OctoInformBeta

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

Critical target: Two-clock mixing bounds: Simultaneous polynomial lower bounds and subexponential upper bounds for continuous site clocks and discrete update attempts under Gaussian disorder

Critical mixing bounds hold with high probability.

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

Critical heat-bath dynamics can be timed in physical time or by counting individual site update attempts. The question is how its worst-start mixing time grows with the number of spins under Gaussian disorder.

Let n≥1n\geq1. A spin configuration x∈Ωn={−1,+1}nx\in\Omega_n=\{-1,+1\}^n assigns a spin xix_i to each site i∈{1,…,n}i\in\{1,\ldots,n\}.✦ A coupling array W=(Wij)i<jW=(W_{ij})_{i<j} assigns a real number to each unordered pair of sites.✦ Under the Gaussian disorder law ℙn\mathbb P_n, these coordinates are independent with Wij∼𝒩(0,1/n)W_{ij}\sim\mathcal N(0,1/n).✦ The Hamiltonian HWH_W is the pair interaction sum in (1); its partition function ZWZ_W normalizes the Gibbs distribution μW\mu_W: HW(x)=∑1≤i<j≤nWijxixj,ZW=∑x∈ΩneHW(x),μW(x)=eHW(x)ZW.\begin{aligned} H_W(x)&=\sum_{1\leq i<j\leq 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}(1) Inverse temperature β∈ℝ\beta\in\mathbb R is represented by βW\beta W, whose couplings have variance β2/n\beta^2/n, with zero external field.

Write y−i=x−iy_{-i}=x_{-i} when configurations x,y∈Ωnx,y\in\Omega_n agree at every site other than ii.✦ The heat-bath kernel PiP_i resamples site ii from its conditional Gibbs distribution, with matrix entries and action on an observable ff given by (2): Pi(x,y)=𝟏{y−i=x−i}μW(y)∑z:z−i=x−iμW(z),(Pif)(x)=∑y∈ΩnPi(x,y)f(y).\begin{aligned} P_i(x,y)&=\mathbf1_{\{y_{-i}=x_{-i}\}} \frac{\mu_W(y)}{\sum_{z:z_{-i}=x_{-i}}\mu_W(z)},\\ (P_if)(x)&=\sum_{y\in\Omega_n}P_i(x,y)f(y). \end{aligned}(2)✦ For nonnegative physical time tt, the continuous heat-bath semigroup PtWP_t^W has generator LWL_W, with rate one per site. The random-scan kernel KWK_W makes one update attempt at a uniformly chosen site.✦ With II the identity kernel on Ωn\Omega_n, the heat-bath generator LWL_W, its continuous semigroup PtWP_t^W and the random-scan kernel KWK_W are the operators in (3): LW=∑i=1n(Pi−I),PtW=etLW,KW=1n∑i=1nPi.L_W=\sum_{i=1}^n(P_i-I), \qquad P_t^W=e^{tL_W}, \qquad K_W=\frac1n\sum_{i=1}^nP_i.(3)

For p:Ωn→ℝp:\Omega_n\to\mathbb R, the total variation functional dW(p)d_W(p) and its continuous and discrete transition distances dcW(t,x)d_c^W(t,x) and ddW(k,x)d_d^W(k,x) are specified in (4), for t≥0t\geq0 and k∈ℕk\in\mathbb N: dW(p)=12∑y∈Ωn|p(y)−μW(y)|,dcW(t,x)=dW(PtW(x,⋅)),ddW(k,x)=dW(KWk(x,⋅)).\begin{aligned} d_W(p)&=\frac12\sum_{y\in\Omega_n}|p(y)-\mu_W(y)|,\\ d_c^W(t,x)&=d_W\bigl(P_t^W(x,\cdot)\bigr), &d_d^W(k,x)&=d_W\bigl(K_W^k(x,\cdot)\bigr). \end{aligned}(4) The worst-start mixing times tc(W)∈ℝt_c(W)\in\mathbb R and td(W)∈ℕt_d(W)\in\mathbb N are the threshold times in (5): tc(W)=inf⁡{t≥0:dcW(t,x)≤1/4 for every x∈Ωn},td(W)=inf⁡{k∈ℕ:ddW(k,x)≤1/4 for every x∈Ωn}.\begin{aligned} t_c(W)&=\inf\{t\geq0:\ d_c^W(t,x)\leq1/4\text{ for every }x\in\Omega_n\},\\ t_d(W)&=\inf\{k\in\mathbb N:\ d_d^W(k,x)\leq1/4\text{ for every }x\in\Omega_n\}. \end{aligned}(5) The unmixed Gibbs masses mcW(t)m_c^W(t) and mdW(k)m_d^W(k) are the Gibbs masses of starting states still farther than 1/41/4 from equilibrium, as defined in (6): mcW(t)=∑x∈ΩnμW(x)𝟏{dcW(t,x)>1/4},mdW(k)=∑x∈ΩnμW(x)𝟏{ddW(k,x)>1/4}.\begin{aligned} m_c^W(t)&=\sum_{x\in\Omega_n}\mu_W(x)\mathbf1_{\{d_c^W(t,x)>1/4\}},\\ m_d^W(k)&=\sum_{x\in\Omega_n}\mu_W(x)\mathbf1_{\{d_d^W(k,x)>1/4\}}. \end{aligned}(6)

The Gibbs mean ⟨f⟩W\langle f\rangle_W, variance Var⁡W(f)\operatorname{Var}_W(f) and Dirichlet form DW(f)D_W(f) of a real observable f:Ωn→ℝf:\Omega_n\to\mathbb R are defined in (7): ⟨f⟩W=∑x∈ΩnμW(x)f(x),Var⁡W(f)=⟨(f−⟨f⟩W)2⟩W,DW(f)=∑i=1n⟨(f−Pif)2⟩W.\begin{aligned} \langle f\rangle_W&=\sum_{x\in\Omega_n}\mu_W(x)f(x),\\ \operatorname{Var}_W(f)&=\bigl\langle(f-\langle f\rangle_W)^2\bigr\rangle_W,\\ D_W(f)&=\sum_{i=1}^n\bigl\langle(f-P_if)^2\bigr\rangle_W. \end{aligned}(7) For real β,C\beta,C, the tempered Poincaré event An(β,C)A_n(\beta,C) is the set of couplings satisfying the variance bound in (8) for every real observable: An(β,C)={W:Var⁡βW(f)≤CDβW(f) for every f:Ωn→ℝ}.A_n(\beta,C)=\{W:\ \operatorname{Var}_{\beta W}(f)\leq C D_{\beta W}(f) \text{ for every }f:\Omega_n\to\mathbb R\}.(8)✦

For ε>0\varepsilon>0, the joint mixing event Bn(ε)B_n(\varepsilon) is the set of couplings satisfying the four bounds in (9): Bn(ε)={W:n2/3−ε≤tc(W)≤eεn,n5/3−ε≤td(W)≤eεn}.B_n(\varepsilon)=\left\{W: \begin{aligned} n^{2/3-\varepsilon}&\leq t_c(W)\leq e^{\varepsilon n},\\ n^{5/3-\varepsilon}&\leq t_d(W)\leq e^{\varepsilon n} \end{aligned}\right\}.(9)✦ The reduction error rn(ε)r_n(\varepsilon) is the quantity in (10), where ⌈a⌉+\lceil a\rceil_+ denotes the least nonnegative integer at least a∈ℝa\in\mathbb R: rn(ε)=2e−3n/2+ℙn{W:|mcW(n2/3−ε)−1|≥1}+ℙn{W:|mdW(⌈n5/3−ε⌉+)−1|≥1}.\begin{aligned} r_n(\varepsilon)={}&2e^{-3n/2} +\mathbb P_n\!\left\{W:\left|m_c^W(n^{2/3-\varepsilon})-1\right|\geq1\right\}\\ &+\mathbb P_n\!\left\{W: \left|m_d^W(\lceil n^{5/3-\varepsilon}\rceil_+)-1\right|\geq1\right\}. \end{aligned}(10)✦

The Claim

Theorem 1. For every ε>0\varepsilon>0, the joint and marginal mixing probabilities have the limits (11) and (12): limn→∞ℙn(Bn(ε))=1.\lim_{n\to\infty}\mathbb P_n\bigl(B_n(\varepsilon)\bigr)=1.(11)✦ limn→∞ℙn{W:n2/3−ε≤tc(W)≤eεn}=1,limn→∞ℙn{W:n5/3−ε≤td(W)≤eεn}=1.\begin{aligned} \lim_{n\to\infty}\mathbb P_n\{W:\ n^{2/3-\varepsilon}\leq t_c(W) \leq e^{\varepsilon n}\}&=1,\\ \lim_{n\to\infty}\mathbb P_n\{W:\ n^{5/3-\varepsilon}\leq t_d(W) \leq e^{\varepsilon n}\}&=1. \end{aligned}(12)✦

Let n>0n>0, β≤1\beta\leq1 and C≥1C\geq1. For every W∈An(β,C)W\in A_n(\beta,C) with |HW(x)|≤3n|H_W(x)|\leq3n for every x∈Ωnx\in\Omega_n, the finite comparison bound is (13): max⁡{tc(W),td(W)}≤44Cn2e6(1−β)n.\max\{t_c(W),t_d(W)\}\leq44Cn^2e^{6(1-\beta)n}.(13)✦

Fix ε>0\varepsilon>0, β≤1\beta\leq1 and C≥1C\geq1 with 6(1−β)≤ε/26(1-\beta)\leq\varepsilon/2. For all sufficiently large nn, every W∈An(β,C)W\in A_n(\beta,C) with |HW(x)|≤3n|H_W(x)|\leq3n for every x∈Ωnx\in\Omega_n satisfies (14): max⁡{tc(W),td(W)}≤eεn.\max\{t_c(W),t_d(W)\}\leq e^{\varepsilon n}.(14)✦

For every ε>0\varepsilon>0, the reduction error has the limit (15): limn→∞rn(ε)=0.\lim_{n\to\infty}r_n(\varepsilon)=0.(15)✦ For fixed ε>0\varepsilon>0, β≤1\beta\leq1 and C≥1C\geq1 with 6(1−β)≤ε/26(1-\beta)\leq\varepsilon/2, the probability comparison (16) holds for all sufficiently large nn: ℙn(An(β,C))−rn(ε)≤ℙn(Bn(ε)).\mathbb P_n\bigl(A_n(\beta,C)\bigr)-r_n(\varepsilon) \leq\mathbb P_n\bigl(B_n(\varepsilon)\bigr).(16)✦

What Is Checked

The claims are stated over the Gaussian coupling law and its Gibbs heat-bath dynamics, with the tempered Poincaré event defined by (8).