Iterative refinement enhances the stability of \(QR\) factorization methods for solving linear equations (Q1176472): Difference between revisions

The use of iterative refinement in order to improve a computed solution \(\hat x\) of a linear system of equations \(Ax=b\) is an important technique in numerical analysis. Basically, it goes as follows: Firstly, one computes the residue \(r=b-A\hat x\). Secondly, one solves \(Ad=r\). Finally, one updates the computed solution with \(y=\hat x+d\). This process is repeated as long as necessary. The method of iterative refinement is traditionally used in conjunction with Gaussian elimination for the solution of ill-conditioned systems. In this case, one usually computes \(\hat x\) in working precision, and performs the iterative refinement in double precision. However, in the past few years iterative refinement in the working precision has become more widely used. It is under this framework of fixed precision that this paper is written. The author shows that an arbitrary linear equation solver can be made stable in the strong, componentwise sense of \textit{R. D. Skeel} [Math. Comput. 35, 817-832 (1980; Zbl 0441.65027)]\ by means of fixed precision iterative refinement provided certain conditions are satisfied. Amongst the consequences of the results in this paper are interesting applications to QR factorizations, least-square problems, and Vandermonde-like systems.