florpi/sunbird

COsmOlogical inference with dEnsity Split cluStering

SUNBIRD

SUNBIRD is a Python package that provides routines to train neural-network-based models for galaxy clustering. It also incorporates pre-trained models for different summary statistics, including:

• Galaxy two-point correlation function
• Density-split clustering statistics
• Void-galaxy cross-correlation function.

These models have been trained on mock galaxy catalogues based on the AbacusSummit simulations. The models are described in detail in Cuesta-Lazaro et al. (in preparation).

Documentation

Documentation is hosted on Read the Docs, pysunbird.readthedocs.io.

Requirements

The following packages are required to run the code:

• black
• pytorch
• pandas
• numpy
• matplotlib
• xarray
• pytorch-lightning
• optuna
• joblib

Installation

Cloning from repository

First

git clone https://github.com/florpi/sunbird.git

To install the code

python setup.py install --user

Or in development mode (any change to Python code will take place immediately)

python setup.py develop --user

sunbird's Issues

Implementing number density as an observable

• The observed galaxy number density measured from spectroscopic surveys places a strong constrain on the range of viable HOD models for a given summary statistic.

• We should be able to produce a dense sample of HOD catalogues for each of the AbacusSummit cosmologies, saving the (unfiltered) number densities to disk. We can then train a simple FCN to learn to predict this number density based on the HOD + cosmological parameters.

• During the MCMC, we can calculate the usual \chi^2 for the multipoles, but now we additionally add a \chi^2 that compares the observed v/s predicted number density. This can be a truncated Gaussian distribution as in Eq. (18) from http://arxiv.org/abs/2110.11412, which penalizes number densities that are lower than the target, but for higher number densities we invoke an incompleteness factor f_ic that in practice downsamples the HOD catalogues to the correct number density (this step already takes place when we create the HODs to train the summary statistics, so it's just a matter of reflecting this in the likelihood).

I'll go ahead and measure the HOD number densities that we can use to train the FCN, and then we can decide on what's the best way to incorporate this into the pipeline.

Save likelihood/posterior of HMC chains

For the HMC sampler, Is it possible to save the likelihood and/or posterior values for each step of the chain? This is needed to report the best-fit / MAP values as we currently do in the papers with the Dynesty sampler.