michaeljwilson / desi-berkeley Goto Github PK

When computing the correlation function, it looks like the NbodyKit code is doing the full Nran x Nran pairs for the "RR" term; This is unnecessary. We're limited by the statistics of the DR term in any case, so we can just choose an "Ndat" sized (random) subset of the randoms and compute "RR" from the cross between that subset and the full random set, then scale up by Nran/Ndat. This will save a factor of Nran/Ndat in time, which could be something like 10x or so. Then both DR and RR will scale as Ndat times Nran. If we want to play it safe we could choose 2x or 4x Ndat of the randoms.

It would be nice to have a "convenience function" which computes multipoles of the Landy-Szalay correlation function given RA, DEC and Z of two data sets and randoms. The SurveyDataPairCount method of nbodykit.algorithms.paircount does the hard work, we just need to normalize the counts by the weight-sums, and integrate against Legendre polynomials. That and "wedges" would be the most common use-cases I suspect.

The correlation function BAO fitting code needs a covariance matrix and the code currently just ingests the BOSS mock-based covariance or a simple no-boundary computation of integrating P^2(k) against Bessel functions.

A portable application shall not rely on any user modifying their bash environment.

If the intention is to provide a user enrionment for interactive analysis, then consider adding activate.sh and deactivate.sh for entering and leaving the environment?

If the intention is to provide a single point of entry then perhaps migrate some of the settings to a config file that can be parsed and loaded easily.