[Pw_forum] What happens at REALLY large ectuwfc?

[O. Baris Malcioglu]

Dear Brad,

Apart from what is already mentioned, another contribution to error
you are mentioning might be due to computational/numeric issues.

Run-of-the-mill FFT algorithms has  an error upperbound that goes like
O(sqrt(N)). Even in the best case (least performing) scenarios the
error upper bound is O(log(N)).

These errors are due to numeric precision employed, and mostly bound
by hardware (unless you compromise performance significantly).

In general, in any calculation that involves computers, in a graph
that shows the "error" as a function of some convergent parameter, I
would expect three regions , first the error diminishes due to
numerical reasons, then there is a flat region, then roundoff errors
begin to dominate, and error picks up again.

If interested, please have a look at the (rather rusty but still
valid) concept "Machine constant"

Best,
Baris.



2010/1/25 Paolo Giannozzi <giannozz at democritos.it>:
> Brad Malone wrote:
>> I originally saw this problem with espresso-4.1.1 in a
>> different system (AlAs on a 2x2x2 shifted grid).
>
> start with 'random' or 'atomic+random' initial wavefunctions
> (startingwfc='...'). I didn't find anything wrong with very
> high cutoffs. Occasionally you can end up in the wrong ground
> state, though, especially in highly symmetric cases like the
> Si example you sent.
>
> "Ghost states" are another story. It may happen that a
> (nonlocal, separable) pseudopotential has a localized
> spurious state that doesn't show up (i.e. it is not
> occupied) at low cutoff, but it is occupied at higher
> cutoff. None of the PP on the QE web site should have
> such states, though.
>
> If you have evidence of further problems, please provide
> a test case so that it can be reproduced. In any case,
> thank you for finding a bug in the cvs version!
>
> P.
> ---
> Paolo Giannozzi, Democritos and University of Udine, Italy
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
>