Convergence properties of trust region methods for linear and convex constraints (Q922951)

The identification property is crucial for this theory. It is proved first for the projected gradient method and later generalized for the trust region method. Global convergence is established for general convex constraints. For the case of linear constraints the superlinear convergence rate is proved. The mainly geometric convergence theory is based on a sufficient reduction in the trust region subproblem in each iteration.