OctoInformBeta

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

Critical target: Typical equilibrium starts remain unmixed: For nn sites, unmixed Gibbs mass tends to one at times o(n2/3)o(n^{2/3}) or after o(n5/3)o(n^{5/3}) update attempts

Typical equilibrium starts mix slowly at criticality.

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

Consider Glauber dynamics for the critical Sherrington-Kirkpatrick model with zero external field. The starting configuration is weighted by its Gibbs probability, while its distance to equilibrium is measured from that fixed starting state.

Let n≥1n\geq1 be the number of sites. A spin configuration x∈Ωnx\in\Omega_n is a function on the site set {1,…,n}\{1,\ldots,n\} taking values in {−1,+1}\{-1,+1\}; xix_i denotes its value at site ii.✦ An unordered spin pair is an element of EnE_n, the set of pairs (i,j)(i,j) of sites with i<ji<j.✦ A coupling array W∈ℝEnW\in\mathbb R^{E_n} is a real-valued function on the set EnE_n of spin pairs.✦ The Gaussian disorder law ℙn\mathbb P_n in (1) has independent coordinates with mean zero and variance 1/n1/n: Ωn={−1,+1}{1,…,n},En={(i,j):1≤i<j≤n},ℙn=⨂(i,j)∈En𝒩(0,1/n).\Omega_n=\{-1,+1\}^{\{1,\ldots,n\}},\qquad E_n=\{(i,j):1\leq i<j\leq n\},\qquad \mathbb P_n=\bigotimes_{(i,j)\in E_n}\mathcal N(0,1/n).(1)

The Hamiltonian HW:Ωn→ℝH_W:\Omega_n\to\mathbb R is the function that assigns each configuration its pair interaction energy, as given in (2). The Gibbs distribution μW\mu_W is the probability distribution on Ωn\Omega_n proportional to eHW(x)e^{H_W(x)}✦; the partition function ZWZ_W is the sum of these weights: 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)

Write x\ix_{\setminus i} for the restriction of xx to all sites other than ii. Two configurations x,y∈Ωnx,y\in\Omega_n agree outside ii when these restrictions coincide.✦ The heat-bath kernel PiWP_i^W is the transition matrix on Ωn\Omega_n that resamples the spin at ii from its conditional Gibbs distribution, with entries given by (3): PiW(x,y)=𝟏{y\i=x\i}μW(y)∑z:z\i=x\iμW(z).P_i^W(x,y)= \mathbf 1_{\{y_{\setminus i}=x_{\setminus i}\}} \frac{\mu_W(y)} {\displaystyle\sum_{z:z_{\setminus i}=x_{\setminus i}}\mu_W(z)}.(3)✦

The heat-bath generator LWL_W updates every site at rate one.✦ The continuous-time matrix PtWP_t^W associated with the heat-bath generator LWL_W is the matrix exponential of that generator at a real time parameter tt, and is the transition semigroup for t≥0t\geq0.✦ The random-scan kernel QWQ_W makes one attempt at a uniformly chosen site.✦ With II the identity matrix, their formulas are (4): LW=∑i=1n(PiW−I),PtW=etLW,QW=1n∑i=1nPiW.L_W=\sum_{i=1}^n(P_i^W-I),\qquad P_t^W=e^{tL_W},\qquad Q_W=\frac1n\sum_{i=1}^nP_i^W.(4)

Unmixed Gibbs Mass

For a real function pp on Ωn\Omega_n, the total variation distance ∥p−μW∥TV\|p-\mu_W\|_{\mathrm{TV}} to Gibbs equilibrium is the nonnegative real number in (5): ∥p−μW∥TV=12∑y∈Ωn|p(y)−μW(y)|.\|p-\mu_W\|_{\mathrm{TV}} =\frac12\sum_{y\in\Omega_n}|p(y)-\mu_W(y)|.(5)✦

The continuous-time distance dWc(t,x)d_W^c(t,x), for real tt, is the total variation distance to equilibrium of the row PtW(x,⋅)P_t^W(x,\cdot).✦ The discrete-time distance dWd(k,x)d_W^d(k,x), for a nonnegative integer k∈ℕ0k\in\mathbb N_0, is the distance to equilibrium after kk attempts from xx.✦ dWc(t,x)=∥PtW(x,⋅)−μW∥TV,dWd(k,x)=∥QWk(x,⋅)−μW∥TV.d_W^c(t,x)=\|P_t^W(x,\cdot)-\mu_W\|_{\mathrm{TV}},\qquad d_W^d(k,x)=\|Q_W^k(x,\cdot)-\mu_W\|_{\mathrm{TV}}.(6)

The unmixed Gibbs masses MWc(t)M_W^c(t) and MWd(k)M_W^d(k) are the Gibbs probabilities of starting configurations whose distances at the indicated time or update count are strictly greater than 1/41/4. MWc(t)=∑x:dWc(t,x)>1/4μW(x),MWd(k)=∑x:dWd(k,x)>1/4μW(x).M_W^c(t)=\sum_{x:d_W^c(t,x)>1/4}\mu_W(x),\qquad M_W^d(k)=\sum_{x:d_W^d(k,x)>1/4}\mu_W(x).(7)

The worst-start mixing times τc(W)\tau_c(W) and τd(W)\tau_d(W) are the infima of the nonnegative times and update counts at which every starting state is within 1/41/4 of equilibrium. τc(W)=inf⁡{t≥0:dWc(t,x)≤1/4 for every x∈Ωn},τd(W)=inf⁡{k∈ℕ0:dWd(k,x)≤1/4 for every x∈Ωn}.\begin{aligned} \tau_c(W)&=\inf\{t\geq0:d_W^c(t,x)\leq1/4\text{ for every }x\in\Omega_n\},\\ \tau_d(W)&=\inf\{k\in\mathbb N_0:d_W^d(k,x)\leq1/4 \text{ for every }x\in\Omega_n\}. \end{aligned}(8)

The Claim

Theorem 1. Let tn≥0t_n\geq0 be deterministic times with tn=o(n2/3)t_n=o(n^{2/3}), and let kn∈ℕ0k_n\in\mathbb N_0 be deterministic update counts with kn=o(n5/3)k_n=o(n^{5/3}). The equilibrium-start assertion is (9), where convergence is in probability under the disorder laws ℙn\mathbb P_n: (MWc(tn),MWd(kn))→ℙn(1,1).\bigl(M_W^c(t_n),M_W^d(k_n)\bigr) \xrightarrow{\mathbb P_n}(1,1).(9)✦ Let δ>0\delta>0. ℙn({W:|MWc(tn)−1|≥δ})→0.\mathbb P_n\bigl(\{W:|M_W^c(t_n)-1|\geq\delta\}\bigr)\longrightarrow0.(10)✦ ℙn({W:|MWd(kn)−1|≥δ})→0.\mathbb P_n\bigl(\{W:|M_W^d(k_n)-1|\geq\delta\}\bigr)\longrightarrow0.(11)✦ The worst-start mixing bounds are (12) and (13): ℙn({W:tn≤τc(W)})→1.\mathbb P_n\bigl(\{W:t_n\leq\tau_c(W)\}\bigr)\longrightarrow1.(12)✦ ℙn({W:kn≤τd(W)})→1.\mathbb P_n\bigl(\{W:k_n\leq\tau_d(W)\}\bigr)\longrightarrow1.(13)✦ For every fixed ε>0\varepsilon>0, the power bounds (14) hold: ℙn({W:n2/3−ε≤τc(W)})→1,ℙn({W:n5/3−ε≤τd(W)})→1.\begin{aligned} \mathbb P_n\bigl(\{W:n^{2/3-\varepsilon}\leq\tau_c(W)\}\bigr) &\longrightarrow1,\\ \mathbb P_n\bigl(\{W:n^{5/3-\varepsilon}\leq\tau_d(W)\}\bigr) &\longrightarrow1. \end{aligned}(14)✦ All limits are taken as n→∞n\to\infty.

What Is Checked

The statements concern the Gibbs model (2) under the Gaussian disorder law (1), with the heat-bath dynamics (4) and deterministic nonnegative time or update sequences on the stated scales.