Testing the accuracy of likelihoods for cluster abundance cosmology

TL;DR

This study evaluates the accuracy of different likelihood functions used in galaxy cluster abundance cosmology, finding Gaussian likelihoods robust for large surveys, while Poisson underestimates errors, and more complex likelihoods offer limited benefits.

Contribution

The paper systematically compares Poisson, Gaussian, and Gauss-Poisson likelihoods for cluster abundance analysis using simulated data, clarifying their accuracy and computational efficiency.

Findings

Gaussian likelihood provides robust constraints for large surveys.

Poisson likelihood underestimates parameter errors.

Gauss-Poisson likelihood offers no significant advantage over Gaussian.

Abstract

The abundance of galaxy clusters is a sensitive probe to the amplitude of matter density fluctuations, the total amount of matter in the Universe as well as its expansion history. Inferring correct values and accurate uncertainties of cosmological parameters requires accurate knowledge of cluster abundance statistics, encoded in the likelihood function. In this paper, we test the accuracy of cluster abundance likelihoods used in the literature, namely the Poisson and Gaussian likelihoods as well as the more complete description of the Gauss-Poisson Compound likelihood. This is repeated for a variety of binning choices and analysis setups. In order to evaluate the accuracy of a given likelihood, this work compares individual posterior covariances to the covariance of estimators over the 1000 simulated dark matter halo catalogs obtained from PINOCCHIO algorithm. We find that for Rubin/LSST or Euclid-like surveys the Gaussian likelihood gives robust constraints over a large range of binning choices. The Poisson likelihood, that does not account for sample covariance, always underestimates the errors on the parameters, even when the sample volume is reduced or only high-mass clusters are considered. We find no benefit in using the more complex Gauss-Poisson Compound likelihood as it gives essentially the same results as the Gaussian likelihood, but at a greater computational cost. Finally, in this ideal setup, we note only a small gain on the parameter error bars when using a large number of bins in the mass-redshift plane.