《1999 The Efficiency of Ballstep Subgradient Level Methods for Convex Optimization》.pdf

MATHEMATICS OF OPERATIONS RESEARCH Vol. 24, No. 1, February 1999 Printed in U.S.A. THE EFFICIENCY OF BALLSTEP SUBGRADIENT LEVEL METHODS FOR CONVEX OPTIMIZATION ¨ KRZYSZTOF C. KIWIEL, TORBJORN LARSSON, AND P. O. LINDBERG We study subgradient methods for convex optimization that use projections onto successive approximations of level sets of the objective corresponding to estimates of the optimal value. We establish convergence and efficiency estimates for simple ballstep level controls without requiring that the feasible set be compact. Our framework may handle accelerations based on “cheap” projections, surrogate constraints, and conjugate subgradient techniques. 1. Introduction. We consider methods for the minimization problemf inf f under * S n n the following assumptions. S is a nonempty closed convex set in ,f : 3 is a convex function, for each x S we can computef (x ) and a subgradient g (x ) f (x ) off at x , and f for each x n we can find P x arg min |x y |, its orthogonal projection on S , where S y S | | is the Euclidean norm. The optimal set S * Arg minS f may be empty. Given the kth iterate x k S and a target level f k that estimates f , we may use