WorksPhysics and Astronomy
Constraining the Milky Way Mass profile with data driven distribution function using DESI DR2 halo stars
Abstract
We employ a novel data-driven dynamical modeling method, the empirical distribution function (emPDF), to constrain the Milky Way mass profile using a large sample of halo stars from the Dark Energy Spectroscopic Instrument Milky Way Survey (DESI MWS). The method does not assume any functional form of the tracer distribution. Deviations from the steady-state assumption and the effects of observational errors are further tested and calibrated using a number of realistic mock observations (AuriDESI). Applying the calibrated method to two classes of tracers in DESI, we derive a best-constrained virial mass and concentration of and for BHBs. For RR Lyrae stars we derive and . The uncertainties quoted outside and within brackets are statistical and total errors including systematic errors calibrated from AuriDESI, respectively. The two tracer populations provide overall consistent measurements of the MW halo mass profile. Combining the halo star constraints with the independent constraint from satellite dynamics, we derive a robust and precise final measurement of the MW halo parameters of and , with correlation coefficients of .
(DESI Collaboration)
Introduction
Accurate constraints on the total mass and density distribution of the Milky Way (MW) dark matter halo are fundamental to understanding the formation and evolution of our Galaxy under the cosmological framework. For example, the “too big to fail” problem was raised more than ten years ago by M. Boylan-Kolchin et al. [9]. They pointed out that the most massive satellite galaxies in our MW are less massive than the most massive subhalos predicted by CDM simulations of MW-mass systems. This problem, however, can be explained if the virial mass of our MW is smaller than that assumed by M. Boylan-Kolchin et al. [9]. For example, J. Wang et al. (2012) and M. Cautun et al. [13] reported that the MW satellites are consistent with CDM predictions at a confidence level of 10% if the dark halo of the MW has a virial mass smaller than . This example demonstrates the importance of precisely constraining the MW host halo mass.
The measurement of the dark matter halo mass of the MW is usually based on the modeling of various types of dynamical tracers, including halo stars, satellite galaxies, cold gas, and even the companion Andromeda galaxy. There are also various types of dynamical methodologies that can be used to model these tracer objects. The readers can refer to W. Wang et al. (2020) for a review on different methodologies applied to various types of tracer objects, and also to some more recent studies (e.g. H.-F. Wang et al. 2023; Y. Zhou et al.). 2023; Y. Jiao et al. 2023; X. Ou et al. 2024; A. J. Deason et al. 2021; S. A. Bird et al. [8]; J. Shen et al. 2022; R. Ibata et al. 2024; L. Posti & A. Helmi 2019; G. Eadie et al. 2018; J. Wang et al. 2022; T. M. Callingham et al. [12]; M. Cautun et al. [14]; Z.-Z. Li et al. 2020a; A. Kravtsov & S. Winney 2024; C. Roche et al. 2024; Z.-Z. Li & J. Han 2021).
Among the different types of methodologies, a common approach for dynamical modeling is based on the phase-space distribution function (DF), in which the tracers are considered to be in dynamical equilibrium, and the model gravitational potential parameters are determined by fitting the model to the observed data (e.g. Z.-Z. Li et al. 2020a; A. J. Deason et al. 2021; W. Wang et al. 2015; J. Wang et al. 2022; J. Shen et al. 2022). The DF provides a complete description of a dynamical system via its full 6D phase-space representation (J. Binney & S. Tremaine [6]).
Despite numerous efforts over the past decades devoted to inferring the mass of our MW, the current constraints on the profile of the outer halo show considerable discrepancy (W. Wang et al. 2015, 2020). The discrepancy mainly arises from two aspects: 1) limited halo tracer populations with incomplete and inaccurate kinematic data (e.g. proper motions) at large distances; and 2) systematic uncertainties behind the modeling method that remain uncorrected or uncalibrated (J. Han et al. 2016a,b; G. Eadie et al. 2018; A. J. Deason et al. 2021). To address these issues and establish reliable constraints on the MW mass profile, improved kinematic observations of halo tracers and reliable modelling methods are essential.
Compared with other types of dynamical tracers (satellite galaxies, globular clusters and stellar streams), halo stars are more numerous, though halo stars stripped from the same progenitor satellite galaxy may carry less independent information than the total sample size due to phase-space correlations (e.g. W. Wang et al. 2017, 2018; J. Han et al. 2020).
The Milky Way Survey program of the Dark Energy Spectroscopic Instrument (DESI MWS; A. P. Cooper et al. 2023) is observing a substantial sample of stars in the MW outer stellar halo, which can serve as valuable dynamical tracers with line-of-sight velocity measurements. Among these, blue horizontal branch (BHB) stars and RR Lyrae (RRL) variables 4 are particularly useful due to their nature as distance estimators and their brightness, allowing them to be observed at larger distances from the halo center.
The DESI MWS DR2 dataset contains 12,989 BHB and 12,868 RRL stars distributed throughout the halo. This constitutes the largest spectroscopic sample of BHB and RRL tracers assembled to date. The total number of BHB stars observed by DESI DR1 is almost four times that of previous surveys beyond 50 kpc, with a non-negligible fraction detected beyond 100 kpc (A. Byström et al. [11]; G. E. Medina et al. 2025a).
The large sample of DESI halo stars, especially those probing the more distant outer stellar halo, will provide invaluable new insights into the MW outer halo. By combining DESI line-of-sight velocities and spectrophotometric distances with Gaia proper motions, these halo stars provide a valuable resource with six-dimensional phase-space information. They can potentially help to constrain the dark matter distribution in the more distant outer stellar halo to better precision, in addition to the use of satellite galaxies as tracers (e.g. Z.-Z. Li et al. 2019, 2020a). Nevertheless, as we will show in this study, the Gaia DR3 proper motion measurements are subject to large uncertainties for halo stars beyond 50 kpc, and thus additional corrections or calibrations are required in our measurement.
With the sizable sample of halo stars as tracers, there is a demand for reliable methodologies that can fully exploit the tracer information to estimate the mass of the Galaxy, as well as a good understanding and control of the systematic errors involved. While many DF methods rely on assumed functional forms (P. J. McMillan & J. J. Binney 2013; K. Hattori et al. 2021), a fundamental limitation of such assumed analytical DFs is that the results might be biased if the hypothetical form does not accurately represent the tracer population (W. Wang et al. 2015; J. Han et al. 2016b).
In response to this limitation of conventional DF methods, a data-driven and minimal assumption method for reconstructing the DF from the data itself under the steady-state assumption was first developed by J. Han et al. (2016a) named the Orbital Probability Distribution function (oPDF), with validations and tests on numerical simulations to understand its underlying systematics (J. Han et al. 2016b; W. Wang et al. 2017). The original oPDF method makes use of the full phase space distribution of tracers to construct a model DF. However, to work around discreteness effects in the model DF, only the predicted radial distribution is compared against observations to infer the best potential model through a likelihood analysis. In a recent work, Z. Li et al. (2024) further introduced a smoothing technique to overcome the discreteness in the model DF, which enables the full phase space data to be used when constructing the likelihood. This improved empirical Distribution Function (emPDF hereafter) method is notable for its flexibility, efficiency, and adaptability in managing selection effects, and is about 1.5 times more precise than the traditional Jeans Equation method (Z. Li et al. 2024). It represents a data-driven dynamical modeling method that avoids parametric assumptions on the DF. A related example is the minimum-entropy action-space approach of L. Beraldo e Silva et al. [5], which shares similarities with emPDF but stems from a distinct physical motivation. Inheriting from the oPDF method, emPDF does not rely on any pre-assumed functional form of the DF, and can maximize the advantage of using the large sample of DESI halo stars observed so far.
In this work, we aim to take full advantage of the DESI MWS dataset by applying this data-driven DF method, emPDF, to the BHB and RRL stars from the DESI DR2 observations. We do not only constrain the dark matter distribution around the MW with real DESI halo stars, but also validate the method and calibrate results using mock catalogues presented in A. H. Riley et al. (2026). This work is one component of the broader DESI DR2 Milky Way mass Key Project, with complementary mass profile measurements presented in companion papers using other dynamical methods (G. E. Medina et al. 2026; M. Valluri et al. 2026), and an overview and joint interpretation presented in DESI Collaboration et al. (2026).
The organization of this paper is as follows. In § 2, we describe the data we used, including both the observational data and mock data. § 3 explains how emPDF works. § 4 presents the validation and calibration of the method using mock samples, and then apply it to the DESI observations.
Data
DESI and DESI Milky Way Survey
DESI is one of the foremost multi-object spectrographs for wide-field surveys ([3] (DESI Collaboration et al. 2016a; J. H. Silber et al. 2023; E. F. Schlafly et al. 2023; C. Poppett et al. 2024; J. Guy et al. 2023; T. N. Miller et al. 2024)). Millions of Galactic targets have been observed by DESI (M. Levi et al. 2013), including the Bright Galaxy Survey targets (BGS; C. Hahn et al. 2023), luminous red galaxies (LRG; R. Zhou et al. 2023), emission line galaxies (ELG; A. Raichoor et al. 2023), and quasars (E. Chaussidon et al. 2023). These observations have already produced a wide range of important scientific results, such as measurements of BAO signals from galaxies and the Lyman alpha forest (DESI Collaboration et al. 2024a,b), full-shape galaxy clustering (DESI Collaboration et al. 2024c), two-point statistics (DESI Collaboration et al. 2024d), and cosmological constraints (DESI Collaboration et al. 2024e,f, 2025).
The DESI MW Survey (MWS) is mainly performed at bright time (A. P. Cooper et al. 2023), when the sky conditions are bright due to the moon phase, twilight, and worse seeing conditions, making the observation of high redshift galaxies inefficient. The main MW survey targets million stars selected from the DESI Legacy imaging Survey (A. Dey et al. 2019) down to a limiting magnitude of , with some special types of stars observed down to fainter magnitudes (A. P. Cooper et al. 2023). It covers most of the northern Galactic cap region and a significant fraction of the southern cap region. Beyond observations in bright-time, DESI MWS also has some targets observed in dark-time, including some secondary targets to address specific science goals.
Following the Early Data Release DESI Collaboration et al. (EDR; 2024g) of DESI in June 2023, S. E. Koposov et al. (2024) published a DESI MWS stellar value-added catalog for 400,000 stars, including commissioning and science validation data (DESI Collaboration et al. 2024h). The first data release (DR1; DESI Collaboration et al. 2026) is now publicly available, with the corresponding MWS stellar catalogue and value-added products described by S. Koposov et al. (2026).
In this paper, we use the DESI MWS data from its first three years of observation, corresponding to an internal dataset planned for public release as part of DESI DR2. In particular, our halo star samples are primarily composed of the BHB and RRL stars from the DR2 data.
DESI MWS BHB sample
The BHB samples used in this study are constructed following the same procedure as A. Byström et al. [11], but based on the DESI DR2 data. DESI BHB targets are first selected based on photometry and astrometry according to the criteria described in A. P. Cooper et al. (2023), their Section 4.4.4, and A. Byström et al. [11], their Section 2.1. For completeness, the selection criteria are listed below:
Here, is the Gaia -band magnitude, is the parallax and its uncertainty, while and are optical magnitudes from the DESI Legacy Imaging Surveys.
represents a polynomial fit to the BHB sequence on the color–magnitude diagram, and
where W1 is the WISE 3.4 µm flux and is its error. As a class of high priority samples in MWS, BHBs are also observed as secondary targets in DESI, which are observed in the dark-time program with magnitude limit . Although the secondary BHB sample extends to larger Galactocentric distances (up to kpc), we restrict our analysis to the main survey BHBs observed in bright time, which reach kpc. The secondary BHBs with constitute only a small fraction of the total sample (2.8%), but their inclusion would introduce a more complicated selection function. In addition, the currently available mock halo star catalogs (see Section 2.4) do not provide compatible data products for secondary BHBs in DESI.
After stellar spectra are obtained for the BHB targets, additional cuts are made to remove possible contamination in the final clean BHB sample, as described in § 2.4 of A. Byström et al. [11]. Targets spectroscopically classified as quasars are first removed to ensure a purely stellar sample. Then targets with bad quality of radial velocities are also removed according to the quality flags defined by the the DESI MWS radial-velocity spectroscopy (RVS) pipeline (S. E. Koposov 2019; S. E. Koposov et al. 2024). Further, contamination from blue stragglers (BS) is also removed. BSs are removed in the – (Kiel) diagram, and only stars with
are retained. Finally, only stars in the color range are kept for reliable distances.
For any mass measurement method based on equilibrium assumptions, the presence of unrelaxed substructures is an important concern. Member stars of the Sagittarius (Sgr) stream have been identified in A. Byström et al. [11] (see their Section 2.6 for detailed selection). In our analysis, we choose to remove the entire Sgr footprint to avoid additional radial selection on the distance. Explicitly, we define boundaries in R.A. and Dec along the Sgr stream and exclude all stars within these boundaries from our analysis. There are 6,687 main bright BHB samples from the catalog after deduplication. After applying the RVS_WARN = 0 requirement and removing stars within the Sgr footprint, 4,368 stars remain.5 Applying the apparent magnitude limit of , where is the Galactic extinction corrected -band magnitude, together with the distance cut of kpc yields a final sample of BHB stars that extend up to kpc into the stellar halo. The distance and velocity distributions of the final sample are shown in Figure 1.
Figure 1. Distribution of Galactocentric distances (), radial velocities (), and tangential velocities () for the selected DESI MWS halo stars. Light blue and pink points/histograms represent BHBs and RRLs, respectively.
The line-of-sight velocities of BHBs are given by the DESI MWS RVS pipeline. Proper motions measurements are obtained from Gaia DR3. Distances were derived from the distance modulus using the absolute magnitudes calculated via the color-dependent relation from V. Belokurov & S. E. Koposov [4] and A. Byström et al. [11]:
In this work, Heliocentric quantities are converted into Galactocentric ones using the Galactocentric frame in Astropy(Astropy Collaboration et al. [1, 2, 3]).
DESI MWS RR Lyrae sample
The DESI MWS RRL targets are selected from Gaia Data Release 2 (DR2; G. Clementini et al. 2019) and the Pan-STARRS1 (PS-1; B. Sesar et al. 2017) catalogs. Details about the RRL sample can be found in G. E. Medina et al. (2025a,b). The DR2 RRL sample consists of stars observed by DESI as part of the main and the Survey Validation (SV) programs, crossmatched with the Gaia DR3 RRL catalog (G. Clementini et al. 2023), with G-band absolute magnitudes computed from the luminosity-metallicity relation calibrated by A. Garofalo et al. (2022):
The radial pulsating velocities of DESI RRLs are corrected following the method of G. E. Medina et al. (2025a).
Similar to BHB stars, we also remove the Sgr footprint from our sample of RRL, leaving 4,838 stars. Applying the same apparent magnitude limit of in addition to the distance cut of kpc, we obtain a final sample of 2,282 RRL stars with Galactocentric distances extending to about 80 kpc.
AuriDESI mock sample
For our calibration and validation, we use the DR2 AuriDESI BHB and RRL mock catalogues described in A. H. Riley et al. (2026). The mock halo star catalog is constructed from the cosmological Auriga simulations (R. J. J. Grand et al. 2017) at level-3 resolution. It follows the particle sampling approach of B. Lowing et al. (2015) to decompose each star particle in the parent hydrodynamical simulation into individual stars. The stellar parameters of each individual star are obtained by matching the star particle to the closest PARSEC isochrone ([10]) according to its stellar age and metallicity.
The mock DESI halo star catalogs used in this study are based on the AuriDESI suite (N. Kizhuprakkat et al. 2024), a new set of mocks specifically constructed for DESI observations; see A. H. Riley et al. (2026) for a detailed description of the DR2 BHB and RRL mock catalogs. These mocks provide mock realizations of the DESI MWS dataset based on the previous AuriGaia ICC-MOCKS (R. J. J. Grand et al. 2018), which themselves are derived from the Auriga simulation suite.
There are six MW-like systems in total, each of which includes four sets of catalogs with different viewing directions, corresponding to four different positions of the observer, located in the disc plane of the simulation and along four different directions (, , , ), with respect to the bar major axis. Apparent magnitudes and sky coordinates are computed for each observer position, and the DESI survey flux limits and footprint masks are then applied. The AuriDESI catalogs provide flags identifying stars associated with bound satellite galaxies and stellar streams (A. H. Riley et al. 2025). We exclude both components from the mock samples in order to construct tracer catalogs that more closely approximate dynamical equilibrium.
To select BHBs and RRLs from the mock stellar catalog above, the following criteria are applied to first define the horizontal branch (HB), including cuts in Kiel space with and , an absolute magnitude limit of , and an age selection of Age Gyr. BHBs are then defined as stars along the HB with K, while RRLs are those with K. These selections are made based on the actual stellar parameters in AuriDESI, obtained from the PARSEC isochrones.
With the mock BHB and RRL samples constructed above, the survey selection functions are further incorporated through the AuriDESI observational probability. The DESI spectroscopic completeness component follows the selection-function construction described in Section 2.4 of DESI Collaboration et al. (2026). This component describes the completeness fraction of spectroscopically observed DESI BHBs and RRLs relative to the original photometric targets, which is a function of the number of passes/exposures and Galactic coordinates, with weak dependence on apparent magnitudes or stellar color (we will discuss this in Section 3.2 below). The observational probability is then applied to the mock BHB and RRL samples by randomly discarding stars at each sky pointing, such that the mock completeness matches that of the real DESI observations. For consistency with the observational samples, we apply the same magnitude cuts of for BHBs and for RRLs to the mock samples. We also impose an inner Galactocentric radius cut of 20 kpc. To ensure that the final numbers of mock BHB and RRL tracers match those in the real DR2 samples, further downsampling is applied to the selected mock catalogs. After this step, each mock realization contains approximately 2,000 BHB stars and 2,300 RRL stars, comparable to the corresponding observational samples. The resulting distributions of distances and velocities in the mock BHB sample broadly cover similar radial and kinematic ranges, although the AuriDESI tracers are less centrally concentrated than the DESI data, as shown in Figure 2.
Figure 2. Distribution of Galactocentric distance (), radial velocity (), and tangential velocity () for the mock and observed BHBs in DR2. Blue histograms show mock data distribution based on AuriDESI halo H6 with the mock observer at a viewing direction 30° from the galactic bar axis. Only H6 is shown here for clarity, as other halos exhibit similar distributions. The solid line histograms represent the final used sample, while dashed histograms correspond to the sample before selection (see § 2.2).
To account for observational effects, we have two types of mock samples: 1) error-free samples where the coordinates and velocities are the actual values from the original mock; 2) error-added samples, which consist of 20 DESI-like observational realizations for each halo and viewing angle for each tracer. Comparing the results from the error-free and error-added samples allows us to assess the impact of observational errors on the model inference.
For the error-added realizations, the observed distances are perturbed by adding Gaussian noise with a standard deviation equal to 5% of the distance, consistent with typical distance uncertainties for DESI BHBs and RRLs. The line-of-sight velocity uncertainty of BHBs is assigned from the DESI DR2 BHB error model in the plane. For RRLs, the uncertainty on the observed line-of-sight velocity is drawn directly from the empirical uncertainty distribution of the DESI DR2 RRL catalogue. The proper motion errors, which dominate the uncertainty on the 3-D velocity, are based on the Gaia DR3 error model and increase with apparent magnitude. For each halo, we define the reference NFW parameters used throughout the mock tests by fitting a spherical NFW profile to the true spherically averaged total enclosed-mass profile over 20 to 150 kpc.
Method
emPDF
In this work, we adopt the Empirical Distribution Function (emPDF) method developed by Z. Li et al. (2024). This method allows us to infer the underlying potential from the observed distribution of tracers, without any assumption about the functional form of the DF. It is a minimal assumption method, with the only essential assumption being that the system is in a steady state. Here, we also make an assumption of a spherically symmetric underlying potential.
The main concept behind emPDF is that, in a steady state system, the underlying DF of tracers remains time-invariant. Therefore, the degree of similarity between the instantaneous observed kinematics of tracers and the time-averaged statistics empirically reconstructed under a trial potential reflects how well the model fits the data.
emPDF is an inheritance and extension of the orbital Probability Density Function (oPDF) method introduced by J. Han et al. (2016a), which states that the probability of observing a tracer at a given position is proportional to the time it spends at that position in a steady-state system, , where represents the position in the orbit. In a spherical potential, the distribution along a given orbit can be expressed via a radial coordinate as . Here, and denote the specific orbital energy and the angular momentum, respectively, while is the radial velocity. Normalizing over a radial orbital period yields
This conditional distribution function is equivalent to time-averaging the radial distribution of each particle along its orbit.
Assuming spherical symmetry, we can decompose the DF into the distribution of orbits, , and the radial distribution, , along the orbit,
A smooth orbital distribution in emPDF is approximated by kernel density estimation (KDE; B. W. Silverman 1986),
where is the weight of each tracer, and is the Epanechnikov kernel, for and zero otherwise, where denotes the bandwidth-scaled separation in the orbital-parameter space. The bandwidth of the kernel is determined using Scott’s rule (D. W. Scott 1979). Combining the conditional radial distribution along each orbit with the continuous orbital distribution, a continuous, data-driven steady-state DF is constructed as
facilitating direct comparison with observations. Bayesian inference can be applied to derive the correct potential since the likelihood between the empirical DF and the observations is computable.
The probability of observing a tracer with under a trial potential, , specified by parameters is
The posterior of the model potential parameters for this tracer sample is
where is the prior on the potential parameters. By maximizing the posterior, the best-fit potential can be obtained.
In this study, we adopt the assumption that the total mass distribution of the MW follows a spherical NFW profile (J. F. Navarro et al. 1996), characterized by two parameters: the virial mass , and the concentration . is defined as the total mass enclosed within a radius , where the average density is 200 times the critical density of the universe, and is defined as the ratio of the virial radius to the scale radius in the NFW density profile. For simplicity, we will hereafter refer to these parameters as and , respectively. The likelihood is estimated by 2-dimensional grid scanning over this parameter space. In the default setup used for the mock samples and for the DR2 BHB sample, we adopt and . For the DR2 RRL sample, we extend the mass range to to ensure that the posterior is fully enclosed by the scanned range.
Due to the inability of a single NFW profile to adequately describe the complex stellar components in the central region, including the stellar disk and other aspherical structures, our analysis is restricted to kpc. At these radii, the baryonic contribution is reduced, and hydrodynamical simulations suggest that the total mass profiles of MW-sized halos can be reasonably approximated by an NFW profile (M. Schaller et al. 2015; T. Kelley et al. 2019).
Survey selection function correction
In our analysis, we do not explicitly correct for the angular selection function or the dependence of the completeness fraction on sky coordinates, because we assume spherical symmetry in EMPDF. The observational catalogues provide a detection probability, , for each star. To check whether the final tracer samples used in our analysis show a significant magnitude dependence in this observational probability, we calculate the mean in bins of apparent magnitude. As shown in Figure 3, is nearly independent of apparent magnitude over the adopted flux-limited range. For BHBs, mainly reflects the DESI spectroscopic completeness, while for RRLs it also includes the Gaia RRL target catalogue completeness. Therefore, within the adopted flux-limited range of for BHBs and for RRLs, which is consistently applied to both the observational and mock samples, we do not apply any additional treatment of the selection function in the flux-limit range.
Figure 3. Mean detection probability, , as a function of apparent magnitude for the final BHB and RRL samples used in the EMPDF analysis. Blue points with error bars show as a function of Galactic extinction corrected -band apparent magnitude, , for BHBs. Violet points show as a function of Gaia -band apparent magnitude for RRLs. Points represent the mean value in each magnitude bin, and error bars show the standard error of the mean. For BHBs, mainly reflects the DESI spectroscopic completeness, while for RRLs it also includes the Gaia RRL target catalogue completeness.
Although the observational probability does not show a significant dependence on magnitude within the adopted range, the flux limit of the DESI MWS program excludes BHB and RRL stars at certain distances. Because BHB and RRL lie on the horizontal branch in the color-magnitude diagram with nearly constant absolute magnitudes, the distance modulus allows this flux limit to define a distance window, which has to be properly considered in the modeling. Explicitly, this radial selection function is corrected through the following two aspects:
Each tracer is a biased representative of its orbital parameters . To account for the missed tracers, we reweight each observed tracer with these calculated weights when computing in Equation (9),
when constructing the empirical orbital distribution. Here , denotes the selection function, which quantifies the probability that a tracer at phase-space position is included in the observed sample. In many practical cases, acts as a binary window function determined by survey limits. In our case, this is precisely the situation: the flux limit of the star samples defines a sharp observable distance window for each star, leading to if the star’s distance falls within its window and otherwise (see § 2.3 in Z. Li et al. (2024) for details).
An observed tracer sample follows the underlying DF modulated by the selection function. We therefore need to renormalize the selected DF before evaluating the likelihood:
When considering the selection function correction above, as illustrated in Z. Li et al. (2024), in principle, Heliocentric distance instead of the Galactocentric
radius should be used. Yet here we neglect the offset between the Sun and the Galactic center, since this effect is minimal for stars beyond 20 kpc (see § 2).
In our analysis throughout this paper, the same method of correcting for survey selection function is applied to both real observed BHBs/RRLs and to mock BHBs/RRLs.
Estimator bias and systematic uncertainties
Real data is subject to observational errors. Generally, the observational errors will spread and bias the distribution of the true coordinate, , to the observed coordinate, , which may result in a bias in our best constrained potential model or mass profiles. Similar to the selection function, observational errors affect EMPDF in two aspects: the construction of the empirical DF and the evaluation of the likelihood under this DF. The current implementation of EMPDF does not explicitly model observational errors. This is not an issue in Z. Li et al. (2024), where systematic uncertainties from observational errors are negligible when using bright satellite galaxies and globular clusters as tracers, since their phase-space coordinates have significantly smaller uncertainties than those of individual halo stars. However, this assumption does not hold when employing halo stars as dynamical tracers, where the larger observational errors of halo stars in our analysis could substantially affect the results. As we will show in Section 4.1 using mock observations, in the presence of observational error, emPDF tends to overestimate the slope of the mass profile.
In addition to observational error, other deviations from the model assumptions in emPDF could also introduce systematic biases and uncertainties. Due to the minimal assumption nature of emPDF, the remaining deviations from our model assumptions include deviations from steady-state, spherical symmetry, and the adopted NFW form of the density profile. A number of previous works using oPDF have shown that the net effect of these deviations, primarily due to the deviation from steady-state, is to produce a stochastic bias of individual estimates, thereby increasing the overall uncertainty of the inference, while the averaged best-fit over an ensemble of systems remains unbiased (ensemble-unbiased) (J. Han et al. 2016b; W. Wang et al. 2017, 2018; Q. Li et al. 2021). As we will show in § 4.1, this conclusion also holds for emPDF, which is an extension to the oPDF method.
Accounting for biases and systematic uncertainties, the resultant can be written as:
where is the best-fit parameter estimated by our model. The term represents the net systematic bias derived from our mock analysis, as will be shown in § 4.1. The uncertainty terms are assumed to be normally distributed with , , where and represent the systematic and statistical uncertainties, respectively. Throughout this paper, unless otherwise specified, we use the term “statistical uncertainty” to refer to the formal statistical uncertainty inferred from the emPDF posterior. The underestimation of this statistical uncertainty, which mainly includes the effects of observational error, together with all other systematic uncertainties due to deviations from model assumptions, is incorporated into the term.
In our analysis, the systematic uncertainty is specified with a two-dimensional covariance matrix for and , and the total uncertainty is derived by convolving the posterior distribution of emPDF with the systematic covariance. This convolution is not easy to implement if we only provide our results in the form of Equation (15). To avoid this complication, in our final results we will directly quote the total uncertainty instead of the systematic one. Our final reported parameters will be in the form of
where is the bias-corrected best-fit value (with applied, as calibrated in § 4.1).
Results
Tests with mock samples
In this subsection, we present our best constrained mass profiles and halo parameters based on the mock sample, and compare with the truth. We mainly present results based on the BHB mock, as the RRL mock follows the same qualitative trends and leads to similar conclusions.
Calibration of biases
Figure 4 compares the best-fit mass profiles to the true profiles, using mock BHB samples with (blue curves) and without (green curves) observational errors. The top six panels refer to different halos (see the legend) viewed from the same direction. The bottom panels show the results for one of the halos (H27) viewed from three other directions, and the results for other halos are qualitatively similar. This figure allows us to evaluate and compare the results both across different halos observed from the same viewing direction and for the same halo viewed from different directions. From these plots, we find that without observational errors, the best constraints (green curves) show relatively smaller deviations from the truth (black curves) in each panel. The green curves are about equally likely to lie above or below the true profiles.
Figure 4. Comparison of the enclosed mass profiles derived from error-free and error-added mock BHB samples. In each panel, the black curve shows the true mass profile of the halo, and the vertical dotted line represents the median radius of tracers. The green curve shows the best-fit mass profile from error-free mock sample and associated shaded band is the confidence region. The blue curve and hatched band correspond to the mean and scatter of the best-fits of 20 error-added realizations for each halo and viewing angle. The top six subpanels show results for different halos from the same observing direction (). The three bottom panels show results for the same halo (H27) from the remaining directions.
In the left panel of Figure 5 we show the distribution of best-fit parameters for all the error-free mock BHB samples. The mean of the distribution as marked by the black cross appears slightly biased towards a lower mass and a higher concentration. Possible contributors may include the non-uniform survey footprint and halo sub-structure, and the nature of the selected old, metal-poor tracers. However, given the small number of systems used in this analysis and the correlations among the data points from the same halo, the uncertainty on the mean is likely underestimated, and we avoid over-interpreting this small bias. Previous works using a much larger number of independent halos have clearly shown that the result from oPDF, the precursor of emPDF, is ensemble unbiased (J. Han et al. 2016b; W. Wang et al. 2017; Q. Li et al. 2021). So we conclude that the results here are also consistent with being ensemble unbiased. Moreover, any residual bias due to the emPDF estimator itself will also be accounted for when we calibrate the result using the more realistic error-added mocks.
Figure 5. The best-fit halo mass (, x-axis) and concentration (, y-axis) versus the true parameters for all halos and all viewing directions. Different colors denote different systems, and different symbols mark the four viewing directions (see legend). Left: Application of emPDF to error-free mock data. The black cross and green ellipse represent the mean and uncertainty of all symbols. The best constraints are approximately unbiased after ensemble average. Right: Similar to the left plot, but showing the results of applying emPDF to the mock data with realistic observational errors incorporated. For each halo and viewing direction, there are 20 realizations. The green ellipse represents the average size of formal error contour from the emPDF output of each realization. The pink ellipse represents the average size of statistical error contour of each 20 realizations. In both plots, the gray ellipses mark the total scatter of different systems (all halos, directions, and realizations). The root-mean-squared error relative to true values are illustrated in the bracket on the top left corner.
In contrast to the results from error-free mocks, the inclusion of observational error leads to a significant overall bias, represented as a steeper slope and higher mass at larger radii of the mass profile (blue curves with hatched bands in Figure 4) than the truth (black). The bias mainly arises from the fact that observational errors increase rapidly with distance. In particular, the proper motion errors contribute a significant amount of uncertainty as the distance increases, as shown in Figure 2 in S. A. Bird et al. [7]. The increasing observational errors with distance inflate the observed velocity dispersion, hence biasing the best constrained mass higher in the outer halo.
This bias is also clearly visible in the distribution of the best-fit parameters in the right panel of Figure 5, as an overestimation of the mass and an underestimation of the concentration for the majority of mock halos. The average biases combining all the mock samples are listed in Table 1. We adopt these average bias values to calibrate our measurements based on real DESI halo stars. Note this bias accounts for total bias in fitting emPDF to a realistic mock observation, which is mainly driven by observational error but also includes residual biases from any other sources of systematics that are present in the mock.
| Tracer | Bias | Systematic covariance |
| BHB | ||
| RRL |
Table 1. Systematic bias and systematic error covariance matrix of the best-fit parameters, calibrated using error-added mock BHBs and RRLs (see § 4.1.1). The second and third columns show the biases of and the systematic covariance error matrix formed by and .
Calibration of formal and total uncertainties
Figure 4 also shows that the formal error on the mass profile derived from emPDF is slightly lower than the standard deviation among different realizations when observational errors are added to the same halo. This is also seen in the right panel of Figure 5, where the formal posterior contours is comparable to the scatter among realizations, although the latter is slightly larger. The scatter among error-added realizations provides an empirical estimate of the single-halo, single-view statistical scatter. The formal posterior uncertainty primarily reflects the finite information content of the tracer sample under the adopted likelihood. In addition, the total scatter among the best-fits from all the mocks is larger than the formal or statistical error derived for a single halo. As discussed in § 3.3, this enlarged uncertainty reflects additional stochastic biases that are not accounted for by the net bias calibration, , including both deviations from the model assumptions and effects of observational errors not captured by the formal uncertainty. We quote the formal uncertainty as the representative statistical uncertainty throughout this paper, instead of introducing an additional term alone. Since the additional excess scatter is incorporated into the total uncertainty.
Subtracting the average formal covariance matrix from the total covariance measured from all the error-added mocks, we obtain the residual systematic covariance listed in Table 1. In our subsequent analysis using real DESI observations, we convolve the posterior distribution of model parameters produced by emPDF with a Gaussian distribution characterized by this residual systematic covariance matrix to account for the systematic uncertainty. With this construction, the quantity reported as below represents the uncertainty after convolving the formal posterior derived from emPDF with the residual systematic covariance calibrated from the mock ensemble.
Since we have six halos observed in 4 different directions, there are 24 different systems in total. Figure 5 provides an overall uncertainty, which actually contains direction-to-direction and halo-to-halo scatters. In Appendix § 8, we discuss the contributions from these two sources separately. The analysis demonstrates that the halo-to-halo scatter is larger than the direction-to-direction scatter, indicating that variations among halos dominate the systematic uncertainties.
We also perform the same analysis using mock catalogs of RRL and obtain similar conclusions, with the calibrated systematics listed in Table 1. For brevity, we do not repeatedly show the results for mock RRLs.
Best constraint at the median tracer radius
As seen in Figure 4, the enclosed mass is most tightly constrained near the median radius of the tracers, , where both the scatter and bias reach their minimum. Such a characteristic radius generally exists in the results from various types of dynamical models (J. Wolf et al. 2010; J. Han et al. 2016a; W. Wang et al. 2017; Z.-Z. Li et al. 2019; Q. Li et al. 2021; Z. Li et al. 2024; R. Shi et al. 2024). It can be intuitively understood as that dynamical tracers mostly constrain the gravitational acceleration at a representative tracer radius. In addition to the uncertainty in the mass-concentration plane, it is also informative to examine the uncertainty of enclosed mass within the half-mass or median radius of tracers, which can serve as a concise summary of the highly correlated uncertainties in the mass and concentration parameters.
Figure 6 displays the difference between the estimates and truth at the median radius of tracers, together with the statistical and total uncertainty. The statistical uncertainty is obtained by averaging the formal statistical error of individual systems. The total uncertainty is derived by measuring the scatter among all halos or all directions for each case. Comparison of the two types of uncertainties, therefore, represents the additional variance introduced by the diversity among halos or different viewing directions of the same halo. This aligns with our finding that the scatter in the total mass among different halos is larger than the scatter among different directions.
Figure 6. Constraint of mass at the median tracer radius, . In both plots, the points represent mean values of estimated mass relative to the truth at median radius of tracers, and the size of the errorbars denote the associated uncertainties. Left: The pink and blue errorbars show mean statistical errors and total errors of different halos with a fixed viewing direction. Right: The pink and blue errorbars are mean statistical errors and total errors of a fixed halo with different viewing directions.
As shown in Figure 6, the mean bias of estimated enclosed mass at relative to the truth is dex, the mean formal statistical error on is dex. Both the bias and the uncertainty are quite small compared with those at the virial radius, consistent with the expectation that this radius provides the tightest mass constraint.
The total uncertainties taking into account the halo-to-halo scatter and the direction-to-direction scatter are both larger than the statistical uncertainty, amounting to 0.027 and 0.019 dex, respectively. Furthermore, the bias and uncertainty also vary more significantly across different halos than across different observational directions.
Applying to DESI BHB
With the calibrations of biases and uncertainties derived from the mock analyses above, we apply the EM-PDF method to the observational DESI BHB samples. Figure 7 shows the resulting enclosed mass profiles and the corresponding constraints in the halo mass-concentration plane. The results have been corrected for the estimator bias, and we show both the formal error and the total uncertainty which includes the systematic uncertainty. Tabulated mass profile constraints are also listed in Table 2. The mass profile is derived by evaluating the posterior distribution of enclosed mass over all potential parameters at each radius. The median and interval of this distribution define the curve and its band at that radius. This provides a direct, radially local constraint that is not a simple mapping from the mass–concentration contour. The best-fit halo parameters correspond to the maximum posterior probability within the total uncertainty contour to ensure consistency with our bias correction, which was calibrated on mock data using the mean bias of the total error constraints.
| BHB | RRL | Joint | |
| [] | |||
| [] | |||
| [] | |||
| [] | |||
| [] | |||
| [] | |||
| [] |
Table 2. The first and second columns show the enclosed masses at different radii using BHB and RRL as tracers. The joint column on the right shows the best constraints jointly from BHBs, RRLs and the simDF in Z.-Z. Li et al. (2020a). Formal statistical uncertainties and the total uncertainties (in bracket) are both provided. Here the last row shows the best constrained total mass within the median radius of the tracers, with and kpc for BHBs and RRLs, respectively. Note our tracers extend from 20 kpc to kpc.

Figure 7. Left: Best constrained MW mass profiles from BHBs. The blue curve and shaded blue band correspond to the median estimate and total uncertainties, with bias corrected (see § 4.1.1) and systematic errors included. The bronze shaded band represents the formal statistical uncertainty only. The vertical dashed blue line represents the median radius of BHBs. The thick and thin gray curves show the median and bounds from Z.-Z. Li et al. (2020a) using the simDF method with satellites as tracers. Recent studies using stars as tracers are also shown in the plot for comparison, distinguished in the legend as the stars within 30 kpc (“stars near”) and the more distant halo stars(“stars far”). Right: Best constrained MW halo mass and concentration using BHBs. The thick and thin bronze contours represent the and formal statistical uncertainties, while the blue contours show the calibrated total uncertainties including systematics. Gray contours represent the result from simDF for comparison. The background gray dashed lines denote constant values of enclosed masses within the median tracer radius, , showing the best-fit, and values including systematic errors. The numbers on top of the gray dashed lines indicate the corresponding values of . The upper and right subpanels display the marginalized distributions, with the dashed vertical lines showing the best-fit values in the total posterior and the shaded regions showing the uncertainties.
Figure 7 also shows measurements from some other works for comparison. Satellite-based measurements provide a useful comparison because they probe a complementary tracer population at larger radii (e.g., E. Patel et al. 2018; T. M. Callingham et al. [12]; Z.-Z. Li et al. 2020a). Here we focus on the comparison with that from Z.-Z. Li et al. (2020a), who modeled the phase space distribution of 28 MW satellites using the simulation-informed distribution function (simDF) method (Z.-Z. Li et al. 2019). simDF assumes the satellites are in a steady-state with a distribution of orbital parameters, , following a universal function informed by cosmological distributions. The introduction of a simulation informed helps to tighten the dynamical constraint while avoiding biases due to an unjustified DF form, as long as the simulations provide a faithful representation of the real orbital distribution. Overall, the best-fit mass profile from simDF is consistent with our constraints within , across the radial range covered by BHBs. Comparing the statistical uncertainties, our uncertainties are significantly smaller in the inner regions, benefiting from the large number of stellar tracers. The errors increase at larger radii due to the decrease in the number of BHBs used as tracers.
Several recent measurements using stars as tracers are also shown in the figure. Overall, our results are broadly consistent with these measurements within the quoted uncertainties. For the very inner region (20–30 kpc), our mass estimates align well with studies that model the rotation curve of disk-dominated stellar tracers(A.-C. Eilers et al. 2019; H.-F. Wang et al. 2023; Y. Zhou et al. 2023; Y. Jiao et al. 2023; X. Ou et al. 2024). In the outer halo, our profile remains compatible, within the quoted uncertainties, with previous measurements based on distant halo stars (M. Zhai et al. 2018; A. J. Deason et al. 2021; S. A. Bird et al. [8]; J. Shen et al. 2022). The result from J. Shen et al. (2022) shows much smaller uncertainties than ours, despite that they only modelled 168 stars using a DF method. This difference may partly reflect their adopted parametric model for the tracer distribution function, whereas our method avoids assuming a specific functional form for the tracer DF. We also note that the systematic uncertainties due to deviations from steady state and spherical symmetry are usually not modelled in these previous works. Considering these differences, our measurement with the bias-calibrated data-driven DF method provides a conservative yet robust constraint using halo stars.
In addition to the systematic uncertainties discussed above, the variations between individual measurements also reflect the different tracer populations, radial coverages, and modelling assumptions used in these studies. For example, the constraints from J. Shen et al. (2022) are based on tracer samples distributed in much larger radii, so the inferred inner mass profile depends on extrapolations that rely on the adopted potential model. M. Zhai et al. (2018) modeled K giants over 5–120 kpc using the spherical Jeans equation, which requires assumptions about the tracer density profile and velocity anisotropy. These differences likely contribute to the method and tracer dependent scatter among current halo star mass measurements.
The right panel of Figure 7 shows the posterior distribution of the virial mass and concentration parameters. The constraints show a tight degeneracy between the mass and concentration parameters. This degeneracy is equivalent to a tight constraint on the mass within the median tracer radius, which is also shown via the gray dashed contours. Parameter combinations along the degeneracy direction yield very similar , while those perpendicular to it correspond primarily to variations in .
Compared with the simDF result of Z.-Z. Li et al. (2020a), our constraint shows substantial overlap with the simDF constraint, yet the best-fits exhibit slight offset in virial mass and concentration. Importantly, this offset is largely aligned with the main mass-concentration degeneracy direction, so the two analyses still imply similar enclosed masses near the characteristic radius directly probed by the tracers. Along the degeneracy direction of our constraint, simDF places tighter constraint than ours. This is mainly because simDF also makes use of a simulation-informed in the model, which provides complementary information to the steady-state assumption (Z.-Z. Li et al. 2019; Z. Li et al. 2024). Besides, the two samples span difdifferent radial ranges, which also contributes to the different degeneracy directions as they constrain different . Perpendicular to our degeneracy direction, however, our constraints are much tighter than that from simDF, due to the larger tracer sample size in our work.
The best-fit values of and along with their uncertainties are reported in Table 3. As the posterior distributions are close to Gaussian, we also provide the parameters of a 2-dimensional Gaussian approximation to the joint posterior of , which can be easily reproduced for comparisons with other works.
Summary of the constraints from DESI BHB and RRL stars, and the joint stellar-satellite analysis. For each type of tracers we report (i) the best-fit values of and , with formal statistical and total uncertainties (second and third columns from the left), and (ii) the parameters of a 2-dimensional Gaussian approximation to the posterior of , including the means (), standard deviations (), and correlation coefficient .
| Tracer | (stat / total) | (stat / total) | Gaussian fit, stat () | Gaussian fit, tot () |
| BHB | ||||
| RRL | ||||
| Joint |
Table 3.
Applying to DESI RR Lyrae
We perform a similar mock-calibration analysis to the RRL samples. The best-fit values of and derived from RRLs along with the 2D Gaussian fitting results are listed in Table 3. The inferred enclosed mass profile at different radius are shown in Table 2.
Figure 8 compares the enclosed mass profiles and mass-concentration contours derived from RRLs with those obtained from the BHBs. For clarity, only the total uncertainties are shown here. Overall, the RRL and BHB based constraints remain generally consistent within their total uncertainties, with the RRL result favoring a modestly higher mass profile over most of the radial range shown. This difference is also reflected in the plane, where the RRL constraint is shifted toward slightly higher virial mass and lower concentration relative to the BHB result.
Figure 8. Left: MW mass profiles inferred from DESI main bright BHBs and RRLs. The blue and orange dashed curves with shaded bands in corresponding colors show the median estimate and total uncertainty based on BHBs and RRLs, respectively, with biases corrected (see § 4.1.1) and systematic errors included. The gray line with gray dotted band represent the mean and total uncertainty from the joint constraint of BHBs, RRLs and satellites. The vertical blue and orange dashed lines mark the median radii of BHBs and RRLs. Right: Best constrained MW halo mass and concentration. Blue and orange crosses with shaded contours in corresponding colors denote the best constraints and total uncertainties based on BHBs and RRLs, with systematics errors included. The gray cross and shaded contour indicate the joint constraint from halo stars plus satellites and the total - uncertainty. Marginalized distributions are displayed in the upper and right subpanels.
The enclosed mass inferred from the RRL analysis is correspondingly slightly higher than that from the BHBs, although the two remain consistent within the uncertainties. The two independent tracer populations produce overlapping, though not identical, constraints. This suggests that the overall inference is robust, while some variation between different tracers persists at the current precision. Notably, the BHB and RRL contours are elongated along a very similar mass-concentration degeneracy direction, implying that the difference between them is smaller in terms of enclosed mass near the median tracer radius than it appears in the plane alone.
Joint constraint with satellites and measurements using the oPDF
To take advantage of the complementary information provided by different tracers, we combined the posterior constraints from our BHB and RRL analyses with the satellite-based constraints of Z.-Z. Li et al. (2020a). Specifically, we approximate the satellite posterior with a bivariate Gaussian and multiply it with the calibrated BHB and RRL posterior distributions, assuming that the three constraints are independent, to obtain the joint posterior. Different tracer populations probe different regions of the halo, and the distinct radial coverage and modelling assumptions lead to characteristic degeneracy directions in the parameter plane. Therefore, joint constraints help to significantly tighten the constraints. Such improvements are shown clearly in Figure 8. Combining the BHB and RRL constraints alone does not qualitatively change the degeneracy direction, since the two halo star tracers are elongated along a very similar direction in the plane, although it does improve the overall constraint. Once the satellite constraint which has a more complementary orientation in parameter space is further included, the posterior shrinks substantially and the degeneracy is greatly reduced. The resulting joint mass profile becomes more tightly constrained particularly at larger radii, where the satellites contribute most strongly. The joint constraint is therefore driven both by the consistency of the two stellar tracers and by the complementary parameter space orientation of the satellite-based result. As both methods involve minimal assumptions that are well tested and calibrated in simulations, our combined analysis provides one of the tightest and most robust measurements of the mass profile of the MW.
Finally, we repeat the analysis on the same BHB and RRL samples using the precursor of EMPDF, the oPDF method, which can help to cross check with the constraints obtained above. In Appendix § 9, we show that the results from oPDF are broadly in agreement with those from the EMPDF, but with generally larger uncertainty, consistent with the fact that EMPDF uses a smoothed empirical DF and a full phase-space likelihood.
Conclusions
In this work, we constrain the Milky Way’s mass profile using the data-driven dynamical modeling method EMPDF, applied to a large sample of halo stars from the DESI MWS. Specifically, we utilize BHB and RR Lyrae stars as dynamical tracers owing to their accurate distance estimates and widespread presence in the outer stellar halo. Our analysis begins with tests on the DR2 AurigaDESI mock catalogues presented in Appendix § 8. H. Riley et al. (2026), to calibrate estimator uncertainties and biases and quantify systematics across different halos and viewing angles. The calibrated estimator is then applied to the real DESI observations. This analysis forms part of the broader DESI DR2 Milky Way mass Key
Project, alongside complementary mass profile measurements from other dynamical methods (G. E. Medina et al. 2026; M. Valluri et al. 2026) and the overall synthesis presented in DESI Collaboration et al. (2026). The key findings and conclusions are summarized as follows:
In the absence of observational errors, emPDF can recover the true mass profile in an ensemble unbiased manner. However, when realistic observational errors are incorporated, an overall bias emerges, particularly due to the increasing uncertainties in the tangential velocity measurements at larger distances. Modeling on mock data helps to calibrate the results from real data, including the correction of both systematic bias and scatter.
By applying our method to mock samples, we find that the total scatter of the best-fit parameters is much larger than the average statistical error, indicating the presence of systematic errors. This
primarily originates from the halo-to-halo scatter. Nevertheless, the constraint on the enclosed mass at the median tracer radius remains remarkably tight even after including systematics, confirming that this characteristic radius provides the strongest dynamical constraint.
Applying the calibration derived from the mocks to the observational BHB data, we obtain and . The corresponding enclosed mass at 100 kpc is . The BHB-based profile is broadly consistent with previous constraints from halo stars within the quoted uncertainties.
Applying similar calibration to observational RR Lyrae data, the best constrained mass and concentration are and . The RRL and BHB results are closely consistent within their uncertainties, providing mutually consistent mass profile constraints from two independent halo tracer populations.
The constraints from our analysis are highly complementary to previous constraints using satellite dynamics. The mass profile constraints from our halo star samples are tightest in the inner halo at kpc, with a total uncertainty of dex. This corresponds to a tight degeneracy in the derived mass and concentration parameters, perpendicular to the previous constraint from Z.-Z. Li et al. (2020a) using satellite dynamics. By combining the constraints from the stars with those from the satellites, we obtain a joint posterior that provides one of the tightest yet well-calibrated constraints to date on the MW halo mass, with the best-fit and .
Our results also provide an external perspective on recent claims of a declining Milky Way rotation curve. Several measurements derived from rotation curve have suggested that the circular velocity declines beyond kpc, with some studies finding a particularly steep drop that may even approach a near-Keplerian behaviour (H.-F. Wang et al. 2023; Y. Jiao et al. 2023; X. Qu et al. 2024). Such a rapidly declining outer rotation curve would point to an unusually low Milky Way halo mass, in tension with many independent constraints based on halo stars and satellites. However, recent work has shown that Jeans based inferences of the rotation curve can be sensitive to departures from axisymmetry. and equilibrium, and to assumptions about the tracer density profile, all of which can bias the inferred decline at large radii (O. Koop et al. 2024; X. Ou et al. 2025). In this context, our constraints, which probe the MW mass profile on outer halo scales using halo stars, do not favor such a low Milky Way mass implied by the most extreme declining rotation curve interpretations. The sharpest low mass inferences are likely sensitive to modelling assumptions and systematics. Our results are instead more consistent with the broader set of other dynamical measurements.
We also applied the oPDF method to the same observational samples for validation and comparison. The inferred mass profiles derived from the oPDF are generally consistent with those from the emPDF, demonstrating that our inference is robust. This consistency further supports the stability and reliability of our constraints on the MW mass.
A natural extension to this work would involve implementing a framework to inherently account for observational error, such as the tentative approach outlined in Appendix C in Z. Li et al. (2024). This would eliminate the need for external calibration from specific mock data. However, we note that simulation calibrations will remain crucial for estimating the systematic uncertainties due to deviation from steady state and spherical symmetry. Because of this, mock building efforts which can produce a much larger number of independent mock observations than AuriDESI will prove helpful for future analysis. We have already shown that combining halo stars with satellites yields a much tighter constraint, and additional tracers such as globular clusters could further strengthen the joint analysis. More broadly, the complementarity among different tracers and dynamical methods suggests that a more systematic combination of independent Milky Way mass constraints will be an important direction for future work. Another useful extension would be to adopt more general Galactic potential models. In particular, the centrally concentrated stellar and gaseous mass could be modeled separately from the extended halo, for example through a central monopole term or a more detailed multi-component potential. Such an extension would allow us to assess more directly how the treatment of the central baryonic mass affects the inferred halo scale radius and concentration. Together with analyses based on a larger suite of MW-like simulations, these developments would help better disentangle and quantify the various sources of systematic uncertainty.
In conclusion, this study demonstrates the power of combining the self data-driven method emPDF, which can fully utilize 6D phase space information of dynamical tracers, with the abundant data of halo stars provided by DESI MWS. With careful calibration based on mock analyses, emPDF delivers precise mass estimates that incorporate systematic uncertainties. As the DESI survey progresses, expanding the sample size and refining observational data, a tighter constraint of MW mass is anticipated in the future.
Data Availability
The machine-readable data underlying the figures have been deposited on Zenodo and will be publicly available upon publication (https://zenodo.org/records/22644507).
Acknowledgments
This work is supported by NSFC (12595312, 12573022, 12022307), the National Key R & D Program of China (2023YFA1605600, 2023YFA1605601, 2023YFA1607800, 2023YFA1607801), 111 project (No. B20019), the science research grants from the China Manned Space Project (No.CMS-CSST-2025-A04), and the Office of Science and Technology, Shanghai Municipal Government (grant Nos. 24DX1400100, ZJ2023-ZD-001). We thank the sponsorship from Yangyang Development Fund. Z.Z.L. acknowledges the Fundamental Research Funds for the Central Universities (KG202502), and the Marie Skłodowska-Curie Actions Fellowship (101109759, CuspCore) under Horizon Europe. CAP acknowledges financial support from the Spanish Ministry of Science, Innovation and Universities (MICIU) projects PID2020-117493GB-I00, PID2023-149982NB-I00 and PID2023-146453NB100. AHR was supported by a fellowship funded by the Wenner-Gren Foundations (Wenner-Gren Stiftelserna), a Research Fellowship from the Royal Commission for the Exhibition of 1851, and by the Science & Technology Facilities Council (STFC) through grant ST/T000244/1. GEM acknowledges support from the Dutch Research Council (NWO) through the Veni grant VI.Veni.252.355. GEM and TSL acknowledge financial support from Natural Sciences and Engineering Research Council of Canada (NSERC) through grant RGPIN-2022-04794.
The computations of this work are carried on the Gravity supercomputer at the Department of and the National Energy Research Scientific Computing Center (NERSC).
This material is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of High-Energy Physics, under Contract No. DE–AC02–05CH11231, and by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract. Additional support for DESI was provided by the U.S. National Science Foundation (NSF), Division of Astronomical Sciences under Contract No. AST-0950945 to the NSF’s National Optical-Infrared Astronomy Research Laboratory; the Science and Technology Facilities Council of the United Kingdom; the Gordon and Betty Moore Foundation; the Heising-Simons Foundation; the French Alternative Energies and Atomic Energy Commission (CEA); the Secretariat of Science, Humanities, Technology and Innovation (SECIHTI) of Mexico; the Ministry of Science, Innovation and Universities of Spain (MICIU/AEI/10.13039/501100011033), and by the DESI Member Institutions: https://www.desi.lbl.gov/collaborating-institutions. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the U. S. National Science Foundation, the U. S. Department of Energy, or any of the listed funding agencies.
The authors are honored to be permitted to conduct scientific research on I’oligam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation.
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.
ChatGPT (OpenAI; GPT-5.5) was used during manuscript preparation to assist with language editing and limited code debugging. All scientific analyses, interpretations, references, and final manuscript content were reviewed and verified by the authors.
Software: ASTROPY ([1, 2, 3]), JUPYTER ((T. Kluyver et al. 2016)), NUMPY ((C. R. Harris et al. 2020)), PANDAS ((W. McKinney 2010)), MATPLOTLIB ((J. D. Hunter 2007)), h5PY ((A. Collette 2013)), SCIPY ((P. Virtanen et al. 2020)), ChatGPT (GPT-5.5) ((OpenAI 2026))
Halo-to-halo and directional systematic uncertainties
We further decompose the systematic uncertainty using the error-added BHB mock samples. For each fixed halo and viewing direction, we first average the recovered parameters over the 20 error realizations. This removes the stochastic scatter associated with individual error realizations and allows us to focus on the variations between different halos and between different viewing directions. In Figure 9, we obtain the estimates for different halos while keeping the viewing direction fixed. The estimates shown in this plot have been corrected for the net bias, allowing us to focus on the scatter. For fixed viewing direction, the averaged estimate of different halos remain broadly distributed around the origin, suggesting the presence of halo-to-halo scatter. Similarly, for the same halo observed from different directions, the scatter of the averaged estimates represents the direction-to-direction scatter, as shown in Figure 10. Notably, for each fixed viewing direction, the ensemble of halos does not show a strong coherent offset in any particular viewing direction. While for a fixed halo, the estimates from different viewing directions often remain in a similar region of the plane for a given halo, rather than being symmetrically distributed around the origin. This suggests that each halo can have its own baseline offset.
Figure 9. Bias-corrected residuals in the best-fit halo mass () and concentration () for the error-added BHB mock samples. Each panel fixes one viewing direction and compares the six mock halos. Faded symbols show the individual error realizations, while the larger symbols show the mean over the 20 realizations for each halo. The gray ellipse is computed only from the larger symbols and therefore represents the halo-to-halo scatter at fixed viewing direction. The numbers in brackets give the root-mean-square error of mass and concentration for the realization-averaged estimates in each panel.
Figure 10. Similar to Figure 9, but each panel fixes one halo and compares the four viewing directions. Faded symbols show the individual error realizations, while the larger symbols show the mean over the 20 realizations for each viewing direction. The gray ellipse is computed only from the larger symbols and therefore represents the direction-to-direction scatter at fixed halo. The root-mean-square error of mass and concentration for the larger symbols are provided in the bracket of the upper left corner in each panel.
Changing the viewing direction moves the estimates around this halo-dependent baseline but does not fully eliminate it.
Quantitatively, the mean halo-to-halo scatter measured from the realization-averaged centroids is , where the average is taken over the four fixed viewing directions. The corresponding direction-to-direction scatter is , where the average is taken over the six fixed halos. The halo-to-halo scatter is therefore larger in both parameters, This indicates that variations among halos constitute the dominant source of systematic uncertainty, although the contribution from viewing-direction variation remains non-negligible.
Results from oPDF
We calibrate the oPDF results independently using the AuriDESI mocks, following the same general procedure as for eMPDF. The final results are compared in Figure 11. Overall, the mass profiles inferred from oPDF are broadly consistent with those from eMPDF, although the details of the shifts differ between tracers. The oPDF constraints are generally broader, while the two tracers remain statistically consistent within the total uncertainties for both estimators. The corresponding constraints in the mass-concentration plane are also consistent within the total uncertainty contours. These results support the robustness of the inferred MW mass profile against the choice of dynamical estimator.
Figure 11. Comparison between the results obtained from emPDF and oPDF. Left: MW mass profiles inferred from both methods. The green curve with light green boundary and purple curve with dotted purple band represent the oPDF results for BHBs and RRLs, respectively. Blue and orange dashed curves with shaded band in corresponding colors are based on emPDF and are for BHBs and RRLs. Only the total uncertainties including systematic errors are shown here. Right: Mass and concentration inferred by emPDF and oPDF (see the legend). The filled contours are total uncertainties. The upper and right subpanels show the marginalized distributions.
References
- [1]Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068
- [2]Astropy Collaboration, Price-Whelan, A. M., Sipőcz, B. M., et al. 2018, AJ, 156, 123, doi: 10.3847/1538-3881/aabc4f
- [3]Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167, doi: 10.3847/1538-4357/ac7c74DOI
- [4]Belokurov, V., & Koposov, S. E. 2016, MNRAS, 456, 602, doi: 10.1093/mnras/stv2688DOI
- [5]Beraldo e Silva, L., Valluri, M., Vasiliev, E., et al. 2025, ApJ, 990, 109, doi: 10.3847/1538-4357/adf1e3
- [6]Binney, J., & Tremaine, S. 2008, Galactic Dynamics: Second Edition
- [7]Bird, S. A., Xue, X.-X., Liu, C., et al. 2021, ApJ, 919, 66, doi: 10.3847/1538-4357/abfa9eDOI
- [8]Bird, S. A., Xue, X.-X., Liu, C., et al. 2022, MNRAS, 516, 731, doi: 10.1093/mnras/stac2036DOI
- [9]Boylan-Kolchin, M., Bullock, J. S., & Kaplinghat, M. 2011, MNRAS, 415, L40, doi: 10.1111/j.1745-3933.2011.01074.x
- [10]Bressan, A., Marigo, P., Girardi, L., et al. 2012, MNRAS, 427, 127, doi: 10.1111/j.1365-2966.2012.21948.xDOI
- [11]Byström, A., Koposov, S. E., Lilleengen, S., et al. 2024, arXiv e-prints, arXiv:2410.09149, doi: 10.48550/arXiv.2410.09149DOI
- [12]Callingham, T. M., Cauntun, M., Deason, A. J., et al. 2019, MNRAS, 484, 5453, doi: 10.1093/mnras/stz365DOI
- [13]Cautun, M., Hellwing, W. A., van de Weygaert, R., et al. 2014, MNRAS, 445, 1820, doi: 10.1093/mnras/stu1829DOI
- [14]Cautun, M., Benítez-Llambay, A., Deason, A. J., et al. 2020, MNRAS, 494, 4291, doi: 10.1093/mnras/staa1017DOI









