OctoInform Beta

OpenAI preprint, September 2026 · §3. Ordered arrays and conditional locality · §3

Lemma 3.3 [Local conditioning]: If BB sees OO only through the observations Aj=aj(Ej,Oj)A_j = a_j(E_j, O_j), then given (E,A,B)(E, A, B) the OjO_j stay independent, each updated by (Ej,Aj)(E_j, A_j) alone

Conditioning on local observations preserves conditional independence.

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

Take a random parameter EE and finitely many random objects OjO_j, j∈Jj \in J, that are independent given EE, the conditional law of OjO_j being Kj(Ej,⋅)K_j(E_j, \cdot) for a local parameter Ej=ej(E)E_j = e_j(E). Each OjO_j is seen through an observation Aj=aj(Ej,Oj)A_j = a_j(E_j, O_j). A further random variable BB is then drawn from a law that depends on EE and, of O=(Oj)j∈JO = (O_j)_{j \in J}, only on the observations A=(Aj)j∈JA = (A_j)_{j \in J}. The question is what EE, AA and BB together say about the pieces OjO_j: whether BB couples them, and whether the update of OjO_j uses anything beyond EjE_j and AjA_j.

Conditional Laws and Kernels

A conditional law of YY given XX, for random variables XX and YY on a measure space (Ω,P)(\Omega, P) with values in measurable spaces 𝒳\mathcal X and 𝒴\mathcal Y, is a Markov kernel KK from 𝒳\mathcal X to 𝒴\mathcal Y such that the joint law of (X,Y)(X, Y) is the law PXP_X of XX followed by KK: P(X∈S,Y∈T)=∫SK(x,T)PX(dx)P(X \in S,\ Y \in T) = \int_S K(x, T)\, P_X(dx) for all measurable S⊆𝒳S \subseteq \mathcal X and T⊆𝒴T \subseteq \mathcal Y. We then write Law(Y∣X=x)=K(x,⋅)\mathrm{Law}(Y \mid X = x) = K(x, \cdot).

The product kernel ⨂j∈JKj\bigotimes_{j \in J} K_j of Markov kernels KjK_j from 𝒳\mathcal X to 𝒴j\mathcal Y_j, indexed by a finite set JJ, is the Markov kernel from 𝒳\mathcal X to ∏j∈J𝒴j\prod_{j \in J} \mathcal Y_j that sends each point to the product of the measures: (⨂j∈JKj)(x,⋅)=⨂j∈JKj(x,⋅).\Bigl(\bigotimes_{j \in J} K_j\Bigr)(x, \cdot) = \bigotimes_{j \in J} K_j(x, \cdot).(1)✦

The observation kernel KaK^{a} of a kernel KK from ℰ\mathcal E to 𝒪\mathcal O through a measurable map a:ℰ×𝒪→𝒜a \colon \mathcal E \times \mathcal O \to \mathcal A is the kernel from ℰ\mathcal E to 𝒜×𝒪\mathcal A \times \mathcal O that sends ee to the law of the pair (a(e,y),y)(a(e, y), y) for yy drawn from K(e,⋅)K(e, \cdot): Ka(e,⋅)=∫𝒪δ(a(e,y),y)K(e,dy).K^{a}(e, \cdot) = \int_{\mathcal O} \delta_{(a(e, y),\, y)}\, K(e, dy).(2)✦

The disintegration of KaK^{a}, for KK Markov, 𝒪\mathcal O standard Borel and 𝒜\mathcal A countably generated, is the Markov kernel QQ from ℰ×𝒜\mathcal E \times \mathcal A to 𝒪\mathcal O that gives the law of yy given a(e,y)=ca(e, y) = c for yy drawn from K(e,⋅)K(e, \cdot): Ka(e,dcdy)=K‾a(e,dc)Q(e,c;dy),K^{a}(e, dc\, dy) = \bar K^{a}(e, dc)\, Q(e, c; dy),(3) where K‾a(e,⋅)\bar K^{a}(e, \cdot) is the image of K(e,⋅)K(e, \cdot) under y↦a(e,y)y \mapsto a(e, y).

The Claim

Theorem 1 (Local conditioning). Let JJ be a finite set and PP an s-finite measure on Ω\Omega, that is, a countable sum of finite measures. Let EE, O=(Oj)j∈JO = (O_j)_{j \in J} and BB be random variables on Ω\Omega with values, respectively, in a measurable space 𝒳\mathcal X, in a product ∏j∈J𝒪j\prod_{j \in J} \mathcal O_j of nonempty standard Borel spaces, and in a measurable space ℬ\mathcal B. For each j∈Jj \in J, let ej:𝒳→ℰje_j \colon \mathcal X \to \mathcal E_j be a measurable map to a measurable space, KjK_j a Markov kernel from ℰj\mathcal E_j to 𝒪j\mathcal O_j, and aj:ℰj×𝒪j→𝒜ja_j \colon \mathcal E_j \times \mathcal O_j \to \mathcal A_j a measurable map to a countably generated measurable space; let QjQ_j be the disintegration of the observation kernel KjajK_j^{a_j}, and put Aj=aj(ej(E),Oj)A_j = a_j(e_j(E), O_j) and A=(Aj)j∈JA = (A_j)_{j \in J}. Suppose that Law(O∣E=x)=⨂j∈JKj(ej(x),⋅)\mathrm{Law}(O \mid E = x) = \bigotimes_{j \in J} K_j\bigl(e_j(x), \cdot\bigr)(4) and that, for a Markov kernel LL from 𝒳×∏j∈J𝒜j\mathcal X \times \prod_{j \in J} \mathcal A_j to ℬ\mathcal B, Law(B∣E=x,O=y)=L(x,(aj(ej(x),yj))j∈J;⋅).\mathrm{Law}\bigl(B \mid E = x,\ O = y\bigr) = L\Bigl(x, \bigl(a_j(e_j(x), y_j)\bigr)_{j \in J};\ \cdot\Bigr).(5) Then Law(O∣E=x,A=c,B=β)=⨂j∈JQj(ej(x),cj;⋅).\mathrm{Law}\bigl(O \mid E = x,\ A = c,\ B = \beta\bigr) = \bigotimes_{j \in J} Q_j\bigl(e_j(x), c_j;\ \cdot\bigr).(6)✦

In the paper, the result reads: if Law((Oj)j∣E)=⨂jKj(Ej,⋅)\mathrm{Law}((O_j)_{j} \mid E) = \bigotimes_j K_j(E_j, \cdot) with Ej=ej(E)E_j = e_j(E) and Aj=aj(Ej,Oj)A_j = a_j(E_j, O_j), and the law of BB given (E,O)(E, O) depends only on (E,A)(E, A), then Law((Oj)j∣E,A,B)=⨂jQj(Ej,Aj;⋅)\mathrm{Law}((O_j)_j \mid E, A, B) = \bigotimes_j Q_j(E_j, A_j; \cdot), where QjQ_j is the law of OjO_j given (Ej,Aj)(E_j, A_j).

What Is Checked

The claim is stated over conditional laws under an s-finite measure PP, the product kernel (1), and the disintegrations (3) of the observation kernels (2), for a finite index set JJ, nonempty standard Borel spaces 𝒪j\mathcal O_j and countably generated spaces 𝒜j\mathcal A_j.