RPA calculation gives IPA response

[lilian]

Dear developers and users,

I have observed what might be a bug in the RPA response, or at least something I don't understand.

The variable B=NGsBlkXd determines the size of the dielectric matrix. So if you increase this value, local field effects are taken into account more and the response should move further away from the IPA response.
However, above a certain value of B, which from my observations is lower than the number of G-vectors (determined by FFTGvecs) and depends on other parameters of the simulation (such as the k-point grid used in nscf), the RPA calculation gives the same result as the IPA one.

See the attached figure below, where B= 311 RL (black line) is the last case to give the RPA response. After this threshold, and from the next possible value (313 RL) all RPA responses are perfectly superimposed on the IPA curve.

plot_RPA_IPA_eels.png

Could you please enlighten me on this behavior, as I don't see any physical sense in the RPA response becoming identical to the IPA response beyond a certain matrix size?

Many thanks in advance!

(I can provide a minimal test upon request)

[Daniele Varsano]

Dear Lilian,

please note that B=1 RL is not exactly the IP response, but in this case a scalar Dyson equation is solved. The IP is anyway always reported in the output in last column.
Having said that, your results are rather unusual, can you please post the input/output/report files of your calculations?

Best,

Daniele

[lilian]

Dear Daniele,

Thank you for your answer.
I have attached the inputs, outputs and reports of the three cases shown in the figure in the first post, as well as the QE inputs.

Many thanks in advance!

[Daniele Varsano]

Dear Lilian,

I do not have a rationale on what is happening here.
The only thing I can see in the input file is that you set a quite low FFTGvec value. In general, unless there are serious memory problems, it is safer to leave the FFT dimension to its default otherwise you can have orthonormalisation issues.

My suggestion here is to rerun the setup without setting the MaxGvecs variable, and then repeat your convergences.

Best,

Daniele