The approximate Dirichlet domain decomposition method. I: An algebraic approach (Q1180345)

[For part II see ibid. 47, No. , 153-167 (1991; see below).] Linear systems \(Au=f\), \(A=A^*>0\), \(A=[A_{ij}]\), \(i,j=1,2\) are considered. The taken block structure is especially useful when \(A\) is the Gram matrix for a basis of an Euclidean space \(V\) split into a direct sum \(V_ 1\oplus V_ 2\) and the angle between \(V_ 1\) and \(V_ 2\) is close to \(\pi/2\). Even if the angle is small it seems promising to construct for \(A\) preconditioners \(B=B^*>0\) of the same block structure and under conditions that the systems with \(B_{22}\) and the Schur matrix \(S_ 1(B)\equiv B_{11}-B_{12}B^{-1}_{22}B_{21}\) are ``easy to solve''. Such \(B\) are called cooperative operators and were key subjects of several papers. The given references can be essentially widen by references in the reviewer's paper in ``Computer Mechanics of Solids'', Moscow, 1991, Issue 2, p. 133-157 (ISBN 5-02-007319-9), devoted to multigrid preconditioners. The main algebraic problem is to give suitable conditions on \(B_{22}\) as approximations to \(A_{22}\) and \(S_ 1(A)\) leading to the spectral equivalence of the grid operators \(A\) and \(B\). The authors suggest such conditions, especially suited for the case of domain decomposition methods with the space \(V_ 1\equiv V_ C\), generated by basis finite element functions corresponding to nodes on cutting lines (on interfaces) of the given region.