The Setting
Fix , an even integer , nonzero vectors in the coordinate space of real -tuples, and widths . Each pair cuts out the slab , and the slabs meet in a symmetric convex polyhedron. The feature of the th slab is the polynomial (1) and is a soft indicator of the slab, sharper as grows.
The soft slab law is the measure on whose density is proportional to the product of these soft indicators with the Gaussian factor : its soft slab potential , its partition function and itself are (2)
The question is how much a linear combination of the features, with real coefficients , can fluctuate under , measured against the means of the features. The library bounds this fluctuation for , and in every dimension when has moment coordinates, a notion defined in the section Moment Coordinates below.
Means and Variances
The mean and the variance of a real function on under are the numbers (3)
For real polynomials on , real rates and a real parameter , the regularized potential is the function , and its Gibbs density, partition function, Gibbs mean and Gibbs variance are formed from it as , , and are formed from . For and all rates zero, the Gibbs density, partition function, Gibbs mean and Gibbs variance of the regularized potential are , , and .
For a nonnegative integer , the derivative polynomial of order of the polynomials with rates and parameter is the polynomial on , the th derivative in of .✦ The polynomial variation of order at a point is the value at of the derivative polynomial of order .✦
Moment Coordinates
For a function on and a vector , the directional derivative is the derivative of at applied to ✦, and the partial derivative is taken along the th standard basis vector ✦. The gradient of at is the vector ✦, and the Hessian of at is the matrix ✦.
The weighted measure of a function on is , Lebesgue measure in the source coordinates with the weight .✦ The function solves the Monge-Ampère equation for when (4) A smooth function on gives moment coordinates for when it solves (4), its Hessian is positive definite and is positive semidefinite at every point , and its gradient map is a homeomorphism of onto itself, with smooth inverse, that pushes forward to , making the moment measure of in the target coordinates .
The Claim
Theorem 1 (Variance bound). Let , let be even, let be nonzero vectors in and positive widths, and let and be the features and the soft slab law of (1) and (2). Let be real numbers.
(Soft variance inequality.) If and , then (5)✦
(Moment variance bound.) If a function on gives moment coordinates for , then (6)✦
What Is Checked
Both bounds are stated over the features and the linear forms of (1); the soft variance inequality (5) is stated over the Gibbs mean and Gibbs variance of the regularized potential with all rates zero, with its Gibbs density, partition function and polynomial variation, and the moment variance bound (6) over the soft slab potential , the law , and a function giving moment coordinates, with its gradient, Hessian and weighted measure .