On the vector \(\varepsilon\)-algorithm for solving linear systems of equations (Q5961063)

The following theory is easily constructed. (i) With \( m,r \geq 0 \), the polynomial \( P(z \|m,r) = [\sum a(m,r|k) z^{k} |0 \leq k \leq r ] \) satisfying the barycentric normalising condition \( P(1 \|m,r) = 1 \) may be associated with a transformation \( {\mathbf t}({\mathbf x} \|m,r) = [\sum a(m,r|k) {\mathbf x}(m+k) |0 \leq k \leq r ] \) of the subsequence \({\mathbf x}(m), \ldots , {\mathbf x}(m+h) \) of a sequence \({\mathbf x} \) of real \( n \)-dimensional vectors. (ii) \( {\mathbf A } \) being a suitable square matrix, \( {\mathbf x } \) may be generated by use of the scheme \( {\mathbf x}(j+1) = {\mathbf A}{\mathbf x}(j) + {\mathbf b} \) \( (j \geq 0) \). \({\mathbf v} \) being a suitable vector, its residual is defined by setting \( res \{ {\mathbf v} \}= ({\mathbf A} - {\mathbf I}) {\mathbf v} + {\mathbf b} \). \( {\mathbf t} \) may then be expressed as a linear combination of residuals: \( {\mathbf t}({\mathbf x} \|m,r) = {\mathbf x}(m) + [\sum c(m,r|k) {\mathbf A}^{k} res \{ {\mathbf x}(m) \}|0 \leq k < h ] \) where \( c(m,r|k) = [\sum a(m,r|j) |k < j \leq r] \); the residual of the transformed estimate \({\mathbf t} \) is a matrix multiple of that of \( {\mathbf x}(m) \): \(res \{ {\mathbf t}({\mathbf x} \|m,r) \} = P({\mathbf A }\|m,r) res\{{\mathbf x}(m) \} \). (iii) Let the \( p \times (r+1) \) matrix \( {\mathbf C}(m,r) \) serve as a homogeneous constraint system in the sense that, the components of the column vector \( {\mathbf a}(m,h) \) being the \( a(m,h|k) \), \( {\mathbf C}(m,h) {\mathbf a}(m,h) = {\mathbf 0} \) (the column having \( p \) zero components). (The definition of \( {\mathbf t} \), the normalising condition and the constraint relationship form a system of linear algebraic equations yielding, under suitable conditions, a determinantal quotient expression for \( {\mathbf t} \).) \( {\mathbf F}(m,r) \) being a suitable \( q \times (r+1) \) transformation matrix, define the transform \( {\mathbf g} \) by setting \( {\mathbf g}(m,r) = {\mathbf F}(m,r) {\mathbf a}(m,r) \), and let the \( (r+1) \times q \) matrix \( {\mathbf H}(m,r) \) satisfy the relationship \( {\mathbf C}(m,r) {\mathbf H}(m,r) {\mathbf F}(m,r) = {\mathbf C}(m,r) \). The \( { p \times q } \) matrix \( {\mathbf K}(m,r) = {\mathbf C}(m,r) {\mathbf H}(m,r) \) is then an annihilator of \( {\mathbf g} \) in the sense that \( {\mathbf K}(m,r) {\mathbf g}(m,r) = {\mathbf 0} \). (With \( r < q \) and \( {\mathbf F} \) being suitable, \( {\mathbf H}(m,r) \) may then be taken to be \( inv\{[transp {\mathbf F}(m,r)] {\mathbf F}(m,r) \} [transp {\mathbf F}(m,r)] \)). (iv) In terms of clause (ii) it is possible to take \( n = q \) and the successive columns \( {\mathbf F}(m,r) \) to be \( { \mathbf A}^{k} res\{{\mathbf x}(m)\} (0 \leq k \leq r) \), \( {\mathbf H} \) being as just defined in terms of \( {\mathbf F} \). In the paper the above results are stated with regard to a determinantal quotient expression involving a quasi skew symmetric constraint system introduced by the second author. No satisfactory proof that the related determinantal quotient expresses vectors produced by means of the vector epsilon algorithm, as is in particular stated in Theorem 1 of the paper under review, has been published; the offered proof of the theorem is referred back to work that is without substance.