KSS calculation: diagonalized eig differ from gradient eig

[Mamikon Gulian]

Hello,

When calculating a KSS file I periodically get this warning every few k-points:

outkss.F90:1403:WARNING
The diagonalized eigenvalues differ by more than 10^-3 Hartree
with respect to the conjugated gradient values.

My questions: (1) If its calculating both the Conjugated Gradient values and Diagonalized values and then comparing, can I just get it to write out two KSS files instead of one, for each method? (that way I can really see what the difference between the two is). (2) How should I alleviate this difference? Should I use tighter convergence parameters? On wavefunctions? Energy?

Thanks,
Mamikon

[bruneval]

Hello,

I think this message is not really robust. You should not carry too much attention to it.
Just make sure that your eigenvalues/wavefunctions are well converged within the CG technique.
Sorry for this sloppy reply.

Fabien

[gmatteo]

The diagonalization step performed during the KSS generation can be seen as an additional SCF iteration.
The diagonalized eigenvalues and eigenfunctions are therefore close to the ones calculated during the
SCF (NSCF) run provided that the SCF (NSCF) cycle is very well converged.

You don't have your input file but I guess that only the high energy states produced
by the direct diagonalization differ from the conjugate gradient ones.
To get rid of the annoying message you might do the following:

* Use a stricter criterion for tolwfr (< 10^-16)
* Increase the number of bands (nband) as well as the value of nbdbuf such that
(nband - nbdbuf) spans the energy range you are planning to use for GW.

The highest states are indeed difficult to converge when iterative techniques (e.g CG) are used.
and the nbdbuf trick partly cures the problem.

Note that tolwfr is the recommended stopping criterion to be used to generate the KSS file,
even when the KSS is produced at the end of a SCF run.
Neither toldfe nor tolvrs are good estimators for the accuracy of the band structure
as they only probe the converge of the occupied states. Besides the energy functional
is variational with respect to density variations...

Hope it helps
Matteo Giantomassi

[Mamikon Gulian]

Matteo, Fabien,

Thanks for your replies. I did feel that the warnings were not very dangerous, but wanted to take care of them anyway. I'll try out your suggestions next time I calculate a KSS file and post the results.

-Mamikon Gulian

[Mamikon Gulian]

Just an update: I increased the convergence parameters and that made the warnings vanish, but as it turned out, this caused no difference in results- that is, the previous settings were already well converged despite the warnings. I guess the warnings were not essential.