Detecting episodes of star formation using Bayesian model selection

TL;DR

This paper demonstrates that Bayesian model comparison effectively identifies the correct number of star formation episodes in galaxy spectral energy distributions, with accuracy improving at higher signal-to-noise ratios.

Contribution

It introduces a Bayesian framework for detecting the number of star formation episodes in galaxy SEDs using simulated data and compares evidence calculation methods.

Findings

Bayes factors reliably identify the correct model in most cases.

Higher S/N improves model selection accuracy.

SDDR can approximate evidence with sufficient sampling.

Abstract

Bayesian model comparison frameworks can be used when fitting models to data in order to infer the appropriate model complexity in a data-driven manner. We aim to use them to detect the correct number of major episodes of star formation from the analysis of the spectral energy distributions (SEDs) of galaxies, modeled after 3D-HST galaxies at z ~ 1. Starting from the published stellar population properties of these galaxies, we use kernel density estimates to build multivariate input parameter distributions to obtain realistic simulations. We create simulated sets of spectra of varying degrees of complexity (identified by the number of parameters), and derive SED fitting results and evidences for pairs of nested models, including the correct model as well as more simplistic ones, using the BAGPIPES codebase with nested sampling algorithm MultiNest. We then ask the question: is it true - as expected in Bayesian model comparison frameworks - that the correct model has larger evidence?} Our results indicate that the ratio of evidences (the Bayes factor) is able to identify the correct underlying model in the vast majority of cases. The quality of the results improves primarily as a function of the total S/N in the SED. We also compare the Bayes factors obtained using the evidence to those obtained via the Savage-Dickey Density Ratio (SDDR), an analytic approximation which can be calculated using samples from regular Markov Chain Monte Carlo methods. We show that the SDDR ratio can satisfactorily replace a full evidence calculation provided that the sampling density is sufficient.