LSMR: An Iterative Algorithm for Sparse Least-Squares Problems
SIAM Journal on Scientific Computing
| Hi-index | 0.00 |
This thesis is concerned with backward perturbation analyses of the linear least squares (LS) and related problems. Two theoretical measures are commonly used for assessing the backward errors that arise in the approximate solution of such problems. These are called the normwise relative backward error (NRBE) and the minimal backward error (MBE). An important new relationship between these two measures is presented, which shows that the two are essentially equivalent. New upper bounds on the NRBE and MBE for the LS problem are given and related to known bounds and estimates. One important use of backward perturbation analysis is to design stopping criteria for iterative methods. In this thesis, minimum-residual iterative methods for solving LS problems are studied. Unexpected convergence behaviour in these methods is explained and applied to show that commonly used stopping criteria can in some situations be much too conservative. More reliable stopping criteria are then proposed, along with an efficient implementation in the iterative algorithm LSQR. In many applications the data in the LS problem come from a statistical linear model in which the noise follows a multivariate normal distribution whose mean is zero and whose covariance matrix is the scaled identity matrix. A description is given of typical convergence of the error that arises in minimum-residual iterative methods when the data come from such a linear model. Stopping criteria that use the information from the linear model are then proposed and compared to others that appear in the literature. Finally, some of these ideas are extended to the scaled total least squares problem.