Practical error estimation in numerical integration (Q1074294): Difference between revisions

There is no shortage of rules and algorithms for one-dimensional quadrature over a finite interval. It is relatively easy to construct automatic quadrature routines. What is very difficult is to determine which of several choices is ''best''. This is mainly because a generally accepted precise definition of ''best'' in this context is lacking, and many proposals in this area contain large subjective components. Nevertheless, the question is discussed in the literature, and the present author has made several interesting suggestions and describes some numerical experiments. His suggestions are mainly about possible ''practical convergence criteria'' and, while this reviewer has reservations about some of these, I am generally in agreement that that all should be investigated. The author gives a very brief outline of his investigations which include Kronrod-Gauss rules and Romberg extrapolation. If I understand him correctly, he concludes that Gaussian rules, used with one of two appropriate convergence criteria, are ''best''. One criterion uses divided differences to approximate the high derivatives in the Peano error bound. The other uses a nonlinear combination of two such criteria.