betaFeedback

WorksPhysics and Astronomy

Simulations agree: the Milky Way's satellites are disrupting

Abstract

Recent analyses of simulated Milky Way-mass galaxies have highlighted the prevalence of satellites undergoing tidal disruption. However, these have focused on single hydrodynamical codes with a number of limitations, making it unclear how generic this result is. We compare satellite disruption in four models: two cosmological hydrodynamical simulations (Auriga and FIRE), cosmological NN-body simulations with stellar particle tagging (Aquarius), and a semi-analytic model with varying host disk properties (SatGen). For all systems, we compute the fraction of stellar mass retained by the progenitor (fboundf_\text{bound}) at the present day. We find a remarkable agreement between FIRE, Aquarius, and SatGen; all predict that ∼\sim20 to 30 per cent of classical-mass satellites have fbound<0.9f_\text{bound} < 0.9 (∼\sim5 to 10 per cent have fbound<0.5f_\text{bound} < 0.5). Satellites in Auriga experience substantially more mass loss than other simulations, which we attribute to their massive central galaxies and large satellite sizes prior to infall. In the ultra-faint regime captured by Aquarius and SatGen, disruption rates are comparable to or exceeding those for classical satellites. The ubiquity of mass loss across models presents both an exciting opportunity (a plethora of stellar streams awaiting discovery) and a nuisance (for theorists using satellite counts to constrain dark matter).

Introduction

Satellite galaxies surrounding Milky Way-mass hosts have been extensively studied in cosmological simulations (e.g., Libeskind et al. 2007; Okamoto et al. 2010; Brooks & Zolotov 2014; Sawala et al. 2016; Wetzel et al. 2016; Applebaum et al. 2021) and semi-analytic models (Kauffmann et al. 1993; Bullock et al. 2000; Benson et al. 2002; Kim et al. 2018, 2026; Nadler et al. 2020; Manwadkar & Kravtsov 2022; Ahvazi et al. 2025) to resolve small-scale challenges to the prevailing Λ\LambdaCDM framework (Sales et al. 2022) or use their observed properties to constrain the nature of dark matter (Nadler et al. 2026). Recent analyses of cosmological hydrodynamical simulations have shown that these satellites are often actively losing stellar mass, forming extended stellar streams (Wang et al. 2017; Panithanpaisal et al. 2021; Riley et al. 2025; Pathak et al. 2025; Meziani et al. 2026). While this may be surprising, given the paucity of tidal tails observed around Milky Way satellites, this debris may be hidden beneath the surface brightness limits of current wide-area photometric surveys (Wang et al. 2017; Shipp et al. 2023).1 It is unclear to what degree these disruption rates are simply reflecting theoretical limitations, either numerical shortcomings of the simulations or improper physical properties of the satellites. Tailored NN-body simulations have highlighted that subhaloes can disrupt artificially due to numerical noise from insufficient number of particles or force resolution (van den Bosch & Ogiya [124]; Errani & Peñarrubia 2020; Errani & Navarro 2021; Chiang et al. 2026; Reddy et al. in prep). Spurious heating due to unequal dark matter and star particle masses can result in galaxies expanding in size (Ludlow et al. 2019, 2023; Liao et al. submitted}), possibly enhancing disruption upon infall to a larger host. Even beyond these numerical effects, simulated systems that do not follow established scaling relations (e.g., stellar mass-halo mass, stellar mass-size) could experience physical disruption that is nonetheless unrealistic (Peñarrubia et al. 2008; {citet}Reddy et al. in prep). On the other hand, many of the careful numerical experiments previously mentioned do not capture physical mechanisms that enhance tidal disruption present in cosmological setups, including preprocessing by massive satellites ({citet}He et al. 2025), the tidal field of a growing host with a central disk ({citet}Garrison-Kimmel et al. 2017; {citet}Kelley et al. 2019; {citet}Grand et al. 2021; {citet}Wang et al. 2025), and satellite-satellite collisions ({citet}Shipp et al. 2025`).

These theoretical uncertainties have led to repeated calls to study satellite disruption across a variety of models (see discussion in Shipp et al. 2023, 2025). One principled approach to this task could be to analyze disrupting satellites in haloes run with the same initial conditions using different numerical prescriptions, in the spirit of the AGORA comparison project (Kim et al. 2014; Jung et al. 2024). However, this may limit the types of models that can be considered, excluding semi-analytic approaches that are not limited by particle resolution. Another approach is to take a variety of available simulations, with their respective strengths and weaknesses, and invest effort to standardize the analysis as much as possible. If a similar result emerged across methods, this would suggest that the high rate of stellar mass loss is not an effect of one theoretical framework or numerical artefact.

In this work we take the latter approach, comparing disrupting satellites in four different publicly available simulation suites: Auriga, FIRE, Aquarius, and SatGen. These models span a variety of galaxy-halo connections for populating satellites, numerical resolution limitations, and models for enhanced tidal disruption from the host galaxy. In § 2 we introduce key aspects of the simulation suites and the corresponding catalogues of accreted satellites. In § 3 we describe how we compute fboundf_{\mathrm{bound}} for the simulations and how we homogenise the merger catalogues. We present our main results in § 4 and discuss their implications in § 5.

Simulations

Auriga

The Auriga simulations ([42, 44]) are cosmological magnetohydrodynamical zoom simulations of isolated Milky Way-mass galaxies. The haloes were selected from the EAGLE dark-matter-only box (Schaye et al. 2015) to have M200c=1−2×1012 M⊙M_{200\mathrm c}=1-2\times10^{12}\,M_\odot and satisfy an isolation criterion. They were resimulated using Arepo (Springel 2010; Pakmor et al. 2016) with a galaxy formation model detailed in Grand et al. [42, 44]. We use the six Auriga haloes that were simulated at ‘level 3’ resolution, with baryonic element (dark matter particle) masses of 6.7(36)×103 M⊙6.7(36)\times10^3\,M_\odot and a minimum softening length of 188 pc. Halo catalogues were constructed using the SUBFIND halo finder (Springel et al. 2001) and connected across time using the LHaloTree merger tree algorithm (Springel et al. 2005).

To identify accreted systems and isolate their constituent star particles, we use the ‘accreted particle lists’ in the Auriga public data release ([44]). To construct these catalogues, star particles are assigned to the (sub)halo they are bound to immediately following their birth. The union of star particles that formed in the main branch of a halo across cosmic time correspond to one system (practically, such particles are assigned the same PeakMassIndex). These are the same catalogues that are used in Riley et al. (2025) and Shipp et al. (2025).

FIRE

We use isolated Milky Way-mass haloes from the Latte suite (Wetzel et al. 2016), run as part of the Feedback in Realistic Environments (FIRE) project. The simulations were run with Gizmo (Hopkins 2015) using the FIRE-2 galaxy formation model (Hopkins et al. 2018). We use the set of ‘m12’ haloes listed in tbl:1 of Wetzel et al. (2023), originally presented across a number of studies (Wetzel et al. 2016; Garrison-Kimmel et al. 2017, 2019; Hopkins et al. 2018; Samuel et al. 2020). We use seven haloes in total, excluding m12z as its merger catalogue is not available. The haloes were simulated with baryonic (dark matter) mass resolution of 7.1(35)×103 M⊙7.1(35)\times10^3\,M_\odot and a softening length of 1 (40) pc. Halo catalogues are produced using the Rockstar algorithm (Behroozi et al. 2013a) and linked into a merger tree using Consistent-Trees (Behroozi et al. 2013b).

To identify stars in each accretion event, we use catalogues assembled in the same manner as Panithanpaisal et al. (2021). They first construct a list of potential stellar stream candidates by identifying satellites between 2.7 and 6.5 Gyr ago whose star particles are within the host’s r200mr_{200\mathrm m} at the present day. To account for satellites that accreted more recently, they augment these catalogues with all present-day satellites (within the host’s r200mr_{200\mathrm m}) that were not accounted for in the previous step. Using the merger trees, they identify the snapshot that each object obtained its maximum stellar mass. They then collect all star particles assigned3 to this object for snapshots within a 200 Myr window of this snapshot.

Aquarius

The Aquarius simulations (Springel et al. 2008) are cosmological zoom simulations of isolated Milky Way-mass haloes. The six haloes were selected from a lower-resolution version of the Millennium-II simulation (Boylan-Kolchin et al. 2009) and resimulated with Gadget-3, an updated version of Gadget-2 (Springel 2005). To produce stellar distributions for these dark-matter-only simulations, Cooper et al. (2010) used the Galform semi-analytic model (Cole et al. 2000; Bower et al. 2006) to model star formation histories for each galaxy and the STINGS particle tagging framework (Cooper et al. 2017, 2025) to assign stars to dark matter particles (for brevity, we refer to these outputs as ‘Aquarius’). The resulting star particle data were released publicly by Pu et al. [95].

Cooper et al. (2010) tagged Aquarius haloes that were simulated at ‘level 2’ resolution, with dark matter particle masses ranging from 6.4−14×103 M⊙6.4-14 \times10^{3}\,\mathrm{M}_{\odot} and a softening length of 65.8 pc. Halo catalogues were constructed using the SUBFIND halo finder (Springel et al. 2001) and connected across time using the D-Trees merger tree algorithm (Jiang et al. 2014). Stars formed in a given halo (according to GALFORM) are assigned to that halo’s most tightly bound particles4 in the following snapshot. By tracking tagged particles to the present day, accreted stellar haloes comprised of satellites, stellar streams, and phase-mixed debris naturally emerge.

To identify accretion events, we use the progenitor catalogues described in Pu et al. [95]. Star particles that become unbound from a given satellite to join the Milky Way-mass host are assigned the same ProgenitorID. Stars that remain bound to the progenitor at the present day are also assigned that corresponding ID.

SatGen

The SatGen code (Jiang et al. 2021; Green et al. 2021) is a semi-analytic framework to generate statistical samples of satellite galaxy populations. Briefly, SatGen constructs merger trees for dark matter host halos in a given cosmology using the extended Press-Schechter (EPS) method of Parkinson et al. (2008). Once the trees are constructed, SatGen assigns stellar masses (see Appendix J of Behroozi et al. 2019) and sizes to progenitor halos at accretion and integrates their orbits accounting for dynamical friction and tidal stripping using prescriptions that are calibrated against a large suite of high-resolution idealized simulations.

Throughout the process, SatGen tracks how much stellar mass each satellite loses using tidal tracks defined in Errani et al. (2018) which assume that the ratios stellar mass (m∗(t)/m∗m_{\ast}(t)/m_{\ast}) and stellar size (reff(t)/reffr_{\mathrm{eff}}(t)/r_{\mathrm{eff}}) can be characterized by

g(x)=(xs+1xs+fb)μfbη.(1)g(x) = \left(\frac{x_s+1}{x_s+f_b}\right)^{\mu} f_b^{\eta}. \tag*{(1)}

Here, g(x)g(x) refers to either m∗(t)/m∗m_{\ast}(t)/m_{\ast} or reff(t)/reffr_{\mathrm{eff}}(t)/r_{\mathrm{eff}}, and fbf_b is the instantaneous bound mass fraction of the corresponding dark matter subhalo. The additional parameters μ\mu, η\eta and xsx_s depend on the ‘cuspiness’ of the subhalo at accretion, parameterized by the inner logarithmic slope of the subhalo’s density profile, and on the ratio reff/rmax⁡r_{\mathrm{eff}}/r_{\max}. Here, rmax⁡r_{\max} is the radius at which the subhalo reaches its maximum circular velocity, vmax⁡v_{\max}, which for an unevolved NFW density profile is ∼2.16 rs\sim2.16\,r_s. SatGen uses a simple interpolation scheme to determine the values of μ\mu, η\eta and xsx_s based on the values inferred by Errani et al. (2018) using a diverse set of simulation results that explored the relevant parameter space.5

We use merger trees processed in the same manner as in Monzon et al. (2024, 2026) that fall in a host mass range Mvir=1011.5−1012.5 M⊙M_{\mathrm{vir}}=10^{11.5}-10^{12.5}\,\mathrm{M}_{\odot}. At every time step we assume that the host’s dark matter potential follows an NFW profile (Navarro et al. 1997). Halo concentrations are computed using the model from Zhao et al. (2009). The galactic disc, when included, is positioned at the centre of the host halo and modeled with the axisymmetric Miyamoto & Nagai (1975) disk that grows with the host (as a constant fraction of the host’s MvirM_{\mathrm{vir}}, see Green et al. 2021). We consider three scenarios: no embedded disc potential, a disc with Md/Mvir=0.05M_d/M_{\mathrm{vir}}=0.05, and a disc with Md/Mvir=0.1M_d/M_{\mathrm{vir}}=0.1. Each sample contains 500 unique accretion histories (specifically, a subset of those used in Dru? et al. 2025) that are identical across the three disc scenarios.

Methods

Homogenising the merger catalogues

While each of our suites produce catalogues that have comparable contents (a list of accreted systems and their associated star particles), they were not constructed in exactly the same way. To facilitate our fboundf_{\mathrm{bound}} comparison, we apply the following selection criteria to all catalogues:

  1. Satellite: has a bound component at the present day

  2. Mass: total stellar mass M⋆>5×105 M⊙M_{\star}>5 \times10^{5}\,\mathrm{M}_{\odot} and bound stellar mass M⋆bound>105 M⊙M_{\star}^{\mathrm{bound}}>10^{5}\,\mathrm{M}_{\odot}

  3. Infall time: fell into the main host within the past 10 Gyr

The first two criteria ensure we are studying surviving classical-mass satellites that are resolved in both Auriga and FIRE. The final one alleviates differences in how far back in time objects need to be tracked. We also consider ultra-faint satellites in Aquarius and SatGen, which capture the tidal evolution of much fainter systems than the hydrodynamical suites. For this, we consider systems with 102 M⊙<M⋆bound<105 M⊙10^{2}\,\mathrm{M}_{\odot}<M_{\star}^{\mathrm{bound}}<10^{5}\,\mathrm{M}_{\odot} instead of the mass criteria listed above. We do not apply a selection in distance from the main host.

Bound fractions

Standard halo finders can miss disrupting subhaloes at the resolution typical of zoom simulations (Han et al. 2018; Diemer et al. 2024; Mansfield et al. 2024). To alleviate these concerns, we apply a novel iterative unbinding approach (implemented in nbody_streams; Arora et al. 2026) to both the Auriga and FIRE catalogues. For each snapshot after falling into the main host, we first determine the center of the satellite using a mass-weighted kernel density estimator (KDE) applied to the star particles. The mean position of particles within the highest 1 per cent of the KDE is taken as the initial center. We then compute the potential of the satellite’s dark matter and star particles using a multipole Poisson solver implemented in AGAMA (Vasiliev 2019). Particles not bound to this potential are temporarily masked and the potential re-computed, iterating until the number of bound particles changes by less than 0.1 per cent. Satellites are tracked forwards in time until fewer than 5 bound particles remain.

This technique results in modestly larger bound fractions at intermediate stages of disruption (0.2<fbound<0.80.2 < f_{\mathrm{bound}} < 0.8) compared to simply relying on the halo finder (see the Appendices of Thoron et al. 2026 for Auriga and Mandacarú Guerra et al. 2026 for FIRE). We note that this method does not compensate for numerical disruption due to low particle count or unequal particle masses (van den Bosch & Ogiya [124]; Ludlow et al. 2019, 2023). If a subhalo disrupts earlier than it should, there is nothing for an improved halo tracker to find.

For the other simulations, fboundf_{\mathrm{bound}} is computed in a more straightforward manner. We assume that the high resolution of Aquarius mitigates concerns of SUBFIND’s ability to track the evolution of subhaloes massive enough to form galaxies.6

SatGen is based on calibrated models (for dark matter) and tidal tracks (for stars), such that identifying the bound remnant is not a concern.

Results

Stellar mass functions

In Figure 1 we show the stellar mass functions for satellite galaxies across all models considered in this work. These include the iterative unbinding corrections to Auriga and FIRE described in § 3.2. We intentionally restrict the results from the hydrodynamical simulations (Auriga and FIRE) to M⋆bound>105 M⊙M_\star^{\mathrm{bound}} > 10^{5}\,\mathrm{M}_\odot, reflecting their resolution limitations. We caution that the number of satellites is sensitive to host halo mass and assembly history (e.g., (Boylan-Kolchin et al. 2010; Joshi et al. 2024)), and it is not our goal to evaluate how well these predictions compare to observed luminosity functions.

Figure 1. The stellar mass functions of satellite galaxies across all simulations considered in this work. Shaded contours correspond to the 1 σ1\,\sigma scatter across haloes, while the solid lines are the medians. For visual clarity, we only show the SatGen scenario for the Md/Mvir=0.1M_{\mathrm d}/M_{\mathrm{vir}}=0.1 model, which is similar across disc scenarios. The models span a variety of galaxy-halo connections and host properties, resulting in different stellar mass functions for satellites.

Overall the predictions span a broad range at a given stellar mass, typically over 0.5 dex. Auriga, FIRE, and Aquarius are in reasonable agreement on the typical number of classical satellites (M⋆bound>105 M⊙M_\star^{\mathrm{bound}} > 10^{5}\,\mathrm{M}_\odot), all falling between 20 to 30. Auriga has the largest number of massive (M⋆bound>107 M⊙M_\star^{\mathrm{bound}} > 10^{7}\,\mathrm{M}_\odot) satellites, with far fewer classical satellites below this mass. This reflects the high stellar mass-halo mass relation in Auriga relative to other simulations, particularly for haloes with Mpeak∼1010M_{\mathrm{peak}} \sim10^{10}–1011 M⊙10^{11}\,\mathrm{M}_\odot, which has been noted in several works (e.g., (Simpson et al. 2018; Sales et al. 2022; Riley et al. 2025)). SatGen predicts fewer classical satellites (∼10\sim10–1515) than the other simulations, with a modest dependence on the adopted central disk mass. We also note that the SatGen results have larger scatter across all masses than any of the other simulation suites. Given that our SatGen results are based on 500 haloes, we expect this is a more accurate representation of the halo-to-halo scatter for a given model than the more limited samples (6 or 7 haloes) from the other suites.

FIND and GALFORM to track the faintest satellites in Aquarius, even at higher resolution than considered here. We remind that satellites with M⋆bound>105 M⊙M_\star^{\mathrm{bound}} > 10^{5}\,\mathrm{M}_\odot are the primary focus of this work.

For ultra-faints (M⋆bound<105 M⊙M_\star^{\mathrm{bound}} < 10^{5}\,\mathrm{M}_\odot) in Aquarius and SatGen, there is interesting agreement in the cumulative number of satellites within the M⋆bound∼103M_\star^{\mathrm{bound}} \sim10^{3}–104 M⊙10^{4}\,\mathrm{M}_\odot range, but the two profiles exhibit different behavior at higher and lower masses. The SatGen results have a constant slope and the largest number of faint satellites, reflecting the adopted stellar mass-halo mass relation with no cutoff for forming galaxies in the lowest-mass haloes. Comparing versions with different central stellar disks, the strongest reduction in the total number of satellites comes from introducing a disk at all, while doubling the mass of the disk does not have as much of an impact.

Meanwhile, Aquarius exhibits a bimodal distribution of faint and bright satellites, with peaks near M⋆bound∼3×103 M⊙M_\star^{\mathrm{bound}} \sim3 \times10^{3}\,\mathrm{M}_\odot, ∼3×106 M⊙\sim3 \times10^{6}\,\mathrm{M}_\odot and a dearth of satellites with M⋆bound∼105 M⊙M_\star^{\mathrm{bound}} \sim10^{5}\,\mathrm{M}_\odot. This is driven by the reionization physics implemented in GALFORM, specifically the parameters that control when reionization takes place and characteristic halo mass scale below which galaxy formation is suppressed (Bose et al. (2018)). The suppression of galaxy formation in low-mass haloes results in a divergence between the Aquarius and SatGen counts at the faintest end. We note that the GALFORM model that underpins the Aquarius predictions is a modified version of that presented in Bower et al. (2006) and newer versions of tagged Aquarius haloes with the latest GALFORM parameters produce similar results (Riley et al. in prep.).

Bound fractions for classicals

In Figure 2 we present the distribution of fboundf_{\mathrm{bound}} for classical-mass satellites across all of the simulation suites. As we show the cumulative distribution, simulations closer to the lower right corner of the figure (fbound∼1f_{\mathrm{bound}} \sim1, P∼0P \sim0) have the least disrupted satellite populations. For visual clarity, we only show the scatter for the SatGen run with Md/Mvir=0.1M_{\mathrm{d}}/M_{\mathrm{vir}} = 0.1. As supplementary material, we provide a figure that has the same information with simulations shown in separate panels, showing individual hosts (for Auriga, FIRE, and Aquarius) and the scatter for all SatGen models.

Figure 2. The cumulative distribution of fboundf_{\mathrm{bound}} for classical-mass satellites across different simulation setups. Simulations with satellite systems that are the least disrupted overall will be towards the lower-right corner. We only include systems that have a total stellar mass M⋆>5×105 M⊙M_\star> 5 \times10^{5}\,\mathrm{M}_\odot and a bound stellar mass of M⋆bound>105 M⊙M_\star^{\mathrm{bound}} > 10^{5}\,\mathrm{M}_\odot. Shaded contours correspond to the 1σ1\sigma scatter across haloes, while the solid lines are the result considering all systems ‘stacked’. All of the models predict substantial stellar mass loss, with three (FIRE, Aquarius, and SatGen) further agreeing on the amount (within host-to-host scatter).

The result is striking – most models essentially agree, within their host-to-host scatter, across nearly the full range of disruption (Auriga is the exception, which we discuss in a moment). Between ∼\sim20 and 30 per cent of satellites have lost at least 10 per cent of their stellar mass (fbound≤0.9f_{\mathrm{bound}} \leq0.9), while ∼\sim10 per cent of satellites have lost half or more of their stellar mass (fbound≤0.5f_{\mathrm{bound}} \leq0.5). This is in spite of FIRE, Aquarius, and SatGen sharing very few modeling choices among all three, beyond grounding their structure formation within a Λ\LambdaCDM cosmology (though with slightly different adopted parameters).

On closer inspection, a number of interesting features emerge. The SatGen models with no stellar disk and Md/Mvir=0.05M_{\mathrm{d}}/M_{\mathrm{vir}} = 0.05 are incredibly close to each other, while the Md/Mvir=0.1M_{\mathrm{d}}/M_{\mathrm{vir}} = 0.1 model yields the expected result that increasing the stellar disk mass results in more stellar mass loss. The Md/Mvir=0.1M_{\mathrm{d}}/M_{\mathrm{vir}} = 0.1 model is also very close to the FIRE result. We note that the present-day M⋆/MvirM_{\star}/M_{\mathrm{vir}} ratio ranges from 0.05 to 0.09 for FIRE and 0.06 to 0.07 for Auriga, with Auriga disks forming earlier in cosmic time ((Belokurov & Kravtsov 2022)). For the Milky Way, this value is likely between 0.02 and 0.05, with uncertainties in both the stellar ((Imig et al. 2025; Lian et al. 2025)) and dark matter (e.g., (Wang et al. 2020)) components. Aquarius acts as an interesting middle ground between these cases, agreeing with FIRE and the massive SatGen disk results near fbound∼0.5f_{\mathrm{bound}} \sim0.5 and with the other SatGen models for fbound≥0.9f_{\mathrm{bound}} \geq0.9.

The exception to the broad agreement across models is Auriga, which has a much more disrupted satellites. For this suite, ∼\sim60 per cent of satellites have fbound≤0.9f_{\mathrm{bound}} \leq0.9, while nearly 40 per cent of satellites have fbound≤0.5f_{\mathrm{bound}} \leq0.5. While previous analysis of Auriga satellites led to the provocative claim that ‘most satellites are streams’ ((Riley et al. 2025; Shipp et al. 2025)), it is clear that this result is not universal among models. However, while Auriga disagrees with other simulations on the amount of disruption that the Milky Way satellites should be experiencing, it agrees that the disruption is higher than conventionally expected or observationally detected (so far). In § 5 we discuss potential reasons for this enhanced disruption.

Bound fractions for ultra-faints

In Figure 3 we show the equivalent fboundf_{\mathrm{bound}} distributions for ultra-faint satellites. We only show results for Aquarius and SatGen, which are the models with sufficient numerical resolution and physical prescriptions to analyze satellites with M⋆bound<105 M⊙M_{\star}^{\mathrm{bound}} < 10^{5}\,M_{\odot}.

Figure 3. The cumulative distribution of fboundf_{\mathrm{bound}} for ultra-faint satellites, shown in a similar manner as Figure 2. We only include systems that have a bound stellar mass of 102 M⊙<M⋆bound<105 M⊙10^{2}\,M_{\odot} < M_{\star}^{\mathrm{bound}} < 10^{5}\,M_{\odot}. The result for FIRE classical-mass satellites is also shown for comparison. Ultra-faint satellites experience stellar mass loss comparable to, or exceeding, those for classicals.

The qualitative trends for ultra-faints are surprisingly similar to those for classicals. The SatGen models with no or lower disk mass are incredibly similar to each other, while the more massive disk model results in a more disrupted satellite population. Aquarius also agrees with the SatGen higher disk mass model for fbound≤0.7f_{\mathrm{bound}} \leq0.7 and transitions to agree with the other models for fbound≥0.9f_{\mathrm{bound}} \geq0.9. The scatter between Aquarius and SatGen are comparable, potentially indicating that improved statistics (many more ultra-faints than classicals) makes the halo-to-halo scatter for fboundf_{\mathrm{bound}} distributions converge with fewer haloes.

Perhaps the most surprising result is that, for a fixed model, the ultra-faints show higher levels of disruption than the classicals. Between ∼\sim30 and 45 per cent of satellites have fbound≤0.9f_{\mathrm{bound}} \leq0.9, while ∼\sim10 to 20 per cent of satellites have fbound≤0.5f_{\mathrm{bound}} \leq0.5. We stress that though

Aquarius and SatGen have different galaxy-halo connections resulting in quite different predictions for the total number of ultra-faints (Figure 1), their fboundf_{\mathrm{bound}} distributions are much more similar.

Discussion and concluding remarks

It has been widely noted that satellites in cosmological (typically hydrodynamical) simulations are experiencing tidal disruption, even invoking this as a mechanism to alleviate small-scale problems of Λ\LambdaCDM ((Brooks & Zolotov 2014; Wetzel et al. 2016; Fattahi et al. 2018)). Several studies using controlled experiments have claimed that such results are subject to numerical effects ([124] (Ludlow et al. 2019; Ludlow et al. 2023; Errani & Navarro 2021; Chiang et al. 2026)) and that the results from hydrodynamical simulations in particular should be treated with caution.

Our baseline result – that several models and methodologies all produce disrupting satellite populations – is a substantial deviation from this conventional wisdom. It is worth noting the numerical strengths and weaknesses of each simulation in turn.7 Auriga and FIRE are likely most impacted by particle resolution and spurious heating, but include ‘messy’ cosmological effects like a turbulent assembly history and satellite-satellite interactions. Aquarius is not subject to spurious heating and is higher resolution than Auriga and FIRE, but does not include additional disruption due to a central stellar disk. SatGen does not suffer from numerical noise and can include the growth of a host disk (in a relatively simplified manner), but does not capture satellite-satellite interactions. Despite this mixture of strengths and weaknesses, three of these models (FIRE, Aquarius, SatGen) not only qualitatively agree that their satellites have experienced disruption, but on the degree of that disruption.7

The remaining model (Auriga) predicts a substantially higher level of disruption than the others. As noted in § 4, the stellar mass-halo mass relation for Auriga satellites is higher than comparable simulations and observational constraints (see Section 6.4 of Riley et al. 2025 for a summary). Low-mass field galaxies in Auriga are also larger than observations at a fixed stellar mass, particularly for M⋆<108 M⊙M_{\star} < 10^{8}\,M_{\odot} (e.g., Figure 1 of Bose et al. 2019). The Auriga central galaxies also mostly fall above the typical M⋆−M200cM_{\star} - M_{200\mathrm{c}} relation (e.g., Figure 10 of Fragkoudi et al. 2025). Taken together, this means that the typical ∼1010 M⊙\sim10^{10}\,M_{\odot} halo is populated with a more massive and more extended galaxy (larger satellites lose stellar mass more readily; Reddy et al. in prep), then falls onto the main Auriga host to evolve in a stronger tidal field. This combination could provide a physical explanation of the enhanced disruption relative to FIRE, even though the two share many of the same potential numerical effects.

It is worth considering that both Aquarius and SatGen find that ultra-faints are more disrupted than classicals. Efforts to model Milky Way satellites typically assign stellar masses (or luminosities) to subhaloes based on their peak halo mass, then assume that this mass does not change unless the satellite is fully disrupted and removed from the sample (e.g., (Nadler et al. 2020; Manwadkar & Kravtsov 2022; Ahvazi et al. 2025)). This decision is typically justified by comparing notes with high-resolution numerical experiments. However, we note that the tidal tracks used in SatGen are calibrated to similar results (Errani et al. 2018), suggesting that substantial stellar mass loss is entirely possible with such frameworks. We encourage such efforts to not only consider increases in galaxy size from tidal heating (which are typically implemented), but also stellar mass loss in their forward modeling.

Stellar mass loss of ultra-faints has been relatively unexplored, since these systems are difficult to model in a Milky Way-like environment. Taking the most conservative of these models (SatGen with no disk) implies that of the 33 spectroscopically-confirmed ultra-faint galaxies discovered to date (Pace 2025), on average ∼10\sim10 have lost 10 per cent of their stellar mass and ∼3\sim3 have lost over half of their stellar mass. Meanwhile, only Tucana III exhibits clearly detected tidal tails (Li et al. [67]), while claims surrounding Boötes III’s tidal tails (Carlin & Sand 2018; Jensen et al. 2026; Li et al. [68]) and nature as a galaxy or star cluster (Zaremba et al. 2025; Cerny et al. 2026) are disputed. If these models are correct, many ultra-faints have tidal tails yet undetected. Recent detections (Jensen et al. 2024) of extended stellar distributions surrounding 7 Milky Way ultra-faint satellites (including Boötes III and Tucana III) may be an encouraging sign. We note that diagnostics of disruption used in the literature (e.g., shape, size, proper motion alignment) may be unreliable indicators of a satellite’s present-day disruption state (Thoron et al. 2026).

We close on the following – a variety of theoretical models predicting tidal disruption does not mean that these predictions reflect reality. It is even possible that the recently unravelled debris could be at such a low surface brightness to never be detectable, even with upcoming facilities including NSF-DOE Rubin, Euclid, Roman, and Arakhis. Detailed mock observations, in both the resolved star (Shipp et al. 2023) and integrated light (Miró-Carretero et al. 2025) regimes, will play a crucial role in determining if satellite disruption rates in the Milky Way and external galaxies are consistent with these models. It is even possible that all of these models are wrong, requiring either new numerical prescriptions or even new physics that shield satellites from forming tidal streams and other debris. However, the alternative is thrilling – an abundance of low surface brightness tidal substructures that will reveal itself within the next decade, if galactic archaeologists search for it.

Acknowledgements

We thank Andrew Pace and Sarah Pearson for enlightening discussions. AHR thanks Taylor Swift’s The Life of a Showgirl and Phoebe Bridgers’s Stranger in the Alps for providing a soundscape conducive to writing this manuscript. This research made extensive use of the Science Explorer, funded by NASA under Cooperative Agreement 80NSSC21M00561, and arXiv.org for bibliographic information.

AHR was supported by a fellowship funded by the Wenner Gren Foundation, a Research Fellowship from the Royal Commission for the Exhibition of 1851, and by STFC through grant ST/T000244/1. AD is supported by the European Union (ERC, BeyondSTREAMS, 101115754) grant. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. OA acknowledges support from the Knut and Alice Wallenberg Foundation, the Swedish Research Council (grant 2025-04892), the Swedish National Space Agency (SNSA grants 2023-00164 and 2025-00405), the LMK foundation, and eSSENCE, a Swedish strategic research programme in e-Science. SR-F acknowledges support from the SNSA (grant 2025-00181). We acknowledge support from the DiRAC Institute in the Department of Astronomy at the University of Washington. The DiRAC Institute is supported through generous gifts from the Charles and Lisa Simonyi Fund for Arts and Sciences, Janet and Lloyd Frink, and the Washington Research Foundation.

This work used the DiRAC@Durham facility managed by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). The equipment was funded by BEIS capital funding via STFC capital grants ST/K00042X/1, ST/P002293/1, ST/R002371/1 and ST/S002502/1, Durham University and STFC operations grant ST/R000832/1. DiRAC is part of the National e-Infrastructure.

For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission.

Data availability

We have used simulations from the Auriga Project public data release (Grand et al. 2024) available on the Globus platform.8 We use the publicly-available FIRE-2 cosmological zoom-in simulations (Wetzel et al. 2023), from the Feedback In Realistic Environments (FIRE) project, generated using the GIZMO code (Hopkins 2015) and the FIRE-2 physics model (Hopkins et al. 2018).9 The Aquarius data is publicly available on Zenodo10 and detailed in Pu et al. (2025). The SatGen code is publicly available.11

High-level catalogues produced in this work (stellar masses, bound fractions) will be made publicly available upon journal acceptance.

Software

This research made use of the Python programming language, along with many community-developed or maintained software packages including: astropy ([5, 7] (Astropy Collaboration et al. 2018)), cmasher ([125, 126]), cython (Behnel et al. 2011), h5py (Collette 2013; Collette et al. 2022), jupyter (Perez & Granger 2007; Kluyver et al. 2016), matplotlib (Hunter 2007), nbody_streams (Arora 2026), numpy (Harris et al. 2020), pandas (McKinney 2010; pandas development team 2022), scikit-learn (Pedregosa et al. 2011; Buitinck et al. 2013; Grisel et al. 2022), and scipy (Virtanen et al. 2020; Gommers et al. 2022). We also thank the maintainers of the arepo-snap-util package. Parts of the results in this work make use of the colormaps in the cmasher package. Software citation information aggregated using The Software Citation Station (Wagg & Broekgaarden 2024; Wagg et al. 2024).

Supplemental figure

In Figure A1 we show the fboundf_{\mathrm{bound}} distributions with simulations separated by panel to aid visual comparison. This contains the same data as presented in Figure 2 and Figure 3.

Figure A1. The cumulative distribution of fboundf_{\mathrm{bound}} for satellites across different simulation setups. The information is the same as is presented in Figure 2 and Figure 3, but each panel corresponds to a different simulation. Simulations that produce particle distributions (Auriga, FIRE, Aquarius) have each individual haloes shown as thin lines. The SatGen realizations are shown as shaded contours, corresponding to the 1σ1\sigma percentiles across all 500 haloes. The medians across all haloes are shown as thick dashed lines, while the stacked distributions (what is shown in the main text) are shown as thick solid lines. Despite substantial differences the galaxy-halo connections and other modeling choices, all of the simulations predict substantial stellar mass loss for satellites of Milky Way-mass haloes.

This paper has been typeset from a \TeX/\LaTeX file prepared by the author.

References

  1. [5]Astropy Collaboration et al., 2013, A&A, 558, A33
  2. [7]Astropy Collaboration et al., 2022, ApJ, 935, 167
  3. [42]Grand R. J. J., et al., 2017, MNRAS, 467, 179
  4. [44]Grand R. J. J., Fragkoudi F., Gómez F. A., Jenkins A., Marinacci F., Pakmor R., Springel V., 2024, MNRAS, 532, 1814
  5. [56]Ji A. P., et al., 2021, ApJ, 921, 32
  6. [67]Li T. S., et al., 2018, ApJ, 866, 22
  7. [68]Li T. S., et al., 2026, arXiv e-prints, p. arXiv:2607.07803arxiv.org/abs/2607.07803
  8. [95]Pu S.-Y., Cooper A. P., Grand R. J. J., Gómez F. A., Monachesi A., 2025, ApJ, 980, 63
  9. [99]Santos-Santos I. M. E., Frenk C. S., Navarro J. F., Cole S., Helly J., 2025, MNRAS, 540, 1107
  10. [100]Santos-Santos I. M. E., Frenk C. S., Navarro J. F., 2026, arXiv e-prints, p. arXiv:2604.09539
  11. [124]van den Bosch F. C., Og iya G., 2018, MNRAS, 475, 4066
  12. [125]van der Velden E., 2020, The Journal of Open Source Software, 5, 2004
  13. [126]van der Velden E., Robert C., Batten A., Clauss C., beskep (Daniel) H., Thyng K., 2024, 1313e/CMasher: v1.8.0, doi:10.5281/zenodo.10677366, https://doi.org/10.5281/zenodo.10677366

Paper details

vlnc:W_yh2wre2u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

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