OctoInformBeta

OpenAI preprint, September 2026 · Section 1. The theorem and the structure of the proof

Section 1 setting: heat-bath clocks, variance and Dirichlet form: Single-site resampling, continuous and discrete time, and Gibbs variance with the unscaled Dirichlet form

Heat-bath clocks and their quadratic forms.

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

An Ising system with nn sites has configuration space Ωn={−1,+1}n\Omega_n=\{-1,+1\}^n, where nn is a nonnegative integer and xix_i is the spin at site ii.✦ A disorder realization W=(Wij)i<jW=(W_{ij})_{i<j} assigns a real coupling to each unordered pair of distinct sites.✦ Its zero-field Hamiltonian HW:Ωn→ℝH_W:\Omega_n\to\mathbb R is the pair interaction energy HW(x)=∑i<jWijxixj.H_W(x)=\sum_{i<j}W_{ij}x_ix_j.(1)✦ The partition function ZWZ_W for the Hamiltonian HWH_W on Ωn\Omega_n is the normalizing constant of the Gibbs probability measure μW\mu_W on Ωn\Omega_n, with both given by (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)

For a site ii and a configuration xx, the set Ωi(x)\Omega_i(x) consists of the configurations that agree with xx outside ii: Ωi(x)={y∈Ωn:yj=xj for every j≠i}.\Omega_i(x)=\{y\in\Omega_n:y_j=x_j\text{ for every }j\ne i\}.(3)✦

The Claim

Definition 1. Fix a coupling array WW and a real observable f:Ωn→ℝf:\Omega_n\to\mathbb R. For a coupling array WW and a site ii, the single-site heat-bath kernel PiP_i on Ωn\Omega_n is the transition kernel that resamples the spin at ii from μW\mu_W conditional on the other spins. Its transition probabilities and its action on ff use the set in (3): Pi(x,y)={μW(y)∑z∈Ωi(x)μW(z),y∈Ωi(x),0,y∉Ωi(x),(Pif)(x)=∑y∈ΩnPi(x,y)f(y).\begin{aligned} P_i(x,y)&= \begin{cases} \displaystyle\frac{\mu_W(y)}{\sum_{z\in\Omega_i(x)}\mu_W(z)}, &y\in\Omega_i(x),\\ 0,&y\notin\Omega_i(x), \end{cases}\\ (P_if)(x)&=\sum_{y\in\Omega_n}P_i(x,y)f(y). \end{aligned}(4)✦ For a coupling array WW on nn sites, the discrete heat-bath kernel PP on Ωn\Omega_n is the average of the single-site kernels in (5), corresponding to one uniform-site attempt when n>0n>0; the reciprocal is taken to be zero when n=0n=0: P=1n∑i=1nPi.P=\frac1n\sum_{i=1}^nP_i.(5)✦ For a coupling array WW on nn sites, the generator LL of clocks of rate one at each site is the linear operator on real observables on Ωn\Omega_n given by (6), where II is the identity kernel: L=∑i=1n(Pi−I).L=\sum_{i=1}^n(P_i-I).(6)✦ The continuous kernel PtP_t is the matrix exponential in (7), defined for real tt and used at t≥0t\ge0 for the continuous-time dynamics: Pt=etL.P_t=e^{tL}.(7)✦ The Gibbs mean μW(f)\mu_W(f) is the expectation in (8) under the measure (2): μW(f)=∑x∈ΩnμW(x)f(x).\mu_W(f)=\sum_{x\in\Omega_n}\mu_W(x)f(x).(8)✦ The Gibbs variance VW(f)V_W(f) is the centered second moment (9): VW(f)=μW((f−μW(f))2).V_W(f)=\mu_W\bigl((f-\mu_W(f))^2\bigr).(9)✦ The unscaled Dirichlet form DW(f)D_W(f) is the sum of the squared resampling residuals in (10), using the operators in (4): DW(f)=∑i=1nμW((f−Pif)2).D_W(f)=\sum_{i=1}^n\mu_W\bigl((f-P_if)^2\bigr).(10)✦

In the conditional-expectation formulation of the zero-field dynamics, write EifE_i f for the conditional expectation given all spins outside ii, Var⁡μW(f)\operatorname{Var}_{\mu_W}(f) for Gibbs variance and ℰW(f)\mathcal E_W(f) for the heat-bath Dirichlet form.

Theorem 1. Fix a coupling array WW, a real observable f:Ωn→ℝf:\Omega_n\to\mathbb R and a site ii. The conditional-expectation and kernel formulations satisfy (11) at every configuration xx, with their variances and Dirichlet forms identified by (12) and (13): (Eif)(x)=∑y∈ΩnPi(x,y)f(y).(E_if)(x)=\sum_{y\in\Omega_n}P_i(x,y)f(y).(11)✦ Var⁡μW(f)=VW(f).\operatorname{Var}_{\mu_W}(f)=V_W(f).(12)✦ ℰW(f)=DW(f).\mathcal E_W(f)=D_W(f).(13)✦

What Is Checked

The definitions and identifications are stated for arbitrary real pair couplings on a finite set of sites and real observables on its spin configurations, with a site chosen for each single-site statement and a real time for the matrix exponential.