A discussion of a new error estimate for adaptive quadrature (Q1123534): Difference between revisions

``A critical part of any automated integration routine is the way we estimate the error. When constructing an error estimate one has to compromise between two contradictory aims: reliability and economy.'' These are the authors' opening sentences in a paper which approaches this problem in a realistic way, preserving a fine balance between theoretical expectation and numerical experiment. The authors treat the problem of how to handle four numerical approximations \(B_ 1,B_ 2,A_ 1\), and \(A_ 2\) to the same integral I (perhaps over a small interval); here \(B_ 1\) and \(A_ 1\) employ two different rules (for example the Gauss- Legendre rule and the corresponding Gauss-Kronrod rule), and \(B_ 2\) and \(A_ 2\) employ the two copy versions of the respective rules. In a contrived situation in which convergence is monotonic, the estimate \(I\cong A_ 2+\epsilon\) with \(\epsilon =(A_ 2-B_ 2)(A_ 2-A_ 1)/(B_ 2-B_ 1-A_ 2+A_ 1)\) is reasonable, and in a previous paper by \textit{D. P. Laurie} [J. Comput. Anal. Math. 12-13, 425-431 (1985; Zbl 0589.65020)] conditions are given for I to lie in \([A_ 2,A_ 2+\epsilon]\). (These conditions depend on I, whose numerical value is of course not available.) The authors discuss previous error estimates based on these conditions. These previous estimates work well in ``the asymptotic region''. However, the point of adaptive quadrature is to economize in function values and this has the effect of arranging the calculation so that this asymptotic region is rarely encountered. With this in mind, the authors propose a more sophisticated error estimate; they confirm by numerical experiment their expectation that this one is more expensive and much more reliable. The reviewer recommends this paper to any quadrature routine constructor who is about to impose a practical convergence criterion.