A projection method of the Cimmino type for linear algebraic systems (Q1179244)

The classical method of \textit{G. Cimmino} [Ric. Sci. Progr. Tecn. Econom. Naz. 1, 326-333 (1938; Zbl 0018.41802)] for solving square systems of linear equations \(Ax=b\) in the least-squares sense is \(x_{p+1}=x_ p+(2/n), (DA)^ T(b-Ax_ p)\) where \(D\) is a diagonal matrix whose entries are the reciprocals of the norms of the row vectors of \(A\). Here, \(n\) is the format of the matrix, so that for large systems, the stepsize is small and thus the rate of convergence is slow. The author suggests to replace the step size \(2/n\) by a parameter \(\alpha\) and proves that (for general rectangular matrices which fulfill certain diagonal dominance conditions), the algorithm converges for \(\alpha=1\). Moreover, \(\alpha=1\) is the optimal step size. Finally, a few straightforward remarks concerning the implementation on vector or on parallel machines are made.