Effective condition number and its applications (Q5962003)

For a linear system of equations \(Ax = b\), with a nonsingular \(n\times n\) matrix \(A\), it is well known that the standard condition number of \(A\) might overestimate the actual sensitivity of the solution \(x\) with respect to perturbations, in particular if the right-hand side \(b\) is perturbed to a much larger extent than \(A\). In such a setting, the so called effective condition number \(\|b\| / ( \sigma_{\min} \|x\| )\) can be expected to provide tighter estimates. The paper under review addresses an analogous question for overdetermined linear systems of the form \(Fx = b\) with an \(m\times n\) matrix \(F\), \(m\geq n\). The notion of effective condition number is extended and leads to an expression which requires the computation of the left singular vectors and singular values of \(F\). A forward error bound based on this effective condition number is derived, which is demonstrated to be much tighter than the traditional bound for the so called Motz problem discretized by a Trefftz method.