Some estimates of the rate of convergence for the cascadic conjugate-gradient method (Q1361289)

The paper deals with a cascadic conjugate gradient (CG) method proposed by \textit{P. Deuflhard} [Contemp. Math. 180, 29-42 (1994; Zbl 0817.65090)]. The idea of the algorithm is as follows: Given a sequence of finite-dimensional vector spaces \(M_{0},M_{1},\ldots, M_{l}\) equipped with inner products \((\cdot,\cdot)_{i}\), \(i=0,1,\ldots,l\). Let also linear ``interpolation'' operators \(I_{i}:M_{i}\rightarrow M_{i+1}\) and linear invertible operators \(L_{i}:M_{i}\rightarrow M_{i}\), \(i=0,1,\ldots,l\) be given. For a given \(f_{l}\in M_{l}\), we want to find a \(v_{l}\in M_{l}\) such that \(L_{l}v_{l}=f_{l}\). For this purpose we solve successively the problems of the form: find \(v_{i}\in M_{i}\) satisfying \(L_{i}v_{i}=f_{i}\) , \(i=0,1,\ldots,l\), where \(f_{i}\in M_{i}\) is given. Let \(u_{0}=L_{0}^{-1}f_{0}\). Starting from \(w_{i}=I_{i-1}u_{i-1}\) as initial approximation, the \(i-\)th equation (successively for \(i=1,2,\ldots,l\)) is approximately solved by a number of steps of the conjugate gradient method. The author defines this method in detail for discrete selfadjoint positive definite problems on a sequence of grids and proves the convergence of the algorithm for this case. As main result (Theorem 4.1), he gives an estimation of the error in the \(i\)th CG process as a function of the executed CG-iteration steps number and the differences between the exact solutions \(v_{j}\) of each \(j-\)th (\(j=1,2,\ldots,i\)) equation and the value of the ``interpolation'' operator \(I_{j-1}\) on the exact solution \(v_{j-1}\) of the former equation. The algebraic properties of the algorithm are verified for a finite element formulation for the elliptic second-order Dirichlet problem in a convex polygon. The paper concludes with some estimations of the number of iteration steps in the CG algorithm for the \(i-\)th equation, necessary for getting an acceptable accuracy of its approximate solution.