Vector extrapolation methods. Applications and numerical comparison (Q1590781)

The paper is concerned with methods for accelerating the convergence of a sequence \( {\mathbf s}(n) ( n \geq 0) \) of finite dimensional real vectors. The first three methods involve the transform \[ {\mathbf t} ( k||m) = \Bigl\{ \sum \beta (m|j) {\mathbf s} (m+j) \mid 0 \leq j \leq k \Bigr\} \] where the barycentric coordinates \( \beta \) satisfy the relationship \( \{ \sum\beta(m|j) \mid 0 \leq j \leq k \} = 1 \) and the further orthogonality conditions \[ \Bigl\{ \sum \alpha(m||i|j) \beta(m|j) \mid 0 \leq j \leq k \Bigr\} = 0 \] for \( 0 \leq i \leq k \) where, \( \Delta \) being the operator for which \( \Delta {\mathbf s}(n+i) \) is \( {\mathbf s}(n+i+1) - {\mathbf s}(n+i) \) and \( \Delta^{2} {\mathbf s} (n+i) \) is \( \Delta {\mathbf s}(n+i+1) - \Delta {\mathbf s}(n+i) \), in the first method \( \alpha(m||i|j) \) is \( (\Delta {\mathbf s} (m+i), \Delta {\mathbf s}(m+j)) \), in the second it is \( (\Delta^{2} {\mathbf s} (m+i), \Delta {\mathbf s}(m+j)) \) and in the third it is \( ( {\mathbf y}(i), \Delta {\mathbf s}(m+j)) \), the \( {\mathbf y } \) being prescribed. The transform is applied in iterated delayed form to the recursive scheme \( {\mathbf s}(n+1) =\psi ({\mathbf s}(n)) (n \geq 0) \) as follows: (i) determine \( {\mathbf s}(i) ( 0 \leq m) \) from \( {\mathbf s}(0) \); (ii) determine \( {\mathbf s}(m+j) (0 < j \leq k) \) from \( {\mathbf s}(m) \) and from \( {\mathbf t} (k||m) \), then set \( {\mathbf s}(0) = {\mathbf t}(k||m) \) and return to (i). The computations may be terminated by inspecting \(||{\mathbf s}(m+1) - {\mathbf s}(m) ||\) at the beginning of stage (ii). In the use of the vector epsilon - algorithm, stage (ii) becomes: (ii) determine \( {\mathbf s}(m+j) ( 0 < j \leq 2r) \) from \( {\mathbf s}(m) \) and form \( {\mathbf \varepsilon}(2r||m) \). Use of a constrained version of the vector epsilon - algorithm is similar. The vector sequences considered arise from the solution of the equation \( {\mathbf Ax} ={\mathbf b} \) in the form \( {\mathbf s}(n+1) = {\mathbf Bs}(n) + {\mathbf d} (n \geq 0) \) where \( {\mathbf B} = {\mathbf I}-\text{inv}({\mathbf M}) {\mathbf A} \) and \( {\mathbf d}=\text{inv}({\mathbf M}) {\mathbf b} \), \({\mathbf M}\) being a suitable matrix. The sparse equations treated derive from discretized boundary value problems.