Relaxation of imbalance in a disordered XX model with on-site dephasing
Extracted source text ยท arXiv 2301.05611
%\documentclass[aps,prb,showpacs,twocolumn,floats,epsfig]{revtex4}
\documentclass[aps,prb,twocolumn,floats,epsfig,groupaddress]{revtex4-1}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{amsmath}
\usepackage{xcolor}
\usepackage{enumerate}
\usepackage{bm}
\usepackage{epsfig}
\usepackage{makecell}
\usepackage{color}
\usepackage{soul}
\usepackage{float}
\usepackage{braket}
%\usepackage{subfig}
\usepackage[colorlinks,linktocpage,bookmarks=false,citecolor=blue,linkcolor=red,urlcolor=blue]{hyperref}
\newcommand\mybar{\kern1pt\rule[-\dp\strutbox]{.8pt}{\baselineskip}\kern1pt}
\newcommand{\wt}{\widetilde}
\newcommand{\ing}{\includegraphics}
\newcommand{\bib}{\bibitem}
\newcommand{\beq}{\begin{equation}}
\newcommand{\eeq}{\end{equation}}
\newcommand{\bea}{\begin{eqnarray}}
\newcommand{\eea}{\end{eqnarray}}
\newcommand{\al}{\alpha}
\newcommand{\de}{\delta}
\newcommand{\De}{\Delta}
\newcommand{\ep}{\epsilon}
\newcommand{\ga}{\gamma}
\newcommand{\ka}{\kappa}
\newcommand{\lamb}{\lambda}
\newcommand{\si}{\sigma}
\newcommand{\ta}{\theta}
\newcommand{\om}{\omega}
\newcommand{\calC}{{\cal C}}
\newcommand{\calD}{{\cal D}}
\newcommand{\calG}{{\cal G}}
\newcommand{\w}{\omega}
\newcommand{\non}{\nonumber}
\newcommand{\noi}{\noindent}
%\newcommand{\red}{\textcolor{red}}
%\newcommand{\blue}{\textcolor{blue}}
%\newcommand{\redbar}[1]{\textcolor{red}{\st{#1}}} % requires package soul
\newcommand{\Tr}{{\rm Tr}}
\newcommand{\mean}[1]{\langle #1 \rangle}
\newcommand{\half}{\frac{1}{2}}
\newcommand{\SN}{S_{\rm N}}
\newcommand{\tr}{{\rm tr}}
\newcommand{\bracket}[3]{\langle #1 | #2 | #3 \rangle}
%\newcommand{\red}{\color{red}}
\usepackage{xcolor}
\newcommand{\marko}[1]{{\color{blue} [MZ: #1]}}
\begin{document}
\title{Relaxation of imbalance in disordered XX model with on site dephasing}
\author{Roopayan Ghosh}
\affiliation{Department of Physics and Astronomy, University College London,Gower Street, London}
\affiliation{Department of Physics, Faculty of Mathematics and Physics, University of Ljubljana, Jadranska 19, SI-1000 Ljubljana, Slovenia}
\author{Marko \v Znidari\v c}
\affiliation{Department of Physics, Faculty of Mathematics and Physics, University of Ljubljana, Jadranska 19, SI-1000 Ljubljana, Slovenia}
\begin{abstract}
The relaxation of observables to their non-equilibrium steady states in a disordered XX chain subjected to dephasing at every site has been intensely studied in recent years. We comprehensively analyse the relaxation of staggered magnetization, i.e., imbalance, in such a system, starting from the N\'eel initial state. We analytically predict emergence of several timescales in the system and extract results which match with large-system numerics without any extra fitting parameter till a universal timescale. An often reported stretched exponential decay is just one of the regimes which holds in a finite window of time and is therefore in fact not a true stretched exponential decay. Subsequently, the asymptotic decay of imbalance is governed by a power law irrespective of the disorder. We show that this emerges from the continuum limit of the low magnitude eigenspectrum of the Liouvillian. However, for finite systems, due to discreteness of the spectrum, the final phase of relaxation is governed by the relevant smallest Liouvillian gap.
\end{abstract}
\maketitle
\section{Introduction}
Non-interacting disordered systems in one-dimension, isolated from external environment, are known to exhibit Anderson localization. \cite{AndersonPhysRev.109.1492}. One expects that if a generic initial state is allowed to evolve in such a system, at long time scales, the wavefunction will not show significant change. Among the several quantifiers of this phenomenon\cite{MirlinRevModPhys.80.1355}, one of the most commonly used is imbalance, a quantity easy to measure in experiments. Imbalance $I$ is the staggered magnetization and captures the difference in orientation of spins on adjacent sites,
\begin{equation}
I=(1/L)\sum_{j=1}^L(-1)^j \langle\sigma_j^z\rangle,
\end{equation}
where $L$ is the length of the lattice, $\sigma^z$ is the Pauli spin z operator, and the expectation is taken with the state we want to measure the quantity. It should be evident that for computational states, the state with the largest $I$, which equals $1$, is the N\'eel state.
One can also surmise that for an Anderson localized system, if one starts from this state, one should see $I\sim 1$ at large timescales. However, if the localized system is no longer isolated, then the external degrees of freedom typically serve to break Anderson localization. The system eventually forgets the memory of its initial state. Accordingly, $I$ would also evolve with time. In the previous works, the focus has mainly been to find the non equilibrium steady state (NESS) in such systems\cite{Znidaric_2010,ScarPhysRevB.103.184202,PolettiPhysRevA.95.052107,PhysRevB.98.020202Flach,Wu_2019,Monthus_2017,PhysRevLett.118.070402Iva,Vershinina_2017}, or perturbations around NESS\cite{SchiroPhysRevB.104.144301}, while recently there has been some work studying how different observables relax to NESS in such systems\cite{MariyaMarkoPhysRevB.93.094205,GopalKnapPhysRevLett.119.046601,PolettiPhysRevLett.114.170401,PhysRevLett.116.237203Garrahan,10.21468/SciPostPhys.12.5.174BarLev,PhysRevB.95.024310Garrahan,PhysRevB.100.165144Poletti} or in Stark localized systems\cite{EckardtPhysRevLett.123.030602}. Of particular focus was the decay of imbalance. A slow stretched exponential ($A \exp(-t^{\alpha})$) decay of $I(t)$ has been reported\cite{AltmanPhysRevLett.116.160401,CaiBarthelPhysRevLett.111.150403,CaiPhysRevLett.124.130602,SchneiderPhysRevB.106.134211,PhysRevLett.116.237203Garrahan}, though the value of $\alpha$ obtained has some dispute. For example, a short theoretical analysis in Ref.~\onlinecite{CaiPhysRevLett.124.130602} finds $\alpha\sim 0.33$ whereas a different analysis in Ref.~\onlinecite{AltmanPhysRevLett.116.160401} puts $\alpha\sim 0.5$. Numerical fits in Refs.~\onlinecite{AltmanPhysRevLett.116.160401} puts $\alpha \sim 0.38$ and Refs.~\onlinecite{SchneiderPhysRevB.106.134211,PhysRevLett.116.237203Garrahan} put $\alpha \sim 0.42$. Additionally, the analytical expressions suggested usually require at least one fitting parameter to be matched with the numerical results.
In our work, we seek to understand and resolve this inconsistency. Upon carefully analyzing the relaxation, we see that several time scales emerge. Broadly speaking one has a regime where off-diagonal matrix elements of $\rho(t)$ are large, then a regime in which $\rho(t)$ becomes increasingly diagonal due to dephasing and the scaling variable is~\cite{MariyaMarkoPhysRevB.93.094205} $\tau=8\gamma t/W^2$, and finally a regime where the system starts to feel its finite size $L$. In fact, we argue that the `stretched exponential' is not a true (asymptotic) description for the relaxation. It holds only in a finite window, and is universal and independent of disorder type just in a linear-expansion regime, after which a non-generic disorder-dependent behaviour follows. Additionally, beyond a timescale, $\tau=8 \gamma t/W^2\sim1$, we show that the relaxation smoothly changes to $1/(L\sqrt{t})$ from the non-generic behaviour in open boundary conditions, which is also the standard result for clean systems with dephasing. Finally, for finite size systems, an exponential decay occurs at the longest timescale, governed by the finite Liouvillian gap.
Using a toy model involving three lattice sites, we find an expression of $I(\tau)$ analytically, using second order perturbation theory, that matches with numerics very accurately, till $\tau\sim 1$, without any fitting parameters. From the model, we show that the universal behaviour $I(\tau)=(1-\beta \sqrt{\tau})$ occurs for very short times $\tau\ll 1$, beyond which the non universal behaviour follows till $\tau\sim 1$.
%After that non-local relaxation causes power law decay and . Certain nuances regarding boundary conditions are also involved during relaxation, which we shall discuss in some detail along with an analysis of Liouvllian spectrum of such systems.
We first provide a brief description of our main results in Sec.~\ref{summary} before going into the details later.
\section{Summary of Results}
\label{summary}
We begin by summarizing our findings. Starting from the N\'eel initial state, in the XX model with on-site disorder and dephasing, the behaviour of $-\ln I(t)$ can be schematically represented by Fig.~\ref{fig1}(a), where $\ln$ denotes the natural logarithm, $\gamma$ is the strength of dephasing, $W$ is the strength of disorder and $J$ is the exchange interaction strength of XX model. We use open boundary conditions (obc) unless specified.
\begin{figure}
\centering
\includegraphics[width=0.85 \columnwidth]{plots/schematic.pdf}
\includegraphics[width=0.9 \columnwidth]{plots/fig1b.pdf}
\caption{(a) A schematic figure denoting the five timescales of imbalance decay for an initial N\'eel state. (a) shows schematic diagram with timescales $t_0\sim {\rm Min}(1/\gamma,1/W)$, $t_1\sim 0.1 W^2/(8 \gamma J^2 )$, $t_2 \sim W^2/(8 \gamma J^2 )$ and $t_3 \sim L^2$, whereas in (b) disorder averaged exact numerical data for different system sizes $L$ is plotted with $\tau=8 \gamma t/W^2$ and $J=1$, while $\beta$ is obtained from Taylor expansion of Eq.~\ref{box} with $\gamma=1$ and $W=8$. The green vertical lines denote the timestamps of the schematic diagram for $L=50$ (open boundary conditions are used).
}
\label{fig1}
\end{figure}
As evident from the figure, five timescales emerge during the evolution of $I(t)$. They are,
\begin{enumerate}[I.]
\item $t < t_0 \sim {\rm Min} (1/\gamma,1/W)$: The shortest timescale in the problem. Here, $ I(t) \sim 1-t^2$ and hence $-\ln I(t) \sim t^2$. During this time period, off diagonal correlations first develop in the system and then start decaying after reaching a maxima.\footnote{Recall that since we start from the N\'eel state, a computational state, and hence at $t=0$ the system is completely uncorrelated.} When the disorder strength $W$ is small compared to the dephasing $\gamma$, $t_0$ is given by $\sim 1/\gamma$, akin to clean systems. However when $W \gg \gamma$, then this behaviour exists till $t \sim 1/W$.
\item $t_0 < t < t_1 \sim 0.1 W^2/(8 \gamma J^2 )$: This is where local relaxation of $\sigma^z$ starts, and most (but not all) off-diagonal correlations become negligible. In this timescale one sees $-{\rm \ln} I(t)\sim -{\rm \ln}(1-\beta \sqrt{t}) \sim \beta \sqrt{t}$ behaviour irrespective of the nature of disorder chosen ($\beta$ can depend on nature of disorder). This is the short time, linear regime of the stretched exponential behaviour noted in previous works\cite{AltmanPhysRevLett.116.160401,CaiBarthelPhysRevLett.111.150403,CaiPhysRevLett.124.130602,SchneiderPhysRevB.106.134211,PhysRevLett.116.237203Garrahan}. Unlike previous results, we find that this regime is not asymptotic and is only valid till $\tau\sim\tau_1=(8 \gamma J^2 t_1)/W^2 = 0.1$. Therefore, the ``stretched exponential'' is not really a true stretched exponential decay as it does not hold asymptotically, i.e., at arbitrarily small values of $I$.
\item $ t_1<t<t_2\sim W^2/( 8 \gamma J^2)$: In this regime, which holds till $\tau\sim\tau_2= (8 \gamma J^2 t_2)/W^2 = 1$ a non-generic decay dependent on the choice of disorder distribution is seen. $t_2$ also marks the end of local relaxation in the system.
\item $t_2<t< t_3$: In this regime, the decay smoothly changes to a power law due to the bunching of low magnitude eigenvalues of Liouvillian.\cite{CaiBarthelPhysRevLett.111.150403,MedvKehrinPhysRevB.90.205410} This is the asymptotic behaviour in the thermodynamic limit for finite $W$. Here, $I(t)\propto 1/(L\sqrt{t})$. The exponent -- $1/2$ in our case -- can depend on boundary conditions. The change of the nature of decay from Region III to IV occurs slowly at a timescale $\propto L$, not shown in the schematic diagram. For finite systems, due to the discrete nature of the Liouvillian spectrum, this behaviour is not asymptotic and there is another timescale.
\item $t> t_3 \sim L^2$: The final timescale in finite sized systems is governed by the Liouvillian gap, i.e. the largest real non zero eigenvalue of the Liouvillian (since it has an eigenvalue of zero). Here the decay is given by $I(t)\sim \exp(-|\lambda_1| t)$ where $\lambda_1$ is the Liouvilian gap. This timescale begins around $t_3 \sim 1/|\lambda_1|$. As will be discussed in Sec.~\ref{region5}, $\lambda_1\sim 1/L^2$, hence this regime starts around $t_3 \sim L^2$.
\end{enumerate}
It is worth noting that till $t \sim t_2$, i.e. till Region III, there is no system size dependence of the results, neither the time windows nor the behaviour depends on $L$. However, both the behaviour of $I(t)$ and the time span of Region IV is dependent on $L$. This feature is clearly visible in Fig.~\ref{fig1}(b) where we plot numerical results obtained for evolution of $-{\rm \ln} I(t)$ for a disorder strength of $W=8$ for different system sizes. We perform disorder averaging over $10^2$ realizations for $L=50,500$ and $10$ for $L=3000$. We also mark the various timescales in this plot, now with respect to $\tau$. Clearly, for $\tau>1$, different system sizes start showing different behaviour.
Since till $\tau\sim 1$, the rescaled data for the different system sizes clearly overlap with one another and is almost linear in $\ln-\ln$ scale, this prompted the stretched exponential fits in previous works. In Fig.~\ref{fig1}(b), we have added a black dashed line showing the stretched exponential expression obtained from a theoretical computation in Sec.~\ref{region23}. As will be clear from the analysis in that section, this behaviour is expected to hold best till small $\tau \sim 0.1$, which is what is demonstrated in the plot. \footnote{Hence numerical fits of the form $\exp(-t^{\alpha})$ in Refs. \onlinecite{AltmanPhysRevLett.116.160401,CaiBarthelPhysRevLett.111.150403,CaiPhysRevLett.124.130602,SchneiderPhysRevB.106.134211} found different $\alpha$s depending on fit windows as the behaviour is not a true stretched exponential. } Finally for $L=50$ one sees a fast growth of $-\ln I(t)$ (see a slight bend upwards) at large times, which is due to the exponential decay of $I(t)$ from the finite size Liouvillian gap.
In what follows, we shall discuss the above findings in details and provide a theoretical understanding for the same. In the next section, we describe the model in more details and elaborate on the technique used to compute the numerical results. Then in Sec.~\ref{timescales} we elaborate on the behaviour of $I(t)$ in different timescales. Finally, in Sec.~\ref{discussion} we provide some final remarks on our results and possible extensions of this work.
\section{Model}
\label{model}
We take the one dimensional disordered XX chain with spin 1/2 particles,
\begin{equation}
H=-J\sum_{j=1}^{L-1}(\sigma^x_j\sigma^x_{j+1}+\sigma^y_j\sigma^y_{j+1})+\sum_{j=1}^L h_j\sigma^z_j
\label{lind1}
\end{equation}
where $h_j$ are the on-site disorders and $\sigma^{x,y,z}$ denotes the Pauli matrices. In this work, unless otherwise mentioned, $h_j$s are chosen from a uniform distribution in $(-W,W)$, $W$ being the strength of disorder. We set $J=1$ for the rest of our work. To simulate an open system, each spin is exposed to a dephasing. The evolution of the system's density matrix is given by,
\begin{equation}
\frac{{\rm d}\rho}{{\rm d} t}=i [\rho,H]+\mathcal{L}^{{\rm deph}}(\rho)=\mathcal{L}(\rho)
\end{equation}
The non-unitary part can be written in terms of Lindblad operators as, $\mathcal{L}^{{\rm deph}}(\rho)=\sum_k([L_k \rho,L_k^{\dagger}]+[L_k, \rho L_k^{\dagger}])$. We choose $L_k=\sqrt{\frac{\gamma_k}{2}}\sigma_k^z$ to represent the on site dephasing term. For the rest of this work we shall put $\gamma_k=\gamma$, i.e. have a spatially uniform dephasing. This choice of dephasing causes an exponential decay of the off diagonal elements of the density matrix with a strength $\gamma$, in the diagonal basis of $\sigma^z$ in absence of disorder.
If we want to solve Eq.~\ref{lind1} directly, numerically or otherwise, we need to solve a set of $4^L-1$ coupled differential equations. Fortunately due to the choice of dephasing and the quadratic nature of the Hamiltonian, the exponentially many equations can be decoupled into blocks of polynomial complexity. \cite{Znidaric2013,Znidaric_2010,PolettiPhysRevA.98.052126,Eisler_2011,Temme_2012,CiracPhysRevA.87.012108,arxiv.2210.10856,SchiroPhysRevB.104.144301} In other words, observables follow a hierarchy based on the number of Fermionic operators they contain. For example, the block of two point correlators decouples from the rest, and one can write a closed set of equations for these observables. Then this solution serves as a source term for the three point correlations and so on. Since we are interested in imbalance, which can be extracted from a two-point correlator in the Fermionic language $(\sigma^z \sim c^{\dagger} c)$, we will just consider the subspace of two point correlation functions. Following the consideration of. Ref.~\onlinecite{Znidaric_2010,Znidaric2013}, we define the operators,
\begin{eqnarray}
A(t)&=&\sum_{r=1}^L \sum_{j=1}^{L+1-r}a_j^{(r)}(t)A_j^{(r)}\nonumber \\
B(t)&=&\sum_{r=2}^L \sum_{j=1}^{L+1-r}b_j^{(r)}(t)B_j^{(r)},
\label{hierarchy}
\end{eqnarray}
where $A_j^{(r+1)}=\sigma_x^j Z_{j+1}^{(r-1)} \sigma_x^{j+r}+\sigma_y^j Z_{j+1}^{(r-1)} \sigma_y^{j+r}$ and $B_j^{(r+1)}=\sigma_x^j Z_{j+1}^{(r-1)} \sigma_y^{j+r}-\sigma_y^j Z_{j+1}^{(r-1)} \sigma_x^{j+r}$ for $r\ge 2$. $Z_j^{(r)}=\sigma_z^j\hdots \sigma_z^{j+r-1}$ are strings of $\sigma_z$ operators, and $A_j^{(1)}=-\sigma_z^j$. Then the equation governing the time evolution of the set of two point correlation functions can be written compactly as,
\begin{equation}
\frac{{\rm d}\mathbf{C}(t)}{{\rm d} t}+2 {\rm i} (\mathbf{P}\mathbf{C}(t)-\mathbf{C}(t)\mathbf{P}^T)+2(\mathbf{\Gamma}\tilde{\mathbf{C}}(t)+\tilde{\mathbf{C}}(t)\mathbf{\Gamma})=0,
\label{diffeq}
\end{equation}
where $\mathbf{C},\tilde{\mathbf{C}},\mathbf{P},\mathbf{\Gamma}$ are $L \times L$ matrices. Their elements are defined as, $C_{jk}(t)=a_j^{(k-j+1)}(t)+{\rm i} b_j^{(k-j+1)}(t)$ for $k>j$, $C_{jj}(t)=a_j^{(1)}(t)$ and $C_{jk}=C_{kj}^*$. $\tilde{\mathbf{C}}=\mathbf{C}-{\rm diag} (\mathbf{C})$. $\mathbf{P}=\mathbf{W}-\mathbf{T}$ where, $W_{jk}=h_k \delta_{jk}$, the on-site disorders and for our model. Also for our model, $T_{jk}=J(\delta_{j,k-1}+\delta_{j,k+1})$, as we consider only nearest neighbour couplings \footnote{For Fermionic Hamiltonians, if longer range couplings are present then in the most general case, $T_{jk}=J_{jk}$ where $J_{jk}$ denotes the hopping strength between two different sites at positions $j$ and $k$.} and $\Gamma_j^k=\gamma \delta_{jk}$. Clearly, $I(t)=(1/L)\sum_j(-1)^{j+1}C_{jj}(t)$. Hence for the N\'eel initial state $C_{jj}(0)=(-1)^{j+1} $ and $I(0)=1$.
Eq.~\ref{diffeq} can be recast into a linear differential equation with $L^2$ variables, in the form,
\begin{equation}
\frac{d\mathbf{f}}{dt}=\mathcal{Q} \mathbf{f}
\label{linearised}
\end{equation}
where $\mathbf{f}=(C_{11},C_{12},\hdots,C_{1L},C_{21},\hdots,C_{LL})$ and $\mathcal{Q}$ is the $L^2 \times L^2$ matrix governing the evolution. In fact, due to the hierarchical structure of observables, the eigenvalues of $\mathcal{Q}$ are exactly the eigenvalues of the one particle sector of the Liouvillian, $\mathcal{L}$, while the eigenvectors of both are connected via an appropriate rotation. This linearised equation will be useful in understanding the behaviour of $I(t)$ in different regimes.
Unlike most of the previous results for this model which are generated by either an effective Hamiltonian or DMRG based techniques \cite{AltmanPhysRevLett.116.160401,CaiPhysRevLett.124.130602}, we use Eq.~\ref{diffeq} or Eq.~\ref{linearised} for the analysis that follows. This not only allows us to obtain an exact description of the correlators but also provides access to sizes and times an order of magnitude larger than usually accessed by DMRG techniques. Thus, we can make better statements about system size dependence of the results.
\section{Different timescales}
\label{timescales}
Let us now discuss each timescale in more details.
\subsection{Region I}
\label{reg1}
\begin{figure}
\includegraphics[width=0.75 \columnwidth]{plots/ultrashorttime.pdf}
\caption{Figure showing the initial $t^2$ behaviour of $-\ln I(t)$ in region I for different disorder strengths $W$.}
\label{fig2}
\end{figure}
Doing a simple power series expansion of $I(t)$ in $t$, the first non-zero term turns out to be quadratic. Hence, we expect $I(t)\sim 1-\eta t^2$ behaviour at the smallest timescale. As evident from Fig.~\ref{fig2}, this is the case. This behaviour continues till time $t_0$ which is dependent on $\gamma$ and $W$. To be more precise, let us first consider the simpler case, $W\rightarrow 0$. At $t=0$, since we start from a pure state, $\rho(t=0)$ has only one non zero element, which is in the diagonal, and $C(t=0)$ is diagonal as well. Correlations then spread rapidly throughout the system, reach a maxima and start decaying around $t \sim 1/\sqrt{|\gamma^2-4|} \sim 1/\gamma$ for $\gamma \gg 1$. As we see from Fig.~\ref{fig2}, indeed below $t_0 \sim 0.5$ for $\gamma=1$ and $W=0$, one can approximate $I(t)$ as $1-\eta t^2$, and hence $-\ln I(t)\sim \eta t^2$. From the numerical fit in Fig.~\ref{fig2} shown by the black dashed line, we find $\eta \sim 16$. This behaviour can also be qualitatively extracted from a simple two-site model (details in Appendix~\ref{appA1}), and we get, for small $t$, $ I(t) \sim 1-8 t^2 +O(t^3)$. One can then numerically check that on addition of a few more sites to the system $\eta$ approaches $16$ rapidly.
However, on addition of disorder, $t_0$ reduces with increasing $W$, as can be seen Fig.~\ref{fig2}. This can be qualitatively understood from Eq.~\ref{diffeq}, where we see that correlations can develop in the system either due to $\mathbf{P}$, i.e. the terms involving disorder or due to $\mathbf{\Gamma}$ which involves dephasing. The dominant term would then determine the smallest timescale. Also, it can be shown using the two-state model again that, for $W \gg \gamma$, the $t_0$ now becomes $1/\delta$, $\delta$ being the difference of on-site disorders in the two-site problem. Hence several timescales emerge due to different $\delta$s at different lattice sites, and we will get $t_0 \ll 1/\gamma$. If we approximate the width of the distribution of $\delta$s as $W$, we can say that $t_0$ is around $ {\rm Min}(1/\gamma,1/W)$. This is Region I.
\begin{figure*}
\includegraphics[width=0.9 \columnwidth]{plots/parametertheorycompare2.pdf}
\includegraphics[width=0.9 \columnwidth]{plots/distributiontheorycompare.pdf}
\caption{ Comparison of analytical predictions with exact numerics in Regions II and III. (a) Plot showing how the rescaled time causes complete collapse of data with different $L,\gamma, W$, and agreement of the said numerical results for the box distributed random on site disorder, with the theoretical predictions. $f(\tau)$ is defined in Eq.~\ref{firstapprox} and $g(\tau)$ is defined in Eq.~\ref{box}. (b) Plot showing agreement of our theoretical prediction with numerical results derived from for two other disorder distributions, viz. gaussian random and alternating disorder, see Eq.~\ref{gaussian} and Eq.~\ref{alternating} respectively. $f_G(\tau)$ is obtained from Eq.~\ref{firstapprox} where the random numbers are picked from a Gaussian distribution.}
\label{fig3}
\end{figure*}
\subsection{Regions II and III}
\label{region23}
Next, we shall understand the decay of imbalance in Regions II and III. Local evolution is dominant in these regimes and we can use the same analysis to describe both. Hence we discuss them together. A timescale separating the two regimes will emerge naturally from the computation that follows.
We consider $W \gg \gamma$, a regime where localization effects would be prominent in the system. Then, to understand the evolution of $\langle\sigma_j^z \rangle$ for site $j$, we need to take into account the influence of the neighbouring sites, $j-1$ and $j+1$. Matrix $\mathbf{C}$ in Eq.~\ref{hierarchy} for this $3$-site system, becomes a $3 \times 3$ matrix. We further ignore $C_{ij}$s where $|j-i|>1$, as they are negligible compared to the rest in this timescale. We can then use Eq.~\ref{linearised} with
\begin{equation}
\mathbf{f}=\begin{pmatrix}
C_{11}, & C_{12}, & C_{21}, & C_{22}, & C_{23}, & C_{32}, & C_{33},
\end{pmatrix}
\label{threesitef}
\end{equation}
where we have taken $j=2$ without loss of generality. The form of $\mathcal{Q}$ for this system in Eq.~\ref{linearised} is given in Appendix~\ref{appA2}. We use second order perturbation theory to find the eigenvalues as outlined in Appendix~\ref{appA2}, since exact analytical diagonalization is not tractable. We also just need to diagonalize the subspace where we have the smaller modulus eigenvalues, as the larger modulus eigenvalues dictate the behaviour in Regime I. Doing so, we obtain the relevant eigenvalues as given in Eq.~\ref{eigenvalues} in Appendix.~\ref{appA2}. These in the regime $ W \gg \gamma$ can be approximated as $\Delta^{(0)}=0$ and,
\begin{equation}
\Delta^{(\pm)}\sim\frac{8 \gamma (\delta_1^2+\delta_2^2\pm\sqrt{\delta_1^4+\delta_2^4-\delta_1^2\delta_2^2})}{\delta_1^2 \delta_2^2},
\label{approxeigen}
\end{equation}
where $\delta_1=|h_{1}-h_2|$ and $\delta_2=|h_2-h_{3}|$.
Hence the long time behaviour of $C_{22}(t)=-\langle\sigma_2^z(t)\rangle$, can be generically written as,
\begin{equation}
C_{22}(t)=\sum_{m=\pm,0}d^{(m)}_2 \exp(-\Delta^{(m)} t)
\label{C22three}
\end{equation}
where $d^{(m)}_2$ is the corresponding element of the $m^{th}$ eigenvector, weighed by the factors arising from the initial conditions. We shall drop the subscript $2$ in what follows.
Further simplification of Eq.~\ref{C22three} can be made by observing that typically (see Appendix~\ref{appA2}) $d^{(+)}$ is the largest coefficient, hence the principal mode of relaxation is $\Delta^{(+)}$. The contribution from $\Delta^{(0)}$ and $\Delta^{(-)}$ modes weighted by $d^{(0)}$ and $d^{(-)}$ respectively effectively act as perturbations around the principal mode\footnote{This is difficult to rigorously prove analytically, but can be seen numerically} (see Appendix~\ref{appA2}). In the random disorder case since $\Delta^{(+)}$s are different for different sites, we reintroduce the subscript $j$, and write for the N\'eel initial state, $ I(t)=(1/L)\sum_{j=1}^L|C_j(t)| \sim(1/L)\sum_{j=1}^L |d_j^{(+)}| e^{-\Delta^{(+)}_j t}$, where the perturbative corrections due to the other modes can be neglected due to averaging.
Additionally, since $d_j^{(+)}$s are $\mathcal{O}(1)$(see Appendix~\ref{appA2}), it is not essential to keep track of them. Hence, given a disorder distribution one can calculate the evolution of imbalance for the N\'eel initial state in the thermodynamic limit as,
\begin{equation}
I(t)= f(t,W,\gamma)=\lim_{L \to \infty}(1/L)\sum_{j=1}^L e^{-\Delta^{(+)}_j t},
\label{firstapprox}
\end{equation}
where different $\Delta^{(+)}_j$s are obtained from the distribution of $h_j$s, using Eq.~\ref{approxeigen}. From Eq.~\ref{firstapprox} we can see that in this timescale, $I(t)$ is the average over the decay of magnetization at individual sites with effective decay rates at each site controlled by the disorders in its nearest neighbours. From Eq.~\ref{approxeigen}, we also see an emergent energy scale $8 \gamma/W^2$ (recall $J=1$), which gives us the rescaled time $\tau=8 \gamma t/W^2$.
In Fig.~\ref{fig3}(a) we show comparison between exact numerical results with Eq.~\ref{firstapprox} and other approximations described later in the section. The plots for different system sizes $L$, disorder strengths $W$ and dephasing $\gamma$ obtained via exact numerics collapse on each other when we use rescaled $\tau=8 \gamma t/W^2$, confirming the presence of the universal timescale\cite{MariyaMarkoPhysRevB.93.094205}. As shown in Fig.~\ref{fig3}(a) with the black dashed line, computing Eq.~\ref{firstapprox} by sampling a large number of random numbers from the box disorder distribution, we get a result that is a very accurate match with exact numerics till $\tau \sim 1$. Note how a numerical result obtained via solving many coupled differential equations can be replicated by correctly sampling the underlying disorder distribution, due to locality of the evolution.
However, it is difficult to perform the summation analytically to obtain a closed form expression for different disorder distributions. To proceed further we need to go to the continuum limit and write Eq.~\ref{firstapprox} as $I(t)=\int d \Delta^{(+)} p(\Delta^{(+)}) e^{-\Delta^{(+)}t}$, $ p(\Delta^{(+)})$ being the probability distribution of $\Delta^{(+)}$. Even so, computing $p(\Delta^{(+)})$ analytically is a very difficult task. Hence, we need to make further simplifying assumptions. If we approximate $ \delta_1^4+\delta_2^2-\delta_1^2\delta_2^2\sim(\delta_1^2+\delta_2^2)^2$ and hence $\Delta^{(+)} \sim 16\gamma ( \frac{1}{\delta_1^2}+\frac{1}{\delta_2^2})$, the integral becomes tractable. This approximation is valid in the limit of $|\delta_1^2- \delta_2^2| \gg 0$, where we have $\delta_1^2\delta_2^2 \ll \delta_1^4+\delta_2^4$ and hence works in the large $\Delta^{(+)}$ tail of $p(\Delta^{(+)})$ very well, see Fig.~\ref{fig6a} in Appendix.~\ref{appA2}. \footnote{This choice of approximation is used to obtain a tractable form for $p(\Delta^{(+)})$. Other choices are also possible in the said limit, but they do not simplify the calculation. We show a comparison of the probability distribution of the approximated and exact $\Delta^{(+)}$ in Appendix~\ref{appA2}}. Then, we consider that $\delta_1$ and $\delta_2$ are independently drawn from different sets of random numbers to avoid calculating the complicated convolution term. This allows us to write,
\begin{equation}
I(t)=\left(\int d\delta p(\delta) e^{-16 \gamma t/\delta^2}\right)^2
\label{master1}
\end{equation}
where $p(\delta)$ is the probability distribution of $\delta$s. Eq.~\ref{master1} allows us to compute the evolution of $I(t)$ analytically for any given disorder distribution. For example, for a disorder distributed uniformly in $(-W,W)$, i.e. the box distribution, Eq.~\ref{master1} gives the result in terms of known mathematical functions as,
\begin{eqnarray}
I(\tau=8 \gamma t/W^2)=g(\tau)&=&[-\sqrt{2 \pi } \sqrt{\tau } \text{erfc}\left(\sqrt{\frac{\tau }{2}}\right)\nonumber \\
&&+e^{-\frac{\tau }{2}}+\frac{1}{2} \tau \Gamma \left(0,\frac{\tau }{2}\right)]^2
\label{box}
\end{eqnarray}
where $\text{erfc(z)}=1-\text{erf}(z)$, ${\rm erf}$ denotes the Error function and $\Gamma$ is the incomplete gamma function defined by $\Gamma(0,z)=\int_z^{\infty}e^{- t}/t dt$. The green dashed line in Fig.~\ref{fig3}(a), obtained from Eq.~\ref{box} shows good agreement till $\tau \sim 0.1$, which is expected since our approximation works best in the large $\Delta^{(+)}$ tail.
\paragraph*{\textbf{$1-\beta \sqrt{t}$ scaling regime:}} Actually, in the regime of $\tau \ll 1$ or $t \ll W^2$, Taylor expanding Eq.~\ref{box} in $\tau$, we get $I(\tau)=1-\beta\sqrt{\tau} +O(\tau)$ with $\beta=\sqrt{8\pi}$, which is exactly the linear term in the stretched exponential\cite{CaiBarthelPhysRevLett.111.150403} $e^{-\beta \sqrt{\tau}}$. In Fig.~\ref{fig3}(a), we plot $-\ln I(\tau)\sim \beta \sqrt{\tau}$ as the brown dashed line, and it also agrees well with exact numerics till $\tau \sim 0.1$.
Generally, if $p(\delta)$ is an analytic function of $\delta$, then it can be expanded as $p(\delta)=\frac{1}{\mathcal{N}}(c+O(\delta))$, where $\mathcal{N}=\int d\delta (c+O(\delta)) $. Since $\int d\delta e^{-16 t/\delta^2}\sim t^{1/2}+O(t)$, the lowest order term obtained from $p(\delta)$ in this form would be $\sqrt{t}$ (for $c=0$ the result would be different). For many different disorder distributions this is a good approximation of results till higher order terms become important with larger $t$.
To highlight this special scaling, we label the regime where the linear approximation of $\beta \sqrt{\tau}$ is valid as Region II which is approximately till $\tau=0.1$, and denote the non-generic regime during $0.1\lesssim\tau \lesssim 1$ as Region III.
\paragraph*{\textbf{Other disorder distributions:}} To further consolidate our claims, we repeat the above calculations for other disorder distributions, viz. the Gaussian random distribution and staggered on-site or alternating potential, shown in Fig.~\ref{fig3}(b). We have also checked the quasiperiodic Aubre Andre distribution, data not shown to avoid cluttering. In Fig.~\ref{fig3}(b), results from Eq.~\ref{firstapprox} is represented by the green dashed line for the Gaussian disorder case, and the black dashed line for the alternating potential case. In both cases (and the qausiperiodic one not shown), we see that Eq.~\ref{firstapprox}, obtained from the second order perturbation result of the $3$-site model, captures the nuances in the evolution very well till $\tau=1$ (longer for alternating potential see Appendix~\ref{appA2}). In general, our analysis is valid for any disorder distribution where typical $|h_j-h_{j+1}| \gg0$, to make the second order perturbation theory hold.\footnote{One can still numerically solve the small size Hamiltonians to obtain the eigenenergies non-perturbatively and continue with the analysis.}
For the Gaussian distribution it is also possible to compute Eq.~\ref{master1} analytically and the result is,
\begin{eqnarray}
I(\tau)=&e^{- \sqrt{8\tau}},
\label{gaussian}
\end{eqnarray}
a different result to Eq.~\ref{box} and exactly the stretched exponential\footnote{Ref.~\onlinecite{AltmanPhysRevLett.116.160401} arrived at this result using a different approach effectively leading to the same integral. However, under their approximations one arrives at this result irrespective of the disorder chosen and then needs to fix a free parameter via numerical fitting to fit to different disorders}. This is shown as the purple dot-dashed line in Fig.~\ref{fig3}(b) which we again see to be valid till $\tau\sim 0.1$, thus showing a separation of timescales similar to box distribution. Other standard disorder distributions such as the exponential and student's T distribution also show similar timescale separation.
Finally, for the staggered potential, i.e when $h_j=(-1)^j W$, Eq.~\ref{firstapprox} has a simple form. Since we have $\delta_1=-\delta_2=\delta/2= W$, $\Delta^{(+)}=6 \gamma/W^2$ for all sites. In this case the weight of the $\Delta^{(-)}$ mode, $d^{(-)}=0$, hence $I(t)\sim d^{(0)}+d^{(+)} e^{-\Delta^{(+)} t}$. For large $W$, this can be approximated as $f_{{\rm alt}}=e^{-\Delta_{{\rm eff}} t}$, where $\Delta_{{\rm eff}}=\frac{8 \gamma}{W^2}$ and hence
\begin{equation}
I(\tau) \sim e^{-\tau}.
\label{alternating}
\end{equation} See Appendix~\ref{appA2} for more details. In Fig.~\ref{fig3}(b), we see exactly the expected behaviour. The almost perfect agreement between our theoretical predictions with the numerical data provides a good benchmark about the generality of our theory.
\paragraph*{\textbf{Effective Hamiltonian approach:}} Before we end this section we make a digression to briefly discuss an alternative approach. In Ref. \onlinecite{MariyaMarkoPhysRevB.93.094205} it was shown that, for the interacting XX chain with dephasing and strong disorder, beyond a timescale given by $1/\gamma$, off-diagonal terms of the density matrix are negligible and the evolution of the diagonal terms of the density matrix is given by the differential equation, $\frac{d \bm{\rho}_D}{dt}=-\mathcal{H}_{\rm eff}\bm{\rho}_D$, where $\bm{\rho}_D$ denotes the diagonal part of $\bm{\rho}$ and
\begin{equation}
\mathcal{H}_{\rm eff}=\sum_{j=1}^L 2 \frac{(\gamma_j+\gamma_{j+1})}{(h_j-h_{j+1})^2+(\gamma_j+\gamma_{j+1})^2}(1-\bm{\sigma}_j.\bm{\sigma}_{j+1}),
\label{effham}
\end{equation}
where $\bm{\sigma}=(\sigma^x,\sigma^y,\sigma^z)$. An emergent rescaled time, $ 4\gamma t/(W^2+4 \gamma^2)$ is seen the overall prefactor of $\mathcal{H}_{\rm eff}$, if we consider $\gamma_{j}=\gamma$ and $\delta_j=h_j-h_{j+1}\sim W$, reminiscent of the one we extracted from the three site model, since this is also a second order perturbative description. In our non-interacting XX model, this effective description still holds as there is no term involving any interaction in Eq.~\ref{effham}. Beyond a cutoff timescale, which for $L=24$ and $W=30$ is given by $\tau \sim 0.01$ (check Appendix~\ref{appB}) the agreement of exact numerics and evolution of $I(t)$ with $\mathcal{H}_{\rm eff}$ is excellent. This means the effective Hamiltonian describes the system approximately for $\tau \gg \tau_0$ in our model ($\tau_0 \sim 0.001$ for $W=30$), while before that timescale the off diagonal elements of the density matrix still has significant contribution to the evolution. While, a direct use of $\mathcal{H}_{\rm eff}$ to repeat the analysis of this section to understand Region II and III are a bit involved, we shall use this effective Hamiltonian to explain the behaviour in Region IV and V in the following sections.
Finally, it is worth noting that For $W=0$, Regions II and III no longer exist and there is an oscillatory behaviour in $-{\rm \ln} I(t)$ as can be seen in Fig.~\ref{fig2}, with an envelope growing as $4 \gamma t$. For large $\gamma$ one can again apply the effective Heisenberg model to obtain these results for this system as in Ref.~\onlinecite{CaiBarthelPhysRevLett.111.150403}. Upon analysis via our three site model, we observe that the magnitude of the real part of eigenvalues involved in the evolution $\sim 4\gamma$ as the corresponding eigenvectors carry almost all the weight, a distinct shift from what happens in the disordered case, where lower magnitude eigenvalues carry most of the weight. Near the end of this time-scale the system begins to realize its non-local nature, and the evolution slowly changes to what is seen in Region IV, discussed below, irrespective of the presence of disorder.
\subsection{Region IV}
\label{region4}
\begin{table*}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Case & disorder & $\alpha$ & $\beta$& $I(t)$ &$-\ln I(t)$\\
\hline
pbc, even $L$ &$0$& $2$ &no overlap & $e^{-4 \gamma t}$& $4 \gamma$\\
\hline
pbc, odd $L$ &$0$& $2$ & 0 & $\frac{1}{L\sqrt{8 \pi t}}$ &$\frac{1}{2}\ln ( t L^2)+\frac{1}{2}\ln (8 \pi)$\\
\hline
obc, any $L$ &$0$& $2$ & 0 & $\frac{1}{L\sqrt{8 \pi t}}$ &$\frac{1}{2}\ln ( t L^2)+\frac{1}{2}\ln (8 \pi)$\\
\hline
obc, any $L$ &$W$& $2$ & 0 & $\frac{\sqrt{\pi}}{2 L\sqrt{a t}}$&$\frac{1}{2}\ln (tL^2)+\frac{1}{2}\ln \frac{4a}{\pi}$\\
\hline
pbc, any $L$ &$W$& $2$ & 2 & $\frac{\sqrt{\pi}}{2 (at)^{3/2}}$ &$\frac{3}{2}\ln t+\frac{1}{2}\ln \frac{4 a^3}{\pi}$\\
\hline
\end{tabular}
\caption{Table showing the decay of $I(t)$ in Regime IV for different cases, $\alpha$ and $\beta$ are exponents defined in Eq.~\ref{alphabeta}. For obc, the scaling of $I(t)$ with $t$ do not qualitatively change on addition of disorder, whereas for pbc they are remarkably different. See text for details}
\label{table1}
\end{table*}
In this section, we shall discuss the behaviour of $I(t)$ in the fourth time-window, where we observe an asymptotic power law decay irrespective of the nature of disorder distribution. This is the first regime where we see $L$ dependent behaviour, as evident from Fig.~\ref{fig1}(b), where the evolution of $-\ln I(t)$ is plotted for different $L$s with open boundary conditions. Additionally, unlike the previous regimes which did not depend on the boundary conditions, behaviour of $I(t)$ in Region IV is strongly dependent on such factors. As shown in Fig.~\ref{fig4}(c), for open boundary conditions, we see a $I(t)\propto 1/(L\sqrt{t})$ or $-\ln I(t)\sim \frac{1}{2}\ln (tL^2)$ behaviour(dependence on $\gamma$ is more complicated, discussed later in the section). However for periodic boundary conditions (pbc), as plotted in Fig.~\ref{fig4}(d), $-\ln I(t)\sim \frac{3}{2}\ln t$, show a completely different behaviour with a different exponent and no system size dependence!
To explain this behaviour, we first realize that in this regime charge is transported on longer scales and the system can no longer be described by three site models of the previous section. Hence, we need to study the eigenspectrum of the Liouvillian. Additionally, we need to focus on the low magnitude eigenvalues as this regime is asymptotic. However, since we have the hierarchical structure of observables, we do not need to study the full $4^L \times 4^L$ Liouvillian, but the eigenspectrum of $\mathcal{Q}$ in Eq.~\ref{linearised}, which, as mentioned earlier, is related to the one-particle Liouvillian of the problem. We can write the solution of Eq.~\ref{linearised} generically as,
\begin{equation}
\mathbf{f}(t)=\mathbf{f}_{{\rm NESS}}+\sum_j \mathbf{\delta f}(j) e^{(\lambda_j+i \Omega_j) t},
\label{rhoness}
\end{equation}
where $\mathbf{f}$ is defined under Eq.~\ref{linearised}, $\mathbf{f}_{{\rm NESS}}$ is the steady state value of the observables, $\lambda_j$ and $\Omega_j$ are the real and imaginary parts of the eigenvalues of $\mathcal{Q}$ and $\mathbf{\delta f}(j)$ contains the corresponding eigenvectors weighted by the initial conditions. Typically, the weights are similar to the corresponding elements of the eigenvector, i.e. $\mathbf{\delta f}(j) \sim m_j^2$ where $m_j$ is the $j^{{\rm th}}$ eigenvector (See Appendix~\ref{appC} for more details).
From Eq.~\ref{rhoness}, it might naively seem that any observable should show exponential decay with different rates at different timescales, depending on $\lambda_j$s. However, in reality, in the thermodynamic limit the eigenspectrum becomes continuous, and this leads to a power law approach of observables towards NESS. To demonstrate this, assume without loss of generality that
\begin{equation}
\lambda_j\sim -j^\alpha, \hspace{0.2in} \mathbf{\delta f}(j) \sim j^{\beta}
\label{alphabeta}
\end{equation} and\footnote{$\Omega_j$ gives an oscillatory term which is usually $\sim 0$ for the relevant low magnitude eigenvalues of the Liouvillian spectrum. If they are included the power law reduces by 1\cite{MedvKehrinPhysRevB.90.205410}} $\Omega_j \ll \lambda_j$. Then Eq.~\ref{rhoness} can be written in the continuum limit as\cite{CaiPhysRevLett.124.130602,CaiBarthelPhysRevLett.111.150403,MedvKehrinPhysRevB.90.205410},
\begin{equation}
\mathbf{f}(t)-\mathbf{f}_{{\rm NESS}}\sim\int dj j^{\beta} e^{-j^\alpha t}\sim t^{-\frac{\beta+1}{\alpha}}.
\label{scaling}
\end{equation}
In what follows, we shall understand the behaviour of $I(t)$ for different systems using Eq.~\ref{rhoness}.
In Table~\ref{table1} we first summarize the results of various cases which shall be discussed in this section.
\begin{figure*}
\includegraphics[width=0.9 \columnwidth]{plots/pbcobceigenvalues.pdf}
\includegraphics[width=0.9 \columnwidth]{plots/pbcobceigenvectoroverlap.pdf}
\includegraphics[width=0.9 \columnwidth]{plots/inversesqrtt.pdf}
\includegraphics[width=0.9 \columnwidth]{plots/periodicpowerlaw.pdf}
\caption{Power law decay of $I(t)$ in Region IV. (a) Plot showing $j^2$ scaling of low magnitude Liouvillian eigenvalues averaged over $100$ disorder realizations compared with approximate theoretical prediction in Eq.~\ref{eigenvaluesapp}. (b) Plot showing scaling of $\mathcal{I}(j)$ defined in Eq.~\ref{support} averaged over $100$ disorder realizations. obc and pbc denote open and periodic boundary conditions respectively. See text for details. (c) Plots showing the scaling of $-\ln I(t)$ with $t$ for obc at different system sizes $L$. (d) Same as (c) but for periodic boundary conditions.}
\label{fig4}
\end{figure*}
Since $\mathbf{f}$ contains all the two particle Fermionic operators in the system, while $I(t)=\frac{1}{L}\sum_{k=1}^L(-1)^{k+1} C_{kk}(t)$, we can rewrite Eq.~\ref{rhoness} in the continuum for $I(t)$ as,
\begin{equation}
I(t)=I_{{\rm NESS}}+\int dj \mathcal{I}(j) e^{\lambda_j t}
\end{equation}
where
\begin{equation}
\mathcal{I}(j)=\frac{1}{L}\sum_{k=1}^L (-1)^{k+1} [\delta \mathbf{f}(j)]_{(k-1)L+k},
\label{support}
\end{equation} sums over the relevant operator subspace for the finite $L$ system. Eq.~\ref{scaling} gets appropriately modified.
Let us first look at the simpler case of systems without disorder. It is known that the low lying eigenenergies in the one-particle Liouvillian, in open boundary conditions are approximately \cite{MarkoPhysRevE.92.042143},
$\lambda_j=2 \left(\sqrt{4 \cos \left(\frac{ \pi j}{L}\right)-3}-1\right), j=0\hdots L-1$.
For periodic boundary conditions the expression becomes, $2 \left(\sqrt{4 \cos \left(\frac{ 2 \pi j}{L}\right)-3}-1\right), j=0\hdots L-1$, i.e. there is a degeneracy of $2$. Since we are interested in the regime $j/L \ll 1$, we can approximate obc eigenvalues as $\frac{2 \pi^2 j^2}{L^2 \gamma}$. For pbc this gets multiplied by a factor of $4$, and hence in both cases $\alpha=2$.
$\mathcal{I}(j)$ can be computed from the relevant elements of the eigenvectors of the one-particle Liouvillian. It turns out that the relevant elements of the $j^{{\rm th}}$ eigenvector can be approximated to be exactly the free fermion wavefunction with the corresponding boundary condition, i.e. for pbc, they are $\frac{1}{\sqrt{L}}\sum_{k=1}^{L}\exp(2 i \pi j k/L)$. Hence $\mathcal{I}(j)\sim\frac{1}{L}[\sum_{k=1}^{L}\frac{1}{\sqrt{L}} \cos(\pi k)\exp(2 i \pi j k/L)]^2$ is identically $0$ for even $L$ due to the symmetry of the wavefunction.\footnote{The square comes from the weightage of the initial conditions, whose weights are of the same order as the element of the eigenfunction. Hence we use $\sim$ instead of $=$.} Since $I$ has no support on the low magnitude eigenspectrum of the Liouvillian, it does not show a power law decay. Instead it decays exponentially with the the largest $\lambda_j=4 \gamma$.
The situation for odd $L$ in pbc is different as the one site is unpaired. In this case $\mathcal{I}(j)\sim\frac{1}{L}[\sum_{k=1}^{L} \frac{1}{\sqrt{L}} \cos(\pi k)\exp(2 i \pi j k/L)]^2 \sim \frac{1}{L^2}$ for $j \ll L$. This means $\beta=0$, and hence $I(t)\sim 2\int dj \frac{1}{L^2} e^{-8 \pi^2 j^2 /L^2}= \frac{1}{L\sqrt{8 \pi t}}$, where the factor $2$ in front comes from the degeneracy of the eigenvalues.
For obc the eigenvectors are given by, $\sqrt{\frac{2}{L}}\sum_{k=1}^{L}\cos( \pi j k/L)$. Hence $\mathcal{I}(j)=\frac{1}{L}[\sum_{k=1}^{L}\sqrt{\frac{2}{L}} \cos(\pi k)\cos( \pi j k/L)]^2$. This equals $0$ when both $L$ and $j$ are even or odd, and $\frac{2}{L^2}$ otherwise. This gives $I(t)\sim \int dj \frac{2}{ L^2} e^{-2 \pi^2 (2j)^2 /L^2}= \frac{1}{L\sqrt{8 \pi t}}$. These results for the clean systems have been matched with exact numerics, data not shown.
Now we focus our attention to disordered systems. Intuitively, we can predict that since the behaviour in this regime involves the low energy eigenspectrum of the Liouvillian, it should not drastically change on addition of disorder. If we look into the obc case, we indeed see $ \propto 1/\sqrt{t}$ behaviour for both clean and disordered systems, but not so for the pbc case where we see a $t^{-3/2}$ behaviour. Let us investigate why.
Unfortunately it is very difficult to find an analytical expression for the eigenspectrum when we add the disorder in the system, so we have to resort to exact diagonalization. However, we can make some crude approximations about the low lying eigenvalues and their scaling with $L,W$ and $\gamma$ since the evolution is described by the effective Heisenberg Hamiltonian, Eq.~\ref{effham} in this time regime. Firstly, since the effective hamiltonian is a Heisenberg model, we conclude that the low energy distribution would be $\propto 1/L^2$. We make a further educated guess by borrowing the result $\lambda_j=\frac{2 \pi^2 j^2}{L^2 \gamma}$ for clean systems from Ref.~\onlinecite{MarkoPhysRevE.92.042143} and then introducing the relevant prefactor in $\mathcal{H}_{{\rm eff}}$ for disordered systems. Thus, we have disorder averaged
\begin{equation}
\lambda_j \sim -a(\gamma,W) j^2/L^2,
\label{eigenvaluesapp}
\end{equation}
for $j \ll L$, where
\begin{equation}
a(\gamma, W)=\frac{8 \pi^2 \gamma }{ (\frac{2W^2}{3}+4 \gamma^2)}
\label{a}
\end{equation}
and we have used $\langle (h_{j}-h_{j+1})^2 \rangle\sim 2 W^2/3$ for the box distribution $(-W,W)$. Henceforth we shall write $a(\gamma, W)$ as $a$.
\begin{figure}
\includegraphics[width=0.9 \columnwidth]{plots/finalgapimbalance.pdf}
\caption{Plot showing the linear growth of $-\ln I(t)$ in Region V for finite size systems of different lengths $L$ and at disorder strengths $W$. The black dashed lines denote the expected growth of $-\ln I(t)$ with the rate given by corresponding $\lambda_1$ dependent on $L,W$ and $\gamma$.}
\label{fig5a}
\end{figure}
\begin{figure*}
\includegraphics[width=0.9 \columnwidth]{plots/lowestgapvsL.pdf}
\includegraphics[width=0.9 \columnwidth]{plots/gapvsWandgamma.pdf}
\caption{ Scaling of $\lambda_1$, the smallest Liouvillian gap, with different parameters. (a) Plot showing $1/L^2$ scaling of $\lambda_1$, for different disorder strengths, $W$. $L_c$ from Eq.~\ref{criticallength} is marked by appropriate coloured gridlines in the plot. Averaging is done over $100$ realizations. (b) Plot showing dependence of $\ln |\lambda_1 L^2|$ with $W$ and $\gamma$ for $L=3000$ averaged over 100 disorder realizations. The red dashed lines indicate three constant contours at $0,2.5$ and $3.5$ for $\ln a$ obtained from Eq.~\ref{a}.}
\label{fig5}
\end{figure*}
In Fig.~\ref{fig4}(a) we see a very good agreement between our approximate prediction in Eq.~\ref{eigenvaluesapp} and numerical results for the low magnitude eigenvalues of the Liouvillian. Now we turn our attention to the eigenvectors. As shown before, imbalance has a constant support which was significant for odd/even $j$ in clean obc systems. Small perturbations around this value due to disorder, does not affect the overall scheme of things upon averaging over many disorder realizations, as can be verified from Fig.~\ref{fig4}(b). As seen in Fig.~\ref{fig4}(b), for odd $j$, $\mathcal{I}_j$ is almost the same for clean and disordered obc systems with even $L$. For even $j$ on the other had this quantity is no longer identically $0$ as the symmetry of the system is broken due to disorder, and shows a $(\sqrt 2j)^2/L^2$ scaling. However, this does not seriously affect the evolution as their magnitude is much smaller than for odd $j$. Consequently, for disordered obc systems,
\begin{eqnarray}
I(t)\sim&& \frac{1}{L} \int dj \frac{2}{L} e^{-a (2j)^2 /L^2}\nonumber \\
&&= \frac{\sqrt{\pi}}{2L\sqrt{a t}}.
\end{eqnarray}
This is the behaviour seen in Fig.~\ref{fig4}(c), where disorder, $W=4$ , is chosen to be smaller to highlight Region IV (data for $W=10$ is similar at larger times, not shown).
However, for the periodic case, we have a different scenario. Previously $\mathcal{I}(j)$ had no support for any $j$ in this regime for clean systems, but now, due to breaking of symmetry, $\mathcal{I}(j)\sim (\sqrt{2}j)^2/L^2$ as evident from Fig.~\ref{fig4}(b). The eigenvalues as seen in Fig.~\ref{fig4}(a) follow the same scaling law as obc (the degeneracy in the smallest magnitude eigenvalue is resolved as we increase $j$). Hence, we get,
\begin{eqnarray}
I(t)\sim&& \frac{1}{L} \int dj \frac{2j^2}{L^2} e^{-a (j)^2 /L^2}\nonumber \\
=&& \frac{\sqrt{\pi}}{2 (a t)^{3/2}}.
\end{eqnarray} Indeed this is what we see in Fig.~\ref{fig4}(d), including the lack of $L$ dependence in the result.
\subsection{Region V}
\label{region5}
As mentioned in Sec.~\ref{summary}, Region IV becomes asymptotic in the thermodynamic limit, when we have a continuum in the eigenspectrum. For finite size systems though, the spectrum eventually gets resolved and hence for larger $t$ we observe a different behaviour, which we classify as region V. In this region, $I(t)$ shows an exponential decay, with the rate governed by the Liouvillian gap $\lambda_1$.\footnote{the exception being clean pbc systems with even $L$ where imbalance has no support on the eigenfunction as shown before} $\lambda_1$ is thelargest real part of non-zero Liouvillian eigenvalue for even $L$ systems. For odd $L$ systems, it is the next largest as the imbalance operator does not have support for the largest non-zero eigenvalue in this case. In Fig.~\ref{fig5a}, we see that $-\ln I(t)$ grows linearly with time at large times, with the slope $|\lambda_1(L,\gamma,W)|$ verifying our expectation. As before, analytical results for clean systems are well known\cite{MarkoPhysRevE.92.042143}, but upon addition of disorder one can only resort to numerics for exact results. Nevertheless, the description given by Eq.~\ref{effham} is still valid in this time scale, and explains the scaling behaviour of $\lambda_1$ with $L,W$ and $\gamma$ as discussed below.
In Fig.~\ref{fig5}(a) we show the change of $\lambda_1$ with $L$ for constant $W$ and $\gamma$. It is known that the $1/L^2$ scaling of $\lambda_1$ works for sufficiently large $L$ for clean systems. In fact, it has already been shown in Ref.~\onlinecite{MarkoPhysRevE.92.042143} that the critical length required to observe this scaling is $L_c \sim \frac{\pi \sqrt{2}}{\gamma}$. For $\gamma=0.1$ this is at $L\sim 45$ and is shown by the blue gridline in the Fig.~\ref{fig5}(a). This behaviour also continues when we introduce disorder in the system. It is evident from the figure that upon increasing $W$, $L_c$ decreases. This can be explained as in the previous section using Eq~\ref{effham} and Eq.~\ref{eigenvaluesapp}, which tells us that on addition of disorder of strength $W$, $\gamma$ can be effectively replaced by $(4\gamma^2+\frac{2 W^2}{3})/(4\gamma)$. This gives
\begin{equation}
L_c(W)\sim \frac{ 4 \sqrt{2} \pi \gamma}{4\gamma^2+\frac{2 W^2}{3}},
\label{criticallength}
\end{equation} shown as different coloured gridlines in the plot, and capture the transition point between the two regimes quite well.
Finally in Fig.~\ref{fig5}(b) we show the change of $\lambda_1$ on changing $W$ and $\gamma$ for constant $L=3000$ via a contour plot. We expect $|\lambda_1 L^2| \sim a$ and hence the equation of constant contours can be obtained from Eq.~\ref{a} as
\begin{equation}
\ln a={\rm constant}.
\label{fitgap}
\end{equation}
Three such contours are plotted as red dashed lines in Fig.~\ref{fig5}(b) and they match very well with the contours obtained from exact numerics.
\section{Discussion}
\label{discussion}
In this work we have studied the evolution of imbalance, $I(t)$ for the disordered XX chain with on-site dephasing, starting from the N\'eel initial state. Using the hierarchical nature of equations for the observables, we have computed $I(t)$ for large system sizes ($L\sim 10^3$) and long times ($t \sim 10^3$). Our analysis showed the emergence of five timescales in disordered systems.
The shortest time scale, $t < t_0 \sim {\rm Min(1/\gamma,1/W)}$ where $ I(t) \sim (1-\eta t^2)$ constitutes the linear response regime of the system. Then, due to the localization via disorder, two timescales denoted by Region II and III emerge, which are absent in clean systems. Region II is the linear regime of what has been called a stretched exponential decay in previous works\cite{AltmanPhysRevLett.116.160401,CaiBarthelPhysRevLett.111.150403,CaiPhysRevLett.124.130602,SchneiderPhysRevB.106.134211,PhysRevLett.116.237203Garrahan}, where $I(t) \sim (1-\beta \sqrt{t})$, and is universal irrespective of the nature of disorder chosen, and continues till $t_1 \sim 0.1 W^2/8\gamma$, before smoothly transitioning to a disorder dependent behaviour in region III which continues till$t_2 \sim W^2/(8\gamma J^2)$. We thus conclude that the stretched exponential fits are neither universal nor asymptotic for the system under study. Simple local three-site models describe the behaviour in these two regimes, and a rescaled time emerges during the analysis $\tau= 8 \gamma t/W^2$. An effective Heisenberg model given by Eq.~\ref{effham} also provides the correct description of the evolution from Region II.
For $t > t_2$, where Region IV begins, dephasing breaks localization and the system shows size and boundary conditions dependent power law decay of $I(t)$. In fact for obc it shows a $1/(L\sqrt{t})$ decay and for pbc a $1/\sqrt{t^3}$ decay. These were explained from the continuum limit of the low magnitude eigenspectrum of the Liouvillian. Finally, we discussed that for finite sized systems due to resolution of the eigenspectrum, the decay of $I(t)$ in final Region V at $t>t_3$ is exponential with the rate governed by the relevant Liouvillian gap $\lambda_1$. We also provided analysis of how $\lambda_1$ typically scales with $W,L,\gamma$ using the effective Heisenberg Hamiltonian approach.
Our analysis gives insight to the mechanism behind the evolution of imbalance in such a system. It demystifies the behaviour of the system in Region II and III clearly showing its local origin and gives us the ability to make predictions about the evolution for many different disorder distributions. Additionally, the proper analysis of Liouvillian eigenspectrum provides us with the correct asymptotic description of evolution for different models and boundary conditions.
We would also like to briefly mention the nuances of crossover from one regime to another. Since the transition is smooth, there is a always a finite time taken by the system to go from one regime to another. For example, from Region III to IV, where the finite size effect first appears, the system typically takes $t\sim L$ to transition to a power law for open boundary conditions. Hence, if the thermodynamic limit is taken first, we would see that even if region III ends at $L$ independent $\tau_2=8 \gamma t_2/W^2 \sim 1$, the power law in Region IV is only reached asymptotically. However, this does not seem to be the case for periodic boundary conditions. This is also consistent with the system size dependence of the results in Region IV.
Additionally, it is worthy to remark that the effective Hamiltonian description\cite{MariyaMarkoPhysRevB.93.094205} of Eq.~\ref{effham} is also valid for weakly interacting systems where the interaction strength is much smaller than disorder strength. This means weak interaction does not play much role in evolution in such timescales, so we expect most of our results to hold for weakly interacting systems as well.
Furthermore, while we have discussed the N\'eel initial state in this work, decay of $I(t)$ from other generic computational initial states also share some similar features. The behaviour in regions I, IV and V seen for the N\'eel state is still observed for other typical states. The differences occur in $t_2$ and the behaviour in Regions II and III. The first difference in other initial states is that $I(t=0)< 1$ i.e. $-\ln I(t=0)>0$. Since $t_2$ is defined by a finite non-zero value of $I$ universal for a given disorder distribution, relaxation from different $I(t=0)$ typically reaches this value at different times. Secondly, all the sites of the initial state are not locally equivalent unlike the N\'eel state. Hence, the scaling in Regions II and III show a difference due to variation in local behaviours. A detailed analysis of this aspect is beyond the scope of the present work.
There are still a few open questions left in the context of this work. One natural extension would be to study the evolution of other observables such as current in such a model. The question of whether the scaling laws depend on the type of dephasing is also worth studying. Indeed, the hierarchical structure of observables or the effective Heisenberg model might break down if we choose a different model of dephasing. Additionally, while Region I-III has been studied under different disorder settings, we have provided a general explanation for the power law decay in Region IV, discussing the two cases where it shows different exponents. Specific potentials may give rise to unique behaviour in this time scale, and an analysis of that is left for a future work.
\section*{acknowledgement}
We would like to acknowledge support by Grants No.~J1-1698, J1-4385 and
No.~P1-0402 from the Slovenian Research Agency. RG would also like to acknowledge support from UKRI grant EP/R029075/1.
\bibliography{xxdephasing}
\bibliographystyle{apsrev4-2}
\appendix
\section{Two site model}
\label{appA1}
In this Appendix we shall derive the eigenspectrum expressions for the two site model using Eq.~\ref{linearised} as the starting point of our computation. For Regime I this simple model is enough to qualitatively explain the features seen.
For the two site model, $\mathbf{f}=(C_{11},C_{12},C_{21},C_{22})$ and,
\begin{equation}
\mathcal{Q}=\left(
\begin{array}{cccc}
0 & 2 i & -2 i & 0 \\
2 i & 2 i \delta+4 \gamma & 0 & -2 i \\
-2 i & 0 & -2 i \delta+4 \gamma & 2 i \\
0 & -2 i & 2 i & 0 \\
\end{array}
\right)\label{matrix}
\end{equation}
where we have taken $J=1$ and $\delta=h_1-h_2$. For $\delta=0$ (clean system), one can compute the eigenvalues exactly as ,
\begin{equation}
\left[0,4 \gamma ,2 \left(\gamma -\sqrt{\gamma ^2-4}\right),2 \left(\sqrt{\gamma ^2-4}+\gamma \right)\right],
\end{equation} and the time dependent solutions to the $C_{11(22)}$ can be written as,
\begin{eqnarray}
C_{11(22)}&=&e^{2 \gamma t} [(-)\cosh \left(2 t \sqrt{\gamma ^2-4 }\right)\nonumber \\
&&-(+)\frac{\gamma \sinh \left(2 t \sqrt{\gamma ^2-4 }\right)}{\sqrt{\gamma ^2-4}}].
\end{eqnarray}
Hence,
\begin{equation}
I(t)=\frac{1}{2}(C_{11}(t)-C_{22}(t))\sim 1- 8 t^2 +O(t^3).
\end{equation}
This is the demonstration of the short time quadratic behaviour when $t<1/\sqrt{\gamma^2-4}$.
However when $\delta>0$ the exact solution is given from a cubic equation since one of the eigenvalues of $\mathcal{Q}$ is always $0$. The expressions are complicated and hence not shown. But we can formulate a simple perturbative result for $\delta \gg1,\gamma$. To do so we first rearrange $\mathcal{Q}$ to separate the degenerate block and non-degenerate blocks as,
\begin{equation}
\mathcal{Q}=\left(
\begin{array}{cccc}
0 & 0 & 2 i & -2 i \\
0 & 0 & -2 i & 2 i \ \\
2 i & -2 i & 4 \gamma +2 i \delta & 0 \\
-2 i & 2 i & 0 & 4 \gamma -2 i \delta \\
\end{array}
\right)=\begin{pmatrix}
\bf{O}_{2 \times 2} & \bf{B}_{2 \times 2} \\
\bf{C}_{2 \times 2} &\bf{D}_{2 \times 2}
\end{pmatrix} .
\label{rearrange}
\end{equation}
Then applying second order degenerate perturbation theory to $\bf{O}$ and second order non degenerate perturbation to $\bf{D}$ we have the eigenvalues as,
\begin{equation}
\left[0,\frac{16 \gamma }{4 \gamma ^2+\delta ^2},4 \gamma +2 i \delta +\frac{8}{4 \gamma +2 i \delta },4 \gamma- 2 i \delta-\frac{8}{-4 \gamma +2 i \delta }\right]
\end{equation}
and the solution for $\delta \gg \gamma$ is given by
\begin{equation}
I(t)=-\frac{4 \cos \left(2 \sqrt{\delta^2+4} t\right)+\delta^2}{\delta^2+4}=1-8 t^2+O\left(t^4\right)
\end{equation}
whence the quadratic decay with $t$ remains but it is valid till $t \sim 1/\sqrt{\delta^2+4}$.
\section{Three site model}
\label{appA2}
\begin{figure}
\includegraphics[width=0.9 \columnwidth]{plots/approxcompare.pdf}
\caption{Comparison of distribution of $p(\Delta^{(+)})$, between $\Delta^{(+)}$ given by Eq.~\ref{appeq1} and approximated by Eq.~\ref{approx}.}
\label{fig6a}
\end{figure}
Now we repeat the above calculation for a three site model which describes the behaviour in Regime II and III.
In this case $\mathbf{f}$ is given by Eq.~\ref{threesitef}, while $\mathcal{Q}$ is given by,
\begin{equation}
\left(
\begin{array}{ccccccc}
0 & 2 i & -2 i & 0 & 0 & 0 & 0 \\
2 i & 4 \gamma +2 i \delta_1 & 0 & -2 i & 0 & 0 & 0 \\
-2 i & 0 & 4 \gamma -2 i \delta_1 & 2 i & 0 & 0 & 0 \\
0 & -2 i & 2 i & 0 & 2 i \tau & -2 i & 0 \\
0 & 0 & 0 & 2 i & 4 \gamma +2 i \delta_2 & 0 & -2 i \\
0 & 0 & 0 & -2 i & 0 & 4 \gamma -2 i \delta_2 & 2 i \\
0 & 0 & 0 & 0 & -2 i & 2 i & 0 \\
\end{array}
\right)
\end{equation}
where, $h_1-h_2=\delta_1$ and $h_2-h_3=\delta_2$. Rearranging in the form of Eq.~\ref{rearrange}, we get the general form in this case as,
\begin{equation}
\mathcal{Q}=\begin{pmatrix}
\bf{O}_{3 \times 3} & \bf{B}_{3 \times 4} \\
\bf{C}_{4 \times 3} &\bf{D}_{4 \times 4}
\end{pmatrix} .
\end{equation}
Then using second order degenerate perturbation theory in the subspace of $\bf{O}$, we can compute the eigenvalues as $\Delta^{(0)}=0$ and
\begin{widetext}
\begin{equation}
\Delta^{\pm}=\frac{8 \gamma \left(8 \gamma ^2\pm\sqrt{16 \gamma ^4+4 \gamma ^2 \left(\delta_1^2+\delta_2^2\right)+\delta_1^4-\delta_1^2\delta_2^2+\delta_2^4}+\delta_1^2+\delta_2^2\right)}{\left(4 \gamma ^2+\delta_1^2\right) \left(4 \gamma ^2+\delta_2^2\right)}.
\label{eigenvalues}
\end{equation}
\end{widetext}
Since we are in the regime where $W \gg \gamma$, we have $\delta_1$, $\delta_2 \gg \gamma$, and hence we can approximate Eq.~\ref{eigenvalues} as Eq.~\ref{approxeigen}.
One can also compute the eigenvectors in this subspace and plug in the initial N\'eel state to find the coefficients,
\begin{eqnarray}
d^{(0)}&=& -\frac{1}{3}\nonumber \\
d^{(+)}&=&\frac{(2 \delta_1^2-\delta_2^2+\kappa)(\delta_2^2+\kappa)}{3 \delta_1^2 \kappa}\nonumber \\
d^{(-)}&=&\frac{(2 \delta_1^2-\delta_2^2-\kappa)(-\delta_2^2+\kappa)}{3 \delta_1^2 \kappa}
\end{eqnarray}
where $\kappa=\sqrt{\delta_1^4-\delta_1^2 \delta_2^2+\delta_2^4}$. We can see that when $|\delta_1|=|\delta_2|$, $d^{(+)}=\frac{4}{3},d^{(-)}=0$. In the opposite regime, i.e. when $|\delta_1^2 - \delta_2^2|\gg 0$, $d^{(+)} \rightarrow 1$ and $d^{(-)}\rightarrow -d^{(0)}$. Consequently, it can be shown $1 \le d^{(+)} \le 4/3$, $0 \le d^{(-)} \le 1/3$ and hence $d^{(+)}$ always provides the largest contribution to the evolution.
We use the result for $|\delta_1|=|\delta_2|$ when we compute $I(t)$ for alternating potential plotted in Fig.~\ref{fig3}(b) as,
\begin{eqnarray}
I(t)&=&-\frac{1}{3}+\frac{4}{3} e^{-\Delta^{(+)} t}\nonumber \\
&\sim &e^{-4/3 \Delta ^{(+)} t}=e^{-\tau},
\label{alternating2}
\end{eqnarray}
where we have used $\Delta ^{(+)}\sim 6 \gamma /W^2$ for $W \gg \gamma$. One surprising aspect is that Eq.~\ref{alternating2} is a good fit to exact numerics in Fig.~\ref{fig3}(b) beyond $\tau \sim 1$. This can be explained by observing that if we take local models of higher sizes, then we will find numerically that $d^{(0)}$ decreases and the weight gets shifted to a mode which evolves with $\sim4/3 \Delta ^{(+)}$. Thus Eq.~\ref{alternating2} remains valid even when the transport occurs in a more non-local region (but not in the full lattice) and hence to longer times, before the system shows the asymptotic power law behaviour. Also notice that because there is only one mode of relaxation, the decay is exponential in this timescale.
For the random disorder case we will have a distribution of $d$ and $\Delta^{(\pm)}$ based on the distribution of disorder. However, the principal mode of decay is via $\Delta^{(+)}$. To the leading order, the effect of $d_0$ on this mode becomes progressively smaller as we move away from the limiting alternating potential case $|\delta_1|=|\delta_2|$, as it is countered by the $d^{(-)}$ term. Furthermore $\Delta^{(-)}$ is typically too small to have a significant effect on evolution in Regimes II and III. Finally, the shifting of weight towards a mode $\sim \Delta^{(+)}$ when we take slightly bigger but still local models, is valid for the disordered case as well. Hence, simply averaging over $\Delta^{(+)}$ modes is usually enough to get accurate results as presented in Fig.~\ref{fig3}(a).
The final point we need to address is the validity of the approximation we have used to arrive at Eq.~\ref{master1} from Eq.~\ref{firstapprox}. Let us recall that
\begin{equation}
\Delta^{(+)}\sim\frac{8 \gamma (\delta_1^2+\delta_2^2+\sqrt{\delta_1^4+\delta_2^4-\delta_1^2\delta_2^2})}{\delta_1^2 \delta_2^2}.
\label{appeq1}
\end{equation}
For very large $W$, we can expect terms with $|\delta_1|\sim |\delta_2|$ would be statistically insignificant. Hence to get a tractable expression, we take the approximation $|\delta_1^2 - \delta_2^2|\gg 0$. This would constitute the long tails of the distribution of $\Delta^{(+)}$. Finally, to be able to perform the analytical computation simply, we choose to approximate $\Delta^{(+)}$ by,
\begin{eqnarray}
\Delta^{(+)}&\sim &\frac{16 \gamma(\delta_1^2+\delta_2^2)}{\delta_1^2 \delta_2^2}
\label{approx}
\end{eqnarray}
and show its agreement with full $\Delta^{(+)}$ in Fig.~\ref{fig6a}.
The approximation qualitatively represents the exact curve reasonably well. Also, since it becomes exact in the tail, i.e. for large $\Delta^{(+)}$, consequently the expression computed using these result represent the exact result for $I(t)$ for smaller $t$ more accurately than the rest, a feature seen in Fig.~\ref{fig3}(a) and Fig.~\ref{fig3}(b). The exact stretched exponential result which arises out of this approximation for Gaussian disorders is also valid only for small $t$.
\section{Agreement of exact numerics with evolution by $\mathcal{H}_{\rm eff}$}
\label{appB}
\begin{figure}
\includegraphics[width=0.9 \columnwidth]{plots/Heffcomparison2.pdf}
\caption{ Comparison between exact numerical results for evolution of $I(t)$ with those obtained by solving Eq.~\ref{Heffevolution} averaged over 20 realizations for $L=24$ and $W=30$.}
\label{fig6}
\end{figure}
In the end of Sec.~\ref{region23} we mentioned that we could use an effective Hamiltonian to replicate the behaviour of $I(t)$ from a bit later than $\tau=\tau_0$. Let us discuss this aspect in a bit of detail.
Following Ref.~\onlinecite{MariyaMarkoPhysRevB.93.094205}, at a sufficiently large timescale the density matrix becomes effectively diagonal and the evolution of the diagonal elements $\rho_{i}$ is given by,
\begin{equation}
\frac{d\rho_{i}(t)}{dt}=\sum_j[\mathcal{H}_{{\rm eff}}]_{ij}\rho_{j}(t)
\label{Heffevolution}
\end{equation}
where $\mathcal{H}_{{\rm eff}}$ is given in Eq.~\ref{effham}. $I(t)$ can then be computed from $\rho_i(t)$ as
\begin{equation}
I(t)=\sum_i^{2^L}\rho_i(t) \sum_j^L\bra{\alpha_i}(-1)^j \sigma_j^z\ket{\alpha_i}
\label{Imbalanceheff}
\end{equation}
where $\ket{\alpha_i}$ can be taken as the computational basis states. We plot the comparison of results obtained from exact numerics using Eq.~\ref{master1} and evolution of the density matrix using Eq.~\ref{Heffevolution} and Eq.~\ref{Imbalanceheff} in Fig.~\ref{fig6}. We see that from $\tau \sim 0.01$ the two results are almost equal. This effective description reproduces the low energy eigenspectrum of the Liouvillian quite accurately and hence correctly approximates the evolution from $\tau \gg \tau_0$, thus is effective in capturing the evolution of $I(t)$ from approximately Region II, as expected.
\section{Computation of low magnitude eigenspectrum of Liouvillian}
\label{appC}
\begin{figure}
\includegraphics[width=0.9 \columnwidth]{plots/appendixcompare.pdf}
\caption{Comparison of exactly computed $\mathcal{I}_j$ via solving $L\times L$ linear equations for the initial conditions with our approximation using the squares of the eigenvector elements for a system with $L=80$, $W=30$, $\gamma=1$ and periodic boundary conditions, averaged over $10^3$ disorder realizations. The dashed line is the approximate scaling we have used in Sec.~\ref{region4}. }
\label{fig7}
\end{figure}
Following the arguments of Ref.~\onlinecite{MarkoPhysRevE.92.042143}, we have extracted the low magnitude eigenspectrum of the system by diagonalizing a matrix of size $\sim L \times L$ instead of $\sim L^2 \times L^2$ one-particle Liouvillian. If we denote the computational basis as $|j\rangle$, then the low magnitude eigenspectrum of the one-particle Liouvillian can be approximated by eigenspectrum the following matrix
\begin{equation}
\begin{pmatrix}
0 & \mathbf{R}^T & 0\\
\mathbf{R} & -4\mathbb{I} \gamma & i \mathbb{I} \mathbf{X} \\
0 & -i \mathbb{I} \mathbf{X }& -4\mathbb{I} \gamma
\end{pmatrix}
\end{equation}
written in the basis $|j\rangle\langle j|$,
$|j\rangle\langle j+1|+|j+1\rangle\langle j|$ and $|j\rangle\langle j+1|-|j+1\rangle\langle j|$. Here $\gamma$ is the dephasing, $R_{jk}=-2\rm{i}\sqrt{2}(\delta_{j,k}-\delta_{k,j+1})$ and$X_{jk}=(h_j-h_{j+1}) \delta_{jk}$, $h_i$ being the on site disorders. Note that unlike the clean Hamiltonian case treated in Ref.~\onlinecite{MarkoPhysRevE.92.042143}, $|j\rangle\langle j+1|+|j+1\rangle\langle j|$ are not eigenvectors of $\mathcal{L}$ as disorder breaks translational invariance. Thus we need to effectively solve a $3L-2 \times 3L-2$ [$3L \times 3L$] problem for obc [pbc], to obtain the eigenspectrum, which is the `tridiagonal' approximation. This allows us to compute the eigenspectrum for large systems and results are shown in Fig.~\ref{fig4}(a) and (b). \\
Furthermore, for small system sizes one can numerically show that the $\mathbf{\delta f}(j)$s are proportional to the corresponding eigenvector elements and that approximating $\mathcal{I}(j)$ with the square of corresponding eigenvector terms is justified. In Fig.~\ref{fig7} we show the comparison of our approximation with exact $\mathcal{I}(j)$ found by numerically computing the correct weights due to the initial N\'eel state. We do exact diagonalization for a system size of $L=80$ at large $W=30$, averaged over $10^3$ realizations under periodic boundary conditions to obtain the exact data. As one can see the qualitative nature of both the plots are the same, they vary by a factor of approximately $1.5$ (computed numerically). Hence in Fig.~\ref{fig4}(b) our approximation was able to qualitatively replicate the behaviour of $I(t)$ with insignificant deviation from exact numerics.
\end{document}
%apsrev4-2.bst 2019-01-14 (MD) hand-edited version of apsrev4-1.bst
%Control: key (0)
%Control: author (72) initials jnrlst
%Control: editor formatted (1) identically to author
%Control: production of article title (-1) disabled
%Control: page (0) single
%Control: year (1) truncated
%Control: production of eprint (0) enabled
\begin{thebibliography}{41}%
\makeatletter
\providecommand \@ifxundefined [1]{%
\@ifx{#1\undefined}
}%
\providecommand \@ifnum [1]{%
\ifnum #1\expandafter \@firstoftwo
\else \expandafter \@secondoftwo
\fi
}%
\providecommand \@ifx [1]{%
\ifx #1\expandafter \@firstoftwo
\else \expandafter \@secondoftwo
\fi
}%
\providecommand \natexlab [1]{#1}%
\providecommand \enquote [1]{``#1''}%
\providecommand \bibnamefont [1]{#1}%
\providecommand \bibfnamefont [1]{#1}%
\providecommand \citenamefont [1]{#1}%
\providecommand \href@noop [0]{\@secondoftwo}%
\providecommand \href [0]{\begingroup \@sanitize@url \@href}%
\providecommand \@href[1]{\@@startlink{#1}\@@href}%
\providecommand \@@href[1]{\endgroup#1\@@endlink}%
\providecommand \@sanitize@url [0]{\catcode `\\12\catcode `\$12\catcode
`\&12\catcode `\#12\catcode `\^12\catcode `\_12\catcode `\%12\relax}%
\providecommand \@@startlink[1]{}%
\providecommand \@@endlink[0]{}%
\providecommand \url [0]{\begingroup\@sanitize@url \@url }%
\providecommand \@url [1]{\endgroup\@href {#1}{\urlprefix }}%
\providecommand \urlprefix [0]{URL }%
\providecommand \Eprint [0]{\href }%
\providecommand \doibase [0]{https://doi.org/}%
\providecommand \selectlanguage [0]{\@gobble}%
\providecommand \bibinfo [0]{\@secondoftwo}%
\providecommand \bibfield [0]{\@secondoftwo}%
\providecommand \translation [1]{[#1]}%
\providecommand \BibitemOpen [0]{}%
\providecommand \bibitemStop [0]{}%
\providecommand \bibitemNoStop [0]{.\EOS\space}%
\providecommand \EOS [0]{\spacefactor3000\relax}%
\providecommand \BibitemShut [1]{\csname bibitem#1\endcsname}%
\let\auto@bib@innerbib\@empty
%</preamble>
\bibitem [{\citenamefont {Anderson}(1958)}]{AndersonPhysRev.109.1492}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {P.~W.}\ \bibnamefont
{Anderson}},\ }\href {https://doi.org/10.1103/PhysRev.109.1492} {\bibfield
{journal} {\bibinfo {journal} {Phys. Rev.}\ }\textbf {\bibinfo {volume}
{109}},\ \bibinfo {pages} {1492} (\bibinfo {year} {1958})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Evers}\ and\ \citenamefont
{Mirlin}(2008)}]{MirlinRevModPhys.80.1355}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont
{Evers}}\ and\ \bibinfo {author} {\bibfnamefont {A.~D.}\ \bibnamefont
{Mirlin}},\ }\href {https://doi.org/10.1103/RevModPhys.80.1355} {\bibfield
{journal} {\bibinfo {journal} {Rev. Mod. Phys.}\ }\textbf {\bibinfo {volume}
{80}},\ \bibinfo {pages} {1355} (\bibinfo {year} {2008})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {{\v{Z}}nidari{\v{c}}}(2010)}]{Znidaric_2010}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{{\v{Z}}nidari{\v{c}}}},\ }\href
{https://doi.org/10.1088/1742-5468/2010/05/l05002} {\bibfield {journal}
{\bibinfo {journal} {Journal of Statistical Mechanics: Theory and
Experiment}\ }\textbf {\bibinfo {volume} {2010}},\ \bibinfo {pages} {L05002}
(\bibinfo {year} {2010})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Taylor}\ and\ \citenamefont
{Scardicchio}(2021)}]{ScarPhysRevB.103.184202}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {S.~R.}\ \bibnamefont
{Taylor}}\ and\ \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont
{Scardicchio}},\ }\href {https://doi.org/10.1103/PhysRevB.103.184202}
{\bibfield {journal} {\bibinfo {journal} {Phys. Rev. B}\ }\textbf {\bibinfo
{volume} {103}},\ \bibinfo {pages} {184202} (\bibinfo {year}
{2021})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Guo}\ and\ \citenamefont
{Poletti}(2017)}]{PolettiPhysRevA.95.052107}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {C.}~\bibnamefont
{Guo}}\ and\ \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Poletti}},\
}\href {https://doi.org/10.1103/PhysRevA.95.052107} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. A}\ }\textbf {\bibinfo {volume} {95}},\
\bibinfo {pages} {052107} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Vakulchyk}\ \emph {et~al.}(2018)\citenamefont
{Vakulchyk}, \citenamefont {Yusipov}, \citenamefont {Ivanchenko},
\citenamefont {Flach},\ and\ \citenamefont
{Denisov}}]{PhysRevB.98.020202Flach}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {I.}~\bibnamefont
{Vakulchyk}}, \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Yusipov}},
\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Ivanchenko}}, \bibinfo
{author} {\bibfnamefont {S.}~\bibnamefont {Flach}},\ and\ \bibinfo {author}
{\bibfnamefont {S.}~\bibnamefont {Denisov}},\ }\href
{https://doi.org/10.1103/PhysRevB.98.020202} {\bibfield {journal} {\bibinfo
{journal} {Phys. Rev. B}\ }\textbf {\bibinfo {volume} {98}},\ \bibinfo
{pages} {020202} (\bibinfo {year} {2018})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Wu}\ \emph {et~al.}(2019)\citenamefont {Wu},
\citenamefont {Schnell}, \citenamefont {Tomasi}, \citenamefont {Heyl},\ and\
\citenamefont {Eckardt}}]{Wu_2019}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {L.-N.}\ \bibnamefont
{Wu}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Schnell}},
\bibinfo {author} {\bibfnamefont {G.~D.}\ \bibnamefont {Tomasi}}, \bibinfo
{author} {\bibfnamefont {M.}~\bibnamefont {Heyl}},\ and\ \bibinfo {author}
{\bibfnamefont {A.}~\bibnamefont {Eckardt}},\ }\href
{https://doi.org/10.1088/1367-2630/ab25a4} {\bibfield {journal} {\bibinfo
{journal} {New Journal of Physics}\ }\textbf {\bibinfo {volume} {21}},\
\bibinfo {pages} {063026} (\bibinfo {year} {2019})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Monthus}(2017)}]{Monthus_2017}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {C.}~\bibnamefont
{Monthus}},\ }\href {https://doi.org/10.1088/1742-5468/aa64f4} {\bibfield
{journal} {\bibinfo {journal} {Journal of Statistical Mechanics: Theory and
Experiment}\ }\textbf {\bibinfo {volume} {2017}},\ \bibinfo {pages} {043302}
(\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Yusipov}\ \emph {et~al.}(2017)\citenamefont
{Yusipov}, \citenamefont {Laptyeva}, \citenamefont {Denisov},\ and\
\citenamefont {Ivanchenko}}]{PhysRevLett.118.070402Iva}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {I.}~\bibnamefont
{Yusipov}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Laptyeva}},
\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Denisov}},\ and\ \bibinfo
{author} {\bibfnamefont {M.}~\bibnamefont {Ivanchenko}},\ }\href
{https://doi.org/10.1103/PhysRevLett.118.070402} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {118}},\
\bibinfo {pages} {070402} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Vershinina}\ \emph {et~al.}(2017)\citenamefont
{Vershinina}, \citenamefont {Yusipov}, \citenamefont {Denisov}, \citenamefont
{Ivanchenko},\ and\ \citenamefont {Laptyeva}}]{Vershinina_2017}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {O.~S.}\ \bibnamefont
{Vershinina}}, \bibinfo {author} {\bibfnamefont {I.~I.}\ \bibnamefont
{Yusipov}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Denisov}},
\bibinfo {author} {\bibfnamefont {M.~V.}\ \bibnamefont {Ivanchenko}},\ and\
\bibinfo {author} {\bibfnamefont {T.~V.}\ \bibnamefont {Laptyeva}},\ }\href
{https://doi.org/10.1209/0295-5075/119/56001} {\bibfield {journal} {\bibinfo
{journal} {Europhysics Letters}\ }\textbf {\bibinfo {volume} {119}},\
\bibinfo {pages} {56001} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Turkeshi}\ and\ \citenamefont
{Schir\'o}(2021)}]{SchiroPhysRevB.104.144301}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {X.}~\bibnamefont
{Turkeshi}}\ and\ \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Schir\'o}},\ }\href {https://doi.org/10.1103/PhysRevB.104.144301} {\bibfield
{journal} {\bibinfo {journal} {Phys. Rev. B}\ }\textbf {\bibinfo {volume}
{104}},\ \bibinfo {pages} {144301} (\bibinfo {year} {2021})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Medvedyeva}\ \emph {et~al.}(2016)\citenamefont
{Medvedyeva}, \citenamefont {Prosen},\ and\ \citenamefont
{{\v{Z}}nidari{\v{c}}}}]{MariyaMarkoPhysRevB.93.094205}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.~V.}\ \bibnamefont
{Medvedyeva}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Prosen}},\
and\ \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{{\v{Z}}nidari{\v{c}}}},\ }\href {https://doi.org/10.1103/PhysRevB.93.094205}
{\bibfield {journal} {\bibinfo {journal} {Phys. Rev. B}\ }\textbf {\bibinfo
{volume} {93}},\ \bibinfo {pages} {094205} (\bibinfo {year}
{2016})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Gopalakrishnan}\ \emph {et~al.}(2017)\citenamefont
{Gopalakrishnan}, \citenamefont {Islam},\ and\ \citenamefont
{Knap}}]{GopalKnapPhysRevLett.119.046601}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Gopalakrishnan}}, \bibinfo {author} {\bibfnamefont {K.~R.}\ \bibnamefont
{Islam}},\ and\ \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Knap}},\
}\href {https://doi.org/10.1103/PhysRevLett.119.046601} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {119}},\
\bibinfo {pages} {046601} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Sciolla}\ \emph {et~al.}(2015)\citenamefont
{Sciolla}, \citenamefont {Poletti},\ and\ \citenamefont
{Kollath}}]{PolettiPhysRevLett.114.170401}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Sciolla}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Poletti}},\
and\ \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Kollath}},\ }\href
{https://doi.org/10.1103/PhysRevLett.114.170401} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {114}},\
\bibinfo {pages} {170401} (\bibinfo {year} {2015})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Levi}\ \emph {et~al.}(2016)\citenamefont {Levi},
\citenamefont {Heyl}, \citenamefont {Lesanovsky},\ and\ \citenamefont
{Garrahan}}]{PhysRevLett.116.237203Garrahan}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {E.}~\bibnamefont
{Levi}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Heyl}}, \bibinfo
{author} {\bibfnamefont {I.}~\bibnamefont {Lesanovsky}},\ and\ \bibinfo
{author} {\bibfnamefont {J.~P.}\ \bibnamefont {Garrahan}},\ }\href
{https://doi.org/10.1103/PhysRevLett.116.237203} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {116}},\
\bibinfo {pages} {237203} (\bibinfo {year} {2016})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Lezama}\ and\ \citenamefont
{Lev}(2022)}]{10.21468/SciPostPhys.12.5.174BarLev}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {T.~L.~M.}\
\bibnamefont {Lezama}}\ and\ \bibinfo {author} {\bibfnamefont {Y.~B.}\
\bibnamefont {Lev}},\ }\href {https://doi.org/10.21468/SciPostPhys.12.5.174}
{\bibfield {journal} {\bibinfo {journal} {SciPost Phys.}\ }\textbf
{\bibinfo {volume} {12}},\ \bibinfo {pages} {174} (\bibinfo {year}
{2022})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Everest}\ \emph {et~al.}(2017)\citenamefont
{Everest}, \citenamefont {Lesanovsky}, \citenamefont {Garrahan},\ and\
\citenamefont {Levi}}]{PhysRevB.95.024310Garrahan}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Everest}}, \bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Lesanovsky}},
\bibinfo {author} {\bibfnamefont {J.~P.}\ \bibnamefont {Garrahan}},\ and\
\bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Levi}},\ }\href
{https://doi.org/10.1103/PhysRevB.95.024310} {\bibfield {journal} {\bibinfo
{journal} {Phys. Rev. B}\ }\textbf {\bibinfo {volume} {95}},\ \bibinfo
{pages} {024310} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Wolff}\ \emph {et~al.}(2019)\citenamefont {Wolff},
\citenamefont {Bernier}, \citenamefont {Poletti}, \citenamefont {Sheikhan},\
and\ \citenamefont {Kollath}}]{PhysRevB.100.165144Poletti}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Wolff}}, \bibinfo {author} {\bibfnamefont {J.-S.}\ \bibnamefont {Bernier}},
\bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Poletti}}, \bibinfo
{author} {\bibfnamefont {A.}~\bibnamefont {Sheikhan}},\ and\ \bibinfo
{author} {\bibfnamefont {C.}~\bibnamefont {Kollath}},\ }\href
{https://doi.org/10.1103/PhysRevB.100.165144} {\bibfield {journal} {\bibinfo
{journal} {Phys. Rev. B}\ }\textbf {\bibinfo {volume} {100}},\ \bibinfo
{pages} {165144} (\bibinfo {year} {2019})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Wu}\ and\ \citenamefont
{Eckardt}(2019)}]{EckardtPhysRevLett.123.030602}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {L.~N.}\ \bibnamefont
{Wu}}\ and\ \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Eckardt}},\
}\href {https://doi.org/10.1103/PhysRevLett.123.030602} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {123}},\
\bibinfo {pages} {030602} (\bibinfo {year} {2019})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Fischer}\ \emph {et~al.}(2016)\citenamefont
{Fischer}, \citenamefont {Maksymenko},\ and\ \citenamefont
{Altman}}]{AltmanPhysRevLett.116.160401}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.~H.}\ \bibnamefont
{Fischer}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Maksymenko}},\ and\ \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont
{Altman}},\ }\href {https://doi.org/10.1103/PhysRevLett.116.160401}
{\bibfield {journal} {\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf
{\bibinfo {volume} {116}},\ \bibinfo {pages} {160401} (\bibinfo {year}
{2016})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Cai}\ and\ \citenamefont
{Barthel}(2013)}]{CaiBarthelPhysRevLett.111.150403}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {Z.}~\bibnamefont
{Cai}}\ and\ \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont {Barthel}},\
}\href {https://doi.org/10.1103/PhysRevLett.111.150403} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {111}},\
\bibinfo {pages} {150403} (\bibinfo {year} {2013})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Ren}\ \emph {et~al.}(2020)\citenamefont {Ren},
\citenamefont {Li}, \citenamefont {Li}, \citenamefont {Cai},\ and\
\citenamefont {Wang}}]{CaiPhysRevLett.124.130602}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {J.}~\bibnamefont
{Ren}}, \bibinfo {author} {\bibfnamefont {Q.}~\bibnamefont {Li}}, \bibinfo
{author} {\bibfnamefont {W.}~\bibnamefont {Li}}, \bibinfo {author}
{\bibfnamefont {Z.}~\bibnamefont {Cai}},\ and\ \bibinfo {author}
{\bibfnamefont {X.}~\bibnamefont {Wang}},\ }\href
{https://doi.org/10.1103/PhysRevLett.124.130602} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. Lett.}\ }\textbf {\bibinfo {volume} {124}},\
\bibinfo {pages} {130602} (\bibinfo {year} {2020})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Marcantoni}\ \emph {et~al.}(2022)\citenamefont
{Marcantoni}, \citenamefont {Carollo}, \citenamefont {Gambetta},
\citenamefont {Lesanovsky}, \citenamefont {Schneider},\ and\ \citenamefont
{Garrahan}}]{SchneiderPhysRevB.106.134211}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Marcantoni}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Carollo}},
\bibinfo {author} {\bibfnamefont {F.~M.}\ \bibnamefont {Gambetta}}, \bibinfo
{author} {\bibfnamefont {I.}~\bibnamefont {Lesanovsky}}, \bibinfo {author}
{\bibfnamefont {U.}~\bibnamefont {Schneider}},\ and\ \bibinfo {author}
{\bibfnamefont {J.~P.}\ \bibnamefont {Garrahan}},\ }\href
{https://doi.org/10.1103/PhysRevB.106.134211} {\bibfield {journal} {\bibinfo
{journal} {Phys. Rev. B}\ }\textbf {\bibinfo {volume} {106}},\ \bibinfo
{pages} {134211} (\bibinfo {year} {2022})}\BibitemShut {NoStop}%
\bibitem [{Note1()}]{Note1}%
\BibitemOpen
\bibinfo {note} {Recall that since we start from the N\'eel state, a
computational state, and hence at $t=0$ the system is completely
uncorrelated.}\BibitemShut {Stop}%
\bibitem [{\citenamefont {Medvedyeva}\ and\ \citenamefont
{Kehrein}(2014)}]{MedvKehrinPhysRevB.90.205410}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.~V.}\ \bibnamefont
{Medvedyeva}}\ and\ \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Kehrein}},\ }\href {https://doi.org/10.1103/PhysRevB.90.205410} {\bibfield
{journal} {\bibinfo {journal} {Phys. Rev. B}\ }\textbf {\bibinfo {volume}
{90}},\ \bibinfo {pages} {205410} (\bibinfo {year} {2014})}\BibitemShut
{NoStop}%
\bibitem [{Note2()}]{Note2}%
\BibitemOpen
\bibinfo {note} {Hence numerical fits of the form $\protect \qopname \relax
o{exp}(-t^{\alpha })$ in Refs. \protect \rev@citealpnum
{AltmanPhysRevLett.116.160401,CaiBarthelPhysRevLett.111.150403,CaiPhysRevLett.124.130602,SchneiderPhysRevB.106.134211}
found different $\alpha $s depending on fit windows as the behaviour is not a
true stretched exponential.}\BibitemShut {Stop}%
\bibitem [{\citenamefont {{\v{Z}}nidari{\v{c}}}\ and\ \citenamefont
{Horvat}(2013)}]{Znidaric2013}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{{\v{Z}}nidari{\v{c}}}}\ and\ \bibinfo {author} {\bibfnamefont
{M.}~\bibnamefont {Horvat}},\ }\href
{https://doi.org/10.1140/epjb/e2012-30730-9} {\bibfield {journal} {\bibinfo
{journal} {The European Physical Journal B}\ }\textbf {\bibinfo {volume}
{86}},\ \bibinfo {pages} {67} (\bibinfo {year} {2013})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Guo}\ and\ \citenamefont
{Poletti}(2018)}]{PolettiPhysRevA.98.052126}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {C.}~\bibnamefont
{Guo}}\ and\ \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Poletti}},\
}\href {https://doi.org/10.1103/PhysRevA.98.052126} {\bibfield {journal}
{\bibinfo {journal} {Phys. Rev. A}\ }\textbf {\bibinfo {volume} {98}},\
\bibinfo {pages} {052126} (\bibinfo {year} {2018})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Eisler}(2011)}]{Eisler_2011}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {V.}~\bibnamefont
{Eisler}},\ }\href {https://doi.org/10.1088/1742-5468/2011/06/P06007}
{\bibfield {journal} {\bibinfo {journal} {Journal of Statistical Mechanics:
Theory and Experiment}\ }\textbf {\bibinfo {volume} {2011}},\ \bibinfo
{pages} {P06007} (\bibinfo {year} {2011})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Temme}\ \emph {et~al.}(2012)\citenamefont {Temme},
\citenamefont {Wolf},\ and\ \citenamefont {Verstraete}}]{Temme_2012}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {K.}~\bibnamefont
{Temme}}, \bibinfo {author} {\bibfnamefont {M.~M.}\ \bibnamefont {Wolf}},\
and\ \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Verstraete}},\
}\href {https://doi.org/10.1088/1367-2630/14/7/075004} {\bibfield {journal}
{\bibinfo {journal} {New Journal of Physics}\ }\textbf {\bibinfo {volume}
{14}},\ \bibinfo {pages} {075004} (\bibinfo {year} {2012})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Horstmann}\ \emph {et~al.}(2013)\citenamefont
{Horstmann}, \citenamefont {Cirac},\ and\ \citenamefont
{Giedke}}]{CiracPhysRevA.87.012108}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Horstmann}}, \bibinfo {author} {\bibfnamefont {J.~I.}\ \bibnamefont
{Cirac}},\ and\ \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont
{Giedke}},\ }\href {https://doi.org/10.1103/PhysRevA.87.012108} {\bibfield
{journal} {\bibinfo {journal} {Phys. Rev. A}\ }\textbf {\bibinfo {volume}
{87}},\ \bibinfo {pages} {012108} (\bibinfo {year} {2013})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Nava}\ \emph {et~al.}(2022)\citenamefont {Nava},
\citenamefont {Campagnano}, \citenamefont {Sodano},\ and\ \citenamefont
{Giuliano}}]{arxiv.2210.10856}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {A.}~\bibnamefont
{Nava}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Campagnano}},
\bibinfo {author} {\bibfnamefont {P.}~\bibnamefont {Sodano}},\ and\ \bibinfo
{author} {\bibfnamefont {D.}~\bibnamefont {Giuliano}},\ }\href
{https://doi.org/10.48550/arxiv.2210.10856} {\bibinfo {title}
{Arxiv:2210.10856}} (\bibinfo {year} {2022})\BibitemShut {NoStop}%
\bibitem [{Note3()}]{Note3}%
\BibitemOpen
\bibinfo {note} {For Fermionic Hamiltonians, if longer range couplings are
present then in the most general case, $T_{jk}=J_{jk}$ where $J_{jk}$ denotes
the hopping strength between two different sites at positions $j$ and
$k$.}\BibitemShut {Stop}%
\bibitem [{Note4()}]{Note4}%
\BibitemOpen
\bibinfo {note} {This is difficult to rigorously prove analytically, but can
be seen numerically}\BibitemShut {NoStop}%
\bibitem [{Note5()}]{Note5}%
\BibitemOpen
\bibinfo {note} {This choice of approximation is used to obtain a tractable
form for $p(\Delta ^{(+)})$. Other choices are also possible in the said
limit, but they do not simplify the calculation. We show a comparison of the
probability distribution of the approximated and exact $\Delta ^{(+)}$ in
Appendix~\ref {appA2}}\BibitemShut {NoStop}%
\bibitem [{Note6()}]{Note6}%
\BibitemOpen
\bibinfo {note} {One can still numerically solve the small size Hamiltonians
to obtain the eigenenergies non-perturbatively and continue with the
analysis.}\BibitemShut {Stop}%
\bibitem [{Note7()}]{Note7}%
\BibitemOpen
\bibinfo {note} {Ref.~\protect \rev@citealpnum {AltmanPhysRevLett.116.160401}
arrived at this result using a different approach effectively leading to the
same integral. However, under their approximations one arrives at this result
irrespective of the disorder chosen and then needs to fix a free parameter
via numerical fitting to fit to different disorders}\BibitemShut {NoStop}%
\bibitem [{Note8()}]{Note8}%
\BibitemOpen
\bibinfo {note} {$\Omega _j$ gives an oscillatory term which is usually $\sim
0$ for the relevant low magnitude eigenvalues of the Liouvillian spectrum. If
they are included the power law reduces by 1\cite
{MedvKehrinPhysRevB.90.205410}}\BibitemShut {NoStop}%
\bibitem [{\citenamefont
{{\v{Z}}nidari{\v{c}}}(2015)}]{MarkoPhysRevE.92.042143}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{{\v{Z}}nidari{\v{c}}}},\ }\href {https://doi.org/10.1103/PhysRevE.92.042143}
{\bibfield {journal} {\bibinfo {journal} {Phys. Rev. E}\ }\textbf {\bibinfo
{volume} {92}},\ \bibinfo {pages} {042143} (\bibinfo {year}
{2015})}\BibitemShut {NoStop}%
\bibitem [{Note9()}]{Note9}%
\BibitemOpen
\bibinfo {note} {The square comes from the weightage of the initial
conditions, whose weights are of the same order as the element of the
eigenfunction. Hence we use $\sim $ instead of $=$.}\BibitemShut {Stop}%
\bibitem [{Note10()}]{Note10}%
\BibitemOpen
\bibinfo {note} {The exception being clean pbc systems with even $L$ where
imbalance has no support on the eigenfunction as shown before}\BibitemShut
{NoStop}%
\end{thebibliography}%