The kink phenomenon in Fejér and Clenshaw–Curtis quadrature

Abstract

The Fejer and ClenshawCurtis rules for numerical integration exhibit a curious phenomenon when applied to certain analytic functions. When N (the number of points in the integration rule) increases, the error does not decay to zero evenly but does so in two distinct stages. For N less than a critical value, the error behaves like O(ρ-2N), where ρ is a constant greater than 1. For these values of N the accuracy of both the Fejer and ClenshawCurtis rules is almost indistinguishable from that of the more celebrated GaussLegendre quadrature rule. For larger N, however, the error decreases at the rate O(ρ-N), i.e., only half as fast as before. Convergence curves typically display a kink where the convergence rate cuts in half. In this paper we derive explicit as well as asymptotic error formulas that provide a complete description of this phenomenon.