The Setting
Let be a finite index set and let be a smooth function on whose Hessian is positive definite at every point, so that is strictly convex. Its gradient map carries the measure on the source side to a measure on the target side. When that measure is for a smooth potential and a normalizing constant , it is the moment measure of , and the Monge-Ampère equation behind it reads, for a constant , (1) Here is the source point and the target point, is read at , and .
A test function , smooth with compact support, gives the matrix field , a vector field obtained by one weighted divergence, and the function , where is the weighted divergence for . The question is how, with the mean under , the energy of the Jacobian of in target coordinates and the energy compare with the transported Hessian energy , in which the Hessian of the potential enters.
Derivatives and Weighted Integrals
The coordinate space is the real vector space of all functions from to , added and scaled coordinatewise✦, and its standard basis vector is the point with coordinate at and at every other index.✦
The directional derivative of a function on a real normed space along a vector of is the function whose value at is the Fréchet derivative of at applied to .✦ The directional derivative of a matrix field on is the matrix field whose entries are the directional derivatives of the entries of .✦
On , the partial derivative of a function is the directional derivative ✦, the gradient is the vector with coordinates ✦, and the Hessian is the matrix with entry in row and column .✦ The Jacobian of a vector field is the matrix with entry in row and column .✦ For an index , the Hessian derivative of is the matrix field , with entries .✦
The reference measure on is the measure written on the source side and on the target side. The weight of a potential is the positive function ✦, and the weighted expectation of a function is the integral of against it.✦ For a constant , the normalized expectation of is (2)✦ the mean of under when .
For a potential on and a vector , the weighted adjoint is the operator that sends a function to .✦ The weighted divergence of a family of functions along a family of vectors is the function ; for a vector field on , along the standard basis, it is (3)✦ and with in place of it is the divergence for the source measure.
The Fields of a Test Function
Let be a real function on . In source coordinates, with , the Hessian of pulled by the gradient is the matrix field ✦, and the moment coefficients are the entries of : (4)✦ The moment divergence is the vector field whose component is the divergence of the row , (5)✦ and is its Jacobian in source coordinates.✦ The pulled matrix is the matrix field ✦ and the quadratic matrix is the matrix field .✦
For a homeomorphism of , a function of the source point is read in target coordinates as the target function .✦ The inverse Hessian field is the matrix field of the target point✦; the transported matrix field and the transported field are the pulled matrix and the moment divergence read in target coordinates, and , and is the Jacobian of the latter. With , the transported Hessian energy of is the function of .✦
For a vector field on , the -Hessian energy of is ✦ and the square trace of its Jacobian is ✦.
In source coordinates, the first trace term at a point is the trace of the product , where .✦ The trace term at the same point is the trace of the square of the product of the Jacobian with the inverse Hessian of .✦
The cross term , the derivative term of and the cubic term are the following functions of the source point (6)
The Claim
Theorem 1. Let , and be smooth functions on .
Suppose that has compact support, that the Hessian of is positive definite at every point, that solves (1) for a constant , and that is a homeomorphism of . Then, with and as fields of the source point, (7)✦
Suppose in addition that is smooth and that pushes forward to the target measure for a constant . Then, with the normalized expectation (2), , , and read at the target point, and , (8)✦
For every vector field on whose components are smooth with compact support, with , (9)✦
If the Hessian of is nonsingular at every point and solves (1) for a constant , then, with and as fields of the source point, for all indices and and at every point, (10)✦
If the Hessian of is positive definite at every point, then at every point (11)✦
In the paper, the result reads: for smooth probability measures and with positive definite and a smooth diffeomorphism pushing to , every real satisfies and , with , , , , and .
What Is Checked
The claim is stated over smooth functions , and on the coordinate space , the Monge-Ampère equation (1), the gradient map of as a homeomorphism relating the source and target measures, the weighted divergence (3) with the weighted and normalized expectations, and the fields built from : the moment coefficients (4), the moment divergence (5), the pulled, quadratic and inverse Hessian matrices with their transported versions, and the terms (6).