The Setting
Deleting a centroid from an optimal Gaussian partition decreases the support function of its centroid family. The Gaussian average of this decrease is the deletion loss. Its bounds constrain the probability of a cell in terms of its centroid norm.
Let be standard Gaussian measure on , with mean zero and identity covariance.✦ A Borel Gaussian partition is a family of Borel sets such that -almost every point belongs to exactly one cell.✦ Empty cells are allowed. The unnormalized Gaussian centroid, cell probability and centroid energy are (1) An active label of a Borel Gaussian partition is a cell index with . The active count of is the number of its active labels.✦ An optimal Borel Gaussian partition of is a partition that maximizes among Borel Gaussian partitions of with labels.✦ A minimal optimal Borel Gaussian partition of is an optimal partition with the least active count among optimal Borel Gaussian partitions of with labels.✦
Write . When , the normalized centroid lengths associated with (1) and the propeller constant are (2) For , put . The support function after deletion of and the deletion loss are (3)
For an indexed family in a real inner product space , its positive halfspace cone consists of the vectors whose inner product with every is nonnegative. For , its translated cone consists of the vectors for which , and its translated Gaussian probability is the Gaussian measure of that set; (4) writes these definitions for and . (4) For such a cone, denotes its unnormalized Gaussian centroid.
The unnormalized Gaussian density , its upper tail and the Gaussian normalizing constant are (5) The standard Gaussian density , distribution and upper tail associated with (5) are (6) The Gaussian quantile is the inverse of the Gaussian cumulative distribution function (CDF) on interior probabilities, extended by zero outside that interval: (7) The inverse Mills ratio is the ratio of the Gaussian density to its upper tail, as given in (8). With in the paper’s convention and a real quantile value, the reflected argument is , and the scalar Gaussian loss is (8) The function in (8) is the inverse Mills derivative in the paper’s convention.
The Claim
Theorem 1. Let and let be an optimal Borel Gaussian partition of . Its deletion loss from (3) satisfies (9)✦
Let be a finite family in a finite-dimensional real Euclidean space with sum zero and for distinct indices. For each there is a unit gap translation direction satisfying (10)✦
For the cone probability in (4) and every , the probability along the translated line has derivative (11)✦ Suppose there is with for every . The quantile probability is concave: (12)✦ For every and , its supporting tangent gives (13)✦
Suppose now that is minimal optimal, , and every cell has positive probability. Set , using (7). Then (14)✦ If , the normalization (2) and the scalar loss (8) satisfy each of the following constraints: (15)✦ (16)✦
In the paper, the result reads: for the extremizer with active cells, fix , let and ; the deletion loss is at least ; there is in the centroid span with for all and ; if , then and is concave; integrating its supporting tangent gives deletion loss at most , where ; hence .
What Is Checked
The partition bounds (9), (14), (15) and (16) are stated over Borel Gaussian partitions, with minimal optimality and positive cell probabilities where specified. The translation statement (10) concerns a finite family with negative pairwise inner products and zero sum. The cone statements (11), (12) and (13) concern intersections of positive halfspaces, with a common strictly positive direction for the quantile and tangent bounds.
The Gaussian CDF in (6) is strictly increasing✦ and has range .✦