PHYSICAL REVIEW E 104, 054105 (2021)

Dynamical obstruction to localization in a disordered spin chain

Dries Sels1,2 and Anatoli Polkovnikov3

1
Department of Physics, New York University, New York, New York 10003, USA
2
Center for Computational Quantum Physics, Flatiron Institute, New York 10010, New York, USA
3
Department of Physics, Boston University, Boston, Massachusetts 02215, USA

(Received 2 June 2021; accepted 5 October 2021; published 8 November 2021)

We analyze a one-dimensional XXZ spin chain in a disordered magnetic field. As the main probes of the
system’s behavior, we use the sensitivity of eigenstates to adiabatic transformations, as expressed through the
fidelity susceptibility, in conjunction with the low-frequency asymptotes of the spectral function. We identify a
region of maximal chaos—with exponentially enhanced susceptibility—which separates the many-body local-
ized phase from the diffusive ergodic phase. This regime is characterized by slow transport, and we argue that
the presence of such slow dynamics highly constrains any possible localization transition in the thermodynamic
limit. Rather, the results are more consistent with absence of the localized phase.

DOI: 10.1103/PhysRevE.104.054105

I. INTRODUCTION Recently, several numerical studies have raised doubts

about the stability of a localized phase in the thermody-
Searching for nonergodic or nonthermalizing systems, sta-
namic limit. Systems with quasiperiodic potentials have been
ble to perturbations, has been a long quest since the seminal
argued to be unstable against interactions by Žnidarič and
work by Fermi, Pasta, Ulam, and Tsingou [1,2]. In low-
Ljubotina [21]. References [22,23], while arguing there is
dimensional setups it is known that such systems exists due
still a transition, revised the critical disorder strength to a
to the Kolmogorov-Arnold-Moser (KAM) theorem [3]. How- significantly higher value then previously thought. Two recent
ever, in extended systems, whether classical or quantum, the papers by Šuntajs et al. [24,25] provide numerical evidence
existence of such stable nonergodic systems largely remains for the drift of the localization transition point, with the
an open problem. There are several known situations when system size towards infinitely large disorder strength in the
integrable systems remain perturbatively stable, i.e., they do thermodynamic limit, by studying the spectral form factor.
not thermalize within any order of perturbation theory in a These findings were challenged in a number of works, in-
small parameter. Just to name a few examples, we mention cluding Refs. [23,26,27], the main critique being that one
periodically driven (Floquet) systems, in which thermaliza- cannot reliably extrapolate the information obtained on small
tion time is exponential (nonanalytic function) either in the systems to the thermodynamic limit. In Ref. [28] an ex-
inverse frequency [4–8] or to the amplitude of perturbation ample of a disordered integrable model was found which
[9,10]. Numerically it was found that certain Floquet quantum does not show localization at any disorder strength. Finally,
systems are exceptionally robust against heating, even in the Kiefer-Emmanouilidis et al. [29,30] provide evidence for slow
absence of obvious small parameters [11–14]; however, it growth of the number entropy at strong disorder; a critique
is still not known what their fate is in the thermodynamic on [29] was formulated in Ref. [31], arguing that no trans-
limit. All the models analyzed in these papers are perturba- port is associated with the increasing entropy. One of the
tively stable. In two seminal papers [15,16] it was argued goals of the present paper is to analyze the stability of the
that disordered systems in the insulating regime are pertur- localized phase to increasing the system size using the sen-
batively stable against adding small short-range interactions. sitivity of eigenstates to infinitesimal local perturbations and
As such they should have strictly zero conductivity at finite the low-frequency behavior of the spectral function. Com-
temperatures in the thermodynamic limit. This phenomenon bining both measures, we provide further arguments for the
was termed many-body localization (MBL). In Ref. [17] it idea that there is no truly localized phase in the thermody-
was proposed that MBL is stable beyond the perturbative namic limit in one-dimensional disordered systems, at least
regime in one-dimensional systems and therefore the MBL in the particular disordered XXZ model we are studying here.
phenomenon might persist in the thermodynamic limit. In These conclusions are supported by a general argument that
Ref. [18] Imbrie proposed a proof for the existence of the proposed earlier Griffith-type scenarios of the MBL transi-
MBL regime in a disordered Ising model using some ex- tion are inconsistent with the spectral sum rule in generic
tra assumptions. After the initial discovery, MBL attracted local interacting models. The ultimate physical reason behind
a lot of attention, both theoretical and experimental, and we this conclusion is the existence of an intermediate maximally
refer to recent reviews [19,20] for further information and chaotic region, found earlier in clean systems [32], and a dis-
references. ordered central spin model [33], which separates the localized

2470-0045/2021/104(5)/054105(11) 054105-1 ©2021 American Physical Society

DRIES SELS AND ANATOLI POLKOVNIKOV PHYSICAL REVIEW E 104, 054105 (2021)

and ergodic regimes in finite systems and prevents the system the anisotropy parameter at  = 1.1 such that the model is
from localization in the thermodynamic limit. close to the extensively studied Heisenberg spin chain [24,42]
In the last decade it was realized that fidelity susceptibility and at the same time has broken SU (2) symmetry, even in the
χ , or more generally, the quantum geometric tensor, is a very absence of disorder. It is expected that these minor modifi-
efficient measure for detecting zero-temperature quantum cations of the model do not affect any results related to the
phase transitions (see, for example, Refs. [34–37]). Fidelity MBL transition apart from a small shift of the critical disorder
susceptibility defines the sensitivity of the (ground) state to strength.
small perturbations. Near the phase transition this sensitivity
is usually amplified, leading to singular behavior of χ ; often a II. FIDELITY SUSCEPTIBILITY
divergence. Physically χ is determined by the low-frequency
tail of the spectral function, which is enhanced near phase The fidelity susceptibility, or equivalently, the diagonal
transition boundaries due to the critical slowing down. At component of the quantum geometric tensor, with respect to
finite energy densities the utility of the fidelity susceptibility some coupling λ of a given eigenstate n is defined as [34,37]
was less obvious because of its exponential divergence with ←−  |n|∂λ H|m|2
the system size in ergodic systems [37]. We note that there are χn = n| ∂λ ∂λ |nc ≡ . (2)
finite-temperature generalizations of the fidelity susceptibility m=n
(En − Em )2
measuring distance between density matrices like the Bures
The Hilbert-Schmidt norm of the AGP is defined as an av-
metric, which in turn is related to the Fisher information (see,
erage over the eigenstates of χn . In order to avoid large
e.g., Refs. [38,39]). However, the latter are not sensitive to
eigenstate-to-eigenstate fluctuations but keeping the exponen-
chaos or integrability, but rather probing equilibrium thermo-
tial sensitivity of this probe, it was proposed in Ref. [32] to
dynamic properties of the system.
additionally regularize χn by introducing an energy cutoff μ
In recent work we have shown that the norm of the adia-
that is exponentially small in the system size.
batic gauge potential (AGP), which is equivalent to the fidelity
As a probe λ we will use the local longitudinal magnetic
susceptibility χn averaged over different eigenstates n, can
field acting on a single spin, i.e.,
serve as a very sensitive probe of quantum chaos [32]. The
AGP norm is able to pick up tiny (exponentially small in H → H + λSlz , ∂λ H = Slz . (3)
the system size) integrability breaking perturbations, which
are not necessarily visible to traditional measures of quantum All calculations will be done at λ = 0, i.e., we will analyze
chaos like the level repulsion [40], the spectral form factor the sensitivity of eigenstates to an infinitesimal increase of a Z
[24], or the survival probability [41]. This sensitivity comes magnetic field on a single site. By direct inspection it becomes
from the fact that the low-frequency tail of the spectral func- clear that the disorder average fidelity susceptibility exponen-
tion defining χn can detect changes at very (exponentially) tially diverges with the system size as long as W < ∞. At
long timescales, where small perturbations should leave the large W , this divergence comes from rare resonances where
most pronounced effect. In particular, dependence of the AGP |Em − En | 2−N , which dominate the average χn in the ab-
norm on the system size changes from polynomial for inte- sence of level repulsion. In contrast, since the resonances are
grable systems to exponential for chaotic systems, and this rare, the typical fidelity susceptibility does not diverge with
change is very easy to detect numerically. One of the key find- the system size at sufficiently large W . As such, it should
ings of that work was that the transition from the integrable become small, since eigenstates are almost polarized along
to ergodic phase across various models happens through an the Z axis and adding an extra magnetic field does not affect
intermediate phase, where the fidelity susceptibility diverges them. For this reason, it is much more convenient to analyze
even faster than in the ergodic regime, and this divergence is the scaling behavior of the typical χn obtained by averaging its
accompanied by exponentially long in the system size relax- logarithm. We will define the typical log-fidelity susceptibility
ation times. In another recent work [33], based on a similar as
analysis, it was further argued that in a particular disordered ζ =  log(χn ) , (4)
central spin model this chaotic behavior coexists with the non-
thermalizing nature of the individual eigenstates, and hence where . . .  stands for averaging over both different dis-
to nonergodic behavior of the system even in the absence of order realizations and different eigenstates.. We note that in
small parameters. the Eigenstate thermalization hypothesis (ETH) regime, ζ is
In this work we adopt the approach of Ref. [32] to analyze equivalent to the logarithm of the AGP norm analyzed in
properties of a disordered one-dimensional XXZ spin chain Ref. [32], since the susceptibility is concentrated around the
and specifically analyze the fate of the many-body localization mean [see, e.g., Fig. 6(A)].
transition and the nature of the nonergodic regime there. We It is a simple exercise to extract asymptotes of ζ at weak
use the standard Hamiltonian and strong disorder. At small W the system is expected to be
ergodic and fully obey ETH [43]. Then (see, e.g., Ref. [32])
1  x x   one expects
H= S j S j+1 + S yj S yj+1 + S zj S zj+1 + h j S zj , (1)
W j j ζ = L log 2 + A. (5)
where S x,y,z
j are the spin-1/2 operators and h j are uncorre- The constant A generally depends on the couplings in the
lated random numbers uniformly distributed in the interval Hamiltonian but is insensitive to the system size (up to possi-
[−1, 1]. We assume periodic boundary conditions, and we fix ble logarithmic corrections). In the opposite regime of strong

054105-2
DYNAMICAL OBSTRUCTION TO LOCALIZATION IN A … PHYSICAL REVIEW E 104, 054105 (2021)

In particular, there is a clear maximum in exp[ζ ] developing

at intermediate values of disorder W ∈ [2, 4]. This maximum
drifts to higher values of disorder as we increase the sys-
tem size, gradually approaching the upper bound (7). As we
discussed above, such scaling implies exponentially long in
the system size relaxation times and correspondingly expo-
nentially large values of the spectral weight at frequencies
of the order of the level spacing. Simultaneously with the
drift of the maximum to higher W with increasing L, the
curves tend to sharpen and the main plot even suggests that
there could be a transition to a new phase (MBL) at disorder
W ∼ 4.5. However, as we discuss later, this conclusion would
be premature and a large part of this work will be devoted to
understanding that drift. In the next section we will analyze
what goes on by directly looking into the spectral function.

III. SPECTRAL FUNCTION

FIG. 1. Typical scaled fidelity susceptibility exp[ζ ]/2L (main To get more insight into the behavior of the system we
plot) and the unscaled one exp[ζ ] (inset) as a function of disorder analyze the spectral function, whose low-frequency nature
for different system sizes L = 12, 14, 16, 18 (yellow to blue). The sets the scaling of the fidelity susceptibility [32]. Our work is
dashed line shows the ETH prediction. The dotted line shows the
by no means the first work to look into the dynamical behavior
predicted peak drift with the system size: χmax = c exp[αW ∗ ], where
of disordered systems through linear response probes. Ample
W ∗ = L log(2)/α [see Eq. (11)]. Yellow, green, and blue shaded
works have numerically investigated the two-point spin cor-
regions correspond to ergodic, glassy, and localized phases. As the
system size increases, the green blue crossover boundary drifts to
relator as well as the optical spin conductivity [44–50], either
stronger disorder values. through real-time dynamics or in Fourier space. This spectral
function is defined as a Fourier transform of the symmetric
correlation function of the local magnetization Slz ≡ ∂λ H (to
disorder it is easy to show using perturbation theory in 1/W simplify notations we suppress the site index l in the notation
that (see Appendix A) for the spectral function). In this paper we will only consider
8 the spectral function averaged over eigenstates:
ζ =B− log W, (6) 
3 1  ∞ dt iωt n
| f (ω)|2 = e Gz (t ), (8)
where B is another purely numerical constant, which neither D n −∞ 2π
depends on the system size nor on the couplings. In Ref. [32]
where D = 2L is the dimension of the Hilbert space and Gnz is
we found that in nondisordered models there is a new robust
the connected correlation function:
intermediate regime separating integrable and ergodic limits,  
where Gnz (t ) ≡ 21 n|Slz (t )Slz (0) + Slz (0)Slz (t )|n c
ζ = 2L log 2 + A , (7) with
     2
and the constant A is also independent of the system size. n|Slz (t )Slz (0)|n c ≡ n|Slz (t )Slz (0)|n − n|Slz |n .
Such scaling of ζ saturates its upper bound. Physically, it cor- The spectral function satisfies a simple f -sum rule, which
responds to the presence of exponentially slow in the system immediately follows from integrating Eq. (8) over frequency
size relaxation times in the system, or equivalently, exponen- ω:
tial concentration of the spectral weight at frequencies of the ∞
order of the level spacing (see next section for more details). 1    z 2 
In Fig. 1 we show ζ as a function of the disorder strength dω | f (ω)|2 = n| Sl |n c
D n
W for different system sizes L = 12, 14, 16, 18. The main plot −∞
shows the fidelity scaled by the expected ETH asymptote: 1   2
exp[ζ ]/2L , while the inset shows the unscaled susceptibility = 1 − 4 n|Slz |n , (9)
4D n
exp[ζ ]. The small and large disorder behavior of ζ agrees with
general expectations. Thus for W < 1/2 all curves collapse which is nothing but the average (over eigenstates) fluc-
into the expected ETH prediction shown in the dashed line. tuations of the local magnetization, which are in turn
It is easy to understand the dependence of the constant A in equivalent to the averaged nonconserved (decaying) part of
Eq. (5) on disorder from the dimensional analysis: A ∝ W 2 . the local magnetization. In the MBL phase it is expected
Likewise, from the inset it is evident that at large disorder that part of the local magnetization is conserved, implying
all curves become system size independent perfectly with the that 0 < n|Slz |nc < 1/4. In contrast, in the ergodic phase
slope close to 8/3, in perfect agreement with the perturbation n|(Slz )2 |nc → 1/4 as L → ∞.
theory (see Appendix A). In the intermediate disorder regime The top panel of Fig. 2 shows the averaged over
0.5  W  10 we see clear deviations from both asymptotes. disorder realizations of spectral function | f (ω)|2 for low-to-

054105-3
DRIES SELS AND ANATOLI POLKOVNIKOV PHYSICAL REVIEW E 104, 054105 (2021)

(a)

FIG. 3. Spectral function II. The Fourier transform of the infinite

(b) (c)
temperature connected two-point correlation of the local z magneti-
zation is shown for two values of the disorder, W = 0.5 and W =
1.75, for chains of size L = 12, 14, 16 in blue dots, yellow dash-dot,
and full red, respectively. Heisenberg time is indicated by the black
√
circles. The black dashed lines show 1/ ω and 1/ω asymptotes.

slower logarithmic in time decay of the local magnetization:

Gz (t ) ∼ A − B log(t ) with A and B constants.
A sharp transition from diffusion to subdiffusion was pre-
viously reported in Ref. [47] based on a study of the stationary
FIG. 2. Spectral function I. The spectral function | f (ω)|2 defined state of a large (400 spin) boundary-driven XXZ chain. Subd-
according to Eq. (8) is shown for various values of the disorder W in iffusion at these values of disorder was also reported before
a L = 16 chain. (a) Disorder increasing logarithmically, i.e., equally in Refs. [44–46,50] with a varying dynamical exponent as
spaced, from W = 0.5 (blue) to W = 2.5 (red). The level spacing, a function of W , which diverges at some critical disorder
or inverse Heisenberg time, is indicated by the black circles, and attributed to the MBL transition. The latter support a Griffiths
an estimate of the Thouless energy is shown by the black squares phase interpretation [52] of the ergodic side of the transition.
[51]. The inset shows an exponential fit of the Thouless energy as Recent works, however, suggest that the subdiffusion might
a function of disorder. (b) The average spectral function with disor- be transient [53,54], and ordinary diffusion might govern the
der increasing logarithmically equally spaced from W = 2.5 (blue) late time behavior of the system in the thermodynamic limit.
to W = 15 (red). The typical level spacing is again shown by the Although no firm statements can be made based on the
black circles. (c) The typical spectral function, defined by averaging present numerics, the data might support the
log | f (ω)|2 over different disorder realizations, for the same disorder √ latter. At weak
disorder W < 0.5 we observe a stable 1/ ω regime which
values as in panel (b). The dashed line is a guide for the eye and becomes flat at the Thouless energy and whose exponent does
shows asymptotic inverse frequency scaling of the spectral function: not drift with system size (see Fig. 3). At larger disorder it
| f (ω)|2 = C/(W ω).
is much harder to identify a stable power law regime, and
the spectral function clearly has a smaller exponent at smaller
frequencies. Moreover, at intermediate frequencies the expo-
nent increases with the system size L approaching 1, as can
intermediate values of disorder, which incrementally changes seen from Fig. 3, where we show the spectral function for two
from W = 0.5 (bottom blue line) to W = 2.5 (top red line). different realizations of disorder: W = 0.5 and W = 1.75 and
Note that all these values of disorder are in the delocalized three different system sizes, L = 12, 14, 16.
regime corresponding to the shaded green color in Fig. 1. One We thus propose the following interpretation of the data.
can observe a clear crossover from diffusive to subdiffusive Beyond a critical value of disorder W ∼ 0.5, corresponding
behavior of the magnetization. In the diffusive
√ regime the cor- to the border between yellow and green regions in Fig. 1, the
relation function Gnz (t ) decays as ∼1/ t at long times, which exponent of the spectral function crosses over from 1/2 to 1
√
is reflected in the | f (ω)|2 ∼ 1/ ω scaling of the spectral and further increase of disorder simply leads to increasing
function, which eventually saturates at low frequencies below range of frequencies with 1/ω asymptotic behavior, which
the inverse Thouless time. Such diffusive scaling is clearly after some possibly diffusive part is ultimately cut off by
seen at W = 0.5. As disorder increases the spectral function system size at small frequencies. Our numerical results, at the
crosses over to 1/ω asymptote, which corresponds to a much present system sizes, do not support the existence of a stable

054105-4
DYNAMICAL OBSTRUCTION TO LOCALIZATION IN A … PHYSICAL REVIEW E 104, 054105 (2021)

portional to the disorder strength such that in this regime

0.0179
| f (ω)|2 ≈ , (10)
Wω
where the constant was extracted from the data in Fig. 2 [51].
One can check that frequencies above this scale, ω > ωuv ,
give small and approximately disorder-independent contribu-
tions to the spectral weight. Therefore it is the subdiffusive
1/ω scaling regime of the spectral function which dominates
the total spectral weight. Because the integral of 1/ω diverges,
and in order to satisfy the sum rule (9), there must be a
low-frequency scale ωth , where the spectral function either
saturates or potentially crosses over to a lower power of the
inverse frequency corresponding to faster dynamics, for ex-
ample, diffusion. Assuming that L is very large, this scale can
be found from
 ωuv
0.0179dω 1
2 ≈ → ωth ≈ ωuv e−αW ,
ωth Wω 4
FIG. 4. Conserved Z magnetization. The conserved part of the
local Z magnetization defined in Eq. (12) as a function of disorder with α ≈ 6.98. Note that the factor of 2 multiplying the inte-
for L = 10, 12, 14, 16 (blue to green). The main plot shows the gral stems from the equal contribution coming from negative
magnetization in the log-linear scale, and the inset does in the log-log frequencies. As we start decreasing L, at some point this scale
scale. The dashed line shows the prediction of perturbation theory in hits the level spacing, after which the spectral weight has to
1/W : Z = 1 − c/W . condense and the system goes into the localized MBL regime.
This happens when
disorder-dependent intermediate exponent but do not rule it ωuv e−αW ≈ uv e
−L log 2
,
out either. We note that this inverse frequency scaling of the
spectral function was found earlier in random regular graph where uv is set by the many-body bandwidth and hence is
models [55,56]. bounded by uv < cL. This gives the approximate relation
As disorder increases beyond W = 2.5 (bottom panel in between the critical disorder strength and the system size at
Fig. 2), the slope of the spectral function decreases again. the transition point:
Simultaneously, a difference between typical and average L log 2 log( uv /ωuv )
spectral functions starts to develop, signaling ergodicity is W ∗ (L) ≈ − . (11)
α α
broken. The typical spectral function (right panel) stays closer
to the 1/ω scaling with a rapidly decreasing coefficient as At large L the critical disorder increases linearly in L, since
a function of W . This change in behavior occurs when the log( uv /ωuv ) only generates logarithmic corrections, such
Thouless energy ωTh becomes comparable to the typical level that W ∗ ≈ 0.099L + c. The latter is consistent with the pre-
spacing [51]. From conservation of the spectral weight (9), diction of Ref. [24], in which the critical disorder strength
we conclude that the system goes into the localized (MBL) was found to shift like W ∗ ≈ 0.098L + c based on a scaling
regime. This conclusion is further supported by analyzing the analysis of a many-body version of the logarithmic Thouless
conserved part of the local magnetization (see Fig. 4). Re- conductance.
markably, even at the strongest disorder shown, W = 15, there Note that we extract a lower bound on the Thouless energy
is a clear tail down to zero frequency, which is well below by using the f -sum rule, since we assumed all of the spectral
the mean level spacing. So the localized phase consists of a weight to be in the 1/ω regime of the spectral function. Nu-
mixture of localized and delocalized degrees of freedom. It is merically we find for L = 16 that the UV part of the spectral
∞
this low-frequency tail which is responsible for a noticeable weight, defined as ωuv dω| f 2 (ω)|, ωuv ≈ 0.1/W takes up a
system size dependence of the fidelity susceptibility, even at significant fraction of the spectral weight, slowly reducing
strong disorder (see Fig. 1). with W from 74% at W = 1.5 to 46% at W = 2.5. It is
expected that as W increases, this UV part goes down to zero
and Eq. (11) is asymptotically exact. But for available system
IV. DYNAMICAL OBSTRUCTION
sizes, in which the range of accessible disorder strengths
Let us now estimate the Thouless energy and argue about showing 1/ω scaling regime is limited, the constant α is
its fate, as well as about the fate of the localization transition significantly renormalized by this spectral weight reduction.
as we increase the system size. First let us note that as dis- If we instead identify the Thouless energy as the point of
order increases, crossing a threshold of the order of W ∼ 1, maximal curvature on Fig. 2, we extract a different estimate
the onset of 1/ω scaling of the spectral function starts at a for the exponent α: α ≈ 3.43 for a system of L = 16 (see
high frequency ωuv ∼ 0.1/W , which is system size indepen- the inset in Fig. 2). This is approximately half the maximal
dent. This observation immediately implies that in this scaling value extracted from the sum rule (11), and it is in good
regime, | f (ω)|2 ≈ C/ω, with the prefactor C independent of agreement with the α = 1/0.29 ≈ 3.45 estimated in [24].
the system size. Numerically we find that C is inversely pro- It should be noted that the Thouless energy has previously

054105-5
DRIES SELS AND ANATOLI POLKOVNIKOV PHYSICAL REVIEW E 104, 054105 (2021)

been numerically extracted from the spectral function (up to where a > 0 is some constant and the constants C and C
L = 20) by Serbyn et al. [50]. Whereas the data is consistent are related by the continuity of the spectral function. For
with ours, it is analyzed differently and the authors had a example, such a spectral function with a = 3 describes the
different interpretation, supporting the Griffiths phase picture zero-temperature disordered model analyzed in Ref. [57] and
on the ergodic side, leading up to the critical fan where the with a = 3/2 describes an infinite temperature-disordered
Thouless energy decays exponentially with the system size spin model on a random regular graph [58]. Then as z di-
ωth ∝ exp(−κL). verges near the critical point, the sum rule is satisfied as
Let us end this section by noting that the commonly ac- long as a > 1. While this behavior of the spectral function is
cepted Griffiths picture, with a critical point at finite disorder, possible, it implies that the dynamical exponent near the tran-
is not internally consistent. The argument requires only three sition, z(ω) ≡ 1/(1 + d log | f (ω)|2 /d log |ω|) = log(1/ω)/a,
inputs: locality, conservation of energy, and bounded norm increases as ω becomes smaller. Physically this behavior of
of local operators. Consider a Griffiths scenario [52] where z(ω) implies that already on the ergodic side, the system’s
local correlations, of a conserved operator like magnetization dynamics is getting slower and slower with time instead of
or simply an operator coupled to the energy, decay in time accelerating as the system “realizes” that it is in the ergodic
like C(t ) ∼ t −1/z , where the dynamical exponent z depends on regime. Also, such scenarios are inconsistent with existing nu-
disorder W . In the Griffiths picture, z would be proportional merical results, both reported here in Fig. 2 and in other works
to the correlation length z ∼ ξ (W ). Since energy is conserved, [53,54]. The only remaining possibility to satisfy the spectral
decaying correlations imply spreading of this magnetization sum rule with the system-size-independent C(W ) at z → ∞ is
in space like x ∼ t 1/z , and the spreading stops at the Thou- when C(W ) → 0 at some critical disorder Wc . This scenario
less time when the dynamics detects the boundaries, i.e., the corresponds to opening a spectral gap in the system. Such
Thouless energy obeys ωth ∼ L −z . In order for ergodicity to be a spectral gap is present in normal Bethe-ansatz integrable
broken, one needs the Thouless energy to reach the level spac- models [32], suppressing the exponential divergence of the
ing, which implies z  L/ log L. In the Griffiths picture, this fidelity susceptibility with the system size. However, this gap
condition is equivalent to saying that the correlation length usually exists only for special classes of perturbations, which
ξ (W ) > L has to exceed the system size, and as such the keep the system integrable. There is no numerical indication
system is critical. In frequency space the spectral function for that a spectral gap opens up in disordered systems.
general subdiffusion with dynamical exponent z simply reads
V. FIDELITY SUSCEPTIBILITY REVISITED
C
| f (ω)|2 = , For the sake of completeness we now revisit the fidelity
ω 1−1/z
susceptibility, as there are a few more interesting points to
where C = | f (ω = 1)|2 is defined by the UV spectral weight. make. From Fig. 1 we clearly discern three different scaling
Note that, thus far, this scenario would also yield exponential regimes of the typical fidelity susceptibility denoted by three
enhancement of the susceptibility in the critical region. How- different colors. These regimes can be loosely labeled as (i)
ever, it suffers from the same problem as before, namely, the completely ergodic or ETH type, (ii) glassy with slow relax-
total amount of spectral weight must be conserved. The latter ation, and (iii) localized. In this section we will analyze these
directly implies three regimes in more detail.
First we analyze the total conserved part of the local Z
Cz  1. magnetization, which is directly related to the integral of the
spectral function [see Eq. (9)]. We define the latter as
This inequality can be satisfied for any finite z, but for the   2 
system to become critical, z has to diverge with L, implying Z = 4 exp log n Slz n . (12)
C has to decay to zero at least like log(L)/L. Not only is this We choose the normalization such that Z = 1 in the fully
inconsistent with the numerical data (see Fig. 3), it is simply localized regime and Z ∼ 2−L in the ergodic regime in the
impossible by locality. Recall that C is the UV spectral weight zero-magnetization sector. To avoid effects of the tail of the
at the onset of the asymptotic hydrodynamic scaling behavior. distribution of magnetization, we first take its logarithm, then
As such it cannot possibly know about the system size L. Of average it over different eigenstates and disorder realizations,
course, the spectral weight can be a function of disorder W , and then exponentiate. In Fig. 4 we show 1 − Z, measuring
and while we argued that C ∼ 1/W such that W ∗ ∼ L, faster the nonconserved part of the local magnetization. It is evident
decay will simply result in a slower drift of the critical point. that as disorder increases, the magnetization starts to localize.
This argument is consistent with the observed existence As we increase the system size, the transition to the localized
of the localization transition in the nonlocal models like the regime happens for larger and larger values of disorder, and
random regular graphs [55,56]. Note that in principle one can the observed drift is consistent with the drift of the peak in the
imagine other scenarios, where the exponent z for example fidelity susceptibility (see Fig. 1). At very large W all curves
would be frequency dependent, approaching unity at the crit- collapse into one, and thus the conserved magnetization be-
ical point only in the limit ω → 0. Such a scenario could comes insensitive to the system size. The black line is shown
correspond to logarithmic corrections to the 1/ω scaling: for reference, representing the leading perturbative correction
to the infinite disorder limit.
C
ω loga (1/ω)
ω > e−az Next in Fig. 5 we show the mean scaled fidelity suscep-
| f (ω)|2 = C
ω1−1/z
0 < ω < e−az , tibility χ̄ /2L ≡ χn /2L and the typical one exp[ζ ] versus

054105-6
DYNAMICAL OBSTRUCTION TO LOCALIZATION IN A … PHYSICAL REVIEW E 104, 054105 (2021)

(a)

(b)

FIG. 5. Ergodicity breaking. The scaled fidelity susceptibility

is shown as a function of disorder for chains of length L =
12, 14, 16, 18 (yellow, red, green, blue). In contrast to Fig. 1, we
show both the typical (circles) and mean (squares). Moreover, by
using a linear scale for the disorder W it becomes directly ap-
parent that both the typical and the mean grow exponentially at (c)
small disorder. After the peak in the typical susceptibility, the mean
susceptibility keeps growing, resulting in an exponentially large sep-
aration between the two. The dashed line shows the predicted drift
of the maximum: χmax = c exp[αW ∗ ], where W ∗ = L log(2)/α [see
Eq. (11)].

disorder. The plot of the latter (solid lines black with circles)
reproduce the data from Fig. 1, except disorder is now shown
in the linear scale. In the ergodic regime at low disorder,
the mean and the typical susceptibilities are parallel to each
other with only a constant offset between them, which comes
from small Gaussian fluctuations of χn around the mean.
As disorder increases the typical susceptibility reaches the
FIG. 6. AGP distribution. The distribution of the fidelity suscep-
maximum and goes down, while χ̄ keeps growing until it
tibility is shown for a chain of L = 16. To make the structure more
saturates at large W . As we discussed earlier, the saturated
apparent we have divided disorder range in three intervals. (a) Disor-
value is determined by rare resonances in Eq. (2), where der increasing from W = 0.5 (blue) to W = 1.75 (red). (b) Disorder
|En − En±1 | 2−L . Such resonances inevitably occur due to increasing from W = 2.1 (blue) to W = 4.3 (red), and (c) shows the
absence of level repulsion (see also Fig. 7). As a result, the disorder increasing from W = 5 (blue) to W = 15 (red). The disorder
mean fidelity susceptibility fluctuates much more at strong is logarithmically equally spaced. Various exponentials are shown by
disorder than the typical one. dashed lines for comparison. The different panels are also labeled on
One can extract additional information about the system Fig. 7.
by analyzing the tail of the probability distribution of χn , as
it was first discussed in Ref. [32]. Both in ETH and localized
regimes, this tail is determined by anomalously small nearest c exp[−z/(2 + α)], we find
 z    z 
energy differences |En − En+1 | and thus contains additional
P(z) ∝ exp − Q c exp − ,
information about the level statistics. For random matrix 2+α 2+α
ensembles, the distribution has been derived in [59] and a where Q(s) is the probability to have the level spacing s. At
numerical survey on disordered systems has been presented in small values of s, corresponding to large z, this probability
Ref. [60]. To avoid exponential factors, it is more convenient usually takes a universal power law form: Q(s) ∝ sβ with
to deal with the distribution of ζn ≡ log(χn ). The probability β = 0, indicating Poisson statistics, β > 0 corresponding to
that ζn = z at large values of z can be estimated as the level repulsion and β < 0 corresponding to the level at-
P(z) ≡ Pr(ζn = z) ≈ Pr(log(| f 2 (s)|/s2 ) = z), traction. Combining all these factors together one finds
 
where s is the smallest of |En − En+1 | and |En − En−1 |. 1+β
P(z) ∝ exp −z . (13)
Assuming that | f (s)|2 = c|s|−α at small s such that s = 2+α

054105-7
DRIES SELS AND ANATOLI POLKOVNIKOV PHYSICAL REVIEW E 104, 054105 (2021)

function α > 0. An independent check on the level statistics

shows that the latter explanation is correct, i.e., β = 0 and
α > 0. In Fig. 7 we show the extracted exponent as a function
of disorder and contrast it with the level spacing statistics.
Clearly the two curves qualitatively follow each other very
well. The exponent is a bit noisier because there are higher
sample-to-sample fluctuations in the distribution tail. As we
emphasized, at large disorder the exponent goes below a naive
Poisson ratio, highlighting exponential enhancement of the
matrix elements of magnetization at small energy differences
compared to the matrix elements of random operators.
It is remarkable that by analyzing the tail of the fidelity sus-
ceptibility we come to the same conclusion as analyzing the
spectral function. Namely, a simple cartoon representation of
the localized phase as a set of weakly coupled ergodic boxes
seems to be invalid. This representation was, for example,
FIG. 7. AGP exponent. The tail of the distribution of the logarith- underlying the renormalization group treatment of the MBL
mic fidelity susceptibility decays with a characteristic exponent. This transition [61,62]. Instead, at small frequencies there is an
figure shows the exponent extracted from the distributions shown exponential enhancement of the matrix elements of local oper-
in Fig. 6, where the different panels are indicated on the top. The ators, as the local magnetization is not special in this respect. It
fitted exponent is shown by the blue circles. For comparison, the red is this enhancement that prevents the system from localization
squares show the level spacing ratio.
in the thermodynamic limit. At this point it is worth nothing
that while phenomenological Renormalization group (RGs)
In the ETH (random matrix) regime we have β = 1 and α = 0 [61,62] support the Griffiths phase, they are not consistent in
such that P(z) ∝ exp[−z], in agreement with the Random ma- that they do not predict the associated enhancement of the
trix theory (RMT) result [59]. In the localized regime, with no susceptibility. However, it is the same susceptibility which
level repulsion (β = 0) and a Gaussian orthogonal ensemble actually controls the RG. More microscopic constructions,
(GOE) random operator ∂λ H with no selection rules α = 0, such as those in [63], are also controlled by the fidelity sus-
we would find P(ζn = z) ∝ exp[−z/2]. Note that this asymp- ceptibility, and similar issues could emerge there as well.
tote gives the slowest decay of the tail of the distribution for
random operators in the absence of level attraction. Any ex-
VI. CONCLUSION
ponent smaller than 1/2 for such operators would necessarily
imply β < 0, i.e., the levels attract. However, for (normalized) We analyzed the sensitivity of eigenstates of a disordered
operators with diverging spectral function | f 2 (s)|, at s → 0 one-dimensional XXZ chain to small perturbations of a local
with a positive exponent α the exponent in the tail of the distri- longitudinal magnetic field. Specifically, we analyzed the typ-
bution can be lower than 1/2. Instead it is bounded from below ical fidelity susceptibility as a function of disorder strength
by 1/3, i.e., P(z) ∝ exp[−z/3], which is realized at β = 0 and and system size. We found that this susceptibility exhibits
the maximum possible value of α = 1. Any exponent less than a maximum near the localization transition, where the sus-
1/3 would necessarily imply level attraction irrespective of ceptibility is exponentially enhanced in the system size. We
the operator ∂λ H. observe a significant drift of this peak to larger disorder with
In Fig. 6 we show the distribution functions of the increasing system size, supporting the conclusions of recent
log-fidelity susceptibility ζn : P(z) for different disorder re- work [21,24,25,29,30].
alizations. We divided the disorder realizations in the three We further analyzed the low-frequency dependence of the
intervals. The first interval [panel (a)] corresponds to disorder spectral function (auto-correlation function) of the local mag-
increasing from W = 0.5 to W = 1.75. In this interval the netization and found that for disorder larger than W √ ∼ 1 it
level spacing distribution is well described by the Wigner- rapidly crosses over from a diffusive | f (ω)|2 ∝ 1/ ω to a
Dyson statistics with β = 1 (see Fig. 7) and α = 0, as the subdiffusive regime with a universal exponent | f (ω)|2 ∝ 1/ω.
spectral function is flat at frequencies of the order of level For the present system sizes, we did not find any evidence
spacing (Fig. 2). So from Eq. (13) we anticipate that P(z) ∝ for existence of robust intermediate disorder-dependent expo-
exp[−z], which perfectly agrees with the numerical results. nents. We argued, based on simple physical principles, that
As we keep increasing disorder, the system starts to lose this regime is inconsistent with localization in the thermo-
level repulsion (Fig. 7), and accordingly, the slope in the tail dynamic limit because of conservation of the total spectral
distribution of ζn goes down. Panel (b) shows the results for weight.
disorder increasing from W = 2.1 to W = 4.3, where this Finally, we identified an issue with earlier analytic treat-
slope clearly changes between 1 and 1/2. As disorder gets ments of the MBL transition. Namely, the maximally chaotic
even larger [panel (c)], the slope of the tail keeps going down critical region separating the ergodic and localized regimes is
until it saturates at a value close to 1/3, clearly below the characterized by a strong divergence of the matrix elements
bound 1/2 for random operators, which we discussed in the of local operators at small energy differences of the order
previous paragraph. This exponent can be explained either of the level spacing. This enhancement is not captured by
by level attraction β < 0 or by divergence of the spectral phenomenological RGs and, according to our arguments, this

054105-8
DYNAMICAL OBSTRUCTION TO LOCALIZATION IN A … PHYSICAL REVIEW E 104, 054105 (2021)

leads to instability of the localized phase at large system sizes. (a)

Despite finding evidence against localization in the thermo-
dynamic limit, we confirmed the existence of a very slow,
nearly localized subdiffusive regime with logarithmic in time
spreading of correlations.
The predictions of our work can be readily tested experi-
mentally. In particular, the spectral function can be extracted
as a Fourier transform of the connected part of the autocorrela-
tion function of the local magnetization, which was measured
already for over three decades in state-of-the-art cold atom
experiments [64]. One can also extract it from transport mea-
surements analyzing spreading of an initially localized spin or
charge, like it was done in Anderson localization experiments
[65]. In time, one might also consider extracting it using a
quantum computer [66,67]. While extracting the very low-
frequency part of the spectral function might be challenging,
the universal 1/ω scaling starts at accessible frequencies and
should be measurable.
(b)
ACKNOWLEDGMENTS
The authors thank Mohit Pandey for his collaboration dur-
ing the early stages of this work. They further acknowledge
inspiring discussions with Ehud Altman, Anushya Chandran,
Soonwon Choi, Phil Crowley, Eugene Demler, David Huse,
Frank Pollmann, Tomaž Prosen, Marcos Rigol, Giuseppe
De Tomasi, and Lev Vidmar. The authors also acknowledge
useful feedback on the manuscript by Dima Abanin, Ana-
toly Dymarsky, Sarang Gopalakrishnan, Maksym Serbyn, and
Michiel Wouters. A.P. was supported by NSF under Grant
No. DMR-1813499 and DMR-2103658 and by the AFOSR
under Grant No. FA9550-16-1-0334 and FA9550-21-1-0342.
The Flatiron Institute is a division of the Simons Foundation.

APPENDIX: STRONG DISORDER SCALING

Here we will have a look at the behavior of the typical FIG. 8. Log W prefactor. Figure shows the behavior of the
fidelity susceptibility ζ for W → ∞. Recall that we are inves- constant c in the asymptotic behavior of the logarithmic fidelity sus-
tigating the susceptibility with respect to changing the local ceptibility ζ = c log W . (a) The numerical evaluation of expression
magnetic field, such that (A8) and (b) ζ extracted from exact diagonalization for systems sizes
 |n|S z |m|2 L = 4 (blue) to L = 12 (red) for very large disorder W . The black
χn = l
, (A1) lines are fits of the form ζ = C log W + B, where C is fixed from
m=n
(En − Em )2 expression (A8) and only the offset B is fitted.

and ζ =  log(χn ) . When W → ∞ the Hamiltonian be-

comes diagonal in the S z basis, and consequently the suscep- It is directly clear from Eq. (A2) that all contributions to the
tibility tends to zero. As long as there are no (many-body) susceptibility in which m and n have the same value of the
resonances, which do not affect the typical susceptibility any- spin on site l vanish. Since HX X simply swaps neighboring
way, we can do perturbation theory in 1/W . To first order we spins, the only states that contribute to the sum in expression
find (A2) are those which have at least one of the neighbors of l
antialigned with l, i.e.,
1  n 2 |zn |HX X |zm |2  2 n 2 
χn(1) = 2 sl − slm  4 , (A2) 1 sln − sl+1
n
sl − sl−1
n
W m=n En(0) − Em(0) χn =
(1)
+ . (A4)
4W 2 (hl − hl+1 )4 (hl − hl−1 )4
where |zn  are z-polarized product states, sln = ±1/2 is the
value of Slz inthis eigenstate, i.e., Slz |zn  = sln |zm , and the Considering all eight possible combinations of sl and sl±1 ,
energy En = i hi sin . Finally, the flip-flop Hamiltonian HX X
(0) there are four states for which one has a contribution of
is defined as 1/4W 2 h4 , two states for which both neighbors contribute
  and we have 1/2W 2 h4 , and two states states for which
HX X = S xj S xj+1 + S yj S yj+1 . (A3) we get 0. The latter, which are completely polarized states
j | · · · ↑↑↑ · · ·  and | · · · ↓↓↓ · · · , need to be handled with

054105-9
DRIES SELS AND ANATOLI POLKOVNIKOV PHYSICAL REVIEW E 104, 054105 (2021)

care in computing the typical susceptibility ζn = log χn , as size system L from our previous argument, as it becomes more
this would diverge. For those states we need to consider higher and more unlikely to have have an all polarized region since
order corrections in perturbation theory, since the first order the total magnetization ultimately needs to sum up to zero.
contribution vanishes. Fortunately that is straightforward to Consider the ith step, in which we are looking at the faith of
do, as long as there are no resonances such that we can neglect a polarized region consisting of 2i − 1 spins. The probability
renormalization of the denominators. To first order we had that both the left and right neighbors have the same polariza-
tion is
ζ = 3
4
log W −2 + 41 log O(W −4 ) + B, (A5)
[L/2 − (2i − 1)] (L/2 − 2i)
where B is some universal constant that is independent of pi = . (A7)
L − (2i − 1) L − 2i
L, W and comes from the disorder average over h. In order
for the polarized states to contribute in the next order, they One can go through exactly the same argument as before,
need to move an antialigned spin to the center by an extra resulting in a generalization of expression (A6):
application of the flip-flip Hamiltonian HX X . Consequently,
either the left or right must have opposite spin to the polarized  
L/4 i−1

region, otherwise also the O(1/W 4 ) contribution vanishes. ζ =− [p j ] (1 − pi ) 2i log W + B, (A8)

i=1 j=0
Once again, we have a probability of 1/4 for both of the ends
to be aligned with the central piece. This argument can be where one should set p0 = 1. The result is plotted in Fig. 8,
applied ad infinitum, resulting in showing how the exponent increases from 2 to 8/3 with in-
  creasing system size L. Due to the constraint, the convergence
3 −2 1 3 −4 1 −6
ζ = log W + log W + log O(W ) + B to the thermodynamic limit is only algebraic and corrections
4 4 4 4
are of O(1/L). Note that this causes a pronounced and rather
∞
3 1   i−1
8 counterintuitive phenomenon for small systems, that in this
=− 2i log W + B = − log W + B. (A6) asymptotic scaling regime the susceptibility of larger systems
i=1
4 4 3
decays to zero faster than that of smaller systems.
The typical eigenstate susceptibility thus decays to zero like Finally, note that expression (A2) applies more generally in
W −8/3 for sufficiently strong disorder, in contrast to the av- the MBL regime, as long as one simply replaces sln = ±1/2
erage over eigenstates which behaves as W −2 . Finally, let us with the nonzero expectation values of the local z magneti-
briefly look at finite size L effects. Given that we are growing zation in the dressed l-bit eigenbasis. It is easy to see that
the polarized region from both sides in every order of pertur- irrespective of the assumptions about the matrix elements, this
bation theory, finite L effects can be estimated by replacing perturbation theory is exponentially divergent with the block
the upper bound in (A6) by L/2. The latter would result size at small energy differences unless (sln − sm ) vanishes
in corrections of order L2−L , and hence convergence to the with the energy difference (En − Em ). As long as the latter
thermodynamic limit is fast. However, since the Hamiltonian is not true, this perturbation theory is unstable, leading to
under consideration conserves total magnetization, we have the spectral weight generation at small frequencies and decay
been primarily concerned with the zero-magnetization sector of conserved magnetization. Further details of this instability
of the problem. The latter modifies the probabilities in a finite will be reported elsewhere.

[1] G. P. Berman and F. M. Izrailev, Chaos: An Interdisciplinary [12] L. D’Alessio and A. Polkovnikov, Ann. Phys. 333, 19 (2013).
Journal of Nonlinear Science 15, 015104 (2005). [13] A. Haldar, R. Moessner, and A. Das, Phys. Rev. B 97, 245122
[2] M. A. Porter, N. J. Zabusky, B. Hu, and D. K. Campbell, Am. (2018).
Sci. 97, 214 (2009). [14] M. Heyl, P. Hauke, and P. Zoller, Sci. Adv. 5 (2019).
[3] V. Arnol’d, Mathematical Methods of Classical Mechanics [15] I. V. Gornyi, A. D. Mirlin, and D. G. Polyakov, Phys. Rev. Lett.
(Springer-Verlag, New York, 1989). 95, 206603 (2005).
[4] D. A. Abanin, W. De Roeck, and F. Huveneers, Phys. Rev. Lett. [16] D. Basko, I. Aleiner, and B. Altshuler, Ann. Phys. 321, 1126
115, 256803 (2015). (2006).
[5] T. Mori, T. Kuwahara, and K. Saito, Phys. Rev. Lett. 116, [17] V. Oganesyan and D. A. Huse, Phys. Rev. B 75, 155111 (2007).
120401 (2016). [18] J. Z. Imbrie, J. Stat. Phys. 163, 988 (2016).
[6] P. Weinberg, M. Bukov, L. D’Alessio, A. Polkovnikov, S. [19] R. Nandkishore and D. A. Huse, Annu. Rev. Condens. Matter
Vajna, and M. Kolodrubetz, Phys. Rep. 688, 1 (2017). Phys. 6, 15 (2015).
[7] A. Rajak, R. Citro, and E. G. D. Torre, J. Phys. A: Math. Theor. [20] D. A. Abanin, E. Altman, I. Bloch, and M. Serbyn, Rev. Mod.
51, 465001 (2018). Phys. 91, 021001 (2019).
[8] O. Howell, P. Weinberg, D. Sels, A. Polkovnikov, and M. [21] M. Žnidarič and M. Ljubotina, Proc. Natl. Acad. Sci. 115, 4595
Bukov, Phys. Rev. Lett. 122, 010602 (2019). (2018).
[9] S. Vajna, K. Klobas, T. Prosen, and A. Polkovnikov, Phys. Rev. [22] E. V. H. Doggen, I. V. Gornyi, A. D. Mirlin, and D. G. Polyakov,
Lett. 120, 200607 (2018). arXiv:2101.05651.
[10] W. D. Roeck and V. Verreet, arXiv:1911.01998. [23] P. Sierant, D. Delande, and J. Zakrzewski, Phys. Rev. Lett. 124,
[11] T. Prosen, Phys. Rev. Lett. 80, 1808 (1998). 186601 (2020).

054105-10
DYNAMICAL OBSTRUCTION TO LOCALIZATION IN A … PHYSICAL REVIEW E 104, 054105 (2021)

[24] J. Šuntajs, J. Bonča, T. Prosen, and L. Vidmar, Phys. Rev. E [48] R. Nandkishore, S. Gopalakrishnan, and D. A. Huse, Phys. Rev.
102, 062144 (2020). B 90, 064203 (2014).
[25] J. Šuntajs, J. Bonča, T. Prosen, and L. Vidmar, Phys. Rev. B [49] O. S. Barišić, J. Kokalj, I. Balog, and P. Prelovšek, Phys. Rev.
102, 064207 (2020). B 94, 045126 (2016).
[26] D. Abanin, J. H. Bardarson, G. D. Tomasi, S. Gopalakrishnan, [50] M. Serbyn, Z. Papić, and D. A. Abanin, Phys. Rev. B 96,
V. Khemani, S. A. Parameswaran, F. Pollmann, A. C. Potter, M. 104201 (2017).
Serbyn, and R. Vasseur, Ann. Phys. 427, 168415 (2021). [51] The Thouless energy is extracted numerically by identifying the
[27] R. K. Panda, A. Scardicchio, M. Schulz, S. R. Taylor, and M. point of maximal curvature of the data on log-log scale. This
Žnidarič, Europhys. Lett. 128, 67003 (2020). procedure is robust up to W ≈ 1.5. The coefficient C, for the
[28] F. H. L. Essler, R. van den Berg, and V. Gritsev, Phys. Rev. B asymptotic C/W ω behavior of the spectral function, is extracted
98, 024203 (2018). by multiplying the spectral function by W ω and extracting the
[29] M. Kiefer-Emmanouilidis, R. Unanyan, M. Fleischhauer, and J. value of the inflection point. This estimate smoothly goes from
Sirker, Phys. Rev. Lett. 124, 243601 (2020). 0.0245 to 0.0179 when disorder changes from W = 1 to W =
[30] M. Kiefer-Emmanouilidis, R. Unanyan, M. Fleischhauer, and 2.5.
J. Sirker, Absence of true localization in many-body localized [52] K. Agarwal, E. Altman, E. Demler, S. Gopalakrishnan, D. A.
phases, Phys. Rev. B 103, 024203 (2021). Huse, and M. Knap, Ann. Phys. 529, 1600326 (2017).
[31] D. J. Luitz and Y. B. Lev, Phys. Rev. B 102, 100202(R) (2020). [53] S. Bera, G. De Tomasi, F. Weiner, and F. Evers, Phys. Rev. Lett.
[32] M. Pandey, P. W. Claeys, D. K. Campbell, A. Polkovnikov, and 118, 196801 (2017).
D. Sels, Phys. Rev. X 10, 041017 (2020). [54] G. De Tomasi, D. Hetterich, P. Sala, and F. Pollmann, Phys. Rev.
[33] T. Villazon, P. W. Claeys, M. Pandey, A. Polkovnikov, and A. B 100, 214313 (2019).
Chandran, Sci. Rep. 10, 16080 (2020). [55] V. E. Kravtsov, I. M. Khaymovich, E. Cuevas, and M. Amini,
[34] L. Campos Venuti and P. Zanardi, Phys. Rev. Lett. 99, 095701 New J. Phys. 17, 122002 (2015).
(2007). [56] V. Kravtsov, B. Altshuler, and L. Ioffe, Ann. Phys. 389, 148
[35] A. F. Albuquerque, F. Alet, C. Sire, and S. Capponi, Phys. Rev. (2018).
B 81, 064418 (2010). [57] O. Motrunich, K. Damle, and D. A. Huse, Phys. Rev. B 63,
[36] M. Kolodrubetz, V. Gritsev, and A. Polkovnikov, Phys. Rev. B 134424 (2001).
88, 064304 (2013). [58] K. S. Tikhonov and A. D. Mirlin, Phys. Rev. B 99, 024202
[37] M. Kolodrubetz, D. Sels, P. Mehta, and A. Polkovnikov, Phys. (2019).
Rep. 697, 1 (2017). [59] P. Sierant, A. Maksymov, M. Kuś, and J. Zakrzewski, Phys.
[38] J. S. Sidhu and P. Kok, AVS Quantum Sci. 2, 014701 Rev. E 99, 050102(R) (2019).
(2020). [60] A. Maksymov, P. Sierant, and J. Zakrzewski, Phys. Rev. B 99,
[39] D. Liska and V. Gritsev, SciPost Phys. 10, 020 (2021). 224202 (2019).
[40] L. D’Alessio, Y. Kafri, A. Polkovnikov, and M. Rigol, Adv. [61] R. Vosk, D. A. Huse, and E. Altman, Phys. Rev. X 5, 031032
Phys. 65, 239 (2016). (2015).
[41] F. Borgonovi, F. M. Izrailev, L. F. Santos, and V. G. Zelevinsky, [62] A. C. Potter, R. Vasseur, and S. A. Parameswaran, Phys. Rev. X
Phys. Rep. 626, 1 (2016). 5, 031033 (2015).
[42] M. Serbyn, Z. Papić, and D. A. Abanin, Phys. Rev. X 5, 041047 [63] T. Thiery, M. Müller, and W. D. Roeck, arXiv:1711.09880.
(2015). [64] A. Rubio-Abadal, J.-y. Choi, J. Zeiher, S. Hollerith, J. Rui, I.
[43] We exclude very small values of disorder W  0.1, where the Bloch, and C. Gross, Phys. Rev. X 9, 041014 (2019).
system becomes very close to the integrable XXZ model. [65] J. Billy, V. Josse, Z. Zuo, A. Bernard, B. Hambrecht, P. Lugan,
[44] K. Agarwal, S. Gopalakrishnan, M. Knap, M. Müller, and E. D. Clément, L. Sanchez-Palencia, P. Bouyer, and A. Aspect,
Demler, Phys. Rev. Lett. 114, 160401 (2015). Nature (London) 453, 891 (2008).
[45] I. Khait, S. Gazit, N. Y. Yao, and A. Auerbach, Phys. Rev. B 93, [66] D. Sels and E. Demler, Quantum generative model for sam-
224205 (2016). pling many-body spectral functions, Phys. Rev. B 103, 014301
[46] D. J. Luitz and Y. B. Lev, Ann. Phys. 529, 1600350 (2017). (2021).
[47] M. Žnidarič, A. Scardicchio, and V. K. Varma, Phys. Rev. Lett. [67] D. V. Vasilyev, A. Grankin, M. A. Baranov, L. M. Sieberer, and
117, 040601 (2016). P. Zoller, PRX Quantum 1, 020302 (2020).

054105-11