OctoInformBeta

OpenAI preprint, September 2026 · Section 1. The theorem and the structure of the proof

Theorem 1.1: Dimension-independent Poincare inequality: A constant independent of dimension for each fixed inverse temperature 0<β<10<\beta<1

A Poincaré bound at high temperature.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

For a spin system at high temperature, a Poincaré inequality compares the variance of an observable with its fluctuations under single-site heat-bath updates. The constant is required to work for every observable.

For n≥0n\geq 0, the spin space is Ωn={−1,+1}n\Omega_n=\{-1,+1\}^n✦, and the set of interacting pairs is Jn={(i,j):1≤i<j≤n}J_n=\{(i,j):1\leq i<j\leq n\}.✦ A coupling array is an element W=(Wij)(i,j)∈JnW=(W_{ij})_{(i,j)\in J_n} of 𝒲n=ℝJn\mathcal W_n=\mathbb R^{J_n}.✦ For n>0n>0, the Gaussian disorder law ℙn\mathbb P_n on 𝒲n\mathcal W_n makes these coordinates independent, with Wij∼N(0,1/n),i<j.W_{ij}\sim N(0,1/n),\qquad i<j.(1)✦ For real β\beta, let ℚn,β\mathbb Q_{n,\beta} be the law of βW\beta W when WW has law ℙn\mathbb P_n in (1); its coordinates have variance β2/n\beta^2/n. At n=0n=0, both laws put mass one on the empty coupling array.

The zero-field Gibbs distribution μW\mu_W on Ωn\Omega_n has Hamiltonian HWH_W and partition function ZWZ_W given by HW(x)=∑i<jWijxixj,ZW=∑x∈ΩneHW(x),μW(x)=eHW(x)ZW.H_W(x)=\sum_{i<j}W_{ij}x_ix_j,\qquad Z_W=\sum_{x\in\Omega_n}e^{H_W(x)},\qquad \mu_W(x)=\frac{e^{H_W(x)}}{Z_W}.(2) Inverse temperature β\beta means using μβW\mu_{\beta W} in (2). For a real observable f:Ωn→ℝf:\Omega_n\to\mathbb R, its Gibbs expectation and variance are ⟨f⟩W=∑x∈ΩnμW(x)f(x),Var⁡μW(f)=∑x∈ΩnμW(x)(f(x)−⟨f⟩W)2.\langle f\rangle_W=\sum_{x\in\Omega_n}\mu_W(x)f(x),\qquad \operatorname{Var}_{\mu_W}(f) =\sum_{x\in\Omega_n}\mu_W(x)\bigl(f(x)-\langle f\rangle_W\bigr)^2.(3)

Configurations x,y∈Ωnx,y\in\Omega_n agree outside site ii, denoted x∼iyx\sim_i y, when xj=yjx_j=y_j for every j≠ij\ne i.✦ The heat-bath kernel PiWP_i^W on Ωn\Omega_n is the transition kernel that resamples spin ii conditional on all other spins: PiW(x,y)={μW(y)∑z∼ixμW(z),y∼ix,0,y≁ix,PiWf(x)=∑y∈ΩnPiW(x,y)f(y).P_i^W(x,y)= \begin{cases} \displaystyle\frac{\mu_W(y)}{\sum_{z\sim_i x}\mu_W(z)},&y\sim_i x,\\[6pt] 0,&y\not\sim_i x, \end{cases} \qquad P_i^Wf(x)=\sum_{y\in\Omega_n}P_i^W(x,y)f(y).(4)✦ For n>0n>0, the uniform-site discrete heat-bath kernel is n−1∑iPiWn^{-1}\sum_iP_i^W, using the kernels in (4). The Dirichlet form DWD_W for μW\mu_W at rate one per site is the quadratic form on real observables f:Ωn→ℝf:\Omega_n\to\mathbb R given by DW(f)=∑i=1n⟨(f−PiWf)2⟩W.D_W(f)=\sum_{i=1}^n \left\langle\bigl(f-P_i^Wf\bigr)^2\right\rangle_W.(5)✦

For a real constant CC, the Poincaré event 𝒜n(β,C)⊆𝒲n\mathcal A_n(\beta,C)\subseteq\mathcal W_n is the set of coupling arrays for which the variance in (3) is bounded by CC times the Dirichlet form in (5) for every real observable: 𝒜n(β,C)={W:Var⁡μβW(f)≤CDβW(f) for every f:Ωn→ℝ}.\mathcal A_n(\beta,C)= \left\{W:\ \operatorname{Var}_{\mu_{\beta W}}(f) \leq C D_{\beta W}(f) \text{ for every }f:\Omega_n\to\mathbb R\right\}.(6)✦ Write ℬn(C)=𝒜n(1,C)\mathcal B_n(C)=\mathcal A_n(1,C) for the event in (6) with the inverse temperature absorbed into the coupling array.

The Claim

Theorem 1 (Theorem 1.1). Fix 0<β<10<\beta<1. There is a real constant C≥1C\geq1, depending only on β\beta, such that limn→∞ℙn(𝒜n(β,C))=1.\lim_{n\to\infty}\mathbb P_n\bigl(\mathcal A_n(\beta,C)\bigr)=1.(7)✦

In the paper, the result reads: for each fixed 0<β<10<\beta<1, there is a finite constant CβC_\beta, depending only on β\beta, such that the disorder probability of Var⁡μβW(f)≤CβDβW(f)\operatorname{Var}_{\mu_{\beta W}}(f)\leq C_\beta D_{\beta W}(f) simultaneously for every real function ff on Ωn\Omega_n tends to one as n→∞n\to\infty, and consequently the uniform-site discrete heat-bath chain has spectral gap at least 1/(Cβn)1/(C_\beta n) with probability tending to one.

What Is Checked

The statement (7) is over Gaussian coupling arrays and every real observable on their finite spin spaces, with the Dirichlet form (5).

The rescaling map Sβ:𝒲n→𝒲nS_\beta:\mathcal W_n\to\mathcal W_n is defined by Sβ(W)=βWS_\beta(W)=\beta W.

Theorem 2 (Rescaling of Poincaré events). For every n≥0n\geq0 and real β,C\beta,C, the Poincaré events satisfy Sβ−1(ℬn(C))=𝒜n(β,C),S_\beta^{-1}\bigl(\mathcal B_n(C)\bigr)=\mathcal A_n(\beta,C),(8)✦ and their disorder probabilities satisfy ℚn,β(ℬn(C))=ℙn(𝒜n(β,C)).\mathbb Q_{n,\beta}\bigl(\mathcal B_n(C)\bigr) =\mathbb P_n\bigl(\mathcal A_n(\beta,C)\bigr).(9)✦