An optimal approximation formula for functions with singularities (Q723767)

Given a \(f\) analytic in a lens shaped region \(\Lambda_d\) with tops in \(\pm1\) and thickness \(d\) at 0, the problem is to interpolate it on \([-1,1]\) in \(N\) points by a rational function \(f_N\). The latter is of the form \(f_N(x)=\sum_{|k|\in\{1,\ldots,N\}} f(\beta_k)\phi_{N,k}(x)\). The \(\beta_k\) are (modified) Ganelius points (depending on \(N\) and \(d\)) [\textit{T. Ganelius}, Ann. Acad. Sci. Fenn., Ser. A.I. Math. 2, 129--145 (1976; Zbl 0354.30026)]. The \(\phi_{N,k}\) form a Lagrange-like interpolation basis, namely \(\phi_{N,k}=c_k\frac{1-x^2}{x-\beta_k}B_n(x;\beta)\) with \(B_N(x;\beta)\) a generalization of a Blaschke product associated with the \(N\) interpolation points. The error is measured in the Hardy space \(H^\infty(\Lambda_d,w_\mu)\), with weight \(w_\mu(z)=(1-z^2)^{\mu/2}\), \(0<\mu\) which should compensate for singularities in \(\pm1\). An error bound for \(0<\mu<\min\{2,\pi/d\}\) was given in [\textit{T. Ushima} et al., ``A function approximation formula using the Ganelius sampling points'', Trans. Japan Soc. Ind. Appl. Math. 27, 1--20 (2017; \url{doi:10.11540/jsiamt.27.1_1})]. The main result of this paper is to generalize this estimate by removing the upper bound on \(\mu\). Therefore a \(\mu\)-dependency is introduced in the \(\phi_{N,k}\) by replacing the factor \((1-x^2)\) by \((1-x^2)^\nu\), \(\mu/2<\nu<\mu/2+1\). The error is then bounded by \(C\exp(-\sqrt{\pi d\mu N})\) with \(C>0\) independent of \(N\). Numerical experiments show that this approximant is better than the nearly optimal sinc interpolation and approaches the minimum error norm.