The Setting
A Gaussian partition assigns a centroid to each cell. At an energy maximum with as few positive-measure cells as possible, the geometry concerns the angles between these centroids and the regions on which each centroid has maximal linear score.
The space has its usual Euclidean inner product, and is standard Gaussian measure, with mean zero and identity covariance. A Gaussian measurable partition of is a family of Borel sets for which -almost every point belongs to exactly one cell.✦ Empty cells are allowed. The unnormalized Gaussian centroid of a set is the vector in , and the centroid energy of a finite labelled family of subsets of is the real number, given in (1): (1)
An optimal partition is a Gaussian measurable partition that maximizes among Gaussian measurable partitions in the same dimension with the same number of labels; write for its value.✦ For a finite labelled family of subsets of , the active labels form the set of labels of positive Gaussian measure, their number is the active count , and their centroids span the subspace of given in (2): (2) A minimal optimal partition is an optimal partition whose active count is no larger than that of any optimal partition with the same dimension and labels.✦
For a finite vector family in , the closed maximizing cone is the set on which has maximal score, as defined in (3): (3)✦
Normalized Configurations
The constant is defined in (4): (4)
For vectors , the negative projection coefficient and orthogonal residual are the scalar and vector in (5): (5)
Write and for the standard Gaussian density and distribution function on . The unnormalized density , its upper tail and the Gaussian normalizing constant are defined in (6): (6) The inverse Mills ratio and the inverse Mills derivative , written with the paper’s argument convention, are the functions on given in (7); the reflected argument is : (7) The Gaussian upper-tail probability and Gaussian loss are the functions defined by (8): (8)
A normalized obtuse configuration of size in is a pair of families and of vectors and real weights satisfying (9), (10) and (11), together with (12) when ; write and for the set of these configurations. (9) (10) (11) (12)✦
The Claim
Theorem 1. Let be a minimal optimal Gaussian partition, with active labels , active count and active centroid span . For distinct , (13)✦ If is nonempty, then (14)✦
Let be a finite family in a real inner product space, with for distinct , and fix . Then (15)✦ If also , then (16)✦ For any finite family in a real vector space whose sum is zero, and any , (17)✦
If the minimal optimal partition has , then for each , (18)✦
For every Gaussian measurable partition with labels, (19)✦
For normalized obtuse configurations in , (20)✦
In the paper, the result reads: if , and an optimal partition of with labels has value and is chosen to minimize the number of positive-measure cells, then, after its null cells are discarded and the remaining centroids are denoted , distinct centroids have negative inner product, their span has dimension , every proper subfamily is linearly independent, for each the vectors with form a basis of , its cells agree -almost everywhere with , and .
What Is Checked
The partition assertions (13), (14), (18) and (19) are stated over Borel Gaussian partitions with fixed dimension and labels; (15) and (16) concern finite families in real inner product spaces under their stated hypotheses, (17) concerns finite zero-sum families in real vector spaces, and (20) concerns the normalized obtuse configurations defined above.