betaFeedback

WorksPhysics and Astronomy

Stress-Testing Small-Scale Cosmological Inference With Empirical Galaxy–Halo Connection Models

Abstract

Small-scale galaxy clustering contains substantial cosmological information, but extracting it reliably requires robust modeling of the galaxy–halo connection. In this work, we compare two widely adopted families of empirical models, subhalo abundance matching (SHAM) and the halo occupation distribution (HOD), testing their predictions for galaxy clustering and their recovery of the input cosmology. Using Aemulus-ββ, a new multi-cosmology simulation suite with substructure merger trees, we fit multiple SHAM and HOD models to the small-scale clustering of a stellar-mass-selected mock galaxy catalog generated with the more physical semi-empirical model UniverseMachine. We then use cosmological evidence modeling to infer the structure growth parameter S8S_8. Our results show that the SHAM model, regardless of the treatment of numerically disrupted subhalos, is able to offer tight, unbiased constraints on S8S_8 with ~2-3% precision, using simulations of relatively small volumes that span a limited number of cosmologies. In contrast, HOD models, even when accounting for assembly bias, yield weaker and potentially biased cosmological constraints that are sensitive to implementation choices: the inferred S8S_8 degrades in precision to ~6-11% and can be offset by ~1σσ. This suggests that the excessive parametric flexibility and model misspecification of most commonly adopted HOD models can degrade and bias the resulting cosmological constraints. We further evaluate each model's performance in predicting higher-order statistics not used in the fit and find that SHAM generally reproduces them more accurately than the tested HOD models. These results highlight the potential of leveraging predictive subhalo-level information from simulations to enable robust cosmological inference on small scales.

Introduction

Understanding the full information content in the distribution of galaxies is one of the main goals of modern astrophysics and cosmology. With the new generation of large sky surveys, such as the Dark Energy Spectroscopic Instrument (DESI, [13 b]), the Euclid mission ([16]), the Roman Space Telescope ([51]), and the Rubin Observatory Legacy Survey of Space and Time (LSST, [26]), we are entering an era of high-precision studies of galaxy clustering across a wide range of scales with galaxy distribution maps of unprecedented volume and density.

To fully exploit these datasets and enhance cosmological constraints, extracting information from small, non-linear scales is essential. These scales contain a large number of measurable modes and can therefore add substantial constraining power (e.g., [28, 65]). However, since analytic perturbation theory breaks down in this regime, dark matter simulations together with robust modeling of the relationship between galaxies and halos ([60]) are necessary to make accurate and robust predictions for relevant statistics. Such models need to be predictive enough to include the full cosmological information, accurate enough to characterize the dependence of the spatial distribution on key observables, and flexible enough to marginalize over uncertainties in galaxy formation.

While hydrodynamical simulations (e.g. [54, 45, 46]) directly model the physical processes of galaxy formation, including baryonic effects, they remain computationally too expensive for iterative evaluations across cosmologies. Semi-analytic models (e.g. [49, 50, 24, 25, 30]) and semi-empirical forward models (e.g. [40, 4, 2]) aim to approximate these physical processes more efficiently and can be calibrated against observational data. However, the large number of model parameters has so far made it difficult to explore a broad galaxy formation parameter space, and the computational cost of constructing these models over large cosmological volumes remains nontrivial for inference applications. Simple empirical galaxy–halo connection models are therefore widely used for their low computational cost and their flexibility to fit observational data.

Two of the most widely adopted empirical frameworks for the galaxy–halo connection are the subhalo abundance matching model (SHAM) and the halo occupation distribution model (HOD). SHAM assigns galaxy properties by ranking (sub)halos according to a chosen halo property and matching their abundances to those of galaxies, generally allowing scatter in the relation ([29, 52, 9, 44]). HOD models instead specify the probability distribution of galaxy occupation as a parametric function of host halo properties, usually beginning with halo mass ([41, 47, 8, 7, 67]). Standard HOD models depend only on host-halo mass and do not explicitly model how occupation at fixed mass varies with secondary halo properties such as formation history or concentration, an effect known as assembly bias ([58, 17, 61, 38]). Such dependencies have been incorporated into HOD models more commonly in recent years (e.g. [22, 63] (Xu et al. 2021)).

SHAM makes use of more information from the simulation, such as the properties and assembly histories of subhalos, and naturally incorporates assembly bias, while HOD models rely only on host-halo catalogs and are thus less computationally demanding. Both empirical models have been broadly employed to fit observational data and study properties of different galaxy samples, and have demonstrated success in reproducing clustering statistics. However, it remains unclear whether they recover cosmology equally reliably when small-scale measurements are included, and how effectively each can properly marginalize over uncertainties in the galaxy–halo connection.

In this work, we investigate requirements for galaxy–halo connection modeling within these empirical frameworks in order to support cosmological inference on small scales. Using AEMULUS-β\beta, a new multi-cosmology simulation suite with sufficient resolution for reliable merger trees and subhalo catalogs, we stress-test multiple SHAM and HOD models against a realistic mock galaxy catalog constructed with the more complex semi-empirical model UniverseMachine ([4]), which provides a physically motivated and observationally grounded galaxy–halo connection. At a fixed fiducial cosmology, we evaluate the accuracy of each model in reproducing two-point clustering down to small scales. We then investigate whether the modeling frameworks can be employed to infer unbiased cosmological parameters using the Cosmological Evidence Modeling (CEM, [34]) method, by fitting the target clustering with models built on all AEMULUS-β\beta cosmologies. The high resolution of AEMULUS-β\beta allows SHAM to be built directly on simulated substructures across different cosmologies for the first time, enabling subhalo-level information to be propagated into the cosmological inference. As an additional predictive test, we compare each model’s predictions for two higher-order statistics not used in the fits, counts-in-cylinders and multiplet clustering, with actual measurements from the target catalog.

This paper is structured as follows. § 2 describes the AEMULUS-β\beta simulation suite and the target mock galaxy catalog constructed using UniverseMachine. § 3 introduces the SHAM and HOD models, and § 4 the statistical methods. In § 5, we examine assembly bias in the UniverseMachine mock and present model fits at fixed cosmology. In § 6, we present and discuss cosmological constraints from different models. We extend the analysis to higher-order statistics in § 8. We discuss limitations and implications in § 9 and conclude in § 10. A schematic overview of the workflow of the paper is shown in Figure 1.

Schematic overview of the workflow of this paper

Figure 1. Schematic overview of the workflow of this paper. Starting from the Aemulus-β\beta simulation suite, we construct a target mock galaxy catalog using UniverseMachine, and build two classes of empirical galaxy–halo connection models, SHAM and HOD, as prediction pipelines. We then fit the empirical models to the clustering of the target catalog via nested sampling, testing their ability to describe the underlying galaxy–halo connection and offer unbiased cosmological constraints.

Simulations and Galaxy Catalogs

Aemulus-β\beta Simulations

We use dark matter catalogs from the AEMULUS-β\beta simulation suite (Alvarez et al. 2026, in prep) to construct various empirical galaxy–halo connection models. The AEMULUS-β\beta suite consists of 25 simulations that span a 7-dimensional wwCDM cosmological parameter space: the baryon density ωb\omega_b, the cold dark matter density ωc\omega_c, the Hubble constant H0H_0, the spectral index nsn_s, the amplitude of scalar perturbations AsA_s, the dark energy equation of state ww, and the effective number of neutrino species NeffN_{\mathrm{eff}}. The 25 cosmologies span the same design volume as the original Aemulus suite ([12]): a Latin hypercube constructed in the eigenbasis of the parameter covariance of combined WMAP9 and Planck 2013 chains with BOSS DR11 BAO and Type Ia supernova measurements, with each hypercube dimension spanning ±4σ\pm4\sigma along its corresponding eigenvector.

Each simulation contains 204832048^{3} particles with a particle mass of mp≈6.2×108(Ωm/0.3) h−1 M⊙m_{\mathrm{p}} \approx6.2 \times10^{8}(\Omega_{\mathrm{m}}/0.3)\,h^{-1}\,M_{\odot} in a periodic volume of (400 h−1Mpc)3(400\,h^{-1}\mathrm{Mpc})^{3}. The Plummer-equivalent softening length is 3 kpc h−13\,\mathrm{kpc}\,h^{-1}, corresponding to roughly 0.0150.015 times the mean-interparticle spacing. This high resolution allows us to construct subhalo catalogs and merger trees over a broad range of different cosmologies, enabling cosmological analysis with reliable small-scale information. Dark matter halos are identified using the Rockstar halo finder ([5]), and merger trees are constructed using consistent-trees ([6]). We use the z=0.07z = 0.07 redshift snapshots throughout the work. Of the 25 Aemulus-β\beta simulations, currently 22 have usable catalogs and merger trees; the remaining are excluded due to corrupted data.

UniverseMachine & the Mock Galaxy Catalog

We use UniverseMachine ([4]) to construct a physically motivated, observationally grounded mock galaxy catalog against which to test the SHAM and HOD models. UniverseMachine ties star formation directly to halo assembly history and is well-calibrated to observations, making it a plausible realization of the real universe with substantially more complex galaxy–halo connection than the models we test.

Specifically, UniverseMachine assigns star formation rates (SFRs) to individual galaxies using the maximum circular velocities, assembly histories, and redshifts of their halos, and then integrates these SFRs along merger trees to self-consistently predict the galaxy stellar masses. Thereafter, UniverseMachine forward models a comprehensive list of summary statistics, including stellar mass functions, specific SFRs, quenched fractions, UV luminosity functions, and clustering. The predicted statistics are then compared with observations to calibrate the model parameters.

We populate one Aemulus-β\beta simulation with the UniverseMachine Data Release 1 model to construct the target mock galaxy catalog. Building the test and target models on the same simulation suite ensures that they share the mass resolution and halo identification. We choose to generate the target mock at a cosmology close to Planck 2018 ([43]) and near the center of the parameter space spanned by Aemulus-β\beta. The final mock galaxy catalog is a stellar-mass-selected sample with log⁡(M⋆/M⊙)≥10.5\log(M_{\star}/M_{\odot}) \ge10.5, yielding a number density of ng∼2×10−3 Mpc−3n_{\mathrm{g}} \sim2 \times10^{-3}\,\mathrm{Mpc}^{-3}. This number density is comparable to a volume-limited DESI BGS-like sample to z∼0.2z \sim0.2 (estimated by integrating the observed luminosity function). A denser sample would improve access to small-scale information, but we do not adopt this for two reasons: first, in real observations a lower stellar mass threshold restricts the volume over which the sample remains complete, leaving the clustering measurements cosmic-variance limited; and second, we set the threshold conservatively above the mass scale at which the Aemulus-β\beta resolution begins to affect the subhalo population. For validation tests at fixed simulation volume, we would expect cosmological constraints to broaden mildly for lower-density samples.

Galaxy–Halo Connection Models

We populate the simulated halo catalogs with two families of empirical models: subhalo abundance matching (SHAM, §3.1), which uses individual (sub)halo properties, and the halo occupation distribution (HOD, §3.2), which assigns galaxies statistically to host halos.

Subhalo Abundance Matching

The SHAM framework assumes a monotonic relationship between a galaxy proxy and a halo proxy, with some scatter in the galaxy–halo relation. Here we adopt a log-normal scatter ss in stellar mass at fixed halo proxy. Building on the interpolation method introduced in Lehmann et al. [35], we propose an extended two-parameter SHAM model that uses a generalized halo proxy

vα=vvir′(vvirvvir′)α,(1)v_{\alpha}=v_{\mathrm{vir}}^{\prime}\left(\frac{v_{\mathrm{vir}}}{v_{\mathrm{vir}}^{\prime}}\right)^{\alpha}, \tag*{(1)}

where vvirv_{\mathrm{vir}} is the virial velocity evaluated at each halo’s epoch of peak mass aMpeaka_{M_{\mathrm{peak}}}, and vvir′v_{\mathrm{vir}}^{\prime} is equivalent to a unit conversion of the peak halo mass MpeakM_{\mathrm{peak}} using the present-day virial density threshold, as defined in Eq. (15) and Eq. (16). This parametrization combines the mass proxy MpeakM_{\mathrm{peak}} with an adjustable dependence on when each halo reaches its peak mass through the additional parameter α\alpha, as the ratio vvir/vvir′v_{\mathrm{vir}}/v_{\mathrm{vir}}^{\prime} depends only on aMpeaka_{M_{\mathrm{peak}}}. Since subhalos tend to reach MpeakM_{\mathrm{peak}} earlier, this model effectively tunes the satellite fraction of the resulting galaxy sample. When α=0\alpha=0, ranking with vα=0v_{\alpha=0} is equivalent to ranking with MpeakM_{\mathrm{peak}}, one of the most popular halo proxies for SHAM; as α\alpha increases, earlier-peaking objects, typically subhalos, are ranked higher, increasing the satellite fraction and the clustering amplitude. This halo proxy parameterization provides a smooth interpolation between ranking by MpeakM_{\mathrm{peak}} and ranking by vvir(aMpeak)v_{\mathrm{vir}}(a_{M_{\mathrm{peak}}}), spanning a range that encompasses widely adopted halo proxies such as vpeakv_{\mathrm{peak}}. Appendix A discusses this parametrization further and compares it with the original proxy of Lehmann et al. [35].

One complexity regarding SHAM models is the treatment of systems that, due to tidal stripping, fall below the resolution limit or the tracking limit of the halo finder after accretion, commonly referred to as “orphan satellites”. In such systems, an associated satellite galaxy could be present even after the subhalo is lost. The treatment of orphans has been shown to affect predictions from semi-analytic, semi-empirical, and SHAM models, with the effect depending on simulation resolution and halo finder (e.g., Guo & White [19]; Behroozi et al. [4]; et al. 2019; Contreras et al. [10]). A variety of orphan prescriptions exist in the literature, but many have not been carefully tested for convergence or for consistency with the full simulation dynamics. Orphan modeling thus introduces additional modeling uncertainty.

In this work, we adopt the galaxy disruption model of UniverseMachine to incorporate orphans when building the SHAM model described above. UniverseMachine follows subhalos until they lose enough mass to drop out of the halo catalog. At this point, subhalos are replaced by massless orphans which are evolved in idealized approximations of their hosts’ potentials. Along these orbits, orphans experience mass loss according to a population-averaged prescription and evolve vmax⁡v_{\max} according to simulated tidal tracks. Orphans are removed from the simulation once vmax⁡v_{\max} falls below a disruption threshold relative to its infall value. See Appendix B of Behroozi et al. [4] for details. To investigate the impact of orphans, we treat the inclusion of orphans as a modeling choice to be tested rather than fixed: we construct the SHAM model both on the raw Rockstar halo catalog and on the modified halo catalog that incorporates orphans, and fit each to the target clustering statistics. We discuss the performance of both SHAM models in fitting the target clustering and inferring cosmology respectively in Section 5 and Section 6.

Halo Occupation Distribution

The HOD framework is a probabilistic model that describes occupation statistics, namely the mean number of galaxies per halo and its moments, assuming that each halo can host at most one central galaxy and some unbounded number of satellite galaxies. Here we use the “vanilla” HOD model, which depends only on halo mass, following Zheng et al. [66]:

nˉcent(M)=12(1+erf⁡[log⁡10(M/Mcut)2σ]),(2)\bar{n}_{\mathrm{cent}}(M)=\frac{1}{2}\left(1+\operatorname{erf}\left[\frac{\log_{10}(M/M_{\mathrm{cut}})}{\sqrt{2}\sigma}\right]\right), \tag*{(2)}
nˉsat(M)=[M−κMcutM1]αHODnˉcent(M),(3)\bar{n}_{\mathrm{sat}}(M)=\left[\frac{M-\kappa M_{\mathrm{cut}}}{M_{1}}\right]^{\alpha^{\mathrm{HOD}}}\bar{n}_{\mathrm{cent}}(M), \tag*{(3)}

with five free parameters McutM_{\mathrm{cut}}, M1M_{1}, σ\sigma, αHOD\alpha_{\mathrm{HOD}}, and κ\kappa. In this parametrization, the central galaxy occupation follows a Bernoulli distribution and the satellite occupation follows a Poisson distribution.

Galaxy occupation can depend on secondary halo properties at fixed halo mass (assembly bias; Wechsler [58]; Gao et al. [17]; Wechsler et al. [61]; Mao et al. [38]), and this effect can bias inference based on mass-only HOD models (e.g., Zentner et al. [64]). We therefore extend the vanilla HOD to include galaxy assembly bias; and to investigate the impact of different modeling approaches, we implement this extension using two widely adopted formulations: the decorated HOD (dHOD) formalism Hearin et al. [22] and the AbacusHOD formalism, which is developed to meet the mock requirements of DESI Yuan et al. [63].

For each HOD formalism, we test two secondary properties, halo concentration and local environment, and always consider them separately. Halo concentration is a natural choice since it encodes information about halo assembly history and is strongly correlated with halo age Wechsler et al. [59, 61]. The choice of local environment is motivated by recent studies of hydrodynamic simulations and observations, which show that considering the local dark matter overdensity can improve the fits to the measured clustering statistics Hadzhiyska et al. (2020); ); Xu et al. (2021); ); Yuan et al. [63]. Specifically, we use rvir/rsr_{\mathrm{vir}}/r_{\mathrm{s}} as the halo concentration, and define the local environment for each halo as the total mass of all neighboring host halos enclosed within 3 h−1Mpc3\,h^{-1}\mathrm{Mpc}. The choice of 3 h−1Mpc3\,h^{-1}\mathrm{Mpc} is empirical, as it yields the best fit to the target clustering statistics among four tested scales (33, 55, 88, 10 h−1Mpc10\,h^{-1}\mathrm{Mpc}).

The dHOD framework implements assembly bias through the “swapping” technique, in which galaxies are redistributed among halos of the same mass according to the rank of a secondary property, preserving the mean occupation ⟨Ngal∣M⟩\langle N_{\mathrm{gal}}|M\rangle. In practice, halos in the same narrow mass bin are divided into an upper and a lower percentile according to a chosen secondary property, and the average number of galaxies in the upper percentile halos is shifted by some δNgal\delta N_{\mathrm{gal}}, with a compensating shift in the lower percentile. The sign and amplitude of the shift are set by the assembly bias parameters AcenA_{\mathrm{cen}} and AsatA_{\mathrm{sat}}, which act on centrals and satellites independently with Acen,Asat∈[−1,1]A_{\mathrm{cen}}, A_{\mathrm{sat}} \in[-1,1]. Positive values increase the mean occupation of halos with larger values of the secondary property, and negative values decrease it. The default split of the halos occurs at the median of the secondary property at fixed halo mass. However, since assembly bias may affect only a small fraction of halos Mansfield & Kravtsov [37], we instead treat the split for centrals as a free parameter. § 5 directly visualizes the contribution of each secondary property to assembly bias, motivating this additional flexibility. We fix the split at the median for satellites, as satellite assembly bias is expected to be much weaker.

The AbacusHOD framework incorporates assembly bias by shifting the characteristic central and satellite mass scales according to each halo’s rank in the secondary property:

log⁡10Mcutmod=log⁡10Mcut−Bcen(crank−0.5),(4)\log_{10} M_{\mathrm{cut}}^{\mathrm{mod}} = \log_{10} M_{\mathrm{cut}} - \mathcal{B}_{\mathrm{cen}}(c^{\mathrm{rank}} - 0.5), \tag*{(4)}
log⁡10M1mod=log⁡10M1−Bsat(crank−0.5),(5)\log_{10} M_{1}^{\mathrm{mod}} = \log_{10} M_{1} - \mathcal{B}_{\mathrm{sat}}(c^{\mathrm{rank}} - 0.5), \tag*{(5)}

where cc denotes the chosen secondary property. This property is ranked within narrow halo mass bins and normalized to crankc^{\mathrm{rank}}, which ranges from 0 to 1. As a result, assembly bias parameters in the AbacusHOD framework modify the mean galaxy occupation at fixed mass directly. We note that the sign convention for the secondary bias parameters in our implementation is opposite to that adopted in Yuan et al. [63], to keep the physical implications of assembly bias consistent with the dHOD framework: specifically, a positive bias parameter increases the occupation of halos with higher concentration or in denser local environments. We also denote the AbacusHOD bias parameters by B\mathcal{B} to distinguish them from the dHOD parameters A\mathcal{A}, as their numerical values are not directly comparable.

All HOD models in this work are implemented with the Python package HALOTOOLS Hearin et al. [23].

Methodology

Clustering Statistics

We use the projected correlation function as the summary two-point statistic throughout this work. As we only work with periodic boxes, the redshift space correlation function can be estimated analytically as

ξ(rp,π)=DD(rp,π)RR(rp,π)−1=DD(rp,π)N2VboxδV(rp,π)−1,(6)\xi(r_{\mathrm{p}},\pi) = \frac{DD(r_{\mathrm{p}},\pi)}{RR(r_{\mathrm{p}},\pi)} - 1 = \frac{DD(r_{\mathrm{p}},\pi)}{N^2}\frac{V_{\mathrm{box}}}{\delta V(r_{\mathrm{p}},\pi)} - 1, \tag*{(6)}

where rpr_{\mathrm{p}} is the transverse distance and π\pi is the distance along the line of sight (LOS), which we always assume to be the zz-axis. DDDD is the normalized pair counts between galaxies and NN is the total number of galaxies. Specifically, we adopt 15 logarithmic bins along the projected separation rpr_{\mathrm{p}} spanning the range of 0.5 Mpc to 30 Mpc, with a maximum LOS separation of πmax⁡=60\pi_{\max} = 60 Mpc.

The projected correlation function is then obtained by integrating the anisotropic 2D correlation function ξ(rp,π)\xi(r_{\mathrm{p}},\pi) along the LOS,

wp(rp)=∫−πmax⁡πmax⁡ξ(rp,π) dπ.(7)w_{\mathrm{p}}(r_{\mathrm{p}}) = \int_{-\pi_{\max}}^{\pi_{\max}} \xi(r_{\mathrm{p}},\pi)\,\mathrm{d}\pi. \tag*{(7)}

As wpw_{\mathrm{p}} marginalizes over the LOS, it is less informative than the full-shape ξ(rp,π)\xi(r_{\mathrm{p}},\pi), but it is unaffected by redshift-space distortions, which introduce extra model complexity. We choose wpw_{\mathrm{p}} as a minimal data vector that could contain cosmological information.

All correlation calculations in this work are performed with the Python package CORRFUNC Sinha & Garrison [48].

Bayesian Analysis

We fit the target wpw_p from UniverseMachine using importance nested sampling to obtain posteriors for the parameters of the galaxy–halo connection models, as well as the model evidence, which is calculated by integrating the likelihood over the model’s parameter space. To perform the fitting, we assume a multivariate Gaussian likelihood for wpw_p, together with an independent error σn\sigma_n for the galaxy number density:

χ2=(wp,data−wp,model)TΣ−1(wp,data−wp,model)+(ndata−nmodelσn)2(8)\chi^2 = (w_{p,\mathrm{data}} - w_{p,\mathrm{model}})^T \Sigma^{-1}(w_{p,\mathrm{data}} - w_{p,\mathrm{model}}) + \left(\frac{n_{\mathrm{data}} - n_{\mathrm{model}}}{\sigma_n}\right)^2 \tag*{(8)}

where the covariance matrix Σ\Sigma is estimated using the jackknife method, and σn=0.05ndata\sigma_n = 0.05n_{\mathrm{data}}. As the Aemulus-β\beta boxes are relatively small, we estimate the covariance matrix with MultiDark Planck 2 (MDPL2, Klypin et al. [27]), a 1 h−1Gpc1\,h^{-1}\mathrm{Gpc} N-body simulation at Planck cosmology, whose larger volume provides a less noisy and better converged covariance estimate.

We run UniverseMachine on MDPL2 to generate a similar galaxy catalog as our target sample, and then divide this larger-volume catalog into 535^3 jackknife subvolumes to compute the covariance following the standard procedure:

Σij=Njack−1Njack∑k=1Njack(wp,i(k)−w‾p,i)(wp,j(k)−w‾p,j)(9)\Sigma_{ij} = \frac{N_{\mathrm{jack}}-1}{N_{\mathrm{jack}}}\sum_{k=1}^{N_{\mathrm{jack}}}\left(w_{p,i}^{(k)}-\overline{w}_{p,i}\right)\left(w_{p,j}^{(k)}-\overline{w}_{p,j}\right) \tag*{(9)}

where w‾p\overline{w}_p is the average of all projected correlation functions measured from the 535^3 jackknife samples. We adopt the corrections from Mohammad & Percival [39] and find the jackknife covariance to be broadly stable across different numbers of jackknife regions. To account for the smaller volume of the Aemulus-β\beta box used to generate the target clustering data, we scale the covariance by the ratio of the simulation volumes before applying it in the likelihood analysis.

We adopt nested sampling because it estimates the model evidence as a byproduct, which is essential for model comparison and for the cosmological analysis presented later in § 4.3. Specifically, we employ Nautilus (Lange [31]), a deep-learning-based package for Bayesian posterior and evidence estimation, with 3000 live points and a target effective sample size of 20,000. We test several choices of hyperparameters and find the estimates of model evidence to be stable across runs.

Cosmological Evidence Modeling

To obtain cosmological constraints from a relatively small number of simulations, we adopt the method of Cosmological Evidence Modeling (CEM) introduced by Lange et al. [34]. The key idea is that the posterior on cosmological parameters C\mathcal{C}, marginalized over the galaxy–halo connection parameters G\mathcal{G}, is directly proportional to the Bayesian evidence of the galaxy–halo connection model at a given cosmology:

P(C∣D)=∫P(C,G∣D)dG(10)P(\mathcal{C}|D) = \int P(\mathcal{C},\mathcal{G}|D)d\mathcal{G} \tag*{(10)}
∝∫L(D∣C,G)P(C)P(G)dG∝Z(D∣C)(11)\propto\int\mathcal{L}(D|\mathcal{C},\mathcal{G})P(\mathcal{C})P(\mathcal{G})d\mathcal{G} \propto\mathcal{Z}(D|\mathcal{C}) \tag*{(11)}

where DD denotes the observed data vector, P(C∣D)P(\mathcal{C}|D) is the posterior on cosmology, L(D∣C,G)\mathcal{L}(D|\mathcal{C},\mathcal{G}) is the likelihood of the data at fixed cosmology and galaxy–halo connection, and Z(D∣C)\mathcal{Z}(D|\mathcal{C}) is the evidence for the observation DD given a cosmology C\mathcal{C}. The marginalized likelihood or Bayesian evidence Z(D∣C)\mathcal{Z}(D|\mathcal{C}), referred to as the cosmological evidence, can be acquired directly by fitting the target wpw_p using nested sampling. For each galaxy–halo connection model, we repeat the fitting with halo catalogs from all simulations with different cosmologies to obtain a distribution of Z(D∣C)\mathcal{Z}(D|\mathcal{C}), which will be used to model the cosmological posterior. The evidence for the box used to build the target galaxy catalog with UniverseMachine is always discarded.

After obtaining the Bayesian evidence for all cosmological simulations, we model Z(D∣C)\mathcal{Z}(D|\mathcal{C}) as a function of the structure growth parameter S8≡σ8(Ωm/0.3)0.5S_8 \equiv\sigma_8(\Omega_m/0.3)^{0.5} alone. We restrict the dimensionality because 21 independent cosmologies are insufficient to characterize a 7-dimensional parameter space and because the projected correlation function is most sensitive to the amplitude of matter fluctuations and the matter energy density. We then describe the distribution of Z(D∣C)\mathcal{Z}(D|\mathcal{C}) with a simple analytic form, namely a normal distribution,

Z(D∣C)∼Z(D∣S8)∝exp⁡[−(S8−μ)2/2σ2].(12)\mathcal{Z}(D|\mathcal{C}) \sim\mathcal{Z}(D|S_8) \propto\exp\left[-(S_8-\mu)^2/2\sigma^2\right]. \tag*{(12)}

We note that the posterior P(S8∣D)P(S_8|D) must satisfy the normalization condition ∫P(S8∣D)dS8=1\int P(S_8|D)dS_8 = 1, so only the shape of Z(D∣S8)\mathcal{Z}(D|S_8) matters. Following Lange et al. [34], we fit Z(D∣S8)\mathcal{Z}(D|S_8) to obtain a posterior sample of normal distributions that could fit the evidence values, and take the final cosmological posterior to be the average of all normalized Gaussians, in order to account for uncertainties in the fit. While the Bayesian evidence is expected to be primarily a function of S8S_8, we also explore this approach for other cosmological parameters, which will be discussed in § 9.

This methodology reduces the expensive full joint inference over cosmology and the galaxy–halo connection to a set of independent and computationally cheaper fits of the galaxy–halo connection at each cosmology separately, allowing cosmological analysis with a limited number of simulations.

Previous CEM analyses have adopted the HOD framework for modeling the galaxy–halo connection; we verify in Appendix B that CEM is also able to recover cosmological information when assuming the SHAM framework.

Modeling Choices and Fits

Assembly Bias in UniverseMachine

Before constructing the empirical galaxy–halo connection models to fit target clustering, we first directly evaluate the existence and structure of assembly bias in the predictions of UniverseMachine. Specifically, we explore whether the assembly bias signal generated by the full forward model UniverseMachine can be reasonably represented by dependencies on a small set of parameters. We choose halo concentration and local environment as the secondary halo properties following the conventions of formulating assembly bias, and evaluate how the galaxy occupation varies as a function of these secondary properties conditioned on halo mass.

To remove the dependence on mass and isolate the influence of secondary halo properties on the expected number of galaxies, we compute the conditional percentile P(<c∣Mhalo)P(<c|M_{\mathrm{halo}}), the probability of halos of mass MhaloM_{\mathrm{halo}} with a secondary property below cc, for each halo. Figure 2 shows the distribution of galaxy occupation as a function of the conditional percentiles. All halos are ranked by their secondary property cc within narrow mass bins, and a halo of larger P(<c∣Mhalo)P(<c|M_{\mathrm{halo}}) has a higher value of cc compared to other halos of the same mass. Specifically, we divide halos within each mass bin further into ten equally spaced bins of P(<c∣Mhalo)P(<c|M_{\mathrm{halo}}) ranging from 0 to 1, count the number of galaxies residing in halos within each percentile bin separately, and plot the probability density of the fractional contribution from galaxies in each percentile of the secondary property to the total number of galaxies in halos of such mass. A flat distribution at 1 across all percentiles would correspond to no additional dependence on the chosen secondary property, while non-flat distributions indicate that galaxies preferentially inhabit halos with certain secondary characteristics and thus experience assembly bias. Conditional percentiles are calculated using the sliding_conditional_percentile function from Halotools.

Figure 2. Probability density of galaxy occupation as a function of the percentile of a secondary halo property at fixed halo mass, P(<c∣Mhalo)P(<c|M_{\mathrm{halo}}), measured from the target UniverseMachine catalog. The distribution of galaxies within halos of the same mass is characterized by quantiles of secondary halo properties. The top row shows the dependence on halo concentration, and the bottom row shows the dependence on local environment, defined as the total mass of neighboring halos enclosed within 3 h−1Mpc3\,h^{-1}\mathrm{Mpc}. The left panels show how the secondary halo property impacts the occupation of central galaxies in the halo mass regime of Mhalo≲3×1012M⊙M_{\mathrm{halo}}\lesssim3\times10^{12}M_{\odot}, which roughly corresponds to where ⟨Ncen⟩≲1\langle N_{\mathrm{cen}}\rangle\lesssim1; the right panels show how satellite occupation varies with the secondary property in the halo mass regime of Mhalo≳3×1012M⊙M_{\mathrm{halo}}\gtrsim3\times10^{12}M_{\odot}. The dashed gray lines at 1 demonstrate the baseline when galaxy occupation has no additional dependence on the chosen secondary property; note that the yy-axis limits are different between the left and right panels. Central occupation is positively correlated with both halo concentration and environment, while satellite occupation is negatively correlated with halo concentration only.

We evaluate assembly bias for central and satellite galaxies separately. The left column of Figure 2 shows the impact of secondary properties on the occupation of centrals in the halo mass regime where ⟨Ncen⟩≲1\langle N_{\mathrm{cen}}\rangle\lesssim1 in the HOD measured from the UniverseMachine mock. This separation is motivated by the fact that central assembly bias can only affect halos of Mhalo≲3×1012M⊙M_{\mathrm{halo}} \lesssim3 \times10^{12} M_{\odot}, as the mean central occupation saturates at 1 for more massive halos, leaving no room for variation due to assembly bias. Halos more massive than this transition across different secondary property percentiles contribute equally to the total number of centrals, leading to flat fractional contributions. Conversely, the right column of the figure shows the impact of secondary properties on the occupation of satellites in the halo mass regime of Mhalo≳3×1012M⊙M_{\mathrm{halo}} \gtrsim3 \times10^{12} M_{\odot}, as satellite galaxies only become abundant in this regime and thus the assembly bias signal is statistically important.

The top row of Figure 2 demonstrates how galaxy occupation depends on the halo concentration. The left panel shows a positive correlation between central occupation and halo concentration, with a strong preference for central galaxies to reside in the highest-concentration halos, especially at lower halo masses. This is consistent with higher-concentration halos forming earlier and having deeper potential wells at fixed mass, making them more efficient at hosting a central galaxy ([59, 18]). We suspect the strong signal in the highest concentration percentile is potentially due to a population of misclassified subhalos in the simulation, the splashback halos located outside the virial radius ([15]). These halos efficiently form stars at their peak mass but subsequently undergo severe mass loss, leading to low Mvir/MpeakM_{\mathrm{vir}}/M_{\mathrm{peak}} ratios and hence large concentrations; meanwhile, they are known to contribute significantly to halo assembly bias ([37]). This observation also motivates us to include the split as a free parameter for central assembly bias in the dHOD framework as discussed in § 3.2, since the default 50% split clearly does not reflect the true assembly bias structure. We also note that this evaluation of assembly bias will not be accessible in real observations, so we allow the split to vary freely in the fit rather than fixing it to a value inferred from these plots. In general, the observed trend that centrals preferentially occupy higher-concentration halos at fixed mass is consistent with the clustering amplitude of UniverseMachine being higher than the corresponding measurement in a catalog constructed from the direct tabulation of P(N∣Mhalo)P(N|M_{\mathrm{halo}}) in UniverseMachine, since higher-concentration low-mass halos tend to be more strongly clustered.

On the other hand, the right panel shows a negative correlation between satellite counts and halo concentration, indicating that satellites prefer to occupy lower-concentration host halos at fixed halo mass. This makes sense, as higher-concentration halos tend to form earlier, allowing more time for processes such as tidal stripping and dynamical friction to occur, and thus host fewer satellites ([53]). Meanwhile, the magnitude of concentration-based assembly bias for satellites is much smaller than for centrals, which agrees with expectations.

Similarly, the bottom row of Figure 2 shows how galaxy occupation varies with the local environment, defined as the total mass of neighboring halos enclosed within 3 h−1Mpc3\,h^{-1}\mathrm{Mpc}. We observe that central galaxies tend to occupy halos in denser environments, potentially reflecting more efficient mass assembly supported by large scale overdensities beyond single-halo scale. While the centrals preferentially occupy halos in denser environments, the trend is weaker than in the concentration case. In particular, the preference for halos in the uppermost percentiles of environment is significantly weaker, with central occupation increasing more gradually with the large-scale environment. Meanwhile, satellite occupation appears unaffected by the local environment at fixed halo mass, with halos across all environment percentiles hosting approximately equal numbers of satellites, as indicated by the flat distribution in the right panel.

To summarize, we find that the complex galaxy–halo connection physics in UniverseMachine gives rise to a clear dependence of galaxy occupation on halo properties beyond mass. This is reflected in the consistent variations of galaxy counts with concentration and environment (especially for centrals), suggesting both properties as reasonable and meaningful proxies to encode assembly bias in galaxy occupation, though the bias is more strongly correlated with the halo concentration than with the local environment in UniverseMachine. To recover the UniverseMachine galaxy–halo connection and perform robust cosmological analysis with empirical galaxy–halo connection models, we need to capture the assembly bias signal as faithfully as possible to describe the clustering and to avoid biasing the cosmological inference. Different assembly bias implementations allow the extended HOD models to predict different shapes for the occupation distribution of Figure 2. The dHOD formalism produces a step function, with the split parameter setting the location of the transition, whereas the AbacusHOD formalism produces a roughly linear and more gradual variation with the conditional percentile. However, compared to the actual measurements from UniverseMachine, neither form is able to capture the real assembly bias structure. Each may be better suited to a particular case, for example, the step function to the sharper concentration dependence. The range of secondary dependence that is allowed in each model is closely connected to whether they are flexible enough to describe real galaxy populations connected to halos and subhalos, and to the cosmological results we present in § 6.

Model Comparison

We now assess the flexibility of the SHAM and HOD models studied in this work to describe galaxy clustering properties of the UniverseMachine catalog at the true cosmology. We fit each model to the target UniverseMachine wpw_p using the halo catalog of the Aemulus-β\beta box on which UniverseMachine was built to isolate the effects of the galaxy–halo connection. Figure 3 presents the predicted wpw_p from the best-fit SHAM and HOD models compared to the UniverseMachine measurement, with error bars from the diagonal of the covariance; the lower panel shows the residuals of each model. Parameters of each model and the best-fit χ2\chi^2 are summarized in Table 1. The χ2\chi^2 values are consistently small for all models that fit the target clustering well, which we attribute to the models being sufficiently flexible, and to the test models and the target being built on the same simulation box and therefore sharing the same realization of cosmic variance. We thus include these values only to compare models fit to the same target. When built on the modified halo catalog that incorporates orphans using UniverseMachine’s galaxy disruption model, the two-parameter SHAM model that

Figure 3. Fit to the UniverseMachine wpw_p from models introduced in § 3, including two-parameter SHAM models built on halo catalogs with and without orphans, four HOD models with different implementations of assembly bias, and the vanilla HOD model. The lower panel shows the fractional residuals between the models and the data. All models except the vanilla HOD fit the target clustering well and yield comparable χ2\chi^2 values.1

ModelParametersBest-fit χ2\chi^2Inferred S8S_8
SHAM (with orphans)α\alpha, ss4.80.844−0.026+0.0280.844^{+0.028}_{-0.026}
SHAM (no orphans)α\alpha, ss4.50.843−0.015+0.0150.843^{+0.015}_{-0.015}
Vanilla HODMcutM_{\mathrm{cut}}, M1M_1, σ\sigma, αHOD\alpha_{\mathrm{HOD}}, κ\kappa33.30.987−0.059+0.0680.987^{+0.068}_{-0.059}
dHOD concentration+Acen+A_{\mathrm{cen}}, AsatA_{\mathrm{sat}}, psplit,cenp_{\mathrm{split,cen}}4.20.879−0.044+0.0680.879^{+0.068}_{-0.044}
AbacusHOD concentration+Bcen+B_{\mathrm{cen}}, BsatB_{\mathrm{sat}}4.00.903−0.045+0.0700.903^{+0.070}_{-0.045}
dHOD environment+Acen+A_{\mathrm{cen}}, AsatA_{\mathrm{sat}}, psplit,cenp_{\mathrm{split,cen}}3.50.946−0.105+0.1260.946^{+0.126}_{-0.105}
AbacusHOD environment+Bcen+B_{\mathrm{cen}}, BsatB_{\mathrm{sat}}5.10.936−0.084+0.1160.936^{+0.116}_{-0.084}

Table 1. Model parameters (§ 3), best-fit χ2\chi^2 at fixed cosmology (§ 5), and inferred cosmological constraints (§ 6) for each empirical galaxy–halo connection model tested in this work.

effectively adjusts satellite fraction fits the target wpw_p well. Surprisingly, the SHAM model built on the original Rockstar halo catalog, which does not include orphans, is also able to fit the target wpw_p comparably well. This observation suggests that the SHAM model parametrization introduced in § 3.1 is able to emulate the effect of orphan modeling for the specific data vector wpw_p; this may indicate that two-point clustering of the stellar-mass-selected sample with log⁡M∗≥10.5\log M_* \ge10.5 only offers modest constraining power on this low-mass galaxy population with large orphan fraction. In Appendix § 15, we further investigate the impact of orphans for SHAM modeling by comparing the galaxy–halo connection predicted by these two implementations. We proceed to cosmological analysis using both models as they both demonstrate enough flexibility to capture the galaxy–halo connection encoded in wpw_p, raising the question of how well they marginalize over uncertainty in the galaxy–halo connection while preserving cosmological information.

As expected, the vanilla HOD model without assembly bias fails to reproduce the clustering of UniverseMachine, while all four extended HOD models with different assembly bias prescriptions provide reasonable fits to the target wpw_p. There is no significant difference in the quality of the fits between the dHOD and AbacusHOD formalisms. Both SHAM models and the four HOD models that incorporate assembly bias all yield comparable χ2\chi^2 values.1 This result indicates that both empirical methods offer comparable flexibility in reproducing the clustering of galaxies in an observationally realistic mock dataset with assembly bias.

Of course, the purpose of this modeling effort is not merely to reproduce an observed data vector, but to derive unbiased constraints on cosmology and the galaxy–halo connection; we present a comparative analysis of this capability of these models in the next section.

Cosmological Results

In this section, we present the results of cosmological inference using CEM for each empirical model.

SHAM

We build the two-parameter SHAM model introduced in § 3.1 on halo catalogs from each of the 22 Aemulus-β\beta simulations, using two halo catalogs per simulation: the standard Rockstar halo catalog and the modified halo catalog that incorporates orphans following UniverseMachine’s disruption model. This is done to assess the impact of uncertainty in orphan modeling independently. We fit both SHAM models to the target UniverseMachine wpw_p and record the Bayesian evidence Z(D∣C)Z(D|C) for all 22 cosmologies. The lower panel of Figure 4 shows the values of cosmological evidence for the with-orphans model (blue) as a function of S8S_8, where ln⁡Z\ln Z of the correct cosmology has been subtracted as a reference and is marked with a red star at zero. The val-

Figure 4. CEM results for the SHAM model built on halo catalogs with orphans (blue) and without orphans (purple). The lower panel plots the distribution of cosmological evidence from the Aemulus-β\beta cosmologies as a function of S8S_8, subtracting ln⁡Z\ln Z of the correct cosmology as a reference. Δln⁡(Z)\Delta\ln(Z) from the with-orphans model are shown as blue dots, while those from the no-orphans model are plotted as purple squares and shifted down by 10 for visual clarity, and the solid curves demonstrate the best-fit normal distributions to the evidence values for both models. The upper panel shows the modeled posterior constraints on S8S_8, with the red dashed vertical line indicating the truth and the shaded region denoting the 1σ1\sigma posterior interval for the with-orphans model. Both SHAM models yield unbiased constraints on S8S_8, while the constraint from the no-orphans model is tighter.

ues of cosmological evidence for the no-orphans model (purple) are shifted down by 10 for better visualization. For both models, we find that ln⁡Z\ln Z is maximized at the correct cosmology, and the remaining evidence values follow an approximately parabolic shape, suggesting that the choice of S8S_8 as the primary variable is reasonable and that SHAM remains sensitive to the underlying cosmology even after marginalizing over the galaxy–halo connection.

Following the CEM methodology outlined in § 4.3, we derive the posterior of the cosmological parameter S8S_8. For each model, the solid curve in the lower panel of Figure 4 shows the best-fit Gaussian distribution to the evidence values Z(D∣C)\mathcal{Z}(D|C), excluding that of the true cosmology. The resulting CEM posteriors for both models are shown in the upper panel and listed in Table 1. In both cases, we find the CEM posterior of S8S_8 to be unbiased, with the input value lying well within the 1σ1\sigma interval; the shaded region shows this interval for the with-orphans model as an example. We obtain S8=0.844±0.027S_8 = 0.844 \pm0.027 for the with-orphans model and S8=0.843±0.015S_8 = 0.843 \pm0.015 for the no-orphans model, while the input truth is S8=0.842S_8 = 0.842. The no-orphans model yields a tighter constraint while both remain unbiased. In general, we are able to constrain S8S_8 to within ∼1.8−3.2%\sim1.8 - 3.2\% when using SHAM in the CEM analysis.

Interestingly, the two-parameter SHAM model, whether built on the halo catalog with or without orphans, is able to infer unbiased cosmology, which complements the finding in § 5.2 that both models can fit the target UniverseMachine clustering well. This suggests that our SHAM parametrization is able to account for a range of models for subhalo stripping and still remain sensitive to cosmology, offering sufficient flexibility to absorb the complexity in both the halo catalog and the galaxy–halo connection without conflating it with cosmological information. The different widths of the two posteriors are connected to the interpolation ranges allowed by the with-orphans and no-orphans models, which we discuss further in Appendix § 13.

Our results demonstrate that the two-parameter SHAM model serves as a robust framework for marginalizing over galaxy–halo connection uncertainties, providing sufficient parametric flexibility to capture the essential features of the complex galaxy assembly in UniverseMachine while remaining sensitive to cosmology. Importantly, as a modeling framework that naturally leverages more subhalo-level information from the simulations, SHAM is able to produce tight and unbiased cosmological constraints even with a relatively small number of simulations spanning a limited number of cosmologies.

HOD

We repeat the same CEM analyses for all four extended HOD formulations studied in this work: each model incorporates assembly bias through either halo concentration or local environment as the secondary dependency and is implemented using either the dHOD or AbacusHOD framework. The vanilla HOD model is also included in the cosmological analysis as a benchmark for comparison.

For all five models, we fit them to the target UniverseMachine clustering wpw_p and record the Bayesian evidence Z(D∣C)\mathcal{Z}(D|C) for all AbacusSummit-β\beta cosmologies. We find that ln⁡Z\ln\mathcal{Z} is maximized at the true cosmology for all four extended HOD models that incorporate assembly bias, while for the vanilla HOD model, the evidence peaks at a different cosmology from the input truth, indicating that the vanilla HOD model attempts to compensate for the impact of assembly bias through tweaking cosmological parameters. This behavior is linked to the results shown in § 5.2 that all four extended HOD models are able to fit a non-trivial set of mock data from UniverseMachine with more complicated underlying physics, while the vanilla HOD model fails the test. These results indicate the necessity of including assembly bias when modeling galaxy clustering and performing cosmological inference using HODs.

We then fit the evidence values for each model with a normal distribution and calculate the final cosmological posterior of S8S_8. Figure 5 shows the posteriors constructed using different HOD models, along with the SHAM results for comparison, with the true S8S_8 value indicated by the red dashed vertical line. Parameters of each model and the resulting constraints on S8S_8 are summarized in Table 1.

Figure 5. Comparison of cosmological constraints on S8S_8 inferred using HOD models with different assembly bias implementations and SHAM models built on halo catalogs with and without orphans. Extended HOD models incorporating secondary dependence on the same halo property are shown in the same color, and the two linestyles denote the two implementation techniques. The right panel shows the 1σ1\sigma posterior range for each model compared to the truth. All HOD models considered here yield potentially biased and weaker cosmological constraints compared to the SHAM models.

From the comparison, we first note that the vanilla HOD model yields significantly biased cosmological constraints, as indicated by the dashed gray line in Figure 5. As the vanilla HOD model ignores assembly bias completely, the best-fit model predicts a projected correlation function approximately 10% lower than the measurements from UniverseMachine. Consequently, the vanilla model demonstrates a strong preference for larger values of S8S_8 to compensate for the higher magnitude of galaxy clustering due to assembly bias through enhanced clustering of dark matter.

Although all assembly bias implementations considered here reproduce the clustering signal similarly well, their performance diverges when inferring cosmology. The extended HOD models yield potentially biased constraints on S8S_8 toward higher values, although less severely than the vanilla model, with less constraining power than the SHAM models. The bias is at the ∼1σ\sim1\sigma level, with S8S_8 constrained to ∼6%\sim6\% for the concentration-based implementations and ∼11%\sim11\% for the environment-based implementations. The asymmetry in the posteriors is due to the fact that the high-S8S_8 tail is poorly constrained, since HOD models can fit the target clustering more easily at higher S8S_8. HOD models tend to compensate for the enhanced galaxy clustering caused by galaxy formation physics in UniverseMachine through adjustments to cosmological parameters, indicating that the implementations tested here do not fully capture the structure of assembly bias in the target model.

Among implementations of concentration-based biases, the constraint on S8S_8 with the AbacusHOD implementation is biased towards slightly higher values compared to the dHOD case. This difference is likely attributable to the limitations of the AbacusHOD formalism to capture UniverseMachine-like assembly bias as well as the dHOD formalism. Support for this interpretation is reflected in Figure 2, which shows that galaxies in UniverseMachine preferentially occupy halos in the highest concentration percentile, and so the dHOD approach of shuffling galaxies with a split at large percentile is more effective in describing this signal than the AbacusHOD approach of smoothly varying galaxy occupation with halo concentration.

When implementing assembly bias based on the local environment, the dHOD and AbacusHOD formalisms generally agree on the inferred posterior of S8S_8. Posteriors derived with environment-based assembly bias models are wider and possibly more biased than those derived with concentration-based models; this suggests that environment-based models are less effective at capturing the true assembly bias signal in UniverseMachine, and less capable of distinguishing between different underlying cosmologies. These results offer an explicit demonstration that both the choice of secondary halo property, and the implementation technique, are of consequential importance for the capability of an HOD model to derive unbiased constraints on cosmology.

In general, we find that while the inclusion of assembly bias in HOD models appears both necessary and sufficient to reproduce the projected correlation function when assuming the correct cosmology, none of the implementations tested in this work is sufficient to fully capture the actual assembly bias signal in the target model, which results in biased cosmological constraints. At the same time, the extra freedom in the assembly bias formalism and the subhalo population leaves HOD models with weaker constraining power on cosmology than the SHAM models. This reveals a fundamental tension in assembly bias modeling for HODs: the formalism can be simultaneously too flexible and physically unrealistic—flexible enough to reproduce reasonable clustering even at incorrect cosmologies, yet not sufficiently physically grounded to recover the true assembly bias signal and cosmological information. HOD models therefore suffer from both excessive parametric flexibility and model misspecification, which together degrade and bias the cosmological constraints they yield. This emphasizes the need to use models with more physically motivated assembly bias when pursuing cosmological analyses that extend into the one-halo regime.

Higher-Order Statistics

To complement the comparison between the empirical models, we further evaluate their performance in predicting higher-order statistics at the true cosmology as an additional test of their ability to capture the full complexity of the galaxy–halo connection.

Methodology & Results

We choose to focus on two higher-order statistics as examples, counts-in-cylinders (CiC) and multiplet clustering. CiC measures the number of companion galaxies in a cylindrical volume of a given projected radius and line-of-sight length centered on each galaxy (e.g. Wang et al. [56]; Pearl et al. [42]). We quantify this statistic through the cumulative distribution function (CDF) for a galaxy to have fewer than NN such companions, P(≤NCiC)P(\leq N_{\mathrm{CiC}}), measured in 10 linearly spaced bins between 0 and 9 and 10 logarithmically spaced bins between 10 and 100. Since we do not incorporate redshift space distortions in this work, we adopt a projected radius of rCiC=1.5 Mpcr_{\mathrm{CiC}} = 1.5\,\mathrm{Mpc} and a half-length of lCiC=2 Mpcl_{\mathrm{CiC}} = 2\,\mathrm{Mpc} along the LoS.

Multiplet clustering describes how small groups of nearby galaxies, or multiplets, trace dark matter the large-scale structure [55]. We identify multiplets by grouping galaxies within a projected separation of 1.5 Mpc and a LoS separation of 2 Mpc, comparable to the cylinder used to define the CiC statistic, and measure the projected cross-correlation wpw_p between the full galaxy sample and all multiplets containing more than one galaxy. In other words, while CiC is sensitive to the local number density of galaxies, multiplet clustering describes the clustering patterns of these small-scale configurations with respect to the large-scale density field.

For each galaxy–halo connection model, we take the best-fit parameters from the fit to the UniverseMachine wpw_p at the true cosmology (§ 5) and use them to predict the higher-order statistics not used in the fit. The top panels of Figure 6 show the predictions for P(≤NCiC)P(\leq N_{\mathrm{CiC}}) and multiplet clustering from the SHAM and HOD models compared to actual UniverseMachine measurements, while the middle panels present the fractional residuals between the predictions and the truths. To facilitate comparison between model predictions and the data, we include jackknife error bars for both statistics, computed in the same manner as for wpw_p. However, we note that the jackknife covariance for CiC is likely unreliable and underestimated, as we observe strong scale dependence and noisy off-diagonal terms in multiple robustness tests. The errors are also strongly correlated between bins for both statistics. We therefore show the error bars here for visual guidance only and note that robust statistical inference will require more careful covariance estimation in the future.

Figure 6. Higher-order statistics predictions compared to measurements from the UniverseMachine mock (black points), for the CDF of counts-in-cylinders (CiC, left) and multiplet clustering (right). For the left panels, the xx-axis is linearly spaced between 0 and 9 and then logarithmically spaced between 10 and 100. The top panels show predictions from the best-fit galaxy–halo connection models at the true cosmology, while the middle panels show their fractional residuals with respect to UniverseMachine. Jackknife error bars are included for visual guidance only. The bottom panels show the fractional difference between the posterior predictive distributions (PPDs) and the UniverseMachine measurements, for the two SHAM models and the concentration-based dHOD model as an example, using samples randomly drawn from the posterior chains. Shaded regions correspond to the 1σ1\sigma ranges of the PPDs, with solid curves showing their means. Both SHAM models predict reasonable higher-order statistics; all HOD models yield predictions that diverge more from the measurements and have wider PPDs.

In order to further explore the information content of each higher-order statistic for different models, we also compute the posterior predictive distributions (PPDs), the distribution of model predictions obtained by evaluating each statistic using samples drawn from the posterior chains. The bottom panels of Figure 6 show the fractional difference between the PPDs and the UniverseMachine measurements, for the two SHAM models and one example HOD model, which implements concentration-based assembly bias using the dHOD framework. The solid curves show the mean of each PPD, and the shaded regions show the 1σ1\sigma intervals.

For both statistics, we observe that the best-fit predictions from the SHAM models generally agree well with the UniverseMachine measurements, while the predictions from all HOD models deviate more significantly from the measurements. Moreover, the PPD of the HOD model is considerably wider than those of the SHAM models, indicating stronger parameter degeneracy in HODs that wpw_p alone cannot resolve and explicitly demonstrating the greater parametric flexibility.

Implications

The ability of SHAM to reasonably predict both higher-order statistics that the model is not fit to demonstrates its robustness and its capacity to capture the essential features of the complex halo-assembly-based UniverseMachine model, further reinforcing its potential to support unbiased cosmological inference.

While the SHAM predictions are generally consistent with the measurements, we observe hints that SHAM models built on halo catalogs with and without orphans are intrinsically distinct given their different predictions for the higher-order statistics. In particular, the counts and clustering patterns of small galaxy configurations respond differently to orphan modeling. The two orphan prescriptions used here represent two bracketing cases by construction: UniverseMachine’s disruption model likely overpredicts the orphan population, while the raw halo catalog ignores them entirely, so a realistic treatment plausibly lies between them. The higher-order statistics are consistent with this, as the measurements and their uncertainties largely overlap the range spanned by the two predictions. This highlights the need for more careful marginalization over orphan prescriptions or for more predictive subhalo tracking with well-characterized uncertainties, and motivates further investigation of subhalo tracking and higher-order statistics sensitive to it in future small-scale analyses.

The narrow PPDs of the SHAM models suggest that the galaxy–halo connection is more tightly constrained within the SHAM framework, since SHAM leverages additional information from subhalo properties and merger trees from the simulation, supporting more reliable descriptions of the galaxy–halo connection and greater sensitivity to cosmology.

On the other hand, the failure of the best-fit HOD models to reproduce these higher-order statistics again suggests that they are not sufficient to capture the full complexity in the galaxy–halo connection, such as non-Gaussianities revealed by beyond-two-point statistics. The discrepancies between the predicted higher-order statistics from HOD models with different secondary bias implementations indicate that these statistics can potentially discriminate between the models and provide insight on assembly bias. Carefully selecting and combining higher-order statistics that are sensitive to different physical aspects of the galaxy–halo connection will be important for robustly characterizing assembly bias and its impact on small-scale cosmological inference.

The wide PPD also demonstrates that these higher-order statistics can effectively break degeneracies in the HOD parameters and contain new information along directions in the galaxy–halo connection parameter space that wpw_p alone is not sensitive to. These higher-order statistics therefore offer additional constraining power beyond the two-point statistics, motivating their inclusion in future cosmological inference. However, while tuning HOD models to these statistics may further constrain the models and improve agreement for specific observables, comprehensive model calibration might remain challenging given the large variety of statistics required to fully characterize non-Gaussian features in the galaxy distribution.

Discussion

Here we test cosmological inference with projected clustering down to small scales (0.5 Mpc) and evaluate the empirical models against higher-order statistics. In this context, we demonstrate that SHAM captures properties of the galaxy–halo connection effectively and gains sensitivity to cosmology by leveraging predictive information from simulations about substructure and physically motivated assembly bias. In contrast, the excessive

parametric flexibility of existing HOD models likely allows them to fit the clustering signal despite a misspecified assembly bias formalism, limiting their robustness for cosmological inference that extends to small scales.

In particular, many analyses that fit observed clustering statistics with HOD models find that including environment-based secondary bias is important. We caution, however, that the environment-based implementations tested here substantially degrade the constraining power on cosmology — Table 1 shows that they cost nearly a factor of two in the constraints on S8S_8 compared to their concentration-based counterparts. This weaker constraining power indicates a stronger degeneracy between environment-based assembly bias and S8S_8, which is itself a plausible explanation for their improved fits to observations. We note that a change in the choice of secondary property alone can substantially affect cosmological inference, and that a better fit to clustering is not necessarily evidence that the model captures the underlying galaxy–halo connection.

We note that the analyses presented in this work use only one target model, UniverseMachine, which provides a specific choice for galaxy assembly and galaxy–halo connection. UniverseMachine has several advantages as the target model: it is one of the very few approaches that have been directly constrained by clustering data similar to that used here, it can be run quickly on large boxes, and it naturally incorporates a physically motivated assembly bias prescription. Although UniverseMachine is a robustly calibrated empirical model based on a wide range of observables, it represents only one point among plausible representations of the real universe. In addition, UniverseMachine and SHAM share the same subhalo-level approach using dark-matter-only simulations, as the core of UniverseMachine is similar to conditional abundance matching, although with different parameters and significantly greater complexity. As a result, treating UniverseMachine as the ground truth is not a fully independent stress test for cosmological inference with SHAM. Nevertheless, UniverseMachine directly ties star formation to the full assembly histories of halos through specific assumptions and flexible functional forms, and thus the secondary dependencies in the resulting mock are not necessarily reproducible by SHAM models. Given that UniverseMachine is one of the few models directly tuned to a range of observational data, with a physically motivated and substantially more complex treatment of assembly bias than the simple empirical models, we regard it as a reasonable validation target.

More comprehensive stress tests of the empirical models would require a broader set of target models. These could include a broader parameter space within UniverseMachine-type empirical models, semi-analytic galaxy formation models, and hydrodynamical simulations, which use fundamentally different methods to predict assembly bias, quenching, and stellar populations. We plan to validate our method against different suites of hydrodynamical simulations in future work, although we note that current models have difficulty simultaneously modeling sufficient volume and physics with sufficient resolution. In addition, we plan to expand the scope of validation tests using Diffsky ([2, 21, 1]) to mock the physical properties of galaxy populations from various numerically expensive simulations and potentially real observations. Diffsky-based mocks will be able to span a wide range in the galaxy formation parameter space, supporting more comprehensive tests of the generalizability of our method to real observations.

We also note that for all CEM analyses, we model the posterior as a normal distribution and fit Z(D∣C)\mathcal{Z}(D|C) as a function of S8S_8 alone. These choices are largely empirically motivated as they appear sufficient to describe the distribution of evidence values we measure, while they are also limited by the small number of available simulations and by sample variance from relatively small simulation volumes. We test both choices as follows.

We try fitting a skew-normal distribution to the evidence values, and find no preference for nonzero skewness. Although genuine asymmetry may exist for the HOD models (§ 6), we are not sensitive to such an asymmetry with a limited number of cosmologies. We then explore potential dependence on cosmological parameters other than S8S_8, such as ww and H0H_0, by studying Z(D∣C)\mathcal{Z}(D|C) as a function of each individually. We find the dependence on these parameters to be much weaker and that these parameters are unable to capture the full variation in the evidence values. We also explore including an additional parameter beyond S8S_8 in the model, by fitting a 2D Gaussian distribution and marginalizing over the second dimension. The greater parametric flexibility of HOD models leaves their evidence values varying too weakly with cosmology to support the additional dimension; for the more constraining SHAM models, we find the S8S_8 posterior to be generally consistent with the 1D result, though slightly broader. We therefore adopt the 1D Gaussian model as a function of S8S_8 throughout the paper, noting that a more flexible form may broaden the posteriors. We expect that incorporating more informative data vectors would enable investigation of asymmetric posterior models and higher-dimensional cosmological inference, and even non-parametric approaches such as Gaussian process emulation ([31]). We leave a more comprehensive exploration of these modeling choices to future work.

Here we have demonstrated the possibility of performing a robust cosmological analysis with SHAM using a small number of high-resolution simulations. While recent works have carried out cosmological inference with SHAM using the full emulator approach together with scaled simulations ([11, 36]), our CEM approach directly uses high-resolution simulations, avoiding numerical errors from scaling and emulation. With increasing data quality from current and next-generation galaxy surveys, this complementary approach can support high-precision cosmological inference. In future work, we plan to validate the scaling method ([3]) with the high-resolution AEMULUS-β\beta simulations, testing whether the scaling technique can predict reliable small-scale information required by empirical models. We have also shown that extending the CEM framework to subhalo-based models like SHAM can advance cosmological inference beyond previous applications using HODs (e.g., [34, 32, 31]). In particular, our cosmological constraints are comparable to or tighter than CEM results from HODs built on a larger number of simulations and results from emulation-based methods.

Therefore, we emphasize that high-resolution simulations across cosmologies, even a limited number, offer valuable information, and the advantages of CEM combined with subhalo-based models will become increasingly valuable for precise cosmological inference. Given the lower computational cost of CEM, it is also conceivable to adopt more complex forward models like UniverseMachine directly as the galaxy–halo connection framework, enabling more complete marginalization over uncertainties in the galaxy–halo connection against a wider range of observables and supporting more robust cosmological inference.

Going forward, we plan to apply the method to real observations. The DESI BGS sample is an immediate and well-suited target, while the method can be generalized to other samples with more careful modeling of survey-specific selection effects. In particular, with the high-density sample expected from DESI Run 2, the method could potentially extract significantly more constraining power from small scales. Here we have chosen a minimal data vector for the cosmological constraints; redshift-space multipoles would add anisotropic information, constraining the modeling of velocity space. Furthermore, incorporating higher-order statistics directly into the inference pipeline would exploit their additional constraining power on the galaxy–halo connection and cosmology, and allow us to investigate optimal combinations of statistics that capture the most information.

Conclusion

In this work, we test the ability of two widely adopted families of empirical galaxy–halo connection models, subhalo abundance matching (SHAM) and the halo occupation distribution (HOD), to predict galaxy clustering and support robust cosmological inference using small-scale clustering statistics. Using the Aemulus-β\beta simulation suite, we generate a realistic stellar-mass-selected mock galaxy catalog with the more complex semi-empirical model UniverseMachine, which provides a physically motivated and observationally grounded galaxy–halo connection and is a reasonable target for testing the empirical models.

We propose a two-parameter SHAM model that allows flexible dependence on when each halo reaches its peak mass, through a halo proxy parameterization that smoothly interpolates between ranking by MpeakM_{\mathrm{peak}} and ranking by vvir(aMpeak)v_{\mathrm{vir}}(aM_{\mathrm{peak}}), effectively adjusting the satellite fraction. We build this model on halo catalogs both with and without orphans. On the HOD side, we focus on a “vanilla” five-parameter HOD model where occupation depends only on host mass, and four extended HOD models that implement assembly bias using different secondary halo properties in different ways (detailed in § 3). We directly investigate assembly bias in the UniverseMachine mock and test whether our empirical models can reproduce the projected clustering wpw_p of this mock down to small scales (∼0.5Mpc\sim0.5\mathrm{Mpc}) at the true cosmology.

Our primary focus is to determine whether fitting these models to the target clustering statistics over the range of cosmologies sampled by Aemulus-β\beta can lead to unbiased parameter recovery. In particular, using cosmological evidence modeling, we construct constraints on the structure growth parameter S8S_8 using evidence values obtained from sampling at each different cosmology. Finally, we predict two higher-order statistics, counts-in-cylinders and multiplet clustering, using best-fit parameters from the fit to wpw_p for each model, and compare the predictions to actual UniverseMachine measurements.

Our results show that:

  • Assembly bias in UniverseMachine is more strongly correlated with the halo concentration than with the local environment (characterized by the total mass of neighboring halos enclosed within 3h−1Mpc3h^{-1}\mathrm{Mpc}), and appears to be more prominent for central galaxies than for satellites (Figure 2). None of the HOD assembly bias formalisms tested in this work can capture the structure of assembly bias in UniverseMachine.

  • At fixed cosmology, the two-parameter SHAM model is able to reproduce UniverseMachine’s projected two-point correlation function wpw_p both with the raw Rockstar halo catalog and with a modified halo catalog that incorporates orphans (Figure 3), suggesting that our SHAM parameterization can potentially encapsulate the still-uncertain effects of subhalo tidal stripping for the data vector considered here.

  • At fixed cosmology, the four extended HOD models with assembly bias are all flexible enough to reproduce UniverseMachine’s projected two-point correlation function wpw_p, while the vanilla HOD model does not have this capability (Figure 3).

  • The two-parameter SHAM model is able to offer tight, unbiased cosmological constraints on the structure growth parameter S8S_8 even with simulations of relatively small volumes that span a limited number of cosmologies (Figure 4). This result holds for two assumptions about orphan modeling that likely bracket the true subhalo tidal stripping, and suggests that SHAM is a reliable galaxy–halo connection framework to support robust cosmological inference on small scales.

  • While HOD models that incorporate assembly bias are able to reproduce wpw_p at fixed cosmology, they result in potentially biased cosmological constraints that are weaker than those of SHAM models and sensitive to how assembly bias is implemented. HOD models attempt to compensate for complexity in the galaxy–halo connection through adjustments in cosmological parameters; the combination of excessive parametric flexibility and model misspecification degrades and biases cosmological constraints from HODs (Figure 5).

  • The SHAM models again demonstrate robustness by successfully reproducing higher-order statistics (counts-in-cylinders and multiplet clustering) that are not used in the fit, whereas HOD models fail this validation test, revealing limitations in their ability to capture higher-order aspects of the galaxy–halo connection with sufficient accuracy (Figure 6). This also indicates that including beyond-two-point statistics can further support cosmological analyses on small scales by distinguishing between models with different assembly bias treatments.

In summary, projected two-point clustering alone can robustly constrain cosmology when using predictive subhalo-level information, while more parametric models such as HODs provide comparatively weaker and potentially biased constraints, motivating the use of subhalo-based models for cosmological inference on small scales.

Acknowledgments

This work received support from the Kavli Institute for Particle Astrophysics and Cosmology at Stanford and SLAC, and from the U.S. Department of Energy under contract number DE-AC02-76SF00515 to SLAC National Accelerator Laboratory. HW received support from the Stanford Graduate Fellowship. Work done at Argonne was supported under the DOE contract DE-AC02-06CH11357.

We thank Joe DeRose, Yao-Yuan Mao, Shanthala Ashwath, and Shawn McLaughlin for contributions to running the Aemulus-β simulation suite. We thank the KIPAC Galaxy Formation and Cosmology group for useful discussions, Joe DeRose for helpful comments on a draft, Tara Dacunha for helpful advice about the visualizations in Figure 1, and Yao-Yuan Mao for useful discussions about the two-parameter SHAM model.

This research made use of computational resources on the Sherlock cluster at Stanford and the S3DF System at SLAC. S3DF is a shared High-Performance Computing facility that supports the scientific and data-intensive computing needs of all experimental facilities and programs of the SLAC National Accelerator Laboratory. SLAC is operated by Stanford University for the U.S. Department of Energy’s Office of Science. We thank Stanford University and the Stanford Research Computing Center, as well as the SLAC computational team, for computational resources and support that contributed to these results.

Appendix

Choice of Interpolation for SHAM

In this work, we employ a SHAM model that includes one additional parameter beyond the standard SHAM approach to introduce a dependence beyond MpeakM_{\mathrm{peak}}, allowing a smooth interpolation between two halo proxies. This is inspired by the similar approach of Lehmann et al. [35] (hereafter L17), which we have modified slightly to expand the range of clustering amplitudes the model can predict. In this Appendix, we investigate the impact of different choices of halo proxies, particularly those related to the proxies used by L17.

The L17 model proposes a generalized halo proxy

vα=vvir(vmax⁡vvir)α,(13)v_{\alpha}=v_{\mathrm{vir}}\left(\frac{v_{\max}}{v_{\mathrm{vir}}}\right)^{\alpha}, \tag*{(13)}

where vmax⁡v_{\max} is the maximal circular velocity and vvirv_{\mathrm{vir}} is the virial velocity of the halo. The authors note that this allows an adjustable dependence on halo concentration through the additional parameter α\alpha, as the ratio vmax⁡/vvirv_{\max}/v_{\mathrm{vir}} can be understood as a proxy for concentration. They evaluate this proxy at the epoch of MpeakM_{\mathrm{peak}}, aMpeaka_{M_{\mathrm{peak}}}, for each halo, and propose that this parametrization allows the model to be an interpolation between the two most commonly used SHAM proxies: the peak halo mass MpeakM_{\mathrm{peak}} and the peak halo maximal circular velocity vpeakv_{\mathrm{peak}}. Specifically, when α=0\alpha=0, ranking with vα=0v_{\alpha=0} is equivalent to ranking with MpeakM_{\mathrm{peak}}, and when α=1\alpha=1, ranking with vα=1v_{\alpha=1} is equivalent to ranking with vmax⁡v_{\max} at MpeakM_{\mathrm{peak}}, which yields nearly identical galaxy samples as ranking with vpeakv_{\mathrm{peak}}.

A subtlety arises in the evaluation of vvirv_{\mathrm{vir}} at aMpeaka_{M_{\mathrm{peak}}}, as the definition of vvirv_{\mathrm{vir}} depends on the virial overdensity Δvir\Delta_{\mathrm{vir}} and the critical density ρcrit\rho_{\mathrm{crit}} in addition to halo mass,

vvir=(4π3ΔvirρcritG3)1/6Mvir1/3,(14)v_{\mathrm{vir}}=\left(\frac{4\pi}{3}\Delta_{\mathrm{vir}}\rho_{\mathrm{crit}}G^{3}\right)^{1/6}M_{\mathrm{vir}}^{1/3}, \tag*{(14)}

where GG is the gravitational constant, and the density threshold ρvir(a)≡Δvir(a)ρcrit(a)\rho_{\mathrm{vir}}(a)\equiv\Delta_{\mathrm{vir}}(a)\rho_{\mathrm{crit}}(a) evolves strongly with the scale factor: it is larger in the early universe so that an object of fixed mass has higher virial velocity at higher redshift. Thus, vvirv_{\mathrm{vir}} at aMpeaka_{M_{\mathrm{peak}}} could be the actual virial velocity of the halo at the epoch of MpeakM_{\mathrm{peak}} adopting the density threshold at the same epoch,

vvir(aMpeak)=(4π3ρvir(a=aMpeak)G3)1/6Mpeak1/3;(15)v_{\mathrm{vir}}(a_{M_{\mathrm{peak}}})=\left(\frac{4\pi}{3}\rho_{\mathrm{vir}}(a=a_{M_{\mathrm{peak}}})G^{3}\right)^{1/6}M_{\mathrm{peak}}^{1/3}; \tag*{(15)}

or we can remove the redshift evolution of the density threshold by defining

vvir′(aMpeak)=(4π3ρvir(a=1)G3)1/6Mpeak1/3,(16)v'_{\mathrm{vir}}(a_{M_{\mathrm{peak}}})=\left(\frac{4\pi}{3}\rho_{\mathrm{vir}}(a=1)G^{3}\right)^{1/6}M_{\mathrm{peak}}^{1/3}, \tag*{(16)}

so that this quantity serves only as a unit conversion of MpeakM_{\mathrm{peak}}, and thus ranking with vα=0=vvir′(aMpeak)v_{\alpha=0}=v'_{\mathrm{vir}}(a_{M_{\mathrm{peak}}}) is reduced to be exactly equivalent to ranking with MpeakM_{\mathrm{peak}}. This definition vα=0=vvir′(aMpeak)v_{\alpha=0}=v'_{\mathrm{vir}}(a_{M_{\mathrm{peak}}}) is the one intended and used by L17 (private communication). For brevity, we write vvirv_{\mathrm{vir}} and vvir′v'_{\mathrm{vir}} for vvir(aMpeak)v_{\mathrm{vir}}(a_{M_{\mathrm{peak}}}) and vvir′(aMpeak)v'_{\mathrm{vir}}(a_{M_{\mathrm{peak}}}) in what follows.

Because aMpeaka_{M_{\mathrm{peak}}} varies across the halo population, the two definitions of virial velocity can produce different clustering amplitudes when used as the halo proxy. As shown in Figure 7, abundance matching by vvir′v'_{\mathrm{vir}} leads to lower clustering amplitude than abundance matching by vmax⁡(aMpeak)v_{\max}(a_{M_{\mathrm{peak}}}), but abundance matching by vvirv_{\mathrm{vir}} predicts higher clustering amplitudes. Since α\alpha enables a smooth interpolation between two different halo proxies, this leads to different interpolation ranges allowed by the models, shown by the shaded regions in Figure 7. In particular, when using the actual virial velocity at aMpeaka_{M_{\mathrm{peak}}} (Eq. (15)), both the clustering amplitude and the satellite fraction decrease with larger α\alpha, while these quantities increase with α\alpha when adopting the definition vvir′v'_{\mathrm{vir}} (Eq. (16)).

Figure 7. Projected correlation functions predicted by SHAM models using different halo proxies at fixed scatter. Left: the model built on the halo catalog without orphans; right: the same model built on halo catalog including orphans. The target UniverseMachine measurement is shown in black. The blue shaded region marks the interpolation range between MpeakM_{\mathrm{peak}} and vpeakv_{\mathrm{peak}} (reported by L17), and the green shaded region shows the interpolation range between vvir(aMpeak)v_{\mathrm{vir}}(aM_{\mathrm{peak}}) and vpeakv_{\mathrm{peak}}. The SHAM implementation adopted in this work spans the full interpolation range, combining both shaded regions.

These differences are closely connected to the physical interpretation of the models, as the halo proxies differ in how they rank the halos at fixed MpeakM_{\mathrm{peak}}. Matching by vvir′v'_{\mathrm{vir}} is equivalent to matching by MpeakM_{\mathrm{peak}} and gives the lowest clustering. Matching by vvirv_{\mathrm{vir}} assigns satellites higher stellar masses than centrals at the same MpeakM_{\mathrm{peak}}, because satellites tend to reach MpeakM_{\mathrm{peak}} earlier than centrals and the virial density threshold is larger at higher redshifts. This leads to an increase in satellite fraction across the sample, which increases the clustering amplitude. Matching by vmax⁡(aMpeak)v_{\max}(a_{M_{\mathrm{peak}}}) increases the satellite fraction relative to matching by vvir′v'_{\mathrm{vir}}, but less than by vvirv_{\mathrm{vir}}, giving intermediate clustering. The opposite trends of the satellite fraction and the clustering amplitude with α\alpha then follow from which proxy is adopted at α=0\alpha=0. L17 interpret α\alpha as tuning the dependence on halo concentration in their model, with higher-concentration halos ranked higher at larger α\alpha, boosting the satellite fraction and increasing the clustering amplitude. We note, however, that the ratio being tuned, vmax⁡(aMpeak)/vvir′v_{\max}(aM_{\mathrm{peak}})/v'_{\mathrm{vir}}, differs from the conventional concentration proxy, vmax⁡/vvirv_{\max}/v_{\mathrm{vir}} with both evaluated at the same epoch, in that it combines each halo’s concentration at MpeakM_{\mathrm{peak}} and the amount of evolution in the virial density threshold between aMpeakaM_{\mathrm{peak}} and the present day. We find that this ratio correlates much more weakly with the present-day concentration than vmax⁡/vvirv_{\max}/v_{\mathrm{vir}} evaluated at the present day.

In this work, we choose an alternative halo proxy parametrization vα=vvir′(vvir/vvir′)αv_{\alpha}=v'_{\mathrm{vir}}(v_{\mathrm{vir}}/v'_{\mathrm{vir}})^{\alpha} (see § 3.1). This parametrization combines the interpolation ranges allowed by both definitions of vvirv_{\mathrm{vir}}, offering greater flexibility in the SHAM model compared to both L17 and the standard approach using individual halo proxies. Here the ratio adjusted by α\alpha, vvir/vvir′v_{\mathrm{vir}}/v'_{\mathrm{vir}}, directly captures the amount of evolution in the virial density threshold ρvir\rho_{\mathrm{vir}} between each halo’s aMpeakaM_{\mathrm{peak}} and the present day, and depends on aMpeakaM_{\mathrm{peak}} alone monotonically. The model thus introduces a clean adjustable dependence on the epoch of peak mass, which correlates with halo formation time, with larger α\alpha favoring halos that reach MpeakM_{\mathrm{peak}} earlier. Since subhalos tend to reach their peak mass earlier than host halos, increasing α\alpha increases their proxy values relative to centrals at fixed MpeakM_{\mathrm{peak}}. This parameterization therefore allows satellites to follow a different stellar-mass-to-halo-mass relation than centrals: the two follow the same relation for α=0\alpha=0 and satellites receive larger stellar masses as α\alpha increases. As a result, the model can effectively tune the satellite fraction, and hence the clustering amplitude.

Figure 7 compares the interpolation ranges allowed by this SHAM model when built on halo catalogs with and without orphans. When built on the halo catalog without orphans, we observe that the effective tuning of satellite fraction of this SHAM parameterization allows it to emulate the effect of orphan modeling in UniverseMachine. When built on the halo catalog including orphans, the interpolation range is noticeably wider. This makes sense as the orphan population is preferentially older and associates with host halos that are more strongly clustered, allowing extra flexibility in the clustering properties. This wider interpolation range leads to the weaker cosmological constraint from the with-orphans model in § 6.1.

Recovery test for SHAM

Lange et al. [34] demonstrates that the CEM method is able to recover cosmological information when assuming the HOD framework for the galaxy–halo connection. In this appendix, we investigate the ability of the SHAM framework to be applied as the galaxy–halo connection model to infer cosmology using CEM, in order to demonstrate the robustness of our analysis in this paper.

We generate fiducial projected correlation functions with SHAM assuming different values of the two model parameters ss and α\alpha as the target data vectors, and repeat the CEM analysis to test whether the methodology is able to recover unbiased cosmology when the model used in the inference is the same as the one used to construct the target. The left panel of Figure 8 shows the modeled posterior of S8S_8 when the scatter is fixed at s=0.2s=0.2 and α\alpha is varied for different clustering amplitudes, while the right panel shows the modeled posterior when ss and α\alpha are tuned simultaneously to generate similar target clustering. For all target data vectors using different combinations of SHAM parameters, we see that CEM together with the SHAM framework is able to recover the input S8S_8, inferring unbiased cosmology.

Figure 8. Cosmological constraints on S8S_8 from CEM with SHAM, when fiducial wpw_p predicted by the same SHAM model using different parameter values are treated as the target data vector. The left panel demonstrates the recovery when the target SHAM model assumes a fixed scatter of s=0.2s=0.2 and different values for α\alpha; the right panel demonstrates the case when ss and α\alpha are tuned simultaneously for similar target clustering amplitudes. Cosmological constraints on S8S_8 are unbiased for all of the target SHAM-based data vectors investigated here.

We note that the constraint on S8S_8 tightens as α\alpha increases at fixed ss from the left panel of Figure 8. This is because the clustering amplitude increases for larger α\alpha, while higher clustering is intrinsically harder to fit since more physical properties of the galaxy–halo connection, such as assembly bias, become important and present in the statistics, leaving narrower parameter space consistent with the measurement. We further test this reasoning by tuning ss and α\alpha simultaneously to generate similar wpw_p as the target data vectors. As shown in the right panel of Figure 8, the cosmological constraints are approximately equivalent in these cases, indicating that the constraining power of CEM on cosmology is dependent on the complexity of galaxy formation encoded in the data vector.

In summary, for a wide range of combinations of SHAM parameters, the CEM method assuming the SHAM framework is able to recover unbiased cosmological information, suggesting that the method is robust enough to apply to cosmological analysis targeting more complex galaxy samples in this work.

Galaxy–Halo Connection from SHAM with Different Orphan Modeling

As presented in the main text, both SHAM models built on halo catalogs with and without orphans are able to fit the target wpw_p and offer unbiased cosmological constraints on S8S_8. However, they are intrinsically different models, as the underlying subhalo populations are certainly distinct. In this appendix, we demonstrate a difference between these two satellite models that is not reflected in the two-point summary statistic.

Figure 9 shows the stellar-mass to halo-mass relation (SHMR) for centrals and satellites predicted by the two SHAM models, compared to the actual measurements from UniverseMachine. The solid curves show the mean of the posterior predictive distribution for the SHMR from each model, with shaded regions giving the 1σ1\sigma ranges. Error bars for the UniverseMachine measurements are estimated analytically as σ/N\sigma/\sqrt{N}, where σ\sigma is the scatter in stellar mass and NN is the number of objects in each halo mass bin. The two SHAM models predict noticeably different SHMRs, demonstrating that the with-orphans and no-orphans models are distinct even though they offer comparable predictions for two-point clustering. For instance, the no-orphans model in general predicts higher satellite stellar masses than the with-orphans model, which is consistent with previous literature that satellite fractions can be boosted to the required level either by adding orphans or increasing the stellar mass at fixed subhalo mass ([44, 57, 4]). The difference in the predicted SHMRs suggests that wpw_p only weakly distinguishes between the orphan treatments for our galaxy sample, though this conclusion may not extend to lower stellar-mass samples or other simulation setups. Meanwhile, more complex statistics are necessary to break this degeneracy. We attempt to evaluate several higher-order statistics in § 8, but reaching a quantitative conclusion would require more careful calibrations of the covariance matrix.

Figure 9. The stellar-mass to halo-mass relation for centrals (left) and satellites (right) predicted by the two SHAM models built on halo catalogs with and without orphans, compared to the measurements from UniverseMachine. The solid curves show the mean of the posterior predictive distribution for each model, and the shaded regions are the 1σ1\sigma interval; error bars on the UniverseMachine measurements correspond to the standard error of the mean in each halo mass bin. The with-orphans and no-orphans SHAM models predict distinct stellar-mass to halo-mass relations, suggesting the inherent difference between the models.

Even though the summary statistic wpw_p only weakly constrains the orphan population and the best-fit SHAM models do not infer exactly the correct galaxy–halo connection (e.g. as reflected in the predicted SHMRs), SHAM as the modeling framework successfully marginalizes over uncertainties in the galaxy–halo connection and supports unbiased cosmological inference. We leave a comprehensive exploration of more complex summary statistics and validation tests to quantify degeneracies between cosmological parameters and galaxy biases for the future.

References

  1. [1]Alarcon, A., Hearin, A. P., Becker, M. R., et al. 2026, The Open Journal of Astrophysics, 9, 61920, doi: 10.3332/001c.161920
  2. [2]Alarcon, A., Hearin, A. P., Becker, M. R., & Chaves-Montero, J. 2023, MNRAS, 518, 562, doi: 10.1093/mnras/stac3118DOI
  3. [3]Angulo, R. E., & White, S. D. M. 2010, MNRAS, 405, 143, doi: 10.1111/j.1365-2966.2010.16459.xDOI
  4. [4]Behroozi, P., Wechsler, R. H., Hearin, A. P., & Conroy, C. 2019, MNRAS, 488, 3143, doi: 10.1093/mnras/stz1182DOI
  5. [5]Behroozi, P. S., Wechsler, R. H., & Wu, H.-Y. 2013a, ApJ, 762, 109, doi: 10.1088/0004-637X/762/2/109DOI
  6. [6]Behroozi, P. S., Wechsler, R. H., Wu, H.-Y., et al. 2013b, ApJ, 763, 18, doi: 10.1088/0004-637X/763/1/18DOI
  7. [7]Berlind, A. A., & Weinberg, D. H. 2002, ApJ, 575, 587, doi: 10.1086/341469DOI
  8. [8]Bullock, J. S., Wechsler, R. H., & Somerville, R. S. 2002, MNRAS, 329, 246, doi: 10.1046/j.1365-8711.2002.04959.xDOI
  9. [9]Conroy, C., Wechsler, R. H., & Kravtsov, A. V. 2006, ApJ, 647, 201, doi: 10.1086/503602DOI
  10. [10]Contreras, S., Angulo, R. E., & Zennaro, M. 2021, MNRAS, 508, 175, doi: 10.1093/mnras/stab2560DOI
  11. [11]Contreras, S., Angulo, R. E., Springel, V., et al. 2023, MNRAS, 524, 2489, doi: 10.1093/mnras/stac3699DOI
  12. [12]DeRose, J., Wechsler, R. H., Tinker, J. L., et al. 2019, ApJ, 875, 69, doi: 10.3847/1538-4357/ab1085DOI
  13. [13]DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016a, arXiv e-prints, arXiv:1611.00036, doi: 10.48550/arXiv.1611.00036
  14. [15]Diemer, B., & Kravtsov, A. V. 2014, ApJ, 789, 1, doi: 10.1088/0004-637X/789/1/1DOI
  15. [16]Euclid Collaboration, Mellier, Y., Abdourrouf, et al. 2025, A&A, 697, A1, doi: 10.1051/0004-6361/202450810
  16. [17]Gao, L., Springel, V., & White, S. D. M. 2005, MNRAS, 363, L66, doi: 10.1111/j.1745-3933.2005.00084.xDOI
  17. [18]Gao, L., & White, S. D. M. 2007, MNRAS, 377, L5, doi: 10.1111/j.1745-3933.2007.00292.xDOI
  18. [19]Guo, Q., & White, S. 2014, MNRAS, 437, 3228, doi: 10.1093/mnras/stt2116DOI
  19. [21]Hearin, A. P., Chaves-Montero, J., Alarcon, A., Becker, M. R., & Benson, A. 2023, MNRAS, 521, 1741, doi: 10.1093/mnras/stad456DOI
  20. [22]Hearin, A. P., Zentner, A. R., van den Bosch, F. C., Campbell, D., & Tollerud, E. 2016, MNRAS, 460, 2552, doi: 10.1093/mnras/stw840DOI
  21. [23]Hearin, A. P., Campbell, D., Tollerud, E., et al. 2017, AJ, 154, 190, doi: 10.3847/1538-3881/aa859fDOI
  22. [24]Henriques, B. M. B., White, S. D. M., Thomas, P. A., et al. 2015, MNRAS, 451, 2663, doi: 10.1093/mnras/stv705DOI
  23. [25]Henriques, B. M. B., Yates, R. M., Fu, J., et al. 2020, MNRAS, 491, 5795, doi: 10.1093/mnras/stz3233DOI
  24. [26]Ivezić, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042cDOI
  25. [27]Klypin, A., Yepes, G., Gottlöber, S., Prada, F., & Heß, S. 2016, MNRAS, 457, 4340, doi: 10.1093/mnras/stw248DOI
  26. [28]Krause, E., & Eifler, T. 2017, MNRAS, 470, 2100, doi: 10.1093/mnras/stx1261DOI
  27. [29]Kravtsov, A. V., Berlind, A. A., Wechsler, R. H., et al. 2004, ApJ, 609, 35, doi: 10.1086/420959DOI
  28. [30]Lagos, C. d. P., Bravo, M., Tobar, R., et al. 2024, MNRAS, 531, 3551, doi: 10.1093/mnras/stae1024
  29. [31]Lange, J. U. 2023, MNRAS, 525, 3181, doi: 10.1093/mnras/stad2441DOI
  30. [32]Lange, J. U., Hearin, A. P., Leauthaud, A., et al. 2022, MNRAS, 509, 1779, doi: 10.1093/mnras/stab3111DOI
  31. [34]Lange, J. U., van den Bosch, F. C., Zentner, A. R., et al. 2019, MNRAS, 490, 1870, doi: 10.1093/mnras/stz2664DOI
  32. [35]Lehmann, B. V., Mao, Y.-Y., Becker, M. R., Skillman, S. W., & Wechsler, R. H. 2017, ApJ, 834, 37, doi: 10.3847/1538-4357/834/1/37DOI
  33. [36]Mahony, C., Contreras, S., Angulo, R. E., et al. 2026, MNRAS, 545, staf2143, doi: 10.1093/mnras/staf2143DOI
  34. [37]Mansfield, P., & Kravtsov, A. V. 2020, MNRAS, 493, 4763, doi: 10.1093/mnras/staa430DOI
  35. [38]Mao, Y.-Y., Zentner, A. R., & Wechsler, R. H. 2018, MNRAS, 474, 5143, doi: 10.1093/mnras/stx3111DOI
  36. [39]Mohammad, F. G., & Percival, W. J. 2022, MNRAS, 514, 1289, doi: 10.1093/mnras/stac1458DOI
  37. [40]Moster, B. P., Naab, T., & White, S. D. M. 2018, MNRAS, 477, 1822, doi: 10.1093/mnras/sty655DOI
  38. [41]Peacock, J. A., & Smith, R. E. 2000, MNRAS, 318, 1144, doi: 10.1046/j.1365-8711.2000.03779.xDOI
  39. [42]Pearl, A. N., Zentner, A. R., Newman, J. A., et al. 2024, ApJ, 963, 116, doi: 10.3847/1538-4357/ad1ffdDOI
  40. [43]Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020, A&A, 641, A6, doi: 10.1051/0004-6361/201833910
  41. [44]Reddick, R. M., Wechsler, R. H., Tinker, J. L., & Behroozi, P. S. 2013, ApJ, 771, 30, doi: 10.1088/0004-637X/771/1/30DOI
  42. [45]Schaye, J., Crain, R. A., Bower, R. G., et al. 2015, MNRAS, 446, 521, doi: 10.1093/mnras/stu2058DOI
  43. [46]Schaye, J., Kugel, R., Schaller, M., et al. 2023, MNRAS, 526, 4978, doi: 10.1093/mnras/stad2419DOI
  44. [47]Seljak, U. 2000, MNRAS, 318, 203, doi: 10.1046/j.1365-8711.2000.03715.xDOI
  45. [48]Sinha, M., & Garrison, L. H. 2020, MNRAS, 491, 3022, doi: 10.1093/mnras/stz3157DOI
  46. [49]Somerville, R. S., Hopkins, P. F., Cox, T. J., Robertson, B. E., & Hernquist, L. 2008, MNRAS, 391, 481, doi: 10.1111/j.1365-2966.2008.13805.xDOI
  47. [50]Somerville, R. S., Popping, G., & Trager, S. C. 2015, MNRAS, 453, 4337, doi: 10.1093/mnras/stv1877DOI
  48. [51]Spergel, D., Gehrels, N., Baltay, C., et al. 2015, arXiv e-prints, arXiv:1503.03757, doi: 10.48550/arXiv.1503.03757DOI
  49. [52]Vale, A., & Ostriker, J. P. 2006, MNRAS, 371, 1173, doi: 10.1111/j.1365-2966.2006.10605.xDOI
  50. [53]van den Bosch, F. C., Tormen, G., & Giocoli, C. 2005, MNRAS, 359, 1029, doi: 10.1111/j.1365-2966.2005.08964.xDOI
  51. [54]Vogelsberger, M., Genel, S., Springel, V., et al. 2014, MNRAS, 444, 1518, doi: 10.1093/mnras/stu1536DOI
  52. [55]Wang, H., Eisenstein, D. J., Aguilar, J. N., et al. 2026, MNRAS, 545, staf2069, doi: 10.1093/mnras/staf2069DOI
  53. [56]Wang, K., Mao, Y.-Y., Zentner, A. R., et al. 2022, MNRAS, 516, 4003, doi: 10.1093/mnras/stac2465DOI
  54. [57]Watson, D. F., & Conroy, C. 2013, ApJ, 772, 139, doi: 10.1088/0004-637X/772/2/139DOI
  55. [58]Wechsler, R. H. 2001, PhD thesis, University of California, Santa Cruz
  56. [59]Wechsler, R. H., Bullock, J. S., Primack, J. R., Kravtsov, A. V., & Dekel, A. 2002, ApJ, 568, 52, doi: 10.1086/338765DOI
  57. [60]Wechsler, R. H., & Tinker, J. L. 2018, ARA&A, 56, 435, doi: 10.1146/annurev-astro-081817-051756DOI
  58. [61]Wechsler, R. H., Zentner, A. R., Bullock, J. S., Kravtsov, A. V., & Allgood, B. 2006, ApJ, 652, 71, doi: 10.1086/507120DOI
  59. [63]Yuan, S., Garrison, L. H., Hadhazyshka, B., Bose, S., & Eisenstein, D. J. 2022, MNRAS, 510, 3301, doi: 10.1093/mnras/stab3355DOI
  60. [64]Zentner, A. R., Hearin, A. P., & van den Bosch, F. C. 2014, MNRAS, 443, 3044, doi: 10.1093/mnras/stu1383DOI
  61. [65]Zhai, Z., Tinker, J. L., Becker, M. R., et al. 2019, ApJ, 874, 95, doi: 10.3847/1538-4357/ab0d7bDOI
  62. [66]Zheng, Z., Coil, A. L., & Zehavi, I. 2007, ApJ, 667, 760, doi: 10.1086/521074DOI
  63. [67]Zheng, Z., Berlind, A. A., Weinberg, D. H., et al. 2005, ApJ, 633, 791, doi: 10.1086/466510DOI

Paper details

vlnc:W_9j8kmn3g
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents