Robust Approximate Zeros in Banach Space

Abstract

We extend Smale’s concept of approximate zeros of an analytic function on a Banach space to two computational models that account for errors in the computation: first, the weak model where the computations are done with a fixed precision; and second, the strong model where the computations are done with varying precision. For both models, we develop a notion of robust approximate zero and derive a corresponding robust point estimate. A useful specialization of an analytic function on a Banach space is a system of integer polynomials. Given such a zero-dimensional system, we bound the complexity of computing an absolute approximation to a root of the system using the strong model variant of Newton’s method initiated from a robust approximate zero. The bound is expressed in terms of the condition number of the system and is a generalization of a well-known bound of Brent to higher dimensions.