A new approach to acceleration of convergence of a sequence of vectors (Q1911449)

In the following \(\mathbb{K}\) is a field and \(\mathbb{M}\) is a right module over \(\mathbb{K}\) for which a \(\mathbb{K}\)-inner product possessing appropriate properties and denoted by open brackets \(({\underset \sim v}, {\underset \sim w} \in \mathbb{M}\) produce \(({\underset \sim v}, {\underset \sim w} \in \mathbb{K}))\) is defined; \(z\) either features as a formal variable in algebraic extensions of \(\mathbb{K}\) and \(\mathbb{M}\) or is given a value in \(\mathbb{K}\); \(N > 1\) is a fixed integer, \({\underset \sim u} (\omega)\) \((0 \leq \omega < N)\) is a sequence in \(\mathbb{M}\) and \({\underset \sim S} (z \mid \omega)\) is \(\{\sum {\underset \sim u} (\tau)z^\tau \mid 0 \leq \tau < \omega \}\) for \(0 \leq \omega \leq N\); \({\underset \sim u} [m,m + r - 1]\) and \({\underset \sim S} (z \mid [m,m + r])\) are the sequences \({\underset \sim u} (\omega)\) \((m \leq \omega < m + r)\) and \({\underset \sim S} (z \mid \omega)\) \((m \leq \omega \leq m + r)\). The paper is concerned with a barycentric form \[ \Lambda (z,S \mid k,m) : = \biggl\{ \sum{\underset \sim S} (m + 2k - \omega) \bigl[ \lambda (\omega) z^\omega/ \sigma (z) \bigr] \mid 0 \leq \omega \leq k \biggr\} \] where \(\lambda (\omega) \in \mathbb{K}\) \((0 \leq \omega \leq k)\) and \(\sigma (z) : = \{\sum \lambda (\omega) z^\omega \mid 0 \leq \omega \leq k\}\). \(\Lambda (z, {\underset \sim S} \mid k,m)\) lies in the space spanned by \({\underset \sim S} [m + k,m + 2k]\), but it is supposed that each \( \lambda (\omega)\) depends upon \({\underset \sim u} [m,m + 2k - 1]\). It is required that, subject to appropriate existence conditions, the form should satisfy three requirements: [i] That \(\Lambda (z,S \mid 1, m)\) should reduce to the form in which \(\lambda (0) = ({\underset \sim u} (m), {\underset \sim u} (m))\) and \(\lambda (1) = - ({\underset \sim u} (m), {\underset \sim u} (m + 1))\) (this is a variant of Aitken's \(\delta^2\)-process for vectors). [ii] That if, with \(k > 1\) fixed, \({\underset \sim u} (\omega) = \{\sum {\underset \sim v} (\tau) \chi (\tau )^\omega \mid 0 < \tau \leq k\}\) with \({\underset \sim v} (\tau) \in \mathbb{M}\), \(\chi (\tau) \in \mathbb{K}\) \((0 < \tau \leq k)\), the \(\chi (\tau)\) being distinct and unequal to 1, then \(\Lambda ({\underset \sim S}, z \mid k,m) = {\underset \sim L} (z)\) for \(0 \leq m \leq N - 2k\), where \({\underset \sim L} (z) = \{\sum {\underset \sim v} (\tau)/ \{1 - \chi (\tau) z\} \mid 0 < \tau \leq k\}\) \((\mathbb{K} \) being a suitable field, the \(\chi (\tau)\) being appropriately restricted and the integer bound \(N\) being discarded, \({\underset \sim L} (z)\) is the limit of the sequence \({\underset \sim S} (\omega)\) \((\omega \geq 0))\). [iii] That when \(\mathbb{M}\) is \(\mathbb{K}\), \(\Lambda ({\underset \sim S}, z \mid k,m)\) (now, together with the \({\underset \sim u} (\omega)\), in \(\mathbb{K})\) should be the approximating fraction whose denominator and numerator are of degrees \(k\) and \(k + m \) respectively in \(z\) and is derived from the ascending power series whose leading coefficients are \({\underset \sim u} (\omega)\) \((0 \leq \omega < N)\). A closed expression, involving a \(2k\)-fold integral, is given for \(\sigma (z)\). (We remark in passing that all of the preceding has a dual counterpart in which \(\Lambda (z,S \mid k,m)\) lies in the space spanned by \({\underset \sim S} [m,m + k].)\) The work is motivated by a misrepresentation of the stability theory of the vector \(\varepsilon\)-algorithm occurring in an earlier paper [*]: [\textit{I. D. Coope} and \textit{P. R. Graves}-\textit{Morris}, Numer. Algorithms, 5, No. 1-4, 275-286 (1993; Zbl 0798.65005)] and an evident lack of computational experience: it is clear that, as presented by the author, computation of the \(\Lambda\)-forms is not a viable acceleration technique. In this regard very few will take these two papers seriously but, for the benefit of those who are momentarily misled, it is remarked that cases in which instability occurs in the application of the \(\varepsilon\)-algorithm are best dealt with by recourse to a preliminary auxiliary sequence transformation. In an accompanying review of a paper by \textit{A. Salam} [Numer. Algorithms, 11, No. 1-4, 327-337 (1996) reviewed below] the reviewer offers a list of expository stratagems which may be adopted by an author who experiences difficulty in producing a proof, an example of each stratagem being present in the paper concerned. In the present paper, the author adopts a further stratagem which is more clever. A vital formula is referred (p. 195, l. 7 et seq.) to a paper which has appeared (it is [*] above); but no formula even remotely resembling that in question occurs in the paper cited.