Robust Padé approximants may have spurious poles (Q476129)

The author addresses an open question posed by \textit{P. Gonnet} et al. [SIAM Rev. 55, No. 1, 101--117 (2013; Zbl 1266.41009)] concerning the influence of decreasing the degree of approximants (leading to fewer Froissart doublets and better approximation) on the existence of pointwise convergence. Exhibiting an example where the approximants generated by the techniques introduced in Gonnet et al. may not converge to the function being approximated and where spurious poles may arise even when the matrices involved have smaller condition number than \(5\); therefore the answer to the open question is negative. \vskip0.3cm The example is an adaptation of the classic one due to J. L. Gammel (given in [\textit{G. A. Baker jun.} and \textit{P. Graves-Morris}, Padé approximants. Part I: Basic theory. Reading, Massachusetts, etc.: Addison-Wesley Publishing Company (1981; Zbl 0468.30032)]). The function used is \[ f(z)=1+\sum_{k=2}^{\infty}\,16^k\left(z^{n_k-1}+z^{2n_k}{(z/z_k)^{n_k}-(z/z_k)^{2n_k+1}\over 1-z/z_k}\right),\;n_k=2^k-2, \] with \(\{z_k,\,k\in\mathbb{N}\}\) any sequence with \(0<|z_k|<1/3\). The function \(f(z)=\sum\,c_jz^j\) than has coefficients satisfying \(0<|c_j|\leq (j+3)^4\) and the \((n_k,n_k)\) Padé approximants have a spurious pole at \(z_k\) (in fact, the denominator of the approximant is \(1-z/z_k\) and the numerator is different from zero there). This, and information about the singular values of the Hankel-type determinants involved, are stated in Theorem 1 in the paper.