Function classes for double exponential integration formulas (Q1006844)

Double exponential (DE) transformations \(\psi:\mathbb{R}\to I\) have been successfully applied by \textit{H. Takahasi} and \textit{M. Mori} [Publ. Res. Inst. Math. Sci., Kyoto Univ. 9, 721--741 (1974; Zbl 0293.65011)] for estimating the integral \(\int_I f(x)dx\) by means of the formula \[ \int_I f(x)dx=\int_{-\infty}^\infty f(\psi(t))\psi'(t)dt\approx T_{N,\psi}(f)= h \sum_{k=-N}^N f(\psi(kh))\psi'(kh). \] An example of DE transformation \(\psi=\psi_{DE}\) is \(\psi(t)=\sinh((\pi/2)\sinh(t))\), whereas \(\psi_{SE}(t)=\sinh(t)\) is a single exponential (SE) transformation. This paper shows that the integrands suited to the DE formulas are those which are analytic in the region \(\mathcal{D}_{\psi_{DE}} (d):=\{z=\psi_{DE}(w)|w\in \mathcal{D}_d\}\), where \(\mathcal{D}_d:=\{z\in \mathbb{C}|\,|Im(z)|<d\}\), a fact already known for the SE formulas. Given a suitable function \(E(z;\beta)\), for instance, \(E(z;\beta)=(1-z^2)^{\beta/2-1}\), the authors prove that for \(h=N^{-1}\log(8dN/\beta)\), the error for DE formulas is \[ \left|\int_I f(x)dx- T_{N,\psi}(f)\right|=\mathcal{O}\left(\exp\left(-\dfrac{2\pi dN}{\log(8dN/\beta)}\right)\right), \] where \(\psi=\psi_{DE}\) is a DE transformation, \(f\) is analytic in \(\mathcal{D}_\psi (d)\), \(0<d<\pi/2\), and \(|f(z)|\leq C|E(z;\beta)|\), \(z\in \mathcal{D}_\psi (d)\). The previous estimate reconfirms the superiority of the DE formulas with respect to the SE formulas.