar to electron scattering experiments.
We used a kinematic selection to reduce backgrounds from the carbon nucleus in the CH$_2$
measurements to about 5-20\%, depending upon the specific event topology considered.
The measurement of this background is entirely data-driven by subtracting it using the
corresponding graphite target measurements.
The availability of large control samples of $\nu(\bar \nu)$-H CC interactions is a necessary
condition for any program of precision measurements with (anti)neutrino interactions
in order to constrain nuclear effects, as well as to achieve the accuracies required
by next-generation long-baseline oscillation experiments.
The simple and safe approach we proposed appears to be a viable alternative
to liquid H$_2$ detectors and a realistic opportunity
to obtain such high statistics samples, since safety and practical arguments
make other techniques unfeasible.
The use of exclusive $\nu(\bar \nu)$-H CC processes allows a determination of
$\nu_\mu$ and $\bar \nu_\mu$ relative fluxes in conventional wide band (anti)neutrino
beams to a precision better than 1\%, which cannot be achieved by other known techniques.
In combination with nuclear targets integrated within the same detector,
$\nu(\bar \nu)$-H CC interactions provide
a direct measurement of nuclear effects resulting in the smearing of the
observed $\nu(\bar \nu)$-nucleus interactions.
Finally, they represent the key to access a sensible program of precision tests
of fundamental interactions and of the structure of nucleons and nuclei.
\begin{acknowledgments}
The authors express their gratitude to the LBNE and DUNE experiments for the use of some
detector simulation tools. We thank L. Camilleri, R. Ent, X. Lu, and X. Qian for fruitful discussions.
We thank F. Ferraro, L. Di Noto, P. Sala, and M. Torti for providing the reconstruction
smearing of $n$ and $\pi^0$ with the FLUKA simulation package.
This work was supported by Grant No. DE-SC0010073 from the Department of Energy, USA.
\end{acknowledgments}
|
\section*{Introduction}
\noindent Biomolecules are the organic molecules present in the living organisms, including large macromolecules: proteins, carbohydrates, lipids and nucleic acids, and small molecules such as metabolites~\cite{wikibio}. Amino acids, important biomolecules~\cite{wikiamino}, are the organic compounds containing amine ($\--$NH$_2)$ and carboxyl ($\--$COOH) as
functional groups as well as hydrogen (H) or alkyl (R) as side chain. They are structural units of protein. Cysteine which is abbreviated as Cys or C is a proteinogenic sulfur-containing non-essential amino acid~\cite{wikicys}. Its linear formula is HOOCCH(NH$_2$)CH$_2$SH, which contains an alpha amino group, a carboxyl group, and a side chain consisting of a mercaptomethyl group. The IUPAC name of cysteine is 2-amino-3-mercaptopropanoic acid.
\noindent Cysteine is a white crystalline solid having molar mass 121.15 gram per mole and melting point is 513 K~\cite{wikicys}. Its solubility in water is 16 gram per 100 mL at 288 K~\cite{lcystop}. But cysteine exhibits hydrophobic nature, due to which it generally resides in the interior of proteins~\cite{biocys}. Cysteine is a non-essential amino acid as it is synthesized in the human body under the ordinary physiological state if there is adequate amount of methionine, which is also a sulfur-containing amino acid. By virtue of ability to form disulfide bonds (methionine can't form this bond), cysteine plays a crucial role in protein structure and in protein-folding mechanisms~\cite{wikicys}. Also cysteine is essential for the synthesis of highly anti-oxidative Glutathione which is important in the detoxification and protection of various tissues and organs in the body~\cite{aminolcys}. Further, cysteine supports in the absorption of nutrients from intestine, metabolism of lipids, enhancing the fertility, strengthening the immune system, preventing from the diseases like dementia, multiple sclerosis as well as Parkinson~\cite{aminolcys}, etc. It is also recognized as anti-aging amino acid. All these activities put the cysteine in special position that can not be substituted by any other amino acid.
\noindent The term transport phenomena means the process by which the mass, linear momentum, angular momentum, energy and charge are transferred from one part of the system to another due to non-uniformity or inhomogeneity in the system. Diffusion, an important transport property, is the phenomenon of transfer of mass as a result of random molecular motion. Various experimental techniques, viz., Peak-Height method~\cite{wenrui}, Nuclear Magnetic Resonance (NMR)~\cite{virk}, etc. as well as Molecular Dynamics (MD) simulations~\cite{virk} have been performed to study the diffusion phenomenon of amino acids in water. These studies were mainly concerned about the effect of concentration, polarity, temperature on the diffusion of amino acids. However, to the best of our knowledge, cysteine diffusion in water using MD simulation has not been performed yet.\\
\section*{Diffusion}
\noindent Diffusion is a process by which matter is transported from one part of system to another part as a result of random molecular motion
due to concentration gradient or thermal agitation without any external force and bulk
motion~\cite{J.crank}. Diffusion plays many important roles in non-living substances as well as living organisms. The diffusion in a homogeneous system having no chemical concentration gradient is called self-diffusion and the corresponding diffusion coefficient is termed as self-diffusion coefficient~\cite{H.hira}. In order to calculate the self-diffusion coefficient one of the most common method is exploiting the Einstein's equation which relates diffusion coefficient with mean square displacement (MSD) of the particles as;
\begin{equation}
\label{selfdiff}
D = \lim_{t \to \infty} \frac{\langle \lbrack r(t) - r(0) \rbrack ^2 \rangle}{6 t}.
\end{equation}
In equation (\ref{selfdiff}), $r(t) - r(0)$ is the displacement of particle from reference point during the course of time $t$, $ \lbrack r(t) - r(0) \rbrack ^2$ is the square of displacement and $\langle ... \rangle$ represents the ensemble average and hence ${\langle \lbrack r(t) - r(0) \rbrack ^2 \rangle}$ gives MSD of particle.\\
\noindent Binary diffusion is the diffusion of two different substances in their binary mixture and the resulting diffusion coefficient is called binary diffusion coefficient. The Darken's relation which is used to calculate the binary diffusion coefficient is given by
\cite{lsdarken},
\begin{equation}
\label{darkeneq}
D_{12}=N_2D_1+N_1D_2.
\end{equation}
In equation (\ref{darkeneq}), $D_{12}$ is the binary diffusion coefficient, $D_1$ and $D_2$ are the self-diffusion coefficients of substances 1 and 2 respectively, and $N_1$ and $N_2$ are the corresponding mole fractions.
\section*{Computational Details}
\subsection*{Modeling of System}
\noindent The one of the best simulation technique to study the dynamic properties of substance at molecular level is molecular dynamics. Here we provide the initial positions and velocities of each particles; the new positions of the particles at each time step can be obtained using Leapfrog algorithm in the simulation. Hence, molecular simulation can provide the trajectories of particles which are utilized to find the equilibrium and transport properties of substances~\cite{smith}. For $N$ interacting particles in the classical MD simulation Newton's equation of motion given in equation (\ref{newtoneq}) is solved to find the new position of the particles.
\begin{equation}
\label{newtoneq}
m_i\frac{\partial^2\textbf{r}_i}{\partial t^2} = -\nabla_i U(r) = \textbf{F}_i.
\end{equation}
In equation (\ref{newtoneq}), $m_i$ is the mass and $r_i$ is the position of $i^{th}$ particle. In RHS of this equation, the force on $i^{th}$ particle is equal to negative gradient of the potential.
\noindent We start the simulation work through modeling of system where we assign the molecules in terms of mass, charge, bond, angle, dihedral, etc. All these essential parameters are specified in selected force field, which is OPLS-AA in our case. The atoms in the system is treated as spherically symmetric potentials in classical force fields. The force between each particle is calculated from pairwise additive potential functions~\cite{allen}. The total potential energy $U_{total}$ of the system is due to contributions from bonded interactions and non-bonded interactions between individual atoms/molecules of the system. The bonded interactions are classified as bond stretching, bond angles, dihedral angle interactions. Lennard-Jones interactions and Coulomb interactions come under the category of non-bonded interactions~\cite{gromanual}. Hence, the total potential energy function in our system is written as,
\begin{equation}
\label{totalpotential}
U_{total} = U_{bond} + U_{angle} + U_{proper} + U_{LJ} + U_{Coulomb}.
\end{equation}
\subsection*{Simulation Set Up}
\noindent We have simulated the 3 cysteine and 1039 water molecules at five different temperatures. Cysteine is three carbon containing alpha-amino acid. There are 14 atoms in a single molecule of cysteine. Different atoms possesses different partial charge due to difference in electro-negativity. Also the same atom possesses different partial charge based on the group of attachment. The Coulomb interaction occurs due to the partial charge exist in atoms/molecules. Likewise, the Vander Waal's interaction occurs as a result of induced dipole interaction. The parameters for non-bonded interactions: partial charge, sigma and epsilon as well as for bonded interactions: bond lengths, angles and dihedrals, are assigned in the OPLS-AA force field by default. In case of water, the required force field parameters are specified in water model. We have used SPC/E water model in our simulation. In this model each hydrogen atom of water is specified with a partial charge of +0.4238 e, and the oxygen atom is assigned with the partial charge of $\--$0.8476 e, where e is the electronic charge whose magnitude is $1.6022 \times 10^{-19}$ Coulomb. All the parameters for SPC/E water model are included in the file \textit{spce.itp} inherent to GROMACS.
\begin{table} [H]
\centering
\label{forcefieldforH2O}
\caption{Force field parameters for SPC/E water model.}
\begin{tabular}{|c|c|} \hline
Parameters &Values\\
\hline
$K_{OH}$ & 3.45 \ $\mathrm{x}$ \ 10$^5$ \ kJ mol$^{-1}$ nm$^{-2}$ \\
$b_{OH} $ & 0.1 nm \\
$K_{HOH}$ & 3.83 \ $\mathrm{x}$ \ 10$^2$ \ kJ mol$^{-1}$ rad$^{-2}$\\
$\Theta_o $ & $ 109.47^0$ \\
\hline
\end{tabular}
\end{table}
\noindent The non-bonded parameters are listed in the file \textit{nonbonded.itp} which are presented below:
\resizebox{0.47\textwidth}{!}{\begin{tabular}{ccccccc}
[\textbf{atomtypes}]&& &&&&\\
atoms& at.num & mass & charge & ptype & sigma & epsilon \\
N & 7 & 14.0067 & $\--$0.90 & A & 3.3000e-01 & 7.1128e-01\\
CA & 6 & 12.0110 & 0.12 & A & 3.5000e-01 & 2.7614e-01\\
C& 6 & 12.0110 & 0.52 & A& 3.75000e-01 & 4.3932e-01\\
O & 8& 15.9994& $\--$0.53& A & 3.0000e-01 & 7.1128e-01 \\
CB & 6 & 12.0110& 0.06 & A& 3.5000e-01 & 2.7614e-01\\
SG & 16&32.0600 &$\--$0.45& A &3.5500e-01 &1.0460e+00\\
H1 & 1 &1.0080 & 0.36 & A &0.0000e+00 &0.0000e+00 \\
H2 & 1 &1.0080 & 0.36 & A &0.0000e+00 &0.0000e+00 \\
HA&1&1.0080 & 0.06&A&2.5000e-01 & 1.2552e-01 \\
HO & 1 &1.0080 & 0.45 & A &0.0000e+00 &0.0000e+00 \\
HB1&1&1.0080 & 0.06&A&2.5000e-01 & 1.25520e-01\\
HB2&1&1.0080 & 0.06&A&2.5000e-01 & 1.2552e-01\\
HG & 1 &1.0080 & 0.27 & A &0.0000e+00 &0.0000e+00 \\
OC & 8& 15.9994& $\--$0.44& A & 2.9600e-01 & 8.7864e-01 \\
OW&8&15.9994&$\--$0.82 & A &3.1656e-01 & 6.5019e-01 \\
HW1&1&1.0080 &0.41 & A &0.0000e+00& 0.00000e+00\\
HW2&1&1.0080 &0.41 & A &0.0000e+00& 0.0000e+00\\
\end{tabular}}\\
\noindent In this table, first column represents name of atoms. Second, third and fourth columns respectively give the atomic number, atomic mass (in atomic mass unit) and partial charge (in e) of respective atoms. Sixth column specifies particle type, where A stands for atom. Seventh and eighth column give the sigma and epsilon parameters of corresponding atom respectively. For dissimilar atoms pair, the values of sigma and epsilon are calculated by using following relations in OPLS force field \cite{gromanual}.
\begin{equation}
\sigma_{ij} = (\sigma_{ii} \sigma_{jj})^{1/2}, \quad
\epsilon_{ij} = (\epsilon_{ii} \epsilon_{jj})^{1/2}.
\label{ljconversion}
\end{equation}
Once after enclosing three cysteine molecule in cubic simulation box, we add 1039 water molecules in the box to form cysteine-water system. We employ periodic boundary condition which eliminates the surface effect. After solvation of cysteine in water, system is subjected to the process of energy minimization; which brought the system to one of a local minima where system possesses minimum energy. In GROMACS,
we used the steepest-descent algorithm for energy minimization.
\begin{figure}[H]
\centering
\includegraphics[scale=0.48]{potential.png}
\caption{Potential energy after energy minimization.}
\label{potentialenergy}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.35]{safterem.png}
\caption{System after energy minimization.}
\label{fig:emafter}
\end{figure}
\noindent Figure (\ref{potentialenergy}) is the energy minimization curve for our system. The energy of the system is said to be minimized if the maximum force experienced in the system is less than \textit{emtol}, which is 50 kJ mol$^{-1}$ nm$^{-1}$. Also for a stable system total energy of the system should be negative. In our system we got that:\\
\textbf{Potential energy} = $\--$5.6946953 $\times$ 10$^{4}$ kJ mol$^{-1}$ and steepest descent converged to $ {F_{max}}$ $ < 50$ kJ mol$^{-1}$ nm$^{-1}$ after 6713 steps.
\noindent Since the total potential energy is negative and nearly constant, our system is stable. Figure (\ref{fig:emafter}) shows our system after energy minimization.
\noindent After energy minimization, the system is ready to study the dynamic properties. However, the dynamical properties such as diffusion, viscosity usually vary with the parameters like temperature, pressure, density, etc. Therefore, before starting the production run (where the physical properties will be calculated) these fore-mentioned parameters should be kept constant during simulation and the system under study has to brought in the state of thermodynamic equilibrium. The process of bringing the system in thermodynamic equilibrium state is called equilibration or npt run. The temperature in the system is kept constant (thermal equilibrium) by using thermostat and the pressure is kept constant by using barostat. In GROMACS, the process of adjusting temperature and pressure to the desired values are respectively termed as \textit{temperature coupling} and \textit{pressure coupling}.
\noindent In the equilibration parameter file i.e. \textbf{equilibration.mdp}, we have used \textit{md integrator}, which is leapfrog algorithm, with step size of 0.001 ps for 5 $\times$ 10$^7$ steps in all runs. This results equilibration time 50 ns. We have used PME (Particle-mesh Ewald) for long-range Coulomb interaction with Fourier spacing of 0.12 nm and cut-off distance of 1.0 nm. V-rescaling have been used for \textit{temperature coupling} with time constant of 0.01 ps. Run is carried out at five different temperatures: 288 K, 293 K, 303 K, 313 K and 323 K. \textit{Pressure coupling} have been done through Barendsen barostat to reference pressure of 1 bar. The \textit{pressure coupling} time constant and isothermal compressibility have been respectively taken as 0.8 ps and 4.6 $\times$ 10$^{-5}$ bar$^{-1}$. We have applied the LINCS constraint algorithm to convert all bonds into constrains during equilibration run. LINCS algorithm is default in GROMACS~\cite{gromanual} if there is not specified any algorithm explicitly.
\begin{table}[H]
\caption{Temperatures and densities of system after equilibration run.}
\label{table:npt temp and dens}
\resizebox{0.475\textwidth}{!} {\begin{tabular}{|c|c|c|c|c|}
\hline
Coupling & Equilibrium & Equilibrium & Estimated \\
Temperature (K) & Temperature (K) & Density (kg m$^{-3}$) & Density (kg m$^{-3}$)~\cite{waterdensity} \\
\hline
288 & 287.992 $\pm$ 0.005 & 1001.120 $\pm$ 0.036 & 999.102 \\
\hline
293 & 292.988 $\pm$ 0.005 & 998.586 $\pm$ 0.041 & 998.207 \\
\hline
303 & 302.992 $\pm$ 0.008 & 993.397 $\pm$ 0.031 & 995.650 \\
\hline
313 & 312.994 $\pm$ 0.006 & 987.487 $\pm$ 0.061 & 992.219 \\
\hline
323 & 322.988 $\pm$ 0.012 & 980.920 $\pm$ 0.057 & 988.039 \\
\hline
\end{tabular}}
\end{table}
\noindent Table (\ref{table:npt temp and dens}) listed out the equilibration temperatures and densities of the system at different reference or coupling temperatures. The equilibrium temperature is almost equal to reference temperature. As our system is very dilute we have compared density of system with that of water at different temperatures. We have found that the densities of system are consistent with that of water within error less than 1\%.
\noindent After finishing the equilibration run, the system is ready for production run or the simulation of NVT ensemble. In this phase of simulation, we determined the diffusion coefficient. Basically \textbf{nvt.mdp} file is similar to the parameter file of equilibration run. However, in nvt run we did not apply the \textit{pressure coupling} since we are executing NVT ensemble. Further it is not required to generate the initial velocities in nvt run as the simulation continues with the velocities generated in equilibration run.
\section*{Results and Discussion}
\noindent In this section, we present the energy profile of system obtained after the production run, structural analysis and transport property-diffusion coefficients of constituents of the system at different temperatures.
\section*{Energy Profile}
\noindent Different interactions exist in the system result various types of energies. The manifestation of all the energies present in the system caused by different interactions constitute the energy profile of the system. The interactions associated with bond stretches, angle vibration, dihedral rotations are categorized as bonded interactions and the energies corresponding to these interactions are termed as bonded interaction energies. The energy terms associated with non-bonded interactions are Coulomb potentials and LJ potentials. Both bonded and non-bonded interaction energies contribute to the total potential energy of the system and on addition of total potential energy and kinetic energy finally yield total energy of the system at particular temperature.
\begin{figure}[H]
\centering
\includegraphics[scale=0.35]{enprofile288mod.png}
\caption{Energy profile of system at 288 K.}
\label{enprof288}
\end{figure}
\noindent Figure (\ref{enprof288}) shows the energy profile of the system at 288 K. This energy profile portrays that the Coulomb and LJ interactions are the prominent interactions contributing to the total potential energy. The energy values due to bonded interactions (angle vibrations and R-B dihedrals) and due to pair interactions (LJ-14 and Coulomb-14) are very little. These values almost coincide to zero with respect to other energy term values in the graph. Therefore, the contribution to total potential energy from these interactions is negligible. The magnitude of Coulomb potential energy is very large with negative sign. This drives the total potential energy to be negative. Similarly, large negative value of potential energy and small positive value of kinetic energy lead the system to acquire negative total energy after production run. This implies our system is bound and stable. Also the relationship between kinetic energy and temperature of the system is shown via table (\ref{table:ke and t}).
\begin{table}[H]
\caption{Relationship between kinetic energy and temperature.}
\label{table:ke and t}
\resizebox {0.48 \textwidth }{!} {\begin{tabular}{|c|c|c|c|}
\hline
SN & Temperature (T) & Kinetic Energy (K.E.) & $\frac{K.E.}{T}$ \\
& (K) & (kJ mol$^{-1}$) & (kJ mol$^{-1}$ K$^{-1}$)\\
\hline
1 & 288 & 7564.40 & 26.265 \\
\hline
2 & 293 & 7695.65 & 26.265 \\
\hline
3 & 303 & 7958.39 & 26.265 \\
\hline
4 & 313 & 8221.05 & 26.265 \\
\hline
5 & 323 & 8483.59 & 26.265 \\
\hline
\end{tabular}}
\end{table}
\noindent In the table (\ref{table:ke and t}), we have calculated the ratios of kinetic energies obtained after production run and the corresponding absolute temperatures. All these ratios have given the same value 26.265 kJ mol$^{-1}$ K$^{-1}$. That is,
$\frac{K.E.}{T}$ = a constant\\
\noindent Or, K.E. = a constant $\times$ T\\
\noindent This implies kinetic energy is proportional to the absolute temperature.
\section*{Structure of System}
\noindent The radial distribution functions, RDF's, of pair of atoms are used for the study of structure of system. From the analysis of distribution of neighboring atoms around a reference atom, structural properties of the system can be determined. The effect of temperature in the structure of system is also studied by analyzing RDF's at five different temperatures.
\subsection*{Structural Analysis of Solvent}
\noindent We have used normal water as solvent. Since the LJ-parameters of hydrogen atoms of water are zero, therefore, we study the RDF of oxygen atom of water for the structural analysis. Here we have discussed the RDF of two oxygen atoms of water molecules at five different temperatures. The RDF plots depicted that how the atoms are organized from reference atom.
\begin{figure}[H]
\centering
\includegraphics[scale=0.33]{rdfowcomb.png}
\caption{RDF plot of oxygen-oxygen of water molecules, $g_{OW\--OW}$(r), at different
temperatures.}
\label{rdfow}
\end{figure}
\begin{table}[H]
\centering
\caption{Details of RDF of oxygen-oxygen atoms of water at different temperatures.}
\label{tablerdfow}
\resizebox {0.475 \textwidth }{!}{%
\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline
\multicolumn{8}{|c|}{RDF analysis of OW$\--$OW atoms }\\ \hline
T (K) & ER (nm) & FPP (nm) & FPV & SPP (nm) & SPV & TPP (nm) & TPV \\ \hline
288 & 0.240 & 0.274 & 3.230 & 0.450 & 1.140 & 0.682 & 1.049\\
\hline
293 & 0.240 & 0.274 & 3.175 & 0.450 & 1.129 & 0.680 & 1.048\\
\hline
303 & 0.240 & 0.274 & 3.077 & 0.450 & 1.110 & 0.686 & 1.043\\
\hline
313 & 0.240 & 0.276 & 3.000 & 0.450 & 1.096 & 0.690 & 1.041\\
\hline
323 & 0.240 & 0.276 & 2.945 & 0.450 & 1.091 & 0.686 & 1.037\\
\hline
\end{tabular}}
\end{table}
\noindent In the figure (\ref{rdfow}), there are three distinct peaks. The first peak, which is located at the separation of about 0.27 nm from centered atom's position, is highest and sharpest. This implies that at this position maximum number of oxygen atoms are clustered from the reference oxygen atom. In other words, the probability of finding oxygen atoms at the first peak position is highest. This is the most preferable position or minimum energy position from the centered atom. From the figure (\ref{ljowow}) of LJ potential of OW$\--$OW of water, we fond the $\sigma$ value of water oxygen is 0.316557 nm and the Vander Waal's radius is 2$^\frac{1}{6} \sigma \approx$ 0.36 nm. However, the FPP in our case is 0.27 nm less than 0.36 nm. This reveals the fact that there is not only LJ interaction between oxygen atoms of water but also other interactions such as Coulomb and bonded interactions are present.
\begin{figure}[H]
\centering
\includegraphics[scale=0.37]{ljowow.png}
\caption{LJ potential plot as a function of distance for OW$\--$OW of water.}
\label{ljowow}
\end{figure}
\noindent The second and third peaks are relatively shorter and wider, which are located approximately at positions 0.45 nm and 0.68 nm respectively. Excluded region, in which RDF is zero, has extended up to 0.24 nm from the center of reference oxygen atom. Any other oxygen atom can not exist within excluded region due to strong repulsive forces: namely $r^{-12}$ term of LJ interaction and repulsive Coulomb interactions~\cite{ishworpaper}. We have also studied the effect of temperature on RDF. Figure (\ref{rdfow}) shows that with increase in temperature the peak positions are shifted to right, heights of peaks are decreased and widths are increased. This reflects that our system have became less organized with increase in temperature. The increase in thermal agitation of atoms in the system with rising temperature accounts this fact. Further beyond third peak graph is straight line possessing unit value on average. This indicates there is no pair correlation of oxygen atoms.
\subsection*{RDF of OC and OW}
\noindent Here OC refers to carbonyl oxygen atom of cysteine and OW means oxygen atom of water. The RDF of OC and OW, $g_{OC\--OW}(r)$, gives the insight about how the carbonyl oxygen atoms of cysteine organized around the oxygen atom of water.
\begin{figure}[H]
\centering
\includegraphics[scale=0.33]{rdfoccomb.png}
\caption{RDF plot of carbonyl oxygen of cysteine and oxygen of water, $g_{OC\--OW}$(r), at different
temperatures.}
\label{rdfoc1}
\end{figure}
\begin{table}[H]
\centering
\caption{Details of RDF of carbonyl oxygen of cysteine and oxygen of water at different temperatures.}
\label{tablerdfoc}
\resizebox {0.475 \textwidth }{!}{%
\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline
\multicolumn{8}{|c|}{RDF analysis of OC$\--$OW atoms }\\ \hline
T (K) & ER (nm) & FPP (nm) & FPV & SPP (nm) & SPV & TPP (nm) & TPV \\ \hline
288 & 0.240 & 0.336 & 1.125 & 0.584 & 1.018 & 0.746 & 1.020\\
\hline
293 & 0.242 & 0.338 & 1.094 & 0.590 & 1.020 & 0.778 & 1.024\\
\hline
303 & 0.242 & 0.336 & 1.185 & 0.578 & 1.031 & 0.762 & 1.021\\
\hline
313 & 0.242 & 0.336 & 1.092 & 0.580 & 1.026 & 0.768 & 1.022\\
\hline
323 & 0.242 & 0.336 & 0.985 & 0.586 & 1.003 & 0.766 & 1.020\\
\hline
\end{tabular}}
\end{table}
\noindent Figure (\ref{rdfoc1}) has two distinct peaks (first and second) and third peak is not so sharp. The first peak, which is located at the separation of about 0.33 nm from the position of reference oxygen atom of water, is highest and sharpest. This implies that at this position maximum number of carbonyl oxygen atoms of cysteine clustered from the reference oxygen atom. Therefore, this is the most preferred position of carbonyl oxygen atoms to cluster around the oxygen atom of water. The second and third peaks are relatively shorter and wider, which are located approximately at the positions 0.58 nm and 0.76 nm respectively. Excluded region extends up to 0.24 nm from the center of reference oxygen atom. There is not possible to find any carbonyl oxygen within excluded region due to strong repulsive forces. Figure (\ref{rdfoc1}) shows with increase in temperature the peak positions are shifted to right, heights of peaks are decreased and widths are increased. This reflects that the less organization of carbonyl oxygen atoms around the reference oxygen atom with increase in temperature. Further beyond third peak graph is straight line possessing unit value on average. This indicates there is no pair correlation of carbonyl oxygen atoms and the reference oxygen atom of water.
\subsection*{RDF of HO and OW}
\noindent Here HO refers to hydroxyl hydrogen atom of carboxyl group of cysteine and OW means oxygen atom of water. The RDF of HO and OW, $g_{HO\--OW}(r)$, gives the insight about how the hydroxyl hydrogen atoms of cysteine organized around the oxygen atom of water.
\begin{figure}[H]
\centering
\includegraphics[scale=0.33]{rdfhocomb.png}
\caption{RDF plot of hydroxyl hydrogen of cysteine and oxygen of water, $g_{HO\--OW}$(r), at different
temperatures.}
\label{rdfho}
\end{figure}
\begin{table}[H]
\centering
\caption{Details of RDF of hydroxyl hydrogen of cysteine and oxygen of water at different temperatures.}
\label{tablerdfho}
\resizebox {0.475 \textwidth }{!} {%
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\multicolumn{6}{|c|}{RDF analysis of HO$\--$OW atoms }\\ \hline
T (K) & ER (nm) & FPP (nm) & FPV & SPP (nm) & SPV \\
\hline
288 & 0.142 & 0.172 & 2.198 & 0.386 & 1.055 \\
\hline
293 & 0.140 & 0.172 & 2.043 & 0.378 & 1.043 \\
\hline
303 & 0.140 & 0.172 & 2.468 & 0.382 & 1.061 \\
\hline
313 & 0.140 & 0.172 & 2.055 & 0.380 & 1.030 \\
\hline
323 & 0.140 & 0.172 & 1.640 & 0.380 & 0.995\\
\hline
\end{tabular}}
\end{table}
\noindent Figure (\ref{rdfho}) has two distinct peaks. The first peak, which is located at the separation of about 0.17 nm from the position of reference oxygen atom of water, is higher and sharper. This implies that at this position maximum number of hydroxyl hydrogen atoms of cysteine clustered from the reference oxygen atom. Therefore, this is the most preferred position of hydroxyl hydrogen atoms to cluster around the oxygen atom of water. The second peak is shorter and wider than first one, which is located approximately at the position 0.38 nm. Excluded region extends up to 0.14 nm from the center of reference oxygen atom. There is not possible to find any hydroxyl hydrogen within excluded region due to strong repulsive forces. Figure (\ref{rdfho}) shows with increase in temperature the peak positions are shifted to right, heights of peaks are decreased and widths are increased. This reflects that the less organization of hydroxyl hydrogen atoms around the reference oxygen atom with increase in temperature. Further beyond second peak graph is straight line possessing unit value on average. This indicates there is no pair correlation of hydroxyl hydrogen atoms and the reference water oxygen atom.
\section*{Diffusion Coefficients}
\subsection*{Self-Diffusion Coefficient of Cysteine}
\noindent The self-diffusion coefficients of cysteine have been calculated for five different temperatures by using corresponding MSD curves. We have determined the self-diffusion coefficient by dividing the slope of linear nature of MSD plot by six according to Einstein's equation (\ref{selfdiff}). In order to get the linear nature of MSD curve, we have plotted the MSD curves for 3 ns for all temperatures even though the production run was done for 50 ns.
\begin{figure}[H]
\centering
\includegraphics[scale=0.32]{msdunkcomb.png}
\caption{MSD vs time plot of cysteine at different temperatures.}
\label{msdunkcomb}
\end{figure}
\noindent Figure (\ref{msdunkcomb}) shows the MSD versus time plot at different temperatures. From this figure, it is seen that as the temperature increased slope of the MSD curves have also been increased which in turn increased the self-diffusion coefficient. This is reflected in the following table.
\begin{table}[H]
\caption{Simulated values of self-diffusion coefficient of cysteine at different temperatures.}
\label{table:cys_self_diff}
\resizebox {0.48 \textwidth } {!} { \begin{tabular}{|c|c|}
\hline
Temperature (K) & Self-Diffusion Coefficient $D_C^{sim}$ ( $10^{-10}$ m$^2$/s)\\
\hline
288 & 8.17 $\pm$ 1.10 \\
\hline
293 & 9.01 $\pm$ 0.27\\
\hline
303 & 10.81 $\pm$ 0.41 \\
\hline
313 & 13.78 $\pm$ 2.79 \\
\hline
323 & 16.29 $\pm$ 0.51\\
\hline
\end{tabular}}
\end{table}
\subsection*{Self-Diffusion Coefficient of Water}
\noindent As previously for cysteine, we have studied MSD plots and hence calculated self-diffusion coefficient of water at different temperatures using the same approach.
\begin{figure}[H]
\centering
\includegraphics[scale=0.32]{msdwatercomb.png}
\caption{MSD vs time plot of water at different temperatures.}
\label{msdwatercomb}
\end{figure}
\noindent Figure (\ref{msdwatercomb}) shows the MSD versus time plot at different temperatures. From this figure, it is seen that as the temperature increased slope of the MSD curves have also been increased which in turn increased the self-diffusion coefficient. This is reflected in the following table.
\begin{table}[H]
\caption{Simulated and experimental values of self-diffusion coefficients of water at different temperatures.}
\label{table:wat_self_diff}
\resizebox {0.48 \textwidth } {!} { \begin{tabular}{|c|c|c|c|}
\hline
Temperature (K) & \multicolumn{3}{|c|}{Self-Diffusion Coefficient ($10^{-10}$ m$^2$/s)}\\
\hline
& Simulated & Experimental~\cite{manfred} & Error \\
& ($D_W^{sim}$) & ($D_W^{exp}$) & (\%)\\
\hline
288 & 19.63 $\pm$ 0.01 & 17.66 & 11.16 \\
\hline
293 & 22.12 $\pm$ 0.14 & 20.25 & 9.23 \\
\hline
303 & 27.38 $\pm$ 0.03 & 25.97 & 5.43\\
\hline
313 & 32.95 $\pm$ 0.20 & 32.22 & 2.26\\
\hline
323 & 40.19 $\pm$ 0.05 & 39.83 & 0.90\\
\hline
\end{tabular}}
\end{table}
\noindent Table (\ref{table:wat_self_diff}) demonstrates that the simulated values of self-diffusion coefficients are in agreement with that of experimental values within 12\% error.
\subsection*{Binary Diffusion Coefficient of Cysteine-water System}
\noindent The self-diffusion coefficient of cysteine and water at particular temperature obtained in previous sections are now used for the calculation of binary diffusion coefficient by means of Darken's relation (\ref{darkeneq}). We have simulated 3 cysteine molecules and 1039 water molecules; 1042 molecules in total. Thus the mole fraction of cysteine is 0.003 and that of water is 0.997. The values of binary diffusion coefficients obtained after calculation and the experimental value are presented in table (\ref{table:binary_diff}).
\begin{table}[H]
\caption{Simulated and experimental values of binary diffusion coefficients at different temperatures.}
\label{table:binary_diff}
\resizebox {0.48 \textwidth } {!} { \begin{tabular}{|c|c|c|c|}
\hline
Temperature (K) & \multicolumn{3}{|c|}{Binary Diffusion Coefficient ($10^{-10}$ m$^2$/s)}\\
\hline
& Simulated & Experimental~\cite{wenrui} & Error \\
& ($D_B^{sim}$) & ($D_B^{exp}$) & (\%)\\
\hline
288 & 8.20 & 7.90 & 3.80 \\
\hline
293 & 9.05 & - & - \\
\hline
303 & 10.86 & - & -\\
\hline
313 & 13.84 & - & -\\
\hline
323 & 16.36 & - & - \\
\hline
\end{tabular}}
\end{table}
\noindent From the table (\ref{table:binary_diff}), the comparison of calculated value of binary diffusion coefficient of cysteine in water with that of experimental value available at 288 K in reference~\cite{wenrui} shows an agreement within error of 4\%. To the best of our knowledge, the experimental values of binary diffusion coefficients of cysteine in water at other temperatures are not found till date. However, the calculated values of diffusion coefficients have been increased with increase in temperature. This is because as the temperature increases thermal energy of molecules also increases but density of system decreases; which in turn increases the available space for diffusion. Thus molecular movement in the system become easier. This consequently increases the diffusion coefficient at higher temperatures. By this we can claim that the simulated values of binary diffusion coefficients at higher temperatures are also reasonably good.
\section*{Temperature Dependency of Diffusion}
\noindent As observed in table (\ref{table:binary_diff}) the diffusion phenomenon is strongly dependent on temperature. This temperature dependent behavior of diffusion is given by Arrhenius equation~\cite{udalpaper}:
\begin{equation}
\label{arr equation1}
D = D_0 ~ e^{\frac{\-- E_a}{N_Ak_BT}}.
\end{equation}
In equation (\ref{arr equation1}), $D$ is the diffusion coefficient, $D_0$ represents pre-exponential factor, $E_a$ is the activation energy for diffusion, $N_A$ is Avogardo's number whose value is 6.022 $\times$ 10$^{23}$ mol$^{-1}$, $k_B$ is the Boltzmann's constant whose value is 1.38 $\times$ 10$^{-23}$ JK$^{-1}$~\cite{udalpaper} and $T$ is the absolute temperature.
\noindent On taking natural logarithm in equation (\ref{arr equation1}) we get,
\begin{equation}
\label{arr equation2}
\ln D = \ln D_0 - \frac{E_a}{N_A k_B T}.
\end{equation}
\noindent We obtain the activation energy $E_a$ for diffusion from the slope of $\ln D$ versus ($\frac{1}{T}$) plot (Arrhenius plot) as,
\begin{equation}
E_a = -N_A~k_B \frac{\partial \mathrm{ln~D}}{\partial (1/T)}.
\end{equation}
\noindent The intercept when extrapolated to the 1/$T$ $\rightarrow 0$ in the Arrhenius plot gives the pre-exponential factor.\\
\begin{figure}[H]
\centering
\includegraphics[scale=0.305]{arrplotcyssim.png}
\caption{Arrhenius diagram for self-diffusion coefficients of cysteine.}
\label{arrplotcyssim}
\end{figure}
\noindent Figure (\ref{arrplotcyssim}) depicts the Arrhenius plot of simulated values of self-diffusion of cysteine. The activation energy for self-diffusion of cysteine in water calculated using the slope of linear fit of simulated values is found to be 15.49 kJ mol$^{-1}$.
\begin{figure}[H]
\centering
\includegraphics[scale=0.305]{arrplotwatsimexpt.png}
\caption{Arrhenius diagram for self-diffusion coefficients of water.}
\label{arrplotwatsimexpt}
\end{figure}
\noindent Figure (\ref{arrplotwatsimexpt}) portrays the Arrhenius plot of both the simulated and experimental values of self-diffusion of water.The activation energies for self-diffusion of water calculated using the corresponding slope of linear fit of simulated values and experimental values are found to be 15.67 kJ mol$^{-1}$ and 17.88 kJ mol$^{-1}$ respectively. This diagram (\ref{arrplotwatsimexpt}) also shows that the simulated and experimental values of diffusion coefficient are in excellent agreements specially at higher temperatures.
\begin{figure}[H]
\centering
\includegraphics[scale=0.305]{arrplotcysbinsim.png}
\caption{Arrhenius diagram for binary diffusion coefficients of binary mixture cysteine and water.}
\label{arrplotcysbinsim}
\end{figure}
\noindent Figure (\ref{arrplotcysbinsim}) displays the Arrhenius plot of simulated values of binary-diffusion of cysteine in water. The activation energy for binary diffusion of cysteine in water calculated using the slope of linear fit of simulated values is found to be 15.50 kJ mol$^{-1}$.\\
\noindent Figures (\ref{arrplotcyssim}), (\ref{arrplotwatsimexpt}) and (\ref{arrplotcysbinsim}) demonstrate the temperature dependency of diffusion. From these plots it is seen that the diffusion coefficients have increased with temperature. We have calculated the activation energies for diffusion of cysteine, water and their binary mixture by using the slopes of respective Arrhenius plots which are tabulated below.
\begin{table}[H]
\centering
\caption{Activation Energies for Diffusions.}
\label{table:activationenergy}
\begin{tabular}{|c|c|c|c|}
\hline
System & \multicolumn{3}{|c|} {Activation Energy ($E_a$) (kJ mol$^{-1}$)} \\
\hline
& Simulated & Experimental& Error (\%)\\
\hline
Water & 15.67 & 17.88~\cite{manfred} & 12.36 \\
\hline
Cysteine & 15.49 & & \\
\hline
Binary mixture & 15.50 & & \\
\hline
\end{tabular}
\end{table}
\noindent From this table (\ref{table:activationenergy}), it is observed that the activation energies for self-diffusion of cysteine and for the binary diffusion of cysteine in water are almost same. This implies that the concentration of cysteine in the system is infinitesimal. Further the activation energy calculated for simulated and experimental values of self-diffusion of water are in agreement within the error of 13\%.
\section*{Conclusions and Concluding Remarks}
\noindent In the present work, we performed the molecular dynamics study of transport property- diffusion of 3 cysteine molecules in 1039 SPC/E water molecules at 288 K, 293 K, 303 K, 313 K and 323 K temperatures using GROMACS 4.6.5 package. We used OPLS-AA force field parameters throughout the simulation. The system under study was confined in a cubic box having one side 3.2 nm after making topology file. To account the short-range interactions cut-off distance of 1.0 nm was used where as PME was employed for long-range interactions. All bonds were kept fixed by LINCS algorithm while the bond angles were free to vibrate.
\noindent The energy minimization was carried out using steepest descent method with 0.01 nm step size. After energy minimization, we proceeded towards equilibration run (NPT ensemble), where temperature and pressure couplings were done by using velocity rescaling and Berendsen barostat respectively. The equilibration run was performed for 50 ns taking 0.001 nm step size; which brought the system in equilibrium.
The stable system obtained after equilibration run was subjected for production run (NVT ensemble). It was also done for 50 ns and 0.001 nm step size. The production run yielded the important results which were analyzed through XMGRACE plots. Various energies of the system were portrayed in the energy profile diagrams. The structures of solute and solvent of the system were studied via radial distribution functions. The analysis of RDF plots at different temperatures reveal that the system becomes less organized with increase in temperature.
\noindent The transport property of system, diffusion, was studied. The self-diffusions of both cysteine and water were separately determined by using the Einstein's equation. The diffusion of binary mixture of cysteine and water was calculated by means of Darken's relation. The simulated values obtained were then compared with corresponding experimental values. The simulated values of self-diffusion coefficients of water were showed excellent agreements with that of experimental values specially at higher temperatures and small deviation ($\sim$11\%) at low temperature (288 K). Likewise, the simulated values of binary diffusion coefficient of cysteine in water was compared with that of available experimental value at 288 K. This comparison showed very little deviation of about 4\%. Further the values of diffusion coefficients at other temperatures showed correlation with temperature i.e. higher at higher temperatures. Moreover, the diffusion coefficients of cysteine were smaller than that of water for each corresponding temperature. This is because cysteine molecule is much larger than water molecule, so cysteine suffers more hindrance than water while moving in the system. Thus we can conclude that the simulated values are reasonably good. The temperature dependency of diffusion were studied through Arrhenius plots. The Arrhenius plots were also utilized to calculate the activation energies for diffusion. The simulated and experimental values of activation energies for self-diffusion of water were in agreement within 13\% error.
\noindent We can inferred that the molecular dynamics simulation is very reliable tool for the study of dynamic properties of a system. It is convenient from both time and money and it is free from experimental hazards. Also this study of transport property- diffusion, of cysteine in water will be reference to further study. In near future, we are interested to extend the present study to diffusion of cysteine in heavy water, with varying concentration as well as temperature, and compare the corresponding values in normal water. We are also curious to study the other transport properties such as viscosity and thermal conductivity in the present system.
\section*{Acknowledgements}
\noindent HB acknowledges the financial support from University Grants Commission (UGC), Nepal for M.Sc. thesis. NA acknowledges CRG support from
UGC award no. CRG 2073/74 S\&T -01. We also acknowledge the support from ICQ-13 for Computational facilities.
|
\section{Introduction}
Semantic parsing (SP) is one of the most important tasks in natural language processing (NLP).
It requires both understanding the meaning of natural language sentences and mapping them to meaningful executable queries such as logical forms, SQL queries, and Python code.
Recently, some state-of-the-art methods with Seq2Seq architectures are able to achieve over 80\% exact matching accuracy even on some complex benchmarks such as ATIS and GeoQuery. These models seem to have already solved most problems in this field.
\begin{figure}[!t]
\vspace{-1.5mm}\hspace{-1mm}
\centering
\includegraphics[width=0.48\textwidth]{fig/spider_top_fig.pdf}\vspace{-2mm}
\caption{
Our corpus annotates complex questions and SQLs. The example contains joining of multiple tables, a \texttt{GROUP BY} component, and a nested query.
}
\label{fig:task}
\vspace{-6mm}
\end{figure}
However, previous tasks in this field have a simple but problematic task definition because most of these results are predicted by semantic ``matching" rather than semantic parsing. Existing datasets for SP have two shortcomings.
First, those that have complex programs \cite{zelle96,li2014constructing,yaghmazadeh2017sqlizer,iyer17} are too small in terms of the number of programs for training modern data-intensive models and have only a single dataset, meaning that the same database is used for both training and testing the model. More importantly, the number of logic forms or SQL labels is small and each program has about 4-10 paraphrases of natural language problem to expand the size of the dataset. Therefore, the exact same target programs appear in both the train and test sets. The models can achieve decent performances even on very complex programs by memorizing the patterns of question and program pairs during training and decoding the programs exactly the same way as it saw in the training set during testing. \citet{cathy18} split the dataset by programs so that no two identical programs would be in both the train and test sets. They show that the models built on this question-splitting data setting fail to generalize to unseen programs.
Second, existing datasets that are large in terms of the number of programs and databases such as WikiSQL \cite{Zhong2017} contain only simple SQL queries and single tables. In order to test a model's real semantic parsing performance on unseen complex programs and its ability to generalize to new domains, an SP dataset that includes a large amount of complex programs and databases with multiple tables is a must.
However, compared to other large, realistic datasets such as ImageNet for object recognition \cite{imagenet_cvpr09} and SQuAD for reading comprehension \cite{Pranav16}, creating such SP dataset is even more time-consuming and challenging in some aspects due to the following reasons.
First, it is hard to find many databases with multiple tables online. Second, given a database, annotators have to understand the complex database schema to create a set of questions such that their corresponding SQL queries cover all SQL patterns. Moreover, it is even more challenging to write different complex SQL queries. Additionally, reviewing and quality-checking of question and SQL pairs takes a significant amount of time. All of these processes require very specific knowledge in databases.
To address the need for a large and high-quality dataset for a new complex and cross-domain semantic parsing task, we introduce {\it Spider}, which consists of 200 databases with multiple tables, 10,181 questions, and 5,693 corresponding complex SQL queries, all written by 11 college students spending a total of 1,000 man-hours.
As Figure \ref{fig:task} illustrates, given a database with multiple tables including foreign keys, our corpus creates and annotates complex questions and SQL queries including different SQL clauses such as joining and nested query.
In order to generate the SQL query given the input question, models need to understand both the natural language question and relationships between tables and columns in the database schema.
In addition, we also propose a new task for the text-to-SQL problem. Since Spider contains 200 databases with foreign keys, we can split the dataset with complex SQL queries in a way that no database overlaps in train and test, which overcomes the two shortcomings of prior datasets, and defines a new semantic parsing task in which the model needs to generalize not only to new programs but also to new databases. Models have to take questions and database schemas as inputs and predict unseen queries on new databases.
To assess the task difficulty, we experiment with several state-of-the-art semantic parsing models. All of them struggle with this task.
The best model achieves only 12.4\% exact matching accuracy in the database split setting.
This suggests that there is a large room for improvement.
\section{Related Work and Existing Datasets}
\label{sec:rel}
Several semantic parsing datasets with different queries have been created. The output can be in many formats, e.g., logic forms. These datasets include ATIS \cite{Price90,Dahl94}, GeoQuery \cite{zelle96}, and JOBS \cite{tang01}. They have been studied extensively \cite{zelle96,Zettlemoyer05,wong07,Das10,Liang11,banarescu13,artzi13,Reddy14,Berant14,dong16}. However, they are domain specific and there is no standard label guidance for multiple SQL queries.
Recently, more semantic parsing datasets using SQL as programs have been created. \citet{iyer17} and \citet{Popescu03} labeled SQL queries for ATIS and GeoQuery datasets. Other existing text-to-SQL datasets also include Restaurants \cite{tang2001using,Popescu03}, Scholar \cite{iyer17}, Academic \cite{li2014constructing}, Yelp and IMDB \cite{Yaghmazadeh17}, Advising \cite{cathy18}, and WikiSQL \cite{Zhong2017}. These datasets have been studied for decades in both the NLP community \cite{warren1982efficient,popescu2003towards,popescu2004modern,li2006constructing,giordani2012translating,wang2017synthesizing,iyer17,Zhong2017,Xu2017,Yu18,pshuang2018PT-MAML,2018executionguided,P18-1068,mccann2018natural} and the Database community \cite{li2014constructing, Yaghmazadeh17}. We provide detailed statistics on these datasets in Table \ref{tb:data}.
Most of the previous work train their models without schemas as inputs because they use a single database for both training and testing. Thus, they do not need to generalize to new domains. Most importantly, these datasets have a limited number of labeled logic forms or SQL queries. In order to expand the size of these datasets and apply neural network approaches, each logic form or SQL query has about 4-10 paraphrases for the natural language input. Most previous studies follow the standard question-based train and test split \cite{Zettlemoyer05}. This way, the exact same target queries (with similar paraphrases) in the test appear in training set as well. Utilizing this assumption, existing models can achieve decent performances even on complex programs by memorizing database-specific SQL templates. However, this accuracy is artificially inflated because the model merely needs to decide which template to use during testing. \citet{cathy18} show that template-based approaches can get even higher results. To avoid getting this inflated result, \citet{cathy18} propose a new, program-based splitting evaluation, where the exact same queries do not appear in both training and testing. They show that under this framework, the performance of all the current state-of-the-art semantic parsing systems drops dramatically even on the same database, indicating that these models fail to generalize to unseen queries. This indicates that current studies in semantic parsing have limitations.
We also want the model to generalize not only to unseen queries but also to unseen databases.
\citet{Zhong2017} published the WikiSQL dataset. In their problem definition, the databases in the test set do not appear in the train or development sets. Also, the task needs to take different table schemas as inputs. Therefore, the model has to generalize to new databases.
However, in order to generate 80654 questions and SQL pairs for 24241 databases, \citet{Zhong2017} made simplified assumptions about the SQL queries and databases. Their SQL labels only cover single \texttt{SELECT} column and aggregation, and \texttt{WHERE} conditions. Moreover, all the databases only contain single tables. No \texttt{JOIN}, \texttt{GROUP BY}, and \texttt{ORDER BY}, etc. are included.
Recently, researchers have constructed some datasets for code generation including IFTTT \cite{quirk2015language}, DJANGO \cite{Oda15}, HEARTHSTONE \cite{ling16}, NL2Bash \cite{nl2bash}, and CoNaLa \cite{yin18msr}. These tasks parse natural language descriptions into a more general-purpose programming language such as Python \cite{Allamanis15,ling16, RabinovichSK17,Yin17}.
\section{Corpus Construction}
\label{sec:data_collection}
\begin{figure}[t!]
\centering
\includegraphics[width=0.45\textwidth]{spider_annotation.pdf}\vspace{-2mm}
\caption{The annotation process of our Spider corpus.}
\label{fig:annotation}\vspace{-3mm}
\end{figure}
All questions and SQL queries were written and reviewed by 11 computer science students. Some of them were native English speakers.
As illustrated in Figure \ref{fig:annotation}, we develop our dataset in five steps, spending around 1,000 hours of human labor in total: \S \ref{sec:database_collection} Database Collection and Creation, \S \ref{sec:question_and_sql_annnotation} Question and SQL Annotation, \S \ref{sec:sql_review} SQL Review, \S \ref{sec:question_review} Question Review and Paraphrase, \S \ref{sec:final_question} Final Question and SQL Review.
\subsection{Database Collection and Creation}
\label{sec:database_collection}
Collecting databases with complex schemas is hard. Although relational databases are widely used in industry and academia, most of them are not publicly available. Only a few databases with multiple tables are easily accessible online.
Our 200 databases covering 138 different domains are collected from three resources. First, we collected about 70 complex databases from different college database courses, SQL tutorial websites, online csv files, and textbook examples. Second, we collected about 40 databases from the DatabaseAnswers\footnote{\url{http://www.databaseanswers.org/}} where contains over 1,000 data models across different domains. These data models contain only database schemas. We converted them into SQLite, populated them using an online database population tool\footnote{\url{http://filldb.info/}}, and then manually corrected some important fields so that the table contents looked natural. Finally, we created the remaining 90 databases based on WikiSQL. To ensure the domain diversity, we select about 500 tables in about 90 different domains to create these 90 databases. To create each database, we chose several related tables from WikiSQL dev or test splits, and then created a relational database schema with foreign keys based on the tables we selected. We had to create some intersection tables in order to link several tables together. For most other cases, we did not need to populate these databases since tables in WikiSQL are from Wikipedia, which already had real world data stored.
We manually corrected some database schemas if they had some column names that did not make sense or missed some foreign keys. For table and column names, it is common to use abbreviations in databases. For example, `student\_id' might be represented by `stu\_id'. For our task definition, we manually changed each column name back to regular words so that the system only handled semantic parsing issues.
\subsection{Question and SQL Annotation}
\label{sec:question_and_sql_annnotation}
For each database, we ask eight computer science students proficient in SQL to create 20-50 natural questions and their SQL labels.
To make our questions diverse, natural, and reflective of how humans actually use databases,
we did not use any template or script to generate question and SQL queries.
Our annotation procedure ensures the following three aspects.
\paragraph{A)~ SQL pattern coverage.}
We ensure that our corpus contains enough examples for all common SQL patterns. For each database, we ask annotators to write SQL queries that cover all the following SQL components: \texttt{SELECT} with multiple columns and aggregations, \texttt{WHERE}, \texttt{GROUP BY}, \texttt{HAVING}, \texttt{ORDER BY}, \texttt{LIMIT}, \texttt{JOIN}, \texttt{INTERSECT}, \texttt{EXCEPT}, \texttt{UNION}, \texttt{NOT IN}, \texttt{OR}, \texttt{AND}, \texttt{EXISTS}, \texttt{LIKE} as well as nested queries. The annotators made sure that each table in the database appears in at least one query.
\paragraph{B)~ SQL consistency.}
Some questions have multiple acceptable SQL queries with the same result.
However, giving totally different SQL labels to similar questions can hinder the training of semantic parsing models.
To avoid this issue, we designed the annotation protocol so that
all annotators choose the same SQL query pattern if multiple equivalent queries are possible.
\paragraph{C)~ Question clarity.}
We did not create questions that are (1) vague or too ambiguous, or (2) require knowledge outside the database to answer.
First, ambiguous questions refer to the questions that do not have enough clues to infer which columns to return and which conditions to consider.
For example, we would not ask ``What is the most popular class at University X?'' because the definition of ``popular'' is not clear: it could mean the rating of the class or the number of students taking the course.
Instead, we choose to ask ``What is the name of the class which the largest number of students are taking at University X?''. Here, ``popular'' refers to the size of student enrollment. Thus, the ``student\_enrollment'' column can be used in condition to answer this question.
We recognize that ambiguous questions appear in real-world natural language database interfaces.
We agree that future work needs to address this issue by having multi-turn interactions between the system and users for clarification. However, our main aim here is to develop a corpus to tackle the problem of handling complex queries and generalizing across databases without multi-turn interactions required, which no existing semantic parsing datasets could do. Moreover, the low performances of current state-of-the-art models already show that our task is challenging enough, without ambiguous questions.
In addition, questions are required to contain the specific information to return.
Otherwise, we don't know if class id is also acceptable in the previous case. Most of the questions in the existing semantic parsing datasets are ambiguous. This is not a serious problem if we use one single dataset because we have enough data domain specific examples to know which columns are the default.
However, it would be a serious problem in cross domain tasks since the default return values differ cross domain and people.
Second, humans sometimes ask questions that require common sense knowledge outside the given database.
For instance, when people ask ``Display the employee id for the employees who report to John", the correct SQL is
\begin{quote}
\texttt{SELECT employee\_id \\ FROM employees \\ WHERE manager\_id = (\\ SELECT employee\_id \\ FROM employees \\ \quad WHERE first\_name = `John')}
\end{quote}
which requires the common knowledge that ``X reports to Y" corresponds to an ``employee-manager" relation.
we do not include such questions and leave them as a future research direction.
\paragraph{Annotation tools} We open each database on a web-based interface powered by the sqlite\_web\footnote{\url{https://github.com/coleifer/sqlite-web}} tool. It allows the annotators to see the schema and content of each table, execute SQL queries, and check the returned results.
This tool was extremely helpful for the annotators to write executable SQL queries that reflect the true meaning of the given questions and return correct answers.
\input{data.tex}
\subsection{SQL Review}
\label{sec:sql_review}
Once the database is labeled with question-query pairs, we ask a different annotator to check if the questions are clear and contain enough information to answer the query.
For a question with multiple possible SQL translations, the reviewers double check whether the SQL label is correctly chosen under our protocol.
Finally, the reviewers check if all the SQL labels in the current database cover all the common SQL clauses.
\subsection{Question Review and Paraphrase}
\label{sec:question_review}
After SQL labels are reviewed, native English speakers review and correct each question.
They first check if the question is grammatically correct and natural.
Next, they make sure that the question reflects the meaning of its corresponding SQL label. Finally, to
improve the diversity in questions, we ask annotators to add a paraphrased version to some questions.
\subsection{Final Review}
\label{sec:final_question}
Finally, we ask the most experienced annotator to conduct the final question and SQL review. This annotator makes the final decision if multiple reviewers are not sure about some annotation issues. Also, we run a script to execute and parse all SQL labels to make sure they are correct.
\section{Dataset Statistics and Comparison}
\label{sec:data_analysis}
We summarize the statistics of Spider and other text-to-SQL datasets in Table \ref{tb:data}.
Compared with other datasets, Spider contains databases with multiple tables and contains SQL queries including many complex SQL components.
For example, Spider contains about twice more nested queries and 10 times more \texttt{ORDER BY (LIMIT)} and \texttt{GROUP BY (HAVING)} components
than the total of previous text-to-SQL datasets.
Spider has 200 distinct databases covering 138 different domains such as college, club, TV show, government, etc.
Most domains have one database, thus containing 20-50 questions, and a few domains such as flight information have multiple databases with more than 100 questions in total.
On average, each database in Spider has 27.6 columns and 8.8 foreign keys.
The average question length and SQL length are about 13 and 21 respectively.
Our task uses different databases for training and testing, evaluating the cross-domain performance.
Therefore, Spider is the \textit{only one} text-to-SQL dataset that contains both databases with multiple tables in different domains and complex SQL queries
It tests the ability of a system to generalize to not only new SQL queries and database schemas but also new domains.
\section{Task Definition}
\label{sec:task}
On top of the proposed dataset, we define a text-to-SQL task that is more realistic than prior work. Unlike most of the previous semantic parsing or text-to-SQL tasks, models will be tested on \textit{both different complex SQL queries and different complex databases in different domains} in our task. It aims to ensure that models can only make the correct prediction when they truly understand the semantic meaning of the questions, rather than just memorization. Also, because our databases contain different domains, our corpus tests model's ability to generalize to new databases. In this way, model performance on this task can reflect the real semantic parsing ability.
In order to make the task feasible and to focus on the more fundamental part of semantic parsing, we make the following assumptions:
\begin{itemize}
\setlength{\leftskip}{-3mm}
\item In our current task, we do not evaluate model performance on generating values. Predicting correct SQL structures and columns is more realistic and critical at this stage based on the low performances of various current state-of-the-art models on our task.
In a real world situation, people need to double check what condition values are and finalize them after multiple times. It is unrealistic to predict condition values without interacting with users. In reality, most people know what values to ask but do not know the SQL logic. A more reasonable way is to ask users to use an interface searching the values, then ask more specific questions. Also, other previous work with value prediction uses one single database in both train and test which makes it vulnerable to overfitting. However, SQL queries must include values in order to execute them. For value prediction in our task, a list of gold values for each question is given. Models need to fill them into the right slots in their predicted SQL.
\item As mentioned in the previous sections, we exclude some queries that require outside knowledge such as common sense inference and math calculation. For example, imagine a table with birth and death year columns. To answer the questions like ``How long is X's life length?'', we use \texttt{SELECT death\_year - birth\_year}. Even though this example is easy for humans, it requires some common knowledge of the life length definition and the use of a math operation, which is not the focus of our dataset.
\item We assume all table and column names in the database are clear and self-contained.
For example, some databases use database specific short-cut names for table and column names such as ``stu\_id'', which we manually converted to ``student id'' in our corpus.
\end{itemize}
\section{Evaluation Metrics}
\label{sec:eval}
Our evaluation metrics include Component Matching, Exact Matching, and Execution Accuracy.
In addition, we measure the system's accuracy as a function of the difficulty of a query.
Since our task definition does not predict value string, our evaluation metrics do not take value strings into account.
We will release the official evaluation script along with our corpus so that the research community can share the same evaluation platform.
\paragraph{Component Matching}
To conduct a detailed analysis of model performance, we measure the average exact match between the prediction and ground truth on different SQL components. For each of the following components: \vspace{-2mm}
\begin{itemize}
\setlength{\itemsep}{-1mm}
\item \texttt{SELECT} \quad $\bullet$ \texttt{WHERE} \quad $\bullet$ \texttt{GROUP BY}
\item \texttt{ORDER BY} \quad $\bullet$ \texttt{KEYWORDS} (including all SQL keywords without column names and operators)
\end{itemize}\vspace{-2mm}
we decompose each component in the prediction and the ground truth as bags of several sub-components, and check whether or not these two sets of components match exactly.
To evaluate each \texttt{SELECT} component, for example, consider \texttt{SELECT avg(col1), max(col2), min(col1)}, we first parse and decompose into a set \texttt{(avg, min, col1), (max, col2)}, and see if the gold and predicted sets are the same.
Previous work directly compared decoded SQL with gold SQL.
However, some SQL components do not have order constraints.
In our evaluation, we treat each component as a set so that for example, \texttt{SELECT avg(col1), min(col1), max(col2)} and \texttt{SELECT avg(col1), max(col2), min(col1)} would be treated as the same query.
To report a model's overall performance on each component, we compute F1 score on exact set matching.
\paragraph{Exact Matching}
We measure whether the predicted query as a whole is equivalent to the gold query.
We first evaluate the SQL clauses as described in the last section. The predicted query is correct only if all of the components are correct. Because we conduct a set comparison in each clause, this exact matching metric can handle the ``ordering issue'' \cite{Xu2017}.
\paragraph{Execution Accuracy\footnote{Please check our website for the latest updates on the task at \url{https://yale-lily.github.io/spider}}}
Since Exact Matching is possible to provide false negative evaluation when the semantic parser is able to generate novel syntax structures, we also consider Execution Accuracy.
For Execution Accuracy, the value is a must in order to execute SQL queries. Instead of generating these values, a list of gold values for each question is given. Models need to select them and fill them into the right slots in their predicted SQL.
We exclude value prediction in Component and Exact Matching evaluations and do not provide Execution Accuracy in the current version.
However, it is also important to note that Execution Accuracy can create false positive evaluation when a predicted SQL returns the same result (for example, `NULL') as the gold SQL while they are semantically different.
So we can use both to complement each other.
Finally, our evaluation also considers multiple acceptable keys if \texttt{JOIN} and \texttt{GROUP} are in the query. For example, suppose ``stu\_id'' in one table refers to ``stu\_id'' in another table, \texttt{GROUP BY} either is acceptable.
\iffalse
\fi
\paragraph{SQL Hardness Criteria}
To better understand the model performance on different queries, we divide SQL queries into 4 levels: easy, medium, hard, extra hard.
We define the difficulty based on the number of SQL components, selections, and conditions, so that queries that contain more SQL keywords (\texttt{GROUP BY}, \texttt{ORDER BY}, \texttt{INTERSECT}, nested subqueries, column selections and aggregators, etc) are considered to be harder.
For example, a query is considered as hard if it includes more than two \texttt{SELECT} columns, more than two \texttt{WHERE} conditions, and \texttt{GROUP BY} two columns, or contains \texttt{EXCEPT} or nested queries. A SQL with more additions on top of that is considered as extra hard.
Figure \ref{fig:example} shows examples of SQL queries in 4 hardness levels.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{fig/example.pdf}
\caption{SQL query examples in 4 hardness levels.}
\label{fig:example}
\end{figure}
\input{results_table.tex}
\section{Methods}
\label{sec:methods}
In order to analyze the difficulty and demonstrate the purpose of our corpus, we experiment with several state-of-the-art semantic parsing models.
As our dataset is fundamentally different from the prior datasets such as Geoquery and WikiSQL, we adapted these models to our task as follows.
We created a `big' column list by concatenating columns in all tables of the database together as an input to all models. Also, for each model, we limit the column selection space for each question example to all column of the database which the question is asking instead of all column names in the whole corpus.
\paragraph{Seq2Seq}
Inspired by neural machine translation \cite{sutskever2014sequence}, we first apply a basic sequence-to-sequence model, \textbf{Seq2Seq}.
Then, we also explore \textbf{Seq2Seq+Attention} from \cite{dong16} by adding an attention mechanism \cite{bahdanau2015neural}.
In addition, we include \textbf{Seq2Seq+Copying} by adding an attention-based copying operation similar to \cite{jia2016}.
The original model does not take the schema into account because it has the same schema in both train and test. We modify the model so that it considers the table schema information by passing a vocabulary mask that limits the model to decode the words from SQL keywords, table and column names in the current database.
\paragraph{SQLNet}
introduced by \cite{Xu2017} uses column attention and employs a sketch-based method and generates SQL as a slot-filling task.
This fundamentally avoids the sequence-to-sequence structure when ordering does not matter in SQL query conditions.
Because it is originally designed for WikiSQL, we extend its \texttt{SELECT} and \texttt{WHERE} modules to \texttt{ORDER BY} and \texttt{GROUP BY} components.
\paragraph{TypeSQL} proposed by \cite{Yu18} improves upon SQLNet by proposing a different training procedure and utilizing types extracted from either knowledge graph or table content to help model better understand entities and numbers in the question.
In our experiment, we use the question type info extracted from database content. Also, we extend their modules to \texttt{ORDER BY} and \texttt{GROUP BY} components as well. It is the only model that uses database content.
\section{Experimental Results and Discussion}
We summarize the performance of all models on our test set including accuracy of exact matching in Table \ref{tab:results} and F1 scores of component matching in Table \ref{tab:results_component}.
\paragraph{Data Splits}
For the final training dataset, we also select and include 752 queries and 1659 questions that follow our annotation protocol from six existing datasets: Restaurants, GeoQuery, Scholar, Academic, IMDB, and Yelp.
We report results on two different settings for all models: (1) Example split where examples are randomly split into 8659 train, 1034 dev, 2147 test. Questions for the same database can appear in both train and test. (2) Database split where 206 databases are split into 146 train, 20 dev, and 40 test. All questions for the same database are in the same split.
\paragraph{Overall Performance}
The performances of the Seq2Seq-based basic models including Seq2Seq, Seq2Seq+Attention, and Seq2Seq+Copying are very low. However, they are able to generate nested and complex queries because of their general decoding process.
Thus, they can get a few hard and extra hard examples correct.
But in the vast majority of cases, they predict invalid SQL queries with grammatical errors.
The attention and copying mechanisms do not help much either.
In contrast, SQLNet and TypeSQL that utilize SQL structure information to guide the SQL generation process significantly outperform other Seq2Seq models. While they can produce valid queries, however, they are unable to generate nested queries or queries with keywords such as \texttt{EXCEPT} and \texttt{INTERSECT} because they limit possible SQL outputs in some fixed pre-defined SQL structures.
As Component Matching results in Table \ref{tab:results_component} shows, all models struggle with \texttt{WHERE} clause prediction the most. \texttt{WHERE} clause is more likely to have multiple columns and operations, which makes its prediction the most challenging. The most number of prediction errors for each component is from column prediction.
In general, the overall performances of all models are low, indicating that our task is challenging and there is still a large room for improvement.
\paragraph{Example Split vs Database Split}
As discussed in Section \ref{sec:task}, another challenge of the dataset is to generalize to new databases.
To study this, in Table \ref{tab:results} and Table \ref{tab:results_component} we compare model performances under the two settings.
For all models, the performance under database split is much lower than that under example split.
Especially, TypeSQL utilizes column names as question types, and it outperforms other models with a large margin under the example split. However, its performance drops the most on the database split data set. This indicates that the model does well on complex SQL prediction but fails to generalize to new databases.
In addition, we observe that all models perform much poorer on column selection under database split than example split.
Overall, the result shows that our dataset presents a challenge for the model to generalize to new databases.
\begin{figure}[!t]
\vspace{-1.5mm}\hspace{-1mm}
\centering
\includegraphics[width=0.48\textwidth]{fig/fk_with_acc_typesql_seq2seq_atten.png}\vspace{-3mm}
\caption{Exact matching accuracy as a function of the number of foreign keys.}
\label{fig:complexity}
\vspace{-4mm}
\end{figure}
\paragraph{Complexity of Database Schema}
In order to show how the complexity of the database schema affects model performance, Figure \ref{fig:complexity} plots the exact matching accuracy as a function of the number of foreign keys in a database.
The performance decreases as the database has more foreign keys.
The first reason is that the model has to choose column and table names from many candidates in a complex database schema.
Second, a complex database schema presents a great challenge for the model to capture the relationship between different tables with foreign keys.
SQL answers to questions on the database with more number of foreign keys are more likely to join more tables.
It indicates that this task requires more effective methods to encode the relation of tables with foreign keys.
\section{Conclusion}
\label{sec:conclusion}
In this paper we introduce Spider, a large, complex and cross-domain semantic parsing and text-to-SQL dataset,
which directly benefits both NLP and DB communities.
Based on Spider, we define a
new challenging and realistic semantic parsing task.
Experimental results on several state-of-the-art models on this task suggest plenty space for improvement.
\section*{Acknowledgement}
We thank Graham Neubig, Tianze Shi, Catherine Finegan-Dollak, and three anonymous reviewers for their discussion and feedback.
We also thank Barry Williams for providing part of our database schemas from the DatabaseAnswers.
\section{Introduction}
\input{intro}
\section{Background}
\input{background}
\section{Our Approach}
\input{approach}
\section{Experiments}
\input{experiments}
\section{Discussion}
\input{discussion}
\section{Related Work}
\input{related}
\section{Conclusion}
\input{conclusion}
\subsection{Models}
\textbf{SQL+Q: } Model only uses questions as its input.
\textbf{SQL+QC: } Model only uses question and column names as inputs.
\textbf{SQLKG+QC: } Our main model comparable to previous works on WikiSQL dataset. It uses questions, column names and type attention but no content of databases as inputs.
\textbf{SQLKG+TC: } model uses question, column names, type attention by fully access to content of databases.
\fi
Semantic parsing maps natural language to meaningful executable programs. The programs could be a range of representations such as logic form \cite{zelle96,Zettlemoyer05,wong07,Das10,Liang11,banarescu13,artzi13,Reddy14,Berant14,pasupat2015compositional,Yin15}.
The main focus of our paper is on translating natural language into SQL queries. \cite{Poon13}
\cite{li2014constructing}
\section{Introduction}
Semantic parsing introduction
Most of previous work on text-to-SQL use either simple datasets or simple train-test splits on a single database. WikiSQL \cite{Zhong2017} is an example of a simple SQL dataset. Its databases only contain single tables. This leads to the fact that corresponding SQL queries do not have complex structures such as JOIN and GROUP BY etc. On the other hand, \cite{iyer17} applied a basic seq2seq model on the seq2SQL task with very complex SQL queries. However, their experiments are on one single database and use question-splitting datasets during training and testing. Thus, their model memorizes database-specific SQL templates and only needs to decide which template to use during testing. Unsurprisingly, it performs very bad on unseen SQL queries despite having high test accuracy under the question-splitting setting on one single database. To avoid this issue, Cathy et al. proposed a new way of splitting datasets so that the same SQL queries do not appear in both train and test data. The template-based approach fails under this query-splitting setting.
In the real world, however, we would like to know how good the seq2SQL model performs not only on unseen queries but also on unseen databases. In this project, we are going to introduce the most realistic seq2SQL task and explore new methods to solve this problem. First, we will label a new corpus including about 10000 SQL-question pairs for about 200 databases with multiple tables (about 4000 SQL-question pairs/8 databases already labeled by Cathy et al. and Iyer et al.). In this task, we split the dataset in a way so that different databases are seen during training and testing. Second, we are going to discover new seq2SQL approaches that take not only questions and SQL pairs but also table schemas and database structures as inputs.
How can we learn a seq2sql parser and generalize well to new databases
Building natural language interfaces to relational databases is an important and challenging problem \cite{li2014constructing,pasupat2015compositional,Yin15,Zhong2017,Yaghmazadeh17,Xu2017,Wang2017}.
It requires a system that is able to understand natural language questions and generate corresponding SQL queries.
Interacting with relational databases through natural language helps users of any background easily query and analyze a vast amount of data. This requires a system that understands users' questions and converts them to SQL queries automatically. However, this research area lacks a large corpus of training dataset because of the hardness of labelling. In this paper we introduce a large human labeled corpus of test to SQL dataset to the research community. Also, we define a more realistic and challenging text2SQL task. We show current state-of-art methods fail in this task. Finally, we present a novel approach to solve this problem.
\section{Related Work}
\label{sec:rel}
Semantic parsing maps natural language to meaningful executable programs. The programs could be a range of representations such as logic forms \cite{zelle96,Zettlemoyer05,wong07,Das10,Liang11,banarescu13,artzi13,Reddy14,Berant14,pasupat2015compositional}. Another area close to our task is code generation. This task parses natural language descriptions into a more general-purpose programming language such as Python \cite{Allamanis15,ling16, RabinovichSK17,Yin17}.
As a sub-task of semantic parsing, the text-to-SQL problem has been studied for decades \cite{warren1982efficient,popescu2003towards,popescu2004modern,li2006constructing,giordani2012translating,wang2017synthesizing}. The methods of the Database community \cite{li2014constructing, Yaghmazadeh17} involve more hand feature engineering and user interactions with the systems. In this work, we focus on recent neural network based approaches \cite{Yin15,Zhong2017,Xu2017,Wang2017,iyer17}. \newcite{dong16} introduce a sequence-to-sequence approach to converting text to logical forms. Most of previous work focus on specific table schemas, which means they use a single database in both train and test. Thus, they don't generalize to new databases.
\newcite{Zhong2017} publish the WikiSQL dataset and propose a sequence-to-sequence model with reinforcement learning to generate SQL queries. In the problem definition of the WikiSQL task, the databases in the test set do not appear in the train and development sets. Also, the task needs to take different table schemas into account. \newcite{Xu2017} further improve the results by using a SQL sketch based approach employing a sequence-to-set model.
\section{Semantic Parsing Datasets}
\label{sec:methods}
Previous semantic parsing benchmarks expand their size by paraphrasing many questions for each labeled program. Then, they split the dataset by questions so that almost all programs in test appear in train for mutilple times as well. To make this problem clearer, (Cathy et al, 2018) propose a new evaluation by splitting dataset based on programs so that train and test have no overlap on programs. They show performances of all current state-of-art semantic parsing systems drop dramatically.
\section{Task Definition}
\label{sec:methods}
\section{Systems}
\label{sec:methods}
\subsection{Implementation Details}
\section{Results and Discussion}
\subsection{Results}
\subsection{Error Analysis}
\section{Conclusion and Future Work}
\label{sec:conclusion}
\subsection{Annotation Projection}
\input{annotation_projection}
\subsection{Direct Transfer}
\input{direct_transfer}
\subsection{Transfer Scenarios}\label{sec_scenarios}
We conduct a diverse set of experiments to evaluate our proposed methods. We
experiment with three scenarios:
1) \emph{parallel data}: in this setting, we use both the annotation projection and direct transfer approaches. In order to see the effect of domain and genre, we use different translation datasets: out-of-domain data (religious text), out-of-domain data (contemporary political text) and relatively in-domain data.
\todo{KM: Below not quite clear. I have tried to rephrase. CHeck if OK.
In the direct transfer approach, we produce the intersected alignments from source and target alignment direction and use the Giza++ \cite{och2000giza} tool to extract from these the translation dictionaries.
We restrict ourselves to the most frequent translation for each word in order to prune alignment noise.
\todo{KM: is effect of domain and genre measured for both approaches? I think so but if the sentence occurs here, it implies only for direct transfer. I have moved it above. If I am wrong, move back.}
2) \emph{translation dictionaries}: we use the manual translation dictionaries extracted from Wiktionary\footnote{\url{https://www.wiktionary.org/}}
\todo{KM: check my rewording}
in place of the dictionaries generated from parallel translations
for the direct transfer model. The Wiktionary entries are noisy and for some languages, the coverage of the lexicon is very restricted; and 3) \emph{comparable corpora}: we use the model of \newcite{??} to extract cross-lingual word embeddings. We use the cross-lingual embeddings as features for the direct transfer method.
\todo{KM: Check if my rewording OK. I foudn this awkward.}
We only use the fix pre-trained word embedding features $x_{ce}$ for the direct transfer method in this scenario.
\subsection{Datasets, Tools and Settings}
\todo{KM: wording confusing here. Makes it sound as if only Persian is laveled. I assume not.}
\paragraph{Labeled sentiment data} We downloaded tweets labeled with sentiment for 12 languages
from \newcite{tweet_paper} as well as SentiPers data\footnote{\url{http://dadegan.ir/catalog/sentipers}}, a set of digital products reviews for Persian (fa), as our rich-resource languages. The languages are Bulgarian (bg), German (de), English (en), Spanish (es), Croatian (hr), Hungarian (hu), Polish (pl), Portuguese (pt), Russian (ru), Slovak (sk), Slovene (sl) and Swedish (sv). We split use 80\% of the data as training, 10\% for development and 10\% of the data for testing. We use all development data sets to train the supervised NBSVM models and
\todo{KM: What is slightly tune? Either drop the word or define.}
slightly tune the parameters for the direct transfer and annotation projection on the Persian development data.
\todo{KM: I changed the wording below. Check}
We also use labeled Uyghur, Chinese and Arabic for the target language only, given smaller amounts of labeled data. We manually annotated the Uyghur and Chinese data and use the test section of the Arabic Semeval dataset\footnote{\url{http://alt.qcri.org/semeval2017/task4/}}. In the multi-source setting, we hold out the training dataset for each language and use the other training sets for training. Table~\ref{tab_tweet_size} shows the data sizes.
\todo{KM: It sounds odd to say here that use U, C and A as ONLY the target language. Why? . To address this I have edited above although we could be called out on that for Arabic. Let's discuss which are low resource languages.}
\input{tweet_data_stat}
\paragraph{Parallel data}
\todo{KM: One problem with the discussion as is, is that you don't describe the domain of the test data. So the out-of-domain descriptions require the reader to think: what is the domain? I think it should be clearly stated above. We can discuss hwere. What exactly is in-domain? Is it product reviews? Or is it any twitter data?}
We using the following datasets:
\begin{itemize}
\item \emph{Bible and Quran}: We use the Quran and Bible translations as our out-of-domain parallel datasets. We use the corpus of \newcite{christodouloupoulos2014massively} for the Bible dataset. This dataset has one translation per language; except for English for which it has two translations. We use both translations of the English data. We use the Tanzil translations\footnote{\url{http://tanzil.net/trans/}} for the holy Quran. This dataset has multiple translations for some languages.
\todo{KM: CHeck my rewording.}
We excluded a subset of the translations from Russian and English that were interpretations as opposed to translations.\footnote{For Russian, we use the Krachkovsky, Kuliev, Osmanov, Porokhova, and Sablukov translations and for English, we use the Ahmedali, Arberry, Daryabadi, Itani, Mubarakpuri, Pickthall, Qarai, Qaribullah, Sahih, Sarwar, Shakir, Wahiduddin, and Yusufali translations.}
\item \emph{Europarl }: We use the Europarl data \cite{koehn2005europarl} as out-of-domain contemporary political text. We restricted ourselves to those sentences that are translated to all of the 10 languages (Bulgarian, German, English, Spanish, Hungarian, Polish, Portuguese, Slovak, Slovene, and Swedish). That comprised a total of 294738 sentences for all languages.
\item \emph{Linguistic Data Consortium (LDC) parallel data}: The LDC consists of in-domain parallel data \todo{KM: Why indomain? State} drawn from Twitter???. We use nine English to target parallel translations from the LDC data: Chinese (16440 sentences)\footnote{LDC2016E30\_LORELEI\_Mandarin}, Persian (57087 sentences)\footnote{LDC2016E93\_LORELEI\_Farsi}, Hungarian (157931 sentences) \footnote{LDC2016E99\_LORELEI\_Hungarian}, Arabic (49446 sentences)\footnote{LDC2016E89\_LORELEI\_Arabic}, Russian (193967 sentences)\footnote{LDC2016E95\_LORELEI\_Russian}, Spanish (345940 sentences)\footnote{LDC2016E97\_LORELEI\_Spanish}, and Uyghur (99272 sentences) \footnote{LDC2016E57\_LORELEI\_IL3\_Incident\_Language\_Pack\_\\for\_Year\_1\_Eval}.
\end{itemize}
\paragraph{Monolingual text}
We use the Wikipedia dump data set to create monolingual word embeddings and Brown clusters. We also use the monolingual text to train cross-lingual word embeddings and Brown clusters based on the method of \newcite{rasooli_16}: this is done by randomly swapping words in a language with its translation in another language. We replaced X\% of the words in the monolingual corpora with their translations drawn from either the bilingual dictionary automatically generated from the parallel corpus or the manual Wiktionary bilingual dictionary.
\todo{KM: Since it is so prominent here, I think we cannot rely on people reading your other paper to understand what is done. It should be defined here. So the last part of this sentence should be made more precise. I've tried but you should edit.}
\paragraph{Pos tagging and tokenization} We use the Universal dependencies corpora \cite{universal_deps} to create training data for sentence and word tokenization, and part-of-speech tagging. We trained sentence tokenization with the OpenNLP toolkit\footnote{\url{https://opennlp.apache.org/}}. We also use the OpenNLP toolkit to train word tokenization for the languages that the Europarl package does not provide a tokenization script; except Arabic, Persian and Chinese. For Chinese we use the Stanford Chinese segmenter \cite{chinese_segmentor}, Madamira \cite{pasha2014madamira} with ATB tokenization format for Arabic, and the Hazm toolkit\footnote{\url{https://github.com/sobhe/hazm}} for Persian. We trained the perceptron-based part-of-speech tagger of \newcite{collins2002discriminative}\footnote{\url{https://github.com/rasoolims/SemiSupervisedPosTagger}} for training our tagging models.
\todo{KM: A question will come up here: you needed data labeled with POS for each language, correct? Did you have it?}
\paragraph{Comparable data}\todo{MSR: Noura; please fill this in. If you want to add any table for statistics; please put in in another tex file and use \\input\{\} for that}
\paragraph{Embeddings and Brown clusters} We use the Word2vec tool\footnote{\url{https://code.google.com/archive/p/word2vec/}} with its default setting and dimension of 300 for all of our embeddings (monolingual and cross-lingual). We trained monolingual and cross-lingual Brown clusters with the method of \newcite{stratos2014spectral}\footnote{\url{https://github.com/karlstratos/singular}} with 500 clusters.
\paragraph{Neural network model} We have implemented our neural network model by using the Dynet library \cite{neubig2017dynet}. We use seven iterations for the single-source transfer experiments and two epochs for the multi-source transfer (concatenation of multiple sources). We use the following dimensions for all of our experiments: pre-trained word embeddings $d_{ce}=300$, updatable word embeddings $d_{e}=400$, cross-lingual Brown cluster embeddings $d_{cc}=50$, batch size of 10K, LSTM output dimension $d_{rec}=400$, and hidden layer dimension $d_h = 400$.
\subsection{Baseline approach}
We provide a simple baseline to compare all of our methods: this is done by translating the Sentiwordnet \cite{baccianella2010sentiwordnet} lexicon to each target language by using the dictionaries extracted from the Bible and Quran parallel data. We use a simple threshold to determine the sentiment assignment for each sentence:
\[
l(x)=
\begin{cases}
\text{positive} & s_x[p]-s_x[n]>\delta \\
\text{negative} & s_x[n]-s_x[p]>\delta \\
\text{neutral} & \text{otherwise} \\
\end{cases}
\]
$s_x[p]$ and $s_x[n]$ show the average positive and negative sentiwordnet score for the words in a sentence $x$. We chose $\delta = 0.1$ based on the performance on the Persian development data.
\subsection{English to Target Transfer Results}
In our first set of experiments, we conduct the transfer experiments by transferring only from the English language. In cases where we have more than one translation for a sentence (such as Quran), we apply majority voting to get the most frequent label. The results are depicted in Table~\ref{tab_single_results}. \todo{MSR: after having all numbers, I will write more. the results for the projection will be updated; some numbers are wrong. Will update it asap}
\input{single_table}
\subsection{Multi-Source Transfer Results}
In this set of experiments, we use all training data sets (except the one for the target language) as the source language in the experiments. For annotation projection, we apply majority voting on sentences with more than one translation. For direct transfer we apply three different techniques: 1) Concatenation: we concatenate all the training sets and use it as the training data; 2) Ensemble-flat: for every source language, we train a single-source direct transfer method and run the trained model on the testing data. We then use the most frequent label assignment on the testing data; 3) Ensemble-KL: this is very similar to the Ensemble-flat method, except that we weigh every label assignment proportional to the KL-divergence of the target and source POS trigram distribution. We use the method of \newcite{rosa-zabokrtsky:2015:ACL-IJCNLP} in which we use the fourth power of the inverted value the KL-divergence.\footnote{We also tried the cosine similarity but similar to \cite{rosa-zabokrtsky:2015:ACL-IJCNLP} we did not see any improvement.}
\todo{KM: THink about whether we should include both ensemble-flat and ensemble-KL.}
Table~\ref{tab_multi_source_results} shows the results for different settings.\todo{the results for the projection will be updated; some numbers are wrong. Will update it asap}
\input{multi_source_table}
\subsection{Supervised Sentiment Analysis}\label{sec_sup_sent}
\todo{KM: Agree this section can be removed.}
In a supervised sentiment analysis system, a set of training examples ${\cal X}=\{x_1, x_2, \cdots, x_n\}$ are given with their corresponding labels $\{l_1, l_2, \cdots, l_n\}$. Each label $l_i$ comes from the set of possible sentiment labels ${\cal L}=\{positive, negative, neutral\}$. A machine learning method learns a model ${\cal M}$ from the labeled instances ${\cal X}$. At decoding time, the sentiment analysis system uses a scoring function obtained from ${\cal M}$ and predicts the best label $l^*(x)$ for every sentence $x$:
\[
l^*(x) = \arg\max_{l \in {\cal L}} \theta(\phi(x,l))
\]
where $\theta(x,l)$ is a function that gives the score of label $l$ for the input sentence $x$. The function $\phi(x,l)\in \mathbb{R}^d$ is a generic feature function that extracts features for sentence $x$ and label $l$. The features may include any useful information such as word n-grams, part-of-speech tags or word embeddings.
\subsection{Data Assumption}
Throughout this paper, we will assume that we have $k$ source languages $L_1, \cdots, L_k$ and a single target language $L_{k+1}$. We assume to have the following data:
\begin{itemize}
\item Labeled training examples ${\cal X}_i$ for $i\in \{1,\cdots,k\}$.
\item Small translation data for all language pairs. We can also extract a cross-lingual translation dictionary $t(w, i, j)$ for which gives the most frequent translation of word $w$ in language $i$ to language $j$. The frequency of translations is automatically extracted from word alignment (e.g. Giza++~\cite{och2000giza}) run on translation data.
\item Part-of-speech tagger and word tokenizer for all languages.
\item Monolingual text for all languages. We use the Wikipedia text in this paper.
\end{itemize}
\subsection{A Baseline Approach}\label{sec_baseline}
As a baseline, we use a lexicon-based approach. We use the Sentiwordnet lexicon \cite{baccianella2010sentiwordnet}: this lexicon gives the probability of each word and sense pair for being positive or negative. Formally, we use the lexicon ${\cal S}$ that gives the probability of positive and negative sentiments as a two-dimensional vector ${\cal S}(w,s)\in \mathbb{R}^2$ for a word $w$ and its corresponding sense $s$. Since we do not have word sense information for the target language, we use the averaged vector $\psi(w)\in \mathbb{R}^2$ over all the possible senses of each word $w$:
\[
\psi(w)=
\begin{cases}
\frac{\sum_{s \in \text{senses}(w)} {\cal S}(w,s)}{|\text{senses}(w)|} & \text{if } w\in {\cal S}\\ [0, 0] & \text{otherwise}
\end{cases}\]
For each sentence $x$ in a target language $L_t$, we use the translation dictionary to get the corresponding English Sentiwordnet score and average them by the number of seen entries in the Sentiwordnet dictionary:
\[
f(x) = \frac{\sum_{i=1}^{n} \psi(t(w_i, \text{en}, L_t))}{\sum_{i=1}^{n}\mathbb{I}(t(w_i, \text{en}, L_t)\in {\cal S})}
\]
where $\mathbb{I}$ is the identity function.
Finally we define a simple threshold to assign the label of the sentence:
\[
l(x)=
\begin{cases}
\text{positive} & f(x)[0]-f(x)[1]>\delta \\
\text{negative} & f(x)[1]-f(x)[0]>\delta \\
\text{neutral} & \text{otherwise} \\
\end{cases}
\]
where $f(x)[0]$ and $f(x)[1]$ show the positive and negative scores of sentence $x$ and $\delta$ is a predefined threshold. In this paper we use $\delta = 0.1$.
\end{comment}
\subsection{Annotation Projection}
In single-source annotation projection, we assume that we have translation data where sentences $\{s_1, s_2, \cdots, s_m\}$ from the source rich-resource language $L_s$ are translated to sentences $\{t_1, t_2, \cdots, t_m\}$ in a low-resource language $L_t$; i.e. $s_i$ is a translation for $t_i$ for $i=1,\cdots, m$. We also assume that we have labeled training examples ${\cal X}=\{x_1, x_2, \cdots, x_n\}$ for the source language $L_s$. Thus, we first train a supervised system ${\cal M}_{sup}$ on ${\cal X}$ and afterward, use the trained system to ${\cal M}_{sup}$ predict the labels of the source-side parallel translated text $\{l_{s_1}, l_{s_2}, \cdots, l_{s_m}\}$. We project those labels to the target-side translation by assuming $l_{s_i} \rightarrow l_{t_i}$. After applying projection, we use the target-side translation text $\{t_1, t_2, \cdots, t_m\}$ with the projected labels $\{l_{t_1}, l_{t_2}, \cdots, l_{t_m}\}$ to train the target model ${\cal M}_{proj}$. The trained model ${\cal M}_{proj}$ is the final model for analyzing sentiment in the target language.
Multi-source transfer is an extension to single-source transfer. In this setting, there is more than one source language and the translation data includes a mapping between source sentences in multiple source languages for each target sentence. In this setting, we apply the same technique for each source language and use majority voting the get the most reliable projected label for each target sentence. Finally, a supervised system is trained on the projected data.
\subsection{Direct Transfer}
In direct transfer, we directly use the training instance $X_i$ for the source languages $L_i$ ($i=1,\cdots k$). We can have an unrestricted number of source languages. In other words, a supervised sentiment analysis system can be trained on the concatenation of $k$ labeled datasets $\cup_{i=1}^{k} {\cal X}_i$ or can be the outcome of the ensemble of $k$ single-source direct transfer methods. One benefit of direct transfer is the direct use of gold labels in code-switched data as opposed to the use of automatically transferred labels in annotation projection. The main problem with direct transfer is that most of the common features do not generalize beyond each source language: for example lexical features in one language may not appear in other languages. To address this problem, we design a deep learning neural network model that makes use of cross-lingual word representations and translation dictionaries. More details are given in \S\ref{sec_direct}.
\subsubsection{Feature Adaptation}
As mentioned in \S\ref{sec_background}, the main challenge in direct transfer is the adaptation of features from the source languages to the target language. When dealing with labeled datasets in other languages, most of the words occurring in the training data do not exist in the target language vocabulary. We apply the following techniques to address this problem:
\begin{itemize}
\item {\bf Word representation features:} We train cross-lingual word representations such that words with similar meanings in different language have similar representations. We use the method of \newcite[Figure 1]{rasooli_16} to train cross-lingual word embeddings and Brown clusters.
\item {\bf Lexical features:} We use the translation dictionaries $t(w, i, t)$ that are either extracted from the translation text or from manual dictionaries to translate as many words of the labeled examples ${\cal X}_i \in L_i$ to the target language words.
\todo{KM: Mohammad - do you use any lexical features other than the words themselves? Or are you referring to the word embeddings of the translated words? It's slightly unclear what you mean by lexical features. Perhaps you should define it if you're going to use it often?}
The lexical features of the translated words can be used directly when training the target model.
\end{itemize}
\subsubsection{Deep Learning Model}
The input to our model is a sequence of words in the sentence $x = \{x_1, x_2, \cdots, x_n\}$ where $n$ is the number of words. If any dictionary (either extracted from parallel data or a manual dictionary) is provided, the words that exist in the dictionary, will be translated to the target language. The model uses two representations for the input: a recurrent representation based on long short-term memories (LSTM) \cite{hochreiter1997long} and the other based on averaging over all inputs. A graphical depiction of underlying model is shown in Figure~\ref{fig_neural_fig}.
\todo{KM: In this section, I don't think it is clear that you use code-switched data. Unless perhaps it is in a later section? I am wondering if some examples would help.}
\todo{KM: YOu definitely need an example here. Show the code-switched data and describe how that affects the word embeddings.}
\input{neural_fig}
\paragraph{Embedding Layer}
We use the following features for every word in the sentence:
\begin{itemize}
\item A \emph{fixed} pre-trained cross-lingual word embeddings $x_{ce} \in \mathbb{R}^{d_{ce}}$ extracted from the method of \newcite{rasooli_16}.
\item A randomly initialized word embedding $x_{e} \in \mathbb{R}^{d_{e}}$ for every word in the sentence.
\todo{KM: I would not comment out the following sentence as it's unclear here whether you have a sentence in one language or if you have both.}
If a word has translation, we use the translation; otherwise we use the original word.
\item The cross-lingual word cluster embedding $x_{cc} \in \mathbb{R}^{d_{cc}}$ to represent the word cluster identity of each word. The word clusters are extracted by the method of \newcite{rasooli_16}.
\todo{KM: Again, I would include the sentence you have commented below. I have uncommented both.}
Similar to the cross-lingual embeddings; we first look at the translation and otherwise the original word.
\todo{KM: Are you using sentiwordnet in *all* methods? I thought you were just using it in the baseline.}
\item In the case of single transfer from English, we use the fixed two-dimensional Sentiwordnet \cite{baccianella2010sentiwordnet} score $x_{sw} \in \mathbb{R}^{2}$ that represents the likelihood of a word being positive or negative. We basically translate Sentiwordnet lexicon to the target language by using the translation dictionaries.
\end{itemize}
\paragraph{Intermediate Layer}
The input for every word $x_i$ to intermediate layers is the concatenation of the embedding layers: $f = x_{ce}[i] \circ x_e[i] \circ x_{cc}[i]$ where $\circ$ is the concatenation operator. Thus the dimension of the input is $d_i = d_{ce}+d_{e}+d_{cc}$.\footnote{For the sake of generality, we have not included the Sentiwordnet representation $x_{sw}$ in the notation since it is only used in single-transfer from English to the target and not in the multi-source transfer experiments.} The following representations are used to represent the intermediate layer:
\begin{itemize}
\item {\bf Recurrent layer}: We use a Bidirectional LSTM (BiLSTM) for representing the sequence of words in the sentence: a forward pass $LSTM_f(f_{[1:n]}) \in \mathbb{R}^{d_{rec} \times d_i}$ gives the final representation of the sentence by looking from left to right, and the backward pass $LSTM_b(f_{[n:1]}) \in \mathbb{R}^{d_{rec} \times d_i}$ looks from right to left. Then the output of the two LSTMs are concatenated as $r(x) \in \mathbb{R}^{2\cdot d_{rec}}$:
\[
r(x) = LSTM_f(f_{[1:n]}) \circ LSTM_b(f_{[n:1]})
\]
\item {\bf Average pool layer}: Since there are so many word order inconsistencies from the source to the target languages, we also use an average pool layer $p(x) \in \mathbb{R}^{d_i}$ to average all the input features. This layer is represents the bag-of-word information of the sentence without taking the sequence information into account:
\[
p(x) = \frac{\sum_{i=1}^{n} f[i]}{n}
\]
\end{itemize}
\paragraph{Output Layer}
We use a multilayer perceptron (MLP) to get the likelihood of each label: the two intermediate layers $r(x)$ and $p(x)$ are concatenated and fed to a hidden layer $H \in \mathbb{R}^{d_h \times (2\cdot d_{rec} + d_i)}$ activated by rectified linear units (ReLU) \cite{nair2010rectified}:
\[
H(x) = ReLU (H (r(x) \circ p(x))
\]
Finally the hidden layer output is fed to the output softmax layer with the weight matrix $W \in \mathbb{R}^{|{\cal L}| \times d_h}$ where $|{\cal L}|$ is the number of unique sentiment labels (in our case it is 3):
\[
p(l | x) = \frac{e_l^{W H(x)}}{\sum_{i=1}^{|{\cal L}|} e_i^{W H(x)}}
\]
We use the sum log likelihood function function with the Adam optimizer \cite{adam_paper} to learn the model parameters.
\subsection{Transfer from a Single Source}
The results from single source transfer are mixed, but some trends do appear. In general, in-domain data (LDC) enables significant increases in performance for most languages. For projection, the LDC results are better than projection results using either BQ or EP in all cases except for Uyghur where perhaps the smaller parallel data caused the drop {\bf KM: numbers could help here. How much smaller was the Uyghur dataset than the others?}. For direct transfer, the LDC results are better than the projection results using either Wikipedia, BQ or EP in all cases except for Farsi. {\bf Mohammad - any idea why?}
Looking at direct transfer in isolation, we see that the automatically generated dictionaries produce better results than the manual Wikipedia dictionary in eight out of 14 languages when the BQ parallel data is used {\bf Mohammed - is there less parallel data in the other six languages?} and six out of nine languages when the EP data is used. In general, where more parallel data is available, we can expect automatically generated dictionaries to do better than manual.
In comparing project with direct transfer, the results are more mixed. With LDC data, each approach does better or equal on approximately the same number of languages. With the BQ data (a lower resource setting), annotation does better on average but that is because of some large differences in a small number of languages (sk, sl, sv, ug, and zh). {\bf What is different about these languages? MOre or less data? By the way, in the table there is room to spell out the language. Why not do that?} With the EP data, results are split, with direct transfer doing better on four out of nine languages, projection better on four, and in one case, equivalent performance. {\bf I think it would be worth doing statistical significance tests on the differences to enable better counts.}
{\bf Would be good to conclude this subsection by indicating which approach better if you only have a single source. Let's discuss.}
\section{Evaluation Metrics}
\section{SQL Hardness Criteria}
To better understand the model performance on different queries, we divide SQL queries into 4 levels: easy, medium, hard, extra hard. We define the difficulty as follows.
We first define:
\begin{itemize}
\item SQL components 1: \texttt{WHERE, GROUP BY, ORDER BY, LIMIT, JOIN, OR, LIKE, HAVING}
\item SQL components 2: \texttt{EXCEPT, UNION, INTERSECT, NESTED}
\item Others: number of aggregations $>$ 1, number of select columns $>$ 1, number of where conditions $>$ 1, number of group by clauses $>$ 1, number of group by clauses $>$ 1 (no consider col1-col2 math equations etc.)
\end{itemize}
Then different hardness levels are determined as follows.
\begin{itemize}
\item Easy: if SQL key words have ZERO or exact ONE from [SQL components 1] and SQL do not satisfy any conditions in [Others] above. AND no word from [SQL components 2].
\item Medium: SQL satisfies no more than two rules in [Others], and does not have more than one word from [SQL components 1], and no word from [SQL components 2]. OR, SQL has exact 2 words from [SQL components 1] and less than 2 rules in [Others], and no word from [SQL components 2]
\item Hard: SQL satisfies more than two rules in [Others], with no more than 2 key words in [SQL components 1] and no word in [SQL components 2]. OR, SQL has 2 $<$ number key words in [SQL components 1] $<=$ 3 and satisfies no more than two rules in [Others] but no word in [SQL components 2]. OR, SQL has no more than 1 key word in [SQL components 1] and no rule in [Others], but exact one key word in [SQL components 2].
\item Extra Hard: All others left.
\end{itemize}
\end{document}
|
\section{Introduction \label{sec:intro}}
The purpose of this paper is to present a detailed semi-analytic analysis of models of ultra-relativistic
nuclear collisions where the early production of particles occurs from strings.
The strings are associated with wounded quarks, and both of their end-point positions fluctuate
in spatial rapidity. The model generalizes the analysis of~\cite{Broniowski:2015oif} where only one-end fluctuations were
considered.
The main assumptions are that the strings emit particles independently of one another and
that the production from a string is uniform between its end-points.
We obtain the one-body string emission function from a fit to the experimental data at $\sqrt{s_{NN}}=200$~GeV,
and use it to constrain the freedom in the distribution of the end-point positions. We then
explore in detail the two-body correlations in various scenarios for the fluctuating end-points.
The derived analytic formulas allow for a full understanding of this simple model. In particular,
we show that standard measures applied in analyses of the longitudinal fluctuations,
such as the Legendre $a_{nm}$ coefficients, fall between certain bounds. This
explains why a priori different models may provide quite similar results for these measures of the longitudinal correlations.
We find that the $a_{nm}$ coefficients can be significantly larger (by a factor of $\sim 3$) when one allows for two end-point to fluctuate,
compared to the case of single end-point fluctuations of~\cite{Broniowski:2015oif}. This observation is
relevant for phenomenological studies.
Since the model, despite its simplifications, is generic, sharing features with more complicated
string implementations, our findings shed light on correlations from other string models in application to
ultra-relativistic heavy-ion collisions.
The basic phenomenon explored in this paper and illustrated with definite calculations can be understood in very simple terms.
Consider a string with left and right end-points and an acceptance window in pseudorapidity. If the left end-point
were always left of the acceptance window,
and the right end-point to the right (they may fluctuate or not, but cannot enter the window), then the string seen in the
window is always the same, hence no fluctuations occur. If,
however, an endpoint via fluctuation enters the acceptance
window, then fluctuations occur, as its observed fragment may
be shorter or longer. The fluctuation effect is larger when both end-points fluctuate
into the acceptance window, which is the case explored in detail below.
The concept of wounded sources formed in the initial stages of ultra-relativistic heavy-ion collisions has proven to be phenomenologically successful in
reproducing multiplicity distributions from soft particle production. The idea (see~\cite{Bialas:2008zza} for
a discussion of the foundations), adopts the Glauber model~\cite{Glauber:1959aa} in its variant suitable for inelastic collisions~\cite{Czyz:1969jg}.
Whereas the wounded nucleon scaling~\cite{Bialas:1976ed}, when applied to the highest BNL Relativistic Heavy-Ion Collider (RHIC) or the
CERN Large Hadron Collider (LHC) energies, requires a sizable admixture of binary collisions~\cite{Back:2001xy,Kharzeev:2000ph},
the scaling based on wounded quarks~\cite{Bialas:1977en,Bialas:1977xp,Bialas:1978ze,Anisovich:1977av} works remarkably
well~\cite{Eremin:2003qn,KumarNetrakanti:2004ym,Bialas:2006kw,Bialas:2007eg,Alver:2008aq,Agakishiev:2011eq,Adler:2013aqf,Loizides:2014vua,Adare:2015bua,%
Lacey:2016hqy,Bozek:2016kpf,Zheng:2016nxx,Sarkisyan:2016dzo,Mitchell:2016jio,Chaturvedi:2016ctn,Loizides:2016djv,Tannenbaum:2017ixt}. Another
successful approach~\cite{Zakharov:2016bob,Zakharov:2016gyu} amends the wounded
nucleons with a meson-cloud component.
For mid-rapidity production, the wounded quark scaling takes the simple form
\begin{eqnarray}
N_{\rm ch} = k (\br{N_A}+\br{N_B}), \label{eq:wounded}
\end{eqnarray}
where $N_{\rm ch}$ is the number of charged hadrons in a mid-rapidity bin, and $\br{N_i}$ are the average numbers of wounded quarks in
nucleus $i$ in a considered centrality class.
The proportionality constant $k$ should not depend on centrality or the mass numbers of the nuclei (i.e., on the overall number of participants),
and indeed this requirement is satisfied to expected accuracy~\cite{Bozek:2016kpf,Tannenbaum:2017ixt}. Of course, $k$ increases with the collision energy.
When it comes to modeling the rapidity spectra, formula~(\ref{eq:wounded}) is replaced with
\begin{eqnarray}
\frac{dN}{d\eta} = \br{N_A} f(\eta)+\br{N_B} f(-\eta), \label{eq:woundeta}
\end{eqnarray}
where $f(\eta)$ is a universal (at a given collision energy) profile for emission from a wounded quark
(we adopt the convention that nucleus $A$ moves to the right and $B$ to the left).
For symmetric ($A=B$) collisions one only gets access to the the symmetric part of $f(\eta)$, as then
$\br{N_A}=\br{N_B}$. However, from asymmetric collisions, such as d-Au, one can also extract
the antisymmetric component in the wounded nucleon~\cite{Bialas:2004su} or wounded quark model~\cite{Barej:2017kcw,Adare:2018toe}
(for A-A collisions analogous analyses were carried out in~\cite{Gazdzicki:2005rr,Bzdak:2009dr,Bzdak:2009xq}),
with the finding that $f(\eta)$ is peaked
in the forward region, thus quite naturally emission is in the forward direction. However, $f(\eta)$ is widely spread in the whole kinematically available range.
The phenomenological result of the approximate triangular shape of the emission profile was later used in modeling the initial
conditions for further evolution, see e.g.~\cite{Adil:2005bb,Bozek:2010bi,Bozek:2013uha,Monnai:2015sca,Chatterjee:2017mhc}.
Microscopically, the approximate triangular shape of the emission function finds a natural origin in color string models, where
one end-point of the string is fixed, whereas the location of the other end-point fluctuates.
In particular,
in the basic Brodsky-Gunion-Kuhn mechanism~\cite{Brodsky:1977de}, the emission proceeds from strings in which
one end-point is associated with a valence parton, and the other end-point, corresponding to
wee partons, is randomly generated along the space-time rapidity $\eta$. When the distribution of the fluctuating end-point is
uniform in $\eta$, and so is the string fragmentation distribution, then the triangular
shape for the emission function follows.
Various Monte Carlo codes implementing the Lund string formation and
decays (see, e.g.,~\cite{Andersson:1983ia,Wang:1991hta,Lin:2004en,Sjostrand:2014zea,Bierlich:2018xfw,Ferreres-Sole:2018vgo}) or the dual parton
model/Regge-exchange approach~\cite{Capella:1992yb,Werner:2010aa,Pierog:2013ria}
also introduce strings of fluctuating ends, with various specific mechanisms and effects (baryon stopping, nuclear shadowing) additionally incorporated.
Apart from reproducing the measured one-body spectra, achieved by appropriate tune-ups of parameters, the incorporated initial-state
correlations show up
in event-by-event fluctuations that can be accessed experimentally.
Thus the fluctuating strings are standard objects used in modeling the early phase of high-energy reactions.
Our model joins the concept of wounded sources with strings in the following way:
\begin{enumerate}
\item Each wounded source has an associated string.
\item The strings emit particles independently of each-other.
\item The end-points of a string are generated universally (in the same manner for all wounded objects) from appropriate distributions.
\item The emission of particles from a string occurring between the end-points is homogeneous in spatial rapidity.
\end{enumerate}
In such a model, event-by-event fluctuations take
the origin from fluctuations of the number of wounded objects, as well as from fluctuations of the positions of the end points~\cite{Broniowski:2015oif}.
The goal of this paper is to study this generic model, with the focus on the end-point behavior which probes the underlying physics.
We take a general approach, with no prejudice as to how the end-points are fluctuating, but using the one-body emission profiles
obtained from experiment as a physical constraint.
More complicated mechanisms associated with dense systems,
such as the formation of color ropes~\cite{Biro:1984cf,Sorge:1995dp} or nuclear shadowing, are not incorporated in our picture. Also, we consider
one type of strings, which allows for simple analytic derivations.
We remark that associating a string with a leading quark is in the spirit of the Lund approach (cf. discussion of Sec.~5 in~\cite{Andersson:1983ia}).
So for simplicity we have in each event $N_i$ ``wounded strings'' associated with valence quarks in nucleus $i$. Other more complicated choices
(e.g, including the binary collisions) are also
possible here, but the advantage of our prescription is that by definition it complies with the experimental scaling of multiplicities of Eq.~(\ref{eq:wounded}).
A specific implementation of some ideas explored in this work, with strings that have one end fixed and the other
fluctuating, has been presented in~\cite{Broniowski:2015oif}.
The outline of our paper is as follows:
In Sec.~\ref{sec:wound} we use the rapidity spectra from d-Au and Au-Au reactions
at $\sqrt{s_{NN}}=200$~GeV to obtain the one-body emission profile of the wounded quark.
In Sec.~\ref{sec:model} we explore our generic string model and derive
simple relations between string end-point distributions and n-body-emission profiles for the radiation from
individual strings.
Section~\ref{sec:end} discusses how a given one-body-emission profile can correspond to a
family of different functions for the string end-point distributions.
Two-body correlations from a single string are discussed in Sec.~\ref{sec:corr}, whereas in
Sec.~\ref{sec:corrm} they are combined to form the two-body correlations in nuclear collisions.
Section~\ref{sec:anm} presents the Legendre $a_{nm}$ coefficients of the two-particle correlations. Finally, Sec.~\ref{sec:concl} draws the final conclusions from our work.
Some more technical developments can be found in the appendices.
\section{Emission profiles from wounded quarks \label{sec:wound}}
\begin{figure}
\hfill \includegraphics[trim={22pt 0 0 0},clip,width=0.48\textwidth]{dAu.pdf}\\
\hfill \includegraphics[trim={22pt 0 0 0},clip,width=0.47\textwidth]{dAusym.pdf}\\
\hfill \includegraphics[trim={26pt 0 0 0},clip,width=0.48\textwidth]{dAuasym.pdf}\\
\vspace{-5mm}
\caption{One-particle emission profiles obtained in the wounded quark model via Eqs.~(\ref{eq:woundeta}-\ref{eq:s_a_wound}) from the PHOBOS
rapidity spectra for d-Au collisions at $\sqrt{s_{NN}}=200$~GeV~\cite{Back:2004mr} in the indicated centrality classes (a), together with
the corresponding symmetric (b) and antisymmetric (c) components.
The shaded bands show the experimental uncertainties (propagated via the Gaussian method) for
the $40-60$\% and $60-80$\% centrality classes, as well as for the PHOBOS minimum bias data~\cite{Back:2003hx}. \label{fig:dAu}}
\end{figure}
We begin by obtaining from experimental data the emission profiles of Eq.~(\ref{eq:woundeta}), needed in the following sections.
We use the method of~\cite{Bialas:2004su}, which has also been
applied recently to wounded quarks in~\cite{Barej:2017kcw}. With
\begin{eqnarray}
&& f_s(\eta) = \frac{1}{2}[f(\eta)+f(-\eta)], \;\; f_a(\eta) = \frac{1}{2}[f(\eta)-f(-\eta)], \nonumber \\
&& N_+=N_A+N_B, \;\; N_-=N_A-N_B,
\end{eqnarray}
one gets immediately
\begin{eqnarray}
&& f_s(\eta)=\frac{dN/d\eta(\eta)+dN/d\eta(-\eta)}{\br{N_+}}, \nonumber \\
&& f_a(\eta)=\frac{dN/d\eta(\eta)-dN/d\eta(-\eta)}{\br{N_-}}.
\label{eq:s_a_wound}
\end{eqnarray}
For asymmetric collisions both parts of the profile can be obtained, whereas for symmetric collisions one can only get $f_s(\eta)$.
\begin{figure}
\begin{center}
\includegraphics[trim={10pt 0 0 0},clip,width=0.48\textwidth]{AuAuuniversality.pdf}
\end{center}
\vspace{-7mm}
\caption{One-particle emission profiles obtained in the wounded quark model from the PHOBOS rapidity
spectra for Au-Au collisions at $\sqrt{s_{NN}}=200$~GeV~\cite{Back:2002wb} in the indicated centrality classes. The shaded bands give the experimental uncertainties
(propagated via the Gaussian method) for the most central and the most peripheral case. \label{fig:AuAu}}
\end{figure}
\begin{figure}
\hfill \includegraphics[trim={2pt 0 0 0},clip,width=0.48\textwidth]{dAufit.pdf}\\
\hfill \includegraphics[trim={10pt 0 0 0},clip,width=0.482\textwidth]{AuAufit.pdf}\\
\vspace{-5mm}
\caption{
Comparison of the wounded-quark model predictions (lines) with the experimental rapidity spectra (points) for d-Au~\cite{Back:2004mr} (a) and Au-Au~\cite{Back:2002wb} (b) collisions,
with the experimental uncertainties shown as shaded bands. A universal profile discussed in the text is taken for the model calculations in all cases.
\label{fig:dAu+AuAu}}
\end{figure}
If the wounded-quark scaling works, then the profiles obtained with different centrality classes or mass numbers of the colliding nuclei should be universal, depending
only on the collision energy. To what extent this is the case, can be assessed
from Figs.~\ref{fig:dAu} and \ref{fig:AuAu}, which show the one-particle emission
profiles that were extracted from experimental data on d-Au and Au-Au collisions from the
PHOBOS data~\cite{Back:2003hx,Back:2004mr,Back:2002wb} in the framework of the wounded quark model.
To this end, the symmetric (for both reactions) and antisymmetric components (only in the case of the d-Au
collisions) were obtained from the experimental data on rapidity spectra by means of Eq.~(\ref{eq:s_a_wound}),
where the valence quark multiplicities $\br{N_\pm}$ were obtained from
{\tt GLISSANDO}~\cite{Broniowski:2007nz,Rybczynski:2013yba}, a Monte-Carlo simulator of the Glauber model.
Figure~\ref{fig:dAu} shows the results for the one-particle-emission profiles $f^{\rm dAu}(\eta)$ extracted from the PHOBOS data~\cite{Back:2003hx,Back:2004mr}
for d-Au collisions, together with their symmetric and antisymmetric components.
In general, the curves for various centrality classes, considering the propagated experimental errors, can be viewed as coinciding.
The apparent exception to this behavior is seen in the symmetric part of the profile for the peripheral centrality $60\%-80\%$, which is
significantly larger for $|\eta|<3$, cf. Fig.~\ref{fig:dAu}(b). We note that for d-Au collisions
this peripheral class corresponds to $\langle N_+\rangle$ in the range from six to eight sources, which are tiny values, where the model admittedly does not work.
It can thus confirm the
findings of~\cite{Barej:2017kcw} that the assumption of universality of the one-particle emission profiles works reasonably well for the central to mid-peripheral d-Au
collisions, whereas it starts to differ for more peripheral centrality classes.
Figure~\ref{fig:AuAu} presents our results for the one-particle emission profiles $f^{\rm AuAu}(\eta)$ extracted from the PHOBOS data~\cite{Back:2002wb}
for Au-Au collisions. As already mentioned, in this case only the symmetric parts of the emission profiles can be obtained.
It can be seen that the results for $f^{\rm AuAu}(\eta)$
in various centrality classes agree remarkably well with one another. They also approximately agree with the symmetric profiles for d-Au collisions
of Fig.~\ref{fig:dAu}(b).
Finally, we test if our method reproduces the PHOBOS charged particle rapidity spectra
for combined d-Au and Au-Au collisions. To this end we take a single ``universal'' $f(\eta)$, consisting
of an antisymmetric part extracted from the minimum-bias d-Au spectra and a symmetric part taken as
the average of the different one-particle emission profiles of Au-Au collisions shown in Fig.~\ref{fig:AuAu}.
The charged particle rapidity spectra $dN_{\rm ch}/d\eta$ were calculated by means of Eq.~(\ref{eq:woundeta}) with this universal $f(\eta)$,
where again the numbers $\br{N_A}$ and $\br{N_B}$ were generated with {\tt GLISSANDO}.
Figure~\ref{fig:dAu+AuAu} shows the resulting one-particle-emission spectra for d-Au and Au-Au collisions obtained that way,
together with the corresponding experimental data from PHOBOS~\cite{Back:2003hx,Back:2004mr,Back:2002wb}:
As expected from Fig.~\ref{fig:AuAu}, the rapidity spectra for the Au-Au collisions, which are almost symmetric, are very well reproduced by the chosen $f(\eta)$.
Also the rapidity spectra for the d-Au collisions, which largely depend on both the symmetric and antisymmetric contribution to $f(\eta)$, are qualitatively well reproduced for
$|\eta|<4$, except for the above-discussed case of the peripheral collisions.
Therefore, we conclude that the wounded quark model with the universal profile function $f(\eta)$ reproduces the experimental rapidity spectra
at $\sqrt{s_{NN}}=200$~GeV
in a way satisfactory for our exploratory study.\footnote{We note that the analogous analysis at the LHC leads to somewhat less accurate agreement, which calls for
improvement of the model.}
In the following analysis of the rapidity fluctuations, we use the $f(\eta)$ obtained here to constrain the string end-point distributions.
\section{Generic string model \label{sec:model}}
In this section we describe a model of generic production from a single string formed in the early phase of the collision process.
Suppose the string is pulled by two end-points placed at spatial rapidities $y_{1}$ and $y_{2}$, whose locations are generated according to a probability
distribution $g(y_1,y_2)$ (if the end-points are generated in an uncorrelated manner, then $g(y_1,y_2)=g_1(y_1)g_2(y_2)$, as will
be assumed shortly). The emission of a particle with rapidity $\eta$ from the string fragmentation process
is assumed to be uniformly distributed along the string, i.e., it is equal to
\begin{eqnarray}
\!\!\!\!\! s(\eta;y_1,y_2)= \omega \left [ \theta(y_1<\eta<y_2) + \theta(y_2<\eta<y_1) \right ], \label{eq:uni}
\end{eqnarray}
where $\omega$ is a dimensionless constant determining the production strength and $\theta(c)$ imposes the condition $c$.
Note that we include the cases of $y_2>y_1$ and $y_1>y_2$, which
may seem redundant but which is needed, for instance, when the two end-points correspond to different partons in a given model.
Let us introduce the short-hand notation
\begin{eqnarray}
\int_{\cal Y} dy_1 dy_2 \, g(y_1,y_2) X = \langle X \rangle_{\cal Y},
\end{eqnarray}
with ${\cal Y}$ denoting the two-dimensional range of integration, depending on the kinematic constraints and/or detector coverage, and $X$ meaning
any expression.
The single-particle density for production from a string upon averaging over the fluctuation of the end-points is therefore
\begin{eqnarray}
f(\eta) = \langle s(\eta;y_1,y_2) \rangle_{\cal Y} ,
\label{eq:1}
\end{eqnarray}
Analogously, for the $n$-particle production ($n\ge 2$) from a single string we have
\begin{eqnarray}
f_n(\eta_1,\dots,\eta_n) =\langle s(\eta_1;y_1,y_2) \dots s(\eta_n;y_1,y_2) \rangle_{\cal Y}, \label{eq:n}
\end{eqnarray}
where we have assumed independent production of the $n$ particles.
In case the string ends are generated independently of each other, one has
\begin{eqnarray}
\langle X \rangle_{\cal Y} = \int dy_1 dy_2 g_1(y_1)g_2(y_2) X\, ,
\end{eqnarray}
where the limits of integration in $y_i$ are formally from $-\infty$ to $\infty$, with the support taken care of by the forms
of $g_{i}(y_i)$. Then we readily find that the one-body emission profile is
\begin{eqnarray}
f(\eta) &=& \omega \left \{ G_1(\eta)[1-G_2(\eta)] + G_2(\eta)[1-G_1(\eta)] \right \} \nonumber \\
&=& \omega \left \{ \tfrac{1}{2} - 2 [G_1(\eta)-\tfrac{1}{2}] [G_2(\eta)-\tfrac{1}{2}] \right \}, \label{eq:f1G}
\end{eqnarray}
where the appropriate cumulative distribution functions (CDFs) are defined as
\begin{eqnarray}
G_{i}(y)=\int_{-\infty}^y dy' \, g_{i}(y') .
\end{eqnarray}
The profile $f(\eta)$ acquires a specific value at the arguments $\eta_1$ and $\eta_2$ where the CDFs reach $\tfrac{1}{2}$, i.e.,
\begin{eqnarray}
&& \eta_1^{(0)} \; : \; G_1(\eta_1^{(0)})=\tfrac{1}{2}, \nonumber \\
&& \eta_2^{(0)} \; : \; G_2(\eta_2^{(0)})=\tfrac{1}{2}. \label{eq:e1e2}
\end{eqnarray}
Then from Eq.~(\ref{eq:f1G}) we obtain
\begin{eqnarray}
\omega=2 f(\eta_1^{(0)}) = 2 f(\eta_2^{(0)}). \label{eq:1o2}
\end{eqnarray}
This equation provides a special meaning to the constant $\omega$.
Furthermore, since $0 \le G_{1,2}(\eta) \le 1$, Eq.~(\ref{eq:f1G}) yields the limit
\begin{eqnarray}
0 \le f(\eta)\le \omega.
\label{eq:fupbound}
\end{eqnarray}
The above features will be explored shortly in a qualitative discussion.
Similarly, for the $n$-particle distributions with $n\ge2$ we have
\begin{eqnarray}
&& f_n(\eta_1,\dots,\eta_n) = \omega^n \left \{ \right. \label{eq:fnG} \\
&& ~~G_1({\rm min}(\eta_1,\dots,\eta_n))[1-G_2({\rm max}(\eta_1,\dots,\eta_n))] + \nonumber \\
&& \left . ~~G_2({\rm min}(\eta_1,\dots,\eta_n))[1-G_1({\rm max}(\eta_1,\dots,\eta_n))] \right \}. \nonumber
\end{eqnarray}
We thus see that in the model with two end-points fluctuating (the relevant assumptions are the uniform string fragmentation (\ref{eq:uni}) and the independence
of the two end-point locations)
all the information carried by the $n$-particle densities produced from a single string is encoded solely in the cumulative distributions
functions $G_1$ and $G_2$.
It is obvious, however, that $G_1$ and $G_2$ cannot be separately determined from the one-body distributions in an unambiguous manner,
hence a large degree of freedom is still left in the model after fixing the rapidity spectra. Yet, the one body distribution provides, via Eq.~(\ref{eq:f1G}), an
important constraint. Our method of matching $G_1$ and $G_2$ to the one-body function $f(\eta)$ is explained in detail in Appendix~\ref{app:match}.
As we stress, there is no uniqueness in the procedure, but there is a systematic way of approaching the problem, allowing one to explore the range of possibilities.
We denote the position of the maximum of $f(\eta)$ as $\eta_{\rm max}$.
We consider three cases:
\begin{enumerate}
\item[i)] The distributions of both end-points are equal, $g_1(\eta)=g_2(\eta)$, Eq.~(\ref{eq:case1}).
In this case $\omega=2f(\eta_{\rm max})$, with $\eta_{\rm max}=\eta_1^{(0)}=\eta_2^{(0)}$.
\item[ii)] The supports of distributions $g_1(\eta)$ and $g_2(\eta)$ do not overlap, Eq.~(\ref{eq:case2}).
In this case $\omega=f(\eta_{\rm max})$ and $\eta_2^{(0)} < \eta_{\rm max}<\eta_1^{(0)}$.
\item[iii)] The form of $g_1(\eta)$ is motivated by parton distribution functions (PDFs) of valence quarks, Eq.~(\ref{eq:pdffit}), and $g_2(\eta)$
is adjusted according to Eq.~(\ref{eq:h2}).
\end{enumerate}
Cases i) and ii) are in a sense most different, showing the span of possibilities formally allowed, whereas case iii) is intermediate.
For case iii) we use the parametrization of the valence quark PDF given by Eq.~(\ref{eq:pdffit}) with parameters
$\alpha=-0.5$ and $\beta=3$, which are typical values at low scales.
We have found that using other reasonable parametrizations has very small influence on our results, with case iii) always remaining
close to case i).
We stress that all the considered cases reproduce, by construction, the one-body emission profiles $f(\eta)$.
We end this section with remarks concerning the model with one end of the string fixed and the other one fluctuating, explored in~\cite{Broniowski:2015oif}.
This simplified version
can be obtained as a special limit from Eqs.~(\ref{eq:f1G},\ref{eq:fnG})
by choosing \mbox{$g_1(\eta)=\delta(\eta-y_{\rm max})$}, which is equivalent of taking, correspondingly, $G_1=0$
for $\eta < y_{\rm max}$, {\em i.e.},
\begin{eqnarray}
f(\eta) &=& \omega G_2(\eta), \label{eq:f1Gsingle} \\
f_n(\eta_1,\dots,\eta_n) &=& \omega^n G_2({\rm min}(\eta_1,\dots,\eta_n)). \nonumber
\end{eqnarray}
We note immediately that this model cannot reproduce $f(\eta)$ for $\eta > \eta_{\rm max}$, as $G_2(\eta)$ cannot decrease.
Thus the model is limited to $\eta \le \eta_{\rm max}$, which, however, is not a problem if we are only interested in the mid-rapidity region.
Moreover, in this region the single-end fluctuating model corresponds precisely to case ii) of the two-end fluctuations. This is obvious from the
following argumentation: When the right end of the string is fluctuating outside of the acceptance region, it is irrelevant if it fluctuates or if it is
fixed, as in both cases we only observe the production from the part of the string falling into the acceptance range. In that situation
(or more precisely for $\eta \le \eta_{\rm max}$) Eqs.~(\ref{eq:f1G},\ref{eq:fnG}) reduce to Eqs.~(\ref{eq:f1Gsingle}). Hence, the single
end-point fluctuation model of~\cite{Broniowski:2015oif} corresponds to the present case ii) at $\eta \le \eta_{\rm max}$, and is not applicable for
$\eta > \eta_{\rm max}$.
\section{End-point distributions \label{sec:end}}
We now come to the discussion of the end-point distributions subjected to the requirement that the one-body
emission profiles are reproduced.
\begin{figure}
\hfill \includegraphics[trim={11pt 0 1pt 0},clip,width=0.49\textwidth]{g1g2PDF_witholdmodelv2.pdf}\\ \vspace{-3mm}
\hfill \includegraphics[trim={13pt 0 1pt 0},clip,width=0.49\textwidth]{G1G2_witholdmodelv2.pdf}\\
\vspace{-5mm}
\caption{Distribution functions $g_1$ and $g_2$ (a) and cumulative distribution functions $G_1$ and $G_2$
(b) of the string end-points for the cases of i) $\omega=2f(\eta_{\rm max})$, ii)~$\omega=f(\eta_{\rm max})$, and iii)~$\alpha=-0.5,$ $\beta=3$,
as indicated in the legend. The light dot-dashed line corresponds to the model with one end-point fixed and the other one fluctuating~\cite{Broniowski:2015oif},
which overlaps with case ii) for $\eta \le \eta_{\rm max}$. The vertical line in panel (b) indicates $\eta=\eta_{\rm max}$.
See the text for further details. \label{fig:g1g2}}
\end{figure}
Figure~\ref{fig:g1g2}a) shows the distributions of the string end-points, $g_1$ and $g_2$, for the three cases, and Fig.~\ref{fig:g1g2}b) the corresponding CDFs, $G_1$ and $G_2$.
The shaded bands give an estimate of the errors due to the experimental uncertainty $\Delta f$ for the one-particle emission profile $f$.
In the case of Fig.~\ref{fig:g1g2}b), the upper limit of the shaded bands corresponds to the values of $G_{1,2}$ that are matched to the
one-body profile $f+\Delta f$, whereas the lower limits are matched to $f-\Delta f$.
For these upper and lower limits of $G_{1,2}$, the derivatives in $\eta$ yield the upper and lower limits of the shaded bands
for $g_1$ and $g_2$ depicted in Fig.~\ref{fig:g1g2}a).
For case iii) a shaded band is given only for $g_2$ ($G_2$). This is because by construction $g_1$ ($G_1$) coming from PDFs are
assumed to be accurate and all uncertainty is therefore attributed to $g_2$ ($G_2$).
In case i) $g_1(\eta)=g_2(\eta)$, hence the distributions are indicated with a single curve (solid line) in Figs.~\ref{fig:g1g2}a) and b).
We note that the distribution of $g_{1}(\eta)$ peaks at forward rapidity (the Au side),
as expected from the shape of the one-body profile $f(\eta)$ in Fig.~\ref{fig:dAu}. The CDF crosses
the value $1/2$ at $\eta_1^{(0)}=\eta_2^{(0)}=\eta_{\rm max}\simeq 2.5$,
which coincides with the maximum of $f(\eta)$.
In case ii) (dashed lines in Fig.~\ref{fig:g1g2}) the supports for $g_1$ and $g_2$ are disjoint. In Fig.~\ref{fig:g1g2}a) the left part of the curve, up to the
point $\eta_{\rm max} \simeq 2.5$ (indicated with a vertical line), corresponds to $g_2$, and the right part to $g_1$. Hence,
the string end-points always follow the ordering $y_1 \ge y_2$,
which does not hold in the other cases.
Figure~\ref{fig:g1g2}b) shows the corresponding CDFs, with $G_1=0$ left from $\eta_{\rm max}$, and $G_2=1$ right from $\eta_{\rm max}$.
In Appendix~\ref{app:match}
we show that $G_1$ and $G_2$ from case ii) are the lower and upper limits for any CDFs in the considered problem. Indeed, the CDFs from the other two cases
fall in between these limiting curves.
Case iii), based on a valence quark PDF for $g_1$, represents an intermediate class of distributions falling between
cases i) and ii). The curves corresponding to the valence quark are dotted and with no error bands. The
distribution $g_1$ (valence quark) is peaked in the forward direction, as expected.
We note that $y_1>y_2$ is favored, although $y_2<y_1$ is also possible.
With the parametrization we used of the valence quark distribution, the CDFs in case iii) are not far from case i).
We have checked that this feature holds also for other reasonable parametrizations of the valence quark PDF.
We underline again that all the cases of Fig.~\ref{fig:g1g2}, which exhibit radically different end-point distributions, reproduce by construction the
one-body emission profile $f(\eta)$.
\section{Correlations from a single string \label{sec:corr}}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{cov_sym.pdf}
\includegraphics[width=0.48\textwidth]{cov_disjoint.pdf}
\includegraphics[width=0.48\textwidth]{cov_-p5_3.pdf}
\end{center}
\vspace{-5mm}
\caption{Covariance for the emission from a single string for cases i) (a), ii) (b) and iii) (c). \label{fig:cov}}
\end{figure}
As we show in this section, the two-particle correlation is sensitive to the particular form of the distributions and differs between cases i), ii), and iii).
A convenient quantity is the covariance of the two-particle emission
from a single string, defined as
\begin{eqnarray}
{\rm cov}(\eta_1,\eta_2)={f_2(\eta_1,\eta_2)} - {f(\eta_1)} {f(\eta_2)}, \label{eq:cov}
\end{eqnarray}
where $f_2$ is given by Eq.~(\ref{eq:fnG}).
Explicitly,
\begin{eqnarray}
&& {\rm cov}(\eta_1,\eta_2)=\omega^2 \big \{ G_1({\rm min}(\eta_1,\eta_2))[1-G_2({\rm max}(\eta_1,\eta_2))] \nonumber \\
&& ~~~ + G_2({\rm min}(\eta_1,\eta_2))[1-G_1({\rm max}(\eta_1,\eta_2))] \nonumber \\
&& ~~~ - \left ( G_1(\eta_1)[1-G_2(\eta_1)] + G_2(\eta_1)[1-G_1(\eta_1)] \right ) \nonumber \\
&& ~~~ \times \left ( G_1(\eta_2)[1-G_2(\eta_2)] + G_2(\eta_2)[1-G_1(\eta_2)] \right ) \big \}. \label{eq:cov2}
\end{eqnarray}
A simplification occurs along the diagonal $\eta_1=\eta_2=\eta$, where
\begin{eqnarray}
{\rm cov}(\eta,\eta) &=& \omega^2 \big \{ \tfrac{1}{4}\!-\! 4 [G_1(\eta)-\tfrac{1}{2}]^2 [G_2(\eta)-\tfrac{1}{2}]^2 \big \} \nonumber \\
&=& f(\eta)[\omega-f(\eta)]. \label{eq:diag}
\end{eqnarray}
Also, the leading expansion at the diagonal in the anti-diagonal direction, with $\eta_1=\eta +\delta$ and $\eta_2=\eta -\delta$, yields a very
simple formula,
\begin{eqnarray}
&& {\rm cov}(\eta\!+\!\delta,\eta\!-\!\delta)={\rm cov}(\eta,\eta) - \omega^2 [g_1(\eta)\!+\!g_2(\eta)] |\delta| + {\cal O}(\delta^2). \nonumber \\ \label{eq:delta}
\end{eqnarray}
Figure~\ref{fig:cov} shows the resulting distributions for ${\rm cov}(\eta_1,\eta_2)$ for the three considered cases.
One observes vivid qualitative differences between the covariances in cases i) and ii), cf. Figs.~\ref{fig:cov}a) and b). Whereas in case i)
the covariance exhibits a monotonously increasing ridge along the $\eta_1=\eta_2$ direction, the covariance in case ii) shows a double peak structure,
with a zero at $\eta=\eta_{\rm max} \simeq 2.5$, which corresponds to the zero of $g_1$ and $g_2$ in Fig.~\ref{fig:g1g2}a). At this point
$G_1(\eta)=0$ and $G_2(\eta)=1$, which upon substitution to Eq.~(\ref{eq:cov2}) yields zero.
Another difference is in magnitude of the covariance, which in case i) is significantly larger than in case ii).
The covariance in case iii) is very close to case i) (cf. Figs.~\ref{fig:cov}a) and c)).
Some small difference can be seen where $\eta_1$ is small(large), but $\eta_2$ large(small),
where in case iii) the covariance noticeably drops to negative values.
We also note that in all cases the values on the diagonal is obeying Eq.~(\ref{eq:diag}). The fall-off from the diagonal in the anti-diagonal
direction is given by the second term in Eq.~(\ref{eq:delta}). We note that the slope is proportional to \mbox{$4 f(\eta_{1,2}^{(0)})[g_1(\eta)+ g_2(\eta)]$},
hence two models which have similar values of $\eta_{1,2}^{(0)}$ and close sums of the two end-point distributions, $g_1(\eta)+ g_2(\eta)$, will have
similar covariances in the vicinity of the diagonal. Both conditions are satisfied between models i) and iii). In particular, we can see that the sum
$g_1(\eta)+ g_2(\eta)$ for model iii) in Fig.~\ref{fig:g1g2}a) (dotted lines) is close to twice $g_{1,2}(\eta)$ for model i) (solid line).
Thus the reason for the similarity of correlations in
cases (i) and (iii) may be traced back to Eq.~(\ref{eq:delta}),
which shows that this is the average of $g_1(\eta)$ and $g_2(\eta)$, which controls the fall-off of the
correlation from the diagonal. These averages happen to be very similar when we use any reasonable parametrization of
the parton distribution function giving the PDF of one end-point distribution, and the fluctuations the other end-point are adjusted
to match the profile function $f(\eta)$, as explained in Sec.~\ref{sec:end}.
\section{Correlations from multiple strings \label{sec:corrm}}
As already discussed in the introduction, in our approach the strings ``belong'' to the valence
quarks either from nucleus A or from nucleus B.
With the underlying assumptions of independent wounded sources, the expressions for the
$n$-body distributions account for the combinatorics in a simple manner, with the particles at rapidities $\eta_i$
being products from a string belonging to $A$ or to $B$. For the one-body density
in A-B collisions one finds
\begin{eqnarray}
{f_{AB}(\eta)} = \br{N_A} {f_A(\eta)} + \br{N_B} {f_B(\eta)}, \label{eq:onebody}
\end{eqnarray}
where $\br{N_A}$ and $\br{N_B}$ are the event-by-event
average numbers of wounded sources in nuclei A and B, respectively,
and $f_{A,B}(\eta)=f(\pm \eta)$ denote the profiles for the emission from a single string, as given by Eq.~(\ref{eq:1}), associated with sources from nuclei $A$ or $B$.
We work in the nucleon-nucleon center-of-mass (CM) frame, hence $f_{A}(\eta)=f_{B}(-\eta)$.
Analogously, one can define the two-body distribution for emission from a single string in nuclei $A$ and $B$ as $f_{A,B}(\eta_1,\eta_2)=f_2(\pm\eta_1,\pm\eta_2)$,
and the corresponding covariances as ${\rm cov}_{A,B}(\eta_1,\eta_2)={\rm cov}(\pm\eta_1,\pm\eta_2)$.
Then, one readily obtains the covariance for the production in A-B collisions (see Appendix \ref{app:comb}) in the form
\begin{eqnarray}
&& {\rm cov}_{AB}(\eta_1, \eta_2) \equiv f_{AB}(\eta_1,\eta_2) - f_{AB}(\eta_1) f_{AB}(\eta_2) \nonumber \\
&=& \br{N_A} {\rm cov}_A(\eta_1,\eta_2) + \br{N_B} {\rm cov}_B(\eta_1,\eta_2) \label{eq:gen} \\
&+& {\rm var}(N_A) {f_A(\eta_1)}{f_A(\eta_2)} + {\rm var}(N_B) {f_B(\eta_1)}{f_B(\eta_2)} \nonumber \\
&+& {\rm cov}(N_A,N_B) \left [{f_A(\eta_1)} {f_B(\eta_2)}+ {f_B(\eta_1)} {f_A(\eta_2)} \right ]. \nonumber
\end{eqnarray}
In the special case of symmetric collisions Eq.~(\ref{eq:gen}) simplifies into
\begin{eqnarray}
&& {\rm cov}_{AB}(\eta_1, \eta_2) = \br{N_A} {\rm cov}_A(\eta_1,\eta_2) + \br{N_B} {\rm cov}_B(\eta_1,\eta_2) \nonumber \\
&& + {\rm var}(N_+) {f_s(\eta_1)} {f_s(\eta_2)} + {\rm var}(N_-) {f_a(\eta_1)} {f_a(\eta_2)}, \label{eq:gensym}
\end{eqnarray}
where $N_-=N_A-N_B$.
The moments of $N_A$ and $N_B$ evaluated with {\tt GLISSANDO} are listed in Appendix~\ref{app:GL}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{symcentralcorr.pdf}
\includegraphics[width=0.48\textwidth]{asymcentralcorr.pdf}
\includegraphics[width=0.48\textwidth]{symcentralcorrFB.pdf}
\end{center}
\vspace{-5mm}
\caption{
Correlations $C_{AB}(\eta_1,\eta_2)$ for the $6\%$ most central Au-Au collisions for model cases i) (a) and ii) (b), as well as $C^\ast_{AB}(\eta_1,\eta_2)$
for case i) (c). \label{fig:FBcorr}}
\end{figure}
We also introduce the customary correlation $C$ defined as
\begin{eqnarray}
C_{AB}(\eta_1,\eta_2) =1+ \frac{{\rm cov}_{AB}(\eta_1,\eta_2)}{f_{AB}(\eta_1)f_{AB}(\eta_2)}, \label{eq:C}
\end{eqnarray}
which is a convenient measure due to its intensive property.
For symmetric collisions Eq.~(\ref{eq:C}) becomes
\begin{eqnarray}
C_{AB}(\eta_1,\eta_2) =1+ \frac{{\rm cov}_{AB}(\eta_1,\eta_2)}{\br{N_+}^2 f_s(\eta_1)f_s(\eta_2)}, \label{eq:Csym}
\end{eqnarray}
To separate the contribution from the string end-point fluctuations, we also define
\begin{eqnarray}
C^\ast_{AB}(\eta_1,\eta_2) = \frac{\br{N_A} {\rm cov}_A(\eta_1,\eta_2) + \br{N_B} {\rm cov}_B(\eta_1,\eta_2) }{f_{AB}(\eta_1)f_{AB}(\eta_2)}. \nonumber \\
\label{eq:Cast}
\end{eqnarray}
We note that Eq.~(\ref{eq:gen}) or (\ref{eq:gensym}) contain terms with two classes of fluctuations: those stemming from single string end-point fluctuations, containing
${\rm cov}_i(\eta_1,\eta_2)$, which were the object of study in the previous section,
and the remaining terms~\cite{Bzdak:2012tp} with moments of fluctuations of the numbers of wounded quarks, $N_A$ and $N_B$.
Therefore
the correlation function $C(\eta_1,\eta_2)$ contains a mixture of both effects. In principle, one could
separate these effects via the technique of partial covariance (see, e.g.,~\cite{Cramer:1946,krzanowski:2000}),
which effectively imposes constraints on a multivariate sample. The details of such an analysis, which leads to very simple and practical
expressions, were presented in~\cite{Olszewski:2017vyg}.
In the present case, however, such an analysis is not necessary if we have in mind the standard $a_{nm}$ coefficients discussed in Sec.~\ref{sec:anm}.
As is clear from Eq.~(\ref{eq:Csym}), the term ${\rm var}(N_+) {f_s(\eta_1)} {f_s(\eta_2)}$ in Eq.~(\ref{eq:gensym})
brings in a constant ${\rm var}(N_+)/\br{N_+}^2$ into $C(\eta_1,\eta_2)$.
Therefore it only changes its baseline and does not affect the $a_{nm}$ coefficients (for $n,m\ge 0$).
As we shall shortly see, the string end-point
fluctuations given by the term with $\br{N_A} {\rm cov}_A(\eta_1,\eta_2) + \br{N_B} {\rm cov}_B(\eta_1,\eta_2)$
are largely dominant over the Bzdak-Teaney~\cite{Bzdak:2012tp} term, ${\rm var}(N_-) {f_a(\eta_1)} {f_a(\eta_2)}$,
with the later entering at a level of 10-20\% in $a_{11}$ (cf. Sec.~\ref{sec:anm}). Hence
one may simply take the view that measuring the $a_{nm}$ coefficients associated with $C(\eta_1,\eta_2)$ essentially provides information on
the string end-point fluctuations, with
only a small contamination by the fluctuation of the number of sources.
Panels a) and b) of Fig.~\ref{fig:FBcorr} show our results for $C_{AB}(\eta_1,\eta_2)$ of the $6\%$ most central Au-Au collisions
at $\sqrt{s_{NN}}=200$~GeV in cases i) and ii) of our model.
The correlations exhibit a ridge structure along the $\eta_1=\eta_2$ direction, which
simply reflects the presence of the ridges in the single-string fluctuations displayed in Fig.~\ref{fig:cov}.
The correlation in case iii) is very close to case i), simply reflecting the behavior of Fig.~\ref{fig:cov}, hence we do not include it in the plot.
Panel c) shows the correlation stemming from the fluctuation of the string end-point, $C^\ast_{AB}(\eta_1,\eta_2)$ of Eq.~(\ref{eq:Cast}).
We note that, apart for an overall shift by a constant, it
is very similar to the correlation $C_{AB}(\eta_1,\eta_2)$ of Eq.~(\ref{eq:C}),
which indicates an important feature shown by our study:
The shape of the correlation function $C(\eta_1,\eta_2)$ is largely dominated by the string end-point fluctuations,
whereas the effects of the fluctuations of the number of sources are small.
\section{$a_{nm}$ coefficients \label{sec:anm}}
\begin{figure}
\hfill \includegraphics[width=0.48\textwidth]{a11v2log_witholdmodelv1.pdf}\\ \vspace{3mm}
\hfill \includegraphics[width=0.48\textwidth]{a11FBv2log_witholdmodelv1.pdf}\\ \vspace{3mm}
\hfill \hspace*{-0pt}\includegraphics[trim={3.pt 0 0pt 0},clip,width=0.44\textwidth]{a11FB_a11_ratio_witholdmodelv1.pdf}\:\!\:\!\\
\vspace{-2mm}
\caption{$a_{11}$ (a) and $a^\ast_{11}$ (b) for Au-Au collisions at $\sqrt{s_{\rm NN}}=200$~GeV as
a function of $\br{N_+}$ (the selected values for $\br{N_+}$ correspond to the $6$ centrality classes $0-6\%$, $6-15\%$, $15-25\%$, $25-35\%$,
$35-45\%$, and $45-55\%$) in cases i) with $\omega=2f(\eta_{\rm max})$, ii) with $\omega=f(\eta_{\rm max})$, iii) with $\alpha=-0.5,\,\beta=3$, together with the model of~\cite{Broniowski:2015oif} with one end point fixed, as indicated in the legend. Panel c) displays the ratio $a^\ast_{11}/a_{11}$.
To enhance visibility, the markers for the overlapping cases are slightly shifted to the left or right along the abscissa. \label{fig:a11}}
\end{figure}
For a given correlation function $C(\eta_1,\eta_2)$, the $a_{nm}$ coefficients are defined
as~\cite{Bzdak:2012tp,ATLAS:2015kla,ATLAS:anm}
\begin{eqnarray}
a_{nm} &=& \int_{-Y}^Y \frac{d \eta_1}{Y} \int_{-Y}^Y \frac{d \eta_2}{Y}\frac{1}{\mathcal{N}_C} C(\eta_1,\eta_2)
T_n\left(\frac{\eta_1}{Y}\right) T_m\left(\frac{\eta_2}{Y}\right), \nonumber \\ \label{eq:anmC}
\end{eqnarray}
with the normalization constant
\begin{equation}
\mathcal{N}_C=\int_{-Y}^Y \frac{d \eta_1}{Y} \int_{-Y}^Y \frac{d \eta_2}{Y} C(\eta_1,\eta_2),
\end{equation}
where $[-Y,Y]$ is the covered pseudorapidity range. Having in mind the typical pseudorapidity acceptance at RHIC, we use $Y=1$.
The functions $T_n(x)$ form a set of orthonormal polynomials.
The choice used in~\cite{ATLAS:2015kla,ATLAS:anm,Jia:2015jga} is
\begin{eqnarray}
T_n(x)=\sqrt{n+1/2}P_n(x),
\end{eqnarray}
where $P_n(x)$ are the Legendre polynomials.
Analogously, we define
\begin{eqnarray}
a_{nm}^\ast &=& \int_{-Y}^Y \frac{d \eta_1}{Y} \int_{-Y}^Y \frac{d \eta_2}{Y}\frac{1}{\mathcal{N}_C} C^\ast(\eta_1,\eta_2)
T_n\left(\frac{\eta_1}{Y}\right) T_m\left(\frac{\eta_2}{Y}\right), \nonumber \\ \label{eq:anmCast}
\end{eqnarray}
which focuses on the fluctuations of the strings
(note that the normalization constant ${\mathcal{N}_C}$ is evaluated with $C(\eta_1,\eta_2)$ as in Eq.~(\ref{eq:anmC})).
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{npa11v2_witholdmodelv1.pdf}\\
\end{center}
\vspace{-5mm}
\caption{The product of $a_{11}$ and $\br{N_+}$, showing the scaling discussed in the text. \label{fig:npa11}}
\end{figure}
Figure~\ref{fig:a11} shows our results for $a_{11}$ (panel a) and $a^\ast_{11}$ (panel b) obtained for Au-Au collisions
at $\sqrt{s_{\rm NN}}=200$~GeV and plotted as functions of the average
number of wounded quarks $\br{N_+}$ in selected centrality classes.
We note that the results for model cases i) and iii) are essentially identical, reflecting the feature seen already in Fig.~(\ref{fig:FBcorr}).
The result for case ii) is about a factor of 3 smaller. In this and following figures we also indicate the results for the model with single end-point fluctuations,
which is identical to case ii) in the considered acceptance region.
In view of the discussion of Sec.~\ref{sec:end}, cases i) and ii) in Fig.~\ref{fig:a11} represent the upper and lower bounds for
the admissible values of the
$a_{11}$ coefficients. This is an important result, as it provides the possible range for this quantity in approaches sharing the
features of our model.
In panel c) of Fig.~\ref{fig:a11} we present the ratio $a^\ast_{11}/a_{11}$, which shows the announced dominance of the string end-point fluctuations
over the fluctuation of the numbers of sources. In model cases i) and iii) the former account for 90\% of the effects, whereas in case ii) they account for 75-85\%.
From Eqs.~(\ref{eq:gensym},\ref{eq:Cast}) it is clear that $a^\ast_{11}$ scales as $1/\br{N_+}$.
For $a_{11}$ there is a small departure of a relative order ${\rm var}(N_-)/\br{N_+}$.
Numerically, for models i) and ii) $a^\ast_{11}\sim 0.08/\br{N_+}$, whereas the leading term of expansion (\ref{eq:delta}) yields a close
result $a^\ast_{11}\sim 0.1/\br{N_+}$.
The approximate scaling for $a_{11}$ is exhibited in Fig.~\ref{fig:npa11}.
A similar analysis of the $a_{11}$ coefficients for the d-Au collisions yields qualitatively similar
results, shown in Fig.~\ref{fig:a11dAu}.
Here, the coefficients $a_{11}^\ast$ account for more than $90\%$ of the total,
hence the dominance of string end-point fluctuations is even more pronounced in d-Au than in Au-Au collisions.
For that reason we present only the results for $a_{11}$.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{a11v2logdAu_witholdmodelv2.pdf}
\end{center}
\vspace{-5mm}
\caption{Same as in Fig.~\ref{fig:a11}a) but for d-Au collisions at $\sqrt{s_{\rm NN}}=200$~GeV, plotted as
a function of the average number of wounded quarks in Au, $\br{N_B}$ (selected values for $\br{N_B}$ correspond to
centrality classes $0-20\%$, $20-40\%$, $40-60\%$, $60-80\%$). \label{fig:a11dAu}}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{a22log_witholdmodel.pdf}\\
\includegraphics[width=0.48\textwidth]{a13log_witholdmodel.pdf}
\end{center}
\vspace{-5mm}
\caption{Same as in Fig.~\ref{fig:a11}a) but for $a_{22}$ (a) and $-a_{13}$ (b). \label{fig:anm}}
\end{figure}
In addition to $a_{11}$ coefficients, one may study the higher-order $a_{nm}$ coefficients. We
give our results for $a_{13}$ and $a_{22}$ from Au-Au collisions
in Fig.~ \ref{fig:anm}. While these coefficients are considerably suppressed as compared to $a_{11}$, shown in Fig.~\ref{fig:anm}, they exhibit the same qualitative behavior.
In particular, they scale almost exactly as $1/\br{N_+}$.
Finally, we remark that when the model results are to be compared to experimental values, one needs to relate the
space-time rapidity of the initial stage, $\eta_{PS}=\frac{1}{2} \ln[(t+z)/(t-z)]$ (until now
denoted as $\eta$ in our considerations), to the
momentum pseudorapidity of the measured hadrons, $\eta=\frac{1}{2} \log[(E+p_z)/(E-p_z)]$.
The experience of hydrodynamic simulations shows a mild longitudinal push, yielding $\eta \simeq 1.25 \,\eta_{PS}$.
This effect leads to a quenching factor of about 1.5 to be applied to the model $a_{nm}$ coefficients before comparing to the data.
\section{Conclusions \label{sec:concl}}
We have analyzed a model where strings are associated with wounded quarks and their end-points fluctuate.
We have used the data for the pseudo-rapidity spectra for d-Au and Au-Au collisions from the PHOBOS Collaboration at
$\sqrt{s_{NN}}=200$~GeV to impose constraints on the one-body distributions in the model.
We have selected a RHIC energy for our study, since the wounded quark model works very well in this case.
We first confirmed the results of~\cite{Barej:2017kcw} that a thus extracted one-body emission function reproduces reasonably well the
experimental rapidity spectra and therefore is universal in the sense that it can be applied to different centrality classes and collision systems
for the considered collision energy.
Then we showed that there remains a substantial freedom in string end-point distributions $G_{1,2}$, which gives rise to a family of possible solutions.
Specifically, we have discussed three cases of solutions: the limiting cases i) and ii) and an intermediate case iii), inspired by the valence quark
parton distribution function. We have argued that case ii) is equivalent to the model with single end-point fluctuations of~\cite{Broniowski:2015oif},
if the acceptance window at mid-rapidity is sufficiently narrow.
The analysis was carried out analytically, which has its obvious merits.
We obtained formulas for the $n$-body distributions of the produced particles.
In the study of the two-body correlations, we have examined the effects from string end-point fluctuations and from the
fluctuation of the number of sources. The former largely dominate in the corresponding Legendre coefficients $a_{nm}$.
We have found that the range for fluctuations is limited by two extreme cases. The lower limit, where the domains of the
fluctuations of both ends do not overlap, coincides (for sufficiently narrow acceptance windows in pseudorapidity)
with the model with single-end fluctuations considered earlier in~\cite{Broniowski:2015oif}. Allowing for both
ends to fluctuate increases significantly the fluctuations, raising the $a_{nm}$ coefficients by a factor of $\sim 3$.
A variant of the model where the distribution of one end of the string follows the valence quark PDF,
is very close to the case giving maximum correlation (our case i)).
Our results, in particular the presented bounds,
can serve as a baseline for future data analysis of the forward-backward fluctuations in rapidity at $\sqrt{s_{NN}}=200$~GeV.
Our simple approach, while neglecting many possible effects such as mutual influence of the strings (merging into color ropes, nuclear shadowing),
short range correlations of various origin, or
assuming strings of only one type, incorporates two basic and generic features: fluctuation of the number of strings and fluctuation of the
location of the string end-points. This makes its predictions valuable for understanding the underlying mechanisms.
It remains to be seen to what extent our analytic approach can be extended to more general models,
in particular going beyond the simple Glauber wounded picture.
\begin{acknowledgments}
Research supported by the Polish National Science Centre (NCN) Grant No. 2015/19/B/ST2/00937.
\end{acknowledgments}
|
\section{SUPPLEMENTARY MATERIAL}
In this Supplementary Material, we {provide additional analysis related to the phenomenon of non-critical slowing in photon condensates presented in the Main Text. We start in Sec.~\ref{sec1} with a
discussion on the effect of truncating our numerical simulations in the Main Text to low-energy cavity modes. We also provide evidence of the persistence of the slowing phenomena for pump powers as high as $10^3/\kappa$. In Sec.~\ref{sec2}, we show that the interesting} slowing down identified as {\it non-critical} in the Main text cannot be attributed to a long tail of the slow dynamics due to an additional phase transition resulting in {\it critical} slowing down.
In Sec.~\ref{sec3} we provide some technical details on the definition of equilibration time in our work. {We finally conclude with a short analysis on the identification of critical exponents in Sec.~\ref{sec4}.}
\section{I. Truncation of the number of modes in the cavity}
\label{sec1}
The present numerical simulations require a restriction to a small number of cavity modes which can result in truncation errors.
In general, the inclusion of more cavity modes can lead to changes in the cavity dynamics, because of the competition over the excitations in the molecular reservoirs that {\it all} cavity modes participate in.
As we will show here, {for a fixed set of system parameters, the $15$ cavity modes and a molecular environment with components up to level-2, as taken in the numerical simulations to obtain the slowing phenomenon,} captures the dynamics of all quantities of interest very well. {We note that this includes up to 3 orders (level-0, level-1 and level-2) of hierarachy in the approximation of the reduced molecular environment, as discussed in the Main Text.}
{We compare these results with those obtained for simulations that include $10$ and $21$ modes and find that the enhanced competition in higher-energy modes does not lead to any qualitative or quantitative changes in the behavior of the system.}
\begin{figure}[h]
\includegraphics[width=0.48\textwidth]{off_diag_compare_modes.png}
\caption{{Simulation showing the slowing phenomenon (top) and the population (below) of the condensed and decondensed modes when simulations include $10$, $15$ and $21$ modes and molecular environment truncated up to level-1 and level-2 hierearchies. The plots here are for the same set of system parameters as those used in Fig.~1 of the Main Text, but for larger intervals in pump power. The labels $\{m,i\}$ refer to the use of $m$ cavity modes and level-$i$ hierarchy of molecular environment in the simulations.}}
\label{10_21}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.48\textwidth]{two_modes.png}
\caption{{Equilibration time for individual cavity modes and mode populations for simulations performed using $15$ and $21$ cavity modes, with a level-2 truncated molecular environment, for pump powers as high as $10^3/\kappa$.}}
\label{15_21}
\end{figure}
{Fig.\ref{10_21}, depicts the net equilibration times together with the corresponding mode populations for simulations with $10$, $15$ and $21$ cavity modes, with the same system parameters considered in Fig.1 of the Main Text. The cut-off wavelength considered here (and in the Main Text) is $580.25$ nm and the plots are shown for both level-1 and level-2 approximations for the molecular environment.
The simulations agree very well in the prediction for all the condensation and decondensation events taking place. For $10$ modes and level-1 molecular approximation, the observed critical pump powers are slightly different compared to the other simulations. However, excellent convergence is observed for $15$ modes and more for both the critical points and the equilibration times. In particular, the slow equilibration in interval `E' is not affected at all by further inclusion of higher cavity modes. Very similar results (not shown) are also be obtained for $28$, $36$ and $45$ modes.}
{We note that the height of the peaks in the equilibration time extremely close to the critical pump powers are dependent on how finely the points are chosen, i.e., how close one is to the critical point. This proximity can change with slight numerical changes that arise from change in system parameters such as spatial resolution of molecules or number of modes. For a sufficiently fine set of pump powers, the height in equilibration peaks at critical points will converge, but this can be numerically challenging. However, this is not relevant to explaining the non-critical equilibration times and the important slowing away from criticality, where excellent convergence is already observed.}
{Fig.\ref{15_21} shows the results obtained for simulations with $15$ and $21$ cavity modes, with level-2 approximation for the molecular environment, and now for pump powers as high as $10^3/\kappa$.
As one can see, the equilibration time and slowing behavior for all individual modes agree very well. Moreover, for the same set of system parameters, the slow equilibration in interval `E' is persistent even for higher pump powers regardless of the number of modes in the simulation. Therefore, as justified by the analysis here, all the results in the Main Text were obtained using simulations with 15 cavity modes and molecular environment containing up to level-2 components.}
\section{II. Non-critical slowing and further condensation peaks}
\label{sec2}
The non-critical slowing is most dramatic for dye detunings, and consequently absorption and emission rates, deep into the decondensed regime.
In this regime however there are no {additional condensation/decondensation events taking place. Therefore, no alternative explanation is possible for the observed slowing down behavior away from the critical pump powers different to the one given in the Main Text.}
{In order to confirm our explanation of non-critical slowing down, we consider a different cutoff wavelength, $579.50$~nm, with $15$ cavity modes and keeping up to level-2 components of the molecular environment in our simulation. Fig.~\ref{3_15_v2} shows equilibration times for individual modes and mode populations.}
{In addition to the slowing behavior observed so far,} there is an additional phase transition with the condensation of mode $[0,3]$ resulting in an additional instance of critical slowing down above the decondensation threshold of mode $[0,1]$.
As one can see, all modes, including mode $[0,1]$ feature critical slowing down around this phase transition, but the plateau of slow dynamics identified as non-critical slowing down in the Main Text is clearly not the tail of this peak. {Interestingly, another region of non-critical slowing down emerges in the interval `F', which is slower than the equilibration time at the previous interval `E'. The reason for further slowing away from critical pump powers is that there are is an additional condensed mode in this region, which introduces a new competition for the molecular excitation.}
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{off_diag_add_condense.png}
\caption{{Equilibration time for individual cavity modes and mode populations as a function of pump power up to $10^3/\kappa$. The
simulations here correspond to $15$ cavity modes and truncation after the level-$2$ components of the molecular environment. The label $[x,y]$ here refers to cavity modes, where $x$ and $y$ are the low-lying energy levels of the 2D harmonic oscillator, similar to those shown in Fig.~1 of the Main Text.}}
\label{3_15_v2}
\end{figure}
\begin{figure}[]
\includegraphics[width=0.48\textwidth]{end_t_v_decay_t.png}
\caption{Comparison between exponential definition of equilibration time and equilibration time based on a fixed deviation from steady state, showing the two are related by a simple scaling factor.}
\label{end_t_v_decay_t}
\end{figure}
\section{III. Definition of equilibration time}
\label{sec3}
Equilibration time is defined in the Main Text as the time for the slowest mode to arrive at a deviation of $10^{-6}$ from its steady state value. For systems close to their steady state value, the decay towards the steady state will be exponential. In such circumstances, the exponential decay time would provide a robust definition of the equilibration time, and would be a feature of the final pump power alone, independent of the starting pump power.
Figure~\ref{end_t_v_decay_t} shows that an exponential definition of decay time gives qualitatively the same results as those presented in Fig.~1 of the Main Text, with the two definitions related by a simple scaling factor. Later in the Main Text however, we consider equilibration times after large quenches in pump power. Under these conditions, much of the dynamics from initial state to final state is transient and non-exponential. We therefore choose a consistent definition of equilibration time which captures the key features in both cases.
\section{IV. Fitting critical power laws}
\label{sec4}
Fig.~\ref{pl_fit} depicts the equilibration rate (inverse of equilibration time) as a function of the power, $P-P_{crit}$, where $P_{crit}$ is the critical power, corresponding to the four phase transitions in figure 1 of the Main Text. For critical slowing down with an exponent of -1, each phase transition should be a pair of straight lines meeting at a rate of zero for $P=P_{crit}$, as is clearly seen in figure~\ref{pl_fit}. For each critical point, because of the different power scales involved for different peaks, the values have been rescaled in $|P-P_{crit}|$ and rate to make all lines visible, and in such a way that the straight line gradient should be 1 as shown by the solid line guide to the eye in figure~\ref{pl_fit}. The coefficients of determination ($R^2$ values) for each transition, below and above threshold, are given in table~\ref{r2}.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{Eq_rate_vs_P_minus_Pcrit.png}
\caption{Demonstration of the fit of the critical slowing down peaks to a critical exponent of -1. All critical peaks are rescaled in rate and $P-P_{crit}$ to make all data visible on the same plot. The four colours of data points correspond to four critical peaks in figure 1 of the Main Text. The solid line shows $|P-P_{crit}|$ as a guide to the eye.}
\label{pl_fit}
\end{figure}
\begin{table}[h]
\begin{tabular}{ c | c | c }
~~~~Critical~~~~ & ~~~~Below~~~~ & ~~~~Above~~~~ \\
point & Threshold & Threshold \\ \hline
& & \\
A-B & -0.9998 & 0.9985 \\
B-C & -0.9996 & 0.9992 \\
C-D & -0.9996 & 0.9999 \\
D-E & -0.9995 & 0.9998 \\
\end{tabular}
\caption{Coefficients of determination for equilibration rate as a function of $|P-P_{crit}|$ just below and just above the four critical peaks in Fig.~\ref{10_21}.}
\label{r2}
\end{table}
\bibliographystyle{prsty}
|
\section{Introduction}
The most known estimate of tail probabilities for quadratic forms is the Hanson--Wright inequality regarding independent centered sub-gaussian random variables. A version of this inequality was first proved in \cite{HanWri,Wri} and recently derived by Rudelson and Vershynin in \cite{RudVer}. In this article we show estimates for quadratic forms in dependent random variables. Our main technique will be estimates of sub-exponential norms of quadratic forms in sub-gaussian random variables.
A random variable $\xi$ is called sub-gaussian if it is dominated by Gaussian random variable. For centered $\xi$ it can be expressed by requiring that
there is a constant $K$ such that $\mathbb{E}\exp(t\xi)\le \exp(K^2t^2/2)$ for all $t\in\mathbb{R}$ (see Kahane \cite{Kahane}). It means that its moment generating function
is majorized by the moment generating function of the centered Gaussian variable with the standard deviation $K$. The infimum of such $K$ is a norm on the space of centered sub-gaussian random variables. It is standardly denoted by $\tau(\cdot)$, and this space itself by $Sub(\Omega)$ on some probability space $(\Omega,\mathcal{F},\mathbb{P})$ (see Buldygin and Kozachenko \cite[Def.1.1 in Ch.1]{BulKoz}).
Non-necessarily centered sub-gaussian random variable $\xi$ can be defined by requiring that $\mathbb{E}\exp(\xi^2/K^2)\le 2$ for some $K>0$; the infimum of such $K$ is the Luxemburg norm on the Orlicz space generated by the function $\psi_2(t)=\exp(t^2)-1$. We will denote this space by $L_{\psi_2}(\Omega)$ and the Luxemburg or $\psi_2$-norm by $\|\cdot\|_{\psi_2}$. Let us note that the norms $\tau(\cdot)$ and $\|\cdot\|_{\psi_2}$ are equivalent on the space
$Sub(\Omega)=\{\xi\in L_{\psi_2}(\Omega):\;\mathbb{E}\xi=0\}$ (compare Vershynin \cite[Prop. 2.5.2]{Ver}).
From now on let $C_2$ denote a universal constant such that $\tau(\xi)\le C_2\|\xi\|_{\psi_2}$ for $\xi\in Sub(\Omega)$.
A number of other equivalent definitions of norms are used in the literature, for instance by using estimates of moments of sub-gaussian random variables, but for our purposes, the above-mentioned norms will be sufficient.
One can similarly define sub-exponential random variables, i.e. by requiring that their $\psi_1$-norm $\|\xi\|_{\psi_1}:=\inf\{K>0:\;\mathbb{E}\exp|\xi/K|\le 2\}<\infty$, where the function $\psi_1(t)=\exp(|t|)-1$. Let us note that if $\xi$ is a sub-gaussian random variable then $\xi^2$ is sub-exponential one and moreover $\|\xi^2\|_{\psi_1}=\|\xi\|_{\psi_2}^2$ (see \cite[Lem.2.7.6]{Ver} for instance).
Let us emphasize that for centered sub-exponential random variables $\xi$ there is not a global estimate of the moment generating function as in sub-gaussian case. We can only formulate the following inequality
\begin{equation}
\label{estMgf}
\mathbb{E}\exp(t\xi)\le \exp(C_1^2\|\xi\|_{\psi_1}^2t^2)
\end{equation}
for $|t|\le 1/(C_1\|\xi\|_{\psi_1})$, where $C_1$ is an universal constant \footnote{The universal constants $C_1$, $C_2$ and, in consequence, $C_3$, $C_4$ will be the same in each occurrence and universal constants $c$ may be different.}; compare Vershynin \cite[Prop. 2.7.1]{Ver}.
Let us recall also the Hanson--Wright inequality (see Rudelson and Vershynin \cite[Th.1.1]{RudVer}).
\begin{thm}
({\rm Hanson--Wright inequality})
Let $\xi=(\xi_1,...,\xi_n)\in \mathbb{R}^n$ be a random vector with independent coordinates $\xi_i$ which satisfy $\mathbb{E}\xi_i=0$ and $\|\xi_i\|_{\psi_2}\le K$.
Let $A=[a_{ij}]_{i,j=1}^n$ be an $n\times n$ matrix. Then, for every $t\ge 0$,
$$
\mathbb{P}\Big(\big|\xi^TA\xi-\mathbb{E}(\xi^TA\xi)\big|\ge t\Big)\le
2\exp\Big(-c\min\Big\{\frac{t^2}{K^4\|A\|^2_{HS}}, \frac{t}{K^2\|A\|} \Big\}\Big),
$$
where $c$ is a universal constant, $\|A\|_{HS}=(\sum_{i,j=1}^na_{ij}^2)^{1/2}$ is the Hilbert-Schmidt norm of $A$, whereas $\|A\|=\sup_{|x|_2\le 1}|Ax|_2$ is the operator norm ($|\cdot|_2$ denotes the standard Euclidean norm in $\mathbb{R}^n$).
\end{thm}
We will need a notion of sub-gaussian random vectors. One of the ways to define such vectors is by requiring that one dimensional marginals
$\left\langle \xi,{\bf t}\right\rangle$ are sub-gaussian for all ${\bf t}\in\mathbb{R}^n$. The sub-gaussian norm of $\xi$ is defined as
$\|\xi\|_{\psi_2}=\sup_{{\bf t}\in S^{n-1}}\|\left\langle \xi,{\bf t}\right\rangle\|_{\psi_2}$ (see \cite[Def. 5.22]{Ver}. Using the Luxemburg norm we can rewrite it as follows
$$
\|\xi\|_{\psi_2}=\inf\Big\{K>0:\;\sup_{{\bf t}\in S^{n-1}}\;\mathbb{E}\exp\big(\left\langle \xi,{\bf t}\right\rangle^2/K^2\big)\le 2\Big\}.
$$
The space of $n$-dimensional sub-gaussian random vectors we will denote by $L_{\psi_2}^n(\Omega)$.
Similarly as in one-dimensional case we can also introduce the definition of a norm for centered sub-gaussian random vectors $\xi$ in the form
$$
\tau(\xi)=\inf\Big\{K>0:\;\forall_{{\bf t}\in\mathbb{R}^n}\;\mathbb{E}\exp\left\langle \xi,{\bf t}\right\rangle\le \exp\big(K^2|{\bf t}|_2^2/2\big)\Big\};
$$
as in the case of the norm $\|\cdot\|_{\psi_2}$ we keep the notation $\tau(\cdot)$ in the multidimensional case.
The space of centered $n$-dimensional sub-gaussian random vectors we will denote by
$Sub^n(\Omega)=\{\xi\in L_{\psi_2}^n(\Omega):\;\mathbb{E}\xi={\bf 0}\in \mathbb{R}^n\}$.
\begin{exa}
Recall that the moment generating function of $g=(g_i)_{i=1}^n\sim\mathcal{N}({\bf 0},I_n)$ ($g_i$ are independent standard normally distributed) equals
$\mathbb{E}\exp\left\langle g,{\bf t}\right\rangle
=\exp(|{\bf t}|_2^2/2)$. It means that $\tau(g)=1$. Let $A$ be $n\times n$ matrix. The random vector $Ag$ has the centered Gaussian distribution with the covariance matrix $Cov(Ag)=AA^T$. The moment generating function of $Ag$ can be estimated as follows
$$
\mathbb{E}\exp\left\langle Ag,{\bf t}\right\rangle=\exp\Big(|A^T{\bf t}|_2^2/2\Big)\le \exp\Big(\|A\|^2|{\bf t}|_2^2/2\Big).
$$
It means that in this case we have $\tau(Ag)\le\|A\|$.
\end{exa}
\begin{rem}
For ${\bf t} \in \mathbb{R}^n$ and a sub-gaussian random vector $\xi\in Sub^n(\Omega)$ the random variable
$\left\langle \xi,{\bf t}\right\rangle\in Sub(\Omega)$. Thu
$$
\mathbb{E}\exp(\left\langle \xi,{\bf t}\right\rangle)\le \exp\big(\tau(\left\langle \xi,{\bf t}\right\rangle)^2/2\big).
$$
By equivalence of norms $\tau(\cdot)$ and $\|\cdot\|_{\psi_2}$ we get
\begin{eqnarray*}
\exp\big(\tau(\left\langle \xi,{\bf t}\right\rangle)^2/2\big)&\le & \exp(C_2^2\|\left\langle \xi,{\bf t}\right\rangle\|_{\psi_2}^2/2\Big)=
\exp\Big(C_2^2\|\left\langle \xi,{\bf t}/|{\bf t}|_2\right\rangle\|_{\psi_2}^2|{\bf t}|_2^2/2\Big)\\
\; & \le & \exp\Big(C_2^2\|\xi\|_{\psi_2}^2|{\bf t}|_2^2/2\Big),\\
\end{eqnarray*}
where $C_2$ is the same as in one-dimensional case. The obtained estimate
$\mathbb{E}\exp(\left\langle \xi,{\bf t}\right\rangle)\le \exp\Big(C_2^2\|\xi\|_{\psi_2}^2|{\bf t}|_2^2/2\Big)$ means that
$\tau(\xi)\le C_2\|\xi\|_{\psi_2}$ also in the multi-dimensional case.
\end{rem}
\section{Results}
To prove our main results we will need some technical lemma.
\begin{lem}
\label{lem2}
Let $\eta$ be a centered random variable. If there exist positive constants $a,b$ such that
$\mathbb{E}\exp(t\eta)\le \exp(a^2t^2/2)$ for $|t|\le b$
then for every $s\ge 0$ we have
$$
\mathbb{P}\big(|\eta|\ge s\big)\le 2e^{-g(s)},
$$
where
$$
g(s)=\left\{
\begin{array}{ccl}
\frac{s^2}{2a^2} & {\rm if} & 0\le s \leq a^2b,\\
bs-\frac{a^2b^2}{2} & {\rm if} & a^2b<s.
\end{array}
\right.
$$
\end{lem}
\begin{proof}
Define a function
$$
\varphi_{a,b}(t)=\left\{
\begin{array}{ccl}
\frac{a^2t^2}{2} & {\rm if} & |t|\le b,\\
\infty & {\rm otherwise} & .
\end{array}
\right.
$$
Observe that $\ln\mathbb{E}\exp(t\eta)\le \varphi_{a,b}(t)$ for every $t\in\mathbb{R}$. One can check that the convex conjugate of $\varphi_{a,b}$, i.e.
$\varphi_{a,b}^\ast(s)=\sup_{t\in\mathbb{R}}\{ts-\varphi_{a,b}(t)\}$, equals
$$
\varphi_{a,b}^\ast(s)=\left\{
\begin{array}{ccl}
\frac{s^2}{2a^2} & {\rm if} & |s|\le a^2b,\\
b s-\frac{a^2b^2}{2} & {\rm if} & |s| > a^2b.
\end{array}
\right.
$$
By exponential Markov's inequality and the estimate on the cumulant generating function $\ln\mathbb{E}e^{t\eta}$, the inequality
$$
\mathbb{P}\big(\eta\ge s\big)\le e^{-st}e^{\ln\mathbb{E}e^{t\eta}}\le e^{-\{st-\varphi_{a,b}(t)\}},
$$
holds for $s,t>0$, giving
$$
\mathbb{P}\big(\eta\ge s\big)\le \inf_{t>0} e^{-\{st-\varphi_{a,b}(t)\}}=e^{-\sup_{t>0}\{st-\varphi_{a,b}(t)\}}=e^{-\varphi_{a,b}^\ast(s)},
$$
since $\sup_{t>0}\{st-\varphi_{a,b}(t)\}=\varphi_{a,b}^\ast(s)$ for the even function $\varphi_{a,b}$. The inequality
$\mathbb{P}(\eta\le -s)\le \exp(-\varphi_{a,b}^\ast(s))$ is proved similarly. Combining ones and taking $g=\varphi_{a,b}^\ast$ we get the proof.
\end{proof}
\begin{rem}
\label{remlem2}
Let us observe that $g(t)\ge \min\{t^2/(2a^2),bt/2\}$ and we may rewrite the claim of the above lemma in a weaker but more traditional, for the Bernstein-type inequality, form as follows
$$
\mathbb{P}\big(|\eta|\ge t\big)\le 2\exp\Big(-\min\Big\{\frac{t^2}{2a^2},\frac{bt}{2}\Big\}\Big).
$$
\end{rem}
\begin{rem}
\label{rem2lem2}
By virtue of (\ref{estMgf}) we know that centered sub-exponential random variables satisfy the assumption of the above lemma with $a=\sqrt{2}C_1\|\eta\|_{\psi_1}$ and $b=1/(C_1\|\eta\|_{\psi_1})$.
So, for such variables, we can rewrite the above estimate in the following way
$$
\mathbb{P}\big(|\eta|\ge t\big)\le 2\exp\Big(-\min\Big\{\frac{t^2}{4C_1^2\|\eta\|_{\psi_1}^2},\frac{t}{2C_1\|\eta\|_{\psi_1}}\Big\}\Big).
$$
Let us emphasize that one of the ways to obtain Bernstein-type inequalities for sub-exponential random variables is to find (estimate) of their $\psi_1$-norms.
\end{rem}
We show that, for $\xi\in L^n_{\psi_2}(\Omega)$ and $n\times n$ matrix $A$, a random variable $\xi^TA\xi-\mathbb{E}(\xi^TA\xi)$ is centered and sub-exponential.
We estimate its $\psi_1$-norm and, in this way, we get Bernstein-type inequality.
We will use the standard inner product notation to write quadratic form, i.e. $\xi^TA\xi=\left\langle A\xi, \xi \right\rangle$. To obtain the Bernstein-type estimations, it is enough to estimate the $\psi_1$-norms of centered quadratic forms $\left\langle A\xi, \xi \right\rangle$.
We will also need some estimate for the $\psi_1$-norm of $\mathbb{E}\left\langle A\xi, \xi \right\rangle$ by the $\psi_1$-norm of the quadratic forms $\left\langle A\xi, \xi \right\rangle$.
By the definition of the Luxemburg norm
and the Jensen inequality applied to a convex function $\exp\{|\cdot|/a\}$ ($a>0$) we get
\begin{equation*}
\label{estE}
2\ge \mathbb{E}\exp\Big(\frac{|\left\langle A\xi, \xi \right\rangle|}{\|\left\langle A\xi, \xi \right\rangle\|_{\psi_1}}\Big)\ge \exp\Big(\frac{|\mathbb{E}\left\langle A\xi, \xi \right\rangle|}{\|\left\langle A\xi, \xi \right\rangle\|_{\psi_1}}\Big)=
\mathbb{E}\exp\Big(\frac{|\mathbb{E}\left\langle A\xi, \xi \right\rangle|}{\|\left\langle A\xi, \xi \right\rangle\|_{\psi_1}}\Big),
\end{equation*}
which means that $\|\mathbb{E}\left\langle A\xi, \xi \right\rangle\|_{\psi_1}\le\|\left\langle A\xi, \xi \right\rangle\|_{\psi_1}$.
It follows that
\begin{equation}
\label{estcent}
\|\left\langle A\xi, \xi \right\rangle-\mathbb{E}\left\langle A\xi, \xi \right\rangle\|_{\psi_1}\le 2\|\left\langle A\xi, \xi \right\rangle\|_{\psi_1}.
\end{equation}
Now we formulate and prove our main results.
\begin{pro}
\label{lemtr}
Let $\xi\in L^n_{\psi_2}(\Omega)$ and $A$ be $n\times n$ matrix. Then $\left\langle A\xi,\xi\right\rangle\in L_{\psi_1}(\Omega)$ and
$$
\|\left\langle A\xi,\xi\right\rangle\|_{\psi_1}\le \|A\|_{tr}\|\xi\|_{\psi_2}^2,
$$
where $\|A\|_{tr}=trace(AA^T)^{1/2}$ is the trace norm of $A$.
\end{pro}
\begin{proof}
Since $\left\langle A\xi,\xi\right\rangle=\left\langle 1/2(A+A^T)\xi,\xi\right\rangle$, we may assume that $A$ is symmetric matrix. Moreover, because we can present it as the difference of two symmetric and nonnegative definite matrices $A_1$ and $A_2$ such that $\|A\|_{tr}=\|A_1-A_2\|_{tr}=\|A_1\|_{tr}+\|A_2\|_{tr}$, then
$$
\|\left\langle A\xi,\xi\right\rangle\|_{\psi_1}\le \|A_1\|_{tr}\|\xi\|_{\psi_2}^2+\|A_2\|_{tr}\|\xi\|_{\psi_2}^2=\|A\|_{tr}\|\xi\|_{\psi_2}^2
$$
and, without loss of generality, we may also assume that this matrix is nonnegative definite.
Let $US U^T$ be the singular-value decomposition (eigendecomposition) of symmetric and nonnegative definite matrix $A$,
where $U$ is a unitary matrix and $S=diag(s_1,...,s_n)$ is a diagonal matrix with singular values (eigenvalues) $s_i$ of $A$ on the diagonal. This allows us to describe
$\left\langle A\xi,\xi\right\rangle$ as $|S^{1/2}U^T\xi|^2_2=\sum_{i=1}^ns_i\left\langle \xi,Ue_i\right\rangle^2$. Thus
\begin{equation*}
\mathbb{E}\exp\Big(\frac{\left\langle A\xi,\xi\right\rangle}{\|A\|_{tr}\|\xi\|_{\psi_2}^2}\Big)=
\mathbb{E}\exp\Big(\frac{\sum_{i=1}^ns_i\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{tr}\|\xi\|_{\psi_2}^2}\Big)=
\mathbb{E}\Big(\prod_{i=1}^n\exp\frac{ s_i\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{tr}\|\xi\|_{\psi_2}^2}\Big).
\end{equation*}
Using the multi-factorial H\"older inequality with exponents $p_i=\sum_{k=1}^ns_k/s_i$, $1\le i\le n$, we get
\begin{eqnarray*}
\mathbb{E}\Big(\prod_{i=1}^n\exp\frac{ s_i\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{tr}\|\xi\|_{\psi_2}^2}\Big)&\le&
\prod_{i=1}^n\Big(\mathbb{E}\exp\Big(\frac{(\sum_{k=1}^ns_k)\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{tr}\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s_i}{\sum_{k=1}^ns_k}}\\
\; &=& \prod_{i=1}^n\Big(\mathbb{E}\exp\Big(\frac{\left\langle \xi,Ue_i\right\rangle^2}{\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s_i}{\sum_{k=1}^ns_k}},
\end{eqnarray*}
since $\|A\|_{tr}=\sum_{k=1}^ns_k$. Because $|Ue_i|_2=1$ then we can estimate the right hand side as follows
$$
\prod_{i=1}^n\Big(\mathbb{E}\exp\Big(\frac{\left\langle \xi,Ue_i\right\rangle^2}{\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s_i}{\sum_{k=1}^ns_k}}\le
\prod_{i=1}^n\Big(\sup_{|{\bf t}|_2=1}\mathbb{E}\exp\Big(\frac{\left\langle \xi,{\bf t}\right\rangle^2}{\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s_i}{\sum_{k=1}^ns_k}}\le 2,
$$
since each factor is less than or equal to $2^{s_i/\sum_{k=1}^ns_k}$, $i=1,...,n$.
In summary, we get
$$
\mathbb{E}\exp\Big(\frac{\left\langle A\xi,\xi\right\rangle}{\|A\|_{tr}\|\xi\|_{\psi_2}^2}\Big)\le 2,
$$
which means that $\|\left\langle A\xi,\xi\right\rangle\|_{\psi_1}\le \|A\|_{tr}\|\xi\|_{\psi_2}^2$.
\end{proof}
Immediately by condition (\ref{estcent}),
Proposition \ref{lemtr} and Remark \ref{rem2lem2} we get our first tail estimate for quadratic forms in sub-gaussian dependent random variables.
\begin{cor}
Let $\xi\in L^n_{\psi_2}(\Omega)$ and $\|\xi\|_{\psi_2}\le K$. Then, for $n\times n$ matrix $A$, we have
$$
\mathbb{P}\Big(\big|\left\langle A\xi,\xi\right\rangle-\mathbb{E}\left\langle A\xi,\xi\right\rangle\big|\ge t\Big)\le
2\exp\Big(-\min\Big\{\frac{t^2}{16C_1^2\|A\|_{tr}^2K^4},\frac{t}{4C_1\|A\|_{tr}K^2}\Big\}\Big).
$$
\end{cor}
If we additionally assume that $\mathbb{E}\xi={\bf 0}$, then we can get a better estimate with the Hilbert-Schmidt norm instead of the trace norm of $A$.
\begin{pro}
\label{lemHS}
Let $\xi\in Sub^n(\Omega)$ and $A$ be $n\times n$ symmetric and nonnegative definite matrix. Then
$$
\|\left\langle A\xi,\xi\right\rangle\|_{\psi_1}\le 2C_2\|A\|_{HS}\|\xi\|_{\psi_2}^2.
$$
\end{pro}
\begin{proof}
We start with some form of the decoupling argument.
Consider two independent Bernoulli random variables $\delta_k\in\{0,1\}$ with $\mathbb{E}\delta_k=1/2$, $k=1,2$, which are also independent from $\xi$. Define new random vectors
$\delta_k\xi=(\delta_k\xi_i)_{i=1}^n$.
Note that these vectors are centered, i.e. $\mathbb{E}_{\xi,\delta_k}(\delta_k\xi)={\bf 0}\in\mathbb{R}^n$, where $\mathbb{E}_{\xi,\delta_k}$ denote expectation with respect to both $\xi$ and $\delta_k$. Notice that
\begin{equation}
\label{est1}
\mathbb{E}_{\xi,\delta_k}e^{\left\langle {\bf t},\delta_k\xi\right\rangle}=
\mathbb{E}_{\xi,\delta_k}e^{\delta_k\left\langle {\bf t},\xi\right\rangle}=\mathbb{E}\Big(e^{\left\langle {\bf t},\xi\right\rangle}/2+1/2\Big)
\le e^{\tau(\xi)^2|{\bf t}|_2^2/2}.
\end{equation}
It means that, for $k=1,2$, $\tau(\delta_k\xi)\le \tau(\xi)$.
By $\mathbb{E}_{\delta}$ we will denote expectation with respect to $\delta=(\delta_1,\delta_2)$ and by $\mathbb{E}_{\xi,\delta}$ with respect to both $\xi$ and $\delta$. Since $\mathbb{E}_\delta(\delta_1\delta_2)=1/4$, we have
$
\left\langle A\xi,\xi\right\rangle=4\mathbb{E}_{\delta}\left\langle A(\delta_1\xi),\delta_2\xi\right\rangle.
$
Jensen's inequality yields
$$
\mathbb{E}e^{\left\langle A\xi,\xi\right\rangle}\le \mathbb{E}_{\xi,\delta}e^{4\left\langle A(\delta_1\xi),\delta_2\xi\right\rangle}.
$$
Conditioning with respect $\delta_1\xi$ and using the definition of the norm $\tau(\cdot)$ we get
\begin{eqnarray*}
\mathbb{E}_{\xi,\delta}e^{4\left\langle A(\delta_1\xi),\delta_2\xi\right\rangle} &=&
\mathbb{E}_{\xi,\delta_1}\Big[\mathbb{E}_{\xi,\delta_2}\Big(e^{4\left\langle A(\delta_1\xi),\delta_2\xi\right\rangle}\Big|\delta_1\xi\Big)\Big]\\
\; &\le&
\mathbb{E}_{\xi,\delta_1}e^{2 |A(\delta_1\xi)|_2^2\tau(\delta_2\xi)^2}\le \mathbb{E}_{\xi,\delta_1}e^{2 |A(\delta_1\xi)|_2^2\tau(\xi)^2}.
\end{eqnarray*}
The second inequality follows from $\tau(\delta_2\xi)\le\tau(\xi)$. Similarly as in (\ref{est1}) one can show that the above right hand side is less than or equal to $\mathbb{E}\exp(2 |A\xi|_2^2\tau(\xi)^2)$.
Summing up we get the following
$$
\mathbb{E}\exp\Big(\frac{\left\langle A\xi,\xi\right\rangle}{2\|A\|_{HS}\|\xi\|_{\psi_2}\tau(\xi)}\Big) \le
\mathbb{E}\exp\Big(\frac{2 |A\xi|_2^2\tau(\xi)^2}{2\|A\|_{HS}^2\|\xi\|_{\psi_2}^2\tau(\xi)^2}\Big)=
\mathbb{E}\exp\Big(\frac{ |A\xi|_2^2}{\|A\|_{HS}^2\|\xi\|_{\psi_2}^2}\Big).
$$
Let $USU^T$ be again the eigendecomposition
of $A$, i.e. $U$ is some unitary matrix
and $S=diag(s_1,...,s_n)$ is a diagonal matrix with eigenvalues $s_i$ of $A$ on the diagonal. Since $U$ is the unitary matrix, we have $|A\xi|_2^2=|S U^T\xi|_2^2$. It can be rewritten as
$$
|SU^T\xi|_2^2=
\sum_{i=1}^n\left\langle SU^T\xi,e_i\right\rangle^2=\sum_{i=1}^ns_i^2\left\langle \xi,Ue_i\right\rangle^2.
$$
Thus
\begin{equation*}
\mathbb{E}\exp\Big(\frac{\left\langle A\xi,\xi\right\rangle}{2\|A\|_{HS}\|\xi\|_{\psi_2}\|\xi\|_{S_{(2)}}}\Big)\le
\mathbb{E}\exp\Big(\frac{\sum_{i=1}^ns^2_i\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{HS}^2\|\xi\|_{\psi_2}^2}\Big)=
\mathbb{E}\Big(\prod_{i=1}^n\exp\frac{ s_i^2\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{HS}^2\|\xi\|_{\psi_2}^2}\Big).
\end{equation*}
Recall that the Hilbert--Schmidt norm of $A$ equals $(\sum_{i=1}^n s_i^2)^{1/2}$.
Using the multi-factorial H\"older inequality with exponents $p_i=\sum_{k=1}^ns_k^2/s_i^2$, $1\le i\le n$, we get
\begin{eqnarray*}
\mathbb{E}\Big(\prod_{i=1}^n\exp\frac{ s^2_i\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{HS}^2\|\xi\|_{\psi_2}^2}\Big)&\le&
\prod_{i=1}^n\Big(\mathbb{E}\exp\Big(\frac{(\sum_{k=1}^ns^2_k)\left\langle \xi,Ue_i\right\rangle^2}{\|A\|_{HS}^2\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s^2_i}{\sum_{k=1}^ns^2_k}}\\
\; &=& \prod_{i=1}^n\Big(\mathbb{E}\exp\Big(\frac{\left\langle \xi,Ue_i\right\rangle^2}{\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s^2_i}{\sum_{k=1}^ns^2_k}}.
\end{eqnarray*}
Because $|Ue_i|_2=1$ then we can estimate the right hand side as follows
$$
\prod_{i=1}^n\Big(\mathbb{E}\exp\Big(\frac{\left\langle \xi,Ue_i\right\rangle^2}{\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s^2_i}{\sum_{k=1}^ns^2_k}}\le
\prod_{i=1}^n\Big(\sup_{|{\bf t}|_2=1}\mathbb{E}\exp\Big(\frac{\left\langle \xi,{\bf t}\right\rangle^2}{\|\xi\|_{\psi_2}^2}\Big)\Big)^{\frac{s^2_i}{\sum_{k=1}^ns^2_k}}\le 2,
$$
since each factor is less than or equal to $2^{s^2_i/\sum_{k=1}^ns^2_k}$, $i=1,...,n$.
Summarizing we get
$$
\mathbb{E}\exp\Big(\frac{\left\langle A\xi,\xi\right\rangle}{2\|A\|_{HS}\|\xi\|_{\psi_2}\tau(\xi)}\Big)\le 2,
$$
which means that $\|\left\langle A\xi,\xi\right\rangle\|_{\psi_1}\le 2\|A\|_{HS}\|\xi\|_{\psi_2}\tau(\xi)\le 2C_2\|A\|_{HS}\|\xi\|_{\psi_2}^2$.
\end{proof}
\begin{rem}
\label{estHS}
Considering quadratic forms we can always assume that their generating matrices are symmetric. But now for the Hilbert--Schmidt norm of $A=A_1-A_2$ ($A_1$ and $A_2$ are symmetric and nonnegative definite) we only have $\|A_1\|_{HS}+\|A_2\|_{HS}\le\sqrt{2}\|A_1-A_2\|_{HS}=\sqrt{2}\|A\|_{HS}$. Thus for arbitrary $A$ we get
$$
\|\left\langle A\xi,\xi\right\rangle\|_{\psi_1}\le 2\sqrt{2}\|A\|_{HS}\|\xi\|_{\psi_2}\tau(\xi)\le 2\sqrt{2}C_2\|A\|_{HS}\|\xi\|_{\psi_2}^2.
$$
\end{rem}
By virtue of Remark \ref{rem2lem2},
Proposition \ref{lemHS} and Remark \ref{estHS} we can formulate the following
\begin{cor}
\label{twHS}
Let $\xi\in Sub^n(\Omega)$ and $\|\xi\|_{\psi_2}\le K$. Then, for $n\times n$ matrix $A$, we have
$$
\mathbb{P}\Big(\big|\left\langle A\xi,\xi\right\rangle-\mathbb{E}\left\langle A\xi,\xi\right\rangle\big|\ge t\Big)\le
2\exp\Big(-\min\Big\{\frac{t^2}{C_3^2\|A\|_{HS}^2K^4},\frac{t}{C_3\|A\|_{HS}K^2}\Big\}\Big),
$$
where $C_3=2\sqrt{2}C_1C_2$
\end{cor}
\begin{rem}
The above estimate has the form similar to the Hanson--Wright inequality but with one important difference. In the second term of the minimum in the Hanson--Wright inequality there is the operator norm of $A$. It is an important (better) component, especially in the context of the long-time behavior of investigated estimates.
\end{rem}
\begin{rem}
These type inequalities can be obtained under more general assumption that a random vector $\xi$ satisfy the convex concentration property (see \cite{VuW,Ada} for instance). Adamczak in \cite{Ada} get the form of the Hanson-Wright inequality under assumption that
a random vector $\xi$ satisfy the convex concentration property with constant $K$ (see \cite[Th.2.5]{Ada}).
As the author notes this constant depends on the dimension $n$, even for vectors with independent sub-gaussian coordinates (see \cite[Rem. 2.9]{Ada}).
\end{rem}
Let us emphasize that our results have been obtained without the additionally assumption of the convex concentration of $\xi$. Moreover our $K$
is simply the norm of $\xi$ and not a slightly enigmatic constant for which the convex concentration property holds.
In the paper Hsu et al. \cite{HKZ} one can find some application of the our norm $\tau$ (without using this notion and notation) to the investigation of the quadratic form
of the form
$|A\xi|_2^2$ (see \cite[Th.2.1]{HKZ}). The authors consider only estimation of probability for one side tails of these forms which are sufficient for their applications.
For the sake of completeness, we will present in the next remark a derivation of the Hanson--Wright inequality for independent normally distributed random variables.
\begin{rem}
Decoupling argument applied to $\xi=(\xi_i)\in Sub^n(\Omega)$ with independent coordinates $\xi_i$ implies that we can investigate
$\left\langle Ag,g\right\rangle$ ($g\sim\mathcal{N}({\bf 0},I_n)$) instead of $\left\langle A\xi,\xi\right\rangle$ (up to a product of some universal constant $c$ and $\psi_2$-norm of $\xi$).
Recall that if $g\sim\mathcal{N}(0,1)$ then $g^2$ has $\chi^2_1$-distribution with one degree of freedom, whose
moment generating function is $\mathbb{E}\exp(tg^2)=(1-2t)^{-1/2}$ for $t<1/2$.
Therefore
$$
\mathbb{E}\exp\Big(\frac{g^2}{K^2}\Big)=\Big(1-\frac{2}{K^2}\Big)^{-1/2},
$$
which is less than or equal to $2$ if $K\ge \sqrt{8/3}$.
It implies that $\|g\|_{\psi_2}=\sqrt{8/3}$.
Because for any ${\bf t}\in S^{n-1}$ and $g\sim\mathcal{N}({\bf 0},I_n)$ the inner product $\left\langle {\bf t},g\right\rangle\sim \mathcal{N}(0,1)$ then by the above and the definition of $\psi_2$-norm of random vectors we also have that $\|g\|_{\psi_2}=\sqrt{8/3}$ for $g\sim\mathcal{N}({\bf 0},I_n)$.
Let $A$ be $n\times n$ symmetric nonnegative definite matrix and
$USU^T$ be its singular value decomposition; i.e. $A=USU^T$, $S=diag(s_1,...,s_n)$ and $U^{-1}=U^T$. Then for $g\sim\mathcal{N}({\bf 0},I_n)$ we get
$$
\left\langle Ag,g\right\rangle-\mathbb{E}\left\langle Ag,g\right\rangl
=\sum_{i=1}^ns_i\Big(\left\langle g,Ue_i\right\rangle^2-\mathbb{E}\left\langle g,Ue_i\right\rangle^2\Big)
$$
and
$$
\mathbb{E}\exp\Big(t(\left\langle Ag,g\right\rangle-\mathbb{E}\left\langle Ag,g\right\rangle)\Big) =
\mathbb{E}\exp\Big(t\sum_{i=1}^ns_i\Big(\left\langle g,Ue_i\right\rangle^2-\mathbb{E}\left\langle g,Ue_i\right\rangle^2\Big)\Big).
$$
Rotational invariance of Gaussian distribution implies that $\left\langle g,Ue_i\right\rangle$, $i=1,...,n$, are independent random variables. In consequence,
\begin{equation}
\label{prod}
\mathbb{E}\exp\Big(t(\left\langle Ag,g\right\rangle-\mathbb{E}\left\langle Ag,g\right\rangle)\Big)=
\prod_{i=1}^n\mathbb{E}\exp\Big(ts_i\Big(\left\langle g,Ue_i\right\rangle^2-\mathbb{E}\left\langle g,Ue_i\right\rangle^2\Big)\Big).
\end{equation}
Moreover, the $\psi_2$-norm of each $\left\langle g,Ue_i\right\rangle$, $i=1,...,n$, is equal to the $\psi_2$-norm of whole vector $g$, i.e.
$$
\|\left\langle g,Ue_i\right\rangle\|_{\psi_2}=\|g\|_{\psi_2}=\sqrt{8/3}.
$$
and
$$
\|\left\langle g,Ue_i\right\rangle^2\|_{\psi_1}=\|\left\langle g,Ue_i\right\rangle\|_{\psi_2}^2= \|g\|_{\psi_2}^2=8/3.
$$
By the above and (\ref{estMgf}), for each factor of (\ref{prod}), we get
$$
\mathbb{E}\exp\Big(ts_i\Big(\left\langle g,Ue_i\right\rangle^2-\mathbb{E}\left\langle g,Ue_i\right\rangle^2\Big)\Big)
\le \exp(256C_1^2s_i^2t^2/9),
$$
if $|t|\le 3/(16C_1s_i)$. Substituting of the above estimate into (\ref{prod}), taking into account that $\sum_{i=1}^ns_i^2=\|A\|_{HS}$ and
$\max_is_i=\|A\|$, we obtain the following
$$
\mathbb{E}\exp\Big(t(\left\langle Ag,g\right\rangle-\mathbb{E}\left\langle Ag,g\right\rangle)\Big)\le \exp\Big(256C_1^2\|A\|_{HS}^2t^2/9\Big)
$$
for $|t|\le 3/(16C_1\|A\|)$.
We have proved that for $\eta=\left\langle Ag,g\right\rangle-\mathbb{E}\left\langle Ag,g\right\rangle$ we can take in Lemma \ref{lem2}
$a=16\sqrt{2}C_1\|A\|_{HS}/3$ and $b=3/(16C_1\|A\|)$.
By Remark \ref{remlem2} we obtain the form of Hanson--Wright's inequality for Gaussian random vector $g\sim\mathcal{N}({\bf 0},I_n)$:
$$
\mathbb{P}\Big(\big|\left\langle Ag,g\right\rangle-\mathbb{E}\left\langle Ag,g\right\rangle\big|\ge t\Big)\le
2\exp\Big(-\min\Big\{\frac{9t^2}{512C_1^2\|A\|_{HS}^2},\frac{3t}{32C_1\|A\|}\Big\}\Big).
$$
\end{rem}
We show an application of Corollary \ref{twHS} to fixed design linear regression (compare Hsu et al. \cite{HKZ}).
\begin{exa}
Let ${\bf x}_1,...,{\bf x}_n\in \mathbb{R}^d$ and $X=({\bf x}_1,...,{\bf x}_n)$ denote the $d\times n$ design matrix. Assume that
$\Sigma=n^{-1}\sum_{i=1}^n{\bf x}_i{\bf x}_i^T$ is invertible.
Let $\xi=(\xi_i)_{i=1}^n\in L_{\psi_2}^n$. Define the coefficient vector of the least expected squared error (given the observation $\xi$):
$$
\beta:=n^{-1}\sum_{i=1}^n\mathbb{E}\xi_i\Sigma^{-1}{\bf x}_i,
$$
and its ordinary least squares estimator:
$$
\hat{\beta}(\xi):=n^{-1}\sum_{i=1}^n\xi_i\Sigma^{-1}{\bf x}_i.
$$
The quality of the estimator $\hat{\beta}$ can be judged by the excess loss
$$
R(\xi)=|\Sigma^{1/2}(\hat{\beta}(\xi)-\beta)|_2^2=\left\langle A(\xi-\mathbb{E}\xi),\xi-\mathbb{E}\xi\right\rangle,
$$
where $A=n^{-2}X^T\Sigma^{-1}X$ as can be shown by algebraic calculations.
Suppose $\|\xi\|_{\psi_2}\le K$ (recall that $\|\xi-\mathbb{E}\xi\|_{\psi_2}\le 2\|\xi\|_{\psi_2}$). By an equivalent form of Corollary \ref{twHS} we get the following tail estimate
$$
\mathbb{P}\Big(\big|R(\xi)-\mathbb{E}R(\xi)\big|\ge \|A\|_{HS}K^2\max\big\{\sqrt{t},t\big\}\Big)\le 2\exp(-t/C_4),
$$
where $C_4=2C_3=4\sqrt{2}C_1C_2$.
\end{exa}
\begin{rem}
In the paper \cite{Zaj} one can find another estimate of quadratics forms (even chaoses of higher order) by using other norms of random vectors and, in consequence, new forms of their tail estimates (see \cite[Rem. 3.6]{Zaj}).
\end{rem}
|
\section{Introduction}
Text simplification is an operation in natural language processing to modify an existing corpus of human-readable text so that the grammar and structure of the sentence is greatly simplified, while the underlying meaning and information remains the same. Due to the fact that natural human languages ordinarily contain complex compound constructions~\cite{luan2017scienceie}, text simplification can make texts easier for human readers as well as automatic text processing. Although there are many previous works regarding text simplification e.g.(\cite{callison}, \cite{fung}), which benifit from data-driven machine translation, paraphrasing or grounded language acquisition techniques, works are still limited because available monolingual parallel corpora are limited or automatically generated are noisy.
Wikipedia is potentially a good resource for text simplification, since it includes standard articles and their corresponding simple articles in English. A challenge with automatic alignment is that standard and simple articles can be written independently so they are not strictly parallel, and have very different presentation ordering. A few studies use editor comments attached to Wikipedia edit logs to extract pairs of simple and difficult words \cite{yatskar}. Works such as \cite{zhu}, use text-based similarity techniques to extract pairs of simple and standard sentences. \cite{sim} use a greedy search over the simple and standard Wikipedia documents to align sentence pairs. By taking advantage of a word-level semantic similarity measure built on top of Wiktionary and WordNet, \cite{sim} do not make assumption about the relative order of sentences in standard versus simple Wikipedia articles and obtain the highest performance. The resulting datasets of manually and automatically aligned sentence pairs are made available. This paper uses the automatic aligned sentence pairs of \cite{sim} to do semi-supervise training and manually aligned data to do adaptation and performance evaluation.
Based on this setting of using Wikipedia on text simplification, the goal of this project is to find sentence level alignment. The problem is solved by first matching a sentence to its correponding simple version. In order to capture rich contextual and semantic information in a sentence, we use a Convolutional Neural Network (CNN) structure to learn low-dimensional, semantic vector representation for sentence. CNN is proven successful on image\cite{image} and speech \cite{speech}. In previous studies, convolutional notion of similarity is efficient in capturing contenxtual information and has been vastly applied to natural language sentence matching. In \cite{clsm}, a Convolutional Latent Semantic Model (CLSM) is proposed to model information retrieval, which use one global max-pooling layer and a deep strucure to extract fixed length sentence-level embedding. In \cite{match} a deeper and more interactive convolutional structure is proposed to matching sentences, by applying multiple maxpooling and convolutional layers~\cite{luan2016multiplicative}. The objective function of these models is to maximaze the similarity score between two sentence embeddings againest several negative samples, which is flexible in fitting in different tasks~\cite{luan2016lstm}. Tasks such as paraphrase identification, information retrieval have shown great improvement when applying this deep convolutional structures. Meanwhile, \cite{word2veccnn} trained a simple CNN with one layer of convolution on top of word embeddings obtained from an unsupervised neural language model \cite{word2vec,luan2018multi} and obtained excellent performance in sentence classification when fine tuning the pre-trained word vectors. Therefore, in this paper, we introduce pre-trained word embeddings and fine-tuning into CLSM and compare its performance with traditional letter-trigram CLSM on text simplification task. Our approach gives comparable performance as knowledge-based methods in \cite{sim}. Since the model is pre-trained, the approach is much faster when testing.
CLSM starts with each word within a temporal context window in a word sequence to directly capture contextual features at n-gram levels. Next, salient word n-gram features are extracted and then aggregated to form a sentence-level vector. Sentence level similarity can be calculated by cosine distance of vector representation of parallel sentences. CLSM fits well to our goal due to two reasons. First of all, sentence level alignment data can be used to train CLSM, then we can use CLSM to find similarity between new sentence pairs. Secondly, on max-pooling layer, salient feature are extracted. When tracing back to the neurons at max-pooling layer, n-grams that have high activation values for both sentences would have semantic matching, which could provide information of phrase-level matching, which is our next-step direction~\cite{luan2015efficient,luan2014relating,luan2017multi}.
The reminder of the report is organized as follows. In section 2, CLSM and its application to text simplification is briefly introduced. In section 3, the results of experiments is described. In section 4, experimental results are analyzed, insights for extracting phrase-level matches are described. Finally, Section 5 draws overall conclusions and describes possible future work.
\section{The CLSM Architecture}
\subsection{Letter-trigram based word n-gram representation}
In order to reduce vocabulary size and reduce out-of-vocabulary problem, we use word hashing technique, which represents a word by a letter-trigram. For example, given a word 'boy', the word is represented by '\#-b-o', 'b-o-y', 'o-y-\#'. The \textit{t} th word is represented as a count vector of letter trigrams $f_{t}$. In Figure \ref{1}, the letter trigram matrix $W_f$ denotes the letter-trigram transformation. Given the letter-trigram representation, each word trigram is represented by
\begin{equation}
l_t = [f_{t-1}^{T}, ..., f_{t}^{T}, ..., f_{t+1}^{T}], \quad t=1,...,T
\end{equation}
One problem of this method is collision, i.e., two different words could have the same letter n-gram vector representation. According to \cite{DSSM},
500K-word vocabulary can be represented by a 30,621 dimensional vector using letter trigrams, a reduction of 16-fold in
dimensionality with a negligible collision rate of 0.0044\%.
While the number of English words can be unlimited, the number of letter n-grams in English (or other similar languages) is often limited. Moreover, word hashing is able to map the morphological variations of the same word to the points that are close to each other in the letter n-gram space. More importantly, while a word unseen in the training set always cause difficulties in word-based representations, it is not the case where the letter n-gram
based representation is used. Thus, letter n-gram based word hashing is robust to the out-of-vocabulary problem.
\subsection{Word embedding initialization and fine-tuning}
Initializing word vectors with word embeddings obtained from an unsupervised neural language model is a popular method to improve performance in the absence of a large supervised training set (\cite{socher}, \cite{collobert}). One advantage of word embedding initialization compared to letter-trigram or bag of words is that it can significantly reduce feature dimension which could prevent overfitting and reduce computational cost. Word embedding initialization from unsupervised language model could introduce prior knowledge of the dataset, yet introduce mismatch for the specific training task at the same time. One way of solving the mismatching problem is to task specificaly fine-tune the word embeddings, which shows further gain in many previous researches \cite{word2veccnn}.
Among different word embedding approaches, \textit{word2vec}\cite{word2vec} has been widely applied to different NLP tasks, and proved great performance in word similarity task. The code for \textit{word2vec} is publicly available and can be trained efficiently from large corpus of natural language. \textit{word2vec} has two structures: Continuous Bag-of-Words Model (CBOW) and Continuous Skip-gram Model (Skip-gram), as Figure \ref{word2vec}. CBOW predicts the current word based on the context, and the Skip-gram predicts surrounding words given the current word. We test both structure and use Skip-gram since it gives best performance. In order to compare the performance of word embedding and letter-trigram, the context window remains the same across different experiments.
\begin{figure}[tb]
\centering
\includegraphics[width=14cm]{./word2vec.png}
\caption{{\it Two word2vec architectures, figure from \cite{word2vec}}}
\label{word2vec}
\end{figure}
\subsection{Convolutional layer, max pooling layer and latent semantic layer}
The convolutional operation\cite{luan2018uwnlp} can be viewed as sliding window based feature extraction. It is designed to capture the word n-gram contextual features. Consider the \textit{t}-th word n-gram, the convolution matrix projects its letter-trigram representation vector $l_t$ to a contextual feature vector $h_t$, $h_t$ is computed by
\begin{equation}
h_t = tanh(W_c \cdot l_t), \quad t=1,...,T
\end{equation}
where $W_c$ is the feature transformation matrix, as known as the convolution matrix, that are shared among all word n-grams. $tanh$ is used as the activation function of the neurons:
\begin{equation}
tanh(x) = \frac{1-e^{-2x}}{1+e^{-2x}}
\end{equation}
In order to aggregate the salient word into sentence-level feature vector, CLSM use max pooling to force the network to retain only the most useful local features produced by convolutional layers. As Figure \ref{1}, the max-pooling layer is
\begin{equation}
v(i) = max_{t=1,...,T}\{h_t(i)\}, \quad i=1,...,K
\end{equation}
where $v(i)$ is the \textit{i}-th element of the max pooling layer $v$, $h_t(i)$ is the \textit{i}-the element of the \textit{t}-th local feature vector $h_t$. $K$ is the dimensionality of the max pooling layer, which is the same as the dimensionality of the local contextual feature vectors \{$h_t$\}.
After the sentence-level feature is produced by the max-pooling operation, one more non-linear transformation layer is applied to extract the high-level semantic representation, denoted by $y$. As shown in Figure \ref{1}, we have
\begin{equation}
y=tanh(W_s\cdot v)
\end{equation}
where $v$ is the global feature vector after max pooling, $W_s$ is the semantic projection matrix, and y is the vector representation of the sentence in the latent semantic space, with a dimensionality of $L$.
\begin{figure}[tb]
\centering
\includegraphics[width=8cm]{./CLSM.png}
\caption{{\it The CLSM structure}}
\label{1}
\end{figure}
\subsection{Using CLSM for sentence similarity modeling}
We use cosine similarity to measure the similarity between two sentences. The score between a standard and a simple sentence is defined as:
\begin{equation}
\label{similarity}
R(sim, std)=cos(y_{sim},y_{std}) = \frac{y_{sim}^T y_{std}}{||y_{sim}||||y_{std}||}
\end{equation}
where $y_{sim}$ and $y_{std}$ are the semantic vector of the simple sentence and the standard sentence, respectively.
The data for training CLSM is the parallel simple and standard sentences from Wikipedia. CLSM is trained in a way that the score of matched sentences are maximized. In order to align sentences in simple document with sentences in standard document, for each simple sentence, 4 randomly selected standard sentences (matched sentence excluded) is collected together as its matched standard sentence. Therefore, the probability of the simple sentence paired with its matched standard sentence is given through softmax:
\begin{equation}
P(std^{+}|sim)=\frac{\exp{(R(sim, std^{+}))}}{\sum_{std^{'}\in\bm{std}}\exp(R(sim, std^{'}))}
\end{equation}
Likewise, the probability of the standard sentence paired with its matched simple sentence is given through softmax:
\begin{equation}
P(sim^{+}|std)=\frac{\exp{(R(std, sim^{+}))}}{\sum_{sim^{'}\in\bm{sim}}\exp(R(std, sim^{'}))}
\end{equation}
In training, the model parameters are learned to maximize the likelihood of the best match standard sentences given the target simple sentences and the best simple sentences given the target simple sentences. That is, we minimize the following loss function
\begin{equation}
L(\Lambda) = -\log \prod_{(sim, std^{+})}P(std^{+}|sim) \prod_{(std, sim^{+})}P(sim^{+}|std)
\end{equation}
\section{Baseline methods}
As a previous study, this work use the same dataset and evualuation metrics as \cite{sim}, which contains a sentence-level similarity score that builds
on a new word-level semantic similarity, described
below, together with a greedy search over the article.
\subsection{Word-level similarity}
\label{baseline}
Word-level similarity functions return a similarity
score $\sigma(w_1, w_2)$ between words $w_1$ and $w_2$. \cite{sim} introduces
a novel similarity metric: structural semantic similarity. The paper compare Wordnet structural semantic similarity scores \cite{wordnet} with Wiktionary structural semantic similarity scores. We therefore use these two methods as our baseline. For details of extracting Wordnet and Wiktionary similarity scores, please refer to \cite{sim} and \cite{wordnet}.
\subsection{Structural semantic similarity}
\cite{sim} extend the word-level similarity metric to account for both semantic
similarity between words, as well as the
dependency structure between the words in a sentence.
A triplet is created for each word using Stanford’s
dependency parser \cite{parser}.
Each triplet $t_w = (w, h, r)$ consists of the given
word $w$, its head word $h$ (governor), and the dependency
relationship (e.g., modifier, subject, etc)
between $w$ and $h$. The similarity between words
$w_1$ and $w_2$ combines the similarity between these
three features in order to boost the similarity score
of words whose head words are similar and appear
in the same dependency structure: $\sigma_{ss_{wk}} (w_1, w_2) =
\sigma_{w_k}(w_1, w_2) + \sigma_{w_k}(h_1, h_2)\sigma_r(r_1, r_2)$ where $\sigma{w_k}$ is
the WikNet similarity and $\sigma_{r}(r_1, r_2)$ represents dependency
similarity between relations $r_1$ and $r_2$
such that $\sigma_r = 0.5$ if both relations fall into the same
category, otherwise $\sigma_r = 0$.
\subsection{Greedy sequence-level alignment}
\label{greedy}
To avoid aligning multiple sentences to the same
content, one-to-one matches is required between
sentences in simple and standard Wikipedia articles
using a greedy algorithm. Similarities
between all sentences $S_{std}^{(j)}$ in the simple article
and $S_{sim}^{(i)}$ in standard article using a sentence level
similarity score is first computed. Then, iteratively
selects the most similar sentence pair $S_{std}^{*}, S_{sim}^{*} =\arg\!\max s(S_{std}^{(j)} , S_{sim}^{(i)}) $ and removes all other pairs associated
with the respective sentences, repeating until
all sentences in the shorter document are aligned.
\section{Experiments}
\subsection{Data sets and experimental setup}
The data for experiments is from \cite{sim}. Both manually and automatically
aligned sentence pairs are available.
The manually annotated dataset is used as a test set for
evaluating alignment methods as well as tuning parameters
for generating automatically aligned pairs
across standard and simple Wikipedia.
46 article pairs are randomly selected from Wikipedia (downloaded in June
2012) that started with the character 'a'. The annotators were given three choices: Good (complete match in semantics), Good partial (contains additional clause or phrase), Partial (discuss unrelated concepts, but share a short related phrase) and Bad (completely unrelated concepts).
The whole dataset include 67853 sentence pairs (277 good, 281 good partial, 117 partial and 67178 bad). The
kappa value for interannotator agreement is 0.68
(13\% of articles were dual annotated). Most disagreements
between annotators are confusions between
‘partial’ and ‘good partial’ matches. An example of good, good partial, partial pairs is listd in Table \ref{good}.
\cite{sim} release a dataset of automatically aligned sentence pairs,
with a scaled threshold greater than 0.45.1 In addition, around 51.5 million potential
matches, with a scaled score below 0.45, are pruned
from the dataset.
\begin{table*}[tb]
\centering
\caption{\label{good} Annotated examples: the matching regions for partial and good partial are italicized (table from \cite{sim})}
\begin{tabular}{c|c c}\hline
& standard & simple \\ \hline
good & \pbox{20cm}{Apple sauce or applesauce is a puree \\ made of apples.} & \pbox{20cm}{Applesauce (or applesauce) is a sauce that is \\made from stewed or mashed apples.}\\ \hline
\pbox{20cm}{good \\partial} & \pbox{20cm}{Commercial versions of applesauce \\ are really available in supermarkets.} & \pbox{20cm}{It is easy to make at home, and \textit{it is also sold} \\ \textit{already made in supermarkets as a common food}}.\\ \hline
partial & \pbox{20cm}{\textit{Applesauce} is a sauce that is \textit{made from}\\ stewed and mashed apples.} & \pbox{20cm}{\textit{Applesauce is made} by cooking down apples \\ with water or apple cider to the desired level.}\\ \hline
\end{tabular}
\end{table*}
In particular, the dataset use their best case method (structured Wordnet) to align sentences from 22k standard and simple articles, which were downloaded in April 2014.
Based on the precision-recall tested on annotated data, a scaled threshold
of 0.67 (P = 0.798, R = 0.599, F1 = 0.685)
for good matches is selected, which gives the highest f1 score. The selected
thresholds yield around 150k good matches,
130k good partial matches, and 110k uncategorized
matches.
In order to provide more discriminitave information to the model, we trace back to the article that contains the good sentence pairs and randomly select 9 sentences that have similarity score lower than 0.67 (the threshold that generate the wikipedia dump) as negative samples.
Each sample contains one simple sentence, one corresponding matched standard sentence (positive), and 9 randomly selected standard sentences (negative).
\subsection{Model variations}
\begin{description}
\item[CLSM9-letter] \hfill \\ Use letter trigram representation for each word. For the letter trigram that have frequency less than 5, replace the token as $<UNK>$. The total letter trigram size is 11024.This model is trained by randomly select 9 negative samples from the same article as positive sample, as described above.
\item[CLSM12-letter] \hfill \\ Use letter trigram representation for each word. This model is trained by randomly select 9 negative samples from the same article as positive sample and 3 negative samples randomly selected from different articles.
\item[CLSM9-word2vec] \hfill \\ Use Skip-gram \cite{word2vec} to represent each word. The word embeddings are fixed while training. If the word frequecy is less than 5, the word token is replaced by $<UNK>$. The total vocabulary size is 108874. 100, 150, 200, 250 dimension of word embeddings are tested. Word embeddings are pre-trained by all Wikipedia articles, including bothe simple and standard articles.
\item[CLSM9-word2vec-ft] \hfill \\ Use Skip-gram to represent each word. The word embeddings are fine tuned during training. 100, 150, 200, 250 dimension of word embeddings are tested.
\end{description}
In order to maintain consistency, the context window is kept 3 (word-trigram) across different models.
Hidden layer dimension use exactly the same parameters as \cite{clsm}. The dimension of convolutional layer and semantic layer are 300 and 120 respectively.
Learning rate are tuned from 0.001 to 0.1 for different models.
The model is trained by stochastic gradient descent.
\subsection{Experimental Results}
The performance of the four different models tested on automaticly aligned Wikipedia sentence pairs are shown in Table \ref{result1}.
The whole data set is split into three parts: 150k samples for training, 5k samples for validation and 5k for testing. The parameters with the best result are used for aligning annotated dataset. The validation and testing task is the identify the best matched pair given the target simple/standard sentence and the 9 negative samples (select 1 out of 10 sentences). For CLSM12-letter model, the task is to select 1 out of 13 sentences, the within article negative sentences remains the same across different experiments.
\begin{table*}[htbp]
\centering
\caption{\label{result1} Experimental results for automatic aligned Wikipedia sentence pairs}
\begin{tabular}{c|c|c}\hline
Model & validation accuracy (\%) & test accuracy (\%)\\ \hline
CLSM9-letter & $\bm{93.8}$ & $\bm{93.1}$ \\ \hline
CLSM12-letter & 93.2 & 92.8 \\ \hline
CLSM9-word2vec (100 dimension) & 92.3 & 92.8\\ \hline
CLSM9-word2vec (150 dimension) & 92.3 & 92.1\\ \hline
CLSM9-word2vec (200 dimension) & 92.7 & 92.4\\ \hline
CLSM9-word2vec (250 dimension) & 92.1 & 92.7\\ \hline
CLSM9-word2vec-ft (100 dimension) & 92.3 & 92.5\\ \hline
CLSM9-word2vec-ft (150 dimension) & 92.6 & 92.7\\ \hline
CLSM9-word2vec-ft (200 dimension) & 92.1 & 92.7\\ \hline
CLSM9-word2vec-ft (250 dimension) & 92.6 & 92.2\\ \hline
\end{tabular}
\end{table*}
From Table \ref{result1}, we could see letter trigram based models (CLSM9-letter and CLSM12-letter) give the best result. Word2vec initialiation and fine tuning are around the same performance. One possible reason is the data for pre-training word embedding is similar to the data training CLSM (CLSM data is a subset). Even though letter-trigram based models have much more parameters ($30k \times 300$) than word embedding based models on convolution layer ($30k \times 300$ for 100 dimension word embedding), the performance is not significantly improved. The computational time for CLSM9-letter on all Wikipedia pairs are over 72H, while the training time for word2vec embedding initialization is less than 4H.
In order to test the performance of sentence level alignment, we use best parameters in Table \ref{result1} to test the annotated dataset. The similarity score is calculated using Eq. \ref{similarity}. Then we do greedy sequence-level alignment based on the extracted score as described in Section \ref{greedy}. The precision, recall and f1 score are in Table \ref{result2}. The baseline methods are described in Section \ref{baseline}. Note that Structured WordNet is the method that generates our training data for CLSM.
\begin{table*}[htbp]
\centering
\caption{\label{result2} Experimental results for aligning annotated data}
\begin{tabular}{c|c|c|c}\hline
Model & precision & recall & f1 \\ \hline
WordNet (Baseline) & 0.80 & 0.60 & 0.69 \\ \hline
CLSM9-letter & 0.80 & 0.57 & 0.67 \\ \hline
CLSM12-letter & 0.81 & 0.55 & 0.66 \\ \hline
CLSM9-word2vec (100 dimension) & 0.73 & 0.56 & 0.63\\ \hline
CLSM9-word2vec-ft (150 dimension) & 0.76 & 0.56 & 0.64\\ \hline
\end{tabular}
\end{table*}
From Table \ref{result2}, CLSM give us a close performance as baseline method with the same precision and sligtly lower recall. This indicates that our model could learn the underlying relations between sentence pairs in training data.
We then tried rescoring similarity scores generated from CLSM9-letter model and WordNet, we further get a gain with 0.71 f1 score.
Since the training data is noisy, we could get a higher performance, therefore our next step is to use part of annotated data to adapt models trained from automatically generated data.
\section{Analysis}
According to \cite{clsm}, different words with related semantic meanings activate the similar set of neurons, resulting to a high overall matching score. This fact could give us inspiration to further pinpointing phrase-level matches within sentence pairs. Therefore, for each matched sentence pairs, we first project two sentences to max-pooling layer. Then, we evaluate the activation values of neurons at the max-pooling layer, and show the indice of the neurons that have high activation values of both sentences. After that, we trace back to the words that win these neurons in both sentences. An example of high activation neurons and the words in sentence pairs are shown in Figure \ref{2}. The simple sentence is `The books and poems that he changed into esperanto from other languages helped to make esperanto more well-known and used'. The corresponding standard sentence is `His translation had a influential impact on the development of esperanto into a language of literature'. From Figure \ref{2}, `books and poems' matches with `his translation', while `make esperanto more well-known and used' matches with `had an influential impact on the development of esperanto'. In this example, though there is no overlap between the two phrases, they both have high activation values at a similar set of neurons, thus lead to a sentence-level match in the semantic space. In future work, we could do phrase-level alignment based on this property of CLSM.
\begin{figure}[tb]
\centering
\includegraphics[width=8cm]{./analysis.png}
\caption{{\it An example of word-level semantic matching in two matched sentence pairs}}
\label{2}
\end{figure}
\section{Conclusion and future work}
In this project, we implemented a deep learning architecture called CLSM, the higher layer of the entire architecture is effective in generating latent semantic vector representations. CLSM is used for sentence level alignment in simple and standard Wikipedia.
CLSM gives good performance in matching standard and simple sentences. However, since the training data we use is automatically generated and is thus noisy. Our next step is to adapt the model to annotated data.
|
\subsection{Nonlinear compression with categorical variable}
Another common type of data has publishable features $X$ are high-dimensional continuous and the private labels $Y$ are discrete, for instance, images with some discrete labels (e.g. gender). Generally speaking, a sample $i$ has $y_i \in \mathcal{Y}=\{-1, +1\}$ and $x_i \in \mathbb{R}^p$ where $x_i^T$ is the $i$th row of the data matrix $X$. The data holder designs a nonlinear compression mechanism to reduce the classification accuracy of $y_i$ given $\tilde{x}_i$, where $\tilde{x}_i=g(x_i, y_i)$. We assume the attacker can use an advanced model, e.g. neural networks, to estimate the private labels. We further specify that $h$ and $g$ are functions parametrized by $\theta_h$ and $\theta_g$. The attacker minimizes the estimation loss, that is, $\min_{\theta_h}\ell(h_{\theta_h}(g_{\theta_g}(X, Y)), Y)$. The data holder designs a compressive function $g$ to maximize the attacker's loss as well as maintain a certain distortion $\gamma$ as aforementioned in equations (\ref{cdc:prob:formulation:general1}),(\ref{cdc:prob:formulation:general2}). This min-max game is difficult to find its equilibrium point in the context of neural networks with constraints, because the objective functions are non-convex with respect to parameters. Therefore, we use a heuristic way to cast the constrained optimization into a unconstrained optimization with regularization as follows:
\begin{align}
\max_{\theta_g}\Big\{\min_{\theta_h}\frac{1}{n}\sum_{i=1}^n\ell\Big(h_{\theta_h}\big(g_{\theta_g}(x_i, y_i)\big), y_i\Big) \\
- \beta\big( (\frac{1}{n}\sum_{i=1}^n\| g_{\theta_{g}}(x_i, y_i) - x_i\|^2) - \gamma \big)^2 \\
+ \rho \min\{0, \gamma - (\frac{1}{n}\sum_{i=1}^n\| g_{\theta_{g}}(x_i, y_i) - x_i\|^2) \} \Big\}, \label{cdc:prob:nonlinear:minmax:augment:term2}
\end{align}
where $\beta$ and $\rho$ are the hyper parameters controlling the iterates satisfied by the constraints. The distortion is characterized by the averaged Euclidean norm of the difference in samples. We propose a simple min-max alternative algorithm (Algorithm \ref{cdc:priv:alg:nn:gen:X_tilde}) to obtain the parameter $\theta_g$ for the function $g$ and yield the corresponding $\tilde{X}$.
\begin{algorithm}[!ht]
\caption{Generating $\tilde{X}$ (Neural Net attacker)}
\label{cdc:priv:alg:nn:gen:X_tilde}
\begin{algorithmic}[1]
\STATE \textit{Input:} dataset $\mathcal{D}$, parameter $\gamma$, iteration number $T$
\STATE \textit{Output:} Optimal data publisher parameters $\theta_{g}$
\STATE Initialize $\theta^t_{g}$ and $\theta^t_{h}$ when $t=0$
\FOR{$t=0,...,T$}
\STATE take minibatch of $n$ samples $\{x_{(1)}, \hdots,x_{(n)}\}$ drawn randomly from $\mathcal{D}$
\STATE Generate $\tilde{x}_{(i)}=g_{\theta_g}(x_{(i)},y_{(i)})$ for $i=1,\hdots,n$
\STATE Compute the parameter $\theta^{t+1}_h$ for the adversary {$\quad \theta^{t+1}_h=\arg\min_{\theta_{h}}\frac{1}{n}\sum\limits_{i=1}^{n}\ell(h_{\theta_h}(\tilde{x}_{(i)}),y_{(i)})$}
\STATE Compute the descent direction $\nabla_{\theta_{g}} \mathcal{L}(\theta_{g}, \theta^{t+1}_h)$, where
{$$
\nonumber
\mathcal{L}(\theta_{g},\theta^{t+1}_h)= -\frac{1}{n}\sum\limits_{i=1}^{n}\ell(h_{\theta^{t+1}_h}(g_{\theta_g}(x_{(i)},y_{(i)})),y_{(i)}) $$}{$$
+ \beta\big( (\frac{1}{n}\sum_{i=1}^n\|g_{\theta_g}(x_{(i)}, y_{(i)}) - x_{(i)}\|_2^2) - \gamma \big)^2 $$}
{$$ + \rho \max\{0, (\frac{1}{n}\sum_{i=1}^n\| g_{\theta_{g}}(x_i, y_i) - x_i\|^2) - \gamma \}
$$}
\STATE Perform backtracking line search along $ \nabla_{\theta_{g}} \mathcal{L}(\theta_{g}, \theta^{t+1}_h)$ and update
{$\theta^{t+1}_{g}= \theta^{t}_{g}-\alpha_t \nabla_{\theta_{g}} \mathcal{L}(\theta_{g}, \theta^{t+1}_h), \quad\alpha_t>0$}
\STATE Exit if solution converged
\ENDFOR
\STATE \textbf{return} $\theta^{t+1}_g$;
$\tilde{X} = g_{\theta^{t+1}_g}(X, Y)$
\end{algorithmic}
\end{algorithm}
Similar to the idea of the Augmented Lagrangian method\cite{wu2010augmented}, the scale of $\beta$ and $\rho$ are gradually increasing as the iteration step increases. The term (\ref{cdc:prob:nonlinear:minmax:augment:term2}) is added to ensure the solution strictly satisfies the constraint mentioned in expression (\ref{cdc:prob:formulation:general2}). Other alternative approaches are also proposed in \cite{goodfellow2014generative,hamm2016minimax,huang2017context}. Distinguished from those works, we construct a convex approximation with distortion constraints that is applied in privacy games.
\subsection{Case study: Images of people}
To perform our experiment of the nonlinear compressive model with a categorical response variable, we use the Groups of People dataset\cite{gallagher_cvpr_09_groups}. The dataset contains 4550 images from Flicker of human faces with labeled attributes such as age and gender. These images are $61 \times 49$ in grayscale pixels ranging from 0 to 255, with 3500 training and 1050 testing samples respectively. In this experiment, the images are $X$ and the label of gender, which is evenly spread in both the training and testing sets, is $Y$. We label female or male as 1 or -1. Sampled raw images are shown in Appendix \ref{cdc:appx:original:sampled:imgs}. %
For the data holder to perform nonlinear compression, we implement a three-layer neural network, which shares the similar concept of the autoencoder\cite{hinton2006reducing}. The first two layers serve as an encoder. The initial layer has 2989 units that takes original vectorized images [$2989=61 \times 49$], followed by a ReLU activation and batch normalization. We vary the second layer units from $2048, 512$, and $128$ for several cases, which are denoted as compression-rank. We define the corresponding compression-rank rate 0.685, 0.171, and 0.043 to be \emph{high}, \emph{medium} and \emph{low} respectively\footnote{the compression-rank rate is obtained by number of bottleneck units divided by input units. e.g. $\frac{2048}{2989} = 0.685$}. The last layer, connected with ReLU activation, has the same dimension as the vectorized image input that performs the role of a decoder.
The attacker is represented by a 3-layer neural network, comprised of an initial 2989 units layer, followed by 2048 units layer, and lastly a two units layer as softmax output. We apply leaky ReLU activation and batch normalization between each layer. %
Before considering adversarial compression, we first classify reconstructed images with different compression-rank rates without having a min-max game. This operation serves two purposes: a) investigating the accuracy of gender classification; b) fetching the minimum distortion threshold $\gamma$ in the context of mean squared error loss (i.e. min $\frac{1}{n}\sum_{i=1}^{n} \|\tilde{x}_i - x_i\|^2$ yields the smallest $\gamma$). The following results are evaluated based on the testing set. Figure~\ref{cdc:fig:img:wo:gan:rank} displays a sampled image associated with different scenarios. A lower compression rank rate yields a worse image quality. Table \ref{cdc:table:face:compress:wo:gan:clf:res} shows that compressing images with the high and medium ranks doesn't reduce the gender classification accuracy too much, yielding a relatively low image quality loss. In the example of high compression-rank, the average distortion per pixel is $0.0166 * 10^{2} \approx 1.6 \% $ which is not too large.
\begin{figure}[!ht]
\begin{center}
\centerline{\hspace{0.0in}
\includegraphics[width=0.23\columnwidth]{fig_faces_img/ex2_sample2_raw_2989.png}
\hspace{-0.2in}
\includegraphics[width=0.23\columnwidth]{fig_faces_img/ex2_sample2_h_2048.png}
\hspace{-0.2in}
\includegraphics[width=0.23\columnwidth]{fig_faces_img/ex2_sample2_h_512.png}
\hspace{-0.2in}
\includegraphics[width=0.23\columnwidth]{fig_faces_img/ex2_sample2_h_128.png}
}
\centering{\includegraphics[width=0.3\columnwidth]{fig_faces_img/ex2_sample2_diff_h_2048.png}
\hspace{-0.2in}
\includegraphics[width=0.3\columnwidth]{fig_faces_img/ex2_sample2_diff_h_512.png}
\hspace{-0.2in}
\includegraphics[width=0.3\columnwidth]{fig_faces_img/ex2_sample2_diff_h_128.png}
}
\caption{A sampled face. \textbf{Top row}: From left to right are an original image, and then decoded images with \emph{high, medium} and \emph{low} compression-rank rate. \textbf{Bottom row}: From left panel to right panels are the result of the pixel difference between decoded and raw image projected to 0-255 with \emph{high}, \emph{medium} and \emph{low} compression-rank rates respectively}
\label{cdc:fig:img:wo:gan:rank}
\end{center}
\end{figure}
\begin{table}[!ht]
\vspace{-0.15in}
\caption{Classification results of gender with the raw data under different compression-rank cases}
\label{cdc:table:face:compress:wo:gan:clf:res}
\centering
\begin{tabular}{cccc}
\hline
compression-rank & accuracy(gender) & distortion/pixel & distortion \\
\hline
raw (2989) & 0.692 & 0 & 0\\
high (2048) & 0.685 & 0.0166 & 0.195 \\
medium (512) & 0.664 & 0.0259 & 0.304 \\
low (128) & 0.627 & 0.0312 & 0.365\\
\hline
\end{tabular}
\end{table}
\vspace{-0.05in}
\par{
\setlength\intextsep{0mm}
Utilizing the previous result as a reference, we pick several proper values of $\gamma$ to further understand the adversarial privacy compression. In the high compression-rank case, we test three scenarios where $\gamma$ is $0.3, 2$ and $4$ respectively. We discover that the encoder-decoder tends to alternate pixels near eyes, mouths, and rims of hair. A similar patten can also be observed when we test the low compression-rank case where $\gamma$ is $1,2$ and $4$. We also notice that the low compression-rank scenario has a more scattered dotted patten of black/gray pixels at the large tolerance level, whereas the high compression-rank case has more concentrated black pixels, as shown in Figure~\ref{cdc:fig:altered:face:high:low:imgs}. We believe the reason is that the data holder always adjusts the pixels that are highly correlated with gender. Since the high compression-rank encoder-decoder preserves more information than the low compression-rank one, it's much easier for the data holder to alter the target pixel features within limited total distortion. The privatized images generated through min-max training indeed yield lower prediction accuracy of gender than the original encoded-decoded images. Table \ref{cdc:table:face:compress:gan:clf:res} depicts the gender classification results indicating that it is harder to predict gender with increased distortion. The table also reveals that higher compression rank performs better in terms of decreasing the accuracy if the distortion is sufficiently large. }
\begin{figure}[!ht]
\centerline{
\hspace{-0.12in}
\includegraphics[width=0.17\columnwidth]{fig_faces_img/ex1_adv_output_ae_2048_tol_0_002.png}
\hspace{-0.18in}
\includegraphics[width=0.17\columnwidth]{fig_faces_img/ex1_adv_output_ae_2048_tol_0_01.png}
\hspace{-0.18in}
\includegraphics[width=0.17\columnwidth]{fig_faces_img/ex1_adv_output_ae_2048_tol_0_021.png}
\hspace{-0.18in}
\includegraphics[width=0.22\columnwidth]{fig_faces_img/ex1_adv_diff_ae_2048_tol_0_002.png}
\hspace{-0.2in}
\includegraphics[width=0.22\columnwidth]{fig_faces_img/ex1_adv_diff_ae_2048_tol_0_01.png}
\hspace{-0.2in}
\includegraphics[width=0.22\columnwidth]{fig_faces_img/ex1_adv_diff_ae_2048_tol_0_021.png}
}
\centerline{
\hspace{-0.11in}
\includegraphics[width=0.17\columnwidth]{fig_faces_img/ex1_adv_output_ae_128_tol_0_0049.png}
\hspace{-0.18in}
\includegraphics[width=0.17\columnwidth]{fig_faces_img/ex1_adv_output_ae_128_tol_0_0063.png}
\hspace{-0.18in}
\includegraphics[width=0.17\columnwidth]{fig_faces_img/ex1_adv_output_ae_128_tol_0_0301.png}
\hspace{-0.2in}
\includegraphics[width=0.22\columnwidth]{fig_faces_img/ex1_adv_diff_ae_128_tol_0_0049.png}
\hspace{-0.2in}
\includegraphics[width=0.22\columnwidth]{fig_faces_img/ex1_adv_diff_ae_128_tol_0_0064.png}
\hspace{-0.2in}
\includegraphics[width=0.22\columnwidth]{fig_faces_img/ex1_adv_diff_ae_128_tol_0_0301.png}
}
\caption{A sampled image with different compression rank and distortion tolerance $\gamma$. \textbf{Top left}: From left to right are visual output when $\gamma$ equals 0.3, 2, 4 in the high compression-rank case. \textbf{Top right}: From left to right, these images show the difference between output images and raw images for corresponding $\gamma$ value in the high compression-rank case. \textbf{Bottom left}: From left to right are visual output when $\gamma$ equals 1, 2, 4 in the low compression-rank case. \textbf{Bottom right}: From left to right are difference between output images and raw images for corresponding $\gamma$ value in the low compression-rank case.}
\label{cdc:fig:altered:face:high:low:imgs}
\end{figure}
\begin{threeparttable}[!hbpt]
\vspace{-0.05in}
\caption{Classification accuracy of gender with the data released under adversarial privacy}
\label{cdc:table:face:compress:gan:clf:res}
\centering
\begin{tabular}{ccccc}
\hline
compression-rank & $\gamma=0.3$ & $\gamma=1$ & $\gamma=2$ & $\gamma=4$ \\
\hline
high (2048) & 0.628 & 0.600 & 0.573 & 0.486 \\
medium (512)\tnote{1} & & 0.607 & 0.594 & 0.512 \\
low (128) & & 0.602 & 0.585 & 0.521 \\
\hline
\end{tabular}
\begin{tablenotes}
\item[1] \small{$\gamma=0.3$ is unattainable, since the compression rank is small enough so that the minimum reconstruction loss (Mean Squared Error) already reachs to the 0.3.}
\end{tablenotes}
\end{threeparttable}
\subsection{Linear Compression with Categorical Label}
Here we consider a scenario that the private labels are discrete, and more specifically, say a sample $i$ that $y_i \in \mathcal{Y}=\{-1, +1\}, x_i \in R^p$. And $x_i^T$ is the $i$-th row of data matrix $X$. Data holder will design a linear compression mechanism to reduce the classification accuracy of $y_i$ given $\tilde{x}_i$, where $\tilde{x}_i=g(x_i, y_i)$, where $g$ is a function applies linear compression. For the attacker here we assume the loss function is logistic loss $\ell_{\theta}(h(\tilde{x}), y) = \sum_{i=1}^N \log(1+\exp(-y_i\theta^T\tilde{x_i})) $.
The compression operation is performed such that $\tilde{x}_i = AA^{\dagger}{x}_i$ where $A \in R^{p\times k}$ and $A^{\dagger} \in R^{k\times p}$ given the result in equation \ref{linear:sudo:inverseA}.
\begin{align}
& \max_{A}\bigg\{ \min_{\theta}\Big( \sum_{i=1}^N \log(1+\exp(-y_i\theta^T AA^{\dagger}x_i))\Big) \bigg\} \\
& s.t. \quad \|XAA^{\dagger} - X\|_F^2 \leq \gamma
\end{align}
Notice the $AA^{\dagger} = A(A^TA)^{-1}A^T$ is positive semidefinite (see appendix \ref{pes:priv:app:linear:tran:M:psd}). Therefore we solve a relaxed Semidefinite Programming with a rank constraint instead of explicitly solve $A$ for outer maximization. Re-parameterizing $M=AA^{\dagger}$ and applying the nuclear norm regularization to relax the rank constraint given $M$ is PSD, we have
\begin{align}
& \max_{M}\bigg\{ \min_{\theta}\Big( \sum_{i=1}^N \log(1+\exp(-y_i\theta^T Mx_i))\Big) - \beta \|M\|_{*} \bigg\} \\
& s.t. \quad \|XM - X\|_F^2 \leq \gamma \\
& \qquad M \succeq 0
\end{align}
we solve this game by using stochastic gradient decent method between attacker and data holder interactively, terminate the procedure when $\theta$ converges and $M$ satisfies the rank $k$. Given a $M$, the attacker's loss gradient is $
\nabla_{\theta}\ell = -\sum_{i=1}^N \sigma(-y_i\theta^T\tilde{x}_i)y_i \tilde{x}_i =-\sum_{i=1}^N \sigma(-y_i\theta^T M {x}_i)y_iM{x}_i $
where $\sigma(z)=1/(1+\exp(-z))$ is the sigmoid function.
The Hessian of logistic loss is $H_{\theta} = \sum_{i=1}^N\sigma(\theta^T\tilde{x}_i)\sigma(-\theta^T\tilde{x}_i)\tilde{x}_i\tilde{x}_i^T = \sum_{i=1}^N\sigma(\theta^TM{x}_i)\sigma(-\theta^TM{x}_i)M{x}_i {x}_i^TM^T$. Since $M$ is positive semidefinite, it's easy to discover that the Hessian $H_{\theta}$ is also positive semidefinite. Thus the inner minimization problem is a convex optimization. It has been discovered that a few samples can influence the attackers model parameter $\theta$ significantly, and with a linear transformation $M$ these effects could be amplified. We solve the $M$ and $\theta$ iteratively. A simple Algorithm \ref{pes:priv:alg:linear:gen:X_tilde:categ} is given as follows
\begin{algorithm}[ht]
\caption{Generating $\tilde{X}$ w/ categorical label (Linear attacker)}
\label{pes:priv:alg:linear:gen:X_tilde:categ}
\begin{algorithmic}[1]
\STATE \textit{Input:} dataset $(X, Y) \in \mathcal{D}$, iteration $T$, desired rank $k$ \\
\STATE \textit{Output:} $M, \tilde{X}$ \\
\STATE partition dataset into several batches ($X, Y$) \\
\FOR {a batch of ($X, Y$)}
\STATE initialize $t=0$, $M_t = I$, $\gamma_t, \beta_t$ for some value \\
\STATE $\theta_t = \arg \min_{\theta}\frac{1}{N} \sum_{i=1}^N \log(1+\exp(-y_i\theta^T M_t x_i))$
\WHILE {$t < T$}
\STATE $t=t+1$\\
\STATE $M_{t} = \big\{ \arg\max_{M} \frac{1}{N} \sum_{i=1}^N \log(1+\exp(-y_i\theta_t^T Mx_i))-\beta_t \|M\|_{*}$ \\
s.t. $\sum_{i=1}^N \sigma(-y_i\theta_t^T M {x}_i)y_iM{x}_i = 0 $;\\
$\quad \|XM - X\|_F^2 \leq \gamma \big\}$
\STATE $\theta_t \gets \arg \min_{\theta}\frac{1}{N} \sum_{i=1}^N \log(1+\exp(-y_i\theta^T M_{t} x_i))$
\STATE $U_t, \Lambda_t \gets $ SVD($M_t$) where $M_t = U_t\Lambda_tU_t^T$
\STATE $ \hat{k} = \tilde{rank}(\Lambda_t)$. check the rank of the matrix $\Lambda$ with non trivial eigenvalues. (e.g $\lambda_j > \eta \lambda_{max}, \text{where } \forall j =1,\hdots,n; \eta = 0.01$.)
\IF {$\hat{k} = k $ and $\|M_t - {M_{t-1} }\|_F \leq \epsilon$ }
\STATE break
\ELSIF {$\hat{k} > k $}
\STATE $\beta_{t+1} \gets \beta_t + \frac{\beta_t}{2}$
\ELSIF {$\hat{k} < k $}
\STATE $\beta_{t+1} \gets \beta_t - \frac{\beta_t}{4} $
\ENDIF
\ENDWHILE
\STATE $M = \frac{1}{T}\sum_{t=1}^T M_t$; $\tilde{X} = XM$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\section{Conclusion}
\label{cdc:conclusion:s5}
Recent breakthroughs in artificial intelligence require a huge amount of data to support the learning quality of various models. Yet the risk of data privacy is often overlooked in the current data sharing processes. The recent news of data leakage by Facebook shows that privacy risk could significantly impact some issues in politics. Therefore, securely designing a good data privatization mechanism is important in the context of utilizing machine learning models. By thorough evaluations, our new min-max adversarial compressive privacy framework provide an effective and robust approach to protect private information. We leverage the data-driven approach without posing assumptions on data distribution. It's crucial during practical implementation, since the real data is often more complicated than a simple characterization of a parametric probability distribution. Along this line of research, many interesting extensions can be built on our framework to create a robust privacy protector for data holders.
\section{Introduction}
Machine learning has progressed dramatically in many real-life tasks such as classifying image\cite{Oshri:2018:IQA:3219819.3219924}, processing natural language\cite{sutskever2014sequence}, predicting electricity consumption\cite{sevlian2018scaling}, and many more. These tasks rely on large datasets that are usually saturated with private information. Data holders who want to apply machine learning techniques may not be cautious about what additional information the model can capture from training data, as long as the primary task can be solved by some model with high accuracy.
In this paper, we propose a privatization mechanism to avoid the potential exposure of the sensitive information while still preserving the necessary utility of the data that is going to be released. This mechanism largely leverages the concept of the game-theoretic approach by perturbing the data and retraining the model iteratively between data holder and malicious data attacker. Such a data perturbation idea is highly correlated with feature transformation and selection of the raw data input that correlated with private labels. %
Protecting privacy has been extensively explored in myriad literature. A popular procedure is to anonymize the identifiable personal information in datasets (e.g. removing name, social security number, etc.). Yet anonymization doesn't provide good immunity against correlation attacks. A previous study \cite{narayanan2008robust} was able to successfully deanonymize watch histories in the Netflix Prize, a public recommender system competition. Another study designed re-identification attacks on anonymized fMRI (functional magnetic resonance imaging) imaging datasets\cite{finn2015functional}. On the other hand, the Differential privacy (DP) \cite{dwork2008differential} has a strong standard of privacy guarantee and is applicable to many problems beyond database release \cite{dwork2014algorithmic}. This DP mechanism has been introduced in data privacy analysis in control and networks\cite{cortes2016differential,katewa2015protecting, huang2015controller, koufogiannis2017differential}. In particular, \cite{cortes2016differential} gave a thorough investigation on performing the centralized and distributed optimization under differential privacy constraints. In this line of research, \cite{huang2015controller} and \cite{koufogiannis2017differential} focused on the cases of dynamic data perturbations in control systems. \cite{katewa2015protecting} presented a noise adding mechanism to protect the differential privacy of network topology.
However, training machine learning models
with DP guarantees using randomized data often leads to a significantly reduced utility and comes with a tremendous hit in sample complexity\cite{duchi2013local, fienberg2010differential}. A recent work \cite{abadi2016deep} applied the DP concept on a deep neural network to demonstrate that a modest accuracy loss can be obtained at certain worst-case privacy levels. However, this was still a ``context-free'' approach that didn't leverage the full structure between the data input and output. %
To overcome the aforementioned challenges, we take a new holistic approach towards enabling private data publishing with consideration on both privacy and utility. Instead of adopting worst-case, context-free notions of data privacy (such as differential privacy), we introduce a context-aware model of privacy that allows the data holder to cleverly alter the data where it matters. %
\input{priv_pgf_scheme_v2}
Our main contributions are listed as follows.
First, with the goal of having a ``distribution free'' data releasing mechanism and inspired by general min-max games, we investigate a typical way of perturbing the data that is the compression using the data-driven approach. As a second contribution, we formulate the interaction between data holders and attackers through convex optimization during the min-max game when both players apply linear models. A corresponding equilibrium can be found and used as the optimal strategy for the data holder to yield the altered data. The third contribution is that our thorough evaluations of realistic datasets demonstrate the effectiveness of our compressive adversarial privacy framework. Finally, we leverage the mutual information to validate that sensitive information can be protected from the privatized data. %
The remainder of our paper is arranged as follows. In section \ref{cdc:priv:game:state:sec:2}, we introduce the general adversarial privacy game. Section \ref{cdc:cap:statement} describes the compressive adversarial privacy game with several cases of realistic data analyses. Section \ref{cdc:quantify:priv:garentee:s4} describes the quantification of privacy. Section \ref{cdc:conclusion:s5} concludes the paper.
\section{Compressive Adversarial Privacy}
\label{cdc:cap:statement}
A typical method to enforce data privacy is data compression. This method is well studied in \cite{zhou2009differential} from a theoretical point of differential privacy. In reality, data compression is used in many applications such as text messaging and video transmission to protect the privacy. In this section, we extend the general min-max framework to a compression approach, namely a compressive adversarial privacy framework, as shown in Figure \ref{cdc:fig:compress:diag}\\
\input{priv_pgf_autoencoder.tex} We focus on two scenarios to illustrate concrete privatization mechanisms. The first one is the linear compression when an attacker takes a linear model. The second one is the non-linear compression when an attacker uses a neural network. We evaluate of both cases based on real data.
\subsection{Linear compression with continuous label}
We introduce the case of an attacker who uses a linear model and the least squared loss function to infer private information through released data. This case is practical especially when private labels are continuous and have a linear relationship with the original data. We denote the original data matrix $X$ and altered data matirx $\tilde{X} \in \mathbb{R}^{n\times p}$, where $n$ is number of samples, $p$ is number of features, and $\mathbb{R}$ is the set of real numbers. Such a data matrix contains individual samples $x_i$ and $\tilde{x}_i \in \mathbb{R}^p$ respectively, where $i=1,\dots, n$. The private-info matrix $Y \in \mathbb{R}^{n\times d}$ consists of $y_i \in \mathbb{R}^d$ that each sample $i$ has $d$ types of private labels. %
Consider a data holder who has a simple data-releasing mechanism that applies a linear transformation of $X$, namely projecting it down to lower dimensions to protect confidential information, i.e. $Z=XA$, where the matrix $A \in \mathbb{R}^{p\times k}, k < p$. Hence, $Z \in \mathbb{R}^{n\times k}$. In order to release meaningful data that still can be utilized by a majority of good users, the data holder performs a linear operation by multiplying $B \in \mathbb{R}^{k\times p}$ on $Z$ and recovers it back to the same dimension as $X$, yielding $\tilde{X} = XAB$. The attacker fits a linear model to minimize the mean squared loss that is $\frac{1}{n} \sum_{i=1}^n \|\tilde{\Theta}^T\tilde{x}_i - y_i\|_2^2 = \frac{1}{n} \sum_{i=1}^n \| \tilde{\Theta}^T B^TA^T x_i - y_i\|_2^2$, where $\tilde{\Theta} \in \mathbb{R}^{p \times d}$. Because the domain of $\tilde{\Theta}^TB^T$ is contained in $\Theta^T$ which is in $\mathbb{R}^{d\times k}$. This attacker's loss is lower bounded by
\begin{align}
\min_{\Theta}\frac{1}{n}\sum_{i=1}^n\|\Theta^T A^Tx_i - y_i \|_2^2 ,
\end{align}
where $\Theta \in \mathbb{R}^{k \times d}$.
Therefore, when the data holder maximizes the attacker's loss, we can maximize this lower bound that automatically maximizes the minimum loss of the attacker. The resulting min-max problem can be formulated as
\begin{align}
\max_{A, B} \min_{\Theta}\frac{1}{n}\sum_{i=1}^n\|\Theta^T A^Tx_i - y_i \|_2^2 \\
s.t. \quad \|XAB - X\|_{F}^2 \leq \gamma ,
\end{align}
where $\| \cdot \|_{F}$ is the Frobenius norm. Given $A$, we further simplify the expression by finding the best recovering matrix $\hat{B}$ in place of $B$ as follows (see \ref{priv:icml2018:appendix:recover:mat:B} for details):
\begin{align}
\hat{B} = (A^TX^TXA)^{-1} A^TX^TX = (A^TA)^{-1}A^T = A^{\dagger} .
\label{linear:sudo:inverseA}
\end{align}
We denote the $A^{\dagger}$ to be the pseudo-inverse of $A$. The best predictor $\Theta$ for the attacker can be expressed as $\Theta = (A^T C_{xx} A)^{-1} A^T C_{xy}$, where $C_{xx}=\frac{1}{n}\sum_{i=1}^n x_ix_i^T, C_{xy}=\frac{1}{n} \sum_{i=1}^n x_iy_i^T $. Substituting $\Theta$ and $B$, we have the following problem:
\begin{align}
\max_{A} \bigg\{ - \mathop{\rm Tr}\Big( C_{xy}^T A(A^T C_{xx} A)^{-1}A^TC_{xy} \Big) \bigg\}\\
s.t. \quad \|X A(A^TX^TXA)^{-1}A^T X^T X - X\|_{F}^2 \leq \gamma .
\end{align}
Notice that $C_{xx} = \frac{1}{n}X^TX$. By flipping the sign of the maximization and denoting $M = A(A^TX^TXA)^{-1}A^T$ which is a positive semidefinite matrix (see the appendix \ref{pes:priv:app:linear:tran:M:psd}), we have the following problem:
\begin{align}
& \min_{M} \frac{1}{n}\mathop{\rm Tr}\big( C_{xy}^TXMX^TC_{xy} \big) \\
& s.t. \qquad M \succeq 0 \\
& \qquad \|XMX^TX - X\|_{F}^2 \leq \gamma \\
& \qquad rank(M) = k . \label{prob:formulate:psd:rank:constraint}
\end{align}
We put a rank constraint (\ref{prob:formulate:psd:rank:constraint}) because the dimension $A$ is $p\times k, (k<p)$. This problem can be further relaxed to a convex optimization by regularizing the nuclear norm of matrix $M$ as follows:
\begin{align}
\min_{M} & \frac{1}{n}\mathop{\rm Tr}\big( C_{xy}^TXMX^TC_{xy} \big) + \beta \|{M}\|_{*}
\label{pes:linear:filter:X:param:M:cvx_form_1} \\
s.t. & \quad \|XMX^TX - X\|_{F}^2 \leq \gamma \\
& \quad {M} \succeq 0 , \label{pes:linear:filter:X:param:M:cvx_form_2}
\end{align}
where $\|\cdot \|_{*}$ is the nuclear norm of a matrix that heuristically controls the rank of a matrix. Such a convex relaxation allows the data publisher to find an optimal solution of $M$, and correspondingly yields the appropriate $\tilde{X}$ (see appendix \ref{priv:icml2018:cast:linear:implicit:solve}). Thus, both players can achieve an equilibrium in this game. To ensure the problem is feasible, one caveat is that we cannot pick arbitrarily small $\gamma$ without considering the aforementioned rank $k$. We note that $\tilde{X} = XAA^{\dagger}$ is a low rank-$k$ approximation of the original data matrix $X$.
\begin{theorem}\label{cdc:svd:err:th1}
Suppose a rank-$p$ matrix $X$ consists of the singular values $\lambda_1, \dots, \lambda_p$. With the best rank-$k$ approximation $\tilde{X}_k$ under Frobenius norm, the distortion threshold $\gamma$ is at least $\sum_{i=k+1}^p \lambda_i^2$.
\end{theorem} %
We put the proof in appendix \ref{cdc:appx:proof:Th1}. The theorem reveals the relationship between setting the distortion tolerance $\gamma$ and the rank $k$.
Hence, a simple algorithm ({Algorithm \ref{pes:priv:alg:linear:gen:X_tilde}}) is proposed for the data holder to generate $\tilde{X}$.
\begin{algorithm}[h!]
\caption{Generating $\tilde{X}$ (Linear attacker)}
\label{pes:priv:alg:linear:gen:X_tilde}
\begin{algorithmic}[1]
\STATE \textit{Input:} dataset $ (X, Y) \in \mathcal{D}$, parameter $\gamma$, $\beta_0$, $k$, $\eta$.
\STATE \textit{Output:} $\tilde{X}$
\STATE partition dataset into several batches $(X,Y)$.
\FOR {a batch of $(X,Y) \in \mathcal{D}$}
\STATE $\hat{k}=0$, $t=0$
\WHILE {$\hat{k} \neq k$}
\STATE $ {M}_{t} = \arg\min_{ {M }}\{ \frac{1}{n} \mathop{\rm Tr}( C_{xy}^TX {M}X^TC_{xy}) + \beta_t \|{M}\|_{*}$ \\
s.t. $\quad \|XMX^TX - X\|_F^2 \leq \gamma \}$ (solving the optimization in equation (\ref{pes:linear:filter:X:param:M:cvx_form_1} - \ref{pes:linear:filter:X:param:M:cvx_form_2}) with certain values of $\gamma, \beta_t, k$).
\STATE $U_{t}, \Lambda_{t} \gets SVD({M}_{t})$ (applying Singular Value Decomposition on ${M}_t = U_t\Lambda_t U_t^T$ to get matrices $U_t$ and $\Lambda_t$.)
\STATE $ \hat{k} = \tilde{rank}(\Lambda_t)$. check the rank of the matrix $\Lambda$ with non trivial eigenvalues. (e.g $\lambda_j > \eta \lambda_{max}, \text{where } \forall j =1,\hdots,n; \eta = 0.01$.)
\IF {$\hat{k} = k $}
\STATE \textbf{break}
\ELSIF {$\hat{k} > k $}
\STATE $\beta_{t+1} \gets \beta_t + \frac{\beta_t}{2}$
\ELSIF {$\hat{k} < k $}
\STATE $\beta_{t+1} \gets \beta_t - \frac{\beta_t}{4} $
\ENDIF
\STATE {$t = t+1$}
\ENDWHILE
\STATE $\tilde{X} \gets X{M}X^TX $
\ENDFOR
\end{algorithmic}
\vspace*{-0.04in}
\end{algorithm}
\begin{remark}
This approach can be interpreted as releasing a low dimensional approximation to a set of data, incorporating the relation between the original data and private labels, while still maintaining a certain distortion between the released data and the original data.
\end{remark}
We also discovered that a similar scheme can be applied on compressing original data with additive Gaussian noise.
See Appendix \ref{icml:appendix:linear:compress:noised:data} for details.
\subsection{Case study: Power consumption data}
The first experiment of our analysis uses the CER dataset, which was collected during a smart metering trial conducted in Ireland by the Irish Commission for Energy Regulation (CER) \cite{CER2009Dataset}. The dataset contains measurements of electricity consumption gathered from over 4000 households every 30 minutes between July 2009 and December 2010. Each participating household was asked to fill out a questionnaire about the households' socio-economic status, appliances stock, properties of the dwelling, etc.\cite{beckel2014revealing}. To demonstrate our concepts, we sampled a portion of the customers who has valid entries of demographic information, e.g. number of appliances and floor area of the individual house. In the following experiment, we treat floor area as private data $Y$. %
Throughout the case simulations, we extract the four-week time series in September 2010. Since the power consumption (in kilowatts) is recorded every 30 minutes, there are $2 \times 24 \times 28 = 1344$ entries for a single household. To simplify the input dimension and avoid the over-fitting issue from raw input, we compute a set of features on the electricity consumption records of a household. The features then serve as the input to the prediction model. Table \ref{priv:num_sim:feature:table} lists all 23 features we calculated from electricity consumption data, which is also used in \cite{beckel2014revealing}. We treat these features as $X$ and normalize them such that they range from 0 to 1. Data normalization is required in our experiment in that it gets rid of the scale inconsistency across the different features.%
\begin{figure}[!ht]
\centering
\framebox{\parbox[b]{1.51in}
{\includegraphics[scale=0.21]
{fig_linear/hist_floor_area.png} }{\includegraphics[scale=0.19]{fig_linear/regression_normed_fl_area_fit.png}}
}
\caption{Display the sampled data. \textbf{Left panel} shows the histogram of \emph{Floor area} of each house. \textbf{Right panel} shows the prediction capability of a linear model that fits on the \emph{Floor area}}
\label{pes:priv:hist:n_fl_area}
\end{figure}
In the linear transformation model, given $Y$ is the private information, we run algorithm \ref{pes:priv:alg:linear:gen:X_tilde} to release $\tilde{X}$. This procedure involves solving semidefinite programming, which could be slow when the dimension of input samples is large. So we partition the samples into several groups with a reasonable number of households in each group (e.g. 30 to 40 as long as the number of households is larger than the number of features). After running experiments on several rank conditions of data matrices, we found that lower rank indicates better privacy (higher prediction error), given in Table~\ref{priv:table:linear:svd:statistics}. With a low rank condition that the data holder maintains, the attacker can barely (see Table~\ref{priv:table:linear:svd:statistics}) predict the private label $Y$. We also partitioned the data into 80\% for training and 20\% for testing. Table \ref{priv:table:linear:svd:statistics} shows the corresponding results with different ranks of the compression matrix for the testing set. A batch of released data differs from the original when rank is $4, 10$ and $18$. The difference is shown in Figure~\ref{cdc:priv:linear:diff:batch:samples}.
\begin{figure*}[!bpht]
\centering
\framebox{\parbox[b]{2.0in}
{\includegraphics[scale=0.26]{fig_linear/margin_orig_alter_rank_4_heatmap.png}
}
{\includegraphics[scale=0.26]{fig_linear/margin_orig_alter_rank_10_heatmap.png}
}
{\includegraphics[scale=0.26]{fig_linear/margin_orig_alter_rank_18_heatmap.png}
}}
\caption{Difference between altered and original data when rank equals 4, 10, and 18}
\label{cdc:priv:linear:diff:batch:samples}
\end{figure*}
\begin{table}[!ht]
\vspace{-0.11in}
\centering
\caption{Metrics of Linear transformed data}
\label{priv:table:linear:svd:statistics}
\begin{tabular}{cccc}
\hline
Rank & RMSE & $R^2$ & distortion \\
\hline
4 & 6.7e+04 & 1.61e-08 & 0.616 \\
10 & 7.1e+01 & 2.28e-03 & 0.081\\
18 & 8.7e-01 & 1.12e-02 & 0.079 \\
23 & 3.9e-02 & 8.01e-01 & 0 \\
\hline
\end{tabular}
\end{table}
\vspace{-0.19in}
\section{Privacy Preserved Releasing}
\label{cdc:priv:game:state:sec:2}
We propose a general data publishing framework by incorporating the following game concept. In general there are two roles in this data releasing game: a data holder and a data consumer. Among data consumers, some people are good users who explore the pattern and extract the value of the data. We merge the good data consumers together with data holders and address their roles from the data holder perspective, since good data user are irrelevant in this game.
Yet there are people who also try to learn personal sensitive information from the data on purpose. We define those malicious users as attackers. We focus on the data holders and attackers for the following description of the game. %
Consider a dataset $\mathcal{D}$ which contains both the original data $X$ and the associated customers' sensitive demographic information $Y$ (e.g. account holder age, house square-footage, gender, etc.). Thus a sample $i$ has a record $(x_i, y_i) \in \mathcal{D}$. We denote the function $g$ as a general mechanism for a data holder to release the data. The released data are denoted as $\tilde{x_i}$ for the customer $i$, which can also be described as $\tilde{x_i}= g(x_i,y_i)$. Notice we don't release $y_i$ to the public because it's private information. Generally speaking, $\tilde{X}=g(X, Y)$. Let the function $h$ represent the adversarial hypothesis, e.g. the estimated outcome $\hat{Y}=h(\tilde{X})$. The attacker would like to minimize the inference loss on private labels, namely $\ell(\hat{Y}, Y)$ given some loss function $\ell$,
while the data holder would like to maximize the attacker's loss, and in principle, also wants to preserve the quality of the released data for research purposes. This data quality is characterized by some distance function measuring between the original data and the altered data. Therefore, we formulate a min-max game between the data holder and the attacker as follows:
\begin{align}
\max_{g \in \mathcal{G}}\Big\{\min_{h \in \mathcal{H}} \ell\Big(h\big(g(X, Y)\big), Y\Big) \Big\} \label{cdc:prob:formulation:general1}\\
s.t. \quad d\big( X, g(X, Y) \big) \leq \gamma \label{cdc:prob:formulation:general2},
\end{align}
where $d()$ could be some distance function, such as Total Variation (TV), Wasserstein-1, or Frobenius norm, etc. \cite{nguyen2009surrogate, arjovsky2017wasserstein}, and $\gamma$ is a hyper-parameter. The constraint ensures that the released data will not be distorted too much from the original data.
This framework allows an attacker to incorporate any loss functions and design various adversarial inference models, which typically take the released data to predict the personal information. Given such a challenge, the data publisher has to design a good privatization mechanism $g$ and $\gamma$ to deteriorate the attacker's performance, which are also data dependent. For simplicity, we focus on the supervised learning setting in this work, but the concept can potentially be extended to the unsupervised learning.
\section{Privacy Guarantee}
\label{cdc:quantify:priv:garentee:s4}
Our previous experiments show that a (local) equilibrium can be achieved through this min-max game approach. While we cannot preclude that there may be some other equilibria in the context of a neural network. Thus, a quantifiable metric is needed to give privacy guarantee between sensitive data and altered data.
In this section we introduce the empirical mutual-information concept to quantify the privatization quality of this min-max game approach, i.e. measuring the correlation between the sensitive response data and the released feature data pre- and post-privatization. Mutual information
(MI) \cite{shannon2001mathematical} is a well established tool that has been widely adopted to quantify the correlation between the two streams of data by a non-negative scalar \cite{cover2012elements}.
From the data driven perspective, we have the empirical MI $\hat{I}(X; Y) = \hat{H}(X) - \hat{H}(X|Y)$, where $\hat{H}$ characterizes the empirical entropy. This empirical entropy can be calculated using the classical nearest $k$-th neighbor method \cite{kraskov2004estimating}.
\textbf{Continuous response label}: Given that $Y$ is continuous, the mutual information can be expressed as
$\hat{I}(X;Y) = \hat{H}(X) + \hat{H}(Y) - \hat{H}(X,Y)$,
where $\hat{H}(X), \hat{H}(Y)$ can be obtained directly from method in \cite{kraskov2004estimating},
given samples $x_i, y_i$. The joint empirical entropy is calculated by concatenating each $x_i$ and $y_i$ together as one sample and using the nearest neighbor entropy estimation again.
\textbf{Categorical response label}:
For the discrete response $Y \in \{-1, +1\}$, we have the mutual information
$\hat{I}(X;Y)= \hat{H}(X)-\hat{H}(X|Y) = \hat{H}(X)- \big(p(Y=-1)\hat{H}(X|Y=-1) + p(Y=1)\hat{H}(X|Y=1) \big)$,
where $p(Y=\pm 1)$ can be approximated by the sample frequency in the dataset, and $\hat{H}(X|Y=\pm 1)$ can be obtained by the aforementioned $k$-th nearest neighbor method with partitioned samples according to the value of $Y$.
For the experiment of the continuous response variable, we first calculate the empirical MI between the power consumption statistics and floor areas. The original MI between power usage statistics data and floor area is $\hat{I}(X;Y) = 2.150$. The resulting MI between altered power usage data and floor areas, which is denoted by $\hat{I}(\tilde{X};Y)$, are 0.995, 0.494, and 0.216 when the rank of compression matrices are $18$, $10$, and $4$. For the categorical response variable experiment, the empirical MI between the images data and gender data is obtained as follows. The original MI between raw images $X$ and gender label $Y$ is $\hat{I}(X;Y) =0.249$. When we pick high compression-rank with $\gamma=0.3, 1, 2$ and $4$, the $\hat{I}(\tilde{X}; Y)$ are $0.217, 0.170, 0.105$, and $0.012$. The medium compression-rank yields $\hat{I}(\tilde{X}; Y)$ to be $0.179, 0.112$, and $0.014$ for $\gamma = 1, 2,$ and $4$ respectively. In the low compression rank case, $\hat{I}(\tilde{X};Y)$ are $0.174, 0.101,$ and $0.017$ with the aforementioned $\gamma$. We notice the empirical MI indeed decreases as the distortion increases. Due to the challenge of high dimensional data, we apply the principal component analysis to project the $\tilde{X}$ down to 16 dimensions and find the approximate $\hat{I}(\tilde{X};Y)$. This is an alternative attempt to demonstrate the effectiveness of using our framework. The changes of mutual information value show the privacy guarantee between the released data and sensitive labels under various distortion conditions. Yet we believe a more advanced architecture of the neural network can be applied to extract the embeddings of semantic features, resulting a better estimates of empirical mutual information. We will explore this potential direction in our future research.
\section{Appendix}
\subsection{Casting linear problem}
\label{priv:icml2018:cast:linear:implicit:solve}
Consider the following problem
\begin{align}
\min_{B} \|XAB - X\|_F^2,
\end{align}
where matrix $X$ is $n\times p$, matrix $A$ is $p \times k$, and matrix $B$ is $k \times p$. We give a brief minimizer derivation as follows:
\begin{align*}
\|XAB - X\|_{F}^2 = \mathbf{Tr}\Big( (XAB - X)(XAB - X)^T \Big) \\
= \mathbf{Tr}\Big( XABB^TA^TX^T - 2XABX^T + XX^T\Big) .
\end{align*}
The derivative of the first term with respect to $B$ is
\begin{align*}
\frac{\partial}{\partial B}\mathbf{Tr} (XABB^TA^TX^T ) = \frac{\partial}{\partial B}\mathbf{Tr} (BB^TA^TX^TXA) \\ = (A^TX^TXA) B + (A^TX^TXA)^T B = 2 (A^TX^TXA) B .
\end{align*}
The derivative of the second term with respect to $B$ is
\begin{align*}
\frac{\partial}{\partial B}\mathbf{Tr}(2XABX^T) = 2\frac{\partial}{\partial B}\mathbf{Tr} (B X^TXA) = 2 (X^TXA)^T
\end{align*}
Thus, we set the derivative equals zero and obtain the minimizer $B$ as follows:
\begin{align}
\frac{\partial}{\partial B } \mathbf{Tr}\Big( XABB^TA^TX^T - 2XABX^T + XX^T\Big) \\
= 2 (A^TX^TXA) B - 2 (X^TXA)^T \triangleq 0 \\
\implies (A^TX^TXA) B = A^TX^TX \\
\implies B = (A^TX^TXA)^{-1}A^TX^TX .
\end{align}
Now we design the $\tilde{X}$ such that
\begin{align}
\tilde{X} = XAB = XA\underbrace{(A^TX^TXA)^{-1}A^TX^TX}_{B} \\
= X\underbrace{A(A^TX^TXA)^{-1}A^T}_{M}X^T X .
\end{align}
Instead of explicitly designing a low rank matrix $A$ , we solve an alternative equivalent problem of determining a low rank matrix ${M}$ to compress the data.
\subsection{Recovering Linear Operation}
\label{priv:icml2018:appendix:recover:mat:B}
Claim: $B$ is pseudoinverse of $A$, i.e. $B = (A^TA)^{-1}A^T = A^{\dagger} $.
\proof We apply Singular Value Decomposition (aka SVD, which is similar to PCA) on data matrix $X = USV^T$. $U$ is $n\times k$, $S$ is $k \times k$, $V$ is $p \times k$. We have
\begin{align*}
B & =(A^TX^TXA)^{-1}A^TX^TX \\
& = \big(A^T(USV^T)^T(USV^T)A \big)^{-1} A^T (USV^T)^T(USV^T)
\\
& = (A^TVSU^TUSV^TA)^{-1} A^T VSU^TU SV^T \\
& = (A^TVS\underbrace{U^TU}_{I} SV^TA)^{-1} A^T VS\underbrace{U^TU}_{I} SV^T \\
& = (A^TVS^2V^TA)^{-1} A^T VS^2V^T \\
& = (\underbrace{S^2A^TV}_{\text{swapped }}V^TA)^{-1} \underbrace{ S^2 A^TV}_{\text{swapped}}V^T \\
& = (S^2 A^T\underbrace{VV^T}_{I}A)^{-1} S^2 A^T \underbrace{VV^T}_{I} \\
& = (S^2A^TA)^{-1} S^2 A^T =
(A^TA)^{-1} A^T .
\end{align*}
\subsection{Proof of positive semidefinite property of a matrix} \label{pes:priv:app:linear:tran:M:psd}
Claim: $M = A(A^TX^TXA)^{-1}A^T$ is Positive Semidefinite.
\proof show $A^TX^TXA$ is positive semidefinite. Since $A \in \mathbb{R}^{p\times k}, X \in \mathbb{R}^{n \times p}$, for any vector $v \in \mathbb{R}^k$, we have
\begin{align*}
v^TA^TX^TXAv = (XAv)^T(XAv) = \|XAv\|_2^2 \geq 0 .
\end{align*}
Therefore, we can apply Signaler Value Decomposition on $(A^TX^TXA)$, we get $(A^TX^TXA)= VSV^T$, where $S = \mathop{\rm diag}(\sigma_1, \sigma_2, ..., \sigma_k)$. The resulting $(A^TX^TXA)^{-1}$ can be expressed as
$(A^TX^TXA)^{-1} = V(S^{-1})V^T$ where \\
$S^{-1} = \mathop{\rm diag}(1/\sigma_1, 1/\sigma_2, ..., 1/\sigma_k)$. Because all $\sigma$ are positive, we denote $\delta_i^2 = \frac{1}{\sigma_i}$. Hence $\Delta = \mathop{\rm diag}(\delta_1, ..., \delta_k)$. And
\begin{align}
M = AV\Delta \Delta^T V^TA^T.
\end{align}
For any $v \in \mathbb{R}^p$, we have
\begin{align}
v^T M v = v^T(\Delta^T V^TA^T)^T \Delta^T V^TA^T v \\
= \|\Delta^T V^TA^T v \|_2^2 \geq 0 .
\end{align}
Thus $M$ is positive semidefinite.
\subsection{Convexity of a re-parameterized problem} \label{pes:priv:app:linear:tran:M:cvx_prob1}
Claim: The following optimization is convex:
\begin{align}
& \min_{M} \mathop{\rm Tr}\big(C_{xy}^TX {M}X^TC_{xy}\big) + \beta \| {M}\|_{*} \\
\text{s.t.} & \quad \|X{M}X^TX - X\|_{F}^2 \leq \gamma .
\end{align}
\proof It is easy to see that the first term $\mathop{\rm Tr}\big(C_{xy}^TX {M}X^TC_{xy}\big)$ is convex, since $M$ is positive semidefinite, trace operator is linear with respect to $M$. The second term and third term are also convex. For any norm, given $0 \leq \alpha \leq 1$ and two matrices $A, B$, we have \vspace*{-0.11in}
\par{\setlength{\abovedisplayskip}{0pt} \small{
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{\abovedisplayskip}
\setlength{\abovedisplayshortskip}{0pt}
\setlength{\belowdisplayshortskip}{1pt}
\begin{align*}
\|\alpha A + (1-\alpha) B\| & \leq \|\alpha A\| + \|(1-\alpha) B\| = \alpha\|A\| + (1-\alpha) \|B\|
\end{align*}
}}
Hence Frobenius norm and Nuclear norm are specific forms of norm that is convex with respect to ${M}$. The first term in the objective is just linear in ${M}$. Thus, the problem is convex.
\subsection{Deriving linear compression with noise} \label{icml:appendix:linear:compress:noised:data}
Consider the case $\varepsilon \sim \mathcal{N}(0, \Sigma)$. We have the min-max game as follows
\begin{align}
& \max_{A, B, \Sigma}\min_{\Theta} \frac{1}{n}\sum_{i=1}^n \mathop{\mathbb{E}} _{\varepsilon \sim P_{\varepsilon} } \|\Theta^T(A^Tx_i + \varepsilon) - y_i \|_2^2 \\
& - \gamma \|XAB - X\|_{F}^2
\end{align}
For attacker, we have the following minimization problem
\begin{align}
& \min_{\Theta} \frac{1}{n}\sum_{i=1}^n \mathop{\mathbb{E}} _{\varepsilon \sim P_{\varepsilon} } \|\Theta^T(A^Tx_i + \varepsilon) - y_i \|_2^2 \\
& = \min_{\Theta} \frac{1}{n}\sum_{i=1}^n
\Big( (A^Tx_i + \varepsilon_i)^T \Theta \Theta^T (A^Tx_i + \varepsilon_i) \\
& - 2 (A^Tx_i + \varepsilon_i)^T \Theta y_i + y_iy_i^T \Big)
\end{align}
We first find the minimizer $\Theta$ for the attacker. By taking the derivative over $\Theta$, we have
\begin{align}
\frac{1}{n}\sum_{i=1}^n\Big( A^Tx_ix_i^TA + \varepsilon_i \varepsilon_i^T\Big) \Theta = A^T \frac{1}{n} \sum_{i=1}^n x_i y_i^T \\
\Theta = (A^TX^TXA + \Sigma)^{-1} A^T C_{xy}
\end{align}
Also we also find the best recover matrix $B$ by considering the following relation $\tilde{X} = XA + \mathbb{\boldsymbol{\varepsilon}}$
\begin{align}
& E_{\varepsilon \sim P_{\varepsilon}}\|\tilde{X} - X\|_{F}^2 \\
& = \mathop{\rm Tr}\Big((XAB + \boldsymbol\varepsilon B - X)(XAB + \boldsymbol\varepsilon B - X)^T\Big) \\
& = \mathop{\rm Tr} (XABB^TA^TX^T + B\Sigma B^T - 2XABX^T)
\end{align}
Taking the derivative over $B$ and set it equals 0, we have $B = (A^TX^TXA + \Sigma)^{-1}A^TX^TX$. Hence the data holder's maximization can be casted into
\begin{align}
& \min_{A, \Sigma} \mathop{\rm Tr}\Big(C_{xy}^TA(A^TX^TXA + \Sigma)^{-1}A^TC_{xy}\Big) \\
& + \|XA(A^TX^TXA + \Sigma)^{-1}A^TX^TX - X\|_F^2
\end{align}
It is not difficult to discover that $A(A^TX^TXA + \Sigma)^{-1}A^T$ is also positive semidefinite. Thus the problem can be relaxed to convex optimization.
\subsection{Proof of Theorem \ref{cdc:svd:err:th1}}
\label{cdc:appx:proof:Th1}
\proof Denote $X = \sum_{i=1}^p \lambda_i u_i v_i^T$, where $\lambda_i$ is singular value, $u_i, v_i$ are corresponding left and right singular vectors. The best rank-$k$ approximation $\tilde{X}_k = \sum_{i=1}^k \lambda_i u_i v_i^T$ is achieved by SVD in Frobenius norm by Eckart-Young theorem\cite{golub1987generalization}. Then $\|X-\tilde{X}_k\|_F^2 = \mathop{\rm Tr}\big( (\sum_{i=k+1}^p\lambda_i u_i v_i^T) (\sum_{i=k+1}^p\lambda_i u_i v_i^T)^T \big) = \mathop{\rm Tr}(\sum_{i=k+1}^p\lambda_i^2 ) = \sum_{i=k+1}^p \lambda_i^2$
\subsection{Features extracted from power data}
Features for power consumptions are displayed in Table~\ref{priv:num_sim:feature:table}.
\begin{table}[!hbpt]
\centering
\caption{Features extracted out of power consumptions}
\label{priv:num_sim:feature:table}
\begin{tabular}{|l|c|}
\hline
Index & Description \\
\hline
1 & Week total mean \\
2 & Weekday total mean \\
3 & Weekend total mean \\
4 & Day (6am - 10pm) total mean \\
5 & Evening (6pm - 10pm) total mean \\
6 & Morning (6am - 10am) total mean \\
7 & Noon (10am - 2pm) total mean\\
8 & Night (1am - 5am) total mean \\
9 & Week max power \\
10 & Week min power \\
11 & ratio of Mean over Max\\
12 & ratio of Min over Mean \\
13 & ratio of Morning over Noon \\
14 & ratio of Noon over Day \\
15 & ratio of Night over Day \\
16 & ratio of Weekday over Weekend \\
17 & proportion of time with $P_t > 0.5 kw$ \\
18 & proportion of time with $P_t > 1 kw$ \\
19 & proportion of time with $P_t > 2 kw$ \\
20 & sample variance of $P_t$ \\
21 & sum of difference $|P_t - P_{t-1}|$ \\
22 & sample cross correlation of subsequent days \\
23 & number of counts that $|P_t - P_{t-1}| > 0.2 kw$ \\
\hline
\end{tabular}
\end{table}
\subsection{Images}
\label{cdc:appx:original:sampled:imgs}
Original people images are shown in Figure~\ref{cdc:appx:original:sampled:group_ppl}
\begin{figure}[!hbpt]
\centerline{\includegraphics[width=0.95\columnwidth]{fig_faces_img/ex1_24samples.png}}
\caption{The sampled images with the dimension of $61 \times 49$ for each one. }
\label{cdc:appx:original:sampled:group_ppl}
\end{figure}
\subsection{Low rank linear transformation}
Singular value matrices for a batch of samples with the low rank transformation are shown in Figure~\ref{cdc:linear:singular:val:matrix}. And the raw and altered power consumption features are shown in Figure~\ref{priv:fig:linear:svd:multi_rank_cases:heatmap}
\begin{figure*}[!bpht]
\centering
\framebox{\parbox[b]{1.92in}
{\includegraphics[scale=0.26]{fig_linear/svd_low_rank_4_patch_0_heatmap.png}
}
{\includegraphics[scale=0.26]{fig_linear/svd_low_rank_10_patch_0_heatmap.png}
}
{\includegraphics[scale=0.26]{fig_linear/svd_low_rank_18_patch_0_heatmap.png}
}}
\caption{Singular values for a sampled data patch when the rank equals 4, 10, and 18}
\label{cdc:linear:singular:val:matrix}
\end{figure*}
\begin{figure}[!ht]
\vskip -0.12in
\framebox{\parbox[b]{1.49in}
{\includegraphics[scale=0.20]{fig_linear/original_data_samples_heatmap.png}}
\hspace{-0.01in}
{\includegraphics[scale=0.20]{fig_linear/altered_data_rank_4_samples_heatmap.png}
}}
\framebox{\parbox[b]{1.485in}
{\includegraphics[scale=0.204]{fig_linear/altered_data_rank_10_samples_heatmap.png}}\hspace{-0.01in}
{\includegraphics[scale=0.204]{fig_linear/altered_data_rank_18_samples_heatmap.png}
}}
\vspace{-0.10in}
\caption{The data holder applies a linear transformation on the data, i.e. compresses it down to low rank matrices and recovers it back to the original dimension. Each figure consists of 155 sampled households with 23 features which has been normalized originally. We fix the distortion tolerance $\gamma$ and tune the nuclear norm coefficient $\beta$ to get the result of various rank scenarios. }
\label{priv:fig:linear:svd:multi_rank_cases:heatmap}
\end{figure}
|
\section{Introduction}
Quantum field theory and General Relativity are two of the most successful physical theories ever developed.
Their combination in a unified body is however still problematic.
In spite of the many attempts, no universally accepted theory of quantum gravity is available in the literature. In the recent years many interesting features of an eventual theory of quantum gravity have been discovered by analysing the unification of the two theories in approximated regimes like semiclassical gravity.
In quantum field theory on curved spacetime, gravity is still described by a classical curved spacetime while
matter is modeled by quantum fields which propagate on that classical spacetime.
In first approximation, their backreaction can be taken into account by means of the semiclassical Einstein equation
\[
G_{ab}= 8\pi\langle T_{ab} \rangle_\omega
\]
which is formally similar to the ordinary Einstein equation where the stress tensor of the classical matter is substituted with the expectation value of the stress tensor of quantum matter in a suitable state $\omega$. This equation is meaningful only if the fluctuations of the stress tensor are small. Interesting and well known effects like Hawking radiation for black holes or particle creation in cosmology can be obtained in this approximation.
A physical scenario where these ideas can be tested is cosmology.
Due to the homogeneity and isotropy hypothesis applicable in this case, the manifold describing the spacetime is $M=I\times \Sigma$ where $I$ is a real interval representing time and $\Sigma$ is the three dimensional manifold of space.
The geometry is described by the Friedmann--Lema\^itre--Robertson--Walker (FLRW) metric
\begin{equation}\label{eq:FLRWmetric}
ds^2 = - dt^2 + a(t)^2 \left(\frac{1}{1-\kappa r^2} dr^2 + r^2d\theta^2 + r^2\sin\theta^2 d\varphi^2 \right)
\end{equation}
where $\kappa\in\{-1,0,1\}$ distinguishes between open, flat or closed spatial sections and the scale factor $a(t)$ describes the way in which the universe evolves with respect to cosmological time. At present our spacetime appears to be almost spatially flat, for this reason we shall consider the case $\kappa=0$ only. Due to the time-dependence present in the metric, concepts like temperature, energy, particles and vacuum states cannot be used -- one is forced to rethink all of them.
In order to fully understand the behavior of quantum matter in cosmological spacetimes, as part of the SFB Project we have analysed the construction of regular states and discussed some of their thermodynamical aspects~\cite{Dappiaggi:2007mx,Dappiaggi:2008dk,Dappiaggi:2010gt}. We have furthermore analysed the backreaction of quantum matter on curvature in the semiclassical approximation~\cite{Dappiaggi:2008mm,Dappiaggi:2009xj,Hack:2010iw}.
As a subsequent step, we have investigated the quantisation of the fluctuation of the matter-gravity system (Einstein--Klein--Gordon system) in the linear regime~\cite{Hack:2014epa} and beyond~\cite{Brunetti:2013maa}.
We have discussed the form of gauge-invariant observables associated to this system when the background is chosen to be an FLRW spacetime~\cite{Brunetti:2016hgw}.
Finally, more recently, we have rigorously constructed equilibrium states for perturbatively constructed interacting field theories~\cite{Lindner:2013ila,Fredenhagen:2013cna}.
The project has been carried out using methods of Algebraic Quantum Field Theory.
In this approach, quantisation is formulated in two steps. The first consists in the determination of the set of observables and of the algebraic relations among them. In particular, the commutation relations of the theory are implemented at this level. The second steps consists in the analysis of the form of physically relevant states of the system.
While the first step is well under control at least for the case of free theories, the choice of a physically relevant state is more difficult and cannot be done using local properties of the spacetime only. The problem of finding regular states also on cosmological spacetimes based on their global properties is not straightforward and was the first point of the analysis performed in this project.
In the next chapter, we briefly outline the obtained results regarding the construction of regular states based on asymptotic spacetime properties and on the local notion of vacuum states.
In the third chapter we review the results obtained in this project concerning the backreaction of quantum matter on the curvature induced by the semiclassical Einstein equation. The forth chapter contains a brief recollection of results obtained regarding the perturbative quantisation of the Einstein--Klein--Gordon system. Finally, we present the construction of equilibrium states for interacting field theory recently obtained in the framework of perturbative algebraic quantum field theory in the fifth chapter.
\section{Regular states on curved spacetimes}
In order to compute correlation functions on a curved spacetime and to analyse their thermal behavior it is necessary to have control on the state of the system.
In particular, expectation values of local fields and their fluctuations need to be finite.
This requirement constrains the possible quantum states and the class of Hadamard states satisfy this constraint~\cite{Kay:1988mu}. It is however not clear how to construct these states starting from initial conditions at some fixed time if the spacetime is not a-priori known.
In this section we discuss various methods developed within this project to construct regular states for a free quantum scalar field $\phi$ on curved backgrounds $(M,g)$ which are globally hyperbolic spacetimes~\cite{Wald:1984rg}. Similar results hold for more general types of fields~\cite{Hack:2010iw}.
The field we have in mind satisfies the equation
\[
P\phi := -\Box \phi + \xi R \phi + m^2\phi = 0
\]
where $\Box=\nabla_a\nabla^a$ is the d'Alembert operator, $\xi$ is the coupling to the scalar curvature $R$ and
$m$ is the mass of the field. The $*-$algebra of observables $\mathcal{A}$ associated to this field is generated by ``smeared fields'' $\phi(f)$ where $f\in\mathcal{D}(M)$, the set of compactly supported smooth functions. The product is such that it satisfies the canonical commutation relations
\[
[\phi(f), \phi(h)] = i\Delta (f,h)
\]
where $\Delta=\Delta_R-\Delta_A$ is the causal propagator of the theory which is equal to the retarded-minus-advanced fundamental solution of $P$, both of which are characterised by $P\Delta_{R/A}(f)=f$, $\text{supp} \Delta_{R/A}(f) \subset J^{+/-}(\text{supp} f)$. We recall that the fundamental solutions exist and are unique on globally hyperbolic spacetimes, see e.g.~\cite{2008arXiv0806.1036B}. Since the field is uncharged, the involution $*$ coincides with the complex conjugation
\[
\phi(f)^* = \phi(\overline{f}).
\]
The algebra $\mathcal{A}$ is called on-shell if $\phi(Pf)=0$ for every $f\in\mathcal{D}(M)$. If this is not the case, the algebra is said to be off-shell. In the latter case, the on-shell projection can be realised by restricting the physically relevant states of the system to those which satisfy the equation of motion
\[
\omega(\phi(f_1)\dots \phi(Pg) \dots \phi(f_n))=0.
\]
In the case of a free scalar field, a state $\omega$ is characterised by its $n-$point functions
which are elements of $\mathcal{D}'(M^n)$, the set of distributions on smooth compactly supported functions, defined as
\[
\omega_n(x_1,\dots, x_n):=\omega(\phi(x_1) \dots \phi(x_n)).
\]
For Gaussian states, also called quasi-free states, the $n-$point functions with odd $n$ vanish while those with even $n$ are obtained from the two-point function as
\[
\omega_{2n}(x_1,\dots, x_{2n}) = \sum_{i\in P_{2n}}\omega_2(x_{i_1},x_{i_2})\dots \omega_2(x_{i_{2n-1}},x_{i_{2n}})
\]
where the sum is taken over all possible ordered sequences $P_{2n}$ of $\{1,\dots,2n\}$. A sequence $i$ is said to be ordered and thus contained in $P_n$ if $i_{2k-1}<i_{2k}$ and $i_{2k'-3}<i_{2k'-1}$ for $k\in \{1,\dots,n\}$ and $k'\in \{2,\dots,n\}$.
Here we shall concentrate our attention on Gaussian states and thus we have to care only about the two-point function
\[
\omega_2(x_1,x_2) = \omega(\phi(x_1)\phi(x_2)).
\]
As said before, the states we are considering need to be sufficiently regular to admit an extension to the algebra of local Wick polynomials which are pointlike products of fields and are realised extending the test functions $(f_1,\dots, f_n)\in \prod C^\infty_0(M)$ to distributions supported on the diagonal in $M^n$. Furthermore, the fluctuations of these fields need to be finite. In order to have these properties under control a sufficient condition is to require the state to be of Hadamard form.
A state of a free scalar field theory is of Hadamard form if on every normal neighborhood its two-point function can be expanded in the following way
\begin{equation}\label{eq:Hadamard}
\omega_2(x_1,x_2) = \lim_{\epsilon\to 0}\frac{U(x_1,x_2)}{\sigma_\epsilon(x_1,x_2)} + V(x_1,x_2)\log\sigma_\epsilon(x_1,x_2) + W(x_1,x_2)
\end{equation}
where $U,V$ are smooth functions on $M^2$ which are universal and depend only on the geometry and on the parameters present in the equation of motion, while $W$ is a smooth function which characterises the state. Furthermore, $\sigma_\epsilon(x_1,x_2) = \sigma(x_1,x_2) +i\epsilon(T(x_1)-T(x_2))$
and $\sigma$ is one half of the square of the geodesic distance taken with sign while $T$ is any time function.
It is important to notice that Hadamard states are equivalently characterised by their wave front set. The wave front set of a distribution encodes its singular structure.
In particular, it is a subset of the cotangent space whose projection on the spacetime gives the singular support of the distribution while the points in the fiber determine the directions of non-rapid decrease of the Fourier transform of the distribution localised in any neighborhood of the corresponding base point. The precise definition can be found in the book of H\"ormander and it is used in the celebrated H\"ormander criterion of multiplication of distributions.
The Hadamard state satisfies the ``microlocal spectrum condition'', namely, its
wave front set is such that
\begin{equation}\label{eq:muc}
WF(\omega_2) = \{ (x_1,x_2;k_1,k_2)\in T^*{M^2}\setminus \{0\} | (x_1,k_1)\sim(x_2,-k_2), k_1\triangleright 0 \}
\end{equation}
where $(x_1,k_1)\sim(x_2,-k_2)$ if $x_1$ and $x_2$ are joined by a null geodesic and $g^{-1}k_1$ and $-g^{-1}k_2$ are the corresponding tangent vector with respect to an affine parametrisation. Furthermore, $k_1\triangleright 0$ if $k_1$ is future directed.
Radzikowski proved in~\cite{Radzikowski:1996pa} that a state is of Hadamard form if and only if it satisfies the microlocal spectrum condition. A generalisation of the microlocal spectrum condition to higher order Wick polynomials can be found in~\cite{Brunetti:1995rf}.
The microlocal spectrum condition can be understood as a covariant remnant of the spectrum condition usually employed in the case of vacuum states for quantum theories on flat spacetimes. It is a remarkable fact that this covariant remnant fixes the singular structure of the two-point function.
The characterisation of Hadamard states by means of the wave front set of the two-point function is quite useful. It is actually much easier to check if a state satisfies the microlocal spectrum condition, then to precisely control the singular structure of its two-point function. Standard results of microlocal analysis like H\"ormander's propagation of singularity theorems can in fact be successfully employed.
\subsection{States from asymptotic properties}\label{se:asymptotic}
As mentioned above it is not straightforward to get Hadamard states out of initial values if the spacetime is not known a priori.
The problem is simpler when the initial values are given on null surface (characteristic initial value problem). In fact, in this situation the positive frequency condition present in the microlocal spectrum condition can be tested directly for the initial values.
This is the basic idea at the heart of the construction of Hadamard states based on asymptotic initial values when the asymptotic past infinity (or more generally the asymptotic past boundary) is a causal cone. In the following we discuss these basic ideas in a more rigorous form.
Consider a spacetime $M$ whose past boundary is a null cone $\mathcal{C}$ in another (conformally related) larger spacetime $\tilde{M}$. Due to the well-posedness of the Goursat problem, it is known that giving initial values on $\mathcal{C}$ fixes uniquely a solution of any hyperbolic equation in $M$.
As an application of the time slice axiom~\cite{2009CMaPh.287..513C}, we may thus construct a map from the on-shell algebra $\mathcal{A}(M) \to \mathcal{A}(C)$ which is an injective $*-$homomorphism. This is essentially done at the level of generators of the algebra $\mathcal{A}(M)$ and
\[
\iota \phi(f) = \Psi(f_{\mathcal{C}}), \qquad f_{\mathcal{C}} = \left. \Delta (f) \right|_{\mathcal{C}}
\]
where the causal propagator $\Delta$ is seen in $\tilde{M}$ and $\Psi(f)$ are the generators of $\mathcal{A}(C)$.
On $C$ there is a notion of causality, it is thus possible to construct states which look like a vacuum, meaning that their spectrum contains only positive frequencies with respect to causal directions on $\mathcal{C}$.
If we indicate such a state by $\Omega^+$, it turns out that its pullback
\[
\omega(A) = \iota^*\Omega^+(A) = \Omega^+(\iota(A))
\]
defines a state in $M$ which is an Hadamard state.
This idea has been used in various physical situation in order to determine the singular structure of known states:
\begin{itemize}
\item in the case of asymptotically flat spacetime to check the Hadamard property of asymptotically vacuum states for massless particles~\cite{Dappiaggi:2005ci};
\item in the case of asymptotically de Sitter spacetimes~\cite{Dappiaggi:2007mx,Dappiaggi:2008dk} to discuss the Hadamard properties of Bunch--Davies-like states;
\item in the case of spacetimes possessing a null big bang scenario to control asymptotically vacuum states~\cite{Pinamonti:2010is};
\item in the case of Schwarzschild spacetime to prove that the Unruh state is of Hadamard form~\cite{Dappiaggi:2009fx}.
\end{itemize}
For a cosmological spacetime $(M,g)$ with $g$ as in \eqref{eq:FLRWmetric}, pure Gaussian states which are invariant under the spatial isometries of the metric have a two-point function of the form:
\begin{equation}\label{eq:two--point-function}
\omega_2(x_1,x_2) := \lim_{\epsilon\to 0^+}\frac{1}{8\pi^3a(t_1)a(t_2)} \int_{\mathbb{R}^3} \chi_k(t_1)\overline{\chi_k(t_2)} e^{i{\mathbf{k}}\cdot (\mathbf{x}_1-\mathbf{x}_2)} e^{- k \epsilon} d\mathbf{k}
\end{equation}
whereby $k = |\mathbf{k}|$ and the mode functions $\chi_k(\tau)$ satisfy the equation
\[
\chi_k''+(m^2 a^2 +k^2)\chi_k + \left(6\xi-1\right) \frac{a''}{a}\chi_k =0,
\]
where $'$ denotes the derivative with respect to conformal time $\tau$ obtained by integrating the relation $dt = a d\tau$, and the modes are normalised such that the following Wronskian condition
\[
\overline{\chi_k}'\chi_k -
\overline{\chi_k}\chi_k' =i \;
\]
holds. The latter condition ensures that the antisymmetric part of $\omega_2$ is proportional to the causal propagator of the theory~\cite{Lueders:1990np}.
As discussed above, it is not easy to find the modes $\chi_k$ for which $\omega_2$ as in \eqref{eq:two--point-function} correspond to an Hadamard state.
The states which are asymptotic vacuum states constructed with the pullback of $\iota$ discussed above, are usually quasi-free states of the from \eqref{eq:two--point-function} and the modes have the asymptotic form
\[
\chi_k(\tau) \underset{{\tau\to-\infty}}{\longrightarrow} \frac{e^{ik\tau}}{\sqrt{2k}}.
\]
The construction of asymptotic vacuum states can be generalised to construct approximate KMS states. In particular, if the boundary state is a KMS state with respect to the characteristic translation, the obtained state in the bulk can be interpreted as an approximate KMS state and its two-point function is
\begin{equation}\label{eq_genth}
\omega_2(x_1,x_2) := \lim_{\epsilon\to 0^+}\frac{1}{8\pi^3a(t_1)a(t_2)}
\int_{\mathbb{R}^3} \left(\frac{\overline{\chi_k(t_1)}\chi_k(t_2)}{1-e^{-\beta k_0}}+\frac{\chi_k(t_1)\overline{\chi_k(t_2)}}{e^{\beta k_0}-1}\right)e^{i{\mathbf{k}}\cdot ({\mathbf{x}}_1-{\mathbf{x}}_2)} e^{- k \epsilon} d\mathbf{k}
\end{equation}
with $k_0:=|\mathbf{k}|$. Other approximate KMS states can now be constructed considering
$k_0:=\sqrt{k^2+m^2a^2_F}$ for a fixed constant $a_F$. Further comments on this can be found in~\cite{Dappiaggi:2010gt}.
\subsection{States of low energy}
Olbermann proposed in~\cite{Olbermann:2007gn} a way to select the mode functions $\chi_k$ in \eqref{eq:two--point-function} in order to obtain a state which satisfies the microlocal spectrum condition. The idea
is to find the modes which minimise the energy density smeared in a chosen smooth way along a fixed time interval.
For this reason these states are called states of low energy. Although they depend on the chosen smooth smearing function they can be interpreted as approximate vacuum states. Excitations of these states are natural candidates for approximate particle states, and first results have already been obtained in~\cite{Degner:2009rq}. An interesting observation is that the number of particles is
oscillating during the cosmological evolution. Whether this is an artefact of
the approximation or a genuine physical effect has to be clarified.
Within this project, a further extension of these ideas to states which are locally in equilibrium has been discussed in the PhD thesis of K\"usk\"u~\cite{Kusku:2008zz}.
\subsection{Sorkin--Johnston states and generalisations thereof}
On a static spacetime, the two-point function of the vacuum state of the theory can be obtained from the spectral decomposition of the causal propagator and its restriction to the positive part of the spectrum.
A local version of this construction was proposed by Sorkin--Johnston, see~\cite{Afshordi:2012jf} and reference therein. The basic difference between their idea and the above is to construct a local vacuum starting from the spectral decomposition of the causal propagator localised in some double cone region. Unfortunately, in general, the states obtained in this way do not satisfy the microlocal spectrum condition \cite{Fewster:2012ew}.
Within the present project, Brum and Fredenhagen in~\cite{Brum:2013bia} showed that a smoothly smeared version of the Sorkin--Johnston construction results in a state which is of Hadamard form.
To construct these states, one mollifies the sharp localisation of the causal propagator to a double cone $N\subset M$ by choosing a real-valued $f\in \mathcal{D}(M)$ such that $f$ is equal to $1$ on $N$. This function is then used to construct the operator
\[
A:=if \Delta f
\]
which can be extended to act on the Hilbert space $L^2(M, d\mu_g)$. The operator obtained in this way is, at least on static and on FLRW spacetimes, bounded and self-adjoint, hence, by means of standard functional calculus, it is possible to extract its positive part
\[
A^+ := \Pi^+A
\]
where $\Pi^+$ is the spectral projection on the interval $[0,|A|]$.
The generalised Sorkin--Johnston state is now obtained by considering the quasifree state characterised by the two-point function associated to $A^+$:
\[
\omega_2(h_1,h_2):=\left(h_1,A^+ h_2\right), \qquad h_i\in\mathcal{D}(N).
\]
The state obtained in this way turns out to be a Hadamard state.
\section{Semiclassical gravity}
At first approximation, the influence of quantum matter on the curvature can be taken into account by means of the semiclassical Einstein equation
\[
G_{ab} = \langle T_{ab} \rangle_\omega
\]
whereby gravity is described by a classical curved spacetime. This equation is formally similar to the classical Einstein equation where the stress tensor of the classical matter is substituted with the expectation value of a the stress-tensor of some quantum field propagating on a curved background.
In the cosmological case, if the conservation equation of the stress tensor
\[
\nabla^a \langle T_{ab} \rangle_\omega = 0
\]
holds, the semiclassical Einstein equation is equivalent to
\[
-{R_a}^a = 8\pi \langle {T_a}^a \rangle_\omega
\]
up to an initial condition which fixes the amount of radiation present in the solution.
In the expectation value $T=\langle{T_a}^a \rangle_\omega$ of the trace of the stress tensor we can recognize three contributions
\[
-{R_a}^a = 8\pi \left( T_{state} + T_{anomalous}+ T_{ren.freed.} \right)
\]
corresponding to a state dependent part, an anomalous contribution due to the requirement of covariant conservation, see e.g.~\cite{PhysRevD.17.1477,Moretti:2001qh,Hollands:2004yh}, and a contribution which can be interpreted as renormalisation freedom~\cite{Hollands:2001nf}.
In the case of $\xi=1/6$, the anomalous contribution is
\[
T_{anomalous} = \frac{1}{8\pi^2} \left( \frac{1}{360} \left( C_{abcd}C^{abcd}+R_{ab}R^{ab}-\frac{R^2}{3}+\Box R \right) + \frac{m^4}{4} \right)
\]
with the Weyl tensor $C_{abcd}$.
The renormalisation freedom is
\[
T_{ren.freed.} = \alpha m^4 +\beta m^2 R +\gamma \Box R,
\]
whereas the state dependent contribution is
\[
T_{state} = \frac{1}{8\pi^2} m^2 [W]
\]
with $[W]$ indicating the coinciding point limit of the smooth part of the state's two-point function \eqref{eq:Hadamard}.
With the choice $\gamma=0$ and a suitable assumption on the state dependent contribution (considering it to be vanishing like in zeroth-order adiabatic states) and in the case of massive, conformally coupled fields, the equation for the trace is an ordinary differential equation
\begin{equation}\label{eq:diff}
\dot{H}(H^2-H_0^2)=-H^4+2H^2H_0^2+M;
\end{equation}
where $H=\dot{a}/a$ is the Hubble parameter and $\dot{\bullet}$ denotes a derivative with respect to cosmological time. The constants
\[
H_0^2 = 180\pi\left(\frac{1}{G} - 8\pi m^2\beta\right),
\qquad
M = \frac{15}{2}m^4-240\pi m^4 \alpha
\]
depend on the renormalisation freedom. $\beta$ can be interpreted as a renormalisation of the Newton constant and for this reason it can be fixed to be $0$, $\alpha$ represents a renormalisation of the cosmological constant and can be set to model the present value of the measured constant.
The solution space of the differential equation \eqref{eq:diff} can be easily obtained. We notice in particular that there are two stable solutions, the first of which could describe the phase of rapid expansion present in the beginning of the universe, whereas the second one can model the present universe and its recent past.
However, in this simplified model there is no mechanism which permits to exit the phase of rapid expansion. The oversimplifications present in the model are the following: the considered matter field is a free field, the role of higher derivatives which are present in the equation for a different choice of $\gamma$ could have an important role and the state dependent contribution is not taken into account properly.
The results presented in this section are summarised in the paper~\cite{Dappiaggi:2008mm}. Within this project it has been shown that similar results hold also when the matter is described by other fields like Dirac fields~\cite{Hack:2010iw}. In particular, the trace anomaly, the form of stress tensor and the extended algebra of fields for Dirac fields has been analyzed in~\cite{Dappiaggi:2009xj}.
\subsection{Rigorous analysis of the semiclassical equation}
The state dependent contribution present in the trace makes the dynamical problem much more complicated, essentially because it is difficult to fix the state with a universal prescription that does not depend on the particular spacetime.
This problem has been rigorously studied in two cases. The first case considered were cosmological spacetimes which possess an initial null big bang phase. In fact, when the past boundary of $M$ is a null surface, it is possible to use similar techniques like the ones discussed above in section \ref{se:asymptotic}. In this way it has been proved in
\cite{Pinamonti:2010is} that the semiclassical Einstein equation leads to a well-posed initial value problem. This means that solutions exists and are unique, at least for small amounts of time after the null big bang.
Relaxing the request of having full Hadamard states, but considering adiabatic states of order $0$ as given in~\cite{Lueders:1990np} at a fixed time $t_0$, it was possible to extend the proof of existence also to this situation. The corresponding results have been summarised in~\cite{Pinamonti:2013wya}, where it is proved that solutions of the semiclassical Einstein equation exist and are unique up to future infinity or up to the occurrence of a future blowup of the scale factor (big rip).
\section{Quantum gravity and cosmological fluctuations}
To go beyond the semiclassical approximation presented above, and to quantise the degrees of freedom of both matter and gravity fields, we have analyzed the quantisation of the fluctuations of the coupled Einstein--Klein--Gordon system, namely, a massive minimally coupled scalar field $\phi$ propagating over a Lorentzian spacetime $(M,g)$ for which
\begin{equation}\label{eq:EKG}
R_{ab}-\frac{1}{2}R g_{ab} = 8\pi T_{ab} ,\qquad - \Box \phi + m^2 \phi + V'(\phi)= 0,
\end{equation}
where $T_{ab}$ are the components of the stress tensor of $\phi$.
To discuss the linearised perturbation of the theory we consider a field configuration $\Gamma=(g,\phi)$ and expand it around a background solution $\Gamma_0$
\[
\Gamma = (g,\phi) = \Gamma_0 + \lambda (\gamma,\varphi) + O(\lambda^2),
\]
equations for $(\gamma,\varphi)$ can now be obtained from \eqref{eq:EKG}.
In the case of perturbations in inflationary cosmological models, the background theory corresponds to an FLRW spacetime plus an inflaton $\phi$ which is constant in space.
The linearised perturbations $(\gamma,\varphi)$ can be quantised using standard methods of algebraic quantum field theory on curved background in a manifestly covariant and gauge-invariant manner. Details of this construction are presented in the paper~\cite{Hack:2014epa}. In particular, a non-local behavior shown by the quantum Bardeen potentials is obtained and it is furthermore shown that a similar effect is present in any local quantum field theory. Finally the known linearised equation for the Mukhanov--Sasaki variable is recovered in that formalism.
\subsection{Beyond the linearised approximation}
The perturbative construction of quantum gravity in a generally covariant way has been presented in~\cite{Brunetti:2013maa}. This construction has been performed in the framework of locally covariant quantum field theory and the gauge problem has been treated by means of the renormalised Batalin--Vilkovisky formalism~\cite{Rejzner:2011jsa}.
The problem of non-renormalisability of the theory persists and, hence, the theory is interpreted as an effective theory at large scales. In spite of this fact, the construction leads to a background-independent theory. Moreover, gauge-invariant observables are considered to be relational observables. This means that they are functionals of four scalar coordinate fields constructed which can be chosen as suitable curvature scalars, see also~\cite{Khavkine:2015fwa} for related work on these functionals. This idea can be used to construct gauge-invariant observables in cosmology in a simple way.
In the case of the Einstein--Klein--Gordon system the procedure works as follows. One selects 4 scalar fields $X_{\Gamma}^a, a=1,\ldots 4$, which are local functionals of the field configuration $\Gamma$ which contains the spacetime metric $g$, the inflaton field $\phi$ and possibly other fields. The fields $X_{\Gamma}^a$ are supposed to transform under diffeomorphisms $\chi$ as
\begin{equation}\label{equivariance}
X_{\chi^*\Gamma}^a=X_{\Gamma}^a\circ\chi\,.
\end{equation}
We choose a background $\Gamma_0$ such that the map
\begin{equation}
X_{\Gamma_0}:x\mapsto (X_{\Gamma_0}^1,\ldots,X_{\Gamma_0}^4)
\end{equation}
is injective. We then set for $\Gamma$ sufficiently near to $\Gamma_0$
\begin{equation}
\alpha_{\Gamma}=X_{\Gamma}^{-1}\circ X_{\Gamma_0}.
\end{equation}
We observe that $\alpha_{\Gamma}$ transforms under diffeomorphisms as
\begin{equation}
\alpha_{\chi^*\Gamma}=\chi^{-1}\circ\alpha_{\Gamma}\,.
\end{equation}
Let now $A_{\Gamma}$ be any other scalar field which is a local functional of $\Gamma$ and transforms under diffeomorphisms as in \eqref{equivariance}. Then the field
\begin{equation}
\mathcal {A}_{\Gamma}:=A_{\Gamma}\circ\alpha_{\Gamma}
\end{equation}
is invariant under diffeomorphisms and may be considered as a local observable. Note that invariance is obtained by shifting the argument of the field in a way which depends on the configuration.
In~\cite{Brunetti:2016hgw} this idea is used to show that the standard theory of cosmological perturbations arises as the linear order of a fully quantised perturbative theory of quantum gravity. In particular, in the way explained above, it was possible to obtain gauge-invariant local observables which can be used to generate perturbative invariant expressions at arbitrary order.
The standard gauge-invariant observables of cosmological perturbation theory were obtained at first order and explicit expressions at second order were explicitly given. Due to the large degree of symmetry present in the case of cosmological backgrounds, the spatial coordinate fields have been selected to be harmonic coordinates. While this choice works perfectly at the classical level, the gauge invariant observables constructed with them display severe non-localities, which could create problems in the
quantisation of the system. Further investigations are needed in order to clarify these points. The ideas of these project have been recently developed further in~\cite{Frob:2017lnt}.
\section{Perturbative algebraic quantum field theory and equilibrium states}
Whenever an interacting theory is treated perturbatively,
its elements are usually given as a formal series in the coupling constant.
Three major problems arises in these kind of treatments:
\begin{itemize}
\item there are ultraviolet problems at each order in perturbation theories;
\item there are infrared problems when the interaction Lagrangian is supported in a non compact domain;
\item the perturbative series does not converge.
\end{itemize}
The first problem is nowadays well understood. The second problem can be treated analysing the adiabatic limits of the theory. Regarding the third problem, very little is known. In some integrable cases, usually in lower dimensional theories, it is nevertheless possible to compare perturbatively obtained results with exact ones. As an example
the Sine--Gordon model can be treated both exactly and perturbatively. In~\cite{Bahns:2016zqj} it is shown that the perturbative methods agrees with the exact treatment. Further developments can be found in~\cite{Bahns:2017mrt}.
During the development of the present project, methods of algebraic quantum field theory have been developed to treat non-linear theories perturbatively. In particular, the framework of perturbative algebraic quantum field theory (pAQFT) was developed in~\cite{Brunetti:2009qc}.
Using this framework, it was possible to clarify the connections between Wilson's concept of the renormalisation group
and the original concept of St\"uckelberg and Petermann. Furthermore, it was recently possible to construct rigorously equilibrium states for interacting theories in the adiabatic limit in~\cite{Lindner:2013ila} and in~\cite{Fredenhagen:2013cna}. Central in this construction was the use of the time slice axiom proved in~\cite{2009CMaPh.287..513C}.
\subsection{Perturbative algebraic quantum field theory}
With the aim of discussing the construction of equilibrium (thermal) states for interacting theories,
we briefly recall the main steps of pAQFT in the simple case of an interacting scalar field theory.
The starting point is a concrete realisation of the observables as functionals of field configurations
\[
\phi\in C^\infty(M;\mathbb{R}).
\]
In order to be able to work with them we require that the observables are smooth local functionals, namely, their functional derivatives exist to all orders and at each order they are described by compactly supported distributions.
Furthermore, the singular structure of these distributions must be under control.
Hence, the observables of the theory are represented by the set of microcausal functionals
\[
\mathcal{F}_{\mu c} = \left\{ F:\mathcal{C}\to\mathbb{C} \left| \forall n \; F^{(n)}\in\mathcal{E}'(M^n) , \text{WF}({F^{(n)}})\cap (\overline{V}_+^n\cup\overline{V}_+^n) =\emptyset \right.\right\}.
\]
Among all possible elements of $\mathcal{F}_{\mu c}$ the regular ones are those whose functional derivatives are smooth functions. They correspond to the elements of the algebra generated by linear fields smeared by smooth test functions.
The set of regular functionals is denoted by $\mathcal{F}_{reg}$ and it does not contain Wick polynomials.
The local functionals are those whose functional derivatives are supported on the diagonals, they are denoted by $\mathcal{F}_{loc}$ and the Wick polynomials are contained in this set.
The free algebra of observables can be constructed equipping microcausal functionals with the following product
\[
F \star_\omega G = \mathcal{M} \exp{\left(\frac{1}{2} \Gamma_\omega \right)} F \otimes G
\]
where $\mathcal{M}$ denotes the pointwise multiplication $\mathcal{M}(F\otimes G) = FG$ and
\[
\Gamma_\omega = \int \omega_2(x,y) \frac{\delta}{\delta \phi(x)}\otimes\frac{\delta}{\delta \phi(y)}
\]
where $\omega_2$ is any two-point function of Hadamard form of the chosen free (linear) quantum theory. A two-point function of Hadamard form solves in particular the free equation of motion up to smooth functions, its antisymmetric part is proportional to the causal propagator of the free theory and its wave front set satisfies the microlocal spectrum condition \eqref{eq:muc}.
The product defined above is well-posed among functionals which possess only a finite number of non-vanishing functional derivatives. It can however be used also among more general functionals if a suitable weaker notion of convergence is considered. This is for example the case if observables are formal power series whose coefficients are elements of $\mathcal{F}_{\mu c}$.
When equipped with the involution defined by means of the complex conjugation as $F^*(\phi) = \overline {F(\phi)}$ the $(\mathcal{F}_{\mu c},\star_\omega,*)$ represent the $*-$algebra of the Wick polynomial normal ordered with respect of $\omega$. These are the observables of a free theory.
The canonical commutation relations are encoded in the form of the product $\star_\omega$, in fact, if one considers linear fields
\[
F_f(\phi)=\int f \phi \,d\mu_g, \qquad f\in \mathcal{D}(M)
\]
it holds that
\[
[F_f,F_h]_{\star_\omega} = F_f\star_\omega F_h - F_h\star_\omega F_h = \omega(f,h) - \omega(h,f) = i\Delta(f,h)
\]
where $\Delta$ is the causal propagator of the free theory, namely, the advanced-minus-retarded fundamental solution. These commutation relations do not depend on the particular form of the chosen Hadamard two-point function $\omega_2$.
Even more, giving two Hadamard functions $H,H'$, the map
\[
\alpha_{H-H'}: (\mathcal{F}_{\mu c},\star_{H'},*) \to(\mathcal{F}_{\mu c},\star_{H},*)
\]
defined by
\[
\alpha_{w} F = \exp \frac{1}{2}\Gamma'_{w} F , \qquad \Gamma'_w = \int w(x,y) \frac{\delta^2}{\delta \phi(x)\phi(y)} d\mu_g(x)d\mu_g(y)
\]
is actually a $*-$isomorphism between the two algebras.
If one restricts attention to the regular functionals, one obtains the well-known Borchers--Uhlmann algebra of fields. We stress that the algebra obtained here is off-shell, this is an essential step in order to be able to treat interactions perturbatively.
The observables associated to interacting field theories can be constructed by means of perturbation theory. To this end one has to consider the time-ordered product.
This product satisfies certain requirements, a corresponding list of axioms can be found in~\cite{Hollands:2001nf}.
The existence of time-ordered products of local Wick polynomials on curved spacetimes has been established in~\cite{Brunetti:1999jn} and a covariant generalisation of this construction can be found in~\cite{Hollands:2001fb}.
These constructive methods are generalisations of the inductive construction of Epstein--Glaser~\cite{Epstein1973}, according to which, at each inductive step $n$,
the causal factorisation property permits to construct the time-ordered product of $n+1$ local fields up to the total diagonal in terms of the time-ordered products of $n$ local fields.
The extension to the total diagonal of the obtained distributions can be done by means of Steinmann-scaling-degree~\cite{steinmann2014perturbation} techniques.
The obtained time-ordered product is a map from multilocal functionals to microcausal functionals
\[
T:\mathcal{F}_{\text{loc}}^{\otimes n} \to \mathcal{F}_{\mu c}
\]
and can be used to construct time-ordered exponentials (local $S$-matrices) of a local functional $V\in\mathcal{F}_{\text{loc}}$
\[
S(V)= \text{exp}_T (V).
\]
This time-ordered exponential is used to represent the interacting local fields in the free algebra by means of the Bogoliubov-map
\[
R_V(F)
:= \left.\frac{d}{d\lambda}S(V)^{-1}\star S(V+\lambda F) \right|_{\lambda=0}.
\]
The causal factorisation property at the level of $S$-matrices is such that for every $F,H,G\in\mathcal{F}_{\text{loc}}$ with $F$ later than $G$
\[
S(F+G+H) = S(F+G) \star S(G)^{-1} \star S(G+H),
\]
whereby $F$ is said to be later then $G$ if $\text{supp} F \cap J^-(\text{supp} G) =\emptyset$.
The interacting algebra is considered to be the algebra generated by the images of local fields under the Bogoliubov map
\[
\mathcal{F}_I\subset \mathcal{F}_{\mu c}.
\]
States for the interacting algebra are characterised by their expectation values on $\mathcal{F}_I$. Hence, every state of the free theory can be promoted to a state for the interacting theory by simply composing it with the Bogoliubov-map (or, rather, products of Bogoliubov-maps applied to the generators of the interacting algebra). However, the physical meaning of a given state changes. In particular, in the case of a Minkowski spacetime, the free vacuum and equilibrium states of the theory are not mapped to vacuum and equilibrium states of the interacting theory.
However, in the adiabatic limit, the vacuum of the interacting theory can be shown to be equivalent to the
free vacuum in this sense, while this is not the case for equilibrium states at finite temperature.
As part of this project, it has been proven that the {\bf time slice axiom} holds both for free and for interacting theories~\cite{2009CMaPh.287..513C}.
This result says that if a state is fixed in an $\epsilon-$neighborhood of a Cauchy surface in a globally hyperbolic spacetime, then it is fixed everywhere.
This observation implies that, if we want to define a state for an interacting field theory, it is sufficient to restrict the attention to field observables supported in an $\epsilon-$neighborhood of a Cauchy surface $\Sigma$. Denoting by $T$ the time function associated to the Cauchy surface $\Sigma$, such that the various Cauchy surfaces are the level sets of $T$, we have
\[
\Sigma_\epsilon := \left\{ p \in M \left |\; |T(p)-T(p_0)|\leq \epsilon , p_0\in\Sigma\right.\right\}.
\]
Thanks to the time slice axiom, to fix a state, it is sufficient to prescribe the expectation values of elements of
\[
\mathcal{F}_{\mu c}(\Sigma_\epsilon) = \{ F\in\mathcal{F}_{\mu c}| \text{supp} F\subset \Sigma_\epsilon \}
\]
for free field theories, or of the subalgebra $\mathcal{F}_{I}(\Sigma_\epsilon)\subset \mathcal{F}_{\mu c}$ generated by $\{R_V(F)|F\in \mathcal{F}_{\text{loc}}(\Sigma_\epsilon)\}$
in the case of interacting theories.
The time slice axiom permits us to restrict the domain of interacting observables. However, although the domain of an interacting field $R_V(F)$ coincides with the domain of $F$, $\text{supp}R_V(F)$ is larger than $\text{supp} F$. In fact, causality implies that
\[
\text{supp}R_V(F) \subset J^{-}(\text{supp} F) \cap J^{+}(\text{supp} V).
\]
Hence, in order to localize the elements of $\mathcal{F}_{I}(\Sigma_\epsilon)$, we need to control the support of $V$.
This last step can be done using the causal factorisation property. For every $F$ supported in $\Sigma_\epsilon$, we have that
\[
R_V(F) = R_{V'}(F)
\]
if $V-V'$ is supported in the future of $\Sigma_\epsilon$ and that
\[
R_V(F) = U^{-1} \star R_{V'}(F) \star U
\]
if $V-V'$ is supported in the past of $\Sigma_\epsilon$, where
\[
U = S^{-1}(V')\star S(V)\,.
\]
If $V$ and $V'$ are local functionals $S^{-1}=S^*$ and thus both $S$ and $U$ are unitary and the algebras generated by
$R_V(\mathcal{F}_{\text{loc}})$ and $R_{V'}(\mathcal{F}_{\text{loc}})$ are $*-$isomorphic with the isomorphism being implemented by the unitary $U$.
With this in mind, restricting the attention to a Minkowski spacetime and considering the interaction Lagrangian
\[
V_\lambda=\int \lambda \mathcal{L}_I d\mu_g
\]
with $\lambda$ a smooth and compactly supported function, it is possible to analyze the adiabatic limit as the inductive limit
\[
R_{V}(F) = \lim_{\lambda\to1} R_{V_\lambda}(F)\,.
\]
This limit is called algebraic adiabatic limit, further details can be found in~\cite{Brunetti:1999jn}.
In Minkowski spacetime we can use the spacetime symmetries to adapt the form of the cutoff $\lambda$ to the geometry of the background. Using a standard Minkowskian coordinate system $(t,\mathbf{x})$ where $t$ parametrises the time and $\mathbf{x}$ the space, we choose
\[
\lambda(t,\mathbf{x}) = \chi(t) h(\mathbf{x})
\]
where $\chi$ is a smooth time cutoff which can be chosen to be supported in $J^+(\Sigma_{2\epsilon})$ and equal to $1$ in $J^+(\Sigma_{\epsilon})$.
The spatial cutoff $h$ is a smooth function which is constant in time and compactly supported in space. Using the time slice axiom and the causality of the $S$ matrix, up to suitable unitaries of the above mentioned type,
$\chi$ can be fixed once and for all and the adiabatic limit amounts to $h\to1$.
For every $F\in \mathcal{F}_I(\mathcal{O})$, the algebraic adiabatic limit $h\to 1$ converges, in fact if $h_n$ is a family of compactly supported smooth functions which are equal to $1$ on the sphere $S_{n} = \{\mathbf{x}\in\mathbb{R}^3 | |x| < n \}$, due to causality, the sequence
$R_{V_{\chi h_n}}(F)$ is constant for $n\geq \overline{n}$ depending on the support of $F$.
In~\cite{Fredenhagen:2013cna} the relation between the free and interacting time evolution has been analysed. The free time evolution corresponds to a translation of the support functions. Indicating by $F_f$ the local field smeared with a test function $f$, the free time evolution is such that
\[
\alpha_s(F_f) = F_{f_s}, \qquad f_s(t,\mathbf{x}):=f(t-s,\mathbf{x}).
\]
The interacting time evolution acts on $\mathcal{F}_I$ as
\[
\alpha^V_s(R_{V}(F_f)) = R_{V}(\alpha_sF_f)=R_{V}(F_{f_s}).
\]
If $F$ is supported in $\Sigma_\epsilon$, the relation between the interacting and free time evolution is described by a cocycle $U_V(t)$ such that
\[
\alpha^V_t(F) = U_V^* \star \alpha_t(F) \star U_V
\]
In~\cite{Fredenhagen:2013cna} it has been shown how to construct the cocycle $U_V$ as a Dyson series of products of its generator $K_V$, see equation (25) in~\cite{Fredenhagen:2013cna}
where, for any $V = \int \chi h\mathcal{L}_{I}$,
\[
K_V = R_V(\mathcal{H}), \qquad \mathcal{H} = \int \dot{\chi} h \mathcal{L}_{I}.
\]
Having this cocycle at our disposal it is possible to construct the equilibrium (KMS) state repeating the so-called Araki-construction also for interacting field theories, when $h$ is of compact support. In particular it holds that
\[
\omega_\beta^V(A) = \frac{\omega(A\star U_V(i\beta))}{U_V(i\beta)}.
\]
The state constructed in this way depends explicitly on $h$, however, the limit $h\to1$ taken in the sense of van Hove is well defined, it is independent on $\chi$ and it constitutes an equilibrium state for the interacting theory in the adiabatic limit~\cite{Lindner:2013ila,Fredenhagen:2013cna} .
The methods developed in this project have opened the path to extending many interesting results of quantum statistical mechanics to quantum field theory.
For example, it was possible to prove return to equilibrium and stability for interaction Lagrangians of compact spatial support
\[
\lim_{t\to\infty} \omega^\beta\circ\alpha_t^V = \omega^{\beta,V}
\]
see~\cite{Drago:2016zlf}. In that paper it is also shown that the return to equilibrium ceases to hold if the adiabatic limit is considered before the large-time limit.
\begin{footnotesize}
\bibliographystyle{sfb676}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.