Regularization properties of LSQR for linear discrete ill-posed problems in the multiple singular value case and best, near best and general low rank approximations**The work was supported by the National Science Foundation of China (No. 11771249).

Abstract

For the large-scale linear discrete ill-posed problem min‖Ax − b‖ or Ax = b with b contaminated by white noise, the Golub–Kahan bidiagonalization based LSQR method and its mathematically equivalent CGLS, the conjugate gradient (CG) method applied to ATAx = ATb, are most commonly used. They have intrinsic regularizing effects, where the iteration number k plays the role of regularization parameter. The long-standing fundamental question is: Can LSQR and CGLS find two-norm filtering best possible regularized solutions? The author has given definitive answers to this question for severely and moderately ill-posed problems when the singular values of A are simple. This paper extends the results to the multiple singular value case, and studies the approximation accuracy of Krylov subspaces, the quality of low rank approximations generated by Golub–Kahan bidiagonalization and the convergence properties of Ritz values. For the two kinds of problems, we prove that LSQR finds two-norm filtering best possible regularized solutions at semi-convergence. Particularly, we consider some important and untouched issues on best, near best and general rank k approximations to A for the ill-posed problems with the singular values with α > 0, and the relationships between them and their nonzero singular values. Numerical experiments confirm our theory. The results on general rank k approximations and the properties of their nonzero singular values apply to several Krylov solvers, including LSQR, CGME, MINRES, MR-II, GMRES and RRGMRES.