The Setting
Take an array object and draw an array of order from its law . Read the first axis of as rows and the last as columns, so that the cell of in row and column is an array of order . Split the columns into groups , , and choose groups , , of rows, with disjoint from and from . The four blocks of cells , , and 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 spaces. A row side joins the two blocks in the same rows, given the parameters of and the cells of those rows in the columns of ; a column side joins the two blocks in the same columns, given the parameters and the cells of the rows of in those columns. The question is whether the two ways around the square from to give the same operator, and whether the two ways from to 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 and the columns of , 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 spaces are complex. The pullback isometry of a measure-preserving map from to is the operator , .✦ The joining operator of measure-preserving maps to and to on the same space is the bounded operator (1)✦ which sends to the conditional expectation of given , read on . The family of joining operators of a kernel from a space to , with reference measures and , assigns to each the joining operator of the coupling , 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 , , and between complex Hilbert spaces such that (2)✦
Array Objects
Fix a finite index set . An array of order over a space is a function from the index tuples to ; these arrays form the product ✦, and those of order zero are the points of ✦. A map acts on arrays entrywise, .✦ For an array of order , an axis and an index , the plane is the array of order obtained by fixing axis of at ✦; the planes of along a subset of are its planes , , on the last axis✦.
Initial array laws of a measurable automorphism of a nonempty standard Borel space are probability measures on , one for each , each invariant under , that are consistent and proper: for every axis , every index and every subset of , (3)✦
An array object over is a pair of nonempty standard Borel spaces, a state space and a parameter space , with automorphisms and , measurable maps and with and , and probability measures on , , such that each is invariant under and carried to by , the family is consistent and proper as in (3), the maps and are independent under read on , and for every subset of and almost every (4)✦ where is the parameter array of , the array applied entrywise, and is the conditional law under of an array given that its parameter array is . The kernel is the conditional law of at order .✦
At each order an array object has data fields, sub--algebras of the Borel sets of : the parameter field and, for each , the outside field and the enlarged field . The data kinds are the labels of the data fields , and , . The projection onto the data of kind is the conditional expectation onto that field, the orthogonal projection of onto the functions measurable for it.✦
An extension of by an array object over the same initial laws is a pair of measurable maps, from the state space of to that of and from the parameter space of to that of , that intertwine the automorphisms, satisfy and , and have for every , where and are the laws of order of and of , and , , , are their maps and .✦ The pullback along an extension at order is the linear isometry , .✦ An array object is saturated if its projections onto data commute with every such pullback: for every extension of by an array object , every , every data kind and every , (5)✦
The Grouped Square
Let be an array of order over the state space of . Its row and its column are arrays of order , and its cell in row and column is an array of order . For a set of rows and a set of columns, is the family of cells , , .✦ The row restriction of rows to the columns of is the family ✦, and the column restriction of columns to the rows of is the family ✦, both indexed by .
A group datum for rows and columns is a pair of a parameter array of order and a family of cells indexed by ✦; the grouped data map sends to ✦. The parent data of for a set of rows are and the parent rows , , and its whole group data for and a set of columns add the columns , : (6) For a whole group datum with parameter array , its group selection for rows is with the column restriction of its columns to the rows of ✦, and its column own data for columns are with the row restriction of its parent rows to the columns of ✦; at these are and .
The row family kernel for a set of rows is the product kernel from parameter arrays of order to families of rows indexed by : (7) The law of cells , the group marginal at for rows and columns , is the image of under the row restriction to .✦ For column groups , , , the row pair kernel sends a group datum for and to the conditional law under of the pair given .✦
The posterior kernel for a set of rows sends a parameter array of order and arrays of order , , to the conditional law of an array drawn from given that its planes are the : the observation kernel, the joint law of these planes and , disintegrated along the planes.✦ The column group kernel for and a set of columns is the product kernel from group data for and to families of columns indexed by , (8)✦ and the column pair kernel is its image under the column restrictions to the rows of and to the rows of ✦.
The row operator for at a group datum for and is the joining operator from to .✦ The column operator for at a group datum for and is from to .✦ The row operator for at a group datum for and is from to , and the column operator for at a group datum for and is from to . At a fixed parameter array , the operators and , as varies, are the fibre row and column operators at .
The Claim
Theorem 1. Let be a saturated array object over initial array laws, and let . Let , , be sets of rows, with disjoint from and from , , , , and neither nor equal to . Let , , be pairwise disjoint sets of columns with union , and nonempty.
For almost every value of the whole group data (6) for and of an array drawn from , the row operators and at the group selections of for and , and the column operators and at its column own data for and , satisfy (9)✦
For almost every parameter array of an array drawn from , and for -almost every array , the fibre operators at satisfy (10)✦
Let and be measure-preserving maps from a probability space to measures and , let and be measure-preserving maps from to and , and and from to and . Let , , , be Markov kernels such that, under , the conditional laws of and given are and , and those of and given are and . Then, with taken under and under , (11)✦
Let , , , and be measurable maps on a probability space that has a disintegration along , and let , , , be measures on the value spaces of , , , . Suppose that given the pairs , , , have conditional laws , , , , which for almost every are couplings of and , of and , of and , and of and . Suppose that given and the conditional law of depends only on and , and that of only on and ; and that given and the conditional law of depends only on and , and that of only on and . Then for almost every value of the joining operators of these couplings, with reference measures among , , , , satisfy (12)✦
In the paper, the result reads: for the grouped square, almost surely and , in every ordering of the groups, with , , and standing for the operators written here , , and .
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.