Lower-order penalization approach to nonlinear semidefinite programming (Q995945)

In this paper a nonlinear semidefinite programming problem is reformulated as an optimization problem with a matrix equality constraint. Then this problem is solved thanks to a lower-order penalty function. The advantage of such a function is that it requires weaker conditions than the usual \(l_1\) penalization. Furthermore the smallest exact penalty parameter of lower-order penalty functions is much smaller that that of \(l_1\) penalty functions. This type of penalty methods has already been applied for solving nonlinear programming problems and programs with complementarity constraints. In a first section, necessary and sufficient conditions are given to ensure the global exactness of the lower-order penalty functions. Then the convergence of the optimal values of the penalty problems is established. Next, since the penalty functions may not be smooth or even locally Lipschitz, the Ekeland variational principle is used to derive necessary optimality conditions for the penalty problems. Finally, under the Robinson condition for semidefinite programming problems, it is shown that any limit point of the penalty problems is a KKT stationary point of the original problem.