The SAMI Galaxy Survey: a statistical approach to an optimal classification of stellar kinematics in galaxy surveys

TL;DR

This paper develops a statistical method using Bayesian mixture models to classify galaxy stellar kinematics from large integral field spectroscopy surveys, accounting for observational effects and improving population separation.

Contribution

It introduces a Bayesian mixture modeling approach for classifying galaxy kinematic populations, validated on SAMI survey data and cosmological simulations, highlighting the importance of data quality in kinematic analysis.

Findings

Bayesian mixture models effectively separate kinematic populations.

Visual classifications complement statistical methods for better accuracy.

The model predicts bimodal distributions consistent with simulations.

Abstract

Large galaxy samples from multi-object IFS surveys now allow for a statistical analysis of the z~0 galaxy population using resolved kinematics. However, the improvement in number statistics comes at a cost, with multi-object IFS survey more severely impacted by the effect of seeing and lower S/N. We present an analysis of ~1800 galaxies from the SAMI Galaxy Survey and investigate the spread and overlap in the kinematic distributions of the spin parameter proxy $\lambda_{Re}$ as a function of stellar mass and ellipticity. For SAMI data, the distributions of galaxies identified as regular and non-regular rotators with \textsc{kinemetry} show considerable overlap in the $\lambda_{Re}$-$\varepsilon_e$ diagram. In contrast, visually classified galaxies (obvious and non-obvious rotators) are better separated in $\lambda_{Re}$ space, with less overlap of both distributions. Then, we use a Bayesian mixture model to analyse the observed $\lambda_{Re}$-$\log(M_*/M_{\odot})$ distribution. Below $\log(M_{\star}/M_{\odot})\sim10.5$, a single beta distribution is sufficient to fit the complete $\lambda_{Re}$ distribution, whereas a second beta distribution is required above $\log(M_{\star}/M_{\odot})\sim10.5$ to account for a population of low-$\lambda_{Re}$ galaxies. While the Bayesian mixture model presents the cleanest separation of the two kinematic populations, we find the unique information provided by visual classification of kinematic maps should not be disregarded in future studies. Applied to mock-observations from different cosmological simulations, the mixture model also predicts bimodal $\lambda_{Re}$ distributions, albeit with different positions of the $\lambda_{Re}$ peaks. Our analysis validates the conclusions from previous smaller IFS surveys, but also demonstrates the importance of using kinematic selection criteria that are dictated by the quality of the observed or simulated data.