Paper detail

The Regularization Theory of the Krylov Iterative Solvers LSQR, CGLS, LSMR and CGME For Linear Discrete Ill-Posed Problems

For the large-scale linear discrete ill-posed problem $\min\|Ax-b\|$ or $Ax=b$ with $b$ contaminated by a white noise, Lanczos bidiagonalization based LSQR and its mathematically equivalent CGLS are most commonly used. They have intrinsic regularizing effects, where the number $k$ of iterations plays the role of regularization parameter. However, hitherto there has been no answer to the long-standing fundamental concern of Björck and Eldén in 1979: {\em for which kinds of problems LSQR and CGLS can find best possible regularized solutions}? Here a best possible regularized solution means that it is at least as accurate as the best regularized solution obtained by the truncated singular value decomposition (TSVD) method or by standard-form Tikhonov regularization and cannot be improved under certain conditions. In this paper we make a detailed analysis on the regularization of LSQR for severely, moderately and mildly ill-posed problems. For the first two kinds of problems, we prove that LSQR finds best possible solutions at semi-convergence and the following results hold until semi-convergence: (i) the $k$-step Lanczos bidiagonalization always generates a near best rank $k$ approximation to $A$; (ii) the $k$ Ritz values always approximate the first $k$ large singular values of $A$ in natural order; (iii) the $k$-step LSQR always captures the $k$ dominant SVD components of $A$; (iv) the diagonals and subdiagonals of the bidiagonal matrices generated by Lanczos bidiagonalization decay as fast as the singular values of $A$. However, for the third kind of problem, the above results do not hold generally. We also analyze the regularization of the other two Krylov solvers LSMR and CGME, proving that LSMR has similar regularizing effects to LSQR for each kind of problem and both are superior to CGME. Numerical experiments confirm our theory on LSQR.