Separable spherical constraints and the decrease of a quadratic function in the gradient projection step

Abstract

We examine the decrease of a strictly convex quadratic function along the projected-gradient path and show that our earlier estimates obtained for the bound constraints are valid for more general feasible sets including those defined by separable spherical constraints. The result is useful for the development of in a sense optimal algorithms for the solution of some QPQC problems with separable constraints and is an important ingredient in the development of scalable algorithms for contact problems with friction.