OctoInform Beta

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

Lemma 5.3 [Counting lift]: Over a saturated array object with |I|≥3|I| \ge 3, measures counting labels, label tuples and label squares, with product conditional laws and Mst=Mab*MstMabM_{st} = M_{ab}^* M_{st} M_{ab}

Parameter laws lifted to counting measures on labels.

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 standard Borel space XX with a measurable automorphism TT, a finite index set II with k=|I|≥3k = |I| \ge 3 elements, and an array object MM over initial array laws of TT, that is, a system of laws of arrays indexed by II on a space YY of states, with a space UU of parameters. Over a parameter uu, a line is a circle-valued function on the states with parameter uu, up to a constant phase; over an array of level 11, a point of YIY^I, a tuple of kk lines, one at each entry, is exact when the product of its functions is almost surely constant. The question is how the laws of the parameter arrays of MM lift to measures on labels of lines, on exact tuples of labels and on square arrays of labels whose rows and columns are exact, and how a label moves from one entry of an exact tuple, that is, one index in II, to another.

An array of level nn over a space YY is a point of YInY^{I^n}, an entry for each index tuple of length nn in II✦, and an array of level 00 is a point of YY✦; an axis of an array of level n+1n+1 is a position aa in its index tuples, a slot is an element ii of II, and the plane restriction fixing axis aa at slot ii takes an array of level n+1n+1 to the array of level nn of its entries with ii inserted at position aa.✦ A family of probability measures mnm_n on YInY^{I^n} is consistent if every plane restriction carries mn+1m_{n+1} to mnm_n, and proper if, for every proper subset JJ of II, the last-axis planes at the slots in JJ, the plane restrictions fixing the last axis at those slots, are independent under mn+1m_{n+1}, each with law mnm_n. Initial array laws mXm^X of TT are consistent and proper laws mnXm^X_n on XInX^{I^n}, invariant under TT applied entrywise.✦

An array object MM over mXm^X consists of standard Borel spaces YY of states and UU of parameters, each with an automorphism, a factor map from YY to XX and a parameter map par:Y→U\mathrm{par}\colon Y \to U intertwining them, and invariant, consistent and proper laws mnm_n on YInY^{I^n} carried to mnXm^X_n by the factor map applied entrywise; it is moreover conditionally local, the last-axis planes at the slots of a proper subset of II staying independent given the parameter array, each with the conditional law of an array of level nn given its parameters, and locally independent, the factor and parameter maps being independent under m0m_0.✦

Applying par\mathrm{par} entrywise carries mnm_n to the parameter law mnUm^U_n on UInU^{I^n}✦, and the own kernel κn\kappa_n is the conditional law under mnm_n of an array of level nn given its parameter array.✦ An array y∈YIy \in Y^I of level 11 is the tuple of its kk entries yjy_j✦, its parameter array p∈UIp \in U^I the tuple of their parameters pjp_j✦, and the raw tuple kernel of MM is the kernel κ1\kappa_1 from UIU^I to YIY^I that sends a parameter array p∈UIp \in U^I to κ1(p)\kappa_1(p), the conditional law of the tuple of entries given pp. For a slot aa, the coordinate kernel ρa\rho_a sends u∈Uu \in U to the conditional law under m1Um^U_1 of pp given pa=up_a = u.✦

For every level nn, an array object MM carries data fields on YIn+1Y^{I^{n+1}}: the parameter field 𝒫n+1\mathcal{P}_{n+1} and, for each index jj in II, the outside field 𝒪n+1,j\mathcal{O}_{n+1,j} and the enlarged field 𝒢n+1,j\mathcal{G}_{n+1,j}, each a sub-σ\sigma-algebra of the Borel sets. The projection onto a data field is the conditional expectation onto it, the orthogonal projection of L2(mn+1)L^2(m_{n+1}) onto the functions measurable for that field.✦

An extension of MM by an array object NN over the same initial laws is a pair of measurable maps, from the states of NN to those of MM and from the parameters of NN to those of MM, intertwining the automorphisms and compatible with the factor and parameter maps, the state map applied entrywise carrying each law of NN to the law of MM of the same level.✦ The array object MM is saturated if, for every extension of MM by an array object NN, every level nn and every data field 𝒫n+1\mathcal{P}_{n+1}, 𝒪n+1,j\mathcal{O}_{n+1,j} or 𝒢n+1,j\mathcal{G}_{n+1,j} on YIn+1Y^{I^{n+1}}, the pullback isometry F↦F∘en+1F \mapsto F \circ e_{n+1} from L2(mn+1)L^2(m_{n+1}) to L2L^2 of the law of NN, with en+1e_{n+1} the state map applied entrywise, intertwines the orthogonal projections onto the functions measurable for that field.✦

Labels and Exact Tuples

Labels are built from codes of vectors. For a finite measure ν\nu on YY, the Hilbert test sequence (hn)(h_n) of ν\nu runs, in a fixed enumeration, through the simple functions ∑j(αj+iβj)𝟏Aj\sum_j (\alpha_j + \mathrm{i}\beta_j)\mathbf{1}_{A_j} with rational αj,βj\alpha_j, \beta_j and sets AjA_j in the algebra generated by a fixed countable generating family of the Borel sets of YY, taken as vectors of L2(ν)L^2(\nu).✦ A Hilbert code for κ0\kappa_0 is a pair q=(u,z)q = (u, z) of a base point u∈Uu \in U and a distance code zz for the test sequence of κ0(u)\kappa_0(u): a nonnegative real sequence with zn≤∥hn−hm∥+zmz_n \le \|h_n - h_m\| + z_m and ∥hn−hm∥≤zn+zm\|h_n - h_m\| \le z_n + z_m for all nn and mm, and with a term below 2−m2^{-m} for every mm.✦ A Hilbert code q=(u,z)q = (u, z) for κ0\kappa_0 decodes to a vector of L2(κ0(u))L^2(\kappa_0(u)) at distance znz_n from hnh_n for every nn✦, and the representative fqf_q of qq is the pointwise limit, as m→∞m \to \infty, of the simple function hnh_n at the first index nn with zn<2−mz_n < 2^{-m}✦. The encoding of a vector vv of L2(κ0(u))L^2(\kappa_0(u)) is the code (u,(∥v−hn∥)n)(u, (\|v - h_n\|)_n).✦

A Hilbert code q=(u,z)q = (u, z) is canonical if the first test vector not orthogonal to its decoded vector has a positive real inner product with it✦, and unimodular if |fq|=1|f_q| = 1 almost everywhere for κ0(u)\kappa_0(u).✦ Write 𝒞\mathcal{C} for the space of Hilbert codes for κ0\kappa_0. A label is a pair (u,z)(u, z) of a parameter u∈Uu \in U and a code z∈𝒞z \in \mathcal{C}, a point of Z=U×𝒞Z = U \times \mathcal{C}.✦ A label tuple is a pair λ=(p,w)\lambda = (p, w) of a parameter array p∈UIp \in U^I of level 11 and a tuple w∈𝒞Iw \in \mathcal{C}^I of codes, with label λi=(pi,wi)\lambda_i = (p_i, w_i) at slot ii; label tuples thus form ZIZ^I.✦ A label square is a pair λ=(e,q)\lambda = (e, q) of a parameter array e∈UI×Ie \in U^{I \times I} of level 22 and a matrix q∈𝒞I×Iq \in \mathcal{C}^{I \times I} of codes, with labels λas=(eas,qas)\lambda_{as} = (e_{as}, q_{as}), and its row λa⋅\lambda_{a\cdot} and column λ⋅s\lambda_{\cdot s} are the label tuples obtained by fixing its first axis at aa or its last axis at ss; label squares form ZI×IZ^{I \times I}.✦

Fix a kernel Λ\Lambda from UIU^I to YIY^I. The exact code set ℰ\mathcal{E} is the set of label tuples (p,w)(p, w) whose codes wiw_i are canonical, unimodular and based at pip_i, and whose exact code mean has modulus one: |∫YI∏i∈Ifwi(yi)dΛ(p)(y)|=1.\Big| \int_{Y^I} \prod_{i \in I} f_{w_i}(y_i) \, d\Lambda(p)(y) \Big| = 1 .(1)✦ The members of the exact code set ℰ\mathcal{E} are called exact code tuples. The line graph GiG_i at a slot ii is the set of pairs (p,z)(p, z) of a parameter array p∈UIp \in U^I and a code zz that is the entry wiw_i of some (p,w)∈ℰ(p, w) \in \mathcal{E}✦, and the exact square set 𝒮\mathcal{S} is the set of label squares whose rows and columns all lie in ℰ\mathcal{E}.

For a slot ii, a matching τi\tau_i assigns to each parameter array p∈UIp \in U^I and each code zz for which one exists an exact code tuple τi(p,z)\tau_i(p, z) over pp with entry zz at slot ii.✦ The label transition kernel KabK_{ab} on ZZ takes a label from slot aa to slot bb: at a label (u,z)(u, z) it is the law of the label at slot bb of the matched tuple over a parameter array pp drawn from ρa(u)\rho_a(u), Kab(u,z)=(p↦(pb,τa(p,z)b))*ρa(u).K_{ab}(u, z) = \big(p \mapsto (p_b, \tau_a(p, z)_b)\big)_* \, \rho_a(u) .(2)✦

Counting Measures and Joining Operators

A Borel fibre enumeration EE of a set S⊆V×WS \subseteq V \times W, for standard Borel spaces VV and WW, is a sequence of measurable sets Dn⊆VD_n \subseteq V and measurable maps gn:V→Wg_n\colon V \to W such that (v,gn(v))∈S(v, g_n(v)) \in S for v∈Dnv \in D_n, every point of SS is (v,gn(v))(v, g_n(v)) for some nn with v∈Dnv \in D_n, and gn(v)≠gm(v)g_n(v) \ne g_m(v) for v∈Dn∩Dmv \in D_n \cap D_m and n≠mn \ne m.✦ The counting measure νE\nu^E of a Borel fibre enumeration EE over a measure ν\nu on VV is the measure that is the image, under (v,n)↦(v,gn(v))(v, n) \mapsto (v, g_n(v)), of the product of ν\nu with counting measure on ℕ\mathbb{N} restricted to ΔE={(v,n):v∈Dn}\Delta_E = \{(v, n) : v \in D_n\}: νE=((v,n)↦(v,gn(v)))*((ν⊗#)|ΔE).\nu^E = \big((v, n) \mapsto (v, g_n(v))\big)_* \big((\nu \otimes \#)|_{\Delta_E}\big) .(3)✦

A set S⊆ZS \subseteq Z of labels matches the line graphs if, for m1Um^U_1-almost every pp and every slot ii, the codes zz with (p,z)∈Gi(p, z) \in G_i are those with (pi,z)∈S(p_i, z) \in S. For such a set and Borel fibre enumerations EE of SS, EπE^\pi of ℰ\mathcal{E} and EϑE^\vartheta of 𝒮\mathcal{S}, the counting measures (3) over the parameter laws are ξ=(m0U)E\xi = (m^U_0)^{E} on labels, π̃=(m1U)Eπ\tilde\pi = (m^U_1)^{E^\pi} on label tuples and ϑ̃=(m2U)Eϑ\tilde\vartheta = (m^U_2)^{E^\vartheta} on label squares.

For measure-preserving maps xx from (Ω,μ)(\Omega, \mu) to (V,ν)(V, \nu) and yy from (Ω,μ)(\Omega, \mu) to (W,η)(W, \eta), the pullback Pxf=f∘xP_x f = f \circ x is an isometry of L2(ν)L^2(\nu) into L2(μ)L^2(\mu)✦, and the joining operator of xx and yy is the bounded operator Ox,y=Px*Py:L2(η)→L2(ν),O_{x,y} = P_x^{*} P_y \colon L^2(\eta) \to L^2(\nu),(4)✦ which sends ff to the conditional expectation of f∘yf \circ y given xx, read on VV. When the slot maps λ↦λi\lambda \mapsto \lambda_i carry π̃\tilde\pi to ξ\xi, MabM_{ab} is the joining operator (4) on L2(ξ)L^2(\xi) of the slot maps at aa and at bb.

The Claim

Theorem 1 (Counting lift). Let MM be an array object whose index set II has at least three elements, and let Λ\Lambda be a kernel from UIU^I to YIY^I such that, for every p∈UIp \in U^I and every slot jj, the coordinate yjy_j has law κ0(pj)\kappa_0(p_j) under Λ(p)\Lambda(p) and the coordinates yty_t with t≠jt \ne j are independent.

  1. Suppose MM is saturated and Λ(p)\Lambda(p) is the raw tuple kernel at pp for m1Um^U_1-almost every pp. Then there are a measurable set SS of labels (u,z)(u, z), each with zz canonical, unimodular and based at uu, and Borel fibre enumerations EE of SS, EπE^\pi of ℰ\mathcal{E} and EϑE^\vartheta of 𝒮\mathcal{S}, such that SS matches the line graphs and, for all i,a,s∈Ii, a, s \in I, (λ↦λi)*π̃=ξ,(λ↦λa⋅)*ϑ̃=π̃,(λ↦λ⋅s)*ϑ̃=π̃.(\lambda \mapsto \lambda_i)_* \tilde\pi = \xi, \qquad (\lambda \mapsto \lambda_{a\cdot})_* \tilde\vartheta = \tilde\pi, \qquad (\lambda \mapsto \lambda_{\cdot s})_* \tilde\vartheta = \tilde\pi .(5)✦

  2. Let SS be a set of labels matching the line graphs, and let EE and EπE^\pi be Borel fibre enumerations of SS and of ℰ\mathcal{E}. For every slot aa, the parameter array pp of a label tuple λ\lambda has, under π̃\tilde\pi, the conditional law given its label λa=(u,z)\lambda_a = (u, z) at aa π̃(p∈⋅|λa=(u,z))=ρa(u).\tilde\pi\big(p \in \cdot \,\big|\, \lambda_a = (u, z)\big) = \rho_a(u) .(6)✦

  3. If moreover MM is saturated, Λ(p)\Lambda(p) is the raw tuple kernel at pp for m1Um^U_1-almost every pp and EϑE^\vartheta is a Borel fibre enumeration of 𝒮\mathcal{S}, then for all slots aa and bb and all slots s≠ts \ne t, the pair of labels (λbs,λbt)(\lambda_{bs}, \lambda_{bt}) of a label square has, under ϑ̃\tilde\vartheta, the conditional law given the row λa⋅\lambda_{a\cdot} ϑ̃((λbs,λbt)∈⋅|λa⋅)=Kab(λas)⊗Kab(λat),\tilde\vartheta\big((\lambda_{bs}, \lambda_{bt}) \in \cdot \,\big|\, \lambda_{a\cdot}\big) = K_{ab}(\lambda_{as}) \otimes K_{ab}(\lambda_{at}),(7)✦ with KabK_{ab} the transition kernel (2), and the joining operators of the slots satisfy Mst=Mab*MstMab.M_{st} = M_{ab}^{*} M_{st} M_{ab} .(8)✦

  4. Let (Ω,μ)(\Omega, \mu), (P,π)(P, \pi) and (W,η)(W, \eta) be σ\sigma-finite measure spaces, with maps ra,ca:Ω→Pr_a, c_a\colon \Omega \to P carrying μ\mu to π\pi and σa:P→W\sigma_a\colon P \to W carrying π\pi to η\eta for a∈Ia \in I, such that σs∘ra=σa∘cs\sigma_s \circ r_a = \sigma_a \circ c_s for all a,s∈Ia, s \in I; let KK be a Markov kernel on WW and a,b,s,t∈Ia, b, s, t \in I. If, under μ\mu, the pair (σs∘rb,σt∘rb)(\sigma_s \circ r_b, \sigma_t \circ r_b) has conditional law K(σs(ra))⊗K(σt(ra))K(\sigma_s(r_a)) \otimes K(\sigma_t(r_a)) given rar_a, then Oσs,σt=Oσa,σb*Oσs,σtOσa,σb.O_{\sigma_s,\sigma_t} = O_{\sigma_a,\sigma_b}^{*}\, O_{\sigma_s,\sigma_t}\, O_{\sigma_a,\sigma_b} .(9)✦

In the paper, the result reads: there are σ\sigma-finite measures π̃\tilde\pi on ZIZ^I and ϑ̃\tilde\vartheta on ZI×IZ^{I \times I} counting matched label tuples and arrays, with marginals ξ\xi and π̃\tilde\pi, and transition kernels KstK_{st} whose operators MstM_{st} are contractions with adjoint MtsM_{ts} commuting with time, satisfying Mst=Mab*MstMabM_{st} = M_{ab}^{*} M_{st} M_{ab} for a≠ba \ne b, and proper families of labelled columns have a product conditional law given a labelled row.

What Is Checked

The claim is stated over array objects over initial array laws, with their extensions and saturation; the parameter laws mnUm^U_n, the own kernels κ0\kappa_0 and κ1\kappa_1, the raw tuple kernel and the coordinate kernels ρa\rho_a; Hilbert codes for κ0\kappa_0, their representatives and the canonical and unimodular codes; the exact code set ℰ\mathcal{E} of (1), its line graphs and the exact square set 𝒮\mathcal{S}; Borel fibre enumerations and their counting measures (3); the transition kernels KabK_{ab} of (2); and the joining operators (4).

About these constructions the library proves that the slot maps λ↦λi\lambda \mapsto \lambda_i on label tuples are measurable✦ and that, under the hypotheses of (6), for every set of labels matching the line graphs and all Borel fibre enumerations EE of it and EπE^\pi of ℰ\mathcal{E}, each slot map carries π̃\tilde\pi to ξ\xi.✦