Random Matrix Product States (RMPS)

• Random Matrix Product States are ensembles of quantum wavefunctions constructed via sequential application of Haar-random unitaries, capturing key statistical properties of thermal and chaotic states.
• They enable polynomially efficient simulations of quantum dynamics by approximating Haar-random ensembles and reproducing typical entanglement behavior with controlled errors.
• RMPS serve as a versatile framework in quantum statistical mechanics and information, bridging low-entanglement ground states and fully random pure states for diverse applications.

Random Matrix Product States (RMPS) are ensembles of quantum many-body wavefunctions constructed by sequentially applying Haar-random unitaries to a set of local sites and ancillary degrees of freedom, leading to tensor network states that efficiently capture core statistical properties of fully random states. RMPS provide a physically motivated, computationally efficient framework whose statistical and entanglement features closely mimic Haar-random vectors in Hilbert space while requiring only polynomially many parameters. RMPS play a foundational role in the study of quantum statistical mechanics, quantum information, and the simulation of chaotic or thermalizing quantum systems, and serve as a controlled interpolation between product states, low-entanglement ground states, and genuinely random pure states.

1. Formal Definition and Ensemble Construction

An RMPS is defined for a 1D quantum chain of $N$ sites, each of local dimension $d$, with maximal bond dimension $D$ (or $\chi$). The canonical construction proceeds as follows:

• Attach a $D$-dimensional ancillary space to the system.
• For each site $k=1,\dots,N$, apply a Haar-random unitary $U^{[k]}$ on the site and the ancilla. The output is mapped to a collection of MPS site tensors $$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$, with physical index $i$ and ancilla indices $\alpha,\beta$.
• The RMPS wavefunction is then

$d$0

where $d$1, $d$2 are fixed boundary ancilla vectors.

Sampling each $d$3 independently from the Haar measure induces the "non-homogeneous" RMPS ensemble. The homogeneous ensemble fixes $d$4 for all $d$5.

Alternative constructions utilize right-normalized or left-canonical MPS forms or block-unitary encodings, embedding each local tensor into a larger Haar-random unitary and projecting along fixed ancilla directions (Garnerone et al., 2010, Lami et al., 2024, Leontica et al., 30 Apr 2025).

2. Statistical and Typicality Properties

Ensemble Average and Concentration

The key feature of RMPS is that, for any such ensemble,

$d$6

independent of the ancilla dimension $d$7 or $d$8 (Garnerone et al., 2010, Garnerone et al., 2013). This establishes that all first-moment (mean) expectation values of observables in RMPS exactly reproduce the Haar (microcanonical) ensemble.

Concentration-of-measure techniques show that for any observable $d$9 with support on $D$0 contiguous sites, the variance of $D$1 over the RMPS ensemble is exponentially suppressed in $D$2 and the local dimension. Explicitly,

$D$3

for suitable constants $D$4 and subsystem size $D$5 (Garnerone et al., 2010, Garnerone et al., 2013).

For subsystem reduced density matrices $D$6 of size $D$7, similar concentration yields

$D$8

with $D$9 scaling as $\chi$0 (Garnerone et al., 2010).

Entanglement Structure and Maximal Entropy

Small subsystems (blocks) of a large RMPS are nearly maximally mixed provided the bond dimension $\chi$1 satisfies $\chi$2, where $\chi$3 is the block size. The normalized reduced state $\chi$4 satisfies

$\chi$5

and the von Neumann entropy of the block approaches $\chi$6 up to corrections vanishing in $\chi$7 (Collins et al., 2012, Haferkamp et al., 2021). There is a volume-law for 2-Rényi entropy in sufficiently disconnected subsystems and almost-maximal entropy for small, connected blocks (Haferkamp et al., 2021).

Higher Moments and Magic

For large $\chi$8, RMPS reproduce not just the first moment (uniform average) but can approximate higher-moment properties of Haar states. Specifically, RMPS ensembles are approximate 2-designs with corrections of $\chi$9 in the operator norm for $D$0 (Garnerone, 2013, Lami et al., 2024).

The so-called "magic" (nonstabilizerness) of RMPS, measured via the $D$1-norm over the Pauli coefficients of $D$2, is extensive with system size $D$3: $D$4 scales linearly in $D$5 with overwhelming probability. Numerical and analytic results confirm that for qubits and moderate $D$6, RMPS are far more magical than product or Clifford states, achieving magic approaching the maximum entropic bound (Chen et al., 2022).

3. Sampling, Computational Methods, and Bias

Sequential and Unbiased Sampling

The canonical procedure for generating RMPS involves product sampling of local Haar-random unitaries (the "sequential RMPS"). However, this measure is not uniform with respect to the restriction of the Fubini–Study measure on the full Hilbert space to the MPS manifold. The sequential method exhibits an entanglement asymmetry under inversion, with marginal distributions of Schmidt eigenvalues distinct for different chain cuts (Leontica et al., 30 Apr 2025).

An unbiased measure can be constructed using the left-canonical form and associated gauge-fixed coordinates. The correct Fubini–Study measure is then

$D$7

where $D$8 is the right-environment matrix at bond $D$9. Numerical and analytic studies confirm that the unbiased ensemble eliminates entanglement asymmetry and modifies the entanglement spectrum compared to the Haar-RMPS case (Leontica et al., 30 Apr 2025).

Metropolis–Hastings Algorithm

Sampling from the unbiased ensemble is achieved via a Metropolis–Hastings scheme, which proposes updates that preserve left-canonical form, with acceptance probability determined by the changes in the measure's weighting factor $k=1,\dots,N$0 (Leontica et al., 30 Apr 2025).

Monte Carlo and Observables

Monte Carlo sampling of RMPS can be used to compute thermal or microcanonical expectation values. Typical observables (e.g., magnetization, correlations) are estimated by evaluating expectation values in sampled RMPS and averaging. For each trial, with $k=1,\dots,N$1 sites and bond dimension $k=1,\dots,N$2, the computational cost per sample is $k=1,\dots,N$3 and $k=1,\dots,N$4 samples suffice for $k=1,\dots,N$5 accuracy, where the error scales as $k=1,\dots,N$6 (Garnerone, 2013).

4. Applications and Physical Implications

Quantum Statistical Mechanics

RMPS efficiently model foundational aspects of statistical physics, including the emergence of microcanonical, canonical, and generalized canonical ensembles from pure states with low circuit complexity. Subsystem thermalization and typicality, normally attributed to highly entangled Haar-random states, are reproduced within the RMPS framework by entanglement and large bond dimension alone (Garnerone et al., 2010, Haferkamp et al., 2021, Collins et al., 2012).

Simulation and Computation

RMPS enable efficient stochastic simulation of thermal and out-of-equilibrium quantum dynamics, as only polynomial resources in $k=1,\dots,N$7 are required to reproduce subsystem properties and expectation values that would conventionally require exponential resources for Haar sampling (Garnerone et al., 2013, Garnerone, 2013). RMPS have proven utility in tensor network algorithms and variational ansätze in ground-state and thermal-state approximations.

Quantum Information and Magic

The extensive magic of RMPS implies that they provide a rich resource for quantum computation beyond the stabilizer (Clifford) regime; RMPS can evade expressivity obstacles seen in variational states limited to Clifford or low-magic subspaces (Chen et al., 2022, Lami et al., 2024).

The action of global Clifford unitaries on RMPS realizes an ensemble of Clifford-enhanced MPS (C MPS) that exactly forms a 3-design and approximates 4-designs to error $k=1,\dots,N$8 (Lami et al., 2024). This demonstrates that moderate $k=1,\dots,N$9 and Clifford post-processing suffice to approximate high-order Haar randomness and magic, relevant for randomized benchmarking and classical simulation tasks.

Gravity and Topological Expansions

RMPS have been leveraged as exact finite-$U^{[k]}$0 models for gravitationally prepared states in 2D quantum gravity path integrals. The transfer-matrix formalism and spectral properties of RMPS directly reproduce key gravitational phenomena such as the bra–ket wormhole phase transition and allow explicit computation of higher topology and off-shell wormhole contributions to all orders (Jung et al., 12 Dec 2025). Continuous RMPS models (cRMPS) generalize these results to the continuum limit.

5. Limitations, Extensions, and Open Directions

Entanglement Scaling and Design Properties

While RMPS are approximate 2-designs for sufficiently large $U^{[k]}$1, they generally cannot reproduce all properties of Haar-random states at finite $U^{[k]}$2. The 2-stabilizer Rényi entropy of RMPS converges to the Haar value as $U^{[k]}$3, and Clifford enhancement is required to reach approximate 4-designs (Lami et al., 2024). For finite $U^{[k]}$4, edge effects and lack of uniform measure in the sequentially generated ensemble may manifest, although unbiased sampling corrects these issues (Leontica et al., 30 Apr 2025).

Connection to Tensor Networks Beyond 1D

Current theory and numerics focus primarily on 1D chains. Extensions to higher-dimensional tensor network states, such as random PEPS or MERA, remain an active research area with preliminary evidence suggesting analogous typicality phenomena but different scaling of fluctuations and correlations (Chen et al., 2022).

Algorithmic and Experimental Implementation

Practical sampling of RMPS is efficient for moderate $U^{[k]}$5 and $U^{[k]}$6, but the construction of unbiased ensembles and scaling to higher dimensions involve algorithmic overhead. Nevertheless, sequential application of Haar-random gates is implementable in quantum circuits, motivating experimental exploration, especially for many-body thermalization and benchmarking tasks (Haferkamp et al., 2021).

6. Representative Results and Quantitative Benchmarks

Table: Scaling Properties and Key Results for RMPS

Property	Scaling / Value	Reference
Mean state $U^{[k]}$7	$U^{[k]}$8	(Garnerone et al., 2010)
Typical variance of observables	$U^{[k]}$9 (for $$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$0 large)	(Garnerone et al., 2010)
2-design approximation error	$$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$1	(Garnerone, 2013)
Purity for block size $$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$2	$$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$3	(Collins et al., 2012)
Magic growth	$$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$4 with $$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$5--$$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$6 for small $$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$7	(Chen et al., 2022)
Convergence of subsystem entropy	$$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$8	(Collins et al., 2012)
4-design error with Clifford	$$A^{i}[k]_{\alpha\beta} = \langle i, \alpha | U^{[k]} | 0, \beta \rangle$$9	(Lami et al., 2024)
MC sample cost per state	$i$0	(Garnerone, 2013)
Metropolis unbiased sampling	Accepts with weighting $i$1	(Leontica et al., 30 Apr 2025)

These results collectively establish that RMPS provide a tractable, flexible, and mathematically rigorous testbed for studying quantum typicality, entanglement, quantum resource theory, and emergent thermodynamics in low- and intermediate-entanglement quantum systems.