How do uncertainties in galaxy formation physics impact field-level galaxy bias?

TL;DR

This work investigates how uncertainties in the galaxy–halo connection influence field-level galaxy bias by comparing two complementary galaxy-formation models, UniverseMachine and IllustrisTNG, across quenched, star-forming, and M*-selected samples at multiple redshifts. Using a second-order Hybrid Effective Field Theory (HEFT) bias expansion with a cubic term and a stochastic component, the authors extract bias parameters and examine assembly bias effects, offering priors to improve cosmological constraints from LRG- and ELG-like tracers. They find that quenched and star-forming populations occupy distinct regions in the bias-parameter space, and that the scatter between UM and TNG is modest when samples are matched; removing assembly bias aligns results more closely with halo-model predictions, particularly for the linear and nonlocal biases. The study provides robust, physically motivated priors on bias relations and time evolution of the linear bias, which can be directly used in EFT-based cosmological analyses to reduce prior-volume effects and tighten constraints for current and future surveys.

Abstract

Our ability to extract cosmological information from galaxy surveys is limited by uncertainties in the galaxy-dark matter halo relationship for a given galaxy population, which are governed by the intricacies of galaxy formation. To quantify these uncertainties, we examine quenched and star-forming galaxies using two distinct approaches to modeling galaxy formation: UniverseMachine, an empirical semianalytic model, and the IllustrisTNG hydrodynamical simulation. We apply a second-order hybrid N-body perturbative bias expansion to each galaxy sample, enabling direct comparison of modeling approaches and revealing how uncertainties in the galaxy-halo connection affect bias parameters and non-Poisson noise across number densities and redshifts. Notably, we find that quenched and star-forming galaxies occupy distinct parts of the bias parameter space, and that the scatter induced from these different galaxy formation models is small when conditioned on similar selections of galaxies. We also detect the signature of assembly bias in our samples; this leads to small but significant deviations from analytic bias predictions, while assembly bias-removed samples match these predictions well. This work indicates that galaxy samples from a spectrum of reasonable, physically motivated models for galaxy formation give a relatively small range of field-level galaxy bias parameters. We estimate a set of priors from these models that should be useful in extracting cosmological constraints from luminous red galaxy- and emission line galaxy-like samples. Looking forward, careful estimates of the range of impacts of galaxy formation, for a given sample and cosmological analysis, will be an essential ingredient for extracting the most precise cosmological information from current and future large galaxy surveys.

Figures (15)

• Figure 1: Visualization of the methodology of this work: starting from the same initial conditions (TNG300), we obtain both the galaxy density field $\delta_g$ for each model (UM and TNG, with and without assembly bias) and the noiseless linear density field $\delta_L$. Using the latter, we construct the bias-weighted fields $\mathcal{O}_i$ at redshift $z$ via advection. These constitute a deterministic model $\delta_{g,\mathrm{ model}}$ that differs from the galaxy density field $\delta_g$ only by a stochastic term $\epsilon$ (see Equation \ref{['eq:bias expansion']}), which we minimize to solve for the best-fit bias parameters of each sample. Finally, we compare results between the models and set priors for use in cosmological inferences. In this figure, we visualize the density field $\delta_g$ and stochasticity $\epsilon$ at $z=0.5$ for a high number density sample in TNG.
• Figure 2: A 2D histogram of $M_*$ and sSFR for TNG and UM galaxies at redshifts $z=0$ (top panel) and $z=1.5$ (bottom panel). The red and blue shaded regions mark the $M_*$ and sSFR cuts which separate the quenched and star-forming samples, respectively. The number density for each population is shown in the legend (medium number density, medium $M_*$).
• Figure 3: The bias parameter relations as a function of $b_1$ at $z=\{0.0, 0.5, 1.0, 1.5\}$ for each sample of galaxies in UM (triangles) and TNG (circles), including quenched (red), star-forming (blue), and the $M_*$-selected population (purple). The color intensity (light to dark) maps to the redshift (low to high). The error bars are calculated according to the steps outlined in Appendix C of Kokron_2022. Best-fit relations for halos 2016JCAP...02..018L and galaxies 2021JCAP...08..029B2022MNRAS.514.5443Z from past works are shown as comparison.
• Figure 4: The bias parameter measurements of TNG galaxies in comparison to UM galaxies for quenched (red), star-forming (blue), and $M_*$-selected samples. The color intensity (light to dark) maps to the redshift (low to high). A 1:1 dashed line is plotted for reference.
• Figure 5: The effect of AB removal on each of the bias parameter measurements. The new measurements without AB are plotted against the original measurements with AB at $z=\{0.0, 0.5, 1.0, 1.5\}$ for each sample of galaxies in UM (triangles) and TNG (circles), including quenched (red), star-forming (blue), and the $M_*$-selected population (purple). The color intensity (light to dark) maps to the redshift (low to high). A 1:1 dashed line is plotted for reference.