arXiv:2610.03683
Modified logarithmic Sobolev inequality for 1D non-commuting Hamiltonians
Ángela Capel, David Pérez-García, Matteo Scandi
Abstract
Given a finite-range, non-commuting Hamiltonian on a 1D chain of finite spins, we prove a modified logarithmic Sobolev inequality with constant $Ω(1/\log n)$ for the heat-bath dynamics, a quasi-local Gibbs sampler, and for the regularised heat-bath dynamics. This implies rapid mixing of both samplers towards the Gibbs state in relative entropy and trace norm, with mixing times of order $\mathcal{O}(\log n\cdot \log(n/ε))$. The interactions may be spatially inhomogeneous. The proof combines weak quasi-factorisation of relative entropy, decay-of-correlation estimates, a uniform conditional local gap, and a buffered comparison between conditional entropy and quadratic energy. The resulting bounded entropy defect is removed using the global gap. To the best of our knowledge, this is the first proof of MLSI in the non-commuting regime.
AI Summary
The paper asks whether Gibbs samplers for noncommuting one-dimensional spin Hamiltonians satisfy a modified logarithmic Sobolev inequality (MLSI) strong enough to imply polylogarithmic mixing. It considers open chains with bounded finite-range, possibly spatially inhomogeneous interactions, fixed local dimension $d=2^q$, and any fixed inverse temperature $\beta>0$. For the heat-bath sampler, which discards a site and reconstructs it using Petz recovery, and its regularised version obtained by iterating each update to a conditional expectation, it proves $\alpha_{\mathrm{MLSI}}\ge c_\beta/\log n$. With each site updated at rate one, relative entropy decays as $e^{-c_\beta t/\log n}$, and both relative-entropy and trace-norm mixing times are $O_\beta(\log n[\log n+\log(1/\epsilon)])$. Heat-bath updates also admit local channel approximations with diamond-norm error exponentially small in buffer width. The proof combines weak entropy quasi-factorisation over overlapping blocks of size $O(\log n)$, one-dimensional correlation and marginal-logarithm locality estimates, and a uniform conditional gap derived from a canonical purified Hamiltonian. Buffered recovery estimates compare block entropy with quadratic energy; modular averaging connects the relevant weighted forms. Summation produces a bounded additive entropy defect, which a global gap and norm-to-entropy argument remove. The logarithmic loss is not shown to be intrinsic, and temperature dependence is absorbed into unspecified constants. The result concerns these particular samplers, not general thermal dynamics. Regularised updates need not be quasi-local. The $1/\log n$ bound is ordinary, not complete, MLSI; a matching memory-uniform bound and extensions beyond one dimension remain open.
We acknowledge the use of LLMs for a proof of MLSI, instead of weak MLSI, and for proofreading, refining proof ideas and creating figures.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 56 pages, 10 figures