[Pw_forum] Problem with nbnd in clusters

[Paolo Giannozzi]

On Nov 20, 2007, at 15:55 , Cyrille Barreteau wrote:

> Does anyone has experience about the behaviour
> of the various diagonalization schemes?
> Which one is better, in which case?

the so-called conjugate-gradient diagonalization basically
never fails; it uses less memory but it is much slower than
the block Davidson algorithm. The latter is also quite
robust but there is a tiny probability to end up in the wrong
ground state. The only such case I know happens in some
high-symmetry perovskites, if you start from superposition
of atomic states, with occupied states only, and if the phase
of the moon is close to 42. The problem disappears if you
add a few more states, or if you start from random wavefunctions.
I am quite surprised to hear that there is a case in which
you get good results with 12 states and bad with 20.
I wouldn't be surprised by the opposite. Could you please
post one jobs showing the problem?

Paolo
---
Paolo Giannozzi, Dept of Physics, University of Udine
via delle Scienze 208, 33100 Udine, Italy
Phone +39-0432-558216, fax +39-0432-558222