by **Julien Langou** » Sun Sep 10, 2006 8:35 pm

Hello

how much "negative" are your eigenvalues?

If the absolute value of your negative eigenvalue is smaller than machine epsilon times the norm of A, you should consider this eigenvalue to be zero. (Which make your matrix numerically not definite so you still might not be happy ...)

If you are looking for more accurate symmetric eigensolvers, you should give a try to xSYEVX, but this will not help that much. The main problem is that with the current LAPACK solvers you loose componentwise accuracy when we go to tridiagonal form. You are normwise accurate. Some new solvers are developped right now to remedy this problem.

Julien Langou