OctoInformBeta

OpenAI preprint, September 2026 · §2. A finite connection comparison

Lemma 2.2: For nonnegative observables at ordered distinct roots on a finite vertex set, with all hyperedge probabilities strictly below one

Nonnegative coordinates for increasing cluster observables.

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

Independent hyperedge percolation permits several vertices to be joined by a single open hyperedge. An ordered list of roots distinguishes clusters according to the first root they contain. Weighting such clusters by a function of their vertex sets gives a vector indexed by the roots.

A labelled hypergraph ℋ\mathcal H on a vertex set VV, with labels EE, is an assignment of a set ℋ(e)⊆V\mathcal H(e)\subseteq V to each e∈Ee\in E; empty hyperedges are permitted.✦ Give each label an opening probability pe∈[0,1]p_e\in[0,1]. The Bernoulli hyperedge measure ℙp\mathbb P_p is the product measure under which the states ω(e)\omega(e) are independent and ℙp(ω(e)=1)=pe\mathbb P_p(\omega(e)=1)=p_e.✦

For a forbidden vertex set DD, vertices uu and vv have open hyperedge adjacency when an open hyperedge disjoint from DD contains them both.✦ Open connectivity in ℋ\mathcal H after deletion of hyperedges meeting DD is connectivity by a finite chain of these adjacencies, including a chain of length zero.✦ For a vertex ss and a source set SS in VV, the cluster CDω(s)C_D^\omega(s) is the set of vertices connected to ss in ℋ\mathcal H using open hyperedges disjoint from DD, and the cluster union CDω(S)C_D^\omega(S) is the union of these clusters over vertices in SS, as in (1). CDω(s)={v∈V:s↔v using open hyperedges disjoint from D},CDω(S)=⋃s∈SCDω(s).C_D^\omega(s)=\{v\in V:s\longleftrightarrow v \text{ using open hyperedges disjoint from }D\}, \qquad C_D^\omega(S)=\bigcup_{s\in S}C_D^\omega(s).(1) Write Cω(s)=C⌀ω(s)C^\omega(s)=C_\varnothing^\omega(s) for the cluster without deletion.

Cluster Columns and Resampling

For roots x1,…,xnx_1,\ldots,x_n, the earlier roots at index kk form the set SkS_k in (2). For a vertex ss and a set SS, the admissible cluster domain 𝒟(s,S)\mathcal D(s,S) is the family of vertex sets containing ss and disjoint from SS. Sk={xj:j<k},𝒟(s,S)={U⊆V:s∈U,U∩S=⌀}.S_k=\{x_j:j<k\}, \qquad \mathcal D(s,S)=\{U\subseteq V:s\in U,\ U\cap S=\varnothing\}.(2)

The first root connection event AkiDA_{ki}^D in ℋ\mathcal H is the event that the cluster of xkx_k after deletion of hyperedges meeting DD avoids the earlier-root set SkS_k and contains xix_i, as in (3). AkiD={CDω(xk)∩Sk=⌀,xi∈CDω(xk)}.A_{ki}^D=\{C_D^\omega(x_k)\cap S_k=\varnothing, \ x_i\in C_D^\omega(x_k)\}.(3) The weighted cluster column wkFw_k^F for an observable F:2V→ℝF:2^V\to\mathbb R is the unnormalized expectation in (4). wkF(i)=𝔼p[𝟏Aki⌀F(Cω(xk))].w_k^F(i)=\mathbb E_p\!\left[ \mathbf 1_{A_{ki}^\varnothing}F(C^\omega(x_k))\right].(4)✦ The unnormalized cluster matrix MM has the constant-observable columns specified in (5). Mik=wk1(i).M_{ik}=w_k^1(i).(5)✦ The cluster coefficients akFa_k^F are the vector defined in (6). akF=M−1wkF.a_k^F=M^{-1}w_k^F.(6)✦

For a root ss, a forbidden set WW, and an observable FF, the cluster expectation after deletion msF(W)m_s^F(W) is the expected value of FF on the cluster of ss after deletion of hyperedges meeting WW, as in (7). msF(W)=𝔼p[F(CWω(s))].m_s^F(W)=\mathbb E_p\!\left[F(C_W^\omega(s))\right].(7) For a root ss and a source set SS in VV, conditional cluster resampling Ts,ST_{s,S} is the operator on observables defined by (8): for a vertex set UU, the forbidden set in the inner expectation is the cluster union from SS after deletion of hyperedges meeting UU. (Ts,SF)(U)=𝔼p[msF(CUω(S))].(T_{s,S}F)(U)=\mathbb E_p\!\left[m_s^F(C_U^\omega(S))\right].(8) For a root ss, source set SS, vertex set UU, and real observable FF on vertex sets, the residual cluster expectation Qs,SF(U)Q_{s,S}^F(U) is the expected correction weight assigned to UU when the weighted root cluster with weight FF is compared with its conditional cluster resampling at ss relative to SS.

The Claim

Theorem 1. Let VV be finite. Suppose that x1,…,xnx_1,\ldots,x_n are distinct, that pe<1p_e<1 for every hyperedge label, and that FF is nonnegative and inclusion-increasing on 𝒟(xk,Sk)\mathcal D(x_k,S_k). The cluster coefficients satisfy (9). akF(i)≥0(1≤i≤n).a_k^F(i)\ge0\qquad(1\le i\le n).(9)✦ For any ordered root list, probabilities pe∈[0,1]p_e\in[0,1], and real observable FF, the column decomposition is (10). wkF(i)=wkTxk,SkF(i)+∑j>kwjQxk,SkF(i)(1≤k,i≤n).w_k^F(i)=w_k^{\,T_{x_k,S_k}F}(i) +\sum_{j>k}w_j^{\,Q_{x_k,S_k}^F}(i) \qquad(1\le k,i\le n).(10)✦ Let s∉Ss\notin S, let pe<1p_e<1 for every hyperedge label, and let FF be nonnegative on 𝒟(s,S)\mathcal D(s,S). There is a constant c≥0c\ge0 for which (11) holds. limℓ→∞(Ts,SℓF)(U)=c(U∈𝒟(s,S)).\lim_{\ell\to\infty}(T_{s,S}^{\,\ell}F)(U)=c \qquad\bigl(U\in\mathcal D(s,S)\bigr).(11)✦

In the paper, the result reads: for a finite hypergraph with independent hyperedge probabilities strictly below one and any ordered list of distinct roots x1,…,xnx_1,\ldots,x_n, not necessarily all vertices, with C(x)C(x) denoting the random cluster of xx, set Ek={C(xk)∩Sk=⌀}E_k=\{C(x_k)\cap S_k=\varnothing\} and dk=ℙp(Ek)>0d_k=\mathbb P_p(E_k)>0, let μk\mu_k be the conditional law of C(xk)C(x_k) given EkE_k on 𝒟k=𝒟(xk,Sk)\mathcal D_k=\mathcal D(x_k,S_k), and put vkF(i)=𝔼μk[𝟏{xi∈U}F(U)]v_k^F(i)=\mathbb E_{\mu_k}[\mathbf 1_{\{x_i\in U\}}F(U)], hk=vk1h_k=v_k^1, and H=(h1⋯hn)H=(h_1\ \cdots\ h_n); for every nonnegative inclusion-increasing FF on 𝒟k\mathcal D_k, the vector vkF−μk(F)hkv_k^F-\mu_k(F)h_k belongs to the nonnegative cone generated by hjh_j with j>kj>k, and in particular H−1vkFH^{-1}v_k^F is coordinatewise nonnegative.

What Is Checked

The statements concern Bernoulli product measures on labelled hypergraphs with a finite vertex set, their cluster columns and resampling operators; the coefficient inequality and the constant limit require pe<1p_e<1 for every label, with distinct ordered roots and a nonnegative increasing observable on 𝒟(xk,Sk)\mathcal D(x_k,S_k) for the former, and s∉Ss\notin S and a nonnegative observable on 𝒟(s,S)\mathcal D(s,S) for the latter.