Quantum computing · Theory (AI-assisted)

Fast mixing and thermal phases from strong decay of correlations

Chao Yin, Vedika Khemani

arXiv:2611.17433v1 ↗PDF ↗Submitted 12 November 2026

Question
Glauber dynamics resamples one site at a time from its conditional distribution; when its spectral gap stays bounded as the system grows, Monte Carlo sampling is fast. The standard sufficient condition, strong spatial mixing, demands that correlations decay even next to a fixed boundary, and it is believed to fail in parts of the high-temperature Potts phase and in the 3D Ising model in a field, where fast mixing was proven only with tools that break under small perturbations. Which property of the Gibbs state alone guarantees a constant gap?
Result
The authors define strong decay of correlations: conditioning any distant region on any configuration, however atypical, changes the distribution of a small region by at most \(f(r)\) at distance \(r\). Along a smooth path of commuting Gibbs states with this property they construct exact, time-\(O(1)\), quasi-local Lindbladians connecting the states in both directions, and quasi-local unitaries connecting their purifications. If the path starts at a product state and \(f\) decays faster than \(r^{-3d}\), single-site Glauber dynamics has a gap independent of system size. This proves fast mixing for the \(q\)-state Potts model throughout its high-temperature phase in any dimension, for low-temperature Ising models in a field with frustrating couplings, and for the 3D \(\mathbb Z_2\) gauge theory in its Higgs regime.

The paper in detail

Glauber dynamics is the simplest model of thermalization and the workhorse of Monte Carlo: pick a site, resample it from its distribution given its neighbours. Its relaxation time is the inverse of its spectral gap, and a gap bounded below independent of the number of sites \(N\) means sampling takes at most polynomial time. Which Gibbs states have such a gap is a long-standing question. The Dobrushin criterion works only at weak coupling. Strong spatial mixing works more generally, but it requires correlations to decay even along a frozen boundary, and boundaries can order on their own.

The paper's condition is point-to-set. For a region \(A\) of bounded diameter and the set \(C\) of sites at distance at least \(r\), strong decay of correlations asks that \(\max_{z_C}\lVert\pi_A(\cdot\mid z_C)-\pi_A\rVert_1\le f(r)\). Conditioning on atypical configurations is what makes it strong: a rare all-minus region in a weak field can flip its surroundings while barely contributing to mutual information. Unlike strong spatial mixing, the condition lets \(A\) be far from the conditioned region only, so boundary order does not violate it.

The first theorem builds an adiabatic connection. Encode the Gibbs distribution as a quantum state \(\lvert\sqrt\pi\rangle=\sum_z\sqrt{\pi(z)}\lvert z\rangle\). As the Hamiltonian changes along a path, \(\partial_s\lvert\sqrt\pi\rangle=-\tfrac12(\partial_sH-\langle\partial_sH\rangle)\lvert\sqrt\pi\rangle\). Each local term is split into pieces of growing range by repeatedly replacing it with its conditional expectation given ever larger spheres; correlation decay makes the long pieces small, and the Markov property makes each piece average to zero inside its sphere. That zero average lets each piece be written as an antisymmetric, hence Hermitian, generator, and also as a resampling Lindbladian, so both connections are exact and quasi-local.

The second theorem turns this into a gap. Glauber dynamics is similar to a frustration-free Hamiltonian whose unique ground state is \(\lvert\sqrt\pi\rangle\). If \(\lvert\sqrt\pi\rangle=U\lvert{+}\cdots{+}\rangle\) for a quasi-local unitary \(U\), then in the rotated frame each local ground-state projector of the Glauber Hamiltonian is close to forcing its interior to \(\lvert+\rangle\). Lieb–Robinson bounds on \(U\) and correlation decay make that precise, and any excitation then costs a constant energy. Applications use known weak spatial mixing for Potts, and Peierls arguments along a path from a strong field for the frustrated Ising model and the gauge theory.

Open questions are whether the mixing is in fact logarithmic, whether correlation decay at a single state suffices without a path, and whether a version of the condition holds throughout every thermal phase, which would make quasi-local connection a universal definition of thermal phases. A companion paper uses the condition to rule out passive thermal quantum memories in stabilizer codes without local relations among checks.

Transferable methods

  1. Telescoping conditional expectations into a quasi-local generator

    The change of the Gibbs state along a path is generated by \(\partial_sH-\langle\partial_sH\rangle\), which is local but not orthogonal to the state in any local sense. Replacing each term by its conditional expectations on growing spheres produces a telescoping sum of pieces that are both short-range-dominated and locally mean-zero, which is exactly what is needed to write them as Hermitian generators or Lindbladians.

    For a local term \(h=\partial_sH_u\), let \(\widetilde V_\ell=\pi(h\mid z_{\partial B_L})\) with \(L=2^\ell-1\), and set \[V_1=h-\widetilde V_1,\qquad V_\ell=\widetilde V_{\ell-1}-\widetilde V_\ell,\qquad V_{\ell_{\max}}=\widetilde V_{\ell_{\max}-1}-\langle h\rangle.\] The sum telescopes to \(h-\langle h\rangle\). Strong decay of correlations bounds \(\lvert\widetilde V_\ell-\langle h\rangle\rvert\le\lVert h\rVert f(L)\), so \(\lVert V_\ell\rVert\) decays with \(\ell\).

    By the Markov property, sampling \(h\) given \(z_{\partial B_L}\) is the same as first sampling the smaller sphere and then \(h\); hence \[\sum_{z_{B_L}}V_\ell\,\pi(z_{B_L}\mid z_{\partial B_L})=0.\] On \(\lvert\sqrt\pi\rangle\) this means \(V_\ell\lvert\sqrt\pi\rangle\) is orthogonal, block by block, to the conditional state \(\lvert\psi\rangle\), so \(W_\ell=-\tfrac i2\sum\lvert z\rangle\langle z\rvert\otimes(\lvert\phi\rangle\langle\psi\rvert-\lvert\psi\rangle\langle\phi\rvert)\) satisfies \(W_\ell\lvert\sqrt\pi\rangle=-\tfrac i2V_\ell\lvert\sqrt\pi\rangle\) with \(\lVert W_\ell\rVert\le\tfrac12\lVert V_\ell\rVert\). For the mixed state, \(-V_\ell\pi=\lVert V_\ell\rVert(\sigma_\ell-\pi)\) with \(\sigma_\ell=\pi(1-V_\ell/\lVert V_\ell\rVert)\), a valid distribution because of the zero mean, so a resampling Lindbladian does the job.

    Editorial check: on a 12-site Ising chain with a field, the telescoping term built from spheres of radius 1 and 3 has conditional mean zero to \(3\times10^{-15}\) given the outer sphere, and its size (0.05) is far below that of the original term (1.4).

    How to reuse it

    Whenever a perturbation of a Markov random field must be generated locally, telescope it through conditional expectations on nested spheres. The Markov property gives local mean-zero pieces for free, and point-to-set correlation decay controls their size; the same pattern underlies quasi-adiabatic continuation for gapped ground states.

  2. A gap from a quasi-local disentangler

    A constant gap for a frustration-free Hamiltonian is usually proved perturbatively. Here the ground state \(\lvert\sqrt\pi\rangle\) is mapped to a product state by a quasi-local unitary, and in that frame each local thermal projector nearly pins its interior to \(\lvert+\rangle\), so any deviation from the product state is penalized locally.

    Detailed balance makes \(\mathcal H_u=-\pi^{-1/2}\mathcal L_u\pi^{1/2}\) Hermitian and positive, \[\mathcal H_u=\sum_{z_{\partial u}}\lvert z_{\partial u}\rangle\langle z_{\partial u}\rvert\otimes\bigl(I-\lvert\sqrt{\pi_u(\cdot\mid z_{\partial u})}\rangle\langle\cdot\rvert\bigr),\] with frustration-free ground state \(\lvert\sqrt\pi\rangle\); its gap equals that of the Glauber generator. A finite-volume Poincaré inequality gives \(\sum_{u'\in B_u}\mathcal H_{u'}\ge\Delta_0(I-\Pi_u)\), where \(\Pi_u\) projects onto conditional Gibbs states in the ball \(B_u\).

    Conjugating by \(U\), Lemma 14 shows \(\sum_rc_d(r+1)^{d-1}\sup_u\lVert Q_u\widetilde\Pi_uQ_{u,r}\rVert\le\tfrac12\), where \(Q_u\) flags a non-\(+\) site at \(u\) and \(Q_{u,r}\) flags the nearest non-\(+\) site at distance \(r\). This needs both the Lieb–Robinson tail of \(U\) and correlation decay, and it is where decay faster than \(r^{-2d}\) (and \(r^{-3d}\) along the path) enters. Summing over \(u\) then gives \(\Delta\ge(2c_dR^d)^{-2}e^{-6\cdot3^d\beta R^d}\).

    The paper argues the decay rate cannot be weakened to \(r^{-1}\): the 3D \(\mathbb Z_2\) gauge theory at low temperature mixes slowly, and in it a flux line pinned at distance \(r\) wanders over an area of order \(r\), giving correlations decaying only as \(1/r\).

    How to reuse it

    For any frustration-free Hamiltonian whose ground state is a quasi-local unitary image of a product state, the same frame change can give a gap beyond perturbation theory, provided local ground-state projectors can be shown to nearly fix their interiors. The authors suggest this route for stability of quantum ground-state phases.

Source: Definition 3 (Eq. (31)); Theorem 10 and Sec. 4.3 (Eqs. (55)–(81)); Theorem 11 and Sec. 5.3 (Eqs. (82)–(100)). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It answers an open question about the Potts model and shows that fast mixing is a property of the thermal phase, not of monotonicity tricks that a small frustrating coupling destroys. The quasi-local connections give a candidate definition of thermal phases analogous to the one for gapped ground states. The proof is quantum in method: Lieb–Robinson bounds control a classical Markov chain.

Limitations. The gap bound is a constant but astronomically small (\(e^{-6\cdot3^d\beta R^d}\)), and it implies polynomial, not logarithmic, mixing time. The results need a path to a product state along which the property holds, and they cover only commuting Hamiltonians with periodic or free boundaries, not arbitrary fixed boundaries. Whether the property for a single state suffices is open. The authors state that ChatGPT was used to help develop the proofs and write the manuscript, and that they take responsibility for correctness.

Description based on the full arXiv HTML of v1, read 2026-11-16: Sections 1–3, Section 4 (Theorem 10 and its proof), Section 5.1–5.2 and the opening of 5.3 (Theorem 11, Lemmas 13–14), Sections 6.3–6.5, Section 8 and the acknowledgements; the proofs of Lemmas 14–15, Section 6.1–6.2 and the quantum extension of Section 7 were not checked. Editorial checks: Example 6 (mutual information about \(1/(4n)\) while the ratio correlation stays at \(1/2\)) and the orthogonality of the telescoping terms on a 12-site Ising chain, exact to \(10^{-15}\).