Abstract
The LSQR iterative method is one of the most popular methods for solving large-scale linear discrete ill-posed problem $Ax=b$ with an error-contaminated right-hand side.
In this paper, we consider the regularization properties of range restricted LSQR (RRLSQR) method. The iteration number $k$ always acts as the regularization parameter because of the semi-convergence. In order to verify whether or not the RRLSQR method finds a 2-norm filtering best regularization solution for severely, moderately and mildly ill-posed problems, we present the $sin \Theta$ theorems for the 2-norm distances between the $k$ dimensional left and right Krylov subspaces generated by Lanczos bidiagonalization and the $k$ dimensional dominant left and right singular subspaces of $A$, and estimate the distances for the three kinds problems assuming that the singular values are simple, and develop a regularized RRLSQR method for solving linear discrete ill-posed problems. Numerical experiments confirm our theoretical results and show the efficiency of the proposed method.
', './data/cmcma/coversheet/cover_en.jpg'))