Some Results on Regularization of LSQR and CGLS for Large-Scale Discrete Ill-Posed Problems
Provides theoretical justification for the regularization behavior of LSQR for different ill-posedness levels, guiding method choice for practitioners solving large-scale discrete ill-posed problems.
The paper establishes quantitative bounds showing that for severely and moderately ill-posed problems, the Krylov subspace captures dominant right singular vectors, explaining LSQR's regularization, but not for mildly ill-posed problems, suggesting hybrid LSQR should be used.
For large-scale discrete ill-posed problems, LSQR, a Lanczos bidiagonalization process based Krylov method, is most often used. It is well known that LSQR has natural regularizing properties, where the number of iterations plays the role of the regularization parameter. In this paper, for severely and moderately ill-posed problems, we establish quantitative bounds for the distance between the $k$-dimensional Krylov subspace and the subspace spanned by $k$ dominant right singular vectors. They show that the $k$-dimensional Krylov subspace may capture the $k$ dominant right singular vectors for severely and moderately ill-posed problems, but it seems not the case for mildly ill-posed problems. These results should be the first step towards to estimating the accuracy of the rank-$k$ approximation generated by Lanczos bidiagonalization. We also derive some other results, which help further understand the regularization effects of LSQR. We draw to a conclusion that a hybrid LSQR should generally be used for mildly ill-posed problems. We report numerical experiments to confirm our theory. We present more definitive and general observed phenomena, which will derive more research.