An algebraic approach to the vector \(\varepsilon\)-algorithm (Q1911459)

In an earlier paper on a related subject [\textit{A. Salam}, Numer. Algorithms, 7, No. 2-4, 225-251 (1994; Zbl 0812.65003)] the author referred the proofs of crucial lemmata to a thesis of restricted access; those parts of other proofs made available for inspection contained what appeared to the reviewer to be errors. In the present paper the performance is repeated. The theory is set in a Clifford algebra \({\mathcal C} (\mathbb{K})\) over a suitable field \(\mathbb{K}\); the algebra is induced by \(h \geq 1\) generators satisfying prescribed relationships and is of dimension \(2^h\). \(V [{\mathcal C} (\mathbb{K})]\) is that part of the algebra consisting of numbers whose components associated with just one generator are possibly nonzero while the remaining components are zero. It is supposed that a sequence \(S(n)\) \((n \geq 0)\) and a double sequence \(g (n \mid \omega)\) \((n \geq 0, \omega \geq 1)\), both consisting of \(V [{\mathcal C} (\mathbb{K})]\)-numbers, are prescribed. The paper deals with the problem of deriving the single \(V [{\mathcal C} (\mathbb{K})]\)-number \(\theta (g, S |r,m)\) from the \(r + 1\) linear equations \[ \biggl\{ \sum c \bigl( g, S |r,m |\omega \bigr) g \bigl( n \mid \omega \bigr) \mid 0 < \omega \leq r \biggr\} + \theta \bigl( g,S |r,m \bigr) = S(n) \] holding for \(m \leq n \leq m + r\), where \(m \geq 0\) and \(r > 0\) are fixed; the coefficients \(c\) are \({\mathcal C} (\mathbb{K})\)-numbers. These equations may be expressed in \({\mathcal C} (\mathbb{K})\)-matrix form as \(G[r,m] \overline c[r,m] = S[r,m]\), where \(G\) is formed from the elements \(g\) and is bordered on the right by a column of \({\mathcal C} (\mathbb{K})\)-units, \(\overline c\) is formed from the coefficients \(c\) followed by \(\theta\), and \(S[r,m]\) is formed from the members of the sequence \(S\). In this equation the matrix product is formed from inner products of rows of \(G\) and the column \(\overline c\), but multiplication of elements takes place in the ring opposite to \({\mathcal C} (\mathbb{K})\): elements of \(G\) postmultiply those of \(\overline c\). The explanation given is tortuous, but in essence the problem is attacked by a direct analogue of reduction to block triangular form in the solution of linear algebraic equations over a field. The above matrix equation is multiplied throughout by a lower triangular matrix \(L\) such that \(LG[r,m]\) is of upper triangular form and the matrix multiplication being as just described. The component equation involving the last row of \(LG[r,m]\) then reduces to the form \[ \theta (g, S \mid r,m)\quad p(r,m) = q(r,m) \] and, if the inverse of \(p(r,m)\) is available, to \(\theta (g,S \mid r, m) = q(r,m) p(r,m)^{-1}\). It is claimed that in the case in which \(g(n \mid \omega)\) has the special form \(S(n + \omega) - S(n + \omega - 1)\) the inversion processes needed to construct \(L\) can be implemented and that \(p(r,m)\) is also invertible, this claim being covered by a lemma (p. 332) whose proof is ``to appear''. It is also stated that, in this case, the quotients for \(\theta\) satisfy the relationships of the \(\varepsilon\)-algorithm, reference being made to the earlier paper, which appears to be unsatisfactory, mentioned at the beginning of this review. In those parts of the reasoning that are visible, clear errors occur. To give just one example, it is stated (p. 332, 2.3) that a certain expression is obviously the product of nonzero \(V [{\mathcal C} (\mathbb{K})]\)-numbers, but it is not shown that the first factor must be nonzero; presumably this error is propagated into the invisible parts of the proof. This paper and its predecessor suggest a strategy of exposition to the mathematician who experiences difficulty in producing proofs. He should demonstrate some minor results; the reasoning may contain errors but at least shows that he has thought about the matter. He should sprinkle the text with results that are classical and correct, but irrelevant. He should refer crucial steps in proofs either to a document of restricted access, or to a further text which is ``to appear'' without stating where, or to previous texts which have been prepared in the same way. There is a possibility that nobody will notice what is going on.