Approximations to galaxy star formation rate histories: properties and uses of two examples

Abstract

Galaxies evolve via a complex interaction of numerous different physical processes, scales and components. In spite of this, overall trends often appear. Simplified models for galaxy histories can be used to search for and capture such emergent trends, and thus to interpret and compare results of galaxy formation models to each other and to nature. Here, two approximations are applied to galaxy integrated star formation rate histories, drawn from a semi-analytic model grafted onto a dark matter simulation. Both a lognormal functional form and principal component analysis (PCA) approximate the integrated star formation rate histories fairly well. Machine learning, based upon simplified galaxy halo histories, is somewhat successful at recovering both fits. The fits to the histories give fixed time star formation rates which have notable scatter from their true final time rates, especially for quiescent and "green valley" galaxies, and more so for the PCA fit. For classifying galaxies into subfamilies sharing similar integrated histories, both approximations are better than using final stellar mass or specific star formation rate. Several subsamples from the simulation illustrate how these simple parameterizations provide points of contact for comparisons between different galaxy formation samples, or more generally, models. As a side result, the halo masses of simulated galaxies with early peak star formation rate (according to the lognormal fit) are bimodal. The galaxies with a lower halo mass at peak star formation rate appear to stall in their halo growth, even though they are central in their host halos.

How to Measure Galaxy Star Formation Histories. I. Parametric Models

“…The four widely used parametric SFH models we consider (and some recent examples of their use in the literature) are: exponentially declining (Mortlock et al 2017;Wu et al 2018;McLure et al 2018), delayed exponentially declining (Ciesla et al 2017;Chevallard et al 2017), lognormal (Diemer et al 2017;Cohn 2018) and double power law (Ciesla et al 2017;Carnall et al 2018). For brevity we will refer to them as tau, delayed, lognormal and DPL models respectively.…”

Learning the relationship between galaxies spectra and their star formation histories using convolutional neural networks and cosmological simulations

“…Extremely Randomised Trees (ERT; Geurts et al 2006) is one such ensemble approach that has been successfully used in a wide range of Astronomy domains (e.g. Kamdar et al 2016;Cohn 2018). It is similar to the popular Random Forest (RF): during training of a RF, a subset of K features is randomly chosen during each split, which reduces the correlation between trees where there are features with a strong correlation with the predictors.…”

Applying unsupervised learning to resolve evolutionary histories and explore the galaxy--halo connection in IllustrisTNG

“…Diemer et al fitting SFHs using log-normals, found that the scatter in star formation histories could not be uniquely determined by a single parameter. Cohn (2018), building off of this work, compared Diemer et al (2017)'s log-normal model to one developed with PCA. Although PCA is able to reproduce simulated SFHs, the authors point out that the underlying assumption of PCA -that scatter seen in the simulation is but a deviation from a single underlying star formation history -is unlikely to be true.…”