The evaluation complexity of finding high-order minimizers of nonconvex optimization
Coralia Cartis
Mathematical Institute, University of Oxford, Woodstock Road, Oxford OX2 6GG, UKNicholas I. M. Gould
Computational Mathematics Group, STFC-Rutherford Appleton Laboratory, Chilton OX11 0QX, UKPhilippe L. Toint
Namur Centre for Complex Systems (naXys), University of Namur, 61, rue de Bruxelles, B-5000 Namur, Belgium
This book chapter is published open access.
Abstract
We introduce the concept of strong high-order approximate minimizers of nonconvex optimization problems. These apply in both standard smooth and composite nonsmooth settings, and additionally allow convex or inexpensive constraints. An adaptive regularization algorithm is then proposed to find such approximate minimizers. Under suitable Lipschitz continuity assumptions, the evaluation complexity of this algorithm is investigated. The bounds obtained not only provide, to the best of our knowledge, the first known result for (unconstrained or inexpensively-constrained) composite problems for optimality orders exceeding one, but also give the first sharp bounds for high-order strong approximate th order minimizers of standard (unconstrained and inexpensively constrained) smooth problems, thereby complementing known results for weak minimizers.