Divide and conquer techniques based on rank-one updating have proven fast,
accurate, and efficient in parallel for the real symmetric tridiagonal
and unitary eigenvalue problems and for the bidiagonal singular value
problem.  Although the divide and conquer mechanism can also be adapted
to the real nonsymmetric eigenproblem in a straightforward way, most
of the desirable characteristics of the other algorithms are lost.
In this paper, we examine the problems of accuracy and efficiency that
can stand in the way of a nonsymmetric divide and conquer eigensolver
based on low-rank updating.