A minimal residual method for large-scale trust-region subproblem

This paper was devoted to solving the large-scale trust-region subproblem (TRS), a core subproblem in nonlinear optimization with extensive applications in scientific computing fields such as the Lorentz eigenvalue problem, Tikhonov regularization, nonlinear least squares, and so on. The generalized Lanczos trust-region (GLTR) method is one of the most popular iterative algorithms for large-scale TRS. In this method, the approximate solution converged more slowly than the approximate Lagrange multiplier. Inspired by this characteristic and the generalized minimal residual method (GMRES) and minimal residual method (MINRES), we proposed a minimal residual method based on the GLTR framework (MRTR). The MRTR method fixed the approximate Lagrange multiplier from GLTR and sought the approximate solution that minimized the residual norm within the Krylov subspace. We established some theoretical results to characterize the relationship between the approximate solutions of MRTR and GLTR. The experimental results showed that the MRTR method outperformed the GLTR method in terms of iteration numbers and CPU running time, and its residual norm convergence curve was smoother with better numerical stability, while achieving approximate optimal values with the same high precision as GLTR.