A short note on the generalized Euler transform for summation of power series (Q1200533)

Let \(S_ n(x)=\sum^ n_{k=0}m_ k(-x)^ k\) be the \(n\)-th partial sum of a power series the coefficients of which are assumed to be the moments \(m_ k=\int^ 1_ 0u^ kf(u)du\) of some known or unknown function. We have, at least formally: \[ S(x)=\sum_{k\geq 0}m_ k(-x)^ k=\int^ 1_ 0{f(u)\over 1+xu}du. \] The generalized Euler transform applied to the first \(n\) terms of this series consists in computing the triangular array of values \(S^{(k)}_ m(x)\) defined by the following algorithm \((x\neq -1)\): \((i)\;S^{(0)}_ m(x)=S_ m(x)\) for \(0\leq m\leq n-1\); \((ii)\text{ for }k=1,2,\dots,n-1\) and \(0\leq m\leq n-k-1\), \[ S_ m^{(k)}(x)=S_ m^{(k-1)}(x)+{1\over 1+x}(S_{m+1}^{(k-1)}(x)-S_ m^{(k-1)}(x)). \] The author is interested in the vector \((S_ 0^{(n- 1)}(x)\), \(S_ 1^{(n-2)}(x),\dots,S^{(0)}_{n-1}(x))\) which forms the inferior diagonal of this triangular array. Some authors have noticed that, in many cases, the best approximation of \(S(x)\) is neither the first value \(S_ 0^{(n-1)}(x)\) nor the last one \(S_{n- 1}^{(0)}(x)\), but an intermediate value \(S_ k^{(n-1-k)}(x)\) for some \(0< k< n-1\). The aim of this article is to propose some estimations of the index \(k\) depending on both \(n\) and \(x\). The simplest one is considered to be \(k=[(n-x)/(x+2)]\) (where \([m]\) is the integer part of \(m\)). Other estimations are given, depending on the behaviour of the function f. The author's results confirm the intuitive fact that partial sums are better when the series converges very fast while values provided by the complete Euler transform are better when the series is strongly divergent (e.g. for large values of \(x\)) or converges very slowly. Although this method is considered to be less efficient than nonlinear transforms like the \(\varepsilon\)-algorithm, the iterated \(\Delta^ 2\)- algorithm or the \(\theta\)-algorithm, it has several advantages: it is very simple to program, numerically stable and it gives upper and lower approximations of \(S(x)\), e.g. \(S_ k ^{(n-1-k)}(x)\) and \(S_{k- 1}^{(n-k)}(x)\) for the best index \(k\). Finally, ten numerical examples confirming the validity of the theoretical estimations are presented.