Generalized multivariate Padé approximants (Q1273408): Difference between revisions

There exist several more or less fortunate extensions of the classical Padé approximants to the multidimensional setting. Let \(f\) be a function given by its power series in \({\mathbb{C}}^d[z]\). For two finite sets of indices, \(M\) and \(N \in {\mathbb{N}}^d\) with the rectangular inclusion property, we can intend to approximate \(f\) with a ratio \(P/Q\), where \[ P \in {\mathbb{P}}_M=\Bigl\{ \sum_{i \in M} \alpha_i z^i: \alpha_i \in {\mathbb{C}}^d\Bigr\}, \quad Q \in {\mathbb{P}}_N= \Bigl\{ \sum_{j \in N} \beta_j z^j: \beta_j \in {\mathbb{C}}^d\bigr\}. \] Following the one-dimensional analogue, several authors define the Padé approximant by imposing the condition that sufficiently many coefficients of the power expansion of \(Qf-P\) vanish. Nevertheless, these definitions may lead to inconsistency (\(P/Q \neq f\) when \(f\) itself is a suitable rational function) or the loss of the classical Montessus de Ballore's theorem. Here a different approach is followed, replacing interpolation by least-square approximation. A generalized Padé approximant is a fraction \(P/Q\) for which a weighted least-square norm of the coefficients of \(Qf-P\) on a set \(E \supset M+N\) is minimal. With this definition both consistency and uniform convergence in the Montessus de Ballore's situation are preserved. The authors present also some numerical results.