OctoInformBeta

OpenAI preprint, September 2026 · Section 1. Introduction

Theorem 1.1: Finite-graph comparison: Zero-field connectivity bounds likelihood ratios with arbitrary finite fields and common pins

Ferromagnetic Ising mixtures satisfy a uniform comparison.

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

The observable is the probability of a prescribed spin configuration on an observed set, when spins on a disjoint source set are sampled from a probability mixture and a third set is pinned in common.

A finite ferromagnetic multigraph G=(V,E)G=(V,E) has a finite vertex set VV and a finite edge list EE, with distinct endpoints ue,veu_e,v_e and a finite coupling Ke≥0K_e\geq 0 for each edge ee; parallel edges are retained.✦ A spin configuration σ∈Ω={−1,+1}V\sigma\in\Omega=\{-1,+1\}^{V} is a function assigning a physical spin to each vertex of VV.✦ For finite real fields h=(hv)v∈Vh=(h_v)_{v\in V}, the interaction exponent EhE_h and Boltzmann weight whw_h are given by (1). Eh(σ)=∑e∈EKeσueσve+∑v∈Vhvσv,wh(σ)=exp⁡(Eh(σ)).E_h(\sigma)=\sum_{e\in E}K_e\sigma_{u_e}\sigma_{v_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(A)Z_h(A) of a set A⊆ΩA\subseteq\Omega is the sum of its Boltzmann weights, as in (2).✦ Zh(A)=∑σ∈Awh(σ).Z_h(A)=\sum_{\sigma\in A}w_h(\sigma).(2)✦

Let I,J,F⊆VI,J,F\subseteq V be the observed, source and pinned sets, respectively. Two configurations σ,η∈Ω\sigma,\eta\in\Omega agree on a set S⊆VS\subseteq V when σv=ηv\sigma_v=\eta_v for every v∈Sv\in S.✦ For configurations a,b,τ∈Ωa,b,\tau\in\Omega, the conditional spin probability Ph(a∣b;τ)P_h(a\mid b;\tau) is the ratio in (3); the signs τ\tau prescribe the common pins. Ph(a∣b;τ)=Zh({σ∈Ω:σ|I=a|I,σ|J=b|J,σ|F=τ|F})Zh({σ∈Ω:σ|J=b|J,σ|F=τ|F}).P_h(a\mid b;\tau)= \frac{Z_h\bigl(\{\sigma\in\Omega: \sigma|_I=a|_I,\ \sigma|_J=b|_J,\ \sigma|_F=\tau|_F\}\bigr)} {Z_h\bigl(\{\sigma\in\Omega: \sigma|_J=b|_J,\ \sigma|_F=\tau|_F\}\bigr)}.(3)✦ A probability mixing weight λ\lambda on Ω\Omega is a nonnegative function with ∑b∈Ωλ(b)=1\sum_{b\in\Omega}\lambda(b)=1. For configurations a,τ∈Ωa,\tau\in\Omega and any real-valued weight λ\lambda on Ω\Omega, the mixture spin probability Mλ(a;τ)M_\lambda(a;\tau) is the weighted sum of conditional spin probabilities in (4).✦ Mλ(a;τ)=∑b∈Ωλ(b)Ph(a∣b;τ).M_\lambda(a;\tau)=\sum_{b\in\Omega}\lambda(b)P_h(a\mid b;\tau).(4)✦

The Comparison Probability

The undeleted vertices form VF=V\FV_F=V\setminus F✦, and the undeleted edge set EFE_F consists of edges with both endpoints in VFV_F.✦ For an open edge set ω⊆EF\omega\subseteq E_F, wired adjacency on VFV_F relates two vertices when both belong to II, both belong to JJ, or an edge of ω\omega joins them.✦ Wired cluster equivalence for an open edge set ω⊆EF\omega\subseteq E_F is the equivalence relation on VFV_F generated by wired adjacency; its classes are the clusters after identifying II and JJ separately.✦ The wired cluster count k(ω)k(\omega) of an open edge set ω⊆EF\omega\subseteq E_F is the number of wired clusters in VFV_F, including isolated vertices.✦

The zero-field Fortuin-Kasteleyn (FK) weight W0(ω)W_0(\omega) has cluster weight 22 and edge parameters pe=1−exp⁡(−2Ke)p_e=1-\exp(-2K_e), as in (5).✦ W0(ω)=2k(ω)∏e∈ωpe∏e∈EF\ω(1−pe),ω⊆EF.W_0(\omega)=2^{k(\omega)} \prod_{e\in\omega}p_e \prod_{e\in E_F\setminus\omega}(1-p_e), \qquad \omega\subseteq E_F.(5)✦ Choose representatives i∈I\Fi\in I\setminus F and j∈J\Fj\in J\setminus F. For any two undeleted vertices i,j∈VFi,j\in V_F, the terminal connection probability q0q_0 is the normalized zero-field FK weight of configurations in which ii and jj belong to the same wired cluster, as in (6).✦ Loops created by contraction contribute constants that cancel from this probability. q0=∑ω⊆EFi↔jW0(ω)∑ω⊆EFW0(ω).q_0= \frac{\displaystyle\sum_{\substack{\omega\subseteq E_F\\i\leftrightarrow j}} W_0(\omega)} {\displaystyle\sum_{\omega\subseteq E_F}W_0(\omega)}.(6)✦

The Claim

Theorem 1. Let GG be a finite loopless multigraph with nonnegative finite couplings KeK_e and arbitrary finite real fields hvh_v. Suppose that I,J,FI,J,F are pairwise disjoint, and choose i∈I\Fi\in I\setminus F and j∈J\Fj\in J\setminus F. Fix a,τ∈Ωa,\tau\in\Omega and two probability mixing weights λ,λ′\lambda,\lambda' on Ω\Omega. The finite graph comparison is (7). exp⁡(−4artanh⁡(q0))≤Mλ(a;τ)Mλ′(a;τ)≤exp⁡(4artanh⁡(q0)).\exp\!\bigl(-4\operatorname{artanh}(q_0)\bigr) \;\leq\; \frac{M_\lambda(a;\tau)}{M_{\lambda'}(a;\tau)} \;\leq\; \exp\!\bigl(4\operatorname{artanh}(q_0)\bigr).(7)✦

In the paper, the result reads: for a finite ferromagnetic graph with pairwise disjoint I,J,FI,J,F, surviving representatives i∈Ii\in I and j∈Jj\in J, arbitrary real fields hh and common signs τ\tau, both inequalities in (7) hold for any two nonnegative mixing weights on JJ of total mass one, with Dirac weights, each placing all mass on one source configuration, recovering individual conditional laws and no criticality, lattice geometry, field-sign or free-neighborhood hypothesis.

What Is Checked

The claim is stated over the conditional Gibbs sums and zero-field FK terminal connection probability on finite loopless ferromagnetic multigraphs, with arbitrary real fields, pairwise disjoint I,J,FI,J,F, surviving representatives i∈Ii\in I and j∈Jj\in J, fixed configurations a,τa,\tau, and two probability mixing weights on Ω\Omega. The zero-field FK terminal connection probability is independent of the choice of representatives within the observed and source sets.