A Superlinearly Convergent Sequential Quadratically Constrained Quadratic Programming Algorithm for Degenerate Nonlinear Programming

We present an algorithm that achieves superlinear convergence for nonlinear programs satisfying the Mangasarian--Fromovitz constraint qualification and the quadratic growth condition. This convergence result is obtained despite the potential lack of a locally convex augmented Lagrangian. The algorithm solves a succession of subproblems that have quadratic objectives and quadratic constraints, both possibly nonconvex. By the use of a trust-region constraint we guarantee that any stationary point of the subproblem induces superlinear convergence, which avoids the problem of computing a global minimum. We compare this algorithm with sequential quadratic programming algorithms on several degenerate nonlinear programs.