The Setting
Fix , an even integer , nonzero vectors of the coordinate space and widths . The features are the polynomials obtained as powers of the linear forms ; the soft slab potential adds them to a Gaussian term, and the soft slab law is the probability measure with Gibbs density normalized by the partition function : (1) Means and variances are taken under : (2)
The question is how much a combination of the features, with real coefficients , can fluctuate under , measured against the means of the features themselves; Proposition 2.1 of the OpenAI preprint on the logarithmic Brunn-Minkowski conjecture answers it with a variance inequality. On the dual side, a function is tested against the features through its correlations ; the dual bound is the bound (9) on these correlations for the functions . Here is a test function, a smooth function with compact support, and is a second-order operator defined below from moment coordinates for .
Moment Coordinates
The partial derivative of a function on is its directional derivative along the standard basis vector , where ✦; the gradient collects the ✦, and the Hessian is the matrix with entries .✦
Moment coordinates for are a smooth function on and a homeomorphism of with smooth inverse such that ; at every point the Hessian is positive definite and is positive semidefinite; solves the Monge-Ampère equation (3) and pushes the weighted source measure forward to .
The source variable carries and the target variable carries ; a function of is read in target coordinates as the target function .✦ The moment matrix of moment coordinates , is the matrix-valued function on given by the Hessian of read in target coordinates, .✦
The Operator and Its Tests
An operator term on is a pair made of a smooth real function on , the coefficient, and a finite list of vectors of , the directions; it acts on a function as , the coefficient times the iterated directional derivative, where stands for .✦ The differential operator of a finite list of terms is the operator on functions that sends to the sum of the actions of on .✦ The operator is the differential operator of the drift terms , one for each coordinate, and the principal terms , one for each pair of coordinates: (4)
The test functions on form a linear space.✦ On it, sends to ✦, and the range of in is the linear subspace (5)✦ where is the class in of a test function .
Weighted Divergence and the Field
For a potential on , with weight , the weighted adjoint of sends a function to ✦, and the weighted divergence of a vector field is (6)✦ For the weighted divergence is the divergence of , and for it is the divergence of .
Given a test function , the moment coefficients are the entries of the matrix , (7)✦ The moment divergence is the divergence under of the th row✦, and the transported field is its reading in target coordinates, .✦ The function is .
The Claim
Theorem 1. Let and be moment coordinates for the soft slab law . For every test function , the function built from satisfies (8)✦ and every in the range of in satisfies (9)✦
Theorem 2. Let be a real inner product space, vectors of and . If for every in a dense subset of , then for all real (10)✦ Let be a measure, square integrable functions under , and real, and let , with centered function . If , then (11)✦
Theorem 3. Let be the soft slab law (1) and real numbers. If and are moment coordinates for , then (12)✦ For the soft slab law on , , and real numbers , (13)✦
What Is Checked
The statements are made over the soft slab law (1) with its features, mean and variance (2); over moment coordinates , and the moment matrix ; over the operator of (4) on test functions, the range of in , the weighted divergence (6) and the transported field ; and, in (10) and (11), over any real inner product space and any measure.
About these constructions the library proves that has smooth entries whenever and are smooth✦, and that is a linear map of the test functions into themselves.✦ It also proves that a test function is square integrable under any measure that is finite on compact sets.