Disordered MPS Decomposition

• Disordered matrix product state decomposition is a tensor-network technique that encodes site-dependent disorder into MPS/MPO representations for many-body systems.
• The method efficiently compresses the full eigenspectrum of a many-body localized phase into a low-bond-dimension MPO using binary local tensor choices.
• It provides practical diagnostics such as bond-dimension scaling and entanglement bounds, and guides numerical simulations of disordered and inhomogeneous quantum systems.

Searching arXiv for the core paper and closely related work on disorder-adapted MPS/MPO representations. Searching for the 2014 MBL–MPO paper. Disordered matrix product state decomposition denotes a family of tensor-network representations in which disorder, inhomogeneity, or realization-dependent local structure is encoded directly into matrix product state (MPS) or matrix product operator (MPO) tensors. In the most specific and influential usage, the decomposition refers to the claim that in a one-dimensional many-body localized (MBL) phase the exact diagonalizing unitary $U$ can be efficiently compressed as a low-bond-dimension MPO, so that the entire many-body spectrum is generated by binary local tensor choices $G_i$ and $E_i$ at each site (Pekker et al., 2014). In broader usage, the same phrase also covers site-dependent random MPS ensembles, fixed-bond-dimension MPS manifolds used for disordered dynamics, and disorder-sensitive tensor decompositions in phases without conventional local order (Haferkamp et al., 2021). Across these settings, the common idea is that disorder does not merely perturb a clean tensor network; it selects a nonuniform local basis, alters entanglement structure, and changes what constitutes an efficient decomposition.

1. Spectrum-wide decomposition in many-body localized phases

The most explicit formulation of a disordered MPS decomposition appears in the proposal to represent the full eigensystem of an MBL Hamiltonian by a single disorder-adapted MPO $U$ (Pekker et al., 2014). The physical motivation is the analogy with Anderson insulators: in a noninteracting localized system, once the localized single-particle orbitals are known, the many-body eigenstates are product states in the occupation basis of those orbitals. The interacting analogue replaces localized orbital occupations by localized tensor choices.

The corresponding MBL ansatz is

$$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$

where $G_i$ is the local tensor corresponding to the local ground-state choice on site $i$, and $E_i$ is the local tensor corresponding to the local excited-state choice (Pekker et al., 2014). For an $n$-site spin-$\tfrac12$ chain, all $G_i$0 eigenstates are obtained by all binary choices of $G_i$1 or $G_i$2 across the chain. This is not merely a statement that each eigenstate is individually approximable by an MPS. Rather, the entire spectrum is organized by one common low-bond-dimension MPO.

The operator $G_i$3 is defined by

$G_i$4

where $G_i$5 are product states in the microscopic $G_i$6 basis and $G_i$7 are the exact many-body eigenstates (Pekker et al., 2014). If $G_i$8 is represented as an MPO with local tensors $G_i$9, then acting on a product basis state selects one of two tensor slices, $E_i$0 or $E_i$1, which are relabeled as $E_i$2 and $E_i$3. In this way, a disordered tensor decomposition becomes a realization-dependent reorganization of the full spectrum into localized binary degrees of freedom.

This formulation is significant because MBL is not primarily a ground-state phenomenon. The relevant structural question concerns highly excited states at finite energy density. The proposal therefore extends tensor-network representability beyond the conventional gapped-ground-state setting and ties it to the existence of quasi-local conserved quantities, area-law entanglement, and failure of thermalization (Pekker et al., 2014).

2. Tensor-network structure, l-bits, and bond-dimension diagnostics

The efficiency of the decomposition is controlled by the MPO bond dimension $E_i$4. Numerically, the MPO for $E_i$5 is obtained by gluing bra and ket indices into a doubled local index,

$E_i$6

compressing the resulting object as an MPS, and then splitting $E_i$7 back into $E_i$8 to recover the MPO (Pekker et al., 2014). The underlying expectation is that if $E_i$9 is quasi-local, then entanglement in the doubled-space representation remains short-ranged, so the required bond dimension stays small.

A central implication is the entanglement bound

$U$0

since the entanglement entropy across a cut is bounded by the logarithm of the Schmidt rank and hence by $U$1 (Pekker et al., 2014). If $U$2 remains bounded in the thermodynamic limit, then all eigenstates generated by the MPO obey an area law and cannot satisfy ETH volume-law scaling. This makes bond-dimension growth itself a localization diagnostic.

The l-bit structure is represented directly in MPO language. The operator

$U$3

is identified as the l-bit raising operator, and a product-basis operator such as

$U$4

is mapped to a conserved quasi-local operator

$U$5

in the physical basis (Pekker et al., 2014). This provides a tensor-network realization of emergent localized integrals of motion.

The same work argues that if the localization length is $U$6, then the MPO bond dimension scales roughly as

$U$7

For strong disorder, $U$8 is finite and $U$9 should saturate with system size; in a delocalized phase $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$0 should grow strongly with $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$1 (Pekker et al., 2014). The numerical fit

$$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$2

introduces a length scale $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$3 interpreted as a saturation length or localization length, and this $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$4 was reported to agree with the localization length extracted from effective l-bit couplings up to an overall rescaling factor of about $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$5 (Pekker et al., 2014).

3. Disorder, rare regions, and the limits of compactness

The canonical model used to test the decomposition is the random-field Heisenberg chain

$$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$6

which is noted to have an MBL transition near $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$7 (Pekker et al., 2014). The construction proceeds by exact diagonalization in each fixed total $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$8 sector, matching product states to eigenstates using the Hungarian algorithm to maximize $$\Psi_\text{MBL}=\prod_i (\psi_i E_i^{\sigma_i}+\phi_i G_i^{\sigma_i})|\sigma_i\rangle,$$9, and compressing the resulting unitary into an MPO. Because the matching is heuristic rather than globally optimal, the resulting bond dimensions are upper bounds on the optimal compression (Pekker et al., 2014).

The reported finite-size behavior separates three regimes. For weak disorder, roughly $G_i$0, the bond dimension grows linearly with $G_i$1. For strong disorder, roughly $G_i$2, the bond dimension has essentially saturated already by $G_i$3. For intermediate disorder $G_i$4, accessible sizes do not support a definitive asymptotic conclusion (Pekker et al., 2014). This establishes the decomposition as efficient only in the localized regime.

An important qualification is that the localized phase is not represented by a uniformly bounded bond dimension over all disorder samples. The paper emphasizes strong Griffiths effects: rare weak-disorder regions can resemble thermal inclusions and require much larger local bond dimension (Pekker et al., 2014). Accordingly, the bond-dimension distribution in the localized phase is peaked at small $G_i$5, around $G_i$6, but has a power-law tail; near the transition the distribution becomes extremely broad. This means that the decomposition is typical rather than uniform. A plausible implication is that disorder-adapted MPS descriptions of MBL should be understood probabilistically across realizations rather than as deterministic worst-case bounds.

The same theme reappears in later hybrid and dynamical contexts. In disordered electron–phonon systems treated by a hybrid MPS–multi-trajectory Ehrenfest scheme, disorder enters as quenched onsite potentials $G_i$7 plus a trajectory-dependent field $G_i$8, so that along one trajectory the effective onsite potential is

$G_i$9

The MPS decomposition itself remains standard, but the numerics become dominated by averaging over both disorder realizations and classical trajectories, with $i$0 and $i$1 to $i$2 in the main disordered calculations (Menzler et al., 11 Dec 2025). This suggests that in disordered tensor decompositions the primary complication may shift from ansatz design to sampling and realization dependence.

4. Random and site-dependent MPS as disordered tensor ensembles

A different meaning of disordered matrix product state decomposition arises in the theory of random matrix product states with site-dependent local tensors (Haferkamp et al., 2021). Here the state is written as a periodic MPS

$i$3

with $i$4 explicitly dependent on site $i$5 (Haferkamp et al., 2021). Disorder means that the tensors are not translation invariant and are generated from independent Haar-random local unitaries $i$6. The resulting ensemble is denoted $i$7.

This framework is not an algorithm for decomposing a given state. It is instead a formal ensemble of disordered tensor decompositions. The local tensors are obtained by feeding a fixed state $i$8 into a Haar-random unitary acting on $i$9, equivalently by drawing tensor cores uniformly from the Stiefel manifold of isometries (Haferkamp et al., 2021). Because the disorder is local and i.i.d., global moments factorize into products of local Haar integrals.

The analytical machinery is built from Weingarten calculus and a mapping of moment tensor networks to one-dimensional classical partition functions. For example, moments such as

$E_i$0

are reduced to sums over local permutation variables in $E_i$1, while subsystem purities become effective one-dimensional statistical mechanics models with local plaquette weights (Haferkamp et al., 2021). This converts a disordered tensor-network problem into a tractable transfer-weight problem.

Several structural consequences follow. The norm concentrates near one,

$E_i$2

and for Hamiltonians with non-degenerate spectrum and non-degenerate spectral gap the normalized state equilibrates exponentially well with overwhelming probability (Haferkamp et al., 2021). For disconnected subsystems, the Rényi-2 entropy is generically extensive, while for connected subsystems of fixed size and sufficiently large bond dimension the reduced state is almost maximally mixed (Haferkamp et al., 2021). These results identify a disordered MPS ensemble as a model of generic states in the trivial phase of matter, with disorder expressed entirely through site-dependent random local tensors.

5. Fixed-bond MPS manifolds and disordered dynamics

In dynamical problems, disordered MPS decomposition often means not a static factorization theorem but a restriction of quantum evolution to a fixed-bond-dimension MPS manifold. The time-dependent variational principle realizes this as

$E_i$3

where $E_i$4 projects the Schrödinger vector field onto the tangent space of the fixed-bond-dimension MPS manifold $E_i$5 (Kloss et al., 2017). The relevant decomposition is therefore the split of $E_i$6 into tangent and discarded components.

The principal result is methodological rather than structural: apparent convergence of long-time observables with bond dimension can be misleading (Kloss et al., 2017). In benchmark cases, TDVP produces incorrect asymptotic behavior even when observables appear stable. For the clean XX chain, exact ballistic transport crosses over under TDVP to apparently diffusive behavior with diffusion constant $E_i$7; for the disordered XX chain, exact Anderson localization is missed, and the mean-square displacement fails to plateau; for the clean XX ladder, the correct $E_i$8 crosses over to an incorrect $E_i$9 (Kloss et al., 2017).

Against this cautionary backdrop, the disordered nonintegrable XXZ chain

$n$0

with $n$1 and $n$2, behaves differently (Kloss et al., 2017). The mean-square displacement

$n$3

obeys $n$4 with fitted exponent $n$5, in excellent agreement with short-time exact and previously numerically exact results (Kloss et al., 2017). This does not establish a new disordered MPS decomposition, but it does delimit when a fixed-complexity disordered MPS representation may be asymptotically faithful for transport.

A common misconception is that weak bond-dimension dependence of long-time observables suffices to validate the method. The explicit counterexamples show that the limits $n$6 and $n$7 need not commute (Kloss et al., 2017). In this sense, disordered MPS manifolds can be reliable dynamical coarse-grainings in some interacting disordered regimes, but not as a universal principle.

6. Other uses of the term and broader methodological context

Outside quenched-random spin chains, the phrase also appears in settings where “disordered” refers to absence of local order or to general site dependence rather than to random couplings. In the one-dimensional extended quantum compass model, the disordered phases are efficiently represented by low-bond-dimension MPS and are characterized by a doubly degenerate entanglement spectrum and nonzero string order parameters rather than by conventional local order parameters (Liu et al., 2012). In the limiting disordered phases with $n$8, the exact ground states have bond dimension $n$9 and doubly degenerate diagonal bond matrix $\tfrac12$0, providing a simple example in which the MPS decomposition exposes hidden order (Liu et al., 2012).

In multimode cavity QED, an MPS-compatible chain decomposition is constructed not for a random many-body state but for an inhomogeneous bosonic environment. Numerical mode decomposition yields arbitrary cavity mode frequencies and couplings, which are then transformed from star geometry to a chain Hamiltonian by an exact orthogonal map

$\tfrac12$1

The resulting state is then represented as an MPS along the chain (Ryu et al., 2022). This is not a disorder paper, but it shows how nonuniform mode structure is converted into an MPS-compatible decomposition.

More abstractly, site-dependent MPS can also be viewed through algebraic and compilation lenses. The algebraic-geometric theory of translation-invariant MPS provides an instructive contrast: many simplifications there depend on cyclic symmetry and fail in genuinely disordered settings (Critch et al., 2012). By contrast, shallow-circuit decompositions of generic site-dependent MPS do not assume translation invariance and therefore extend naturally to inhomogeneous inputs; the best-performing protocol was a sequential analytic circuit growth plus global constrained unitary optimization strategy (Rudolph et al., 2022). A plausible implication is that disorder primarily obstructs symmetry-based reductions, not tensor-network compilation in principle.

Taken together, these works support a broad but technically precise interpretation of disordered matrix product state decomposition. In one line of work, it is a disorder-adapted full-spectrum factorization of localized eigenstates by an MPO of the diagonalizing unitary (Pekker et al., 2014). In another, it is a formal ensemble of site-dependent random tensors with analyzable entropic and equilibration properties (Haferkamp et al., 2021). In a third, it is a constrained manifold for long-time disordered dynamics whose reliability is regime dependent (Kloss et al., 2017). The unifying content is that disorder changes the natural tensor basis, the relevant bond-dimension diagnostics, and the distinction between typical and worst-case representability.