On the Efficient Global Dynamics of Newton's Method for Complex Polynomials

TL;DR

This paper analyzes the theoretical and practical efficiency of Newton's method for finding all roots of complex polynomials, demonstrating near-optimal complexity and exceptional performance on very high-degree polynomials.

Contribution

It provides a near-complete theoretical complexity analysis of Newton's method for polynomial root finding and presents a practical implementation that outperforms existing methods on large degrees.

Findings

Newton's method finds all roots with near-optimal complexity.

Implementation successfully finds roots of polynomials with degrees over 100 million.

Expected total iterations are polynomial in degree, with improved bounds under certain conditions.

Abstract

We investigate Newton's method as a root finder for complex polynomials of arbitrary degree. While polynomial root finding continues to be one of the fundamental tasks of computing, with essential use in all areas of theoretical mathematics, numerics, computer graphics and physics, known methods have either excellent theoretical complexity but cannot be used in practice, or are practically efficient but are a lacking a successful theory behind them. In this manuscript we describe the theoretical complexity of Newton's method for finding all roots of polynomials of given degree and show that it is near-optimal for the known set of starting points that find all roots. This theoretical result is complemented by a recent implementation of Newton's method that finds all roots of various polynomials of degree more than a million, significantly faster than our upper bounds on the complexity indicate, and often much faster than established fast root finders. In some experiments, it was possible to find all roots even with complexity $O(d\log d)$ for degrees exceeding 100 million. Newton's method thus stands out as a method that has merits both from the theoretical and from the practical point of view. Our study is based on the known explicit set of universal starting points that is guaranteed to find all roots of polynomials of degree $d$ (appropriately normalized). We show that this set contains $d$ points that converge very quickly to the $d$ roots: the expected total number of Newton iterations required to find all roots with precision $\eps$ is $O(d^3\log^3d+d\log|\log\eps|)$, which can be further improved to $O(d^2\log^4d+d\log|\log\eps|)$; in the worst case allowing near-multiple roots, the complexity is $O(d^4\log^2d+d^3\log^2d|\log\eps|)$. The arithmetic complexity for all these estimates is the same as the number of Newton iterations steps, up to a factor of $\log^2 d$.