On the duality gap in nonconvex optimization