Nested grids ILU-decomposition (NGILU) (Q1919400)

The authors construct incomplete LU (ILU)-preconditioners \(LU\) for large scale \(N \times N\) matrices \(A\) arizing typically from the finite difference, or finite element discretizations of partial differential equations. The construction of the ILU-preconditioner \(LU\) is based on some multilevel renumbering of the unknowns and on the splitting \(A=LU+R\) such that the entries \(r_{ij}\) of \(R\) satisfy the condition \(|r_{ij} |\leq \varepsilon_{ij}\) with some prescribed threshold parameter \(\varepsilon_{ij}\). Similar to other multilevel precoditioners, the multilevel renumbering of the unknowns takes care of the lower frequencies, too. Therefore, one can expect that the number of conjugate gradient (cg) or cg-like iterations required for solving the system \(Ax=b\) with some prescribed relative accuracy doesn't significantly depend on the number \(N\) of unknowns. Indeed, the numerical examples given in the paper confirm this conjecture.