Unbiased approaches to compressive sensing and low rank matrix estimation
Marcus Carlsson, Lund University
16 December 2024
It is well known that the standard compressive sensing techniques, LASSO for the scalar case and nuclear norm minimization for the matrix case, come with a significant “shrinking bias”. I will introduce a toolbox of techniques to construct non-convex penalties with a number of desirable features, in particular that they provably do not suffer from the “shrinking bias”. The types of penalites that one may construct can be tailormade to the application at hand, and works both for vectors and matrices.