OctoInformBeta

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

Critical target setting: equilibrium starts and linear observables: Unmixed Gibbs mass and variance-to-Dirichlet ratios at fixed couplings

Equilibrium starts and linear spin fluctuations.

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

Equilibrium initialization asks about typical starting configurations under the Gibbs law. Linear combinations of the spins give a family of observables for measuring equilibrium fluctuations and their Dirichlet cost.

Let nn be a nonnegative integer, with 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 xix_i to each site ii.✦ A coupling array W=(Wij)i<jW=(W_{ij})_{i<j} assigns a real number to each unordered pair of distinct sites, represented by i<ji<j.✦ The Hamiltonian HWH_W of a coupling array WW on nn sites is the real function on Ωn\Omega_n giving the pair interaction energy in (1). HW(x)=∑0≤i<j<nWijxixj.H_W(x)=\sum_{0\leq i<j<n}W_{ij}x_ix_j.(1)✦

For a coupling array WW on nn sites, the partition function ZWZ_W is the sum of the weights eHW(x)e^{H_W(x)} over x∈Ωnx\in\Omega_n, and the Gibbs distribution μW\mu_W is the probability distribution on Ωn\Omega_n obtained by dividing these weights by ZWZ_W, as in (2). ZW=∑x∈ΩneHW(x),μW(x)=eHW(x)ZW.Z_W=\sum_{x\in\Omega_n}e^{H_W(x)}, \qquad \mu_W(x)=\frac{e^{H_W(x)}}{Z_W}.(2)

The Gibbs mean ⟨f⟩W\langle f\rangle_W of a real observable ff on Ωn\Omega_n is its average under μW\mu_W; its Gibbs variance Var⁡W(f)\operatorname{Var}_W(f) is its mean squared deviation from that average, 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)=\left\langle(f-\langle f\rangle_W)^2\right\rangle_W.(3)

For a site ii and a configuration xx, let 𝒞i(x)={y∈Ωn:yj=xj for every j≠i}\mathcal C_i(x)=\{y\in\Omega_n:y_j=x_j\text{ for every }j\ne i\} be the set of configurations agreeing with xx outside ii.✦ The heat-bath kernel PiWP_i^W replaces the spin at ii according to the Gibbs conditional distribution; its entries and its action on observables are given in (4). PiW(x,y)={μW(y)∑z∈𝒞i(x)μW(z),y∈𝒞i(x),0,y∉𝒞i(x),(PiWf)(x)=∑y∈ΩnPiW(x,y)f(y).P_i^W(x,y)= \begin{cases} \displaystyle\frac{\mu_W(y)}{\sum_{z\in\mathcal C_i(x)}\mu_W(z)}, &y\in\mathcal C_i(x),\\ 0,&y\notin\mathcal C_i(x), \end{cases} \qquad (P_i^Wf)(x)=\sum_{y\in\Omega_n}P_i^W(x,y)f(y).(4)✦

Each site is updated at rate one in the heat-bath generator LWL_W, whose exponential kernel PtWP_t^W, for a real parameter tt, is defined in (5). Here II is the identity matrix on Ωn\Omega_n. LW=∑i=0n−1(PiW−I),PtW=etLW.L_W=\sum_{i=0}^{n-1}(P_i^W-I), \qquad P_t^W=e^{tL_W}.(5)

For n>0n>0, the random-scan kernel KWK_W makes one heat-bath attempt at a uniformly chosen site.✦ Its definition is (6), with n−1=0n^{-1}=0 when n=0n=0. KW=n−1∑i=0n−1PiW.K_W=n^{-1}\sum_{i=0}^{n-1}P_i^W.(6)✦

The total variation distance ∥p−μW∥TV\|p-\mu_W\|_{\mathrm{TV}} of a real mass function pp on Ωn\Omega_n from μW\mu_W is one half the sum of the absolute mass differences over Ωn\Omega_n, as 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 a starting configuration x∈Ωnx\in\Omega_n, a real parameter tt, and a nonnegative integer kk, the continuous distance dWct(t,x)d_W^{\mathrm{ct}}(t,x) and discrete distance dWdisc(k,x)d_W^{\mathrm{disc}}(k,x) are the total variation distances of the rows PtW(x,⋅)P_t^W(x,\cdot) and KWk(x,⋅)K_W^k(x,\cdot) from the Gibbs distribution μW\mu_W, respectively, as in (8). dWct(t,x)=∥PtW(x,⋅)−μW∥TV,dWdisc(k,x)=∥KWk(x,⋅)−μW∥TV,k∈ℕ.\begin{aligned} d_W^{\mathrm{ct}}(t,x)&=\|P_t^W(x,\cdot)-\mu_W\|_{\mathrm{TV}},\\ d_W^{\mathrm{disc}}(k,x)&=\|K_W^k(x,\cdot)-\mu_W\|_{\mathrm{TV}}, \qquad k\in\mathbb N. \end{aligned}(8)

The Dirichlet form DW(f)D_W(f) of a real observable ff on Ωn\Omega_n is the sum, over sites ii, of the mean squared deviations of ff from PiWfP_i^Wf under the Gibbs distribution μW\mu_W, as in (9). DW(f)=∑i=0n−1⟨(f−PiWf)2⟩W.D_W(f)=\sum_{i=0}^{n-1}\left\langle(f-P_i^Wf)^2\right\rangle_W.(9)✦

The Definition

Definition 1. Fix a real coupling array WW on nn sites. The continuous unmixed mass GWct(t)G_W^{\mathrm{ct}}(t), for t∈ℝt\in\mathbb R, is the Gibbs mass of starting configurations whose distance at tt remains strictly greater than 1/41/4, as in (10). GWct(t)=∑x∈ΩndWct(t,x)>1/4μW(x).G_W^{\mathrm{ct}}(t) =\sum_{\substack{x\in\Omega_n\\d_W^{\mathrm{ct}}(t,x)>1/4}}\mu_W(x).(10)✦

The discrete unmixed mass GWdisc(k)G_W^{\mathrm{disc}}(k), for k∈ℕk\in\mathbb N, is the Gibbs mass of starting configurations whose distance after kk attempts remains strictly greater than 1/41/4, as in (11). GWdisc(k)=∑x∈ΩndWdisc(k,x)>1/4μW(x).G_W^{\mathrm{disc}}(k) =\sum_{\substack{x\in\Omega_n\\d_W^{\mathrm{disc}}(k,x)>1/4}}\mu_W(x).(11)✦

For a coefficient vector a∈ℝna\in\mathbb R^n, the linear spin observable faf_a is the real function on Ωn\Omega_n defined in (12). fa(x)=∑i=0n−1aixi.f_a(x)=\sum_{i=0}^{n-1}a_ix_i.(12)✦

The spin covariance matrix CWC_W is the real n×nn\times n matrix of centered Gibbs spin correlations in (13). (CW)ij=⟨(xi−⟨xi⟩W)(xj−⟨xj⟩W)⟩W.(C_W)_{ij} =\left\langle (x_i-\langle x_i\rangle_W)(x_j-\langle x_j\rangle_W) \right\rangle_W.(13)✦

The covariance norm κW\kappa_W is the Euclidean operator norm of CWC_W, as in (14). κW=∥CW∥2→2.\kappa_W=\|C_W\|_{2\to2}.(14)✦

The linear relaxation time Rlin(W)R_{\mathrm{lin}}(W) is the supremum of the variance-to-Dirichlet ratios of nonzero linear spin observables in (15), with an empty supremum assigned the value zero. Rlin(W)=supa∈ℝn\{0}Var⁡W(fa)DW(fa).R_{\mathrm{lin}}(W) =\sup_{a\in\mathbb R^n\setminus\{0\}} \frac{\operatorname{Var}_W(f_a)}{D_W(f_a)}.(15)✦

What Is Checked

The definitions are stated over real coupling arrays on every finite spin space, with a real parameter for the exponential kernel and a nonnegative integer number of random-scan attempts.