On inner iterations of the joint bidiagonalization based algorithms for solving large scale ill-posed problems

TL;DR

This paper investigates the inner iteration process of joint bidiagonalization algorithms for large-scale ill-posed problems, proposing a stopping criterion that balances solution accuracy and computational efficiency.

Contribution

It introduces a practical stopping criterion for inner least squares problems, enabling relaxed accuracy requirements without compromising the regularized solution quality.

Findings

Relaxed inner problem accuracy does not affect solution quality for moderate noise levels.

Proposed stopping criterion improves overall algorithm efficiency.

Numerical experiments confirm the effectiveness of the new approach.

Abstract

The joint bidiagonalization process of a matrix pair $\{A,L\}$ can be used to develop iterative regularization algorithms for large scale ill-posed problems in general-form Tikhonov regularization $\min_x\left\{\|Ax-b\|_{2}^2+\lambda^2\|Lx\|_{2}^2\right\}$ or the essentially equivalent one $\min\|Lx\|_{2} \ \mbox{{\rm s.t.}} \ x\in\mathcal{S} = \{x|\ \|Ax-b\|_{2}\leq \eta\|e\|_{2}\}$, where $e$ is a Gaussian white noise, $L$ is a regularization matrix and $\eta>1$ slightly. A bottleneck of the algorithms is that a large scale inner least squares problem with $(A^{T}, L^{T})^{T}$ as the coefficient matrix must be solved at each outer iteration, which may be costly, especially when the solution accuracy of these problems is high. In this paper, we give a detailed investigation on the solution accuracy requirement on the inner least squares problems and propose a practical stopping criterion for inner iterations. The results show that for ill-posed problems with not too small noise levels, the solution accuracy of the inner least squares problems can be relaxed considerably while it will not reduce the accuracy of regularized solutions, thus the overall efficiency of the algorithms can be improved substantially. Numerical experiments are made to illustrate our results and show some numerical performance of the algorithms.