A study of semiiterative methods for nonsymmetric systems of linear equations (Q1070763)

If one formulates the linear matrix equation \(Ax=b\) in the standard way in the equivalent fixed point form \(Fx=x\) then given any \(x_ 0\) a sequence of iterates is defined by \(x_{m+1}=Fx_ m\). It is natural to ask whether one can improve on the convergence speed of the sequence by introducing a new sequence \(\{y_ m\}\) defined by \(y_ m:=\sum^{m}_{i=0}\pi_{m,i}x_ i(m\geq 0).\) Any such method which is obtained by transforming the sequence \(\{x_ m\}\) into a sequence \(\{y_ m\}\) is called a semiiterative method (SIM). Such a sequence transformation can be executed in various forms. The second section of the paper under review is devoted to this topic. How to practically compute the iterates \(y_ m\) is the topic of the third section of the paper. It turns out that so-called column-constant nodal matrices lead to a simple computational scheme. Special emphasis is on so-called Euler methods. An Euler method is a SIM in which the coefficients are derived from the Taylor coefficients of the mth power of a certain meromorphic function. It is shown that the matrix of the above mentioned coefficients \(\pi_{m,i}\) can be obtained from the matrix of the Taylor coefficients by a simple similarity transformation. The asymptotic convergence factor is investigated in Section 4. The topic of Section 5 is the construction of SIMs. In Section 6 of the paper regions of convergence are studied which for cyclic methods reduce to lemniscates. In the last section it is shown how to use Faber polynomials for defining asymptotically optimal SIMs. The paper is mainly a survey article with scattered new results.