Extension of a class of periodizing variable transformations for numerical Integration

DOI10.1090/S0025-5718-05-01773-4MaRDI QIDQ5713230zbMATH OpenFDO

Publication date 12 December 2005

Published in Mathematics of Computation (Search for Journal in Brave)

Full work available at URL https://doi.org/10.1090/S0025-5718-05-01773-4

endpoint singularities Euler-Maclaurin expansions numerical examples Numerical Integration numerical quadrature sin\(^m\)-transformation Trapezoidal Rule

Mathematics Subject Classification ID

Approximate quadratures (41A55) Euler-Maclaurin formula in numerical analysis (65B15) Numerical quadrature and cubature formulas (65D32)

Review text

It has been known for many years that the performance of the trapezoidal rule over a finite interval, say [0,1], can be significantly improved when the integrand function \(f(x)\) is transformed to say \(\hat{f}(x) = f(\psi(x))\psi ' (x)\) in such a way that \(\hat{f}(x)\) is ``flat at both ends of the interval. This motivated the development in the 1970s of various numerical quadrature techniques. Examples include the IMT method [\textit{M. Iri, S. Moriquti} and \textit{Y. Takasawa}, J. Comput. Appl. Math. 17, 3-20 (1987); translation from Kokyuroku Ser. Res. Inst. for Math. Sci. Kyoto Univ. 91, 82-118 (1970; Zbl 0616.65023)], the double exponential method [\textit{M. Mori}, Publ. Res. Inst. Math. Sci., Kyoto Univ. 14, 713--729 (1978; Zbl 0402.65012)] and an approach due to [\textit{N. M. Korobov} [Zahlentheoretische Methoden in der Näherungsanalysis. Moskau: Staatsverlag für physikalisch-mathematische Literatur. (1963; Zbl 0115.11703)]. In a landmark paper [A new variable transformation for numerical integration. Numerical integration IV. Proceedings of the conference at the Mathematical Research Institute, Oberwolfach, Germany, November 8-14, 1992. Berlin: Birkhäuser. ISNM, Int. Ser. Numer. Math. 112, 359-373 (1993; Zbl 0791.41027)] the author introduced a class \(S_m\) of transformations \(\psi(t)\), \(m\) being a positive integer. Briefly, \(\psi \in S_m\) when \(\psi (t)/t^{m+1}\) and \(\psi(1-t)/(1-t)^{m+1}\) have even expansions in \(t\) and in \((1-t)\), respectively; \(\psi(0)=0\), together with a symmetry condition (on the derivatives) \(\psi(t) + \psi (1-t) = 1\) are also required. When \(\psi \in S_m\), \(f \in C^\infty\), and \(m\) is even, the \(n\)-panel trapezoidal rule \(Q_n\hat{f}\) gives a discretisation error \(O(h^{2m+2})\) rather than the general \(O(h^2)\) in terms of the steplength \(h = 1/n\). In a different major contribution [Numer. Math. 98, No.~2, 371--387 (2004; Zbl 1081.65003)] the author generalized the Euler-Maclaurin asymptotic expansion to treat more sophisticated endpoint singularities, including \(\sum_{s=0}^{\infty} c_sx^{\gamma_s}\) (and \(\sum_{s=0}^{\infty} d_s(1-x)^{\delta_s}\)) where, with certain minor restrictions, \(\gamma_s\) (together with \(\delta _s\)) is monotonic and general. (Previously, results in this area, e.g., Navot, Lyness and Ninham, essentially restrict \(\gamma_s\) to a sequence of type \(\gamma _s = \alpha+s\).) In the present paper, the author updates his 1993 paper in two ways; by treating \(S_m\) for non-integer \(m\) (in particular, for half integer \(m\)): and by treating his more sophisticated version of the Euler-Maclaurin expansion. He provides a full theory for determining the order of the approximation \(Q_n\hat{f}\) in terms of \(m\) and early members of the set of exponents \(\gamma _s\) and \(\delta _s\). At first sight, these results do not appear to be significantly different from those presented in 1993. However, examination of the relevant expressions shows that a judicious choice of \(m\) (the order of the transformation) may result in the vanishing of the first term in the expansion of the discretisation error, thus increasing the accuracy of \(Q_n\hat{f}\) significantly. The author develops this approach and provides a scheme for choosing \(m\). One of several results, one which is illustrated in both his examples, is that when \(\gamma _0 = \delta_0 = 1\) and \(f \in C^\infty\), i.e., \(f(0) = f(1) = 0\), then the error is \(O(h^{2m+2})\) when \(2m\) is not an odd integer, but \(O(h^{3m+3})\) when \(2m\) is odd, i.e., \(m\) is half an odd integer. In two numerical examples, the author provides enough numerical evidence to convince even the most confirmed sceptic. When \(m\) is an integer an obvious choice for \(\psi(t)\) is available as a simple expression in terms of circular functions. However, for non-integer \(m\), the corresponding choice is essentially a hypergeometric function. The author guides the reader through the theory to obtain two different series representations which may be used to calculate an abscissa \(\psi(ih)\); he also provides an integral representation, which may be evaluated using Gauss-Jacobi quadrature. This reviewer was disappointed to find that while simple expressions for an obvious choice for \(\psi(t)\) and \(\psi '(t)\) are available for integer \(m\), their evaluation when \(m\) is not integer is not immediate. While the author provides information that should be very helpful, it is left to the user to carry out tasks like accelerating slowly convergent series before he can begin to use the method. This paper contains no comparison between numerical results based on the \(S_m\) transformation with those using the more familiar approaches, such as the more accessible Gaussian formulas or the techniques mentioned in the first paragraph of this review. Nevertheless, this paper comprises a major contribution to the theory of one-dimensional numerical quadrature.

Recommendations

Cites work

Cited in

(22)

Describes a project that uses

Uses Software

This page was built for publication: Extension of a class of periodizing variable transformations for numerical Integration

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q5713230)