Superlinearly convergent trust-region method without the assumption of positive-definite Hessian (Q2370038)

The authors study a modified version of the trust-region method for solving unconstrained minimization problems. The modification concerns the up-date rule for the trust-region radius. The radius depends on the gradient and tends to zero. Also for this method global convergence is shown under common assumptions. Moreover, superlinear convergence is proven under (slightly) weaker conditions than in earlier papers. The assumption of the positive definiteness of the Hessian is replaced by a (slightly weaker, but more abstract) condition based on an error bound for the distance of the actual iterate from the solution set. Numerical experiments are presented.