Iterative refinement for constrained and weighted linear least squares (Q1334996)

The behavior of mixed precision iterative refinement applied to constrained and weighted least squares problems is analyzed. It is shown that the initial convergence rate depends on the spectral radius of two matrices whose elements are closely related to the condition of the problem. Computational results show that it is sufficient to consider only one of these matrices. It is pointed out that the method works very well when equations corresponding to large weights are ill-conditioned and the coefficient matrix itself is ill-conditioned.