OctoInform Beta

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

Lemma 4.4 [Active-space reduction]: For operator families satisfying both rectangle identities, and for the grouped square of a saturated array object, where the restrictions again satisfy both identities

Rectangle operators intertwine the projections onto active spaces.

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

A joining of two probability spaces defines a Markov operator between their L2L^2 spaces, sending a function of the second coordinate to its conditional expectation given the first. In Rokhlin’s multiple-mixing problem such operators come in families indexed by conditioning data, and four families connect four Hilbert spaces in a rectangle. The question is how each family acts on the essential ranges of the others and on their intersections at the corners, the spaces on which the rectangle is active.

The library treats an abstract rectangle of operator families over random parameters, and then a grouped square built from a random array, whose corners are L2L^2 spaces of blocks of cells and whose edges are joining operators of conditional laws of those blocks.

Operator Rectangles

Let ν\nu be a measure on a measurable space Θ\Theta, and let AA be a family of bounded linear operators A(t):E→HA(t)\colon E\to H between complex Hilbert spaces, indexed by the points tt of Θ\Theta. The family AA is strongly measurable if, for every vv in EE, the map t↦A(t)vt\mapsto A(t)v is strongly measurable.✦ The essential kernel Ker⁡νA\operatorname{Ker}_\nu A of AA is the subspace of EE of the vectors that A(t)A(t) annihilates for ν\nu-almost every tt✦, and the essential range Ran⁡νA\operatorname{Ran}_\nu A of AA is the closed subspace of HH orthogonal to the essential kernel of the adjoint family A*A^*: Ker⁡νA={v∈E:A(t)v=0 for ν-a.e. t},Ran⁡νA=(Ker⁡νA*)⟂.\operatorname{Ker}_\nu A=\{v\in E : A(t)v=0 \text{ for $\nu$-a.e. } t\},\qquad \operatorname{Ran}_\nu A=(\operatorname{Ker}_\nu A^*)^{\perp}.(1)✦

The family AA has essential total range if the only closed subspace of HH that contains the range of A(t)A(t) for ν\nu-almost every tt is HH itself: {L⊆H closed subspace:A(t)E⊆L for ν-a.e. t}={H}.\{L\subseteq H \text{ closed subspace} : A(t)E\subseteq L \text{ for $\nu$-a.e. } t\}=\{H\}.(2)✦

For a closed subspace LL of a Hilbert space, PLP^L denotes the orthogonal projection onto LL. The compression of a bounded operator A:E→HA\colon E\to H to closed subspaces E0⊆EE_0\subseteq E and H0⊆HH_0\subseteq H is the operator PH0A|E0:E0→H0P^{H_0}A|_{E_0}\colon E_0\to H_0.✦

An operator rectangle is given by random elements xx, yy, cc, dd of a probability space (Ω,μ)(\Omega,\mu), with laws μx\mu_x, μy\mu_y, μc\mu_c, μd\mu_d, such that each of the pairs (c,y)(c,y), (c,x)(c,x), (d,y)(d,y), (d,x)(d,x) is independent, and by strongly measurable families of bounded operators F(x):H2→H1F(x)\colon H_2\to H_1, G(y):H4→H3G(y)\colon H_4\to H_3, K(c):H3→H1K(c)\colon H_3\to H_1 and N(d):H4→H2N(d)\colon H_4\to H_2 between complex Hilbert spaces, indexed by the values of the four random elements and satisfying both rectangle identities almost surely: F(x)N(d)=K(c)G(y),F(x)*K(c)=N(d)G(y)*.F(x)\,N(d)=K(c)\,G(y),\qquad F(x)^*K(c)=N(d)\,G(y)^*.(3)

Essential ranges of these families and of their adjoint families are taken for the law of the indexing random element, Ran⁡F=Ran⁡μxF\operatorname{Ran}F=\operatorname{Ran}_{\mu_x}F and so on. The active intersections are the closed subspaces D1=Ran⁡F∩Ran⁡K⊆H1,D2=Ran⁡F*∩Ran⁡N⊆H2,D3=Ran⁡G∩Ran⁡K*⊆H3,D4=Ran⁡G*∩Ran⁡N*⊆H4,\begin{aligned} D_1&=\operatorname{Ran}F\cap\operatorname{Ran}K\subseteq H_1, &\qquad D_2&=\operatorname{Ran}F^*\cap\operatorname{Ran}N\subseteq H_2,\\ D_3&=\operatorname{Ran}G\cap\operatorname{Ran}K^*\subseteq H_3, &\qquad D_4&=\operatorname{Ran}G^*\cap\operatorname{Ran}N^*\subseteq H_4, \end{aligned}(4) and the restricted operators F̃(x):D2→D1\widetilde F(x)\colon D_2\to D_1, G̃(y):D4→D3\widetilde G(y)\colon D_4\to D_3, K̃(c):D3→D1\widetilde K(c)\colon D_3\to D_1 and Ñ(d):D4→D2\widetilde N(d)\colon D_4\to D_2 are the compressions of F(x)F(x), G(y)G(y), K(c)K(c), N(d)N(d) to them.

Array Objects

Let TT be a measurable automorphism of a standard Borel space XX, a nonempty measurable space isomorphic to a Polish space with its Borel sets, and let II be a finite index set. An array of order nn over a space YY is a function from the index tuples InI^n to YY, a point of YInY^{I^n}.✦ A map f:Y→Zf\colon Y\to Z acts on arrays entrywise✦, and an array of order zero is identified with its one entry, a point of YY.✦ For an axis aa and a slot i∈Ii\in I, the plane restriction fixing axis aa at slot ii sends an array of order n+1n+1 to the array of order nn whose entry at a tuple is the entry at that tuple with ii inserted at position aa✦; the last-axis planes of an array along J⊆IJ\subseteq I are its plane restrictions fixing the last axis at the slots in JJ.✦

Initial array laws for TT are probability measures mnXm^X_n on XInX^{I^n}, one for each order nn, that are invariant under TT acting entrywise, consistent, every plane restriction carrying mn+1Xm^X_{n+1} to mnXm^X_n, and proper: for J≠IJ\neq I, the last-axis planes along JJ are independent under mn+1Xm^X_{n+1}, each with law mnXm^X_n.✦

An array object MM over the initial array laws mnXm^X_n for TT consists of standard Borel spaces YY of states and UU of parameters, each with an automorphism, measurable maps ϕ:Y→X\phi\colon Y\to X intertwining the automorphism of YY with TT and π:Y→U\pi\colon Y\to U intertwining it with that of UU, and probability measures mnm_n on YInY^{I^n}, invariant, consistent and proper in the same sense and carried by ϕ\phi entrywise to mnXm^X_n, such that ϕ\phi and π\pi are independent under m0m_0 and the condition below holds.✦ The parameter array of an array is its image under π\pi entrywise, and κn\kappa_n is the conditional law under mnm_n of an array of order nn given its parameter array.✦ The condition is that for J≠IJ\neq I the last-axis planes along JJ of an array of order n+1n+1 are independent given its parameter array, the plane at each j∈Jj\in J having conditional law κn\kappa_n at the plane at jj of the parameter array.

An extension of MM by an array object NN over the same laws is a pair of measurable maps, one between the state spaces and one between the parameter spaces of NN and MM, that intertwine the automorphisms, are compatible with ϕ\phi and π\pi, and whose state map carries the laws mnNm^N_n of NN entrywise to the laws mnMm^M_n of MM.✦ The pullback along the state map of an extension of MM by NN is composition with that map entrywise, a linear isometry of L2(mnM)L^2(m^M_n) into L2(mnN)L^2(m^N_n).✦ On arrays of order n+1n+1 an array object comes with 2|I|+12|I|+1 sub-σ\sigma-algebras of the Borel sets of YIn+1Y^{I^{n+1}}: the parameter σ\sigma-algebra 𝒫n+1\mathcal P_{n+1} and, for each j∈Ij\in I, the outside σ\sigma-algebra 𝒪n+1,j\mathcal O_{n+1,j} and the enlarged σ\sigma-algebra 𝒢n+1,j\mathcal G_{n+1,j}. The projection onto one of these σ\sigma-algebras is the conditional expectation onto it, the orthogonal projection of L2(mn+1)L^2(m_{n+1}) onto the functions measurable for it✦, and MM is saturated if these projections commute with pullback along every extension of MM.✦

The Grouped Square

A triple partition of II is an ordered triple (P,A,B)(P,A,B) of nonempty, pairwise disjoint subsets of II whose union is II.✦ Fix an array object MM, a triple partition (P,A,B)(P,A,B) of II serving as the rows, a triple partition (P′,A′,B′)(P',A',B') of II serving as the columns and a level nn, the order of the cells below. An array zz of order n+2n+2 over YY has a cell zajz_{aj} for every row aa and column jj in II, the array of order nn obtained by fixing its first axis at aa and its last axis at jj.✦ For sets RR of rows and SS of columns, zRSz_{RS} is the block of the cells zajz_{aj} with a∈Ra\in R and j∈Sj\in S.✦ A group datum for RR and SS is a pair of a parameter array of order n+2n+2 over UU and a block, a family of arrays of order nn over YY indexed by RR and SS✦; the grouped datum of zz for RR and SS is the group datum formed by its parameter array π(z)\pi(z) and the block zRSz_{RS}.✦

The row family kernel is the product kernel that sends a parameter array ee of order n+2n+2 to the product, over the rows a∈Ra\in R, of κn+1\kappa_{n+1} at the plane of ee fixing the first axis at aa.✦ For sets RR of rows and SS of columns, the group marginal kernel is the image of the row family kernel for RR under the map that keeps only the cells in the columns of SS✦; its value at ee is the law λRSe\lambda^e_{RS} of the cells of the block in the rows of RR and the columns of SS.✦

The joining operator of measure-preserving maps uu and ww from a probability space to (Θ1,ν1)(\Theta_1,\nu_1) and (Θ2,ν2)(\Theta_2,\nu_2) is the operator L2(ν2)→L2(ν1)L^2(\nu_2)\to L^2(\nu_1) obtained by composing the pullback isometry f↦f∘wf\mapsto f\circ w with the adjoint of the pullback isometry along uu; it sends ff to the conditional expectation of f∘wf\circ w given uu.✦ A kernel that assigns to each point uu of a parameter space a coupling of two random points with laws ν1\nu_1 and ν2\nu_2 gives a family of joining operators L2(ν2)→L2(ν1)L^2(\nu_2)\to L^2(\nu_1), one for each uu, the joining operator of the coupling at uu, which sends a function of the second point to its conditional expectation given the first.✦

The row pair kernel for rows RR and columns S0S_0, S1S_1, S2S_2 is the conditional law under mn+2m_{n+2} of the pair of blocks (zRS1,zRS2)(z_{RS_1},z_{RS_2}) given the grouped datum of zz for RR and S0S_0.✦ The row joining operator from the columns S2S_2 to the columns S1S_1, conditioned on the columns S0S_0, at a group datum (e,ξ)(e,\xi), with ξ\xi a block in the rows of RR and the columns of S0S_0, is the joining operator L2(λRS2e)→L2(λRS1e)L^2(\lambda^e_{RS_2})\to L^2(\lambda^e_{RS_1}) of the row pair kernel.✦ For fixed ee, the fibre row operators are the row joining operators at the group data (e,ξ)(e,\xi), as a family over the blocks ξ\xi✦; the row-active subspace of L2(λRS1e)L^2(\lambda^e_{RS_1}) is their essential range for λRS0e\lambda^e_{RS_0}.✦

The parent rows of an array of order n+1n+1 along a set R0R_0 are its plane restrictions fixing the first axis at the slots of R0R_0.✦ The posterior kernel takes a parameter array uu of order n+1n+1 and a family of arrays indexed by R0R_0 to the conditional law of an array drawn from κn+1(u)\kappa_{n+1}(u) given that its parent rows along R0R_0 are that family, the disintegration of the observed kernel, the joint law of the parent rows and the array.✦ For columns SS and rows R0R_0, the column group kernel takes a group datum (e,γ)(e,\gamma), with γ\gamma a block in the rows of R0R_0 and the columns of SS, to the product over j∈Sj\in S of the posterior kernel at the plane of ee fixing the last axis at jj and the cells of γ\gamma in column jj.✦ For sets R1R_1 and R2R_2 of rows, the column pair kernel is the image of the column group kernel under the maps that keep only the cells in the rows R1R_1 and only those in the rows R2R_2; it is a coupling of the blocks zR1Sz_{R_1S} and zR2Sz_{R_2S}.✦ The column joining operator from the rows R2R_2 to the rows R1R_1, conditioned on the rows R0R_0, at (e,γ)(e,\gamma) is the joining operator L2(λR2Se)→L2(λR1Se)L^2(\lambda^e_{R_2S})\to L^2(\lambda^e_{R_1S}) of the column pair kernel.✦ For fixed ee, the column joining operators at the group data (e,γ)(e,\gamma) form the fibre column operators, a family over the blocks γ\gamma✦, and the column-active subspace of L2(λR1Se)L^2(\lambda^e_{R_1S}) is their essential range for λR0Se\lambda^e_{R_0S}.✦

Fix ee. At a corner (α,β)(\alpha,\beta) of the square, with α\alpha one of AA, BB and β\beta one of A′A', B′B', and with α‾\bar\alpha, β‾\bar\beta the other ones, VαβV_{\alpha\beta} is the row-active subspace of the rows α\alpha from the columns β‾\bar\beta to β\beta conditioned on P′P', and WαβW_{\alpha\beta} is the column-active subspace of the columns β\beta from the rows α‾\bar\alpha to α\alpha conditioned on PP. Both are closed subspaces of L2(λαβe)L^2(\lambda^e_{\alpha\beta}), and their intersection DαβD_{\alpha\beta} is the common active space of the corner.✦ The edges of the square are the fibre row operators RA(ξ)R_A(\xi) and RB(υ)R_B(\upsilon) of the rows AA and BB from the columns B′B' to A′A' conditioned on P′P', and the fibre column operators CA′(γ)C_{A'}(\gamma) and CB′(δ)C_{B'}(\delta) of the columns A′A' and B′B' from the rows BB to AA conditioned on PP, with ξ\xi, υ\upsilon, γ\gamma, δ\delta blocks in the rows and columns (A,P′)(A,P'), (B,P′)(B,P'), (P,A′)(P,A'), (P,B′)(P,B'). The restricted edge operators are the compressions of the edge operators to the common active spaces, R̃A(ξ):DAB′→DAA′\widetilde R_A(\xi)\colon D_{AB'}\to D_{AA'}, R̃B(υ):DBB′→DBA′\widetilde R_B(\upsilon)\colon D_{BB'}\to D_{BA'}, C̃A′(γ):DBA′→DAA′\widetilde C_{A'}(\gamma)\colon D_{BA'}\to D_{AA'} and C̃B′(δ):DBB′→DAB′\widetilde C_{B'}(\delta)\colon D_{BB'}\to D_{AB'}.

The Claim

Theorem 1. Let xx, yy, cc, dd and FF, GG, KK, NN form an operator rectangle. Then each family intertwines the projections onto essential ranges of the other families and of their adjoints, PRan⁡FK(c)=K(c)PRan⁡G,PRan⁡F*N(d)=N(d)PRan⁡G*,PRan⁡KF(x)=F(x)PRan⁡N,PRan⁡K*G(y)=G(y)PRan⁡N*,\begin{aligned} P^{\operatorname{Ran}F}K(c)&=K(c)\,P^{\operatorname{Ran}G}, &\qquad P^{\operatorname{Ran}F^*}N(d)&=N(d)\,P^{\operatorname{Ran}G^*},\\ P^{\operatorname{Ran}K}F(x)&=F(x)\,P^{\operatorname{Ran}N}, &\qquad P^{\operatorname{Ran}K^*}G(y)&=G(y)\,P^{\operatorname{Ran}N^*}, \end{aligned}(5)✦ and the projections onto the active intersections (4), PD1F(x)=F(x)PD2,PD3G(y)=G(y)PD4,PD1K(c)=K(c)PD3,PD2N(d)=N(d)PD4,\begin{aligned} P^{D_1}F(x)&=F(x)\,P^{D_2}, &\qquad P^{D_3}G(y)&=G(y)\,P^{D_4},\\ P^{D_1}K(c)&=K(c)\,P^{D_3}, &\qquad P^{D_2}N(d)&=N(d)\,P^{D_4}, \end{aligned}(6)✦ each identity holding for almost every value of the random element it involves.

Let MM be a saturated array object, (P,A,B)(P,A,B) and (P′,A′,B′)(P',A',B') triple partitions of II and nn a level. Then for almost every parameter array ee of order n+2n+2, under the law of the parameter array of an array drawn from mn+2m_{n+2}, the row-active and column-active projections commute at the corner (A,A′)(A,A'), PVAA′PWAA′=PWAA′PVAA′,P^{V_{AA'}}P^{W_{AA'}}=P^{W_{AA'}}P^{V_{AA'}},(7)✦ the edge operators intertwine the projections onto the common active spaces, PDAA′RA(ξ)=RA(ξ)PDAB′,PDBA′RB(υ)=RB(υ)PDBB′,PDAA′CA′(γ)=CA′(γ)PDBA′,PDAB′CB′(δ)=CB′(δ)PDBB′,\begin{aligned} P^{D_{AA'}}R_A(\xi)&=R_A(\xi)\,P^{D_{AB'}}, &\qquad P^{D_{BA'}}R_B(\upsilon)&=R_B(\upsilon)\,P^{D_{BB'}},\\ P^{D_{AA'}}C_{A'}(\gamma)&=C_{A'}(\gamma)\,P^{D_{BA'}}, &\qquad P^{D_{AB'}}C_{B'}(\delta)&=C_{B'}(\delta)\,P^{D_{BB'}}, \end{aligned}(8)✦ for almost every value of the block each identity involves, and the restricted edge operators satisfy both rectangle identities almost surely, R̃A(ξ)C̃B′(δ)=C̃A′(γ)R̃B(υ),R̃A(ξ)*C̃A′(γ)=C̃B′(δ)R̃B(υ)*,\widetilde R_A(\xi)\,\widetilde C_{B'}(\delta)=\widetilde C_{A'}(\gamma)\,\widetilde R_B(\upsilon), \qquad \widetilde R_A(\xi)^*\,\widetilde C_{A'}(\gamma)=\widetilde C_{B'}(\delta)\,\widetilde R_B(\upsilon)^*,(9)✦ each restricted family having essential total range (2) for the law of its block.

In the paper, the result reads: in every grouped cell, a corner (α,β)(\alpha,\beta) of the square, the projections PVαβP^{V_{\alpha\beta}} and PWαβP^{W_{\alpha\beta}} commute; row operators intertwine the column-active projections and column operators the row-active ones; all incident operators intertwine the projections PDαβ=PVαβPWαβP^{D_{\alpha\beta}}=P^{V_{\alpha\beta}}P^{W_{\alpha\beta}}; DαβD_{\alpha\beta} is invariant under conjugation; and the restrictions to the spaces DαβD_{\alpha\beta} have essential closed range equal to the target space, also for the reverse operators.

What Is Checked

The claim is stated over strongly measurable families of bounded operators between complex Hilbert spaces, indexed by random elements with four independent pairs and bound by the rectangle identities, with their essential kernels and ranges, active intersections and compressions; and over a saturated array object over initial array laws on a standard Borel space, two triple partitions of its index set, and the laws of cells, row and column joining operators and active subspaces built from its conditional laws.