The error and guaranteed accuracy of cubature formulas in multidimensional periodic Sobolev spaces (Q486360)

Let \(\Omega\) be a bounded domain of \(\mathbb{R}^n\), \(n\geq 1\), and \(\varphi\in X(\Omega)\), where \(X(\Omega)\) is a Banach space embedded into the space \(C(\overline{\Omega})\) of continuous functions on \(\Omega\). The author studies cubature formulas of the form \[ (M_\Omega,\varphi)=\frac{1}{|\Omega|}\int_\Omega \varphi(x)dx \approx \sum_{j=1}^N c_j \varphi(x^{(j)})=(\sum_N,\varphi), \] where \(|\Omega|\) stands for the volume of \(\Omega\), \(x^{(j)}\in \Omega\) (\(j=1,\dots,N\)) are the nodes and the weights \(c_j\) fulfil the condition \(c_1+\cdots+c_N=1\). The theoretical error \((l_N,\varphi)=(M_\Omega-\sum_N,\varphi)\) differs from the practical error \((\tilde{l}_N,\varphi)=(M_\Omega-\widetilde{\sum}_N,\varphi)\). The latter is nonlinear and depends on the floating point representation of reals. The upper estimate of the practical error is studied when the integrands belong to Sobolev spaces on a multidimensional cube. Due to the nature of the issue, three machine constants play a major role; these are \(\varepsilon_0\) and \(\varepsilon_\infty\), respectively underflow and overflow level, and the machine epsilon \(\varepsilon_1\). These epsilon constants are used to define the number \({\mathbf{R_F}}=\sup_{\mathbf{B}_{L_\varepsilon}}\{|M_\Omega(\varphi)-(\widetilde{\sum}_N,\varphi)|/\|\varphi|X(\Omega)\|\}\), where \(\mathbf{B}_{L_\varepsilon}=\{\varphi \in X(\Omega)|\sqrt{2}\varepsilon_1^{9/2}\leq A_n \|\varphi|X(\Omega)\|\leq \varepsilon_\infty^{1/3}\}\). \(\mathbf{R_F}\) is used to estimate the practical error as does \(\|\varphi|X(\Omega)\|\) when it comes to the theoretical one. Theorem 1 provides an estimate of the distance between these two quantities. As corollary, an upper bound for the practical error is found. Extremal functions for cubature formulas of the type given above are described as a linear combination of functions, such that the \(i\)th is extremal for the cubature formula with the single node \(x^{(i)}\).