Quantum Approximate Thermalization

• Quantum approximate thermalization is the phenomenon where local observables in a globally pure quantum state display thermal-like behavior due to entanglement and mechanisms like ETH.
• Methodologies include rigorous high-temperature bounds, analysis of reduced density matrices, and the use of weak ETH to demonstrate local convergence to microcanonical or Gibbs ensembles.
• Implications extend to engineered quantum simulations and variational algorithms, highlighting both practical state preparation techniques and limitations in thermalization under controlled perturbations.

Searching arXiv for recent and foundational work on quantum approximate thermalization. Quantum approximate thermalization denotes the emergence of thermal behavior in a quantum system only at the level of restricted observables, subsystems, asymptotic bounds, or approximate state representations, rather than as exact convergence of the full many-body pure state to a thermal density matrix. In isolated systems, unitary Schrödinger evolution preserves purity and the amplitudes in the energy eigenbasis, so exact global thermalization of a pure state is excluded; what can occur instead is local indistinguishability from a microcanonical or Gibbs ensemble, suppression of temporal fluctuations, or convergence of reduced density matrices and few-body observables to thermal values. Across the literature, this phenomenon is explained by the eigenstate thermalization hypothesis (ETH), weak ETH, entanglement growth, locality, spectral nonresonance conditions, and, in some settings, engineered dissipation or variational state-preparation schemes (Polkovnikov et al., 2017). More recent work has sharpened the notion into explicit finite-size and thermodynamic-limit bounds, for example proving local convergence to Gibbs states in translation-invariant high-temperature qubit systems under nondegenerate spectral gaps (Pilatowsky-Cameo et al., 2024).

1. Conceptual definition and scope

The basic distinction underlying quantum approximate thermalization is between the global state and its local manifestations. In a closed quantum system, if the initial state is pure, the full state remains pure under unitary dynamics. The density matrix of the complete system therefore does not literally relax into a thermal mixed state. Nonetheless, local observables can relax, and reduced density matrices of small subsystems can become mixed and thermal-like because of entanglement with the rest of the system (Polkovnikov et al., 2017).

A standard formulation uses the reduced density matrix

$$\rho_A = \mathrm{Tr}_B \left( |\Psi\rangle \langle \Psi| \right),$$

for a subsystem $A$, where $B$ is the complement. Even when the global state has zero von Neumann entropy, $\rho_A$ can have nonzero entropy, quantified for example by

$S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$

This separation between global purity and local mixedness is the core structural reason that thermal behavior is compatible with reversible microscopic dynamics (Polkovnikov et al., 2017).

The thermal reference state depends on the setting. In isolated systems with fixed energy, the natural benchmark is often the microcanonical ensemble. In high-temperature rigorous results and in open or engineered settings, the Gibbs state

$$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$

is the relevant target (Pilatowsky-Cameo et al., 2024). Approximate thermalization then means that, for a fixed finite subsystem $A$, the evolved state becomes locally indistinguishable from $g_\beta$, or that expectation values of local observables approach the thermal prediction up to an error that vanishes with system size or with an explicit control parameter.

The literature uses the term in several related but non-identical senses. In isolated chaotic systems it means local ETH-type equilibration. In weak-ETH analyses it refers to asymptotic convergence for almost all admissible initial states and nearly all late times (Dabelow et al., 2022). In open-system models it means convergence to a Gibbs-like steady state up to controlled high-temperature corrections (Schrinski et al., 2022). In quantum algorithms it can denote approximate preparation of thermal states or thermal-like distributions rather than physical equilibration dynamics (Foldager et al., 2021, Díez-Valle et al., 2022). This suggests that “quantum approximate thermalization” functions as an umbrella concept rather than a single formal definition.

2. ETH, weak ETH, and the mechanism of local equilibration

The dominant microscopic explanation is ETH. In its simple form, ETH states that a single energy eigenstate is equivalent to a microcanonical ensemble for physical observables. For a local observable $\hat{O}$, the expectation value in an energy eigenstate $|E_n\rangle$ obeys

$A$0

or, in the notation used in one review,

$A$1

If the initial state has a narrow energy distribution, local observables then relax to ensemble values because individual eigenstates already encode thermal behavior (Polkovnikov et al., 2017).

A more refined ETH matrix ansatz appears in studies connecting thermalization, chaos, and information-theoretic structure: $A$2 with $A$3 and $A$4. Here the diagonal term gives the smooth thermal value, while the off-diagonal term controls temporal fluctuations and relaxation rates (Qasim et al., 30 Oct 2025). In this framework, approximate thermalization is directly tied to the entropy-suppressed off-diagonal matrix elements.

Weak ETH relaxes the requirement that every eigenstate in an energy window be thermal. Instead, it requires that the variance of diagonal matrix elements in a microcanonical shell vanish in the thermodynamic limit: $A$5 This weaker condition is sufficient to prove thermalization for two broad classes of initial conditions: typical pure states with a prescribed nonequilibrium local expectation value, and Gibbs states subjected to a local quench $A$6 with $A$7 local (Dabelow et al., 2022). The result is explicitly asymptotic and approximate: local observables relax to the corresponding microcanonical values, corrections vanish with increasing system size, and exceptional states or times are exponentially rare.

One consequence is that integrability does not automatically preclude thermalization after local perturbations. Because weak ETH can hold even in some integrable models, approximate local thermalization can emerge under conditions much broader than those usually associated with strong ETH (Dabelow et al., 2022). A plausible implication is that the relevant distinction is not simply integrable versus nonintegrable, but rather the interplay of locality, the class of observables, and the structure of the initial state.

3. Rigorous high-temperature results and local indistinguishability

A major recent development is the derivation of local thermalization from first principles under explicit assumptions, without positing ETH as an axiom. For translation-invariant local qubit Hamiltonians, thermalization has been proved for typical initially unentangled pure states drawn from a maximum-entropy ensemble, assuming three conditions: high effective temperature, translation invariance, and no perfect resonances in the form of nondegenerate spectral gaps (Pilatowsky-Cameo et al., 2024).

The local thermalization criterion is expressed through the local trace norm

$A$8

where the maximization runs over observables supported on a fixed finite connected region $A$9. The main bound is

$B$0

with $B$1. Thus, in the thermodynamic limit, local observables cannot distinguish the evolved pure state from the Gibbs state (Pilatowsky-Cameo et al., 2024).

The proof decomposes into equilibration to the diagonal or time-averaged state, and comparison of that state to the Gibbs ensemble. The first step uses nondegenerate spectral gaps and small inverse participation ratio (IPR),

$B$2

to show local equilibration. The second step uses a weak-ETH-type statement for diagonal states with finite correlation length. The key intermediate concept is the “energy-delocalized Gibbs ensemble” (EDGE), defined by closeness of the ensemble average to Gibbs and by energy delocalization in the sense of small average IPR (Pilatowsky-Cameo et al., 2024).

High temperature is essential for two reasons. First, the Gibbs state is separable for sufficiently small $B$3, allowing a decomposition into product states over the alphabet

$B$4

Second, the Gibbs state has finite correlation length at sufficiently high temperature, which is needed for the weak-ETH comparison step (Pilatowsky-Cameo et al., 2024). Translation invariance then yields a strong IPR estimate,

$B$5

by distinguishing periodic or approximately periodic states from aperiodic ones. This suggests that symmetry can play a direct role in converting separability plus locality into rigorous thermalization.

4. Entanglement, subsystem thermalization, and finite systems

Entanglement is the operational bridge between unitary many-body dynamics and local thermal behavior. In small and intermediate-size isolated systems, local thermalization is typically diagnosed through reduced density matrices, entanglement entropies, and local occupation statistics rather than through any claim of global thermal mixedness (Polkovnikov et al., 2017).

An experimentally central example is the six-boson, six-site lattice system studied by Kaufman and collaborators, highlighted in a review of chaos and thermalization in small quantum systems. The system had Hilbert-space dimension $B$6, was initialized with one boson per site in two identical copies, and was then subjected to a quantum quench that allowed particle hopping. After a short transient time, the one-site and two-site reduced density matrices became indistinguishable from a thermal ensemble, while direct measurements of particle occupation distributions agreed with equilibrium thermal predictions. At the same time, the full isolated state remained pure (Polkovnikov et al., 2017). This is an archetypal realization of approximate thermalization: local thermalization without global entropy production.

Related numerical work emphasizes that thermalization can occur already inside a Krylov subspace much smaller than the full Hilbert space. For hard-core bosons on a two-dimensional lattice, thermalization was found to take place within a dynamically generated subspace

$B$7

with dimension typically of order $B$8, even when the full Hilbert-space dimension reached $B$9 or larger (Khlebnikov et al., 2013). The projected Hamiltonian in the Lanczos basis assumes tridiagonal form and can be interpreted as a finite one-dimensional Anderson-localization problem. In that setting, ETH-like smoothness of diagonal matrix elements and suppression of off-diagonal ones were observed in the dynamically relevant sector rather than over the entire Hilbert space.

The same work argued that subsystem thermalization can precede full-system equilibration because entanglement entropy grows rapidly. For a bipartition into $\rho_A$0 and $\rho_A$1, the reduced state $\rho_A$2 can become close to a thermal density matrix well before all observables of the full system settle, because particle transport across the interface generates large entanglement. This reinforces a recurring theme: approximate thermalization is often best understood as an entanglement phenomenon rather than as literal state conversion (Khlebnikov et al., 2013).

5. Failures, obstructions, and partial thermalization

Approximate thermalization is neither universal nor observable-independent. Several classes of exceptions or partial failures are emphasized in the literature.

Quantum many-body scars (QMBS) are exceptional eigenstates embedded in otherwise thermal spectra. A recent analysis of perturbed scarred systems distinguishes exact scars protected by a restricted spectrum-generating algebra (RSGA) from approximate scars lacking exact algebraic closure. For exact QMBS under local perturbation $\rho_A$3, an improved lower bound on the thermalization time was derived: $\rho_A$4 improving on a previous $\rho_A$5 bound. For approximate QMBS under generic perturbations, the relaxation time instead follows

$\rho_A$6

consistent with second-order perturbation theory and Fermi’s golden rule (Mao et al., 25 Feb 2026). Boundary conditions can switch a model between exact and approximate scar behavior; in the deformed PXP model, periodic boundary conditions give exact scars with $\rho_A$7, whereas open boundary conditions produce approximate scars with $\rho_A$8 (Mao et al., 25 Feb 2026). This establishes that approximate thermalization can be parametrically delayed by hidden algebraic structure, and that “more exact” scar structure does not necessarily imply slower decay under generic perturbations.

A field-theoretic example of partial thermalization appears in the massive Schwinger model used as a proxy for quark-gluon plasma dynamics. There, the equal-time Wigner function is decomposed into scalar, pseudoscalar, vector, and axial-vector components. Thermalization is diagnosed by comparison to microcanonical and canonical ensemble values and by the dimensionless deviation

$\rho_A$9

In the strong-coupling case $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$0, all Wigner components thermalize. In the weaker-coupling case $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$1, thermalization is partial: the pseudoscalar $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$2 and vector $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$3 components thermalize, while the scalar $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$4 and axial-vector $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$5 do not (Chen et al., 2024). The failure is traced to approximate scar states affecting parity-even sectors, and the thermalization pattern can swap when the topological angle is changed to $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$6. This demonstrates that approximate thermalization can be symmetry-dependent and operator-dependent rather than a blanket property of a state (Chen et al., 2024).

Weak-ETH analyses also emphasize limits. The results apply to local observables, short-range Hamiltonians, and post-quench Hamiltonians satisfying weak ETH; many-body localized systems are not expected to thermalize under analogous conditions (Dabelow et al., 2022). Open-system high-temperature Gibbs approximation for the planar rotor likewise breaks down at low temperature, where the steady state can deviate strongly from Gibbs and may remain localized or require tunneling processes to reach the global minimum (Schrinski et al., 2022). These examples show that approximate thermalization is conditional, not generic in the absence of the relevant dynamical ingredients.

6. Structural reformulations, diagnostics, and related frameworks

One structural reformulation identifies thermality with an approximate Markov property. In one-dimensional systems, any quantum $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$7-approximate Markov chain,

$S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$8

is close in relative entropy to a Gibbs state of a short-range Hamiltonian, and conversely Gibbs states of 1D short-range Hamiltonians have small conditional mutual information (CMI) decaying with the size of the buffer region (Kato et al., 2016). For a tripartition $S_A = -\mathrm{Tr}(\rho_A \ln \rho_A).$9, the CMI is

$$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$0

For 1D Gibbs states, the paper derives decay essentially of the form

$$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$1

and proves recoverability bounds that imply efficient constant-depth preparation of finite-temperature 1D Gibbs states (Kato et al., 2016). This provides a non-ETH route to approximate thermality: locality plus small CMI can characterize thermal states.

Another line of work relates approximate infinite-temperature thermalization to quantum state designs. If a small subsystem $$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$2 of a bipartite pure state is close to maximally mixed,

$$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$3

then measuring the complement in a Haar-random orthonormal basis produces, with high probability, an $$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$4-approximate quantum state $$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$5-design on $$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$6, with

$$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$7

under a dimension condition on the purifying subsystem (Wilming et al., 2022). This result formalizes a pseudorandomness consequence of approximate thermalization rather than a dynamical criterion.

Chaos diagnostics have also become a tool for studying thermalization. In a trapped-ion digital quantum simulation of a $$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$8D $$g_\beta=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}$$9 lattice gauge theory, randomized-measurement protocols were used to learn an approximate entanglement Hamiltonian and extract the gap-ratio distribution and entanglement spectral form factor. The development of level repulsion and a ramp-plateau structure was interpreted as universal early-time evidence of quantum chaos, regarded there as a prerequisite for thermalization (Mueller et al., 2024). Although this is not itself a proof of thermalization, it offers an experimentally scalable diagnostic pathway when full state tomography is inaccessible.

A more speculative but technically explicit connection appears in work linking ETH, approximate quantum error correction, and the chaos bound. There, dynamical fluctuation errors,

$A$0

are bounded by ETH spectral envelopes, and further constrained through the Lyapunov exponent $A$1 using chaos-bound-inspired estimates (Qasim et al., 30 Oct 2025). This suggests that the same off-diagonal ETH data governing approximate thermalization also constrains correctability and scrambling, though this connection is formulated within a narrow regime of chaotic systems with a hierarchy $A$2.

7. Engineered, computational, and variational forms of approximate thermalization

Approximate thermalization also appears as a controlled target in open-system engineering and quantum computing.

An engineered-bath protocol for analog quantum simulation couples local degrees of freedom to driven, dissipative ancilla pseudospins whose energy splittings $A$3 are periodically swept across the system bandwidth. Under the timescale hierarchy

$A$4

one obtains a Born-Markov master equation with frequency-resolved Lindblad operators and Lorentzian transition rates (Metcalf et al., 2019). The Gibbs state

$A$5

is an exact fixed point if detailed balance holds, and an approximate fixed point when detailed-balance violations remain small near resonance (Metcalf et al., 2019). Numerical studies found small trace distance, often $A$6, for a two-spin system and good performance up to $A$7 spins, with the hardest regime at low temperature or for congested spectra (Metcalf et al., 2019).

For open quantum dynamics, the planar rotor with periodic potential provides a complementary example. Its master equation combines decoherence, friction, and momentum diffusion, with $A$8. The Gibbs state $A$9 is stationary up to corrections

$g_\beta$0

so the steady state is Gibbs-like in the high-temperature regime (Schrinski et al., 2022). This is a textbook instance of approximate thermalization to a steady state rather than dynamical local equilibration in a closed system.

On NISQ-oriented quantum hardware, “thermalization” often refers to preparing Gibbs states approximately. The Noise-Assisted Variational Quantum Thermalizer (NAVQT) exploits explicit depolarizing noise,

$g_\beta$1

within a variational circuit and minimizes an approximate free energy

$g_\beta$2

thereby avoiding purification and ancillas (Foldager et al., 2021). On system sizes $g_\beta$3 to $g_\beta$4, NAVQT achieved fidelities above $g_\beta$5 for uniform Ising chains with and without a transverse field, while random Heisenberg models were hardest and could fall below $g_\beta$6 at some temperatures (Foldager et al., 2021). The learned noise level interpolated correctly between $g_\beta$7 at very high temperature and $g_\beta$8 at very low temperature. This is approximate thermalization in an algorithmic state-preparation sense, not a claim about spontaneous equilibration.

Single-layer QAOA yields a different, explicitly non-ETH notion: “pseudo-Boltzmann states.” For certain universal Ising models, the measurement statistics in the computational basis approximately satisfy

$g_\beta$9

even though the final state is pure and the dynamics are short and circuit-based rather than chaotic (Díez-Valle et al., 2022). The effective inverse temperature

$\hat{O}$0

emerges from a hidden correlation between state energy and the covariance of energies with Hamming distances in the energy landscape (Díez-Valle et al., 2022). The paper explicitly distinguishes this from genuine thermal equilibration. It is thermal-like only in output statistics.

A related numerical perspective studies approximate many-body ansatz states themselves. Across tensor-network, neural-network, and quantum-circuit ansatz families, the spectral decomposition of approximate states in the exact eigenbasis often exhibits an empirical exponential law,

$\hat{O}$1

with fitted inverse effective temperature $\hat{O}$2 (Chen et al., 2024). For approximate ground states, $\hat{O}$3 is often below $\hat{O}$4; for imaginary-time-evolved targets, $\hat{O}$5 tracks the target $\hat{O}$6 only up to a crossover $\hat{O}$7, after which a high-energy plateau appears (Chen et al., 2024). This is not thermalization in the dynamical sense, but it indicates that thermal-like spectral organization can emerge as an ansatz artifact or expressibility diagnostic.

Overall, the research landscape shows that quantum approximate thermalization is best understood as a family of rigorously and experimentally distinguishable phenomena: local Gibbs or microcanonical indistinguishability in closed systems, asymptotically vanishing observable deviations under weak ETH, high-temperature theorems derived from translation invariance and spectral nonresonance, Gibbs-like steady states in open systems, and thermal-like probability laws or approximate Gibbs preparation in quantum algorithms. This suggests that future unification will likely proceed not by collapsing these notions into a single definition, but by clarifying the maps between local indistinguishability, entanglement structure, recoverability, chaos diagnostics, and computational state representations (Pilatowsky-Cameo et al., 2024).