Conditioning of pseudospectral matrices for certain domain decompositions (Q1976423)

The authors examine the conditioning of systems resulting from conforming Chebyshev spectral approximations in nonconforming domain decompositions in rectangular domains. The resulting matrices are large, relatively sparse and have block structures. When solving fourth order systems, the resulting matrices suffer from poor conditioning. The authors show how some direct methods of solution are effected by this poor conditioning. They describe an iterative refinement procedure and demonstrate that its use can lead to significant overall improvement in the overall accuray in the results.