The evaluation complexity of finding high-order minimizers of nonconvex optimization

  • Coralia Cartis

    Mathematical Institute, University of Oxford, Woodstock Road, Oxford OX2 6GG, UK
  • Nicholas I. M. Gould

    Computational Mathematics Group, STFC-Rutherford Appleton Laboratory, Chilton OX11 0QX, UK
  • Philippe L. Toint

    Namur Centre for Complex Systems (naXys), University of Namur, 61, rue de Bruxelles, B-5000 Namur, Belgium
The evaluation complexity of finding high-order minimizers of nonconvex optimization cover
Download Chapter PDF

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.