METHODS FOR SPARSE AND LOW-RANK RECOVERY UNDER SIMPLEX CONSTRAINTS

The de facto standard approach of promoting sparsity by means of l(1)-regularization becomes ineffective in the presence of simplex constraints, that is, when the target is known to have non-negative entries summing to a given constant. The situation is analogous for the use of nuclear norm regularization for the low-rank recovery of Hermitian positive semidefinite matrices with a given trace. In the present paper, we discuss several strategies to deal with this situation, from simple to more complex. First, we consider empirical risk minimization (ERM), which has similar theoretical properties w.r.t. prediction and l(2)-estimation error as l(1)-regularization. In light of this, we argue that ERM combined with a subsequent sparsification step (e.g., thresholding) represents a sound alternative to the heuristic of using l(1)-regularization after dropping the sum constraint and the subsequent normalization. Next, we show that any sparsity-promoting regularizer under simplex constraints cannot be convex. A novel sparsity-promoting regularization scheme based on the inverse or negative of the squared l(2)-norm is proposed, which avoids the shortcomings of various alternative methods from the literature. Our approach naturally extends to Hermitian positive semidefinite matrices with a given trace.