Lyman-alpha is applied twice in desisim.templates.QSO.make_templates()

[londumas]

As reported here #234 (comment), the Lyman-alpha absorption is applied twice in desisim.templates.QSO.make_templates().
The issue is that the PCA components used to get the quasar continuum are calculated from BOSS data corrected for absorption by the Lyman-limit but not by the Lyman-series. When adding lyaforest=True, the code applies the Lyman-alpha forest another time.
All of this contributes to produce unrealistic quasar spectra that are hard to interpret in terms of redshift fitting for example.

This can be corrected by computing new PCA from corrected BOSS data.

[andreufont]

@londumas - was this fixed? If so, should we close the issue?

[londumas]

@andreufont, this issue is not fixed. The current PCA use by desisim.QSO have been build without correcting for the Lyman-alpha forest. So the mean transmission is there in a certain way. When the Lyman-alpha fluxtuations are added, then they are present twice. However, it doesn't seem to be a big issue now since we don't use desisim.QSO anymore.

[andreufont]

Thanks @londumas. This seems like a serious issue for desisim.QSO, specially at high-z where the mean flux is quite low and applying twice would make a large difference.

[moustakas]

I'm in the process of updating the stellar templates, and implementing some minor updates to the ELG, LRG, and BGS templates, so this would be a good time to update the QSO templates, if possible.

Does anyone have any suggestions for how to fix this double Lyman-alpha issue and, more importantly, can someone with the expertise implement the fix?

CC @profxj @andreufont @londumas

[andreufont]

@moustakas - I would suggest that we divide the "blue part" of the simulated spectra (i.e., lambda_rest < 1215.67A) by our best estimate of the mean transmitted flux fraction, F(z). For instance, we could use the fitting function from https://arxiv.org/abs/1307.2259.

The problem with that is that we might cause a discontinuity in the Lyman alpha emission line of the quasars. May be we could smooth the transition by gradually switching between F=1 in the "red part" to F(z) in the "blue part"...

Unfortunately I'll be travelling the next two weeks and completely unable to fix this.

[londumas]

@moustakas, This issue was for desisim.templates.QSO, now desisim uses SIMQSO, so this issue is not that important.
@andreufont, we have tried a long time ago what you just said and it did introduce a big discontinuity:
2nd plot of desihub/redrock#104 (comment)

@moustakas, if you are working on ELG, could you see if you have the time to fix desihub/redrock#111. This would be very usefull for the redshift efficiency at low SNR.