arXiv · 2611.15709
Trust-region high-index saddle dynamics for ill-conditioned saddle-point computation
Abstract
We propose a trust-region high-index saddle dynamics (TRHiSD) framework for computing high-index saddle points in ill-conditioned problems. To address the deterioration of high-index saddle dynamics caused by inappropriate step sizes and search directions, we exploit its intrinsic local min--max structure and construct separate trust-region subproblems on the maximization and minimization subspaces to adapt the two trust-region radii to different local curvature scales. Under suitable assumptions and with exact spectral subspaces, we prove local linear convergence of the Cauchy-point variant and local quadratic convergence of the Newton-step variant. Numerical experiments on modified Rosenbrock functions are consistent with the predicted convergence rates. For the Landau--de Gennes model, a TRHiSD variant using a truncated conjugate-gradient step for the minimization subproblem substantially reduces the outer-iteration count and the wall-clock time in the tested cases.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Guanghua Ji, Longjing Li, Zhen Xu, Jianyuan Yin. 2026-11-11. Trust-region high-index saddle dynamics for ill-conditioned saddle-point computation. https://arxiv.science/abs/2611.15709
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.