Minimal polynomial and reduced rank extrapolation methods are related (Q2361155)

Let \(\{\mathbf{x}_m\}\) be a vector sequence in \(\mathbb{C}^N\). It is assumed that this sequence results from the fixed-point iteration solution of the system of linear or nonlinear equations \[ \mathbf{x}=f(\mathbf{x}), \quad \text{solution}\; s; \; \mathbf{x}\in \mathbb{C}^N \; \text{and}\; f:\mathbb{C}^N \to \mathbb{C}^N, \] from \[ \mathbf{x}_{m+1}=f(\mathbf{x}_m), \quad m=0,1,\dots, \] \(\mathbf{x}_0\) being an initial vector chosen by the user. Minimal polynomial extrapolation (MPE) and reduced rank extrapolation (RRE) are two polynomial methods used for accelerating the convergence of sequence of vectors \(\{\mathbf{x}_m\}\). Both methods produce approximations \(\mathbf{s}_k\) to the limit or antilimit of \(\{\mathbf{x}_m\}\) that are of the form \(\mathbf{s}_k=\sum_{i=0}^k \gamma_i \mathbf{x}_i\) with \(\sum_{i=0}^{k}\gamma_i=1,\) for some scalars \(\gamma_i\). This work shows that the vector \(\mathbf{s}_k^{MPE}\) and \(\mathbf{s}_k^{RRE}\) produced by two methods are related in the more than one way, and independently of the way the \(\mathbf{x}_m\) are generated. The results of article contain as special cases the known results pertaining to the connections between the method of Arnoldi and the method of generalized minimal residuals, two important Krylov subspace methods for solving nonsingular linear systems.