Rank-structured QR for Chebyshev rootfinding

TL;DR

This paper introduces a rank-structured QR algorithm for Chebyshev polynomial rootfinding that is efficient, parallelizable, and provides improved backward stability guarantees compared to traditional methods.

Contribution

It extends the QR iteration with an aggressive deflation strategy exploiting rank structure, achieving quadratic complexity and better backward stability for Chebyshev rootfinding.

Findings

Achieves quadratic complexity and linear storage requirements.

Guarantees small backward error up to an explicit amplification factor.

Demonstrates improved accuracy in numerical tests.

Abstract

We consider the computation of roots of polynomials expressed in the Chebyshev basis. We extend the QR iteration presented in [Eidelman, Y., Gemignani, L., and Gohberg, I., Numer. Algorithms, 47.3 (2008): pp. 253-273] introducing an aggressive early deflation strategy, and showing that the rank-structure allows to parallelize the algorithm avoiding data dependencies which would be present in the unstructured QR. We exploit the particular structure of the colleague linearization to achieve quadratic complexity and linear storage requirements. The (unbalanced) QR iteration used for Chebyshev rootfinding does not guarantee backward stability on the polynomial coefficients, unless the vector of coefficients satisfy $\|p\| \approx 1$, an hypothesis which is almost never verified for polynomials approximating smooth functions. Even though the presented method is mathematically equivalent to the unbalanced QR algorithm, we show that exploiting the rank structure allows to guarantee a small backward error on the polynomial, up to an explicitly computable amplification factor $\hat\gamma_1(p)$, which depends on the polynomial under consideration. We show that this parameter is almost always of moderate size, making the method accurate on several numerical tests, in contrast with what happens in the unstructured unbalanced QR.