Tensor Methods for Large, Sparse Unconstrained Optimization (Q4377571)

Tensor methods for unconstrained optimization were first introduced by \textit{R. B. Schnabel} and \textit{T.-T. Chow} [SIAM J. Optim. 1, No. 3, 293-315 (1991; Zbl 0758.65047)]. The author extends these methods to large, sparse problems. Especially he gives algorithms for an efficient finding of a minimizer of the fourth-order tensor approximation of the objective function when the Hessian is sparse. In the line search the new algorithm tries the full tensor step first. If this provides enough decrease in the objective function then the search is terminated; otherwise the Newton direction is calculated additionally and the algorithm selects the direction with the lower function value as the next iterate. A two-dimensional trust region approach is used for robustness. Test results are presented for sets of problems where the Hessian at the minimizer is nonsingular and where it has the rank deficiency 1 or 2. The numbers of gradient and function evaluations and the execution times are reported for the tensor and the Newton method. These results show that tensor methods are more efficient and more reliable than Newton's method.