OctoInform Beta

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

Lemma 4.3 [The rectangle identities]: In the grouped square of a saturated array object, LAJD=JSLBL_AJ_D = J_SL_B and LA*JS=JDLB*L_A^*J_S = J_DL_B^* for row operators LL and column operators JJ

Row and column operators almost surely form a rectangle.

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 an array object and draw an array yy of order n+2n+2 from its law mn+2m_{n+2}. Read the first axis of yy as rows and the last as columns, so that the cell of yy in row aa and column jj is an array of order nn. Split the columns into groups RR, SS, DD and choose groups PP, AA, BB of rows, with PP disjoint from AA and from BB. The four blocks of cells yA×Sy_{A\times S}, yA×Dy_{A\times D}, yB×Sy_{B\times S} and yB×Dy_{B\times D} are the corners of a square.

Each side of the square carries a joining of its two blocks, and with it a Markov operator between their L2L^2 spaces. A row side joins the two blocks in the same rows, given the parameters of yy and the cells of those rows in the columns of RR; a column side joins the two blocks in the same columns, given the parameters and the cells of the rows of PP in those columns. The question is whether the two ways around the square from yB×Dy_{B\times D} to yA×Sy_{A\times S} give the same operator, and whether the two ways from yB×Sy_{B\times S} to yA×Dy_{A\times D} do, each of these crossing one side backwards by its adjoint.

The library states these rectangle identities for a saturated array object, over the whole group data, which consist of the parameter array, the rows of PP and the columns of RR, and fibre by fibre over the parameters, and the identities behind them for the joining operators of a square of kernels, plain and given a parameter.

Joining Operators

All L2L^2 spaces are complex. The pullback isometry of a measure-preserving map ϕ\phi from (Ω,μ)(\Omega,\mu) to (Z1,ν1)(Z_1,\nu_1) is the operator ϕ*:L2(ν1)→L2(μ)\phi^*\colon L^2(\nu_1)\to L^2(\mu), f↦f∘ϕf\mapsto f\circ\phi.✦ The joining operator of measure-preserving maps ϕ\phi to (Z1,ν1)(Z_1,\nu_1) and ψ\psi to (Z2,ν2)(Z_2,\nu_2) on the same space is the bounded operator Qϕ,ψ=(ϕ*)*ψ*:L2(ν2)→L2(ν1),Q_{\phi,\psi} = (\phi^*)^*\,\psi^*\colon L^2(\nu_2)\to L^2(\nu_1),(1)✦ which sends gg to the conditional expectation of g∘ψg\circ\psi given ϕ\phi, read on Z1Z_1. The family of joining operators of a kernel θ\theta from a space VV to Z1×Z2Z_1\times Z_2, with reference measures ν1\nu_1 and ν2\nu_2, assigns to each v∈Vv\in V the joining operator Qθ(v):L2(ν2)→L2(ν1)Q_\theta(v)\colon L^2(\nu_2)\to L^2(\nu_1) of the coupling θ(v)\theta(v), which sends a function of the second point to its conditional expectation given the first.✦

A rectangle of operators is a quadruple of bounded linear operators F:H3→H1F\colon H_3\to H_1, G:H4→H2G\colon H_4\to H_2, K:H2→H1K\colon H_2\to H_1 and N:H4→H3N\colon H_4\to H_3 between complex Hilbert spaces such that FN=KGandF*K=NG*.FN = KG \quad\text{and}\quad F^*K = NG^*.(2)✦

Array Objects

Fix a finite index set II. An array of order nn over a space YY is a function from the index tuples InI^n to YY; these arrays form the product YInY^{I^n}✦, and those of order zero are the points of YY✦. A map f:Y→Zf\colon Y\to Z acts on arrays entrywise, f(n)y=f∘yf^{(n)}y = f\circ y.✦ For an array yy of order n+1n+1, an axis a∈{0,…,n}a\in\{0,\dots,n\} and an index i∈Ii\in I, the plane πa,iy\pi_{a,i}y is the array of order nn obtained by fixing axis aa of yy at ii✦; the planes of yy along a subset JJ of II are its planes πn,jy\pi_{n,j}y, j∈Jj\in J, on the last axis✦.

Initial array laws mXm^X of a measurable automorphism TT of a nonempty standard Borel space XX are probability measures mnXm^X_n on XInX^{I^n}, one for each n≥0n\ge 0, each invariant under T(n)T^{(n)}, that are consistent and proper: for every axis aa, every index ii and every subset J≠IJ\ne I of II, (πa,i)*mn+1X=mnX,((πn,j)j∈J)*mn+1X=(mnX)⊗J.(\pi_{a,i})_*\,m^X_{n+1} = m^X_n, \qquad \bigl((\pi_{n,j})_{j\in J}\bigr)_*\,m^X_{n+1} = (m^X_n)^{\otimes J}.(3)✦

An array object MM over mXm^X is a pair of nonempty standard Borel spaces, a state space YY and a parameter space UU, with 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}, n≥0n\ge 0, such that each mnm_n is invariant under TY(n)T_Y^{(n)} and carried to mnXm^X_n by x(n)x^{(n)}, the family is consistent and proper as in (3), the maps xx and par\mathrm{par} are independent under m0m_0 read on YY, and for every subset J≠IJ\ne I of II and almost every pp Lawmn+1((πn,jy)j∈J|y‾=p)=⨂j∈Jκn(πn,jp),\mathrm{Law}_{m_{n+1}}\bigl((\pi_{n,j}y)_{j\in J}\mathrel{\big|}\bar y = p\bigr) = \bigotimes_{j\in J}\kappa_n(\pi_{n,j}\,p),(4)✦ where y‾\bar y is the parameter array of yy, the array par\mathrm{par} applied entrywise, and κn(e)\kappa_n(e) is the conditional law under mnm_n of an array given that its parameter array is ee. The kernel κn\kappa_n is the conditional law of MM at order nn.✦

At each order n+1n+1 an array object has data fields, sub-σ\sigma-algebras of the Borel sets of YIn+1Y^{I^{n+1}}: the parameter field 𝒫n+1\mathcal{P}_{n+1} 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}. The data kinds are the 2|I|+12|I|+1 labels dd of the data fields 𝒫n+1\mathcal{P}_{n+1}, 𝒪n+1,j\mathcal{O}_{n+1,j} and 𝒢n+1,j\mathcal{G}_{n+1,j}, j∈Ij\in I. The projection Πn,dM\Pi^M_{n,d} onto the data 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 extension of MM by an array object NN over the same initial laws is a pair of measurable maps, σ\sigma from the state space of NN to that of MM and τ\tau from the parameter space of NN to that of MM, that intertwine the automorphisms, satisfy xM∘σ=xNx_M\circ\sigma = x_N and parM∘σ=τ∘parN\mathrm{par}_M\circ\sigma = \tau\circ\mathrm{par}_N, and have σ*(n)mnN=mnM\sigma^{(n)}_*m^N_n = m^M_n for every nn, where mnMm^M_n and mnNm^N_n are the laws of order nn of MM and of NN, and xMx_M, parM\mathrm{par}_M, xNx_N, parN\mathrm{par}_N are their maps xx and par\mathrm{par}.✦ The pullback along an extension at order nn is the linear isometry σn*:L2(mnM)→L2(mnN)\sigma^*_n\colon L^2(m^M_n)\to L^2(m^N_n), f↦f∘σ(n)f\mapsto f\circ\sigma^{(n)}.✦ An array object MM is saturated if its projections onto data commute with every such pullback: for every extension of MM by an array object NN, every nn, every data kind dd and every f∈L2(mn+1M)f\in L^2(m^M_{n+1}), Πn,dNσn+1*f=σn+1*Πn,dMf.\Pi^N_{n,d}\,\sigma^*_{n+1}f = \sigma^*_{n+1}\,\Pi^M_{n,d}f.(5)✦

The Grouped Square

Let yy be an array of order n+2n+2 over the state space of MM. Its row ya⋅=π0,ayy_{a\cdot} = \pi_{0,a}y and its column y⋅j=πn+1,jyy_{\cdot j} = \pi_{n+1,j}y are arrays of order n+1n+1, and its cell yaj=πn,jπ0,ayy_{aj} = \pi_{n,j}\pi_{0,a}y in row aa and column jj is an array of order nn. For a set AA of rows and a set CC of columns, yA×Cy_{A\times C} is the family of cells yajy_{aj}, a∈Aa\in A, j∈Cj\in C.✦ The row restriction of rows (za)a∈A(z_a)_{a\in A} to the columns of CC is the family (πn,jza)(\pi_{n,j}z_a)✦, and the column restriction of columns (wj)j∈C(w_j)_{j\in C} to the rows of AA is the family (π0,awj)(\pi_{0,a}w_j)✦, both indexed by A×CA\times C.

A group datum for rows AA and columns CC is a pair (e,h)(e,h) of a parameter array ee of order n+2n+2 and a family hh of cells indexed by A×CA\times C✦; the grouped data map sends yy to (y‾,yA×C)(\bar y, y_{A\times C})✦. The parent data of yy for a set PP of rows are y‾\bar y and the parent rows yp⋅y_{p\cdot}, p∈Pp\in P, and its whole group data for PP and a set RR of columns add the columns y⋅jy_{\cdot j}, j∈Rj\in R: W(y)=(y‾,(yp⋅)p∈P,(y⋅j)j∈R).W(y) = \bigl(\bar y,\ (y_{p\cdot})_{p\in P},\ (y_{\cdot j})_{j\in R}\bigr).(6) For a whole group datum uu with parameter array ee, its group selection for rows AA is ee with the column restriction of its columns to the rows of AA✦, and its column own data for columns SS are ee with the row restriction of its parent rows to the columns of SS✦; at u=W(y)u = W(y) these are (y‾,yA×R)(\bar y, y_{A\times R}) and (y‾,yP×S)(\bar y, y_{P\times S}).

The row family kernel ρA\rho_A for a set AA of rows is the product kernel from parameter arrays ee of order n+2n+2 to families of rows indexed by AA: ρA(e)=⨂a∈Aκn+1(π0,ae),\rho_A(e) = \bigotimes_{a\in A}\kappa_{n+1}(\pi_{0,a}e),(7) The law of cells λA,C(e)\lambda_{A,C}(e), the group marginal at ee for rows AA and columns CC, is the image of ρA(e)\rho_A(e) under the row restriction to CC.✦ For column groups RR, SS, DD, the row pair kernel θA\theta_A sends a group datum pp for AA and RR to the conditional law under mn+2m_{n+2} of the pair (yA×S,yA×D)(y_{A\times S}, y_{A\times D}) given (y‾,yA×R)=p(\bar y, y_{A\times R}) = p.✦

The posterior kernel βP\beta_P for a set PP of rows sends a parameter array vv of order n+1n+1 and arrays hph_p of order nn, p∈Pp\in P, to the conditional law of an array zz drawn from κn+1(v)\kappa_{n+1}(v) given that its planes π0,pz\pi_{0,p}z are the hph_p: the observation kernel, the joint law of these planes and zz, disintegrated along the planes.✦ The column group kernel γS\gamma_S for PP and a set SS of columns is the product kernel from group data for PP and SS to families of columns indexed by SS, γS(e,h)=⨂j∈SβP(πn+1,je,(hpj)p∈P),\gamma_S(e,h) = \bigotimes_{j\in S}\beta_P\bigl(\pi_{n+1,j}e,\ (h_{pj})_{p\in P}\bigr),(8)✦ and the column pair kernel ωS\omega_S is its image under the column restrictions to the rows of AA and to the rows of BB✦.

The row operator for AA at a group datum p=(e,h)p = (e,h) for AA and RR is the joining operator LA(p)=QθA(p)L_A(p) = Q_{\theta_A}(p) from L2(λA,D(e))L^2(\lambda_{A,D}(e)) to L2(λA,S(e))L^2(\lambda_{A,S}(e)).✦ The column operator for SS at a group datum p=(e,h)p = (e,h) for PP and SS is JS(p)=QωS(p)J_S(p) = Q_{\omega_S}(p) from L2(λB,S(e))L^2(\lambda_{B,S}(e)) to L2(λA,S(e))L^2(\lambda_{A,S}(e)).✦ The row operator LB(p)L_B(p) for BB at a group datum p=(e,h)p=(e,h) for BB and RR is QθB(p)Q_{\theta_B}(p) from L2(λB,D(e))L^2(\lambda_{B,D}(e)) to L2(λB,S(e))L^2(\lambda_{B,S}(e)), and the column operator JD(p)J_D(p) for DD at a group datum p=(e,h)p=(e,h) for PP and DD is QωD(p)Q_{\omega_D}(p) from L2(λB,D(e))L^2(\lambda_{B,D}(e)) to L2(λA,D(e))L^2(\lambda_{A,D}(e)). At a fixed parameter array ee, the operators LA(e,h)L_A(e,h) and JS(e,h)J_S(e,h), as hh varies, are the fibre row and column operators at ee.

The Claim

Theorem 1. Let MM be a saturated array object over initial array laws, and let n≥0n\ge 0. Let PP, AA, BB be sets of rows, with PP disjoint from AA and from BB, |P|+2≤|I||P|+2\le|I|, |P|+|A|+1≤|I||P|+|A|+1\le|I|, |P|+|B|+1≤|I||P|+|B|+1\le|I|, and neither P∪AP\cup A nor P∪BP\cup B equal to II. Let RR, SS, DD be pairwise disjoint sets of columns with union II, SS and DD nonempty.

  1. For almost every value uu of the whole group data (6) for PP and RR of an array drawn from mn+2m_{n+2}, the row operators LA(u)L_A(u) and LB(u)L_B(u) at the group selections of uu for AA and BB, and the column operators JS(u)J_S(u) and JD(u)J_D(u) at its column own data for SS and DD, satisfy LA(u)JD(u)=JS(u)LB(u),LA(u)*JS(u)=JD(u)LB(u)*.\begin{aligned} L_A(u)\,J_D(u) &= J_S(u)\,L_B(u),\\ L_A(u)^*\,J_S(u) &= J_D(u)\,L_B(u)^*. \end{aligned}(9)✦

  2. For almost every parameter array ee of an array drawn from mn+2m_{n+2}, and for κn+2(e)\kappa_{n+2}(e)-almost every array yy, the fibre operators at ee satisfy LA(e,yA×R)JD(e,yP×D)=JS(e,yP×S)LB(e,yB×R),LA(e,yA×R)*JS(e,yP×S)=JD(e,yP×D)LB(e,yB×R)*.\begin{aligned} L_A(e,y_{A\times R})\,J_D(e,y_{P\times D}) &= J_S(e,y_{P\times S})\,L_B(e,y_{B\times R}),\\ L_A(e,y_{A\times R})^*\,J_S(e,y_{P\times S}) &= J_D(e,y_{P\times D})\,L_B(e,y_{B\times R})^*. \end{aligned}(10)✦

  3. Let ss and tt be measure-preserving maps from a probability space (Ω,μ)(\Omega,\mu) to measures νs\nu_s and νt\nu_t, let aa and bb be measure-preserving maps from νs\nu_s to νa\nu_a and νb\nu_b, and cc and dd from νt\nu_t to νc\nu_c and νd\nu_d. Let κac\kappa_{ac}, κbd\kappa_{bd}, κca\kappa_{ca}, κdb\kappa_{db} be Markov kernels such that, under μ\mu, the conditional laws of c∘tc\circ t and d∘td\circ t given ss are κac(a(s))\kappa_{ac}(a(s)) and κbd(b(s))\kappa_{bd}(b(s)), and those of a∘sa\circ s and b∘sb\circ s given tt are κca(c(t))\kappa_{ca}(c(t)) and κdb(d(t))\kappa_{db}(d(t)). Then, with Qa,bQ_{a,b} taken under νs\nu_s and Qc,dQ_{c,d} under νt\nu_t, Qa∘s,c∘tQc,d=Qa,bQb∘s,d∘t,Qa∘s,c∘t*Qa,b=Qc,dQb∘s,d∘t*.\begin{aligned} Q_{a\circ s,\,c\circ t}\,Q_{c,d} &= Q_{a,b}\,Q_{b\circ s,\,d\circ t},\\ Q_{a\circ s,\,c\circ t}^*\,Q_{a,b} &= Q_{c,d}\,Q_{b\circ s,\,d\circ t}^*. \end{aligned}(11)✦

  4. Let aa, bb, cc, dd and uu be measurable maps on a probability space (Ω,μ)(\Omega,\mu) that has a disintegration along uu, and let νa\nu_a, νb\nu_b, νc\nu_c, νd\nu_d be measures on the value spaces of aa, bb, cc, dd. Suppose that given u=qu = q the pairs (a,c)(a,c), (b,d)(b,d), (a,b)(a,b), (c,d)(c,d) have conditional laws ηac(q)\eta_{ac}(q), ηbd(q)\eta_{bd}(q), ηab(q)\eta_{ab}(q), ηcd(q)\eta_{cd}(q), which for almost every qq are couplings of νa\nu_a and νc\nu_c, of νb\nu_b and νd\nu_d, of νa\nu_a and νb\nu_b, and of νc\nu_c and νd\nu_d. Suppose that given uu and (a,b)(a,b) the conditional law of cc depends only on uu and aa, and that of dd only on uu and bb; and that given uu and (c,d)(c,d) the conditional law of aa depends only on uu and cc, and that of bb only on uu and dd. Then for almost every value qq of uu the joining operators of these couplings, with reference measures among νa\nu_a, νb\nu_b, νc\nu_c, νd\nu_d, satisfy Qηac(q)Qηcd(q)=Qηab(q)Qηbd(q),Qηac(q)*Qηab(q)=Qηcd(q)Qηbd(q)*.\begin{aligned} Q_{\eta_{ac}}(q)\,Q_{\eta_{cd}}(q) &= Q_{\eta_{ab}}(q)\,Q_{\eta_{bd}}(q),\\ Q_{\eta_{ac}}(q)^*\,Q_{\eta_{ab}}(q) &= Q_{\eta_{cd}}(q)\,Q_{\eta_{bd}}(q)^*. \end{aligned}(12)✦

In the paper, the result reads: for the grouped square, almost surely LA(x)Jt(d)=Js(c)LB(y)L_A(x)J_t(d) = J_s(c)L_B(y) and LA(x)*Js(c)=Jt(d)LB(y)*L_A(x)^*J_s(c) = J_t(d)L_B(y)^*, in every ordering of the groups, with LA(x)L_A(x), LB(y)L_B(y), Js(c)J_s(c) and Jt(d)J_t(d) standing for the operators written here LA(u)L_A(u), LB(u)L_B(u), JS(u)J_S(u) and JD(u)J_D(u).

What Is Checked

The claims are stated over the row and column operators of the grouped square of a saturated array object, the joining operators of its row pair and column pair kernels, and over the joining operators of measure-preserving maps and of conditional couplings.