New vector sequence transformations (Q1887617)

\( { \mathbf x } \) being a \(p\)-dimensional real vector, \( { \mathbf Z }( n \; | \; i ) \) being a \( p \times q(i) \) real matrix and \( { \mathbf a } (i) \) being \( q(i)\)-dimensional real vectors, the real vector sequence \( { \mathbf y } \) whose terms are, for \( n \geq 0 \), \[ { \mathbf y }(n) = { \mathbf x } + \left\{ \; \sum \;{ \mathbf Z }( n \; | \; i ) \; \cdot \;{\mathbf a }(i) \; \| \; 0 <i\leq k\;\right\} \] converges under suitable conditions to \( { \mathbf x } \). The authors consider the problem of extracting a limit or formal limit of a sequence \( { \mathbf x }(n) \) (\( n \geq 0 \)) by fitting terms of a sequence of the form \( { \mathbf y } \) to those of \( { \mathbf x } \) for \( n = m, m+1, \ldots , m+r \) for suitable \( r \). The \( { \mathbf Z }( n \; | \; i ) \) are taken to be known, \( { \mathbf x } \) and the \( { \mathbf a } (i) \) are determined during the process of fitting. The value of \( { \mathbf x } \) ( which depends of course upon \( m \) and \( r \)) is the estimate of the required limit obtained from the terms \( {\mathbf x}(m), \ldots , {\mathbf x}(m+r) \). The derivation of the limit makes use of expressions of the form \[ { \mathbf A } - { \mathbf B } \cdot inv( { \mathbf D } ) \cdot { \mathbf C } \] obtained from the matrix whose first rows are those of \( { \mathbf A } \), \( { \mathbf B } \) and whose last rows are those of \( { \mathbf C } \), \( { \mathbf D } \), \( { \mathbf A } \) and \( { \mathbf C } \) having equal column numbers; ``inv'' is the generalised inverse of (possibly nonsquare) \( { \mathbf D } \). An account (from which subsequent treatments have been taken) of the use of expressions of this form in which \( { \mathbf D } \) is taken to be square nonsingular is given in the book of \textit{A. C.Aitken} [Determinants and matrices (1939; Zbl 0022.10005); section 29]. There is no consideration at all of the numerical stability of the processes involved in extracting estimates of a limit \( { \mathbf x } \) in the above. (Stability usually determines the usefulness of a convergence acceleration technique; direct use of methods of computational linear algebra is often a straightforward invitation to disaster.)