Improve lss_theory

[damonge]

A possibly not complete list of things that should be implemented in lss_theory.py is:

• generalize to correlation functions. This is not currently in CCL, but will soon be
• convolve multipoles with window function (currently C_ell only computed at band centres)
• add flags for RSDs and magnification (magnification bias should then be passed as parameters too)
• implement a cleverer/more general way of passing parameters (?)
• make bias interpolation order a parameter (?)
• make sure we're happy with the way bias parameters are currently being passed
• do we want to start implementing other tracers (e.g. cmb, lensing...)?
• benchmark likelihood evaluation
• include simple photo-z systematics (?)

[damonge]

I imagine the answer to the penultimate one is no.

Regarding the last one: currently the bottleneck is the calls to CCL (which is good news). Within CCL, if you use a fast power spectrum (e.g. Eisenstein & Hu or BBKS), the bottleneck is the LOS integrals. If you use CLASS, the bottleneck is actually the call to CLASS.

[jeremyneveu]

What does the second point mean exactly ? Sky window or redshift window ?

[damonge]

Basically, after having deconvolved any sky mask/weights/contaminants you'll be left with an estimate of the power spectrum at certain ells which is a convolution of the true C_ell with some window function (in ell). SACC has some capability to store these, but we're not using them (we just assume it's enough to evaluate the theoretical C_ell at the centre of each bandpower).
This is probably not a super bad approximation, so I would focus on the other tasks for the time being.