complex eigenvalues

[Peter Dobcsanyi]

On Sun, Jun 20, 2004 at 04:36:14PM +0100, Steven G Gilmour wrote:
> I was talking to some numerical type people and this came up. I think a
> routine for general matrices is being used,

All numerical routines in LAPACK for eigenvalue computation work as the
matrix was complex.

> when perhaps there is a better one for symmetric matrices.

Depends what "a better one" means in terms of numerical errors.  There
is another routine in the library: Heigenvalues(A) returns the (real
positive) eigenvalues of the square, Hermitian positive definite matrix A.

I have modified my program to use this routine and re-ran the
computations on a few thousands designs in the internal/database
collection. The results are the same, apart from a single case of
"rounding difference". I am convinced that Hermitian() does something
very similar for getting real values to what I did before and outlined
in my previous post instead of using a really different algorithm for
Hermitian matrices.

Anyway, I keep this new version if there is no objection.

--             ,
    Peter Dobcsanyi