Min-Max Elementwise Backward Error for Roots of Polynomials and a Corresponding Backward Stable Root Finder
A new measure called min-max elementwise backward error is introduced for approximate roots of scalar polynomials $p(z)$. Compared with the elementwise relative backward error, this new measure allows for larger relative perturbations on the coefficients of $p(z)$ that do not participate much in the overall backward error. By how much these coefficients can be perturbed is determined via an associated max-times polynomial and its tropical roots. An algorithm is designed for computing the roots of $p(z)$. It uses a companion linearization $C(z) = A-zB$ of $p(z)$ to which we added an extra zero leading coefficient, and an appropriate two-sided diagonal scaling that balances $A$ and makes $B$ graded in particular when there is variation in the magnitude of the coefficients of $p(z)$. An implementation of the QZ algorithm with a strict deflation criterion for eigenvalues at infinity is then used to obtain approximations to the roots of $p(z)$. Under the assumption that this implementation of the QZ algorithm exhibits a graded backward error when $B$ is graded, we prove that our new algorithm is min-max elementwise backward stable. Several numerical experiments show the superior performance of the new algorithm compared with the MATLAB \texttt{roots} function. Extending the algorithm to polynomial eigenvalue problems leads to a new polynomial eigensolver that exhibits excellent numerical behaviour compared with other existing polynomial eigensolvers, as illustrated by many numerical tests.