OctoInformBeta

OpenAI preprint, September 2026 · Section 2. Pointwise comparison through buffers

Lemma 2.2, proof: normalized conditional and mixture ratios: Conditional configuration probabilities and arbitrary source mixtures in finite ferromagnetic Ising models

Extreme log odds control mixture ratios.

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 the conditional distribution of spins on an observed set II as the prescribed spins on a source set JJ vary, with fixed signs τ\tau on a pinned set FF. The source prescription may also be sampled from a probability distribution.

A finite ferromagnetic weighted multigraph GG has a finite vertex set VV, a finite set EE of separately indexed edges, endpoint maps ℓ,r:E→V\ell,r:E\to V with ℓ(e)≠r(e)\ell(e)\ne r(e), and real couplings Ke≥0K_e\geq0.✦ Parallel edges are allowed. A spin configuration σ∈Ω={−1,+1}V\sigma\in\Omega=\{-1,+1\}^{V} assigns a physical spin to every vertex.✦ Two configurations agree on a set S⊆VS\subseteq V when their restrictions to SS coincide.✦

For a real field h=(hv)v∈Vh=(h_v)_{v\in V}, the Ising interaction exponent Eh(σ)E_h(\sigma) and the Boltzmann weight wh(σ)w_h(\sigma) are Eh(σ)=∑e∈EKeσℓ(e)σr(e)+∑v∈Vhvσv,wh(σ)=exp⁡(Eh(σ)).E_h(\sigma)=\sum_{e\in E}K_e\sigma_{\ell(e)}\sigma_{r(e)} +\sum_{v\in V}h_v\sigma_v, \qquad w_h(\sigma)=\exp\bigl(E_h(\sigma)\bigr).(1) The constrained partition sum Zh(P)Z_h(P) for a condition PP on configurations is Zh(P)=∑σ∈ΩP(σ)wh(σ).Z_h(P)=\sum_{\substack{\sigma\in\Omega\\P(\sigma)}}w_h(\sigma).(2)✦

Fix pairwise disjoint sets I,J,F⊆VI,J,F\subseteq V and configurations a,b,τ∈Ωa,b,\tau\in\Omega. The unnormalized joint marginal W(a,b)W(a,b) with pins τ\tau is the constrained sum W(a,b)=Zh(σI=aI,σJ=bJ,σF=τF),W(a,b)=Z_h\bigl(\sigma_I=a_I,\ \sigma_J=b_J,\ \sigma_F=\tau_F\bigr),(3)✦ where subscripts denote restriction and a list of constraints means their intersection.

The Gibbs measure ℙG,h\mathbb P_{G,h} normalizes the weights in (1). The conditional spin probability Ca(b)C_a(b) is the probability of the observed prescription aIa_I, given the source prescription bJb_J and the pins τF\tau_F, expressed through the constrained sums of (2): Ca(b)=ℙG,h(σI=aI∣σJ=bJ,σF=τF)=W(a,b)Zh(σJ=bJ,σF=τF).C_a(b)=\mathbb P_{G,h}\bigl(\sigma_I=a_I\mid \sigma_J=b_J,\ \sigma_F=\tau_F\bigr) =\frac{W(a,b)}{Z_h(\sigma_J=b_J,\ \sigma_F=\tau_F)}.(4)✦

The extreme log odds LL is the following combination of the joint marginals (3), with +\boldsymbol{+} and −\boldsymbol{-} the constant positive and negative configurations: L=log⁡W(+,+)+log⁡W(−,−)−log⁡W(+,−)−log⁡W(−,+).L=\log W(\boldsymbol{+},\boldsymbol{+}) +\log W(\boldsymbol{-},\boldsymbol{-}) -\log W(\boldsymbol{+},\boldsymbol{-}) -\log W(\boldsymbol{-},\boldsymbol{+}).(5)✦

For real mixing weights μ:Ω→ℝ\mu:\Omega\to\mathbb R, the mixture spin probability Mμ(a)M_\mu(a) is the weighted sum of the conditional probabilities (4): Mμ(a)=∑b∈Ωμ(b)Ca(b).M_\mu(a)=\sum_{b\in\Omega}\mu(b)C_a(b).(6)✦

The zero-field Fortuin-Kasteleyn terminal connection probability q0q_0 is the probability that the two terminals are connected in the random-cluster model with cluster weight 22 and edge probabilities pe=1−exp⁡(−2Ke)p_e=1-\exp(-2K_e) on the graph obtained from GG by deleting FF and its incident edges and identifying II and JJ separately as the two terminals. Parallel edges remain distinct; loops created by identification contribute constants and cancel from terminal connection probabilities.

The Claim

Theorem 1. Let GG be a finite ferromagnetic weighted multigraph with real field hh, let I,J,FI,J,F be pairwise disjoint, and fix pin signs τ\tau. With the extreme log odds LL of (5), every observed prescription aa and every pair of source prescriptions b,b′∈Ωb,b'\in\Omega satisfy exp⁡(−L)≤Ca(b)Ca(b′)≤exp⁡(L).\exp(-L)\leq\frac{C_a(b)}{C_a(b')}\leq\exp(L).(7)✦ Let μ,ν:Ω→[0,∞)\mu,\nu:\Omega\to[0,\infty) be arbitrary probability weights, so that ∑bμ(b)=∑bν(b)=1\sum_b\mu(b)=\sum_b\nu(b)=1. For the mixture probabilities of (6), exp⁡(−L)≤Mμ(a)Mν(a)≤exp⁡(L).\exp(-L)\leq\frac{M_\mu(a)}{M_\nu(a)}\leq\exp(L).(8)✦

In the paper, the result reads: Theorem 1.1 bounds the ratio of actual normalized configuration probabilities with arbitrary probability mixtures on both sides by exp⁡(−4artanh⁡(q0))\exp(-4\operatorname{artanh}(q_0)) and exp⁡(4artanh⁡(q0))\exp(4\operatorname{artanh}(q_0)), with intermediate bounds exp⁡(−L)\exp(-L) and exp⁡(L)\exp(L) and the extreme log odds estimate L≤4artanh⁡(q0)L\leq4\operatorname{artanh}(q_0), while mixing preserves the same constants without another factor in the exponent.

What Is Checked

The claims (7) and (8) are stated over finite ferromagnetic weighted multigraphs, arbitrary finite real fields, pairwise disjoint observed, source and pinned sets, and arbitrary spin prescriptions; the mixture claim additionally requires two nonnegative weight families of total mass one.