Error estimation in preconditioned conjugate gradients (Q2490362)

The authors deal with the problem of convergence in the preconditioned conjugate gradient method for the solution of a linear system \(Ax=b\), where \(A\) is a symmetric positive definite \(n\times n\) matrix. Indeed, the problem of this convergence was the subject of many papers in literature. However, they always assume exact arithmetic and consequently, they assume preserving orthogonality and exploiting the finite termination property (i.e. getting the exact solution in a finite number of steps, which does not exceed the dimension \(n\) of the problem). Unfortunately, most practical computations violate these assumptions. Taking this fact into account, the authors focus on estimating the \(A\)-norm of the error and present a practical estimate for such norm. It is simple and numerically stable. Eventually, they propose to combine their results with the standard quantities, which are already used, to establish a convenient stopping criteria. Some nice numerical experiments are reported to illustrate how this new estimate works.