On an algorithm considered by Stieltjes (Q2640799)

The contents of the main part of the paper can be described by the following paraphrase of its introduction: It is very well known that Gauss determined the arithmetic-geometric mean by the algorithm \(p_{n+1}=(p_ nq_ n)^{1/2}\), \(q_{n+1}=(p_ n+q_ n)/2\). In a letter to Hermite, dated Dec. 31, 1891, Stieltjes mentioned the algorithm \(s_{n+1}=(s^ 2_ n+t^ 2_ n)/(s_ n+t_ n)\), \(t_{n+1}=(s_ n+t_ n)/2\) with the limit \(M(s_ 1,t_ 1)\). The principal purpose of the present note is to reconstruct the mathematics of Stieltjes on this subject. In particular, for the coefficients \(a_ n\) of the power series \(f(x)=a_ 0+a_ 1x+a_ 2x^ 2+...\) of the function f satisfying \(M(f(x),f(-x))=1\), the authors give a recursion formula which, they think, must have been known to Stieltjes. On the other hand, contrary to the statement of Stieltjes that ``le rayon de convergence de cette série se reduit à zéro'' they show that it has a positive radius of convergence \(r=(\lim_{n\to \infty} c_ n)^{-1}\approx 0.634584512652...,\) where \(c_ 0=0\), \(c_ n=(2+\exp (2^ n \log c_{n-1}))^{1/2^ n}\) \((n=1,2,...).\) The rest of the paper deals with algorithms of the form \(u_{n+1}=(\epsilon (u^ 2_ n+v^ 2_ n)+(1-\epsilon)2u_ nv_ n)/(u_ n+v_ n),\quad v_{n+1}=(u_ n+v_ n)/2\) [misprint in line 3 of p. 482: \((1+\epsilon)\) should be replaced by (1-\(\epsilon\))], and with more general ones. Already for \(x_{n+1}=((x_ n+y_ n)^ 3+c(x_ n- y_ n)^ 3)(x_ n+y_ n)^{-2}/2\), \(y_{n+1}=(x_ n+y_ n)/2\) they pose the determination of the domain of convergence as an open problem.