Quantum information · Theory (AI-assisted)

High-temperature Gibbs states are uniformly Markovian and trivial

Andrew D. Kim, Tai Hsuan Yang, Jong Yeon Lee

arXiv:2611.26626v1 ↗PDF ↗Submitted 19 November 2026

Question
At high temperature a thermal state should have only short-range correlations, and classifying finite-temperature phases needs this in information-theoretic form: exponential decay of mutual information between distant regions, and of conditional mutual information (CMI) for every arrangement of three regions, including ones where part of the system is traced out. For commuting Hamiltonians or one dimension such bounds were known. For general non-commuting local Hamiltonians the standard cluster expansion of the logarithm of a reduced Gibbs state can diverge, and CMI bounds existed only when the middle region separates the outer two in a full partition of the system. Do high-temperature Gibbs states of arbitrary local Hamiltonians satisfy these bounds?
Result
Yes, above a constant temperature. For \(k\)-local Hamiltonians on qudits of dimension at most \(q\), interaction range \(R\) and interaction strength per site \(J\), if \(\beta J<1/(8e^3k(2q-1)^k)\) then the mutual information obeys \(I(A:C)\le K_{\mathrm{MI}}\min\{|\partial A|,|\partial C|\}e^{-\mathrm{dist}(A,C)/R}\) for all disjoint regions. At a smaller threshold the same form of bound holds for \(I(A:C|B)\) for every disjoint \(A,B,C\) with the rest traced out, a uniform Markov property. The key step writes the Gibbs state as a mixture of products of single-site states drawn from a projective 2-design, with weights obtained by applying the classical-shadow inverse map to \(e^{-\beta H}\). A cluster expansion shows that these weights are positive at high temperature, which also gives separability, and that their logarithms are sums of exponentially decaying local potentials.

The paper in detail

For a state on a lattice of qudits, the mutual information \(I(A:C)\) measures all correlations between regions \(A\) and \(C\), and the conditional mutual information \(I(A:C|B)\) measures what remains once \(B\) is known. Small CMI means \(C\) can be recovered from \(B\) alone, a quantum Markov property. Mixed-state phase classifications ask for both to decay exponentially with distance: MI for regions far apart (no long-range order), CMI for an annulus shielding a disk (local recoverability), and CMI for non-shielding regions with part of a ball traced out (spatial uniformity). The paper proves all of these for Gibbs states \(\rho_\beta\propto e^{-\beta H}\) of any bounded-strength local Hamiltonian at high enough temperature.

Earlier approaches expanded \(\log\rho_{\beta,L}\) for each region \(L\) in clusters of interaction terms, but for non-commuting terms that operator-valued series need not converge uniformly in system size. The paper moves to scalars. On each site it uses a projective 2-design, for qubits the six Pauli eigenstates, and writes \(e^{-\beta H}\) as an average of product projectors weighted by \(\omega_{\mathbf s}(e^{-\beta H})=\operatorname{tr}[(\otimes_xD_{s_x})e^{-\beta H}]\), where \(D_s=(d+1)P_s-I\) is the classical-shadow snapshot. Partial traces act on these weights simply: averaging a site's label replaces its functional by the normalised trace. Every reduced state is therefore described by marginals of a single function of classical labels.

A cluster expansion of each scalar weight converges when \(\beta J(2q-1)^ke^{\beta J+\lambda+1}\le1/(8k)\), uniformly in the region. The weights are then positive, so \(\rho_\beta\) is a mixture of product states with a genuine probability law \(\pi(\mathbf s)\), giving separability as a by-product. For MI, the log-likelihood ratio \(\log[p_{AC}/(p_Ap_C)]\) is a combination of four cluster series in which every cluster missing \(A\) or \(C\) cancels. The survivors span the distance and are exponentially small, and the resulting pointwise bound \(p_{AC}\le e^{\Delta}p_Ap_C\) lifts to \(\rho_{AC}\le e^\Delta\rho_A\otimes\rho_C\) and to the MI bound. For CMI, the label law is a Gibbs distribution with potentials decaying in both size and diameter. Each label's conditional distribution is then, with probability at least \(1/2\), an independent draw, and otherwise a local resampling. Data processing along this resampling contracts the CMI between two sites geometrically in their distance, with the conditioning kept fully quantum.

Stitching two high-temperature Gibbs states across an interface is again a Gibbs state with spatially varying couplings, so the bounds hold across it. Taking one side at infinite temperature shows that the phase is that of the maximally mixed state. On-site noise leaves the classical label law unchanged and only changes the local states, so both bounds survive it. The thresholds are far from the true transition: in a numerical example the weights stay positive at temperatures several hundred times lower than proved. Weak multi-site noise, sharper thresholds, and whether rapid mixing of a Gibbs sampler alone implies the boundary-law CMI bound are left open.

Transferable methods

  1. A positive classical representation from a local 2-design

    The obstruction is that \(\log\) of an operator-valued cluster sum need not converge when the terms do not commute. Measuring every site in a 2-design and reweighting by the shadow inverse turns the Gibbs operator into a function of classical labels. That function is a scalar, so the standard polymer machinery for classical spin systems applies: positivity, convergent \(\log\), and cancellation of clusters that miss a region.

    A 2-design \(\{(\nu(s),P_s)\}\) of rank-one projectors satisfies \(\mathbb E[P_s]=I/d\) and \(\mathbb E[P_s\otimes P_s]=(I+\mathrm{SWAP})/(d(d+1))\). Contracting the second identity with \(I\otimes O\) gives \(\mathbb E[P_s\operatorname{tr}(P_sO)]=(O+\operatorname{tr}(O)I)/(d(d+1))\), hence \[O=d\,\mathbb E_s\bigl[\operatorname{tr}(D_sO)\,P_s\bigr],\qquad D_s=(d+1)P_s-I.\] The functional \(\tau_s(O)=\operatorname{tr}(D_sO)\) is unital, has norm \(2d-1\) (the eigenvalues of \(D_s\) are \(d\) and \(-1\)), and averages to the normalised trace. Applied site by site, \(\rho_\beta=\mathbb E_{\mathbf s}\bigl[p(\mathbf s)\bigotimes_xP_{s_x}\bigr]\) with \(p(\mathbf s)=\omega_{\mathbf s}(e^{-\beta H})/\omega_\varnothing(e^{-\beta H})\).

    Expanding \(e^{-\beta H}\) as a sum over pairwise disjoint polymers (connected sets of interaction terms), the product functional factorises over disjoint supports, \(\omega_{\mathbf s}(e^{-\beta H})=\sum_{\mathcal G}\prod_{\gamma\in\mathcal G}\omega_{\mathbf s}(Q_\gamma)\), with \(|\omega_{\mathbf s}(Q_\gamma)|\le\prod_{X\in\gamma}(2q-1)^k(e^{\beta\|h_X\|}-1)\). This is a classical polymer gas. The Kotecký–Preiss criterion, checked by counting polymers through rooted plane trees (at most \(4^{m-1}\) shapes, each child costing \(k\eta\)), gives an absolutely convergent \(\log\) once \(\eta_\lambda\le1/(8k)\). The weights are then positive, and their logarithm is a sum of potentials \(\Psi_U\) with \(\sup_x\sum_{U\ni x}e^{|U|+2\,\mathrm{diam}(U)/R}\|\Psi_U\|\le2\eta_{k+2}\).

    Editorial check: with the six Pauli eigenstates, the reconstruction formula returns a random complex \(2\times2\) operator exactly, and \(\|\tau_s\|=3=2d-1\). For the Heisenberg chain on six sites (\(J=6\), \(q=2\), \(k=2\)), computing all \(6^6=46{,}656\) weights exactly shows they are all positive up to \(\beta J\approx0.21\). The proven threshold for the MI theorem is \(\beta J<3.5\times10^{-4}\) (and \(9.7\times10^{-6}\) for CMI), so positivity, and hence separability, holds far beyond what the theorem certifies, consistent with the authors' remark that their thresholds are not tight.

    How to reuse it

    Replacing a non-commutative expansion by a positive expansion in an overcomplete basis is a general tool. The same 2-design reweighting can give classical sampling algorithms, learning guarantees and decay-of-correlation proofs for any operator that is close to the identity in a local sense, for example short-time evolutions, weakly noisy states, or high-temperature fermionic states with a suitable local design.

  2. Cluster cancellation for MI, and contraction by resampling for CMI

    Mutual information involves logarithms of three reduced states. With a convergent cluster expansion for each label marginal, the log-likelihood ratio is a signed combination in which every cluster that does not touch both regions appears twice with opposite signs, so only long clusters remain. For CMI the conditioning region is quantum and need not shield, so instead of cancellation the proof uses a dynamical picture: resampling one label at a time is mostly independent of the rest, which forces information between distant sites to decay.

    For labels on \(A\) and \(C\), \[\ell=\log\omega_{(\mathbf s_A,\mathbf s_C)}(e^{-\beta H})+\log\omega_\varnothing(e^{-\beta H})-\log\omega_{\mathbf s_A}(e^{-\beta H})-\log\omega_{\mathbf s_C}(e^{-\beta H}).\] A cluster missing \(C\) contributes equally to the first and third terms and to the second and fourth, so it cancels, and likewise for clusters missing \(A\). Survivors contain an interaction path from \(\partial A\) to \(\partial C\), so their number of terms \(N_\Gamma\) satisfies \(RN_\Gamma\ge\mathrm{dist}(A,C)\). Inserting \(1=e^{-N_\Gamma}e^{N_\Gamma}\) and anchoring at a boundary site gives \(|\ell|\le K_{\mathrm{MI}}|\partial A|e^{-\mathrm{dist}(A,C)/R}=\Delta\).

    The pointwise bound \(p_{AC}\le e^\Delta p_Ap_C\), averaged against the positive product projectors, becomes the operator inequality \(\rho_{AC}\le e^\Delta\rho_A\otimes\rho_C\). Operator monotonicity of \(\log\) gives \(\log\rho_{AC}-\log(\rho_A\otimes\rho_C)\le\Delta I\), and taking the expectation in \(\rho_{AC}\) gives \(I(A:C)\le\Delta\). For CMI, when the potentials satisfy \(\sup_x\sum_{U\ni x}|U|e^{\mu\,\mathrm{diam}(U)}\|\Psi_U\|\le\log(6/5)\), each conditional law decomposes as \[\pi(s_x|\mathbf s_{x^c})=\lambda_{x,0}\nu_x(s_x)+\sum_{U\ni x}\lambda_{x,U}\mathsf K_{x,U}(s_x|\mathbf s_{U\setminus x}),\qquad\lambda_{x,0}\ge\tfrac12.\] Data processing then gives \(I(A:C|B)\le2\log q\sum_{a\in A,c\in C}e^{-\mu\,\mathrm{dist}(a,c)}\), which the boundary argument converts into the stated bound.

    Editorial check: on an eight-site Heisenberg chain, site \(A=\{0\}\), site \(C=\{r\}\), and \(B\) the site just beyond \(C\), so \(B\) does not shield and the sites in between are traced out. The CMI is \(2.5\times10^{-3}\), \(1.0\times10^{-4}\), \(4.2\times10^{-6}\), \(1.7\times10^{-7}\) and \(7.2\times10^{-9}\) for \(r=2,\dots,6\) at \(\beta=0.2\), a factor of about 25 per site. At \(\beta=0.05\) it is \(9.9\times10^{-6}\) at \(r=2\) and falls by about 400 per site. These are illustrations of the predicted behaviour well outside the proven temperature range, not tests of the constants.

    How to reuse it

    The cancellation identity applies to any quantity built from logarithms of marginals of a convergent polymer model (mutual information, Rényi-type correlations, effective boundary Hamiltonians). The resampling decomposition, a Dobrushin-type condition with the conditioning kept quantum, is a template for proving Markov properties of other states with positive local representations, such as noisy stabilizer states or steady states of weakly interacting dissipative dynamics.

Source: Section 2 (Theorems 1–2, Eqs. (2.9)–(2.12)), Section 3 (2-design decomposition, Eqs. (3.1)–(3.16); cluster expansion, Lemma 5 and Corollary 6), Section 4 (proof of Theorem 1, Eqs. (4.2)–(4.10)), Section 5 (Lemmas 7–8), Section 6. 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. This is the first proof of exponentially decaying CMI for non-commuting Gibbs states at high temperature without restricting to shielding partitions, and with a prefactor set by the smaller boundary rather than the region sizes. Together with mutual-information decay it puts high-temperature Gibbs states in the trivial mixed-state phase under both standard criteria, local reversibility and the mixed-state entanglement bootstrap. It also shortens the locally reversible circuits that connect them to the maximally mixed state from polylogarithmic to constant depth. The device of trading an operator expansion for a positive classical one, by measuring in a 2-design, sidesteps the divergence that blocked earlier approaches.

Limitations. The temperature thresholds are far from sharp: they scale as \((2q)^{-k}\), and the 2-design decomposition costs a factor \((2q-1)^{-k}\) relative to the best known separability thresholds, which the authors say is inherent to the method. Nothing is proved near phase transitions or at low temperature, where CMI can carry topological contributions. Bounded local dimension and interaction strength are essential, since bosonic models with entropic order stay ordered at any temperature. Stability under weak local noise beyond on-site channels is stated as a conjecture with a proof sketch. No classical sampling algorithm is given, although the authors expect one to follow. The authors state that the proof was found using GPT-6 Astra under their guidance, that they verified and reorganised it, and that GPT-6 Astra and Claude Opus 5.5 assisted with literature search, technical refinements, writing and editing.

Description based on the full arXiv HTML of v1, read 2026-11-21: Sections 1–4 (setting, Theorems 1–2, the 2-design decomposition, the convergent cluster expansion of Lemma 5 and the mutual-information proof), the statements of Lemmas 7–8 in Section 5, the Section 6 discussion and the acknowledgement; the proofs of Lemmas 7–8 and Appendices A–B were not checked line by line. Editorial checks: the 2-design identities for the six Pauli eigenstates (exact reconstruction of a random operator, \(\|\tau_s\|=3\)). For the six-site Heisenberg chain, all \(6^6=46{,}656\) Gibbs weights are positive up to \(\beta J\approx0.21\), about 600 times the proven mutual-information threshold of \(3.5\times10^{-4}\). For an eight-site chain at \(\beta=0.2\), with the in-between sites traced out and a conditioning site placed outside the gap, the CMI between site 0 and site \(r\) falls from \(2.5\times10^{-3}\) at \(r=2\) to \(7\times10^{-9}\) at \(r=6\).