Nuclear norm minimization and the affine rank minimization problem

Brian Borchers
(New Mexico Tech)

ABSTRACT:

In the affine rank minimization problem, we want to find a matrix X of minimum rank that satisfies a linear matrix equation AX=B.  In general, this is a very hard problem, since the rank of X is a nonconvex function of X.  The nuclear norm of a matrix X is the sum of the singular values of X.  The nuclear norm is a convex function of X, and efficient algorithms exist for the minimization of the nuclear norm of X subject to linear constraints on the elements of X.  It is surprising that in many cases it is possible to recover a minimum rank solution to AX=B by minimizing the nuclear norm of X subject to the linear constraints.  This talk will introduce the affine rank minimization problem, its applications, and the nuclear norm minimization approach to solving the affine rank minimization problem.