OctoInform Beta

OpenAI preprint, September 2026 · §5. Measurable labels and finite blocks · §5

Lemma 5.1 [Active-space exactness]: In a saturated array object on at least three slots, at almost every parameter tuple, with an abstract statement for commuting random projections

Active spaces as closed spans of exact lines.

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

Let TT be a measurable automorphism of a standard Borel space XX and let II be a finite set of slots. The library works with laws mnm_n on arrays indexed by InI^n whose entries are points of a second standard Borel space YY, each point lying over a point of XX and carrying a parameter in a space UU. At order one an array is a tuple y=(yj)j∈Iy = (y_j)_{j \in I} with a parameter tuple p=(pj)j∈Ip = (p_j)_{j \in I}. Given pp, which circle-valued functions of one coordinate yiy_i are almost surely a constant phase times a product of circle-valued functions of the other coordinates? Each such function spans an exact line at slot ii, and the question is whether, for a saturated array object with at least three slots and for almost every pp, the closed linear span of the exact lines is the active space of functions of yiy_i. The library also states an abstract equality for random subspaces of an inner product space whose orthogonal projections commute.

Arrays and Array Objects

An array of order nn over a space YY is a function from the index tuples InI^n to YY, so that the arrays of order nn form the product YInY^{I^n}✦; an array of order zero is a single point of YY✦. For an axis a∈{0,…,n}a \in \{0, \dots, n\} and a slot ii, the plane restriction pla,i:YIn+1→YIn\mathrm{pl}_{a,i} \colon Y^{I^{n+1}} \to Y^{I^n} fixes axis aa at ii: (pla,iy)(v0,…,vn−1)=y(v0,…,va−1,i,va,…,vn−1),(\mathrm{pl}_{a,i}\, y)(v_0, \dots, v_{n-1}) = y(v_0, \dots, v_{a-1}, i, v_a, \dots, v_{n-1}),(1) For an array y∈YIn+1y \in Y^{I^{n+1}} and a subset J⊆IJ \subseteq I, the last-axis planes of yy along JJ are the family (pln,jy)j∈J(\mathrm{pl}_{n,j}\, y)_{j \in J} of arrays of order nn. A map f:Y→Zf \colon Y \to Z acts on arrays coordinatewise: f×In:YIn→ZInf^{\times I^n} \colon Y^{I^n} \to Z^{I^n} sends yy to f∘yf \circ y.✦

Initial array laws of a measurable automorphism TT of a nonempty standard Borel space XX are probability measures mnXm^X_n on XInX^{I^n}, n≥0n \ge 0, invariant under T×InT^{\times I^n}, consistent and proper: for every axis aa, every slot ii and every proper subset JJ of II, (pla,i)*mn+1X=mnX,((pln,j)j∈J)*mn+1X=⨂j∈JmnX.(\mathrm{pl}_{a,i})_* m^X_{n+1} = m^X_n, \qquad \bigl((\mathrm{pl}_{n,j})_{j \in J}\bigr)_* m^X_{n+1} = \bigotimes_{j \in J} m^X_n .(2)✦

An array object MM over initial array laws 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:Y→Xx \colon Y \to X and par:Y→U\mathrm{par} \colon Y \to U with x∘TY=T∘xx \circ T_Y = T \circ x and par∘TY=TU∘par\mathrm{par} \circ T_Y = T_U \circ \mathrm{par}, and probability measures mnm_n on YInY^{I^n} that are invariant under TY×InT_Y^{\times I^n}, carried by x×Inx^{\times I^n} to mnXm^X_n, and consistent and proper as in (2).✦ The parameter law πn\pi_n of MM is the image of mnm_n under par×In\mathrm{par}^{\times I^n}✦, and the own kernel κn\kappa_n of MM is the Markov kernel from UInU^{I^n} to YInY^{I^n} giving the conditional law under mnm_n of an array given its parameter array✦. An array object is moreover conditionally local and locally independent: for every nn and every proper subset JJ of II, the last-axis planes along JJ of an array of order n+1n + 1 drawn from mn+1m_{n+1} have, given its parameter array pp, the conditional law ⨂j∈Jκn(pln,jp)for πn+1-almost every p,\bigotimes_{j \in J} \kappa_n(\mathrm{pl}_{n,j}\, p) \quad \text{for $\pi_{n+1}$-almost every } p,(3)✦ and under m0m_0, read on YY, the maps xx and par\mathrm{par} are independent.

Extensions and Saturation

For measure-preserving maps ϕ\phi from (Ω,η)(\Omega, \eta) to (Z,α)(Z, \alpha) and ψ\psi from (Ω,η)(\Omega, \eta) to (W,β)(W, \beta), the pullback isometry ιϕ:L2(α)→L2(η)\iota_\phi \colon L^2(\alpha) \to L^2(\eta) of complex L2L^2 spaces is the map sending each F∈L2(α)F \in L^2(\alpha) to F∘ϕF \circ \phi, and the joining operator Jϕ,ψ=ιϕ*ιψ:L2(β)→L2(α)J_{\phi,\psi} = \iota_\phi^* \iota_\psi \colon L^2(\beta) \to L^2(\alpha) sends each F∈L2(β)F \in L^2(\beta) to the conditional expectation of F∘ψF \circ \psi given ϕ\phi, read on ZZ.

For an array object MM we write YMY_M and UMU_M for its spaces of states and parameters, xMx_M and parM\mathrm{par}_M for its maps, and mnMm^M_n for its laws mnm_n. An extension of an array object MM by an array object NN over the same initial array laws consists of measurable maps σ:YN→YM\sigma \colon Y_N \to Y_M of states and ρ:UN→UM\rho \colon U_N \to U_M of parameters that intertwine the automorphisms of NN with those of MM, with xM∘σ=xNx_M \circ \sigma = x_N and parM∘σ=ρ∘parN\mathrm{par}_M \circ \sigma = \rho \circ \mathrm{par}_N, such that σ×In\sigma^{\times I^n} carries mnNm^N_n to mnMm^M_n for every nn. The pullback of an extension (σ,ρ)(\sigma, \rho) of MM by NN at order nn is the isometry ισ×In:L2(mnM)→L2(mnN)\iota_{\sigma^{\times I^n}} \colon L^2(m^M_n) \to L^2(m^N_n).✦

There are 2|I|+12|I| + 1 data kinds: the parameter kind and, for every slot jj, the outside kind and the enlarged kind at jj.✦ On arrays of order n+1n + 1 an array object MM has a data field of each kind, a sub-σ\sigma-algebra of the Borel sets of YIn+1Y^{I^{n+1}}: the parameter field 𝒫n+1\mathcal{P}_{n+1}, the outside field 𝒪n+1,j\mathcal{O}_{n+1,j} and the enlarged field 𝒢n+1,j\mathcal{G}_{n+1,j}.✦ The projection En,dME^M_{n,d} onto the data field of kind dd is the conditional expectation onto that field, the orthogonal projection of L2(mn+1)L^2(m_{n+1}) onto the functions measurable for it.✦

An array object MM is saturated when its projections onto data commute with pullback along every extension: for every array object NN over the same initial array laws, every extension of MM by NN with state map σ\sigma, every order nn and every data kind dd, En,dNισ×In+1=ισ×In+1En,dMon L2(mn+1M).E^N_{n,d}\, \iota_{\sigma^{\times I^{n+1}}} = \iota_{\sigma^{\times I^{n+1}}}\, E^M_{n,d} \quad \text{on } L^2(m^M_{n+1}).(4)✦

Slots, Local Laws and Exact Lines

The coordinate yjy_j of an array y∈YIy \in Y^I of order one is its plane pl0,jy\mathrm{pl}_{0,j}\, y, a point of YY, and likewise pjp_j is the coordinate of a parameter tuple p∈UIp \in U^I✦. Fix a slot ii. The cell at ii is a copy of YY holding the coordinate yiy_i, so that functions on the cell are functions of yiy_i.✦ The local law λi\lambda_i at slot ii is the Markov kernel from UU to the cell that sends a parameter uu to κ0(u)\kappa_0(u), the conditional law under m0m_0 of a point of YY given that its parameter is uu.✦

For a measure ν\nu on the cell at ii and a parameter tuple pp, the active space 𝒜i(ν,p)\mathcal{A}_i(\nu, p) is the subspace of L2(ν)L^2(\nu) formed by the functions of the cell that are active under the parameters pp.✦

Let 𝕋⊂ℂ\mathbb{T} \subset \mathbb{C} be the unit circle, with the Borel sets it inherits from ℂ\mathbb{C}, and let ν\nu be a finite measure on the cell at ii. A measurable function ff from the cell to 𝕋\mathbb{T} is exact at p∈UIp \in U^I when there are a constant c∈𝕋c \in \mathbb{T} and measurable functions aj:Y→𝕋a_j \colon Y \to \mathbb{T} for the slots j≠ij \ne i with f(yi)=c∏j≠iaj(yj)for κ1(p)-almost every y∈YI.f(y_i) = c \prod_{j \ne i} a_j(y_j) \quad \text{for $\kappa_1(p)$-almost every } y \in Y^I .(5)✦ The exact lines at pp are the complex lines ℂf\mathbb{C} f in L2(ν)L^2(\nu) through the exact functions ff; their union is the exact set ℰi(ν,p)\mathcal{E}_i(\nu, p)✦, and its closed linear span is the exact span 𝒞i(ν,p)\mathcal{C}_i(\nu, p)✦.

A complement split of a nonempty set P⊆IP \subseteq I of slots is a pair of nonempty sets A,B⊆IA, B \subseteq I such that PP, AA and BB are pairwise disjoint with union II.✦

The Claim

Theorem 1. Let μ\mu be an s-finite measure on a measurable space Θ\Theta and HH a complex inner product space. For each θ∈Θ\theta \in \Theta let C(θ)⊆V(θ)C(\theta) \subseteq V(\theta) be subspaces of HH, with C(θ)C(\theta) closed and V(θ)V(\theta) admitting an orthogonal projection QθQ_\theta. Suppose that for μ⊗μ\mu \otimes \mu-almost every pair (θ,θ′)(\theta, \theta') the projections QθQ_\theta and Qθ′Q_{\theta'} commute and V(θ)∩V(θ′)⊆C(θ)V(\theta) \cap V(\theta') \subseteq C(\theta). Then V(θ)=C(θ)for μ-almost every θ.V(\theta) = C(\theta) \quad \text{for $\mu$-almost every } \theta .(6)✦

Theorem 2. Let MM be a saturated array object whose set of slots II has at least three elements, let ii be a slot, and let (A,B)(A, B) be a complement split of {i}\{i\}. Then for π1\pi_1-almost every parameter tuple p∈UIp \in U^I, with ν=λi(pi)\nu = \lambda_i(p_i) the local law at the coordinate pip_i, 𝒜i(ν,p)=𝒞i(ν,p).\mathcal{A}_i(\nu, p) = \mathcal{C}_i(\nu, p) .(7)✦

In the paper, the result reads: for π\pi-almost every tuple pp and every slot ii, RanΠi(p)=Ci(p)\mathrm{Ran}\, \Pi_i(p) = C_i(p), where the paper’s RanΠi(p)\mathrm{Ran}\, \Pi_i(p) is the active space 𝒜i\mathcal{A}_i, its Ci(p)C_i(p) the exact span 𝒞i\mathcal{C}_i, and its π\pi the parameter law π1\pi_1.

What Is Checked

The claim (7) is stated over a saturated array object with at least three slots, a slot ii with a complement split of {i}\{i\}, the parameter law π1\pi_1, the local law λi\lambda_i, the active space 𝒜i\mathcal{A}_i and the exact span 𝒞i\mathcal{C}_i of the exact lines (5); the claim (6) is stated over an s-finite measure and two families of subspaces of a complex inner product space.

The library proves of these constructions that a measurable function with values in the unit circle is square integrable for a finite measure✦, and that the state map of an extension, applied coordinatewise, is measure preserving from mnNm^N_n to mnMm^M_n for every order nn✦.