A Riemannian Subspace BFGS Trust Region Method
Wei Hejie,Yang Wei Hong,Chai Yinsheng
DOI: https://doi.org/10.1007/s11590-022-01964-9
IF: 1.5288
2023-01-01
Optimization Letters
Abstract:Motivated by the subspace techniques in the Euclidean space, this paper presents a subspace BFGS trust region (RTR-SBFGS) algorithm to the problem of minimizing a smooth function defined on Riemannian manifolds. In each iteration of the RTR-SBFGS algorithm, a low-dimensional trust region subproblem is solved, which reduces the amount of computation significantly for large scale problems. A limited-memory variant of RTR-SBFGS, named LRTR-SBFGS, is introduced also. Both RTR-SBFGS and LRTR-SBFGS are proved to converge globally. Under some mild conditions, we establish the local linear convergence of these two methods. Numerical results demonstrate that, compared to the state-of-the-art algorithms, RTR-SBFGS and LRTR-SBFGS are effective methods and subspace techniques are suitable for Riemannian optimization problems.