Sparsification of sums with respect to convex cones
Abstract
Let $x_1,x_2,\ldots,x_m$ be elements of a convex cone $K$ such that their sum, $e$, is in the relative interior of $K$. An $ε$-sparsification of the sum involves taking a subset of the $x_i$ and reweighting them by positive scalars, so that the resulting sum is $ε$-close to $e$, where error is measured in a relative sense with respect to the order induced by $K$. This generalizes the influential spectral sparsification model for sums of positive semidefinite matrices. This paper introduces and studies the sparsification function of a convex cone, which measures, in the worst case over all possible sums from the cone, the smallest size of an $ε$-sparsifier. The linear-sized spectral sparsification theorem of Batson, Spielman, and Srivastava can be viewed as a bound on the sparsification function of the cone of positive semidefinite matrices. This result is generalized to a family of convex cones (including all hyperbolicity cones) that admit a $ν$-logarithmically homogeneous self-concordant barrier with certain additional properties. For these cones, the sparsification function is bounded above by $\lceil4ν/ε^2\rceil$. For general convex cones that only admit an ordinary $ν$-logarithmically homogeneous self-concordant barrier, the sparsification function is bounded above by $\lceil(4ν/ε)^2\rceil$. Furthermore, the paper explores how sparsification functions interact with various convex geometric operations (such as conic lifts), and describes implications of sparsification with respect to cones for certain conic optimization problems.