OctoInformBeta

OpenAI preprint, September 2026 · Additional results of the selected critical-temperature target

Critical target: Natural covariance and linear-Rayleigh scale: For nn sites, both quantities lie within every diverging deterministic factor of n2/3n^{2/3}, with probability tending to one

Covariance and linear relaxation share a critical scale.

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

At inverse temperature one and zero external field, the spin covariance matrix records collective fluctuations, while linear spin observables test relaxation under heat-bath updates. The question concerns their size as the number of sites tends to infinity.

For n≥1n\geq 1 sites indexed by 0,…,n−10,\ldots,n-1, a spin configuration x∈Ωn={−1,+1}nx\in\Omega_n=\{-1,+1\}^n assigns a spin value to each site.✦ A coupling array W=(Wij)0≤i<j<nW=(W_{ij})_{0\leq i<j<n} assigns a real number to each unordered spin pair.✦ The Gaussian disorder law ℙn\mathbb P_n is the product probability measure in (1), with independent coordinates: ℙn=⨂0≤i<j<n𝒩(0,1/n).\mathbb P_n=\bigotimes_{0\leq i<j<n}\mathcal N(0,1/n).(1)✦

For a coupling array WW, the Hamiltonian HWH_W is the interaction energy in (2); the partition function ZWZ_W is its exponential sum over configurations, and the Gibbs distribution μW\mu_W is the normalized exponential weight: HW(x)=∑0≤i<j<nWijxixj,ZW=∑x∈ΩneHW(x),μW(x)=eHW(x)ZW.\begin{aligned} H_W(x)&=\sum_{0\leq i<j<n}W_{ij}x_ix_j,\\ Z_W&=\sum_{x\in\Omega_n}e^{H_W(x)}, &\mu_W(x)&=\frac{e^{H_W(x)}}{Z_W}. \end{aligned}(2)

For a real observable ff on Ωn\Omega_n, the Gibbs mean ⟨f⟩W\langle f\rangle_W is its expectation under μW\mu_W, and the Gibbs variance Var⁡W(f)\operatorname{Var}_W(f) is its mean squared deviation from that expectation, as in (3): ⟨f⟩W=∑x∈ΩnμW(x)f(x),Var⁡W(f)=⟨(f−⟨f⟩W)2⟩W.\langle f\rangle_W=\sum_{x\in\Omega_n}\mu_W(x)f(x), \qquad \operatorname{Var}_W(f)=\big\langle(f-\langle f\rangle_W)^2\big\rangle_W.(3)

The spin covariance matrix ΣW\Sigma_W is the real matrix whose entries are the Gibbs covariances of the site spins in (4).✦ The covariance norm ∥ΣW∥op\|\Sigma_W\|_{\mathrm{op}} of the spin covariance matrix ΣW\Sigma_W is its operator norm for the Euclidean norm on ℝn\mathbb R^n: (ΣW)ij=⟨(xi−⟨xi⟩W)(xj−⟨xj⟩W)⟩W,∥ΣW∥op=sup∥a∥2=1∥ΣWa∥2.(\Sigma_W)_{ij} =\big\langle(x_i-\langle x_i\rangle_W) (x_j-\langle x_j\rangle_W)\big\rangle_W, \qquad \|\Sigma_W\|_{\mathrm{op}} =\sup_{\|a\|_2=1}\|\Sigma_Wa\|_2.(4)

Heat Bath Relaxation

Two configurations agree outside site ii if their spins coincide at every site j≠ij\ne i.✦ The set Ai(x)A_i(x) in (5) consists of the configurations agreeing with xx outside ii. The heat-bath kernel PiWP_i^W resamples site ii from its conditional Gibbs distribution: Ai(x)={y∈Ωn:yj=xj for all j≠i},PiW(x,y)=𝟏Ai(x)(y)μW(y)∑z∈Ai(x)μW(z).A_i(x)=\{y\in\Omega_n:y_j=x_j\text{ for all }j\ne i\}, \qquad P_i^W(x,y) =\frac{\mathbf 1_{A_i(x)}(y)\mu_W(y)} {\sum_{z\in A_i(x)}\mu_W(z)}.(5)

For a real observable ff on Ωn\Omega_n, the site update average PiWfP_i^Wf is the function giving the expectation of ff after the heat-bath kernel PiWP_i^W updates site ii. For a real observable ff on Ωn\Omega_n, the Dirichlet form DW(f)D_W(f) is the sum of the Gibbs mean squared deviations of ff from its site update averages PiWfP_i^Wf, with rate one per site, as in (6): (PiWf)(x)=∑y∈ΩnPiW(x,y)f(y),DW(f)=∑i=0n−1⟨(f−PiWf)2⟩W.(P_i^Wf)(x)=\sum_{y\in\Omega_n}P_i^W(x,y)f(y), \qquad D_W(f)=\sum_{i=0}^{n-1}\big\langle(f-P_i^Wf)^2\big\rangle_W.(6)✦

For a real coefficient vector a∈ℝna\in\mathbb R^n, the linear spin observable faf_a is the weighted spin sum in (7): fa(x)=∑i=0n−1aixi.f_a(x)=\sum_{i=0}^{n-1}a_ix_i.(7)✦ The linear Rayleigh supremum Rlin(W)R_{\mathrm{lin}}(W) for a coupling array WW is the supremum of the ratio of Gibbs variance to Dirichlet form over linear spin observables faf_a with nonzero coefficient vectors aa in ℝn\mathbb R^n. The full relaxation time trel(W)t_{\mathrm{rel}}(W) for a coupling array WW is the supremum of the ratio of Gibbs variance to Dirichlet form over all real observables on Ωn\Omega_n with positive Gibbs variance, using the Dirichlet form of (6), as specified in (8): Rlin(W)=supa∈ℝn\{0}Var⁡W(fa)DW(fa),trel(W)=supf:Ωn→ℝVar⁡W(f)>0Var⁡W(f)DW(f).R_{\mathrm{lin}}(W) =\sup_{a\in\mathbb R^n\setminus\{0\}} \frac{\operatorname{Var}_W(f_a)}{D_W(f_a)}, \qquad t_{\mathrm{rel}}(W) =\sup_{\substack{f:\Omega_n\to\mathbb R\\ \operatorname{Var}_W(f)>0}} \frac{\operatorname{Var}_W(f)}{D_W(f)}.(8)

The Claim

Theorem 1. Let (Mn)n≥1(M_n)_{n\geq1} be a deterministic sequence of positive real numbers tending to infinity. For Gaussian couplings with law ℙn\mathbb P_n, the covariance bound is (9): ℙn{n2/3Mn≤∥ΣW∥op≤Mnn2/3}→1.\mathbb P_n\!\left\{ \frac{n^{2/3}}{M_n}\leq\|\Sigma_W\|_{\mathrm{op}} \leq M_nn^{2/3} \right\}\longrightarrow1.(9)✦ The linear Rayleigh bound is (10): ℙn{n2/3Mn≤Rlin(W)≤Mnn2/3}→1.\mathbb P_n\!\left\{ \frac{n^{2/3}}{M_n}\leq R_{\mathrm{lin}}(W) \leq M_nn^{2/3} \right\}\longrightarrow1.(10)✦ Their joint event satisfies (11), with all limits taken as n→∞n\to\infty: ℙn{n2/3Mn≤∥ΣW∥op≤Mnn2/3,n2/3Mn≤Rlin(W)≤Mnn2/3}→1.\mathbb P_n\!\left\{ \begin{aligned} \frac{n^{2/3}}{M_n}&\leq\|\Sigma_W\|_{\mathrm{op}}\leq M_nn^{2/3},\\ \frac{n^{2/3}}{M_n}&\leq R_{\mathrm{lin}}(W)\leq M_nn^{2/3} \end{aligned} \right\}\longrightarrow1.(11)✦ For every n>0n>0 and every coupling array WW, the comparison (12) holds: Rlin(W)≤trel(W).R_{\mathrm{lin}}(W)\leq t_{\mathrm{rel}}(W).(12)✦

What Is Checked

The scale statements concern the Gibbs observables ΣW\Sigma_W and Rlin(W)R_{\mathrm{lin}}(W) under the product Gaussian law ℙn\mathbb P_n, for every deterministic positive sequence Mn→∞M_n\to\infty; the relaxation comparison concerns arbitrary couplings on a positive number of sites.