An Exact Penalty Method for Nonconvex Problems Covering, in Particular, Nonlinear Programming, Semidefinite Programming, and Second-Order Cone Programming

This paper presents an exact penalty method for solving optimization problems with very general constraints covering, in particular, nonlinear programming (NLP), semidefinite programming (SDP), and second-order cone programming (SOCP). The algorithm is called the sequential linear cone method (SLCM) because for SDP and SOCP the main cost of computation amounts to solving at each iteration a linear cone program for which efficient solvers are available. Restricted to NLP, SLCM is exactly a sequential quadratic program method. Under two basic conditions which concern only the data, it is proved that the sequence of iterates is bounded. Furthermore, in particular, when the feasible set is nonempty, under two additional constraint qualification conditions, it is proved that the cluster points are stationary points. In that case, it is established also that the sequence of penalty parameters eventually stays constant, and for a particular class of data it is proved that a unit step length can be obtained.