OctoInformBeta

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

Critical target setting: mixing at threshold 1/4: Worst-start total variation and the joint critical-temperature event at threshold 1/41/4

Mixing times for the two clocks.

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

Heat-bath dynamics has two natural clocks: elapsed time when each site updates at rate one, and the number of attempts when a single site is chosen at each step. The critical mixing event concerns both clocks for the same realization of the couplings.

For n∈ℕn\in\mathbb N, let In={0,…,n−1}I_n=\{0,\ldots,n-1\} be the site set, and let Ωn\Omega_n and EnE_n be the configuration space and the set of unordered pairs of distinct sites defined in (1). A spin configuration x∈Ωnx\in\Omega_n assigns a spin value xi∈{−1,+1}x_i\in\{-1,+1\} to each site.✦ A coupling array W∈ℝEnW\in\mathbb R^{E_n} assigns a real number WijW_{ij} to each unordered pair of distinct sites, indexed with i<ji<j in (1).✦ Ωn={−1,+1}In,En={(i,j)∈In×In:i<j}.\Omega_n=\{-1,+1\}^{I_n}, \qquad E_n=\{(i,j)\in I_n\times I_n:i<j\}.(1)

The Hamiltonian HW:Ωn→ℝH_W:\Omega_n\to\mathbb R is the pair interaction energy in (2). HW(x)=∑(i,j)∈EnWijxixj.H_W(x)=\sum_{(i,j)\in E_n}W_{ij}x_ix_j.(2)✦ The partition function ZWZ_W is the sum of the Boltzmann weights over Ωn\Omega_n.✦ The Gibbs distribution μW\mu_W is the probability mass function obtained by normalizing these weights, as in (3). ZW=∑z∈ΩneHW(z),μW(x)=eHW(x)ZW.Z_W=\sum_{z\in\Omega_n}e^{H_W(z)}, \qquad \mu_W(x)=\frac{e^{H_W(x)}}{Z_W}.(3)

For a site i∈Ini\in I_n, configurations x,y∈Ωnx,y\in\Omega_n agree outside ii when xj=yjx_j=y_j at every site j≠ij\ne i.✦ Write Ai(x)A_i(x) for the set of configurations agreeing with xx outside ii. The site update kernel PW,iP_{W,i} is the matrix on Ωn\Omega_n that resamples the spin at ii from the Gibbs distribution conditional on the other spins, with entries given by (4). PW,i(x,y)={μW(y)∑z∈Ai(x)μW(z),y∈Ai(x),0,y∉Ai(x).P_{W,i}(x,y)= \begin{cases} \displaystyle\frac{\mu_W(y)}{\sum_{z\in A_i(x)}\mu_W(z)}, & y\in A_i(x),\\[6pt] 0, & y\notin A_i(x). \end{cases}(4)✦

The heat-bath generator LWL_W is the matrix on Ωn\Omega_n obtained by summing PW,i−IP_{W,i}-I over sites i∈Ini\in I_n. For real tt, the continuous-time evolution PW,tP_{W,t} is the matrix exponential of tLWtL_W in (5).✦ Elapsed time corresponds to t≥0t\ge0. LW=∑i∈In(PW,i−I),PW,t=etLW.L_W=\sum_{i\in I_n}(P_{W,i}-I), \qquad P_{W,t}=e^{tL_W}.(5) The random-scan matrix PWP_W is the average of the site update kernels in (6), with the convention 0−1=00^{-1}=0.✦ For n>0n>0, the selected site is uniform; for k∈ℕk\in\mathbb N, PWk(x,⋅)P_W^k(x,\cdot) is the distribution after kk attempts from xx. PW=n−1∑i∈InPW,i.P_W=n^{-1}\sum_{i\in I_n}P_{W,i}.(6)

The Definition

Definition 1. For couplings W∈ℝEnW\in\mathbb R^{E_n} and a real function p:Ωn→ℝp:\Omega_n\to\mathbb R, the total variation distance ∥p−μW∥TV\|p-\mu_W\|_{\mathrm{TV}} from the Gibbs distribution is the functional in (7). ∥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)|.(7)✦ For t∈ℝt\in\mathbb R and x∈Ωnx\in\Omega_n, the continuous-time distance dWct(t,x)d_W^{\mathrm{ct}}(t,x) is the total variation functional of the corresponding row of PW,tP_{W,t} in (8). dWct(t,x)=∥PW,t(x,⋅)−μW∥TV.d_W^{\mathrm{ct}}(t,x) =\|P_{W,t}(x,\cdot)-\mu_W\|_{\mathrm{TV}}.(8)✦ For k∈ℕk\in\mathbb N and x∈Ωnx\in\Omega_n, the discrete-time distance dWdt(k,x)d_W^{\mathrm{dt}}(k,x) is the distance from equilibrium after kk steps, defined by (9). dWdt(k,x)=∥PWk(x,⋅)−μW∥TV.d_W^{\mathrm{dt}}(k,x) =\|P_W^k(x,\cdot)-\mu_W\|_{\mathrm{TV}}.(9)✦ The continuous-time worst-start mixing time tmix(W)∈ℝt_{\mathrm{mix}}(W)\in\mathbb R is the infimum of the nonnegative times satisfying the condition in (10). tmix(W)=inf⁡{t∈ℝ:t≥0,dWct(t,x)≤14 for every x∈Ωn}.t_{\mathrm{mix}}(W) =\inf\bigl\{t\in\mathbb R:t\ge0,\ d_W^{\mathrm{ct}}(t,x)\le\tfrac14 \text{ for every }x\in\Omega_n\bigr\}.(10)✦ The discrete-time worst-start mixing count kmix(W)∈ℕk_{\mathrm{mix}}(W)\in\mathbb N is the infimum of the update counts satisfying (11). kmix(W)=inf⁡{k∈ℕ:dWdt(k,x)≤14 for every x∈Ωn}.k_{\mathrm{mix}}(W) =\inf\bigl\{k\in\mathbb N: d_W^{\mathrm{dt}}(k,x)\le\tfrac14 \text{ for every }x\in\Omega_n\bigr\}.(11)✦ For ε∈ℝ\varepsilon\in\mathbb R, the joint critical mixing event ℳn,ε\mathcal M_{n,\varepsilon} is the subset of coupling arrays defined by the two-clock bounds in (12). ℳn,ε={W∈ℝEn:n2/3−ε≤tmix(W)≤eεn,n5/3−ε≤kmix(W)≤eεn}.\mathcal M_{n,\varepsilon} =\left\{W\in\mathbb R^{E_n}: \begin{aligned} n^{2/3-\varepsilon} &\le t_{\mathrm{mix}}(W)\le e^{\varepsilon n},\\ n^{5/3-\varepsilon} &\le k_{\mathrm{mix}}(W)\le e^{\varepsilon n} \end{aligned} \right\}.(12)✦

What Is Checked

The definitions are stated for every n∈ℕn\in\mathbb N and every real coupling array W∈ℝEnW\in\mathbb R^{E_n}, with ε∈ℝ\varepsilon\in\mathbb R for the joint event.