Dierk Schleicher

Béla Bollobás, Malte Lackmann, Dierk Schleicher

We specify a small set, consisting of $O(d(\log\log d)^2)$ points, that intersects the basins under Newton's method of \emph{all} roots of \emph{all} (suitably normalized) complex polynomials of fixed degrees $d$, with arbitrarily high probability. This set is an efficient and universal \emph{probabilistic} set of starting points to find all roots of polynomials of degree $d$ using Newton's method; the best known \emph{deterministic} set of starting points consists of $\lceil 1.1d(\log d)^2\rceil$ points.

Dierk Schleicher

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 $\log^2 d$.

Marvin Randig, Dierk Schleicher, Robin Stoll

We present a practical implementation based on Newton's method to find all roots of several families of complex polynomials of degrees exceeding one billion ($10^9$) so that the observed complexity to find all roots is between $O(d\ln d)$ and $O(d\ln^3 d)$ (measuring complexity in terms of number of Newton iterations or computing time). All computations were performed successfully on standard desktop computers built between 2007 and 2012.