OctoInform Beta

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

Proposition 4.8 [Common exact lines]: How exactness along rows and along columns meets column activity in the cells of a square of a saturated array object

Exact lines common to row and column laws.

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 stationary random array of states over a finite index set II, each state carrying a parameter, and cut a square from it: rows and columns indexed by II, each row and each column with its conditional law given its parameters. A circle-valued function of a block of cells may factor almost surely along the rows of the block, into functions of the other columns, or along its columns, into functions of the other rows. The question is when a function of a single cell that factors along its row also factors along its column, and how factoring along rows sits with the Markov operators that pass between rows along the columns.

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.✦ An array of order nn over a space YY is a map from InI^n to YY, a point of YInY^{I^n}✦; an array of order zero is a point of YY✦. For an array yy of order n+1n+1, an axis aa and an index i∈Ii\in I, the plane of yy fixing axis aa at ii is the array of order nn whose entry at a tuple vv is the entry of yy at the tuple obtained by inserting ii into vv at position aa.✦ Initial array laws of TT are probability measures mnXm^X_n on XInX^{I^n}, one for each nn, that are invariant under TT applied entrywise, consistent, in that every plane of an array drawn from mn+1Xm^X_{n+1} has law mnXm^X_n, and proper: for every proper subset JJ of II, the planes fixing the last axis at the indices in JJ are independent under mn+1Xm^X_{n+1}, each with law mnXm^X_n.✦

An array object MM over initial array laws mXm^X of TT consists of a standard Borel space YY of states and a standard Borel space UU of parameters, each with a measurable automorphism, measurable equivariant maps x:Y→Xx\colon Y\to X and par:Y→U\mathrm{par}\colon Y\to U, and probability measures mnm_n on YInY^{I^n} that are invariant, carried to mnXm^X_n by xx applied entrywise, consistent and proper in the same sense, and conditionally local: given the parameter array, obtained by applying par\mathrm{par} entrywise, the planes of properness stay independent, each with the conditional law of an array of order nn given the corresponding plane of the parameter array; at order zero, xx and par\mathrm{par} are independent under m0m_0.✦

The parameter law mnUm^U_n of MM is the law of the parameter array of an array drawn from mnm_n, a probability measure on UInU^{I^n}.✦ The kernel κn\kappa_n is the Markov kernel from UInU^{I^n} to YInY^{I^n} that gives at uu the conditional law under mnm_n of the array given that its parameter array is uu.✦

An extension of MM by an array object NN over the same initial laws is a pair of measurable equivariant maps, σ\sigma from the states of NN to those of MM and ρ\rho from the parameters of NN to 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\sigma applied entrywise carries mnNm^N_n to mnMm^M_n for every nn✦; pullback along it sends FF in the complex space L2(mnM)L^2(m^M_n) to F∘σF\circ\sigma in L2(mnN)L^2(m^N_n), with σ\sigma applied entrywise✦.

The data fields of MM on YIn+1Y^{I^{n+1}} are the parameter field 𝒫n+1\mathcal{P}_{n+1}, the sub-σ\sigma-algebra of the Borel sets generated by the parameter array, and, for each j∈Ij\in I, the outside field 𝒪n+1,j\mathcal{O}_{n+1,j} and the enlarged field 𝒢n+1,j\mathcal{G}_{n+1,j} at the index jj, sub-σ\sigma-algebras of the Borel sets of YIn+1Y^{I^{n+1}}; for a data field dd, EdME^M_d is the conditional expectation onto dd, the orthogonal projection of L2(mn+1M)L^2(m^M_{n+1}) onto the functions measurable for dd. The array object MM is saturated if, for every extension of MM by an array object NN, every nn and every data field dd, EdN(F∘σ)=(EdMF)∘σfor every F∈L2(mn+1M).E^N_d(F\circ\sigma)=(E^M_dF)\circ\sigma\qquad\text{for every }F\in L^2(m^M_{n+1}).(1)✦

The Square

An array yy of order n+2n+2 is read as a square. Its row ya⋅y_{a\cdot} is the plane fixing the first axis at aa, its column y⋅jy_{\cdot j} the plane fixing the last axis at jj, both of order n+1n+1, and its cell ya,jy_{a,j} is the array of order nn obtained by fixing both. For a set AA of rows and a set SS of columns, yA,S=(ya,j)a∈A,j∈Sy_{A,S}=(y_{a,j})_{a\in A,\,j\in S} is the block of cells in those rows and columns✦; yA,jy_{A,j} and ya,Sy_{a,S} are the cells of column jj in the rows of AA and of row aa in the columns of SS. A parameter array ee of order n+2n+2 is read in the same way. At n=0n=0, pa=ea⋅p_a=e_{a\cdot} and qj=e⋅jq_j=e_{\cdot j}, points of UIU^I, are the parameters of row aa and of column jj.

The law of cells λA,Se\lambda^e_{A,S} in the rows of AA and the columns of SS at a parameter array ee is the law of the block yA,Sy_{A,S} when the rows ya⋅y_{a\cdot}, a∈Aa\in A, are drawn independently, row aa from κn+1(ea⋅)\kappa_{n+1}(e_{a\cdot}).✦

A complement split of a nonempty set P⊆IP\subseteq I is an ordered pair (A,B)(A,B) of nonempty subsets of II such that PP, AA and BB are pairwise disjoint with union II.✦ A triple partition of II is an ordered triple of nonempty, pairwise disjoint subsets of II with union II.✦

Let PP, AA, BB be sets of rows, RR a set of columns and cc a family of cells in the rows of AA and the columns of RR. Draw each column y⋅jy_{\cdot j}, j∈Rj\in R, independently from κn+1(e⋅j)\kappa_{n+1}(e_{\cdot j}) conditioned on yA,j=cA,jy_{A,j}=c_{A,j}✦; this couples the blocks yP,Ry_{P,R} and yB,Ry_{B,R}✦. The column operator QceQ^e_c is the joining operator of this coupling: the bounded operator from L2(λB,Re)L^2(\lambda^e_{B,R}) to L2(λP,Re)L^2(\lambda^e_{P,R}) that sends a function of yB,Ry_{B,R} to its conditional expectation given yP,Ry_{P,R}.

The column-active space 𝒞A,Be(P,R)\mathcal{C}^e_{A,B}(P,R) is the essential range of the family c↦Qcec\mapsto Q^e_c with respect to λA,Re\lambda^e_{A,R}, the orthogonal complement in L2(λP,Re)L^2(\lambda^e_{P,R}) of the essential kernel of the adjoint family: 𝒞A,Be(P,R)={w∈L2(λP,Re):(Qce)*w=0for λA,Re-almost every c}⟂.\mathcal{C}^e_{A,B}(P,R)=\bigl\{w\in L^2(\lambda^e_{P,R}) : (Q^e_c)^*w=0\ \text{for } \lambda^e_{A,R}\text{-almost every }c\bigr\}^{\perp}.(2)✦

A measurable map ff to the unit circle 𝕋\mathbb{T} of ℂ\mathbb{C}, with its Borel sets, defines a vector of L2L^2 of a finite measure, written ff again.✦ Let t=(A′,B′)t=(A',B') be a complement split of the set RR of columns. A vector vv of L2(λP,Re)L^2(\lambda^e_{P,R}) is exact for the row split tt if v=dfv=d\,f with d∈ℂd\in\mathbb{C} and ff a measurable map from the blocks of cells in the rows of PP and the columns of RR to 𝕋\mathbb{T} such that, for some measurable maps gg and hh to 𝕋\mathbb{T}, f(yP,R)=g(yP,A′)h(yP,B′)for κn+2(e)-almost every y.f(y_{P,R})=g(y_{P,A'})\,h(y_{P,B'})\qquad\text{for }\kappa_{n+2}(e)\text{-almost every }y.(3)✦ The set 𝒳te(P,R)\mathcal{X}^e_t(P,R) is the set of vectors of L2(λP,Re)L^2(\lambda^e_{P,R}) that are exact for the row split tt.✦

A vector vv of L2(λP,Re)L^2(\lambda^e_{P,R}) is fully exact, exact for the law of the rows of PP and for the law of the columns of RR, if v=dfv=d\,f with d∈ℂd\in\mathbb{C} and ff a measurable map from the blocks of cells in the rows of PP and the columns of RR to 𝕋\mathbb{T} such that, for some constants c,c′∈𝕋c,c'\in\mathbb{T} and measurable maps gig_i, i∉Ri\notin R, and hih_i, i∉Pi\notin P, to 𝕋\mathbb{T}, f(yP,R)=c∏i∉Rgi(yP,i)=c′∏i∉Phi(yi,R)for κn+2(e)-almost every y.f(y_{P,R})=c\prod_{i\notin R}g_i(y_{P,i})=c'\prod_{i\notin P}h_i(y_{i,R})\qquad \text{for }\kappa_{n+2}(e)\text{-almost every }y.(4)✦ The set ℱe(P,R)\mathcal{F}^e(P,R) is the set of fully exact vectors of L2(λP,Re)L^2(\lambda^e_{P,R}).✦

Cells and Lines

At n=0n=0, a block with the single row aa and the single column jj is one point of YY, and γ\gamma is the measurable isomorphism from such blocks onto YY.✦ For a measure ν\nu on YY to which γ\gamma carries λ{a},{j}e\lambda^e_{\{a\},\{j\}}, a vector vv of L2(ν)L^2(\nu) pulls back to v∘γv\circ\gamma in L2(λ{a},{j}e)L^2(\lambda^e_{\{a\},\{j\}}).

A family (fi)i∈I(f_i)_{i\in I} of measurable maps from YY to 𝕋\mathbb{T} is an exact tuple for a law Λ\Lambda on YIY^I if, for some constant c∈𝕋c\in\mathbb{T}, ∏i∈Ifi(yi)=cfor Λ-almost every y.\prod_{i\in I}f_i(y_i)=c\qquad\text{for }\Lambda\text{-almost every }y.(5) For a finite measure ν\nu on YY, a slot i∈Ii\in I and p∈UIp\in U^I, the exact cone ℰi(ν,p)\mathcal{E}_i(\nu,p) is the set of vectors dfid\,f_i of L2(ν)L^2(\nu) with d∈ℂd\in\mathbb{C} and (fk)k∈I(f_k)_{k\in I} an exact tuple for κ1(p)\kappa_1(p).✦

For a measure ν\nu on YY, a slot i∈Ii\in I and p∈UIp\in U^I, the slot-active space 𝒜i(ν,p)\mathcal{A}_i(\nu,p) is the following closed subspace of L2(ν)L^2(\nu).✦ Draw an array yy of order one from κ1(p)\kappa_1(p), with entries yky_k, k∈Ik\in I. If the entry yiy_i has law ν\nu, then 𝒜i(ν,p)\mathcal{A}_i(\nu,p) is the closure of the range of the joining operator from L2L^2 of the law of the other entries (yk)k≠i(y_k)_{k\ne i} to L2(ν)L^2(\nu), the operator that sends a function of (yk)k≠i(y_k)_{k\ne i} to its conditional expectation given yiy_i; otherwise 𝒜i(ν,p)={0}\mathcal{A}_i(\nu,p)=\{0\}.

A line is a measurable map from YY to 𝕋\mathbb{T} up to a constant phase, and lines are labelled by codes. The test vectors are the complex rational combinations of indicators of sets in the algebra generated by a fixed countable generating family of the σ\sigma-algebra of YY, taken in a fixed enumeration.✦ A Hilbert code for κ0\kappa_0 is a pair z=(u,(rk)k≥0)z=(u,(r_k)_{k\ge0}) of a base point u∈Uu\in U and a sequence of nonnegative reals that locates one vector of L2(κ0(u))L^2(\kappa_0(u)) by its distances rkr_k to the test vectors.✦ The representative z̃\tilde z of a Hilbert code zz for κ0\kappa_0 is the pointwise limit, as m→∞m\to\infty, of the first test vector within 2−m2^{-m} of the vector that zz locates.✦ A Hilbert code z=(u,(rk)k≥0)z=(u,(r_k)_{k\ge0}) for κ0\kappa_0 is unimodular if |z̃|=1|\tilde z|=1 almost everywhere for κ0(u)\kappa_0(u)✦, and canonical if the first test vector not orthogonal to the vector that zz locates has a positive real inner product with it.✦

For a Markov kernel Λ\Lambda from UIU^I to YIY^I and a slot ii, the line graph ℒi(Λ)\mathcal{L}_i(\Lambda) is the set of pairs (p,z)(p,z) of p∈UIp\in U^I and a code zz for κ0\kappa_0 such that z=ziz=z_i for an exact family of codes at pp: canonical unimodular codes zkz_k, k∈Ik\in I, each with base point pkp_k, with |∫∏k∈Iz̃k(yk)dΛ(p)(y)|=1.\Bigl|\int\prod_{k\in I}\tilde z_k(y_k)\,d\Lambda(p)(y)\Bigr|=1.(6)✦

The Claim

Theorem 1. Let MM be an array object over initial array laws of TT, with finite index set II, and let a,j∈Ia,j\in I.

  1. Let MM be saturated, let n≥0n\ge0, let PP be a set of rows with a complement split (A,B)(A,B), and let RR be a set of columns that has a complement split. Then for mn+2Um^U_{n+2}-almost every parameter array ee of order n+2n+2, 𝒞A,Be(P,R)∩⋂t𝒳te(P,R)⊆ℱe(P,R),\mathcal{C}^e_{A,B}(P,R)\cap\bigcap_{t}\mathcal{X}^e_t(P,R)\subseteq\mathcal{F}^e(P,R),(7)✦ the intersection running over all complement splits tt of RR.

  2. Let MM be saturated, let II have at least two elements, and let (A,B)(A,B) be a complement split of {a}\{a\}. Then for m2Um^U_2-almost every ee the following holds: if ν\nu is a measure on YY to which γ\gamma carries λ{a},{j}e\lambda^e_{\{a\},\{j\}} and v∈L2(ν)v\in L^2(\nu), then v∈𝒜a(ν,qj)⇔v∘γ∈𝒞A,Be({a},{j}).v\in\mathcal{A}_a(\nu,q_j)\iff v\circ\gamma\in\mathcal{C}^e_{A,B}(\{a\},\{j\}).(8)✦

  3. Let ee be a parameter array of order two such that the column y⋅jy_{\cdot j} of a square yy drawn from κ2(e)\kappa_2(e) has law κ1(qj)\kappa_1(q_j), and let ν\nu be a finite measure on YY to which γ\gamma carries λ{a},{j}e\lambda^e_{\{a\},\{j\}}. Then for every v∈L2(ν)v\in L^2(\nu), v∘γ∈ℱe({a},{j})⟹v∈ℰa(ν,qj).v\circ\gamma\in\mathcal{F}^e(\{a\},\{j\})\implies v\in\mathcal{E}_a(\nu,q_j).(9)✦

  4. Let MM be saturated, let n≥0n\ge0, and let (Pr,Ar,Br)(P_r,A_r,B_r) and (Pc,Ac,Bc)(P_c,A_c,B_c) be triple partitions of II, read as sets of rows and of columns. Then for mn+2Um^U_{n+2}-almost every ee the following holds: if ff, gg, hh are measurable maps to 𝕋\mathbb{T} of the blocks of cells in the rows of ArA_r and the columns of AcA_c, BcB_c, PcP_c respectively with g(yAr,Bc)¯=h(yAr,Pc)f(yAr,Ac)\overline{g(y_{A_r,B_c})}=h(y_{A_r,P_c})\,f(y_{A_r,A_c}) for κn+2(e)\kappa_{n+2}(e)-almost every yy, then f∈𝒞Pr,Bre(Ar,Ac)⇔g∈𝒞Pr,Bre(Ar,Bc).f\in\mathcal{C}^e_{P_r,B_r}(A_r,A_c)\iff g\in\mathcal{C}^e_{P_r,B_r}(A_r,B_c).(10)✦

  5. Let MM be saturated, let II have at least three elements, and let t∈It\in I differ from jj. Then for m2Um^U_2-almost every ee and every exact tuple (fi)i∈I(f_i)_{i\in I} for κ1(pa)\kappa_1(p_a), fj∈𝒜a(κ0(ea,j),qj)⇔ft∈𝒜a(κ0(ea,t),qt).f_j\in\mathcal{A}_a\bigl(\kappa_0(e_{a,j}),q_j\bigr)\iff f_t\in\mathcal{A}_a\bigl(\kappa_0(e_{a,t}),q_t\bigr).(11)✦

  6. Let MM be saturated, let II have at least three elements, and let (fp,i)i∈I(f_{p,i})_{i\in I}, for p∈UIp\in U^I, be maps from YY to 𝕋\mathbb{T} with (p,y)↦fp,i(y)(p,y)\mapsto f_{p,i}(y) measurable for each ii, forming an exact tuple for κ1(p)\kappa_1(p) for m1Um^U_1-almost every pp. Then for m2Um^U_2-almost every ee, fpa,j∈𝒜a(κ0(ea,j),qj).f_{p_a,j}\in\mathcal{A}_a\bigl(\kappa_0(e_{a,j}),q_j\bigr).(12)✦

  7. Let MM be saturated, let II have at least three elements, and let Λ\Lambda be a Markov kernel from UIU^I to YIY^I with Λ(p)=κ1(p)\Lambda(p)=\kappa_1(p) for m1Um^U_1-almost every pp, under which, for every pp and every k∈Ik\in I, the coordinate yky_k has law κ0(pk)\kappa_0(p_k) and the coordinates other than yky_k are independent. Then for m2Um^U_2-almost every ee and every Hilbert code zz for κ0\kappa_0, (pa,z)∈ℒj(Λ)⟹(qj,z)∈ℒa(Λ).(p_a,z)\in\mathcal{L}_j(\Lambda)\implies(q_j,z)\in\mathcal{L}_a(\Lambda).(13)✦

In the paper, the result reads: for almost every parameter square and at every cell (a,j)(a,j), the projections onto the paper’s subspaces Va,jV_{a,j} and Wa,jW_{a,j} commute and their common range is spanned by lines exact for both the full row law κ1(pa)\kappa_1(p_a) and the full column law κ1(qj)\kappa_1(q_j), the row Markov operators intertwine the column-active projections and the column operators the row-active ones, and for an exact row tuple (χi)i∈I(\chi_i)_{i\in I} and partner indices ss and tt in II the row operator carries χt\chi_t to an explicit multiple of χs¯\overline{\chi_s}, so that column activity of one partner line is equivalent to that of any other, and symmetrically.

What Is Checked

The claim is stated over a saturated array object MM with its parameter laws mnUm^U_n and conditional laws κn\kappa_n, the laws of cells λA,Se\lambda^e_{A,S} of a square, the column operators QceQ^e_c and the column-active spaces 𝒞A,Be(P,R)\mathcal{C}^e_{A,B}(P,R), the sets 𝒳te(P,R)\mathcal{X}^e_t(P,R) and ℱe(P,R)\mathcal{F}^e(P,R) of vectors exact for a row split and fully exact, and, at a single cell, the slot-active spaces 𝒜i\mathcal{A}_i, the exact cones ℰi\mathcal{E}_i and the line graphs ℒi\mathcal{L}_i of canonical unimodular Hilbert codes.

Beyond the claim, the library proves that the vectors a family of operators sends to zero almost everywhere form a linear subspace✦, and that a measurable map to 𝕋\mathbb{T} is square integrable for every finite measure✦. It also proves that pullback along an extension is a linear isometry at every order.