Too many quasars near one another

[londumas]

This is linked in a way to the following tickets:

• extra covariance in cross-correlation along the line-of-sight: extra covariance in cross-correlation along the line-of-sight #50
• Improve the smaller scale clustering of quasars auto-correlation: Improve the smaller scale clustering of quasars auto-correlation #31

When comparing the number of pairs for the cross between data and mocks, we observe an important difference for small rt, i.e. along the line of sight, and for rp<0.
This is linked to the important number of pairs of quasars in mocks at r~0, which can also be observed there #31 (comment).
One way to correct that is to improve the small scale 3d clustering of quasars.

[fjaviersanchez]

If I understand correctly the issue and I may weigh in: CoLoRe's small-scale power is not going to be perfect for scales smaller than 30 Mpc/h because it's a lognormal simulation. It also implements linear growth which at scales smaller than that is unclear that will work. Moving to something more complicated like LPT or 2LPT may help with this but won't guarantee having "perfect" small-scales behavior. Plus, using LPT or 2LPT will increase the memory usage at the time of generating the box/quasars, effectively limiting the resolution (which will anyway affect small scales) of the box so this is a tradeoff game. We can try to think of a hybrid approach putting smaller boxes inside of larger boxes but it won't be trivial to do.

Maybe @andreufont wants to comment further about this or has different ideas though.

[andreufont]

Unless the current setting was not acceptable, I'd try to move on with what we have. I agree that it would be nice to have a a better quasar clustering though.

2LPT would be great, since it would also give us the right bispectrum, and I'm sure that at NERSC we could run some of these. May be not 100... Also, it would require modifying the running a bit, since currently we assume we are adding 2 Gaussian fields.

A simpler solution would be to run CoLoRe with a modified input P(k) such that it results in a better quasar clustering. If we worried that this could break the Lya skewers, we could run CoLoRe twice with the same random numbers, and different P(k) that differed at high-k. This would preserve the right 3D largs scale cross-correlation, while allowing for a better qso auto. But I'm afraid the small scale cross would still be off.