scientific article; zbMATH DE number 3390213

From MaRDI portal
Publication:5659030





The paper is mostly a survey, but contains a number of sharpened or even new results. Numerical integration in several dimensions is greatly facilitated if the domain of integration is a multidimensional unit interval and if the integral can be represented as a uniformly convergent Fourier series \(f(\mathbf x) = \sum c_{\mathbf h} \exp(2\pi i\mathbf x \cdot\mathbf h)\), where bold letters denote vectors, and dots denote scalar products. Methods of reducing any sufficiently smooth integrand to such a form without affecting the value of the integral are discussed in detail. The result is a series satisfying \(| c_{\mathbf h}|\leq K_r R(\mathbf h)^{-(r+1)}\), where \(R(\mathbf h)=\prod_{i=1}^s \max(1, | h_j|)\), \(h_1,\ldots, h_s\) are the coordinates of \(\mathbf h\), the integer \(r\) depends on the smoothness of the integrand, and \(K_r\) is a constant depending on the multidimensional variations in the sense of Vitali of some partial derivatives of the periodically extended integrand. Then \[ \int_{Q^s} f(\mathbf x)\,d\mathbf x\sim \frac 1{p} \sum_{k=0}^{p-1} f\left(\frac{k}{p}\mathbf g\right), \] where \(Q^s\) is the \(s\)-dimensional unit interval, the integer \(p\) can be arbitrarily chosen, \(\mathbf g\) is a suitably chosen lattice point, and the coordinates of \(p^{-1}k\mathbf g\) are reduced modulo 1. The absolute error of integration is bounded by \(K_rP^{(r+1)}(\mathbf g)\), where \(P^{(n)}(\mathbf g)=\sum R(\mathbf h)^{-n}\), the sum being extended over all \(\mathbf h\neq (0,\ldots,0)\) satisfying \(\mathbf h\cdot\mathbf g\equiv 0\pmod p\). If \(\rho(\mathbf h)\) is the minimum of \(R(\mathbf h)\) for all the aforesaid lattice points, favourable upper bonds for \(P^{(r+1)}(\mathbf g)\) are obtained in terms of \(\rho(\mathbf g)\). Sharpening previous results by \textit{N. M. Korobov} [Number-theoretical methods in approximate analysis. (Russian) Moscow: FizMatLit (1963; Zbl 0115.11703)] and \textit{E. Hlawka} [Compos. Math. 16, 92--105 (1964; Zbl 0146.27602)], the author shows that for any sufficiently large prime \(p\) and any number \(s >1\) of dimensions there exist lattice points \(\mathbf g\) for which \(\rho(\mathbf g)>(s -1)! p/(2 \log p)^{s-1}\). Such \(\mathbf g\) are described as ``good lattice points. They allow us to integrate sufficiently smooth functions much more efficiently than by the Monte-Carlo method or by conventional methods based on iteration. If the integrand is merely of bounded variation in the sense of Hardy and Krause, the error of integration depends on the discrepancy of the set of points over which \(f\) is averaged. If this set is generated by a good lattice point, \(\rho(\mathbf g)\) determines upper bounds for the discrepancy. Integration over more general convex domains is also briefly discussed in terms of the isotropic discrepancy. [For the entire collection see Zbl 0238.00008.]




Cited in
(63)








This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q5659030)