A case against a divide and conquer approach to the nonsymmetric eigenvalue problem (Q685968)

The divide and conquer method was first studied by \textit{J. J. M. Cuppen} [Numer. Math. 36, 177-195 (1981; Zbl 0431.65022)] for a symmetric tridiagonal eigenproblem. This method is convenient for parallel implementations where smaller subproblems are solved in parallel with one problem per processor. In this paper the symmetric divide and conquer method based on rank-one or rank-two updating is extended in a straightforward way to a nonsymmetric tridiagonal and Hessenberg eigenproblem. In both cases it is supposed that the submatrices (tridiagonal or Hessenberg) formed by matrix tearing are diagonalizable. For an application of this method, an efficient algorithm for finding all roots of a complex rational equation is supposed. The root finding is discussed and the general structure is graphically demonstrated. Stability comments are at the end of this paper.