On the convergence of Padé-type approximants to analytic functions (Q793490)

In a presentation differing from that of the paper under review, the problem considered is as follows. Let \(\pi_{i,j}(i=0,1,...\); \(j=0,...,i)\) be a system of functions of a complex variable and let f(z)\(\sim \sum^{\infty}_{j=0}f_ jz^ j(*)\) as z tends to zero over an open set with limit point at the origin. By imposing suitable conditions upon the \(\pi_{i,j}\) and f, obtain a set T for which, with \(P_ i(z)=\sum^{i}_{j=0}\pi_{i,j}(z)f_ j, \lim P_ i(z)=f(z)\) for all \(z\in T\). In the author's solution, it is supposed that f is analytic over an open domain D containing the origin and that the \(\pi_{i,j}\) are such that a nonvoid set \(\Lambda\) exists for which, with \(p_ i(u,z)=\sum^{i}_{j=0}\pi_{i,j}(z)u^{-j-1}\), for each \(z\in \Lambda\) the sequence \(p_ i(u,z) (i=0,1,...)\) converges uniformly to \((u-z)^{-1}\) for u over a nonvoid set U(z). Evidently a neighbourhood of z does no belong to U(z).For those \(z\in \Lambda \cap D\) such that U(z)\(\cap D\) contains a contour \({\mathcal C}(z)\) enclosing z, \(P_ i(z)=(2\pi i)^{-1}\int_{{\mathcal C}(z)}p_ i(u,z)f(u)du\), and \(P_ i(z)\to f(z)\). When \(D\supset \{{\mathbb{C}}\backslash U(z)\}\), or \(\{ {\bar {\mathbb{C}}}\backslash D\}\subset U(z)\), such a \({\mathcal C}(z)\) exists. Thus having obtained the \(\Lambda\) and all U(z) from the \(\pi_{i,j}\), determine the open domain \(T(\Lambda\),D) composed of all z for which \(\xi\in U(z)\) for all \(\xi\in {\bar {\mathbb{C}}}\backslash D\); then \(lim P_ i(z)=f(z)\) uniformly over every compact subset of \(T(\Lambda\),D). This is the author's principal result. A subsidiary result is that a relationship of the form \(p_ i(u,z)-(u- z)^{-1}=O(\rho_ i(z))\) for large i uniformly for all \(u\in U(z)\) induces a corresponding result \(P_ i(z)-f(z)=O(\rho_ i(z)).\) A sequence to sequence transformation of the form \(B_ i=\sum^{i}_{j=0}b_{i,j}S_ j (i=0,1,...)\) may be expressed in terms of the differences \(t'_ 0=S'_ 0,t'_ j= S_ j-S_{j-1} (j=1,2,...)\) of the members of the sequence being transformed: \(B_ i=\sum^{i}_{j=0}\pi '\!_{i,j}t'\!_ j (i=0,1,...)\) where \(b_{i,j}=\pi '\!_{i,j}-\pi '\!_{i,j+1} (j=1,...,i-1)\), \(b_{i,i}=\pi '\!_{i,i}\). The \(\pi_{i,j} (z)/z^ j\) of the author's result are therefore associated with a summation method of the above type in which the \(b_{i,j}\) are functions of z. In this context, the author's result is an extension of a result of \textit{Y. Okada} [Math. Z. 23, 62-71 (1925; JFM 51.0243.01); see also \textit{W. Gawronski} and \textit{R. Trautner}, Period. Math. Hung. 7, 201-211 (1976; Zbl 0324.40002)]. Let the \(\pi_{i,j}\) be constants, and suppose that lim \(\sum^{i}_{j=0}\pi_{i,j}w^ j=(1-w)^{-1}\) for all \(w\in G\), G being an open set containing the origin. The set \(\Lambda\) now has one member, namely \(z=1\), and \(U(1)=G^{-1}\). Suppose again that f in relationship (*) is analytic over the open domain D containing the origin. Applying the theory sketched above with \(f_ j\) replaced by \(f_ jw^ j\), it follows that the sum \(\sum^{i}_{j=0}\pi_{i,j}f_ jw^{j-1}\) tends, as i increases, to f(w) for all w for which \(D\supset \{wU(1)\}\). All \(w\in Z(G,D)\equiv \{\cap \xi G\) for all \(\xi\in {\mathbb{C}}\backslash D\}\) are characterised by the property that given w, for each \(\xi\in {\mathbb{C}}\backslash D\) a g(w,\(\xi)\in G\) can be found such that \(\xi =w/g(w,\xi)\), so that \(\{{\mathbb{C}}\backslash D\}\subset \{\{wG^{-1}\}\equiv \{wU(1)\}\}\) and in turn \(D\supset wU(1).\) The following corollary to the author's result follows: if the sequence transformation \({\mathfrak B}\) associated with the constants \(\pi_{i,j}\) sums the geometric series \(\sum^{\infty}_{j=0}z^ j\) to \((1-z)^{- 1}\) over an open set G containing the origin, then \({\mathfrak B}\) sums the series (*) to f(z) over Z(G,D). This is Okada's result. The convergence of the \(p_ i(u,z)\) is particularly easy to study in the case in which \(p_ i(u,z)=(u-z)^{-1}\{1-R_ i(u,z)\}\) where \(R_ i(u,z)=z^ iq_ i(u)/u^ iq_ i(z), q_ i\) being an ith degree polynomial. The behaviour of the \(p_ i\), and hence that of the \(P_ i(z)\), may then be studied in terms of the \(q_ i\). (Such a discussion in which \(q_ i(z)=(z-\beta)^ i\) is given.) In this case, \(P_ 0(z)=0\) and \(q_ i(z)P_ i(z)\) is an (i-1)th degree polynomial which has, if f is analytic at the origin, a representation of the form \((2\pi i)^{- 1}\int_{{\mathcal C}}u^{-i}(u-z)^{-1}\{u^ iq_ i(z)-z^ iq_ i(u)\}f(u)du.\) In the special case in which \(q_ i(z)=z^ iQ_ i(z^{-1})\), where \(Q_ i\) is the ith degree orthogonal polynomial generated by the sequence \(\{f_ j\}\), \(P_ i(z)\) is, subject to suitable existence conditions, the ith order convergent \(C_ i(z)\) of the continued fraction associated with the series (*). As is known from the theories of the Stieltjes and Hamburger moment problems, convergence of the \(P_ i(z)\) holds in certain cases in which f is not analytic at the origin, but is represented asymptotically there by the series (*). The suggestion naturally prompts itself that this might also be true when the \(\pi_{i,j}\) are independent of the \(f_ j\). It may also be remarked that, as was shown by Stieltjes in the case of Stieltjes moments and has been shown by the reviewer for other cases (for further references, see the reviewer's paper in Acta Math. Akad. Sci. Hung. 25, 291-297 (1974; Zbl 0323.30043) among all rational functions with numerator and denominator of degrees i-1 and i respectively, \(C_ i(z)\) offers an approximation to f(z) which is best in a certain sense. In this sense the author's approximations, although of great interest, are not better.