English

Backward error analysis of the Lanczos bidiagonalization with reorthogonalization

Numerical Analysis 2022-10-20 v1 Numerical Analysis

Abstract

The kk-step Lanczos bidiagonalization reduces a matrix ARm×nA\in\mathbb{R}^{m\times n} into a bidiagonal form BkR(k+1)×kB_k\in\mathbb{R}^{(k+1)\times k} while generates two orthonormal matrices Uk+1Rm×(k+1)U_{k+1}\in\mathbb{R}^{m\times (k+1)} and Vk+1Rn×(k+1)V_{k+1}\in\mathbb{R}^{n\times {(k+1)}}. However, any practical implementation of the algorithm suffers from loss of orthogonality of Uk+1U_{k+1} and Vk+1V_{k+1} due to the presence of rounding errors, and several reorthogonalization strategies are proposed to maintain some level of orthogonality. In this paper, by writing various reorthogonalization strategies in a general form we make a backward error analysis of the Lanczos bidiagonalization with reorthogonalization (LBRO). Our results show that the computed BkB_k by the kk-step LBRO of AA with starting vector bb is the exact one generated by the kk-step Lanczos bidiagonalization of A+EA+E with starting vector b+δbb+\delta_{b} (denoted by LB(A+E,b+δbA+E,b+\delta_{b})), where the 2-norm of perturbation vector/matrix δb\delta_{b} and EE depend on the roundoff unit and orthogonality levels of Uk+1U_{k+1} and Vk+1V_{k+1}. The results also show that the 2-norm of Uk+1Uˉk+1U_{k+1}-\bar{U}_{k+1} and Vk+1Vˉk+1V_{k+1}-\bar{V}_{k+1} are controlled by the orthogonality levels of Uk+1U_{k+1} and Vk+1V_{k+1}, respectively, where Uˉk+1\bar{U}_{k+1} and Vˉk+1\bar{V}_{k+1} are the two orthonormal matrices generated by the kk-step LB(A+E,b+δbA+E,b+\delta_{b}) in exact arithmetic. Thus the kk-step LBRO is mixed forward-backward stable as long as the orthogonality of Uk+1U_{k+1} and Vk+1V_{k+1} are good enough. We use this result to investigate the backward stability of LBRO based SVD computation algorithm and LSQR algorithm. Numerical experiments are made to confirm our results.

Cite

@article{arxiv.2210.10297,
  title  = {Backward error analysis of the Lanczos bidiagonalization with reorthogonalization},
  author = {Haibo Li and Guangming Tan and Tong Zhao},
  journal= {arXiv preprint arXiv:2210.10297},
  year   = {2022}
}
R2 v1 2026-06-28T03:58:05.117Z