Prediction proberties of Aitken's iterated \(\Delta^2\) process, of Wynn's epsilon algorithm, and of Brezinski's iterated theta algorithm (Q1590791): Difference between revisions

The paper is concerned with double sequences \( f(i,k|z) ( i,k \geq 0) \) of functions obtained from partial sums \( f(i,0|z) = \{ \sum c(\tau) z^{\tau} |0 \leq \tau < i \} \). It is supposed that integers \( \xi(i,k) \geq 0 \) exist such that \[ f(i,k|z) \sim f(\xi(i,k),0|z) + z^{\xi(i,k)} g(i,k|z) \] where \( g(i,k|z) \) is associated with an ascending formal power series whose first coefficient \( d(i,k) \) differs from \( c(\xi(i,k)) \). For the cases in which the functions \( f \) are produced by repeated application of Aitken's process and by use of the epsilon- and theta-algorithms, recursions for the \( g(i,k|z) \) are given. Three remarks are perhaps in order. (a) In the case of the epsilon-algorithm, it follows from Satz 5.6 on p. 243 of \textit{O. Perron}'s book [Die Lehre von den Kettenbrüchen, Bd.\ II, Teubner (1957; Zbl 0077.06602)] that the difference \( d(i,k) - c(\xi(i,k)) \) is expressible as a quotient of two Hankel determinants. Such determinants satisfy a simple recursion, as do the ratios themselves: the cumbersome construction given in the paper may be discarded. Furthermore, when \[ c(\tau) = \Bigl\{ \prod \{ ( A - q^{\alpha + \nu }) / ( B - q^{\beta + \nu }) \} \mid 0 \leq \nu < \tau \Bigr\} \] the Hankel determinant quotients may be expressed in simple closed form of which there exist many special cases obtained by giving \( q \), \( A \), \( \alpha \), \( B \) and \( \beta \) particular values (one of the cases is that used as a numerical illustration in the paper). (b) In a numerical investigation it is supposed that \textit{ceteris paribus } a smaller value of the leading coefficient difference modulus \( |d(i,k) - c(\xi(i,k)) |\) implies a smaller value of \( |f(i,k|z) - C(z) |\), where \( C(z) \) is the sum associated with the power series having coefficients \( c \), but this may not be so. (c) For the numerical example chosen, corresponding repeated application of the epsilon-algorithm causes a dramatic increase in the rate of convergence (inclusion of this variant in the comparison would, of course, greatly have detracted from the interest of the results).