Dense Initializations in Large-Scale Quasi-Newton Trust-Region Methods

This talk proposes dense initial quasi-Newton matrices for large-scale optimization methods. Based on an implicit eigendecomposition of large quasi-Newton matrices, we develop novel dense initial matrices. The proposed approach is specifically advantageous in quasi-Newton trust-region methods, because the partial eigendecomposition of quasi-Newton matrices allows for efficient solutions of trust-region subproblems. The proposed method is tested and compared on standard large-scale problems from the CUTEst collection.