Non-stationary extremal eigenvalue approximations in iterative solutions of linear systems and estimators for relative error

TL;DR

This paper introduces efficient estimators for extremal eigenvalues and relative errors in iterative solutions of linear systems, improving stopping criteria and accuracy without extensive computation.

Contribution

It presents novel estimators for extremal eigenvalues and relative errors in CG methods that do not require heavy computation or prior eigenvalue knowledge.

Findings

Estimators accurately approximate extremal eigenvalues from CG iterates.

Relative error estimators are effective in A-norm and l2 norm.

Numerical results demonstrate improved stopping criteria for large problems.

Abstract

Non-stationary approximations of the final value of a converging sequence are discussed, and we show that extremal eigenvalues can be reasonably estimated from the CG iterates without much computation at all. We introduce estimators of relative error for conjugate gradient (CG)methods that adopt past work on computationally efficient bounds of the absolute errors using quadrature formulas. The evaluation of the Gauss quadrature based estimates though, depends on a priori knowledge of extremal eigenvalues; and the upper bounds in particular that are useful as a stopping criterion fail in the absence of a reasonable underestimate of smallest eigenvalue. Estimators for relative errors in A-norm and their extension to errors in l2 norm are presented with numerical results. Estimating the relative error from the residue in an iterative solution is required for efficient solution of a large problem with even a moderately high condition. Specifically, in a problem of solving for vector x in Ax=b, the uncertainty between the strict upper bound in relative error [{\kappa}*||r(i)||/||b||] and its strict lower bound [||r(i)||/({\kappa}*||b||)] is a factor of {\kappa}^2 (given residue r(i)= b-Ax(i) is the residual vector at ith iteration and {\kappa} the condition number of the square matrix A).