OctoInform Beta

OpenAI preprint, September 2026 · §4. The operator square · §4

Definition 4.1 [Active space, exact tuples and lines]: Active spaces as closed ranges of Markov operators of joinings, and exact lines as circle-valued tuples with almost surely constant product modulo phases

Active spaces and exact lines of a tuple law.

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 (Yi)i∈I(Y_i)_{i\in I} be random variables with values in a standard Borel space YY, indexed by a finite set II, with joint law λ\lambda on YIY^I. The item concerns two kinds of dependence among them. One is linear: which square-integrable functions of a single coordinate YiY_i the other coordinates reach in the mean, the closed subspace of such functions being the active space at ii. The other is multiplicative: which circle-valued functions of the coordinates multiply to an almost surely constant product, such a tuple of functions being an exact tuple, and the tuple with each function taken up to a constant phase an exact line.

The linear question is also asked of arrays: for an array object, an array of order n+2n+2 is a square of cells, and the cells are grouped into blocks of rows and columns, all defined below. The joinings are then conditional laws given a block of cells and the parameter array, which records the parameters of the entries of the array, so that a block carries a measurable family of Markov operators rather than a single one. The library defines the active space of the conditional law of an array of order one at a slot, that is, at an index i∈Ii\in I, the row, column and common active spaces of a block of cells, and the exact tuples and exact lines of a law on YIY^I. Square-integrable functions are complex-valued throughout.

Arrays and Array Objects

An array of order nn over YY assigns an entry in YY to each index tuple in InI^n; the arrays form the product YInY^{I^n}✦, and an array of order zero is a point of YY✦. The plane restriction of an array yy of order n+1n+1 that fixes the axis aa at the index ii 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.✦

Initial array laws of a measurable automorphism TT of a standard Borel space XX are probability measures mnXm^X_n on XInX^{I^n}, one for each nn, each invariant under TT applied entrywise, consistent, in that every plane restriction carries mn+1Xm^X_{n+1} to mnXm^X_n, and proper: for every subset JJ of II other than II, the planes of an array of law mn+1Xm^X_{n+1} that fix the last axis at the indices in JJ are independent, each of law mnXm^X_n.✦

An array object MM over the initial array laws mnXm^X_n of TT consists of a standard Borel space YY with an automorphism TYT_Y, a standard Borel space UU of parameters with an automorphism TUT_U, measurable maps ξ:Y→X\xi\colon Y\to X and par:Y→U\mathrm{par}\colon Y\to U, with ξ\xi intertwining TYT_Y with TT and par\mathrm{par} intertwining TYT_Y with TUT_U, and probability measures mnm_n on YInY^{I^n}, one for each nn.✦ The parameter array par(y)\mathrm{par}(y) of an array yy is par\mathrm{par} applied entrywise. Each mnm_n is invariant under TYT_Y applied entrywise and is carried to mnXm^X_n by ξ\xi applied entrywise, and every plane restriction carries mn+1m_{n+1} to mnm_n. For every subset JJ of II other than II, the planes of an array of law mn+1m_{n+1} that fix the last axis at the indices in JJ are independent, each of law mnm_n, and they stay independent given the parameter array, each with the conditional law of an array of order nn given the matching plane of the parameter array; under m0m_0, a law on YY, the maps ξ\xi and par\mathrm{par} are independent.

For a parameter array u∈UInu\in U^{I^n}, mn(⋅∣u)m_n(\cdot\mid u) is the conditional law under mnm_n of an array given that its parameter array is uu, a Markov kernel from UInU^{I^n} to YInY^{I^n}.✦

The Grouped Square

An array yy of order n+2n+2 is a square of cells. The row ya•y_{a\bullet} of yy at a∈Ia\in I is the plane restriction of yy fixing the first axis at aa, and the column y•jy_{\bullet j} of yy at j∈Ij\in I is the plane restriction of yy fixing the last axis at jj; both are arrays of order n+1n+1. The cell yaj∈YIny_{aj}\in Y^{I^n} of yy at (a,j)(a,j) is the plane restriction of yy fixing the first axis at aa and the last axis at jj, so that its entries are yaj(v)=y(a,v,j)y_{aj}(v)=y(a,v,j). For a set AA of rows and a set SS of columns, yA×S=(yaj)a∈A,j∈Sy_{A\times S}=(y_{aj})_{a\in A,\,j\in S} is the block of cells on A×SA\times S✦, and the grouped data of yy on A×SA\times S is the pair (par(y),yA×S)(\mathrm{par}(y),y_{A\times S})✦.

For a parameter array e∈UIn+2e\in U^{I^{n+2}}, the row family kernel draws arrays waw_a of order n+1n+1, one for each a∈Aa\in A, independently, waw_a from mn+1(⋅∣ea•)m_{n+1}(\cdot\mid e_{a\bullet}).✦ The law of cells λA×Se\lambda^e_{A\times S} is the law of the block on A×SA\times S whose cell at (a,j)(a,j) is the plane of waw_a that fixes the last axis at jj.✦

For a set AA of rows and sets RR, SS, DD of columns, the row joining at grouped data (e,x)(e,x), with xx a block on A×SA\times S, is the row pair kernel there: the conditional law under mn+2m_{n+2} of the pair of blocks (yA×R,yA×D)(y_{A\times R},y_{A\times D}) given that par(y)=e\mathrm{par}(y)=e and yA×S=xy_{A\times S}=x.✦

The parent rows of an array ww of order n+1n+1 on a set AA are its rows wa•w_{a\bullet}, a∈Aa\in A✦, and the posterior kernel sends a parameter array uu and parents zz to the conditional law of ww under mn+1(⋅∣u)m_{n+1}(\cdot\mid u) given that its parent rows on AA are zz✦. For a set RR of columns and sets PP, AA, BB of rows, the column group kernel at grouped data (e,z)(e,z), with zz a block on A×RA\times R, draws arrays wjw_j of order n+1n+1, one for each j∈Rj\in R, independently, wjw_j from the posterior given the parameter column e•je_{\bullet j} and the parents (zaj)a∈A(z_{aj})_{a\in A}.✦ The column joining at (e,z)(e,z), also called the column pair kernel, is the law of the pair of blocks on P×RP\times R and on B×RB\times R whose cell at (b,j)(b,j) is the row of wjw_j at bb.✦

The Definition

Definition 1. Let s:Ω→Zs\colon\Omega\to Z and t:Ω→Wt\colon\Omega\to W be measure-preserving maps from (Ω,μ)(\Omega,\mu) to (Z,ν)(Z,\nu) and to (W,ρ)(W,\rho), and let Us:L2(ν)→L2(μ)U_s\colon L^2(\nu)\to L^2(\mu), h↦h∘sh\mapsto h\circ s, be the pullback isometry of ss.✦ The joining operator of ss and tt, the Markov operator of the joining (s,t)*μ(s,t)_*\mu of ν\nu and ρ\rho, is Js,t=Us*Ut:L2(ρ)→L2(ν).J_{s,t}=U_s^{*}\,U_t\colon L^2(\rho)\to L^2(\nu).(1)✦ For a kernel κ\kappa from a space of parameters to Z×WZ\times W, the family of joining operators assigns to each parameter uu the Markov operator from L2(ρ)L^2(\rho) to L2(ν)L^2(\nu) of the joining κ(u)\kappa(u) of ν\nu and ρ\rho.✦

Let η\eta be a measure on a space VV and let EE and HH be Hilbert spaces. The essential kernel of a family (G(q))q∈V(G(q))_{q\in V} of bounded operators from EE to HH is the subspace of the vectors e∈Ee\in E with G(q)e=0G(q)e=0 for η\eta-almost every qq.✦ The essential range of a family (F(q))q∈V(F(q))_{q\in V} of bounded operators from EE to HH is the orthogonal complement of the essential kernel of the adjoints: ess ran⁡ηF={h∈H:F(q)*h=0for η-almost every q}⟂.\operatorname{ess\,ran}_\eta F=\bigl\{h\in H : F(q)^{*}h=0\ \text{for $\eta$-almost every } q\bigr\}^{\perp}.(2)✦

For a parameter tuple p∈UIp\in U^I, let λp=m1(⋅∣p)\lambda_p=m_1(\cdot\mid p), a law on YIY^I, and let πi(y)=yi\pi_i(y)=y_i and π≠i(y)=(yj)j≠i\pi_{\neq i}(y)=(y_j)_{j\neq i}. The raw active space at the slot ii is 𝒜i∘(p)=ran⁡Jπi,π≠i¯⊆L2(πi*λp).\mathcal A^{\circ}_i(p)=\overline{\operatorname{ran}J_{\pi_i,\pi_{\neq i}}}\subseteq L^2(\pi_{i*}\lambda_p).(3)✦ The slot active space 𝒜i(p)\mathcal A_i(p) of λp\lambda_p at ii is the closure of the range of the joining operator of σi\sigma_i and π≠i\pi_{\neq i}, where σi(y)\sigma_i(y) is the block on {i}×{i}\{i\}\times\{i\} whose single cell is yiy_i: 𝒜i(p)=ran⁡Jσi,π≠i¯⊆L2(σi*λp).\mathcal A_i(p)=\overline{\operatorname{ran}J_{\sigma_i,\pi_{\neq i}}}\subseteq L^2(\sigma_{i*}\lambda_p).(4)✦

Fix e∈UIn+2e\in U^{I^{n+2}}. For a set AA of rows and sets RR, SS, DD of columns, the row joining operator at grouped data on A×SA\times S is the Markov operator from L2(λA×De)L^2(\lambda^e_{A\times D}) to L2(λA×Re)L^2(\lambda^e_{A\times R}) of the row joining there✦; at (e,x)(e,x) it is the fibre row operator 𝖱e(x)\mathsf R^e(x)✦, and the row active space is 𝒜Arow(R,S,D;e)=ess ran⁡λA×Se𝖱e⊆L2(λA×Re).\mathcal A^{\mathrm{row}}_A(R,S,D;e)=\operatorname{ess\,ran}_{\lambda^e_{A\times S}} \mathsf R^e\subseteq L^2(\lambda^e_{A\times R}).(5)✦ For a set RR of columns and sets PP, AA, BB of rows, the column joining operator at grouped data on A×RA\times R is the Markov operator from L2(λB×Re)L^2(\lambda^e_{B\times R}) to L2(λP×Re)L^2(\lambda^e_{P\times R}) of the column joining there✦; at (e,z)(e,z) it is the fibre column operator 𝖢e(z)\mathsf C^e(z)✦, and the column active space is 𝒜Rcol(P,A,B;e)=ess ran⁡λA×Re𝖢e⊆L2(λP×Re).\mathcal A^{\mathrm{col}}_R(P,A,B;e)=\operatorname{ess\,ran}_{\lambda^e_{A\times R}} \mathsf C^e\subseteq L^2(\lambda^e_{P\times R}).(6)✦ For sets PP, AA, BB of rows and RR, SS, DD of columns, the common active space of the block P×RP\times R is 𝒜P×Rcom(e)=𝒜Prow(R,S,D;e)∩𝒜Rcol(P,A,B;e).\mathcal A^{\mathrm{com}}_{P\times R}(e)=\mathcal A^{\mathrm{row}}_P(R,S,D;e)\cap \mathcal A^{\mathrm{col}}_R(P,A,B;e).(7)✦

Let ℳ(Y,𝕋)\mathcal M(Y,\mathbb T) be the group of measurable maps from YY to the circle 𝕋\mathbb T under pointwise multiplication. For a law λ\lambda on YIY^I, the set of exact tuples is ℰ(λ)={(χi)i∈I∈ℳ(Y,𝕋)I:there is c∈𝕋 with ∏i∈Iχi(yi)=c for λ-almost every y}.\mathcal E(\lambda)=\Bigl\{(\chi_i)_{i\in I}\in\mathcal M(Y,\mathbb T)^I : \text{there is } c\in\mathbb T\text{ with }\prod_{i\in I}\chi_i(y_i)=c\text{ for $\lambda$-almost every } y\Bigr\}.(8)✦ For a law ν\nu on YY, the constant phase subgroup Φ(ν)\Phi(\nu) consists of the maps ν\nu-almost everywhere equal to a constant✦, and the local projective group is ℙ(ν)=ℳ(Y,𝕋)/Φ(ν)\mathbb P(\nu)=\mathcal M(Y,\mathbb T)/\Phi(\nu)✦, whose elements are the lines. For laws νi\nu_i on YY, i∈Ii\in I, with ν=(νi)i∈I\nu=(\nu_i)_{i\in I}, write [χ]νi[\chi]_{\nu_i} for the line of χ∈ℳ(Y,𝕋)\chi\in\mathcal M(Y,\mathbb T) in ℙ(νi)\mathbb P(\nu_i). The tuple projectivization is the map from ℳ(Y,𝕋)I\mathcal M(Y,\mathbb T)^I to ∏i∈Iℙ(νi)\prod_{i\in I}\mathbb P(\nu_i) that sends (χi)i∈I(\chi_i)_{i\in I} to ([χi]νi)i∈I([\chi_i]_{\nu_i})_{i\in I}.✦ The set ℒ(λ;ν)\mathcal L(\lambda;\nu) of exact lines of λ\lambda is the image of ℰ(λ)\mathcal E(\lambda) under it: ℒ(λ;ν)={([χi]νi)i∈I:(χi)i∈I∈ℰ(λ)}⊆∏i∈Iℙ(νi).\mathcal L(\lambda;\nu)=\bigl\{([\chi_i]_{\nu_i})_{i\in I} : (\chi_i)_{i\in I}\in \mathcal E(\lambda)\bigr\}\subseteq\prod_{i\in I}\mathbb P(\nu_i).(9)✦

In the paper, the result reads: for the law λ\lambda on S1×⋯×SmS_1\times\cdots\times S_m of a tuple (Y1,…,Ym)(Y_1,\dots,Y_m), with μi\mu_i the law of YiY_i, the active space at an index ii is the closed span of the conditional expectations 𝔼[F(Yj:j≠i)∣Yi]\mathbb E[F(Y_j : j\neq i)\mid Y_i] in L2(μi)L^2(\mu_i), where FF runs over the functions of the other coordinates, an exact tuple is a tuple of maps χi:Si→𝕋\chi_i\colon S_i\to\mathbb T with ∏iχi(Yi)=c\prod_i\chi_i(Y_i)=c almost surely for some constant cc, and an exact line is a line ℂχi\mathbb C\chi_i occurring in an exact tuple, up to constants of modulus one.

What Is Checked

The active spaces (3) to (7) are stated over an array object MM with a finite index set II, its laws mnm_n and its conditional laws mn(⋅∣u)m_n(\cdot\mid u) given parameter arrays, through the joining operators (1) and the essential ranges (2); the exact tuples (8) and the exact lines (9) are stated over a law λ\lambda on YIY^I and laws νi\nu_i on YY.

The library proves that the essential kernel is a linear subspace and that the measurable circle maps form a subgroup of all circle-valued maps on YY, with Φ(ν)\Phi(\nu) a subgroup of it. It proves that ℰ(λ)\mathcal E(\lambda) is a subgroup of ℳ(Y,𝕋)I\mathcal M(Y,\mathbb T)^I and that the tuple projectivization is a homomorphism, whose image ℒ(λ;ν)\mathcal L(\lambda;\nu) is a subgroup of ∏iℙ(νi)\prod_i\mathbb P(\nu_i).