OctoInform Beta

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

Lemma 4.7 [Projection of tensor contractions]: For almost every parameter array, with inputs projected onto their own active spaces; the corner’s common active space is L2L^2 of circle-valued coordinates

Active projections pass through contractions in saturated array objects.

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

Fix a measurable automorphism TT of a standard Borel space XX and a finite index set II. An array of order nn over a space YY is a point of YInY^{I^n}, an entry in YY for every index tuple v∈Inv \in I^n.✦ For an axis a∈{0,…,n}a \in \{0, \dots, n\} and an index i∈Ii \in I, the plane restriction πa,i\pi_{a,i} sends an array yy of order n+1n+1 to 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.✦ A map f:Y→Zf \colon Y \to Z acts on arrays entrywise, and the entrywise map is again written ff.✦

An array yy of order n+2n+2 is an I×II \times I grid: its row cc is the plane π0,cy\pi_{0,c}\,y, its column jj is the plane πn+1,jy\pi_{n+1,j}\,y, and its cell ycjy_{cj} is the array of order nn obtained by fixing the first axis at cc and the last at jj. For a set RR of rows and a set SS of columns, the block yRS=(ycj)c∈R,j∈Sy_{RS} = (y_{cj})_{c \in R,\, j \in S} is the family of cells in R×SR \times S.✦ A triple partition of II is an ordered splitting I=P⊔A⊔BI = P \sqcup A \sqcup B into three nonempty, pairwise disjoint sets.✦ With a triple partition P,A,BP, A, B of the rows and a triple partition P′,A′,B′P', A', B' of the columns the grid falls into nine blocks, and the corner yPP′y_{PP'} has two neighbours in its row of blocks, yPA′y_{PA'} and yPB′y_{PB'}, and two in its column, yAP′y_{AP'} and yBP′y_{BP'}.

In an array object, defined below together with its parameter array, the grid is random, and given its parameter array each block has a law. Conditioning on the corner the product of a function of one neighbour and a function of the other neighbour on the same side gives a function of the corner, their contraction; and each block has a closed subspace of L2L^2 of its law, its common active space. The question is, for a saturated array object (defined below), how orthogonal projection onto the common active space of a block acts on contractions of functions of its neighbours, and what that space is.

Array Objects

Initial array laws mXm^X of TT are probability measures mnXm^X_n on XInX^{I^n}, one for each order nn, that are invariant under TT acting entrywise, consistent, in that every plane restriction πa,i\pi_{a,i} carries mn+1Xm^X_{n+1} to mnXm^X_n, and proper, in that for every subset J≠IJ \neq I of II the last-axis planes (πn,jy)j∈J(\pi_{n,j}\,y)_{j \in J} of y∼mn+1Xy \sim m^X_{n+1} have law (mnX)⊗J(m^X_n)^{\otimes J}.✦

An array object MM over mXm^X consists of nonempty standard Borel spaces YY of states and UU of parameters 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}, one for each nn, such that each mnm_n is invariant under TYT_Y and carried by xx to mnXm^X_n, both acting entrywise, the family is consistent and proper as above, it is conditionally local, in that for J≠IJ \neq I and y∼mn+1y \sim m_{n+1} the planes πn,jy\pi_{n,j}\,y, j∈Jj \in J, are conditionally independent given the parameter array par(y)\mathrm{par}(y), with conditional laws mn(⋅∣πn,jpar(y))m_n(\cdot \mid \pi_{n,j}\,\mathrm{par}(y)), where mn(⋅∣u)m_n(\cdot \mid u) is the conditional law under mnm_n of an array given that its parameter array is uu, and under m0m_0, a law on YY, the maps xx and par\mathrm{par} are independent.✦

An extension of MM by an array object NN over the same initial laws is a pair of measurable maps, ϕ\phi from the states of NN to those of MM and τ\tau from the parameters of NN to those of MM, that intertwine the automorphisms, satisfy xM∘ϕ=xNx_M \circ \phi = x_N and parM∘ϕ=τ∘parN\mathrm{par}_M \circ \phi = \tau \circ \mathrm{par}_N, and carry mnNm^N_n to mnMm^M_n with ϕ\phi acting entrywise, for every nn.✦ Pullback along it, F↦F∘ϕF \mapsto F \circ \phi, is a linear isometry from L2(mnM)L^2(m^M_n) to L2(mnN)L^2(m^N_n); all L2L^2 spaces here are complex.✦

On arrays of order n+1n+1 an array object has 2|I|+12|I|+1 data fields, sub-σ\sigma-algebras of the Borel σ\sigma-algebra 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}. Conditional expectation onto a data field is the orthogonal projection of L2(mn+1)L^2(m_{n+1}) onto the functions measurable for it.✦ An array object MM is saturated if these projections commute with pullback along every extension: for every extension ϕ\phi of MM by an array object NN, every nn, every data field ℱ\mathcal F and every F∈L2(mn+1M)F \in L^2(m^M_{n+1}), 𝔼[F∘ϕ|ℱN]=𝔼[F|ℱM]∘ϕ,\mathbb E\bigl[F \circ \phi \mathrel{\big|} \mathcal F^N\bigr] = \mathbb E\bigl[F \mathrel{\big|} \mathcal F^M\bigr] \circ \phi,(1)✦ where ℱM\mathcal F^M and ℱN\mathcal F^N are the fields of that kind for MM and for NN.

Joining Operators

For a measure-preserving map xx from a measure space (Ω,μ)(\Omega, \mu) to (Ω1,ν)(\Omega_1, \nu), the pullback VxF=F∘xV_x F = F \circ x is a linear isometry from L2(ν)L^2(\nu) to L2(μ)L^2(\mu).✦ For a second measure-preserving map zz from (Ω,μ)(\Omega, \mu) to (Ω2,ρ)(\Omega_2, \rho), the joining operator of xx and zz is the bounded operator Jx,z=Vx*Vz:L2(ρ)→L2(ν),J_{x,z} = V_x^{*} V_z \colon L^2(\rho) \to L^2(\nu),(2)✦ which sends FF to the conditional expectation 𝔼μ[F(z)∣x]\mathbb E_\mu[F(z) \mid x], read as a function on Ω1\Omega_1.

When z=(z1,z2)z = (z_1, z_2) carries μ\mu to a product η1⊗η2\eta_1 \otimes \eta_2, so that z1z_1 and z2z_2 are independent with laws η1\eta_1 and η2\eta_2, the contraction of g∈L2(η1)g \in L^2(\eta_1) and h∈L2(η2)h \in L^2(\eta_2) is C(g,h)=Jx,z(g⊗h)=𝔼μ[g(z1)h(z2)∣x],C(g, h) = J_{x,z}(g \otimes h) = \mathbb E_\mu[g(z_1)\, h(z_2) \mid x],(3)✦ where g⊗hg \otimes h is the function (s,t)↦g(s)h(t)(s, t) \mapsto g(s) h(t) in L2(η1⊗η2)L^2(\eta_1 \otimes \eta_2).

For a measure θ\theta on a space WW and a family (Fw)w∈W(F_w)_{w \in W} of bounded operators from a Hilbert space 𝒦\mathcal K to a Hilbert space ℋ\mathcal H, the essential range of the family is the closed subspace of ℋ\mathcal H ess ran⁡θF={v∈ℋ:Fw*v=0 for θ-almost every w}⟂.\operatorname{ess\,ran}_\theta F = \{v \in \mathcal H : F_w^{*} v = 0 \text{ for $\theta$-almost every } w\}^{\perp}.(4)✦

Cell Laws and Active Spaces

Fix an array object MM over mXm^X, an order nn and a parameter array ee of order n+2n+2, a point of UIn+2U^{I^{n+2}}. For a set RR of rows and a set SS of columns, the cell law λRS\lambda_{RS} at ee is the law of the block yRSy_{RS} when the rows π0,cy\pi_{0,c}\,y, c∈Rc \in R, are drawn independently, row cc from mn+1(⋅∣π0,ce)m_{n+1}(\cdot \mid \pi_{0,c}\,e).✦

For a set RR of rows, sets SS, S1S_1, S2S_2 of columns and a block bb of cells in R×S1R \times S_1, the row coupling at bb is the conditional law under mn+2m_{n+2} of the pair (yRS,yRS2)(y_{RS}, y_{RS_2}) given that par(y)=e\mathrm{par}(y) = e and yRS1=by_{RS_1} = b.✦ The row active space RowR(S∣S1,S2)\mathrm{Row}_R(S \mid S_1, S_2) is the essential range (4), for bb drawn from λRS1\lambda_{RS_1}, of the joining operators L2(λRS2)→L2(λRS)L^2(\lambda_{RS_2}) \to L^2(\lambda_{RS}) of the row couplings at bb.✦

For a set SS of columns, sets RR, R1R_1, R2R_2 of rows and a block bb of cells in R1×SR_1 \times S, the column coupling at bb is the law of the pair of blocks (yRS,yR2S)(y_{RS}, y_{R_2S}) when the columns πn+1,jy\pi_{n+1,j}\,y, j∈Sj \in S, are drawn independently, column jj from mn+1(⋅∣πn+1,je)m_{n+1}(\cdot \mid \pi_{n+1,j}\,e) conditioned on its cells in the rows of R1R_1 being those of bb.✦ The column active space ColS(R∣R1,R2)\mathrm{Col}_S(R \mid R_1, R_2) is the essential range, for bb drawn from λR1S\lambda_{R_1S}, of the joining operators L2(λR2S)→L2(λRS)L^2(\lambda_{R_2S}) \to L^2(\lambda_{RS}) of the column couplings at bb.✦

For sets RR, R1R_1, R2R_2 of rows and SS, S1S_1, S2S_2 of columns, the common active space of the block R×SR \times S is the closed subspace of L2(λRS)L^2(\lambda_{RS}) D(R∣R1,R2;S∣S1,S2)=RowR(S∣S1,S2)∩ColS(R∣R1,R2).D(R \mid R_1, R_2;\, S \mid S_1, S_2) = \mathrm{Row}_R(S \mid S_1, S_2) \cap \mathrm{Col}_S(R \mid R_1, R_2).(5)✦ At the corner and its four neighbours the theorem below uses DPP′=D(P∣A,B;P′∣A′,B′),DPA′=D(P∣A,B;A′∣P′,B′),DPB′=D(P∣A,B;B′∣P′,A′),DAP′=D(A∣P,B;P′∣A′,B′),DBP′=D(B∣P,A;P′∣A′,B′),\begin{aligned} D_{PP'} &= D(P \mid A, B;\, P' \mid A', B'), \\ D_{PA'} &= D(P \mid A, B;\, A' \mid P', B'), & D_{PB'} &= D(P \mid A, B;\, B' \mid P', A'), \\ D_{AP'} &= D(A \mid P, B;\, P' \mid A', B'), & D_{BP'} &= D(B \mid P, A;\, P' \mid A', B'), \end{aligned}(6) Write ΠPP′\Pi_{PP'}, ΠPA′\Pi_{PA'}, ΠPB′\Pi_{PB'}, ΠAP′\Pi_{AP'} and ΠBP′\Pi_{BP'} for the orthogonal projections of L2(λPP′)L^2(\lambda_{PP'}), L2(λPA′)L^2(\lambda_{PA'}), L2(λPB′)L^2(\lambda_{PB'}), L2(λAP′)L^2(\lambda_{AP'}) and L2(λBP′)L^2(\lambda_{BP'}) onto DPP′D_{PP'}, DPA′D_{PA'}, DPB′D_{PB'}, DAP′D_{AP'} and DBP′D_{BP'} respectively.

The Claim

Theorem 1. Let MM be a saturated array object over initial array laws of TT, let I=P⊔A⊔BI = P \sqcup A \sqcup B and I=P′⊔A′⊔B′I = P' \sqcup A' \sqcup B' be triple partitions of the rows and of the columns, and let n≥0n \geq 0. For almost every parameter array ee of order n+2n+2, with respect to the law of par(y)\mathrm{par}(y) for y∼mn+2y \sim m_{n+2}, the following hold, with cell laws and common active spaces taken at ee and yy drawn from mn+2(⋅∣e)m_{n+2}(\cdot \mid e).

  1. Suppose yPP′y_{PP'} has law λPP′\lambda_{PP'}, and yPA′y_{PA'} and yPB′y_{PB'} are independent with laws λPA′\lambda_{PA'} and λPB′\lambda_{PB'}, and let CrowC^{\mathrm{row}} be the contraction (3) with x=yPP′x = y_{PP'}, z1=yPA′z_1 = y_{PA'} and z2=yPB′z_2 = y_{PB'}. Then for all g∈L2(λPA′)g \in L^2(\lambda_{PA'}) and h∈L2(λPB′)h \in L^2(\lambda_{PB'}), ΠPP′Crow(g,h)=Crow(ΠPA′g,ΠPB′h).\Pi_{PP'}\, C^{\mathrm{row}}(g, h) = C^{\mathrm{row}}(\Pi_{PA'} g,\, \Pi_{PB'} h).(7)✦

  2. Suppose yPP′y_{PP'} has law λPP′\lambda_{PP'}, and yAP′y_{AP'} and yBP′y_{BP'} are independent with laws λAP′\lambda_{AP'} and λBP′\lambda_{BP'}, and let CcolC^{\mathrm{col}} be the contraction (3) with x=yPP′x = y_{PP'}, z1=yAP′z_1 = y_{AP'} and z2=yBP′z_2 = y_{BP'}. Then for all g∈L2(λAP′)g \in L^2(\lambda_{AP'}) and h∈L2(λBP′)h \in L^2(\lambda_{BP'}), ΠPP′Ccol(g,h)=Ccol(ΠAP′g,ΠBP′h).\Pi_{PP'}\, C^{\mathrm{col}}(g, h) = C^{\mathrm{col}}(\Pi_{AP'} g,\, \Pi_{BP'} h).(8)✦

  3. There are a countable set JJ and measurable maps α\alpha, β\beta, γ\gamma, δ\delta and χ\chi into the torus 𝕋J\mathbb T^J, defined on the blocks of cells in A×P′A \times P', B×P′B \times P', P×A′P \times A', P×B′P \times B' and P×P′P \times P', such that DPP′=L2(σ(χ),λPP′),χ(yPP′)=α(yAP′)β(yBP′)−1=γ(yPA′)δ(yPB′)−1,D_{PP'} = L^2(\sigma(\chi), \lambda_{PP'}), \qquad \chi(y_{PP'}) = \alpha(y_{AP'})\, \beta(y_{BP'})^{-1} = \gamma(y_{PA'})\, \delta(y_{PB'})^{-1},(9)✦ where L2(σ(χ),λPP′)L^2(\sigma(\chi), \lambda_{PP'}) is the space of functions in L2(λPP′)L^2(\lambda_{PP'}) measurable for the σ\sigma-algebra generated by χ\chi, and the identities hold for almost every yy.

Moreover, let η1\eta_1 and η2\eta_2 be measures, KK a bounded linear operator from L2(η1⊗η2)L^2(\eta_1 \otimes \eta_2) to a complex Hilbert space ℋ\mathcal H, DD a closed subspace of ℋ\mathcal H contained in the closure of the range of KK, and Q1Q_1, Q2Q_2 bounded operators on L2(η1)L^2(\eta_1) and L2(η2)L^2(\eta_2) such that ΠDK(g⊗h)=K(Q1g⊗Q2h)\Pi_D K(g \otimes h) = K(Q_1 g \otimes Q_2 h) for all gg and hh in L2(η1)L^2(\eta_1) and L2(η2)L^2(\eta_2) respectively, where ΠD\Pi_D is the orthogonal projection onto DD. Then span¯{K(Q1g⊗Q2h):g∈L2(η1),h∈L2(η2)}=D.\overline{\operatorname{span}}\,\{K(Q_1 g \otimes Q_2 h) : g \in L^2(\eta_1),\ h \in L^2(\eta_2)\} = D.(10)✦

In the paper, the result reads: for the contraction Cr(g,h)=𝔼[g(Ys)h(Yt)∣Yr]C_r(g, h) = \mathbb E[g(Y_s) h(Y_t) \mid Y_r] of a law on three cells, with YrY_r the corner yPP′y_{PP'}, YsY_s and YtY_t its two neighbours in its row or in its column of blocks, and Πr\Pi_r, Πs\Pi_s, Πt\Pi_t the projections onto their common active spaces, ΠrCr(g,h)=Cr(g,Πth)=Cr(Πsg,h)=Cr(Πsg,Πth)\Pi_r C_r(g, h) = C_r(g, \Pi_t h) = C_r(\Pi_s g, h) = C_r(\Pi_s g, \Pi_t h); contractions with inputs in these spaces span DPP′D_{PP'}; and DPP′=L2(σ(χ))D_{PP'} = L^2(\sigma(\chi)), so that each grouped intersection, the intersection (5) of a row and a column active space, is spanned by lines, circle-valued functions up to a constant phase, exact for both the law of the corner with its two neighbours in its row and the law of the corner with its two neighbours in its column.

What Is Checked

The claim is stated over a saturated array object MM over initial array laws of TT, triple partitions of the rows and of the columns, an order nn and almost every parameter array ee of order n+2n+2, and over the constructions at ee: the cell laws, the common active spaces (6) built from the row and column couplings, and the contractions (3) of the corner with its neighbours; part (10) is stated over measures η1\eta_1 and η2\eta_2, a bounded operator on L2(η1⊗η2)L^2(\eta_1 \otimes \eta_2) and a closed subspace of its target.

The library also proves that the entrywise state map of an extension is measure preserving from mnNm^N_n to mnMm^M_n at every order nn✦, and that g⊗hg \otimes h is square integrable for η1⊗η2\eta_1 \otimes \eta_2 whenever gg and hh are square integrable for η1\eta_1 and η2\eta_2.✦