OctoInform Beta

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

Lemma 4.6 [Normalization of the polar rectangle]: For almost every parameter array, fixed unitaries match the four polar laws in pairs, with common support a closed abelian unitary group

Polar rectangles of a saturated object normalize into one abelian group.

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

Split the rows and the columns of a random array into three blocks each. Given the array of parameters of the entries and the cells of one block, the cells of a row block that lie in the two other column blocks form a joining, and so do the cells of a column block that lie in the two other row blocks; the Markov operators of these joinings make a square of four operator families between four L2L^2 spaces. The question is whether fixed identifications bring the unitary parts of their polar decompositions into one commutative group of unitaries.

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 of YY for every index tuple in InI^n✦; a plane of an array of order n+1n+1 is the array of order nn obtained by holding one of its indices at a slot i∈Ii\in I.✦ The initial array laws of TT are probability measures mnXm^X_n on XInX^{I^n}, one for each order nn, invariant under TT applied entrywise, consistent and proper in the sense defined below for the laws of an array object.✦ An array object MM over the initial array laws mnXm^X_n of TT is a standard Borel space YY of states with an automorphism TYT_Y, an equivariant factor map to XX and an equivariant parameter map par\mathrm{par} to a standard Borel space UU with an automorphism TUT_U, together with probability measures mnm_n on YInY^{I^n}, invariant under TYT_Y applied entrywise, carried to mnXm^X_n by the factor map, and consistent, proper and conditionally local in the sense defined below.✦ The laws mnm_n of an array object are consistent if every plane of an array of order n+1n+1 has law mnm_n under mn+1m_{n+1}, and proper if, under mn+1m_{n+1}, the planes along the last axis at the slots of a proper subset of II are independent. The laws mnm_n are conditionally local if, given the parameter array, obtained by applying par\mathrm{par} entrywise, the planes along the last axis at the slots of a proper subset of II stay independent, each with the conditional law of an array of order nn given its own parameters. Under m0m_0, whose arrays are points of YY, the factor map and par\mathrm{par} are independent.

An extension of MM is an array object NN over the same initial laws with measurable equivariant maps from its states and its parameters to those of MM, compatible with the factor and parameter maps, whose state map, applied entrywise, carries the law mnNm^N_n of NN to mnm_n at every order.✦ For an extension NN of MM, write ι\iota for composition with the entrywise state map, the pullback along it; this is a linear isometry from L2(mn)L^2(m_n) to L2(mnN)L^2(m^N_n).✦ The data fields of an array object MM are σ\sigma-algebras of arrays of order n+1n+1: the parameter field 𝒫n+1\mathcal P_{n+1}, the outside fields 𝒪n+1,j\mathcal O_{n+1,j} and the enlarged fields 𝒢n+1,j\mathcal G_{n+1,j}, with jj the index of the field. An array object MM is saturated if, for every extension NN of MM and every order nn, ιΠ=ΠNι\iota\,\Pi = \Pi^N\iota on L2(mn+1)L^2(m_{n+1}) for each data field, where Π\Pi and ΠN\Pi^N are the orthogonal projections onto the functions measurable for that data field of MM and of NN.✦

For a parameter array uu of order nn, write mnum^u_n for the conditional law under mnm_n of an array given that its parameter array is uu, and m‾n\bar m_n for the law of the parameter array under mnm_n.

A triple partition of II is an ordered triple of nonempty, pairwise disjoint subsets of II with union II.✦ Call the first index of an array of order n+2n+2 its row index and the last its column index, and fix triple partitions I=P⊔A⊔BI = P\sqcup A\sqcup B of the rows and I=P′⊔A′⊔B′I = P'\sqcup A'\sqcup B' of the columns, and an order nn. In an array ω\omega of order n+2n+2, the cell in row rr and column jj is the array of order nn obtained by holding the first index at rr and the last at jj✦, and ωR,S\omega_{R,S} is the family of cells with row in RR and column in SS.✦ The square reads ω\omega through four blocks: x=ωA,P′,y=ωB,P′,c=ωP,A′,d=ωP,B′.x = \omega_{A,P'},\qquad y = \omega_{B,P'},\qquad c = \omega_{P,A'},\qquad d = \omega_{P,B'}.(1)

Fix a parameter array ee of order n+2n+2. The cell law mR,Sem^e_{R,S} is the law of ωR,S\omega_{R,S} when the rows ωr\omega_r, r∈Rr\in R, arrays of order n+1n+1 obtained by holding the first index at rr, are drawn independently, each from mn+1erm^{e_r}_{n+1} with ere_r the row of ee at rr.✦ The row joining operator Fx:L2(mA,B′e)→L2(mA,A′e)F_x\colon L^2(m^e_{A,B'})\to L^2(m^e_{A,A'}) is the Markov operator of the joining of ωA,A′\omega_{A,A'} and ωA,B′\omega_{A,B'} given by their conditional law under mn+2m_{n+2}, given that the parameter array is ee and that the block ωA,P′\omega_{A,P'} is xx; it takes a function of the second member to its conditional expectation given the first. The row joining operator Gy:L2(mB,B′e)→L2(mB,A′e)G_y\colon L^2(m^e_{B,B'})\to L^2(m^e_{B,A'}) is the Markov operator of the joining of ωB,A′\omega_{B,A'} and ωB,B′\omega_{B,B'} given by their conditional law under mn+2m_{n+2}, given that the parameter array is ee and that the block ωB,P′\omega_{B,P'} is yy; it takes a function of the second member to its conditional expectation given the first.

Along the columns, given ee and cc, the columns ωj\omega^j, j∈A′j\in A', arrays of order n+1n+1 obtained by holding the last index at jj, are drawn independently, each from the conditional law under mn+1ejm^{e^j}_{n+1} given that its cells in the rows of PP are those of cc, with eje^j the column of ee at jj.✦ The column joining operator Kc:L2(mB,A′e)→L2(mA,A′e)K_c\colon L^2(m^e_{B,A'})\to L^2(m^e_{A,A'}) is the Markov operator of the resulting joining of ωA,A′\omega_{A,A'} and ωB,A′\omega_{B,A'}. The column joining operator Nd:L2(mB,B′e)→L2(mA,B′e)N_d\colon L^2(m^e_{B,B'})\to L^2(m^e_{A,B'}) is the Markov operator of the joining of ωA,B′\omega_{A,B'} and ωB,B′\omega_{B,B'} obtained from the columns ωj\omega^j, j∈B′j\in B', drawn independently, each from the conditional law under mn+1ejm^{e^j}_{n+1} given that its cells in the rows of PP are those of dd.

Polar Rectangles

With H1=L2(mA,A′e)H_1 = L^2(m^e_{A,A'}), H2=L2(mA,B′e)H_2 = L^2(m^e_{A,B'}), H3=L2(mB,A′e)H_3 = L^2(m^e_{B,A'}) and H4=L2(mB,B′e)H_4 = L^2(m^e_{B,B'}), the four families form a square: H4→NdH2→FxH1,H4→GyH3→KcH1.H_4 \xrightarrow{\;N_d\;} H_2 \xrightarrow{\;F_x\;} H_1,\qquad H_4 \xrightarrow{\;G_y\;} H_3 \xrightarrow{\;K_c\;} H_1.(2)

The essential kernel of a family s↦Qss\mapsto Q_s of bounded operators from EE to HH on a measure space is the subspace of vectors v∈Ev\in E with Qsv=0Q_sv = 0 for almost every ss✦, and its essential range ran⁡Q\operatorname{ran}Q is the orthogonal complement in HH of the essential kernel of the adjoint family s↦Qs*s\mapsto Q_s^*.✦ Write FF, GG, KK, NN for the families x↦Fxx\mapsto F_x, y↦Gyy\mapsto G_y, c↦Kcc\mapsto K_c, d↦Ndd\mapsto N_d. Under the cell laws μx=mA,P′e\mu_x = m^e_{A,P'}, μy=mB,P′e\mu_y = m^e_{B,P'}, μc=mP,A′e\mu_c = m^e_{P,A'} and μd=mP,B′e\mu_d = m^e_{P,B'} of the four blocks, the rectangle spaces D1,…,D4D_1,\dots,D_4 are 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}(3)

The compression of a bounded operator Q:E→HQ\colon E\to H to closed subspaces V⊆EV\subseteq E and W⊆HW\subseteq H is the operator V→WV\to W that applies QQ and projects orthogonally onto WW; compressing gives families F̃x:D2→D1\tilde F_x\colon D_2\to D_1, G̃y:D4→D3\tilde G_y\colon D_4\to D_3, K̃c:D3→D1\tilde K_c\colon D_3\to D_1 and Ñd:D4→D2\tilde N_d\colon D_4\to D_2.✦

Write 𝒰(E,H)\mathcal U(E,H) for the unitary operators from a Hilbert space EE onto a Hilbert space HH, and 𝒰(H)=𝒰(H,H)\mathcal U(H) = \mathcal U(H,H), a topological group in the strong operator topology. A measurable map uu into 𝒰(E,H)\mathcal U(E,H) is a measurable polar decomposition of a family s↦Qss\mapsto Q_s of operators from EE to HH if, for almost every ss, the operator QsQ_s is injective with dense range and Qs=u(s)|Qs|Q_s = u(s)\,|Q_s|, where |Q|=(Q*Q)1/2|Q| = (Q^*Q)^{1/2}.✦

Let μ\mu be a measure on a space Ω\Omega, let x,y,c,dx, y, c, d be maps from Ω\Omega to spaces 𝒳,𝒴,𝒞,𝒟\mathcal X, \mathcal Y, \mathcal C, \mathcal D with measures μx,μy,μc,μd\mu_x, \mu_y, \mu_c, \mu_d, and let F,G,K,NF, G, K, N be operator families indexed by these spaces and placed between Hilbert spaces H1,…,H4H_1, \dots, H_4 as in (2). A measurable polar rectangle over these data is a choice of measurable polar decompositions aa of FF, bb of GG, c0c_0 of KK and d0d_0 of NN, under the respective measures, with a(x)d0(d)=c0(c)b(y)μ-almost everywhere on Ω.a(x)\,d_0(d) = c_0(c)\,b(y)\qquad \mu\text{-almost everywhere on }\Omega.(4)✦ The grouped square at ee is the square of the compressed families F̃,G̃,K̃,Ñ\tilde F, \tilde G, \tilde K, \tilde N between D1,…,D4D_1, \dots, D_4 for the fixed triple partitions. A polar rectangle of the grouped square at ee is a measurable polar rectangle over the compressed families, with Ω\Omega the space of arrays of order n+2n+2 under mn+2em^e_{n+2}, the blocks (1) and the cell laws μx,μy,μc,μd\mu_x, \mu_y, \mu_c, \mu_d.✦

Haar Normalization

In a group, LgL_g is the left translation h↦ghh\mapsto gh. For Hilbert spaces H1,…,H4H_1, \dots, H_4 and fixed unitaries α:H2→H1\alpha\colon H_2\to H_1, β:H4→H3\beta\colon H_4\to H_3 and δ:H4→H2\delta\colon H_4\to H_2, the normalization maps carry the four unitary spaces of the square into 𝒰(H4)\mathcal U(H_4): φa(u)=(αδ)−1uδ,φb(v)=β−1v,φc(u′)=(αδ)−1u′β,φd(v′)=δ−1v′\varphi_a(u) = (\alpha\delta)^{-1}u\,\delta,\qquad \varphi_b(v) = \beta^{-1}v,\qquad \varphi_c(u') = (\alpha\delta)^{-1}u'\beta,\qquad \varphi_d(v') = \delta^{-1}v'(5) for u∈𝒰(H2,H1)u\in\mathcal U(H_2,H_1), v∈𝒰(H4,H3)v\in\mathcal U(H_4,H_3), u′∈𝒰(H3,H1)u'\in\mathcal U(H_3,H_1) and v′∈𝒰(H4,H2)v'\in\mathcal U(H_4,H_2).

Probability measures ρa,ρb,ρc,ρd\rho_a, \rho_b, \rho_c, \rho_d on 𝒰(H2,H1)\mathcal U(H_2,H_1), 𝒰(H4,H3)\mathcal U(H_4,H_3), 𝒰(H3,H1)\mathcal U(H_3,H_1) and 𝒰(H4,H2)\mathcal U(H_4,H_2) have a Haar normalization if there are unitaries α,β,δ\alpha, \beta, \delta as above and a closed abelian subgroup 𝒢\mathcal G of 𝒰(H4)\mathcal U(H_4) with (φa)*ρa=(φb)*ρb=:ρ,(φc)*ρc=(φd)*ρd=:τ,supp⁡ρ=supp⁡τ=𝒢,(Lg)*ρ=ρ(g∈𝒢).\begin{gathered} (\varphi_a)_*\rho_a = (\varphi_b)_*\rho_b =: \rho,\qquad (\varphi_c)_*\rho_c = (\varphi_d)_*\rho_d =: \tau,\\ \operatorname{supp}\rho = \operatorname{supp}\tau = \mathcal G,\qquad (L_g)_*\rho = \rho\quad (g\in\mathcal G). \end{gathered}(6)✦ A measurable polar rectangle (a,b,c0,d0)(a, b, c_0, d_0) has a Haar normalization when the laws a*μxa_*\mu_x, b*μyb_*\mu_y, (c0)*μc(c_0)_*\mu_c and (d0)*μd(d_0)_*\mu_d of its four maps, its polar variables, have a Haar normalization.✦

The Claim

Theorem 1. (i) Let Γ\Gamma be a second-countable Hausdorff topological group in which the indicator of every closed set is a pointwise limit of continuous functions Γ→[0,1]\Gamma\to[0,1] equal to 11 on that set. Let σ\sigma be a measure on a measurable space ZZ, let κ\kappa and λ\lambda be s-finite kernels from ZZ to Γ×Γ\Gamma\times\Gamma, and let ρ\rho and τ\tau be probability measures on Γ\Gamma. Assume that for σ\sigma-almost every zz both marginals of κ(z)\kappa(z) are ρ\rho, both marginals of λ(z)\lambda(z) are τ\tau, and uv′=u′vfor κ(z)⊗λ(z)-almost every ((u,v),(u′,v′));u\,v' = u'\,v\qquad\text{for } \kappa(z)\otimes\lambda(z)\text{-almost every } \big((u,v),(u',v')\big);(7) that ∫Zκ(z)dσ(z)=ρ⊗ρ\int_Z\kappa(z)\,d\sigma(z) = \rho\otimes\rho and ∫Zλ(z)dσ(z)=τ⊗τ\int_Z\lambda(z)\,d\sigma(z) = \tau\otimes\tau; and that at some point z0z_0 the first marginals of κ(z0)\kappa(z_0) and λ(z0)\lambda(z_0) are ρ\rho and τ\tau, the pair (1,1)(1,1) lies in the supports of κ(z0)\kappa(z_0) and of λ(z0)\lambda(z_0), and (7) holds at z0z_0. Then there is a closed abelian subgroup 𝒢\mathcal G of Γ\Gamma with supp⁡ρ=supp⁡τ=𝒢,(Lg)*ρ=ρ(g∈𝒢).\operatorname{supp}\rho = \operatorname{supp}\tau = \mathcal G,\qquad (L_g)_*\rho = \rho\quad (g\in\mathcal G).(8)✦

(ii) Let H1,…,H4H_1, \dots, H_4 be complex Hilbert spaces, σ\sigma a probability measure on ZZ, κ\kappa a Markov kernel from ZZ to 𝒰(H2,H1)×𝒰(H4,H3)\mathcal U(H_2,H_1)\times\mathcal U(H_4,H_3) and λ\lambda one from ZZ to 𝒰(H3,H1)×𝒰(H4,H2)\mathcal U(H_3,H_1)\times\mathcal U(H_4,H_2), and ρa,ρb,ρc,ρd\rho_a, \rho_b, \rho_c, \rho_d probability measures on these four unitary spaces. Assume that for σ\sigma-almost every zz the marginals of κ(z)\kappa(z) are ρa\rho_a and ρb\rho_b, those of λ(z)\lambda(z) are ρc\rho_c and ρd\rho_d, and (7) holds with products read as compositions, and that ∫Zκ(z)dσ(z)=ρa⊗ρb,∫Zλ(z)dσ(z)=ρc⊗ρd.\int_Z\kappa(z)\,d\sigma(z) = \rho_a\otimes\rho_b,\qquad \int_Z\lambda(z)\,d\sigma(z) = \rho_c\otimes\rho_d.(9)✦ Then ρa,ρb,ρc,ρd\rho_a, \rho_b, \rho_c, \rho_d have a Haar normalization.✦

(iii) Let MM be a saturated array object, with triple partitions of the rows and of the columns and an order nn fixed as above. For m‾n+2\bar m_{n+2}-almost every parameter array ee of order n+2n+2, every polar rectangle of the grouped square at e has a Haar normalization.\text{every polar rectangle of the grouped square at } e \text{ has a Haar normalization.}(10)✦

(iv) Let ℋ\mathcal H be a complex Hilbert space, 𝒢\mathcal G a closed abelian subgroup of 𝒰(ℋ)\mathcal U(\mathcal H), and ρ\rho a probability measure on 𝒰(ℋ)\mathcal U(\mathcal H) with support 𝒢\mathcal G and (Lg)*ρ=ρ(L_g)_*\rho = \rho for every g∈𝒢g\in\mathcal G. Then there are a countable set JJ, a Hilbert basis (vj)j∈J(v_j)_{j\in J} of ℋ\mathcal H and continuous homomorphisms χj\chi_j from 𝒢\mathcal G to the unit circle 𝕋\mathbb T of ℂ\mathbb C, with its Borel σ\sigma-algebra, such that gvj=χj(g)vj(g∈𝒢,j∈J).g\,v_j = \chi_j(g)\,v_j\qquad (g\in\mathcal G,\ j\in J).(11)✦

(v) Let ℋ\mathcal H, 𝒢\mathcal G and ρ\rho be as in (iv), and let τ\tau be a probability measure on 𝒰(ℋ)\mathcal U(\mathcal H) with support 𝒢\mathcal G. Let μ\mu be a measure on a space Ω\Omega and x,y,c,d,zx, y, c, d, z maps from Ω\Omega to spaces 𝒳,𝒴,𝒞,𝒟,Z\mathcal X, \mathcal Y, \mathcal C, \mathcal D, Z. Let aa on 𝒳\mathcal X and bb on 𝒴\mathcal Y be measurable maps into 𝒰(ℋ)\mathcal U(\mathcal H) carrying measures νx\nu_x and νy\nu_y to ρ\rho, and c0c_0 on 𝒞\mathcal C and d0d_0 on 𝒟\mathcal D measurable maps into 𝒰(ℋ)\mathcal U(\mathcal H) carrying νc\nu_c and νd\nu_d to τ\tau. Let κ\kappa and λ\lambda be Markov kernels from ZZ to 𝒰(ℋ)×𝒰(ℋ)\mathcal U(\mathcal H)\times\mathcal U(\mathcal H), the conditional laws under μ\mu of (a(x),b(y))(a(x),b(y)) and of (c0(c),d0(d))(c_0(c),d_0(d)) given zz. Assume that for almost every tt under the law of zz both marginals of κ(t)\kappa(t) are ρ\rho, both marginals of λ(t)\lambda(t) are τ\tau, and (7) holds with tt for zz. Then there are a countable set JJ and measurable maps θx,θy,θc,θd,θz\theta_x, \theta_y, \theta_c, \theta_d, \theta_z from 𝒳,𝒴,𝒞,𝒟,Z\mathcal X, \mathcal Y, \mathcal C, \mathcal D, Z to 𝕋J\mathbb T^J such that the square-integrable functions of aa and those of θx\theta_x form the same closed subspace of L2(νx)L^2(\nu_x), those of bb and of θy\theta_y the same closed subspace of L2(νy)L^2(\nu_y), and, μ\mu-almost everywhere, θz(z)=θx(x)θy(y)−1=θc(c)θd(d)−1.\theta_z(z) = \theta_x(x)\,\theta_y(y)^{-1} = \theta_c(c)\,\theta_d(d)^{-1}.(12)✦

In the paper, the result reads: after fixed unitary identifications of the four spaces D1,…,D4D_1, \dots, D_4 with one Hilbert space ℋ\mathcal H, the four polar variables take values in one closed abelian subgroup 𝒢\mathcal G of 𝒰(ℋ)\mathcal U(\mathcal H) with common support 𝒢\mathcal G; aa and bb have a common left-invariant law ρ\rho; a(x)b(y)−1=c0(c)d0(d)−1=h(z)a(x)\,b(y)^{-1} = c_0(c)\,d_0(d)^{-1} = h(z), where hh is a function of the variable zz; 𝒢\mathcal G has a simultaneous orthonormal eigenbasis; and the monomials in its circle coordinates at h(z)h(z) span the square-integrable functions of h(z)h(z) and are exact for the row and column three-cell laws, the laws of three cells along a row and along a column.

What Is Checked

The claim is stated over saturated array objects with their laws and their conditional laws given parameters, triple partitions of the rows and the columns, the cell laws and the row and column joining operators of the grouped square, their compressions to the rectangle spaces, measurable polar decompositions and polar rectangles, and Haar normalization; its general parts are stated over kernels into a topological group or into spaces of unitaries, and over closed abelian groups of unitaries with a left-invariant law.

Beyond the claim, the library proves that the entrywise state map of an extension is measure-preserving from the law of the extension to the law of MM at every order✦, and that the essential kernel of an operator family is a linear subspace. It also proves that the essential range of an operator family is a closed subspace, so that the four rectangle spaces are closed subspaces.