Quantum chaos and the arrow of time

Extracted source text · arXiv 2212.03914

\documentclass[12pt]{article}
%%
\pdfoutput=1
\usepackage{jheppub}
\usepackage{amsfonts}
%\usepackage{showkeys}
\usepackage{amsmath,amssymb} % for mathematical symbols
\usepackage{graphicx} % for \includegraphics command and options
\usepackage[dvipsnames,x11names]{xcolor} % to use color
\renewcommand{\baselinestretch}{1.1}
%%% ToC APPEARANCE
\usepackage[nottoc,notlof,notlot]{tocbibind} % to put the bibliography in the ToC
\usepackage[titles]{tocloft} % to alter the style of the ToC
\usepackage{hyperref}
\usepackage{cleveref}
\usepackage{float}
\usepackage{subcaption}
\usepackage{hhline}
\usepackage{bbold}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\title{\textbf{Positive $\Delta E$ theorem}}

\author{Nilakash Sorokhaibam}\emailAdd{phy\_sns@tezu.ernet.in}

\affiliation{Department of Physics, Tezpur University, Tezpur, Assam, India.}

\abstract{We show that the change in energy ($\Delta E$) of a system is always positive when a highly chaotic closed quantum system at finite positive temperature is perturbed. This result holds true universally at linear order. Assuming eigenstate thermalization hypothesis (ETH), we show that this result holds true at all orders. For these systems, the entropy (S) increases with increasing temperature. Moreover, entanglement entropy (EE) of a subsystem is also equal to the extensive thermal entropy associated with the subsystem. So, $\Delta E$, $\Delta S$ and $\Delta EE$ are always positive at finite positive temperature. If the system is at negative temperature, $\Delta E$ is negative but $\Delta S$ and $\Delta EE$ are positive. If the system is at infinite temperature, $\Delta E$, $\Delta S$ and $\Delta EE$ are all exactly zero.}


\begin{document}
\maketitle

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 
\section{Introduction}

%THE idea for this work comes to my mind when Gautam asked me the following question few months ago.
This paper tries to address the following question:\footnote{We thank Gautam Mandal for raising this question during one of our discussions few months ago.}

\vspace{.4cm}

\begin{center}
\begin{minipage}{14cm}
``Under what conditions can one prove, if at all, that if one starts from a non-equilibrium state, the entanglement entropy of a subsystem cannot decrease?"
\end{minipage}
\end{center}

\vspace{.4cm}

This paper is a partial answer to this question. We are considering highly chaotic closed quantum systems which thermalize. We can say this much that, when we perturb such a system initially in an equilibrium state, the entropy as well as the entanglement entropy of a subsystem always increase. These quantities are calculated when the system is back in equilibrium again. We are considering finite perturbations. The system goes out of equilibrium due to this finite perturbation but it eventually equilibrates back to another equilibrium state. We are not concerned with what happens during non-equilibrium dynamics, say, whether the entanglement entropy of a subsystem oscillates with time before equilibration, etc.

In this paper, we will rather concentrate on the total energy of the system. The non-equilibrium dynamics is due to a finite perturbation for a finite duration of time. Our precise mathematical result is that
\begin{center}
\begin{minipage}{14cm}
\emph{the total energy of the system always increases after such perturbations if the system satisfies eigenstate thermalization hypothesis (ETH).}
\end{minipage}
\end{center}
In other word, the change in energy ($\Delta E$) is always positive. ETH is generally studied for many-body systems. But our result will hold true for single particle systems, few body systems, many-body systems as well as quantum field theories (QFTs) as long as they satisfy ETH. For many-body systems and QFTs, the assertion that the entropy (S) always increases follows from the general feature of highly chaotic many-body systems that the entropy is extensive and increases with increasing (effective) temperature.  The assertion that the entanglement entropy (EE) of a subsystem always increases follows from another general feature of these systems that the observed entropy of a subsystem is actually the entanglement entropy of the subsystem being entangled to the rest of the system. So, the change in entropy ($\Delta S$) and the change in entanglement entropy ($\Delta EE$) of a subsystem are always positive.

The initial equilibrium state could be a thermal state or an energy eigenstate. In section \ref{linear}, we will work with thermal state. In section \ref{all_order}, we will work with an energy eigenstate. The above statements are true for non-zero and positive temperature (or effective temperature in case of energy eigenstate). If the Hilbert space or the Fock space is finite, negative temperature ranges are allowed. If the system was initially at negative temperature, then when disturbed, it actually loses energy and $\Delta E$ is negative. So, the system moves towards the middle of the energy spectrum. The state density is higher near the middle of the spectrum. Correspondingly, $\Delta S$ and $\Delta EE$ are positive. If it was initial at infinite temperature, then $\Delta E$, $\Delta S$ and $\Delta EE$ are all exactly zero.

It is easy to cook up a simple integrable system for which the present result does not hold true. This is expected because integrable systems in general does not satisfy ETH. Consider a single spin system with the Hamiltonian
\begin{equation}
\mathcal{H}_s=B\sigma_z, \qquad \sigma_z=\begin{pmatrix}
1 && 0\\
0 && -1
\end{pmatrix}
\end{equation}
Consider this system was initially in a thermal state with inverse temperature $\beta=1/T$. If we perturbed this system with a time dependent term
\begin{equation}
\mathcal{H}'_s=\lambda \delta(t) \sigma_x, \qquad \sigma_x=\begin{pmatrix}
0 && 1\\
1 && 0
\end{pmatrix}
\end{equation}
the change in the total energy is given by
\begin{equation}
\Delta E=4B\sinh(\beta B)\lambda\sin(2\lambda)=4B\sinh(\beta B)\left(2 \lambda ^2-\frac{4 \lambda ^4}{3}+\frac{4 \lambda ^6}{15}+O\left(\lambda ^7\right)\right)\
\end{equation}
Even for finite and positive temperature, we can see the change in energy can be negative depending on the value of the strength of the perturbation.\footnote{The system also does not thermalize but it is not our main concern in this paper.} In the $\lambda$ series, the first term is always positive independent of the value or sign of $\lambda$. We will show that this is true universally irrespective of whether the system is highly chaotic or not. This term comes from linear response theory. The second term is negative. In general, terms of order $\lambda^{2n}$ are negative (for positive integer $n$). We will show that all of these terms become positive if we consider systems satisfying ETH.

ETH \cite{PhysRevA.43.2046,Srednicki:1994mfb,DAlessio:2015qtq} is concerned with sensible observables for which statistical mechanics applies. It states that the matrix elements for such an observable $\mathcal{O}$ in the energy eigenstates are given by
\begin{equation}
\langle n|\mathcal{O}|m\rangle=O_{nm}=O(\bar{E})\delta_{nm}+e^{-S(\bar{E})/2} f_1(\bar{E},\omega)R_{nm}\
\label{ETH}
\end{equation}
where $|n\rangle$ and $|m\rangle$ are energy eigenstates with energies $E_n$ and $E_m$ respectively and $\bar{E}=(E_n+E_m)/2$, $\omega=E_n-E_m$ and $S(E)$ is the entropy at energy $E$. So that statistical mechanics is applicable, $\mathcal{O}(\bar{E})$ is equal to the expectation value of $\mathcal{O}$ in the microcanonical ensemble at energy $\bar{E}$. $\mathcal{O}(\bar{E})$ and $f_{1}(\bar{E},\omega)$ are also smooth functions of their arguments. $R_{nm}$ is a pseudo-random variable with zero mean and unit variance. ETH is after all a hypothesis which so far has no proof but it has been shown to hold true for physically sensible observables of highly chaotic systems. $\mathcal{O}$ is assumed to be hermitian and hence
\begin{equation}
O_{nm}=O^*_{nm}, \qquad R^*_{mn}=R_{nm}, \qquad f^*_{1}(\bar{E},-\omega)=f_{1}(\bar{E},\omega)
\end{equation}
$\mathcal{O}$ is considered to be a few-body observable in case of many-body systems or quantum field theories but not necessarily a local operator. It also does not commute with the Hamiltonian otherwise the dynamics will be trivial. More importantly but as expected, $\mathcal{O}$ cannot be a fermionic operator.

The system is perturbed by introducing a time-dependent $\mathcal{O}$ term in the Hamiltonian.\footnote{We will not consider coupling the system explicitly to a bath quantum system.} The total time-dependent Hamiltonian is
\begin{equation}
\mathcal{H}(t)=\mathcal{H}_0+\lambda(t)\,\mathcal{O}
\label{Ht}
\end{equation}
where $\mathcal{H}_0$ is the initial chaotic Hamiltonian as well as the final Hamiltonian. $\lambda(t)$ is a function of time $t$ with support for only a finite range of $t$ (say $0<t<s$). This means that the Hamiltonian is time-dependent for only a finite time interval. As we have mentioned above, $\mathcal{O}$ can be a local operator in case of local many-body systems or a quantum field theory. This local disturbance will spread out and the system will thermalize to an equilibrium state with higher total energy. Or, $\mathcal{O}$ could be a global operator. The system will again thermalize to an equilibrium state with higher total energy. We will calculate the energy gained (or lost) from the system due to this perturbation.

\section{General proof at linear order}
\label{linear}
As we have mentioned above, the contribution from linear response theory to $\Delta E$ is universally positive for finite and positive initial temperature irrespective of whether the system is highly chaotic or not. We consider the time dependent Hamiltonian (\ref{Ht}) where the real function $\lambda(t)$ has support for a finite range $0<t<s$. The rate of change of total energy is given by
\begin{equation}
\frac{dE}{dt}=\frac{d}{dt}\langle \mathcal{H}\rangle=\left\langle \frac{d\mathcal{H}}{dt}\right\rangle=\langle\mathcal{O}(t)\rangle\dot{\lambda}(t)\
\label{dEdt}
\end{equation}
We are working in linear response theory regime so we will expand $\mathcal{O}(t)$ only upto linear order in $\lambda(t)$.
\begin{equation}
\langle\mathcal{O}(t)\rangle=\langle\mathcal{O}(0)\rangle-i\int_{-\infty}^{t}\langle[\mathcal{O}_I(t),\mathcal{O}_I(t')]\rangle \lambda(t')dt'=\langle\mathcal{O}(0)\rangle+\int_{-\infty}^{\infty}G_R(t,t')dt'\
\end{equation}
$\mathcal{O}_I(t)=e^{it\mathcal{H}_0}\mathcal{O}(0)e^{-it\mathcal{H}_0}$ implies working in the interaction picture. $G_R(t,t')=-i\theta(t-t')\langle[\mathcal{O}_I(t),\mathcal{O}_I(t')]\rangle$ is the usual definition of the retarded Green's function of the operator $\mathcal{O}$ for a qauntum theory with the Hamiltonian $\mathcal{H}_0$. The expectation values are taken for the initial equilibrium state, not the time evolved state in interaction picture. The total energy gained upto this order is
\begin{equation}
\Delta E=\int_{t_1}^{t_2}\frac{dE}{dt}=\int_{t_1}^{t_2} \langle\mathcal{O}(0)\rangle \dot{\lambda}(t)dt+\int_{t_1}^{t_2}\int_{-\infty}^{\infty} dt dt' \; G_R(t,t')\lambda(t')\dot{\lambda}(t)\
\end{equation}
where $t_1\leq 0$ and $t_2\geq s$. The first term is zero because $\lambda(t_1)=\lambda(t_2)=0$ and $\mathcal{O}(0)$ is a constant.
Let us take $t_1\to-\infty$ and $t_2\to\infty$ so that we can work in the frequency space. Going to Fourier space using the inverse Fourier transform representation
\begin{equation}
\lambda(t')=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}\lambda(\omega)e^{-i\omega't'}d\omega', \quad 
\dot{\lambda}(t')=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}(-i\omega')\lambda(\omega)e^{-i\omega't'}d\omega'
\label{fourierconv}
\end{equation}
we get,
\begin{equation}
\Delta E=\frac{-i}{2\pi}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty} d\omega d\omega' \omega \lambda(\omega)\lambda(\omega')\int_{-\infty}^{\infty}\int_{-\infty}^{\infty} dt dt' G_R(t,t')\;e^{-i\omega't'}e^{-i\omega t}\
\end{equation}
$G_R(t,t')=G_R(t-t')$ because it is calculated in a thermal state. Now after a change of variable to $\tilde{t}=t-t', T=t+t'$, we have
\begin{eqnarray}
&& \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} dt dt' \; G_R(t-t')\;e^{-i\omega't'}e^{-i\omega t}\nonumber\\
&=& \frac{1}{2}\,\sqrt{2\pi} \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} d\tilde{t} dT\; G_R(\tilde{t})\;e^{-i(\omega'+\omega)T/2}e^{i(\omega'-\omega) \tilde{t}/2}\nonumber\\
&=&(2\pi)^{3/2}\; G_R\left(\frac{\omega'-\omega}{2}\right)\delta\left(\omega'+\omega\right)\nonumber\
\end{eqnarray}
Substituting we get,
\begin{eqnarray}
\Delta E=-i\sqrt{2\pi}\int_{-\infty}^{\infty} d\omega' (-\omega') \lambda(\omega')\lambda(-\omega')\; G_R\left(\omega'\right)\
\end{eqnarray}
$\lambda(t)$ is real so $\lambda(\omega')\lambda(-\omega')$ is real, positive and even. So, only the imaginary part of $G_R(\omega)$ will contribute in the integral above. For any operator which are not fermionic, the real part of $G_R(\omega)$ is even and the imaginary part of $G_R(\omega)$ is odd. So, we get
\begin{equation}
\Delta E=-\, 2\, \sqrt{2\pi}\int_{0}^{\infty} d\omega \,\omega \,|\lambda(\omega)|^2\; \text{Im}\; G_R(\omega)\
\end{equation}
The spectral function is defined as
\begin{equation}
A(\omega)=-2\,\text{Im}\;G_R(\omega)
\end{equation}
Again for any operator which are not fermionic, $A(\omega)$ is positive for $\omega>0$ and negative for $\omega<0$ at positive temperature.
So, in the end we get
\begin{equation}
\Delta E=\sqrt{2\pi}\int_{0}^{\infty} d\omega \omega |\lambda(\omega)|^2\, A(\omega)\
\end{equation}
which is a positive quantity.

At negative temperature, $\Delta E$ is negative because $A(\omega)$ sign is flipped. At infinite temperature, $\Delta E$ is zero because $A(\omega)$ is zero.
\section{Proof for all order assuming ETH}
\label{all_order}
In this section, ETH will play a very important role. For cleaner mathematical steps, let use consider that the thermal expectation value $O(\bar{E})$ is zero. It can be easily verify that all the results still hold true with nonzero values of $O(\bar{E})$. We have,
\begin{equation}
\langle n|\mathcal{O}|m\rangle=O_{nm}=e^{-S(\bar{E})/2} f_1(\bar{E},\omega)R_{nm}\
\label{ETH1}
\end{equation}
It is easy to show that $\mathcal{O}^2$, $\mathcal{O}^3$ etc. also satisfy ETH \cite{Srednicki:1995pt}.
\begin{gather}
\langle n|\mathcal{O}^2|m\rangle=O^2_{nm}=O^2(\bar{E})\delta_{nm}+e^{-S(\bar{E})/2} f_2(\bar{E},\omega)R'_{nm}\\
\langle n|\mathcal{O}^3|m\rangle=O^3_{nm}=O^3(\bar{E})\delta_{nm}+e^{-S(\bar{E})/2} f_3(\bar{E},\omega)R''_{nm}\
\label{ETH23}
\end{gather}
We will later exploit the fact that $R_{nm}$, $R'_{nm}$, $R''_{nm}$, etc are in general independent. We will use the following expression of $\Delta E$.
\begin{equation}
\Delta E=\int_{-\infty}^{\infty}\frac{dE}{dt}=\int_{-\infty}^{\infty} \langle\mathcal{O}(t)\rangle \dot{\lambda}(t)dt=-\int_{-\infty}^{\infty}\lambda(t)\,\left\langle\frac{d\mathcal{O}(t)}{dt}\right\rangle\, dt
\end{equation}
where we have used $\lim_{t\to\pm\infty}\lambda(t)=0$, which is the case for the perturbations of finite duration that we are considering.

For simplicity, let us first consider the case in which the perturbing term has delta function time dependence.
\begin{equation}
\mathcal{H}(t)=\mathcal{H}_0+\delta(t)\lambda\mathcal{O}\
\end{equation}
The case of general $\lambda(t)$ will be considered later in this section. From here onwards, $\mathcal{O}=\mathcal{O}(0)$. The initial state of the system is also taken to be an energy eigenstate $|n\rangle$. Now
\begin{eqnarray}
\mathcal{O}(t)&=& e^{i\lambda \mathcal{O}} e^{it\mathcal{H}_0}\mathcal{O}(0)e^{-it\mathcal{H}_0} e^{-i\lambda \mathcal{O}}= e^{i\lambda \mathcal{O}} \mathcal{O}_I(t) e^{-i\lambda \mathcal{O}}\nonumber\\
&=& \mathcal{O}_I(t)+i\lambda [\mathcal{O},\mathcal{O}_I(t)]+\frac{(i\lambda)^2}{2!} [\mathcal{O},[\mathcal{O},\mathcal{O}_I(t)]]+\frac{(i\lambda)^3}{3!} [\mathcal{O},[\mathcal{O},[\mathcal{O},\mathcal{O}_I(t)]]]\nonumber\\
&&+\frac{(i\lambda)^4}{4!} [\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},\mathcal{O}_I(t)]]]+\frac{(i\lambda)^5}{5!} [\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},\mathcal{O}_I(t)]]]+...\nonumber\
\end{eqnarray}
\begin{eqnarray}
\Delta E&=&\int_{-\infty}^{\infty}\frac{dE}{dt}=\int_{-\infty}^{\infty} \langle\mathcal{O}(t)\rangle \dot{\lambda}(t)dt=\int_{-\infty}^{\infty} \lambda(t) \left\langle\frac{d\mathcal{O}(t)}{dt}\right\rangle dt=-\lambda\left\langle\left.\frac{d\mathcal{O}(t)}{dt}\right\rangle\right|_{t=0}\nonumber\\
&=&-\lambda \left[\left\langle\left.\frac{d\mathcal{O}(t)}{dt}\right\rangle\right|_{t=0}+i\lambda \left\langle\left.\left[\mathcal{O},\frac{d\mathcal{O}(t)}{dt}\right]\right\rangle\right|_{t=0}\right.\nonumber\\
&&+\frac{(i\lambda)^2}{2!} \left\langle\left.\left[\mathcal{O},[\mathcal{O},\frac{d\mathcal{O}(t)}{dt}]\right]\right\rangle\right|_{t=0}+\frac{(i\lambda)^3}{3!} \left\langle\left.\left[\mathcal{O},[\mathcal{O},[\mathcal{O},\frac{d\mathcal{O}(t)}{dt}]]\right]\right\rangle\right|_{t=0}\nonumber\\
&&\left.+\frac{(i\lambda)^4}{4!} \left\langle\left.\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},\frac{d\mathcal{O}(t)}{dt}]]]\right]\right\rangle\right|_{t=0}+\frac{(i\lambda)^5}{5!} \left\langle\left.\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},\frac{d\mathcal{O}(t)}{dt}]]]]\right]\right\rangle\right|_{t=0}+...\right]\nonumber\\
&=&-\lambda \left\langle i[H_0,\mathcal{O}]\right\rangle-i\lambda^2 \left\langle\left[\mathcal{O},i[H_0,\mathcal{O}]\right]\right\rangle\nonumber\\
&&+\frac{\lambda^3}{2!} \left\langle\left[\mathcal{O},[\mathcal{O},i[H_0,\mathcal{O}]]\right]\right\rangle+\frac{i\lambda^4}{3!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},i[H_0,\mathcal{O}]]]\right]\right\rangle\nonumber\\
&&-\frac{\lambda^5}{4!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},i[H_0,\mathcal{O}]]]]\right]\right\rangle+\frac{i\lambda^6}{5!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},i[H_0,\mathcal{O}]]]]]\right]\right\rangle+...\nonumber\\
&=&-i\lambda \left\langle [H_0,\mathcal{O}]\right\rangle+\lambda^2 \left\langle\left[\mathcal{O},[H_0,\mathcal{O}]\right]\right\rangle+\frac{i\lambda^3}{2!} \left\langle\left[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]\right]\right\rangle-\frac{\lambda^4}{3!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]]\right]\right\rangle\nonumber\\
&&-\frac{i\lambda^5}{4!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]]]\right]\right\rangle+\frac{\lambda^6}{5!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]]]]\right]\right\rangle+...\nonumber\
\end{eqnarray}
The odd powers of $\lambda$ terms have zero or negligible expectation value compared to the other terms. This can be easily shown by explicitly writing out all the terms of the commutators and take the expectation value in the energy eigenstate $|n\rangle$. For example,
\begin{eqnarray}
&& \langle n|[\mathcal{H}_0,\mathcal{O}]|n\rangle=E_n O_{nn}-O_{nn}E_n=0\nonumber\\
&& \langle n|[\mathcal{O},[\mathcal{O},[\mathcal{H}_0,\mathcal{O}]]]|n\rangle=\langle n|(3\mathcal{O}^2\mathcal{H}_0\mathcal{O}-3\mathcal{O}\mathcal{H}_0\mathcal{O}^2+\mathcal{O}^3\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^3|n\rangle\nonumber\\
&=&3\sum_m E_m(O^2_{nm}O_{mn}-O_{nm}O^2_{mn})+E_n O^3_{nn}-E_nO^3_{nn}\nonumber\\
&=&3\sum_m E_m(O^2_{nm}O_{mn}-O_{nm}O^2_{mn})\
\label{oddzero}
\end{eqnarray}
The diagonal value $O^2(\bar{E})$ of $O^2_{mn}$ dropped out. If $O_{mn}$ and $O^2_{mn}$ are real then the final expression is exactly zero. Even if they are not real, due to the pseudo random variables $R_{nm}$ and $R'_{nm}$, the total sum is negligible, suppressed by a factor of $e^{-S(\bar{E})/2}$ compared to the other terms. So we have,
\begin{eqnarray}
\Delta E&=&\lambda^2 \left\langle\left[\mathcal{O},[H_0,\mathcal{O}]\right]\right\rangle-\frac{\lambda^4}{3!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]]\right]\right\rangle\nonumber\\
&&+\frac{\lambda^6}{5!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]]]]\right]\right\rangle-\frac{\lambda^8}{7!} \left\langle\left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[H_0,\mathcal{O}]]]]]\right]\right\rangle+...\nonumber\\
\end{eqnarray}
As we have shown above, the first term is always positive. In the above form, it means that the expectation value is always positive.
\begin{eqnarray}
\langle E_n|\left[\mathcal{O},[H_0,\mathcal{O}]\right]|E_n\rangle&=&\langle E_n|2\mathcal{O}H_0\mathcal{O}-\mathcal{O}^2H_0-H_0\mathcal{O}^2|E_n\rangle\nonumber\\
&=&\sum_m 2|O_{nm}|^2(E_m-E_n)
\end{eqnarray}
Note that this is due to the Hermitian nature of $\mathcal{O}$ so $O^*_{nm}=O_{mn}$ and $E_n$ is in the lower half of the energy spectrum, meaning the effective inverse temperature $\beta$ is non-zero and positive. Our main result will depend heavily on this mathematical inference. The expectation value will be negative if the state $|n\rangle$ is in the upper half of the spectrum or the effective temperature is negative. If the effective temperature is infinite or $E_n$ is exactly at the middle of the energy spectrum, then the above expectation value will be exactly zero.

Let us now consider the commutator in the $\lambda^4$ term. It can be simplified as
\begin{eqnarray}
&& \left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{H}_0,\mathcal{O}]]]\right]\nonumber\\
&=& 4\mathcal{O}^3 \mathcal{H}_0\mathcal{O}+4\mathcal{O}\mathcal{H}_0\mathcal{O}^3-6\mathcal{O}^2\mathcal{H}_0\mathcal{O}^2-\mathcal{O}^4\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^4\nonumber\\
&=& -3(2\mathcal{O}^2\mathcal{H}_0\mathcal{O}^2-\mathcal{O}^4\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^4)+4(\mathcal{O}^3 \mathcal{H}_0\mathcal{O}+\mathcal{O}\mathcal{H}_0\mathcal{O}^3-\mathcal{O}^4\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^4)\nonumber\\
\label{l4delta}
\end{eqnarray}
The expectation value for the terms in the first set of brackets is positive.
\begin{eqnarray}
&&\langle n|2\mathcal{O}^2\mathcal{H}_0\mathcal{O}^2-\mathcal{O}^4\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^4|n\rangle\nonumber\\
&=& 2\sum_m |O^2_{nm}|^2(E_m-E_n)
\label{l4large}
\end{eqnarray}
$|\mathcal{O}^2_{mn}|$ is positive definite and the rest of the reasoning is same as for the $\lambda^2$ term above. Still the expectation value of the terms inside the second set of brackets can play spoilsport. But we can show that the expectation value of the terms inside the second set of brackets in (\ref{l4delta}) is negligible. The expectation value is
\begin{eqnarray}
&&\langle n|\mathcal{O}^3 \mathcal{H}_0\mathcal{O}+\mathcal{O}\mathcal{H}_0\mathcal{O}^3-\mathcal{O}^4\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^4|n\rangle\nonumber\\
&=& \sum_m (O^3_{nm}O_{mn}+O_{mn}O^3_{mn})(E_m-E_n)\
\label{l4small}
\end{eqnarray}
This sum is much smaller, suppressed by a factor of $e^{-S(\bar{E})/2}$, compared to the sum (\ref{l4large}). The reasoning is similar to that we have given for (\ref{oddzero}). $(m=n)$ terms dropped out. In the remaining terms, $\mathcal{O}^3_{mn}\mathcal{O}_{mn}$ is not positive definite and they have a random component so most of them will cancel out. Note that this is different from the result of higher-point functions of Foini and Kurchan \cite{Foini:2018sdb}. In their paper, the sum for the randomly fluctuating terms consists of much larger number of terms. Here, as we can see, the first sum as well as the second sum is summed over the same number of terms. As a result, the total contribution from order $\lambda^4$ is also positive. Similarly $\lambda^6$, $\lambda^8$ terms always give positive contribution.
\begin{eqnarray}
&& \left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{H}_0,\mathcal{O}]]]\right]\nonumber\\
&=& 10(2\mathcal{O}^3 \mathcal{H}_0 \mathcal{O}^3-\mathcal{O}^6\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^6)-15(\mathcal{O}^2\mathcal{H}_0\mathcal{O}^4+\mathcal{O}^4\mathcal{H}_0\mathcal{O}^2-\mathcal{O}^6\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^6)\nonumber\\
&& \qquad +6(\mathcal{O}\mathcal{H}_0\mathcal{O}^5+\mathcal{O}^5\mathcal{H}_0\mathcal{O}-\mathcal{O}^2\mathcal{H}_0\mathcal{O}^4-\mathcal{O}^4\mathcal{H}_0\mathcal{O}^2)\nonumber\\
&& \left[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{O},[\mathcal{H}_0,\mathcal{O}]]]\right]\nonumber\\
&=& -35(2\mathcal{O}^4\mathcal{H}_0\mathcal{O}^4-\mathcal{O}^8\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^8)+56(\mathcal{O}^5\mathcal{H}_0\mathcal{O}^3+\mathcal{O}^3\mathcal{H}_0\mathcal{O}^5-\mathcal{O}^8\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^8)\nonumber\\
&&+8(\mathcal{O}\mathcal{H}_0\mathcal{O}^7+\mathcal{O}^7\mathcal{H}_0\mathcal{O}-\mathcal{O}^8\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^8)-28(\mathcal{O}^6\mathcal{H}_0\mathcal{O}^2+\mathcal{O}^2\mathcal{H}_0\mathcal{O}^6-\mathcal{O}^8\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^8)\nonumber\\
\end{eqnarray}
In general, the $\lambda^a$ term is
\begin{eqnarray}
&& (-1)^a\frac{1}{(a-1)!} [\mathcal{O},[\mathcal{O},...,[\mathcal{H}_0,\mathcal{O}]...]]\nonumber\\
&=&\frac{C^a_{a/2}}{2}\, (2\mathcal{O}^{a/2}\mathcal{H}_0 \mathcal{O}^{a/2}-\mathcal{O}^a \mathcal{H}_0-\mathcal{H}_0\mathcal{O}^a)\nonumber\\
&& -C^a_{a/2-1}(\mathcal{O}^{a/2}\mathcal{H}_0 \mathcal{O}^{a/2-1}+\mathcal{O}^{a/2-1}\mathcal{H}_0 \mathcal{O}^{a/2}-\mathcal{O}^a \mathcal{H}_0-\mathcal{H}_0\mathcal{O}^a)+...\
\end{eqnarray} 
where $C^a_b=\frac{a!}{b!(a-b)!}$. The first term is always positive and big compared to the rest of the terms. So, $\Delta E$ is positive.

Now let us consider the case of general time dependence but with finite support. So,
\begin{equation}
\mathcal{H}(t)=\mathcal{H}_0+\lambda(t)\mathcal{O}, \quad \text{where} \quad \lambda(t)=0 \quad \text{for} \quad t<0, t>s
\end{equation}
We will work in the interaction picture. The analysis is in a similar line to what we have done above for the delta function perturbation.
\begin{equation}
\Delta E=-\int_{-\infty}^{\infty}\lambda(t)\,\left\langle\frac{d\mathcal{O}(t)}{dt}\right\rangle\, dt=-i\int_{-\infty}^{\infty}\lambda(t)\,\left\langle\mathcal{U}_I^\dagger(t)\left[\mathcal{H}_I(t),\mathcal{O}_I(t)\right]\mathcal{U}_I(t)\right\rangle dt
\end{equation}
where $\mathcal{U}_I(t)$ is the time evolution operator in interaction picture. The expectation value is taken for the initial state at time $t=0$. Working in the interaction picture makes sure that the series representation of $\mathcal{U}_I(t)$ is in powers of $\lambda(t)$.
\begin{eqnarray}
&&\mathcal{U}_I^\dagger(t)\left[\mathcal{H}_I(t),\mathcal{O}_I(t)\right]\mathcal{U}_I(t)\nonumber\\
%&=&\mathcal{U}_I^\dagger(t)\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\mathcal{U}_I(t)\nonumber\\
&=&\left(\mathbb{1}+i\int_0^{t}dt_1 \lambda(t_1)\mathcal{O}_I(t_1)-\int_0^tdt_1\int_0^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)\right.\nonumber\\
&&\left.-i\int_0^tdt_1\int_0^{t_1}dt_2\int_0^{t_2}dt_3\lambda(t_1)\lambda(t_2)\lambda(t_3)\mathcal{O}_I(t_3)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)+...\right)\nonumber\\
&&\times\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\times\left(\mathbb{1}-i\int_0^{t}dt_1 \lambda(t_1)\mathcal{O}_I(t_1)-\int_0^tdt_1\int_0^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)\right.\nonumber\\
&&\left.+i\int_0^tdt_1\int_0^{t_1}dt_2\int_0^{t_2}dt_3\lambda(t_1)\lambda(t_2)\lambda(t_3)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)\mathcal{O}_I(t_3)+...\right)\nonumber\
\label{lgenexp}
\end{eqnarray}
The $\lambda^0$ term has zero expectation value. The terms at $\lambda^1$ order is the linear response theory term. We have already shown that the total contribution to $\Delta E$ from the terms is always positive. The $\lambda^2$ order terms are
\begin{eqnarray}
&&\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\int_{-\infty}^tdt''\lambda(t'')\mathcal{O}_I(t'')\nonumber\\
&&-\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\nonumber\\
&&-\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)
\end{eqnarray}
In the above equation, the lower limits of the integral have been extended to $-\infty$. This is possible because $\lambda(t)$ is zero for $t<0$. Now the expectation value is taken for the state at time $t=-\infty$ which is same as the state at time $t=0$ since it is an equilibrium state. We can show that the total contribution at this order is suppressed by $e^{-S(\bar{E})/2}$, just like with the delta function perturbation. The detailed steps are simple and involve simple algebraic steps and some contour integrations. We will rather work out the details for some terms at $\lambda^4$ order and show that their contribution is always dominant and positive. The terms at $\lambda^3$ in (\ref{lgenexp}) are
\begin{eqnarray}
&&-i\left(\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\int_0^tdt_1\int_0^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)\right.\nonumber\\
&&\qquad -\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')\nonumber\\
&&\qquad +\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\int_{-\infty}^{t_2}dt_3\lambda(t_1)\lambda(t_2)\lambda(t_3)\mathcal{O}_I(t_3)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\nonumber\\
&&\qquad -\left.\left[\mathcal{H}_0,\mathcal{O}_I(t)\right]\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\int_{-\infty}^{t_2}dt_3\lambda(t_1)\lambda(t_2)\lambda(t_3)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)\mathcal{O}_I(t_3)\right)\nonumber\
\end{eqnarray}
Out of these terms, the dominant contribution to $\Delta E$ is from terms which are similar to $2\mathcal{O}^2\mathcal{H}_0\mathcal{O}^2-\mathcal{O}^4\mathcal{H}_0-\mathcal{H}_0\mathcal{O}^4$ in case of delta function perturbation which we have shown in (\ref{l4large}). In the present case, we have
\begin{eqnarray}
&&\int_{-\infty}^{\infty}dt\,\lambda(t)\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')\mathcal{O}_I(t)\mathcal{H}_0\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)\nonumber\\
&&\qquad -\int_{-\infty}^{\infty}dt\,\lambda(t)\int_{-\infty}^tdt'\lambda(t')\mathcal{H}_0\mathcal{O}_I(t')\mathcal{O}_I(t)\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)\nonumber\\
&&+\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)\mathcal{H}_0\mathcal{O}_I(t)\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')\nonumber\\
&&\qquad -\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)\mathcal{O}_I(t)\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')\mathcal{H}_0\nonumber\
\end{eqnarray}
The expectation value in energy eigenstate $|n\rangle$ is given by
\begin{eqnarray}
&&\sum_m(E_m-E_n)\left[\int_{-\infty}^{\infty}dt\,\lambda(t)\int_{-\infty}^tdt'\lambda(t')\langle n|\mathcal{O}_I(t')\mathcal{O}_I(t)|m\rangle\right.\nonumber\\
&&\hspace{4cm}\times\left.\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\langle m|\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)|n\rangle\right]\nonumber\\
&&+\sum_m(E_m-E_n)\left[\int_{-\infty}^{\infty}dt\,\lambda(t)\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\langle n|\mathcal{O}_I(t_2)\mathcal{O}_I(t_1)|m\rangle\right.\nonumber\\
&&\hspace{4cm}\times\left.\langle m|\mathcal{O}_I(t)\int_{-\infty}^tdt'\lambda(t')\mathcal{O}_I(t')|n\rangle\right]\nonumber\
\end{eqnarray}
the rest of the contributions are suppressed by $e^{-S(\bar{E})/2}$, just like in case of the delta function perturbation. Now as in (\ref{l4large}), we have to show that the product of the factors in the square brackets are positive definite. This will complete the proof that the contribution from $\lambda^4$ order in $\Delta E$ is positive (assuming initial positive effective temperature or $E_n$ is in the lower half of the energy spectrum). The steps and the results are similar for the two factors so we will write down only for the first factor. The first step is to insert two complete set of states so that each $\mathcal{O}_I$ operators can be replaced by their matrix elements. The next step is to replace $\lambda(t)$'s with their inverse Fourier transform representation so that we can integrate out the time variables. This process left us with frequency $\omega$ integrals. The frequency integrals can be evaluated using contour integration. The contours have to be closed in the lower half of the complex plane. This is fixed by the fact the time integrals have to be finite as $t\to-\infty$ and we are using the convention (\ref{fourierconv}) for the inverse Fourier transforms. Finally we get,
\begin{eqnarray}
&&\int_{-\infty}^{\infty}dt\,\lambda(t)\int_{-\infty}^tdt'\lambda(t')\langle n|\mathcal{O}_I(t')\mathcal{O}_I(t)|m\rangle\int_{-\infty}^tdt_1\int_{-\infty}^{t_1}dt_2\lambda(t_1)\lambda(t_2)\langle m|\mathcal{O}_I(t_1)\mathcal{O}_I(t_2)|n\rangle\nonumber\\
&=& (2\pi)^2\left(\sum_l\lambda(E_n-E_l)O_{nl}\lambda(E_l-E_m)O_{lm}\right)\left(\sum_k\lambda(E_m-E_k)O_{mk}\lambda(E_k-E_n)O_{kn}\right)\nonumber\
\end{eqnarray}
The two factors inside the brackets are conjugate of each other because $\lambda(-\omega)=\lambda^*(\omega)$ and $O_{nm}=O^*_{mn}$. So, the full expression is positive definite. Hence, $\Delta E$ at $\lambda^4$ order is positive. Similarly, we can show that the contribution to $\Delta E$ at all orders are positive. The delta function perturbation is a good guide in working out the details at higher orders.

Just as in case of the delta function perturbation, if we start from an initial state with negative (effective) temperature then $\Delta E$ is negative. If the initial temperature is infinite then $\Delta E$ is zero.

\section{Conclusions}
\label{conclusions}
We show that the change in energy ($\Delta E$) of a highly chaotic system is positive when the system is disturbed. In case of many-body systems or quantum field theories, this also implies that the change in entropy ($\Delta S$) and the change in entanglement entropy ($\Delta EE$) of a subsystem are also positive. The system is considered to be initially in an equilibrium state with finite and positive temperature. If the system was initially in a state with negative temperature, then $\Delta E$ is negative but $\Delta S$ and $\Delta EE$ are positive. If the system was initially at infinite temperature, then $\Delta E$, $\Delta S$ and $\Delta EE$ are all exactly zero. We would like to emphasize that our result holds true irrespective of the pseudo random nature of $R_{nm}$ in (\ref{ETH}). As explained in \cite{Foini:2018sdb}, this pseudo random nature of $R_{nm}$ gives rise to non-trivial higher-point functions.
\section*{Acknowledgement}
The author would like to thank Gautam Mandal for helpful discussions. The author is fully supported by  Department of Science and Technology (Government of India) under the INSPIRE Faculty fellowship scheme
 (Ref. No. DST/INSPIRE/04/2020/002105).
\appendix

\bibliography{dE_theorem} 
\bibliographystyle{JHEP}

\end{document}


\providecommand{\href}[2]{#2}\begingroup\raggedright\begin{thebibliography}{1}

\bibitem{PhysRevA.43.2046}
J.~M. Deutsch, {\it Quantum statistical mechanics in a closed system},  {\em
  Phys. Rev. A} {\bf 43} (Feb, 1991) 2046--2049.

\bibitem{Srednicki:1994mfb}
M.~Srednicki, {\it {Chaos and Quantum Thermalization}},  {\em Phys. Rev. E}
  {\bf 50} (Aug, 1994) 888--901,
  [\href{http://arxiv.org/abs/cond-mat/9403051}{{\tt cond-mat/9403051}}].

\bibitem{DAlessio:2015qtq}
L.~D'Alessio, Y.~Kafri, A.~Polkovnikov, and M.~Rigol, {\it {From quantum chaos
  and eigenstate thermalization to statistical mechanics and thermodynamics}},
  {\em Adv. Phys.} {\bf 65} (2016), no.~3 239--362,
  [\href{http://arxiv.org/abs/1509.06411}{{\tt arXiv:1509.06411}}].

\bibitem{Srednicki:1995pt}
M.~Srednicki, {\it {Thermal fluctuations in quantized chaotic systems}},  {\em
  J. Phys. A} {\bf 29} (1996) L75--L79,
  [\href{http://arxiv.org/abs/chao-dyn/9511001}{{\tt chao-dyn/9511001}}].

\bibitem{Foini:2018sdb}
L.~Foini and J.~Kurchan, {\it {Eigenstate thermalization hypothesis and out of
  time order correlators}},  {\em Phys. Rev. E} {\bf 99} (2019), no.~4 042139,
  [\href{http://arxiv.org/abs/1803.10658}{{\tt arXiv:1803.10658}}].

\end{thebibliography}\endgroup