OctoInform Beta

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

Lemma 3.4 [Rolling an axis into the local state]: For |R|+2≤|I||R|+2\le|I| and a∉Ra\notin R, the rolled laws are conditional for RR and give an array object extending MM

Rolling an axis of a saturated object gives an 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 measurable automorphism TT of a standard Borel space XX and a finite set II of slots. An array object is a tower of invariant laws on arrays of states lying over XX, indexed by tuples of slots, whose planes, the subarrays obtained by fixing one slot in one coordinate, are conditionally independent given the parameters of the entries. The question is when the laws obtained by grouping the entries of an array along its first axis, with a parameter that remembers part of the resulting row of states, form an array object again, one that extends the object they came from.

Arrays and Array Objects

An array of order nn over a space YY is a family y=(yv)v∈Iny=(y_v)_{v\in I^n} assigning an entry yv∈Yy_v\in Y to every index tuple vv of length nn in II; the arrays form the product space YInY^{I^n}✦, and an array of order zero is identified with its single entry, so that YI0=YY^{I^0}=Y.✦ A map f:Y→Zf\colon Y\to Z acts on arrays entrywise, f(y)v=f(yv)f(y)_v=f(y_v), under the same letter.✦ For an array yy of order n+1n+1, an axis ℓ∈{0,…,n}\ell\in\{0,\dots,n\} and a slot i∈Ii\in I, the plane y|ℓ=iy|_{\ell=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 ℓ\ell.✦ The last-axis plane at jj is y|j=y|n=jy|_j=y|_{n=j}, and the planes of yy along a set JJ of slots are the family (y|j)j∈J(y|_j)_{j\in J}.✦

Initial array laws mXm^X of TT are probability measures mnXm^X_n on XInX^{I^n}, one for each n≥0n\ge0, each invariant under TT applied entrywise, that are consistent and proper: for every nn, every axis ℓ\ell, every slot ii and every proper subset JJ of II, (y↦y|ℓ=i)*mn+1X=mnX,(y↦(y|j)j∈J)*mn+1X=⨂j∈JmnX.\bigl(y\mapsto y|_{\ell=i}\bigr)_*m^X_{n+1}=m^X_n,\qquad \bigl(y\mapsto(y|_j)_{j\in J}\bigr)_*m^X_{n+1}=\bigotimes_{j\in J}m^X_n .(1)✦

An array object MM over mXm^X consists of a standard Borel space YY of states with a measurable automorphism TYT_Y and a measurable factor map x:Y→Xx\colon Y\to X satisfying x∘TY=T∘xx\circ T_Y=T\circ x; a standard Borel space UU of parameters with a measurable automorphism TUT_U and a measurable parameter map u:Y→Uu\colon Y\to U satisfying u∘TY=TU∘uu\circ T_Y=T_U\circ u; and probability measures mnm_n on YInY^{I^n}, the laws of MM.✦ Each mnm_n is invariant under TYT_Y and lies over the base, x*mn=mnXx_*m_n=m^X_n; the laws are consistent and proper as in (1); they are locally independent, the maps xx and uu being independent under m0m_0, a law on YY; and they are conditionally local: with κn(q)\kappa_n(q) the conditional law under mnm_n of an array yy given its parameter array u(y)=qu(y)=q, for every nn and every proper subset JJ of II, Law⁡mn+1((y|j)j∈J|u(y)=p)=⨂j∈Jκn(p|j)for u*mn+1-almost every p.\operatorname{Law}_{m_{n+1}}\bigl((y|_j)_{j\in J}\mathrel{\big|}u(y)=p\bigr) =\bigotimes_{j\in J}\kappa_n\bigl(p|_j\bigr) \quad\text{for $u_*m_{n+1}$-almost every $p$.}(2)✦ The array objects over mXm^X form a class 𝑶𝒃𝒋(mX)\mathbf{Obj}(m^X).

An extension of MM by an array object NN over mXm^X, with states Y′Y', parameters U′U', factor map x′x', parameter map u′u' and laws mn′m'_n, is a pair of measurable maps σ:Y′→Y\sigma\colon Y'\to Y and ρ:U′→U\rho\colon U'\to U that intertwine the automorphisms of NN with those of MM and satisfy x∘σ=x′,u∘σ=ρ∘u′,σ*mn′=mn(n≥0).x\circ\sigma=x',\qquad u\circ\sigma=\rho\circ u',\qquad \sigma_*m'_n=m_n\quad(n\ge0).(3)✦ The pullback of an extension (σ,ρ)(\sigma,\rho) is the map σ*:L2(mn)→L2(mn′)\sigma^*\colon L^2(m_n)\to L^2(m'_n), F↦F∘σF\mapsto F\circ\sigma, with σ\sigma applied entrywise; it is a linear isometry of complex Hilbert spaces.

For n≥1n\ge1 the arrays of order nn of an array object carry 2|I|+12|I|+1 data fields, sub-σ\sigma-algebras of the Borel σ\sigma-algebra of YInY^{I^n}: the parameter field 𝒫n\mathcal{P}_n and, for each slot jj, the outside field 𝒪n,j\mathcal{O}_{n,j} and the enlarged field 𝒢n,j\mathcal{G}_{n,j}.✦ The projection P𝒟MP^M_{\mathcal{D}} onto a data field 𝒟\mathcal{D} is the orthogonal projection of L2(mn)L^2(m_n) onto the closed subspace of 𝒟\mathcal{D}-measurable functions, the conditional expectation onto 𝒟\mathcal{D}.✦ An array object MM over mXm^X is saturated if the projections P𝒟MP^M_{\mathcal{D}} onto its data fields commute with the pullback along every extension of MM: for every extension (σ,ρ)(\sigma,\rho) of MM by an array object NN over mXm^X, every n≥1n\ge1, every data field 𝒟\mathcal{D} and every F∈L2(mn)F\in L^2(m_n), P𝒟N(σ*F)=σ*(P𝒟MF),P^N_{\mathcal{D}}\bigl(\sigma^*F\bigr)=\sigma^*\bigl(P^M_{\mathcal{D}}F\bigr),(4)✦ where on the left 𝒟\mathcal{D} is the field of NN of the same name.

Rolling an Axis

Let MM be an array object over mXm^X. The roll of order nn is the measurable isomorphism roll⁡n:YIn+1→(YI)In\operatorname{roll}_n\colon Y^{I^{n+1}}\to(Y^I)^{I^n} given by (roll⁡ny)w=(y(c,w0,…,wn−1))c∈I,w∈In.(\operatorname{roll}_n y)_w=\bigl(y_{(c,w_0,\dots,w_{n-1})}\bigr)_{c\in I},\qquad w\in I^n .(5)✦ The rolled state space is YIY^I✦ with the automorphism TYT_Y in every coordinate✦, and the rolled law of order nn is the probability measure μn=(roll⁡n)*mn+1\mu_n=(\operatorname{roll}_n)_*m_{n+1} on (YI)In(Y^I)^{I^n}; at order zero, μ0\mu_0 is m1m_1 on YIY^I.✦ For a set RR of slots, the rolled parameter space is UI×YRU^I\times Y^R✦, with TUT_U in every coordinate of UIU^I and TYT_Y in every coordinate of YRY^R✦, and the rolled parameter map uR:YI→UI×YRu^R\colon Y^I\to U^I\times Y^R is uR(y)=((u(yc))c∈I,(yc)c∈R).u^R(y)=\bigl((u(y_c))_{c\in I},\,(y_c)_{c\in R}\bigr).(6)✦ The rolled laws are conditional for RR if they satisfy (2) with μ\mu, uRu^R and κR\kappa^R in place of mm, uu and κ\kappa, where κnR(q)\kappa^R_n(q) is the conditional law under μn\mu_n of an array yy given its parameter array uR(y)=qu^R(y)=q.✦

Let the rolled laws be conditional for RR, and let aa be a slot outside RR with R∪{a}≠IR\cup\{a\}\ne I. The rolled object MR,aM^{R,a} is the array object over mXm^X with states YI,parameters UI×YR,factor map y↦x(ya),parameter map uR,laws μn,\text{states } Y^I,\quad \text{parameters } U^I\times Y^R,\quad \text{factor map } y\mapsto x(y_a),\quad \text{parameter map } u^R,\quad \text{laws } \mu_n,(7)✦ and the coordinatewise automorphisms of its state and parameter spaces.

The Claim

Theorem 1. Let MM be a saturated array object over mXm^X, and let RR be a set of slots with |R|+2≤|I||R|+2\le|I|. Then the rolled laws are conditional for RR: for every n≥0n\ge0 and every proper subset JJ of II, Law⁡μn+1((y|j)j∈J|uR(y)=p)=⨂j∈JκnR(p|j)for u*Rμn+1-almost every p.\operatorname{Law}_{\mu_{n+1}}\bigl((y|_j)_{j\in J}\mathrel{\big|}u^R(y)=p\bigr) =\bigotimes_{j\in J}\kappa^R_n\bigl(p|_j\bigr) \quad\text{for $u^R_*\mu_{n+1}$-almost every $p$.}(8)✦ For every slot a∉Ra\notin R, the rolled object is an array object over mXm^X, MR,a∈𝑶𝒃𝒋(mX),M^{R,a}\in\mathbf{Obj}(m^X),(9)✦ and the coordinate maps form an extension of MM by it, the coordinate extension: σa:YI→Y,σa(y)=ya,ρa:UI×YR→U,ρa(p,z)=pa.\sigma_a\colon Y^I\to Y,\quad \sigma_a(y)=y_a,\qquad \rho_a\colon U^I\times Y^R\to U,\quad \rho_a(p,z)=p_a .(10)✦

In the paper, with k=|I|k=|I|, the result reads: for R⊂IR\subset I with |R|≤k−2|R|\le k-2 and a∉Ra\notin R, the saturated object MM has an extension with state space YIY^I, law m1m_1 at order zero, parameter ((u(yc))c∈I,(yc)c∈R)\bigl((u(y_c))_{c\in I},(y_c)_{c\in R}\bigr), factor map y↦yay\mapsto y_a, law mn+1m_{n+1} at order nn, and automorphisms acting coordinatewise.

What Is Checked

The claim is stated over saturated array objects (4) over initial array laws of a measurable automorphism of a standard Borel space, and over the roll (5), the rolled laws μn\mu_n, the rolled parameter map (6) and the rolled object (7).

For every array object MM and every slot aa, the library also proves that the rolled laws are invariant under TYT_Y in every coordinate✦, consistent✦ and proper as in (1)✦, and carried to mnXm^X_n by y↦x(ya)y\mapsto x(y_a)✦ and to mnm_n by σa\sigma_a, both applied entrywise.✦ The rolled parameter map uRu^R is measurable✦ and intertwines the coordinatewise automorphisms of YIY^I and of UI×YRU^I\times Y^R.✦ Under μ0\mu_0 the maps y↦x(ya)y\mapsto x(y_a) and uRu^R are independent when a∉Ra\notin R and R∪{a}≠IR\cup\{a\}\ne I✦, and R∪{a}≠IR\cup\{a\}\ne I holds whenever a∉Ra\notin R and |R|+2≤|I||R|+2\le|I|.✦