OctoInform Beta

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

Proposition 3.2 [Saturation]: One on which conditional expectations onto data fields commute with pullback along every further extension; limits of towers of extensions are array objects

Every array object has a saturated extension.

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

Fix a standard Borel space XX with a measurable automorphism TT, and a finite index set II. The objects here are laws of random arrays (y(v))v∈In(y(v))_{v\in I^n}, indexed by tuples in InI^n and with entries in a state space YY over XX, invariant under an automorphism of YY acting entrywise and carrying Οƒ\sigma-algebras of data on the arrays of each order. Passing to an extension, a larger state space mapping to YY, may change the conditional expectation of a function of the arrays onto these Οƒ\sigma-algebras. The question is whether every such object has an extension past which no further extension changes them, and whether the inverse limit of a tower of extensions is again an object of the same kind.

Arrays and Array Objects

An array of order nn over a space YY is a function y:Inβ†’Yy\colon I^n\to Y that assigns an entry y(v)y(v) to every index tuple v=(v0,…,vnβˆ’1)v=(v_0,\dots,v_{n-1}) in InI^n, so that the arrays of order nn form the product YInY^{I^n}✦; an array of order zero is a single point of YY.✦ A map f:Yβ†’Zf\colon Y\to Z acts on arrays entrywise, by fn(y)(v)=f(y(v))f_n(y)(v)=f(y(v)).✦ For an array yy of order n+1n+1, an axis a∈{0,…,n}a\in\{0,\dots,n\} and an index i∈Ii\in I, the plane y|a=iy|_{a=i} is the array of order nn whose entry at vv is the entry of yy at the tuple obtained by inserting ii into vv at position aa.✦ For a set JβŠ†IJ\subseteq I, the last-axis planes of yy along JJ form the family Ξ J(y)=(y|n=j)j∈J\Pi_J(y)=(y|_{n=j})_{j\in J}.✦

Initial array laws mXm^X of (X,T)(X,T) over II are probability measures mnXm^X_n on XInX^{I^n}, nβ‰₯0n\ge0, each invariant under TT acting entrywise, that are consistent and proper: for every nn, every axis aa, every index ii and every proper subset JJ of II, (y↦y|a=i)*mn+1X=mnX,(Ξ J)*mn+1X=(mnX)βŠ—J.(y\mapsto y|_{a=i})_*\,m^X_{n+1}=m^X_n, \qquad (\Pi_J)_*\,m^X_{n+1}=(m^X_n)^{\otimes J}.(1)✦

Laws mnm_n on YInY^{I^n}, nβ‰₯0n\ge0, are conditionally local for a measurable map Ο€:Yβ†’U\pi\colon Y\to U if, for every nn, every proper subset JJ of II and almost every value pp of the parameter array Ο€n+1(y)\pi_{n+1}(y) under mn+1m_{n+1}, the planes y|n=jy|_{n=j} with j∈Jj\in J are independent under mn+1m_{n+1} given Ο€n+1(y)=p\pi_{n+1}(y)=p, the plane at jj having law ΞΊn(p|n=j)\kappa_n(p|_{n=j}), where ΞΊn\kappa_n is the conditional law under mnm_n of an array of order nn given its parameter array Ο€n(y)\pi_n(y).

An array object MM over mXm^X consists of nonempty standard Borel spaces YY of states and UU of parameters with measurable automorphisms TYT_Y and TUT_U, measurable maps x:Yβ†’Xx\colon Y\to X and Ο€:Yβ†’U\pi\colon Y\to U with x∘TY=T∘xx\circ T_Y=T\circ x and Ο€βˆ˜TY=TUβˆ˜Ο€\pi\circ T_Y=T_U\circ\pi, and probability measures mnm_n on YInY^{I^n}, nβ‰₯0n\ge0, that are invariant under TYT_Y acting entrywise, lie over the initial array laws mXm^X, (xn)*mn=mnX(x_n)_*\,m_n=m^X_n, are consistent and proper as in (1) and are conditionally local for Ο€\pi, with xx and Ο€\pi independent under m0m_0, a measure on YY.✦

Extensions and Saturation

The kinds of data field over II are the parameter, the outside at jj and the enlarged at jj, for j∈Ij\in I, so that there are 2|I|+12|I|+1 kinds.✦ An array object MM has, at every order nβ‰₯1n\ge1, one data field of each kind, a sub-Οƒ\sigma-algebra of the Borel Οƒ\sigma-algebra of YInY^{I^n}: the parameter field 𝒫n\mathcal P_n and, for each j∈Ij\in I, the outside field π’ͺn,j\mathcal O_{n,j} and the enlarged field 𝒒n,j\mathcal G_{n,j}.✦ For a data field π’Ÿ\mathcal D of order nn, the conditional expectation EM[β‹…βˆ£π’Ÿ]E_M[\,\cdot\mid\mathcal D] is the orthogonal projection of the complex Hilbert space L2(mn)L^2(m_n) onto the closed subspace of π’Ÿ\mathcal D-measurable functions.✦

An extension of an array object MM is an array object Mβ€²M' over the same initial laws, with states Yβ€²Y', parameters Uβ€²U', automorphisms TYβ€²T_{Y'} and TUβ€²T_{U'}, maps xβ€²x' and Ο€β€²\pi' and laws mnβ€²m'_n, together with measurable maps ΞΈ:Yβ€²β†’Y\theta\colon Y'\to Y and ρ:Uβ€²β†’U\rho\colon U'\to U such that θ∘TYβ€²=TY∘θ\theta\circ T_{Y'}=T_Y\circ\theta, ρ∘TUβ€²=TU∘ρ\rho\circ T_{U'}=T_U\circ\rho and x∘θ=xβ€²,Ο€βˆ˜ΞΈ=Οβˆ˜Ο€β€²,(ΞΈn)*mnβ€²=mn(nβ‰₯0).x\circ\theta=x', \qquad \pi\circ\theta=\rho\circ\pi', \qquad (\theta_n)_*\,m'_n=m_n\quad(n\ge0).(2)✦ We write ΞΈ:Mβ€²β†’M\theta\colon M'\to M. For an extension ΞΈ:Mβ€²β†’M\theta\colon M'\to M and nβ‰₯0n\ge0, the pullback along ΞΈ\theta is the map F↦F∘θnF\mapsto F\circ\theta_n, composition with the measure-preserving map ΞΈn\theta_n; it is a linear isometry of L2(mn)L^2(m_n) into L2(mnβ€²)L^2(m'_n).✦

An array object MM is saturated if, for every extension ΞΈ:Mβ€²β†’M\theta\colon M'\to M, every order nβ‰₯1n\ge1, every data field π’Ÿ\mathcal D of MM of order nn, with π’Ÿβ€²\mathcal D' the field of Mβ€²M' of the same kind, and every F∈L2(mn)F\in L^2(m_n), EMβ€²[F∘θnβˆ£π’Ÿβ€²]=EM[Fβˆ£π’Ÿ]∘θn.E_{M'}[F\circ\theta_n\mid\mathcal D']=E_M[F\mid\mathcal D]\circ\theta_n.(3)✦

Towers and Their Limits

For measurable maps Ξ²s:Ws+1β†’Ws\beta_s\colon W_{s+1}\to W_s, sβ‰₯0s\ge0, called the bonding maps, the inverse limit lim←(Ws,Ξ²s)\underset{\longleftarrow}{\lim}(W_s,\beta_s) is the set of paths, the sequences ΞΎ=(ΞΎs)sβ‰₯0\xi=(\xi_s)_{s\ge0} in ∏sWs\prod_s W_s with Ξ²s(ΞΎs+1)=ΞΎs\beta_s(\xi_{s+1})=\xi_s for every ss, with the projections ξ↦ξs\xi\mapsto\xi_s. Maps of the stages that commute with the bonding maps act on paths stagewise✦; in particular, automorphisms of the stages that commute with the bonding maps give an automorphism of the inverse limit.

Given probability measures ΞΌs\mu_s on WsW_s, let Ξ»s(w)=ΞΌs+1(β‹…βˆ£Ξ²s=w)\lambda_s(w)=\mu_{s+1}(\,\cdot\mid\beta_s=w) be the conditional law under ΞΌs+1\mu_{s+1} of a point of Ws+1W_{s+1} given its image in WsW_s.✦ The sequential limit law is the law on ∏sWs\prod_s W_s, given by the Ionescu-Tulcea theorem, of the Markov chain with ΞΎ0∼μ0,Law⁡(ΞΎs+1∣ξ0,…,ΞΎs)=Ξ»s(ΞΎs),\xi_0\sim\mu_0, \qquad \operatorname{Law}(\xi_{s+1}\mid\xi_0,\dots,\xi_s)=\lambda_s(\xi_s),(4)✦ and the inverse limit law is its restriction to the set of paths.✦ For nβ‰₯0n\ge0, applying Ξ²s\beta_s entrywise makes the arrays WsInW_s^{I^n} an inverse system whose paths are the arrays of paths✦. The inverse limit array law of probability measures Ξ½s\nu_s on WsInW_s^{I^n} is the inverse limit law of (Ξ½s)sβ‰₯0(\nu_s)_{s\ge0} for this system, a measure on (lim←Ws)In(\underset{\longleftarrow}{\lim} W_s)^{I^n}.

A tower of array objects over mXm^X is a sequence M0,M1,…M_0,M_1,\dots of array objects over mXm^X, the stage MsM_s with states YsY_s, parameters UsU_s, automorphisms TYsT_{Y_s} and TUsT_{U_s}, maps x(s)x^{(s)} and Ο€(s)\pi^{(s)} and laws mn(s)m^{(s)}_n, together with extensions Οƒs:Ms+1β†’Ms\sigma_s\colon M_{s+1}\to M_s with parameter maps ρs:Us+1β†’Us\rho_s\colon U_{s+1}\to U_s.✦ The limit M∞M_\infty of the tower consists of the states Y∞=lim←(Ys,Οƒs)Y_\infty=\underset{\longleftarrow}{\lim}(Y_s,\sigma_s) and the parameters U∞=lim←(Us,ρs)U_\infty=\underset{\longleftarrow}{\lim}(U_s,\rho_s), the automorphisms TY∞T_{Y_\infty} and TU∞T_{U_\infty} acting stagewise by TYsT_{Y_s} and TUsT_{U_s}, the factor map x∞(y)=x(0)(y0)x_\infty(y)=x^{(0)}(y_0)✦, the parameter map Ο€βˆž(y)=(Ο€(s)(ys))sβ‰₯0\pi_\infty(y)=(\pi^{(s)}(y_s))_{s\ge0}✦, and at each order nn the law mn∞m^\infty_n on Y∞InY_\infty^{I^n}, the inverse limit array law of (mn(s))sβ‰₯0(m^{(s)}_n)_{s\ge0}.✦

The Claim

Theorem 1. Let mXm^X be initial array laws of a measurable automorphism TT of a standard Borel space XX, over a finite index set II.

(a) Every array object MM over mXm^X has an extension ΞΈ:M*β†’Mwith M* saturated.\theta\colon M^{*}\to M \quad\text{with } M^{*} \text{ saturated}.(5)✦

(b) For every tower of array objects over mXm^X, the limit M∞=(Y∞,U∞,TY∞,TU∞,x∞,Ο€βˆž,(mn∞)nβ‰₯0)M_\infty=\bigl(Y_\infty,\,U_\infty,\,T_{Y_\infty},\,T_{U_\infty},\,x_\infty,\,\pi_\infty,\, (m^\infty_n)_{n\ge0}\bigr)(6)✦ is an array object over mXm^X.

(c) In particular the laws of M∞M_\infty are conditionally local: for every order nn, every proper subset JJ of II and almost every value pp of the parameter array (Ο€βˆž)n+1(y)(\pi_\infty)_{n+1}(y) under mn+1∞m^\infty_{n+1}, Law⁡mn+1∞(Ξ J(y)|(Ο€βˆž)n+1(y)=p)=⨂j∈JΞΊn∞(p|n=j),\operatorname{Law}_{m^\infty_{n+1}}\bigl(\Pi_J(y)\mathrel{\big|}(\pi_\infty)_{n+1}(y)=p\bigr) =\bigotimes_{j\in J}\kappa^\infty_n\bigl(p|_{n=j}\bigr),(7)✦ where ΞΊn∞\kappa^\infty_n is the conditional law under mn∞m^\infty_n of an array of order nn given its parameter array.

In the paper, the result reads: every array object has an extension such that, for every nβ‰₯1n\ge1, every F∈L2(mn)F\in L^2(m_n) and every π’Ÿ\mathcal D among 𝒫n\mathcal P_n, 𝒒n,j\mathcal G_{n,j} and π’ͺn,j\mathcal O_{n,j}, no further extension changes E[Fβˆ£π’Ÿ]E[F\mid\mathcal D], that is, E[F∘θnβˆ£π’Ÿβ€²]=E[Fβˆ£π’Ÿ]∘θnE[F\circ\theta_n\mid\mathcal D']=E[F\mid\mathcal D]\circ\theta_n.

What Is Checked

The claim is stated over array objects over initial array laws of a measurable automorphism of a standard Borel space with a finite index set, over their extensions with the pullbacks along them and the conditional expectations onto the parameter, outside and enlarged fields in complex L2L^2, and over towers of extensions with their limits, built from inverse limits of standard Borel spaces and the sequential limit laws of conditional laws.