
3
As a next step, we demonstrate how to find pairs of en-
ergy eigenstates whose equal superpositions are well approx-
imated by product states. It is reasonable to hypothesize that
such states must have a low entanglement with respect to
any bipartition. Therefore we performed a structured search
on small systems using exact diagonalization and targeting
energy eigenstates whose sub-lattice entanglement entropy
(ABABAB...-bipartition) is small, see Supplementary Ma-
terial for more details. Targeting small sub-lattice entangle-
ment is a heuristic choice motivated by the following con-
siderations: i) Product states have vanishing sub-lattice en-
tanglement entropy and therefore any state sufficiently close
to a product state should have small sublattice entanglement
and ii) even generic translationally invariant matrix-product
states (MPS) [69, 70], which are commonly considered to
be low-entangled, have extensive sub-lattice entanglement en-
tropies [71]. Therefore small sub-lattice entanglement heuris-
tically indicates an amount of entanglement that is small
even compared to MPS. Our preliminary analysis showed that
pairs of energy eigenstates whose equal superpositions are
well approximated by product states exist and that one class
of them comes in the form of deformed domain walls (see
Fig. 1). This knowledge then allows us to devise an effi-
cient tensor-network based algorithm to study large systems,
which we now briefly explain (further details may be found in
Sec. IV B).
B. Numerical construction
At sufficiently strong disorder, the eigenstates of ˆ
Hfeature
an area-law entanglement and may be represented faithfully
as MPS [57], whose explicit representation can be determined
using the DMRG-X algorithm [72]. The algorithm starts with
a “seed” state |m1⟩⊗···⊗|mL⟩, where |mj⟩ ∈ {|↑⟩,|↓⟩}
denote the eigenstates of ˆ
S(z)
j. These seeds are the eigenstates
of ˆ
Hin the limit of W→ ∞. DMRG-X then iteratively de-
termines an (approximate) eigenstate at finite Wthat is, in a
sense, closest to the initial seed. The main numerical con-
trol parameter is the so-called bond dimension χ, which we
choose so that high-energy eigenstates are obtained up to ma-
chine precision.
In our case, we find the energy eigenstates |E:k⟩associ-
ated with seeds in domain-wall form
|dw : k⟩=|↓⟩ ⊗ ··· ⊗ |↓⟩
| {z }
ktimes
⊗|↑⟩ ⊗ ··· ⊗ |↑⟩.(8)
We then form the superposition of the energy eigenstates re-
sulting from neighboring domain-walls,
Ψ(k)
±E=1
√2(|E:k⟩ ± |E:k+ 1 ⟩),(9)
and finally construct their product-state approximation |Φ(k)
±⟩.
This allows us to calculate the certified amplitude Acert.of
Eq. (7). All of these operations can be implemented efficiently
and accurately in the MPS representation (see Sec. IV B for
further details). We stress that at this point it is not clear why
the states Ψ(k)
±Eshould be close to product states apart from
the fact that we found revolving product states with a sim-
ilar structure in our small scale exact-diagonalization numer-
ics (see Supplementary Material). Our main results in the next
section show that for domain-wall seeds, closeness to a prod-
uct state is indeed a generic case for sufficiently strong dis-
order. This in turn immediately implies the non-equilibrating
behavior for the associated product states.
C. Main results
In Fig. 2 our aggregated numerical data for the certified am-
plitude at varying system sizes up to L= 160 and at a disor-
der strength W= 8 with 100 disorder realizations per system
size is depicted (the corresponding fidelities are discussed in
Supplementary Note 1). We find median certified amplitudes
of the order of 0.7, essentially independent of the system size
with decreasing fluctuations as Lincreases. Moreover, the
maximum certified amplitudes for domain-wall states with in-
terface in the middle half of the system (sites k=L/4to
k= 3L/4) slowly increase with system size, with all sam-
pled realizations reaching Acert.>0.91 for L= 160. The
restriction to states with the interface in the middle half of the
system excludes states that can be interpreted as being close
to single-particle excitations (see below and Supplementary
Note 2). We emphasize that the certified amplitude provides
a lower bound to the magnitude of the oscillations of ˆ
Aand
that there may exist local operators which oscillate with even
higher amplitude.
In a nutshell, Fig. 2 conclusively demonstrates the (generic)
existence of initial product states that host high-fidelity re-
vivals and the existence of local, indefinitely-oscillating, ob-
servables in a system of up to 160 sites. The overall shape of
these product states is of the form of two domain walls sep-
arated by a spin pointing roughly in ±x-direction at their in-
terface. Moving away from the interface, the spins still point
away from their original ±z-directions, but with decreasing
components in the x−y-plane. This is visualized in Fig. 1. As
a side remark, we mention that the Hamiltonian ˆ
Hmay also
be interpreted as a Hamiltonian of interacting fermions by a
Jordan-Wigner transformation. However, in this picture the
parity super-selection rule forbids our reviving product states,
since they correspond to super-positions of states with differ-
ent fermion-number parity.
The fact that we find oscillating deformed domain walls
is particularly interesting since previous results indicate that
the bare domain-wall states |dw : k⟩approach a steady-state
with a smeared-out interface, a process known as domain-wall
melting [73]. An interface spin pointing away from the z-axis
therefore protects against this mechanism.
Since the DMRG-X algorithm outputs the energy eigen-
states |E:k⟩as MPS, we can compute the expectation values
⟨Ψ(k)
±|ˆ
B(t)|Ψ(k)
±⟩of any local operator ˆ
Bexactly for arbitrary
times t(see Sec. IV B). This in turn allows us to quantitatively
estimate the finite-time expectation value ⟨Φ(k)
±|ˆ
B(t)|Φ(k)
±⟩