> > This is an optimistic assumption; in my modest experience I have found > that a ghosts at E < E_atomic_state always breaks the potential, at E < 1 > Ry usually breaks it, and at E < 5~6 Ry causes erratic and slow > convergence. > Interesting .. Have you found this problem only with pw ? I guess that ghost states that introduce large differences between the log derivatives in the bonding energy window, will create spurious features in the potential like more minima, walls, etc, that break the code. But if in the presence of ghosts the log derivatives are the same in the bonding energy window then the ghost should be benign and the code should run fine.
> Try to test it in pw in a single atom configuration: it may converge fine > as long as you start from atomic wavefunctions. But if you start with some > random wavefunctions it converge slowly to a deeper "haunted" > configuration. > Well in this case there is plenty of room for improvements. Probably just increasing rcloc as you said and/or including a d projector will bust the ghosts. I was more interested in understanding what may be the reasons for pw and the method in general to break down in the presence of ghosts like those in this example. Thanks for the exchange .. bye, L.
