Draft version February 2, 2024

Typeset using LATEX twocolumn style in AASTeX63

The Power of High Precision Broadband Photometry: Tracing the Milky Way Density Profile with
Blue Horizontal Branch Stars in the Dark Energy Survey
Fengqing Yu (余枫青),1, 2 Ting S. Li,3, 4, 2 Joshua S. Speagle (沈佳士),5, 3, 4, 2 Gustavo E. Medina,3, 4
Sergey E. Koposov,6, 7, 8 Joss Bland-Hawthorn,9, 10 Lara R. Cullinane,11 Gwendolyn M. Eadie,3, 5, 2
Denis Erkal,12 Geraint F. Lewis,9 Guilherme Limberg,13 and Daniel B. Zucker14, 15
1 Department of Computer Science, University of Toronto, 40 St. George Street, Canada
arXiv:2402.00104v1 [astro-ph.GA] 31 Jan 2024

2 Data Sciences Institute, University of Toronto, 17th Floor, Ontario Power Building, 700 University Ave, Toronto, ON M5G 1Z5,
Canada
3 David A. Dunlap Department of Astronomy & Astrophysics, University of Toronto, 50 St George Street, Toronto ON M5S 3H4, Canada
4 Dunlap Institute for Astronomy & Astrophysics, University of Toronto, 50 St George Street, Toronto, ON M5S 3H4, Canada
5 Department of Statistical Sciences, University of Toronto, 9th Floor, Ontario Power Building, 700 University Ave, Toronto, ON M5G

1Z5, Canada
6 Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh EH9 3HJ, UK
7 Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK
8 Kavli Institute for Cosmology, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK
9 Sydney Institute for Astronomy, School of Physics, A28, The University of Sydney, NSW 2006, Australia
10 Centre of Excellence for All-Sky Astrophysics in Three Dimensions (ASTRO 3D), Australia
11 Department of Physics and Astronomy, Johns Hopkins University, 3400 N. Charles St, Baltimore, MD 21218, USA
12 Department of Physics, University of Surrey, Guildford GU2 7XH, UK
13 Universidade de São Paulo, IAG, Departamento de Astronomia, SP 05508-090, São Paulo, Brazil
14 School of Mathematical and Physical Sciences, Macquarie University, Sydney, NSW 2109, Australia
15 Macquarie University Research Centre for Astrophysics and Space Technologies, Sydney, NSW 2109, Australia

ABSTRACT
Blue Horizontal Branch (BHB) stars, excellent distant tracers for probing the Milky Way’s halo
density profile, are distinguished in the (g − r)0 vs (i − z)0 color space from another class of stars, blue
straggler stars (BSs). We develop a Bayesian mixture model to classify BHB stars using high-precision
photometry data from the Dark Energy Survey Data Release 2 (DES DR2). We select ∼ 2100 highly-
probable BHBs based on their griz photometry and the associated uncertainties, and use these stars
to map the stellar halo over the Galactocentric radial range 20 ≲ R ≲ 70 kpc. After excluding known
stellar overdensities, we find that the number density n⋆ of BHBs can be represented by a power law
density profile n⋆ ∝ R−α with an index of α = 4.28+0.13 −0.12 , consistent with existing literature values.
In addition, we examine the impact of systematic errors and the spatial inhomogeneity on the fitted
density profile. Our work demonstrates the effectiveness of high-precision griz photometry in selecting
BHB stars. The upcoming photometric survey from the Rubin Observatory, expected to reach depths
2-3 magnitudes greater than DES during its 10-year mission, will enable us to investigate the density
profile of the Milky Way’s halo out to the virial radius, unravelling the complex processes of formation
and evolution in our Galaxy.
1. INTRODUCTION merged and assembled (Fukushima et al. 2019; Searle
The Milky Way halo contains fundamental informa- & Zinn 1978). Hence, the stellar halo preserves a fossil
tion about the evolution history of our Galaxy and na- record of the Galaxy’s formation history and past ac-
ture of dark matter (Helmi 2008). It is generally believed cretion events (Helmi 2020). Moreover, the stellar halo
that the Milky Way is formed via hierarchical formation, can provide insight into the structure of the dark mat-
where small stellar systems such as dwarf galaxies are ter halo through stellar dynamics (Gerhard 2012). In
upcoming surveys, it may even be possible to measure
the Galaxy’s change in mass with cosmic time by using
Corresponding author: Ting S. Li the kinematics of a smoothly distributed halo popula-
[email protected]
2 Yu, Li, Speagle et al.

tion (Sharma et al. 2023), assuming such a population color space. However, this separation becomes more dif-
exists. ficult for fainter stars because of the increase in uncer-
The Milky Way stellar halo can be directly probed tainties, and the decreased fraction of BHBs at fainter
with tracer populations such as red giant-branch (RGB) magnitudes.
stars, RR Lyrae (RRL) stars, and blue horizontal- In this study, we exploit the BHB/BS separation seen
branch (BHB) stars. These stars are bright, and can be in Li et al. (2019) and present a statistical model to
observed even in the very periphery of the Milky Way. classify BHB stars using photometric data from Dark
Among them, BHB stars are frequently used because Energy Survey Data Release 2 (DES DR2; Abbott et al.
their absolute magnitudes (and thus distances) are rela- 2021), incorporating photometric uncertainties. Using
tively straightforward to calibrate (Preston et al. 1991; BHB candidates selected from our model, we measure
Fukushima et al. 2018). The stellar density profile of the the Galactic stellar halo density profile n⋆ (R) and com-
halo n⋆ (R) is often fitted with an inverse power law with pare its slope with literature values.
index α, such that n⋆ ∝ R−α , where R is the Galacto- In Section 2, we present an overview of the photomet-
centric radius. Deason et al. (2011a) mapped BHB us- ric data and preprocessing methods used in this study.
ing Sloan Digital Sky Survey Data Release 8 from 1 - 40 In Section 3, we describe our statistical model that pre-
kpc with a broken power-law model; they found an in- dicts the probability of a star being a BHB based on its
ner slope αin = 2.3 and αin = 4.6 with the break radius colors and associated uncertainties. In Section 4, we use
≈ 27 kpc. Deason et al. (2018) subsequently mapped the selected BHB candidates to derive the Milky Way
BHB stars starting at 50 kpc using Hyper Suprime-Cam density profile. In Section 5, we discuss our findings and
(HSC) photometry; they measured a slope of 4 when ex- implications, and we conclude in Section 6.
cluding the Sagittarius (Sgr) stream, consistent with the
power law from smaller distances. 2. DATA
On the other hand, Thomas et al. (2018) showed an 2.1. High-quality Photometry
inner slope of α ≈ 4.2 and a shallower outer slope of
High-precision photometry is required for teasing out
α ≈ 3.2 beyond a radius of about 40 kpc by mapping
the BHB signal from the contaminants for a large
BHB stars in the Canada-France Imaging Survey I. More
dataset. Here, we consider the photometry from the
recently, addressing the dichotomy in literature values of
Dark Energy Camera (DECam; Flaugher et al. 2015),
the single breaking radius, Han et al. (2022) showed a
given its wide sky coverage, large telescope aperture,
doubly broken power law with break radii at 12 kpc and
and the clear BHB/BS sequence separation identified
28 kpc using the H3 Survey (Conroy et al. 2019). Be-
by Li et al. (2019). We utilize two DECam-based sur-
cause of the discrepancies in the slope and break radius,
veys, the Dark Energy Survey (DES) and the DECam
the stellar halo density profile continues to be an active
Legacy Survey (DECaLS), and compare the results from
area of research.
the public data releases of these two surveys. As part of
While BHBs are incredibly useful, finding them in
the evaluation, we use the BHB/BS classification based
photometric data can be challenging. To select BHBs
on the measurements from S 5 . Throughout the paper,
from photometric data, the main obstacle lies in the re-
the subscript 0 indicates our use of extinction-corrected
moval of other contaminants that have similar colors and
photometry; this correction is made possible using the
magnitudes to BHBs, such as blue straggler stars (BSs),
dust map derived by Schlegel et al. (1998).
white dwarfs (WDs), and quasi-stellar objects (QSOs).
The DES is a wide-field optical/near-infrared imaging
Among them, the removal of BSs remains the most diffi-
survey that contains 400 million astronomical objects
cult as these are much closer to BHBs in the color-color
(Dark Energy Survey Collaboration et al. 2016), cover-
space than other contaminants. Previous works have
ing ∼5000 deg2 of the South Galactic Cap region and
largely relied on the u − g color to distinguish BHBs and
obtaining photometry in g, r, i, z, Y bands. In this com-
BSs (Yanny et al. 2000; Deason et al. 2011a). Deason
parison, we include both DES DR1 and DES DR2, the
et al. (2018) used a combination of g, r, i, z photometry
former of which is also used by Li et al. (2019).
from HSC to identify BHB stars. Using the BHB and
We also exploit the DECaLS survey, which forms part
BS classifications based on surface gravity and effective
of the DESI Legacy Imaging Surveys (Dey et al. 2019).
temperature from the Southern Stellar Stream Spectro-
The latter mapped 14,000 deg2 of the extragalactic sky
scopic Survey (S 5 ) and photometry from the Dark En-
visible from the northern hemisphere in three optical
ergy Survey Data Release 1 (DES DR1; Abbott et al.
bands (g, r, and z). DECaLS data release 9 (DECaLS
2018), Li et al. (2019) identified a distinct and separable
DR9) incorporates the DES imaging, which is used in
sequence for BHB and BS stars in (i−z) vs (g −r) color-
the comparison here.
 Mapping the Galactic Halo with BHBs 3

0.10 BHB 0.10 BHB

0.15 BS 0.15 BS

0.20 0.20
0.25 0.25
(r z)0

(r z)0
0.30 0.30
0.35 0.35
0.40 0.40
0.45 DES DR1 0.45 DECaLS DR9
0.50 0.50
0.4 0.3 0.2 0.1 0.0 0.1 0.4 0.3 0.2 0.1 0.0 0.1
(g r)0 (g r)0
0.10 BHB 0.10 BHB
0.15 BS 0.15 BS
Outlier
0.20 0.20
0.25 0.25
(r z)0

0.30 (r z)0 0.30

0.35 0.35
0.40 0.40
0.45 DES DR2 (S5-Based Classification) 0.45 DES DR2 (Model-Based Classification)
0.50 0.50
0.4 0.3 0.2 0.1 0.0 0.1 0.4 0.3 0.2 0.1 0.0 0.1
(g r)0 (g r)0
Figure 1. Photometry from DES DR1 (top left), DECaLS DR9 (top right), and DES DR2 (bottom left) cross-matched with S 5
survey targets, with target classes of BHB and BS as classified by S 5 surface gravity and temperature measurements (see Section
2). The bottom right panel uses the same photometry as the bottom left, but classified according to the photometric mixture
model described in Section 3. The improved data quality of DES DR2 relative to DES DR1 and DECaLS DR9 is clearly visible,
with both BHB and BS sequences becoming tighter and their separation more apparent. The high quality photometry in DES
DR2 is one of the key factors that enables our mixture model to accurately recover the underlying BHB and BS populations
from photometric colors alone. We note that we do not include i−band photometry here for comparison since DECaLS DR9
does not include i−band data.
2.2. Spectroscopic Crossmatching We cross-match stars in S 5 DR1 with DES DR1, DES
S 5 (Li et al. 2019) is a spectroscopic survey with an DR2, and DECaLS DR9 respectively. Following Li et al.
initial focus on identifying stream member stars within (2019), We restrict our selection to stars with effective
the footprint of the DES, and an eventual goal of map- temperature 6000K < Teff < 10000K and surface grav-
ping the entire Southern sky. Its first public data re- ity 2.5 < log g < 6. We then apply photometric cuts
lease (S 5 DR1; Li & S5 Collaboration 2021) includes a of 16 < g0 < 19, −0.35 < (g − r)0 < −0.05 and
total of ∼ 31000 stars, most of which are concentrated −0.5 < (i − z)0 < 0.1. To ensure reliable stellar pa-
on streams within the DES footprint. In S 5 DR1, the rameters, we require the stars in S 5 to have high signal-
survey measures stellar parameters by fitting interpo- to-noise ratio (S/N) spectra (sn 1700d > 5) and small
lated stellar atmosphere models to the IR spectra in the uncertainties on surface gravity (logg std < 0.5). Fur-
Calcium Triplet region. As shown in Figure 11 of Li thermore, we require that all g, r, i, z photometry are
et al. (2019), the stellar parameters of surface gravity available in DES DR2. RR Lyrae stars are removed by
and effective temperature from S 5 DR1 can effectively cross-matching with the RR Lyrae catalog from Gaia
provide a distinction between BHB stars and BS stars, Data Release 3 (Gaia Collaboration et al. 2023). Af-
with BS stars having higher surface gravity at the same ter the cut, a total of 365 stars remain. We adopt the
temperature. same criterion defined by Li et al. (2019) for BHB and
 4 Yu, Li, Speagle et al.

0.10 the selected magnitude ranges in Figure 3. From left

0.15 to right, as the magnitude increases and stars become
fainter, the BHB sequence and BS sequence merge to-
0.20
gether and a clear separation is no longer observed. This
0.25 phenomenon is mostly due to the increase in photomet-
(r z)0

0.30 ric uncertainty at fainter magnitudes. While uncertain-

0.35 ties in both (i−z)0 and (g−r)0 increase as the magnitude
increases, uncertainties in (i − z)0 are approximately 3
0.40 times larger than those in (g − r)0 for stars with BHB-
0.45 like colors. At g0 ≈ 21, the mean uncertainty for (i−z)0
0.50 additional color cut reaches 0.035, but the mean uncertainty for (g − r)0 is
0.35 0.30 0.25 0.20 0.15 0.10 0.05 only around 0.012. Thus, we enforce an additional mag-
(g r)0 nitude cut of 16 < g0 < 21. In total, we selected 46031
sources for our mixture model described in Section 3.
Figure 2. Extinction-corrected (g − r)0 vs (i − z)0 of DES
DR2 photometry for stars with magnitudes in the range 18 ≤
3. MIXTURE MODEL
g0 ≤ 20). The color-color cut used to remove contamination
from WDs and quasars (QSOs) (lower sequence) is shown as In this section, we describe a mixture model that pre-
a dashed red line. dicts the probability of a star being a BHB, dependent
on the g0 magnitude, (g − r)0 and (i − z)0 color.
BS classification in log g − Teff space. In Figure 1, we
depict the stars in (r − z)0 vs (g − r)0 color-color space1 , 3.1. General Form of the Likelihood
with photometry from DES DR1, DECaLS DR9, and We assume that our observed data can be classified
DES DR2 respectively. into three categories: BHBs, BSs, and other contaminat-
Compared to both DES DR1 and DECaLS DR9, DES ing sources (which we will refer to as “outliers”). Each
DR2 shows tighter sequences of BHB and BS stars and of these groups occupies fractions fBHB , fBS and fout in
a clearer separation. The improvement in photometric our dataset, respectively, subject to the constraint that
calibration in DES DR2 enables our mixture model de- fBHB +fBS +fout = 1. We assume that the BHB and BS
scribed in Section 3 to differentiate the BHB and BS stars lie along sequences in (g − r)0 and (i − z)0 color-
sequences with photometry alone. This is an important color space centered around central ridgelines µBHB and
result. µBS , with some amount of intrinsic scatter σBHB and
σBS respectively, while outliers are drawn from an alter-
2.3. Data Preparation
nate distribution which we take to be uniform between
Our ultimate goal is to develop a model which allows some lower and upper color bounds cmin and cmax . Fi-
a purely photometric selection of BHB stars, since this nally, we assume our observed colors and magnitudes
allows us to select a much larger sample of these stars have uncertainties σobs which are accurately estimated
than those observed by spectroscopic surveys. We build by the photometric pipelines.
our model with DES DR2. We first apply a magnitude This gives us the following statistical model. First,
cut between 15 mag and 24 mag for (g, r, i, z) photome- we sample the object class Y based on the fractions of
try. Next, high-quality photometry is selected by using objects in each group given some extra data values D
flags g,r,i,z< 4. We use extended class coadd ≤ 1 (e.g., magnitudes):
to select only objects highly likely to be stars. To select
blue stars, we apply color cuts of −0.35 < (g − r)0 < Y |D ∼ Categorical(fBHB , fBS , fout ) (1)
−0.05 (as motivated by S 5 DR1 x DES DR2 in Figure
1) and −0.2 < (i − z)0 < 0.05. One additional color cut, Then, given the result, we assign one of three different
(r − z)0 ≥ −0.17 + (g − r)0 × 0.09, is used to remove color-color distributions to the intrinsic colors X:
outliers likely to be WDs and QSOs, as shown in Figure X|Y = BHB, D ∼ Normal(µBHB , σBHB ) (2)
2.
The stars are then separated into different magnitude X|Y = BS, D ∼ Normal(µBS , σBS ) (3)
bins, based on g0 band magnitude. We show stars at X|Y = Outlier, D ∼ Uniform(cmin , cmax ) (4)

For example, if the star is categorized as a BS, then the

1 Note the i-band is not used here because DECaLS DR9 does not data is assumed to be distributed as normal with mean
contain i-band photometry µBS and standard deviation σBS .
 Mapping the Galactic Halo with BHBs 5
0.05 0.05 0.05
16 g0 17 19 g0 20 20.5 g0 21
0.00 0.00 0.00

0.05 0.05 0.05

(i z)0

(i z)0

(i z)0
0.10 0.10 0.10

0.15 0.15
0.15

0.20 0.20
0.20
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.35 0.30 0.25 0.20 0.15 0.10 0.05
(g r)0 (g r)0 (g r)0

Figure 3. As Figure 1, but now showing the distribution of stars in three different extinction-corrected g0 magnitude ranges.
As we extend into fainter magnitudes, the larger measurement errors cause the two well-defined sequences of BHB and BS
stars to blend together, making it harder to disentangle potential BHB candidates. This motivates the use of the probabilistic
mixture model described in Section 3.
Finally, we add in observational uncertainties to get colors given our observations. Since it is not guaranteed
the distribution of the observed colors Xobs : to obtain an analytic solution, we will write out the full
likelihood below:
Xobs |X ∼ Normal(X, σobs ) (5)
P (Xobs,i |Di , θ) =
Given a collection of n independent and identically Z
distributed (iid) observations {Xobs,i }ni=1 and combin-
X
fj (Di ) P (Xobs,i |Xi , Di , θ)Pj (Xi |Di , θ) dXi (9)
ing all our parameters that characterize the detailed re- j
lationships in our model into the vector θ, this gives us
the total (combined) log-likelihood: This gives us a tractable form of the likelihood to eval-
uate, where:
n
X
ln Ptot (θ) = ln P (Xobs,i |Xi , Yi , Di , θ). (6) • P (Xobs,i |Xi , Di , θ) is the noise distribution of the
i=1 observed colors given the true colors, which we as-
The individual likelihood for each object can be written sume follow a Normal distribution with unknown
as a sum of the class-dependent log-likelihoods over j = mean and known measurement uncertainties σobs .
{BHB, BS, Outlier}: • Pj (Xi |Di , θ) describes the intrinsic color-color re-
lationships for BHB and BS stars as well as out-
P (Xobs,i |Xi , Yi , Di , θ) = liers, which we take to also be Normal distributions
1(Yi = j)Pj (Xobs,i |Xi , Di , θ)
X
(7) centered around some mean ridgeline µj with some
j intrinsic scatter σint (for BHB and BS classes) or
Uniform within some color boundaries (for out-
where 1(Yi = j) is the indicator function which gives 1 liers).
when Yi = j and 0 otherwise. Pj (·) is the probability
distribution for X within each class (either Normal or • The integral over all possible intrinsic colors Xi is
Uniform). the explicit way in which we marginalize over the
Of course, in practice the particular class Yi for object unknown true colors for a given object.
i is unknown. Therefore, we marginalize over all possible
classes, which allows us to replace our indicator function 3.2. Detailed Model Implementation
with the expected fraction of objects in each class: To determine the final functional form for our model,
we fit individual models over objects from distinct mag-
P (Xobs,i |Xi , Di , θ) = nitudes bins over g0 = (16, 21) with a width of 0.5 mag,
X as well as across the entire g0 magnitude range. Af-
fj (Di ) Pj (Xobs,i |Xi , Di , θ) (8)
j
ter exploring a number of different models with varying
complexity, we find a fixed third-order polynomial can
In addition, the true color Xi of a given object is not accurately describe the ridgeline for the BHB and BS
known. Thus, We must also marginalize over all possible populations at all magnitudes. We take this to be a
6 Yu, Li, Speagle et al.

model for (i − z)0 color as a function of (g − r)0 color, 1D and 2D marginal posterior distributions of the 14
giving us four parameters each: parameters in our photometric mixture model. There is
some covariance among a3 , a2 , a1 , a0 . As those parame-
3
X ters collectively describe the position of the BHB ridge-
µBHB ≡ (i − z)0,BHB = ak [(g − r)0 ]k (10)
line, they are expected to be dependent on each other.
k=0
3
This holds true for parameters b3 , b2 , b1 , b0 as well. In
addition, the coefficients of the polynomial that models
X
µBS ≡ (i − z)0,BS = bk [(g − r)0 ]k (11)
k=0 the outlier ratio should also be dependent on each other,
corresponding to the observation of covariance in c1 , c0 .
We also confirm that the data are fully consistent with This applies to the pair d1 , d0 as well. Other than those
constant intrinsic scatter terms σBHB and σBS across all pairs, the rest of the parameters should not be depen-
magnitudes, giving us two additional parameters. We dent on each other, and indeed there is no evidence of
find the probability of having a source belong to each covariance shown in the plot.
group can be described using a linear function of g0 mag- Using the best-fit parameters in Table 1, the proba-
nitude, giving us four additional parameters: bility of each star being a BHB is calculated via

fout (g0 ) = c0 + c1 × g0 (12) fBHB LBHB

pBHB = (15)
fBHB (g0 ) = (1 − fout (g0 )) × (d0 + d1 × g0 ) (13) fBHB LBHB + fBS LBS + foutlier Loutlier

with the probability of being in the BS class determined where Lj is the corresponding likelihood described in
via: Section 3.1. pBS and poutlier are defined similarly.
fBS (g0 ) = 1 − fBHB (g0 ) − fout (g0 ) (14) In the top right of Figure 4, we show a visualiza-
tion of the predicted class probabilities of the stars used
This gives us a total of 14 unique parameters in our in our model in color-color space. Stars classified as
model to parametrize the distribution of BHB, BS, and BHBs, BSs, and outliers are shown in blue circles, pur-
outlier sources. ple squares, and red stars, respectively. The predicted
To marginalize over the observational uncertainties in class is defined to be the one that shows the highest
color, we solve the integral outlined in equation (9) in posterior-marginalized probability, and the color is used
two parts. First, we note that the distribution of (i−z)0 to represent the probability (one of pBHB , pBS , poutlier ).
color at a given (g − r)0 color is either Normal (BS The blue solid line and purple solid line show the ridge-
or BHB) or uniform (outlier). The integral of a prod- lines µBHB , µBS , calculated from the best-fit parameters
uct of two normal distributions is analytic and gives a in Table 1. The shaded regions with dashed lines on
new
q normal distribution with a standard deviation of their boundaries along each ridgeline show the intrin-
2
σ(i−z) 2
+ σBHB . This leaves us with just the integral
0 sic scatters σint for each sequence. Furthermore, the
over (g − r)0 . For each object, we define a grid of 50 predicted ratios of BHB stars (blue), BS stars (purple),
points in (g − r)0 evenly spaced between ±5σ(gr )0 , com- and outliers (red) as a function of magnitude are shown
pute the integral over (i − z)0 given each (g − r)0 , and in the middle right of Figure 4.
then compute the remaining 1-D integral using a simple Using the mixture model, we predict the probabilities
Riemann sum over the associated (g − r)0 likelihood. of being a BHB star from 16 mag to 21 mag. Figure 5
For the uniform case, the integral always gives a con- shows stars from selected magnitude bins. The model
stant value that does not depend on the input parame- distinguishes BHB versus BS easily at brighter magni-
ters (and therefore can simply be ignored). tudes (g0 < 19) with a clear decision boundary. At
We assume broad, uniform priors on all parameters fainter magnitudes, we see there are substantially fewer
as listed in Table 1. We sample from posterior using BHBs, because of the larger uncertainties and decreas-
the dynesty Nested Sampling package (Speagle 2020; ing ratio of BHB versus BS at these fainter magnitudes.
Koposov et al. 2023) version 1.2.3 in Dynamic Nested
Sampling mode (Skilling 2004, 2006; Higson et al. 2019) 3.4. Validation from cross-matching with S 5 Data
under default settings that used multi-ellipsoidal bounds We apply our mixture model to S 5 x DES DR2 data to
and uniform sampling (Feroz et al. 2009). compute the BHB probability and use a threshold cutoff
on the probability of pBHB ≥ 0.5 to distinguish between
3.3. Results BHB and BS stars. The resulting predictions are shown
The result of the sampling process is shown in the bot- in the bottom right panel of Figure 1. It matches the
tom left panel of Figure 4. The corner plot shows the prediction from the spectroscopic parameters in S 5 well
 Mapping the Galactic Halo with BHBs 7

Table 1. Uniform prior range, best-sampled results with the maximum likelihood, and uncertainties of the 14 parameters in
the photometric mixture model.
Parameter Uniform Prior Range Best Fita 95% Credible Interval
a3 (−0.2, 0.2) −0.0088 (−0.0019, 0.0020)
a2 (−1.0, 1.0) 0.438 (−0.037, 0.037)
a1 (−10, 10) 2.68 (−0.21, 0.21)
a0 (−25, 25) 6.88 (−0.36, 0.36)
σBHB (0, 0.015) 0.0049 (−0.0001, 0.0001)
b3 (−0.2, 0.2) −0.0512 (−0.0018, 0.0018)
b2 (−1.0, 1.0) 0.357 (−0.037, 0.037)
b1 (−10, 10) 1.54 (−0.22, 0.22)
b0 (−25, 25) 3.74 (−0.40, 0.39)
σBS (0, 0.015) 0.0107 (−0.0002, 0.0002)
c1 (−3.5, 3.5) 0.149 (−0.015, 0.017)
c0 (−0.6, 0.6) -0.0071 (−0.0008, 0.0007)
d1 (−3.5, 3.5) 3.022 (−0.070, 0.072)
d0 (−0.6, 0.6) -0.1450 (−0.0037, 0.0036)
a Samples with the maximum likelihood

(see the bottom left panel), with recall 0.962, precision NGC 300 as there is an excess of blue stars in the DES
0.953, and false positive rate 0.047, suggesting the ro- DR2 photometry in this region. For Fornax and Sculp-
bust performance of our model. tor, we use a 3-deg radius cut centered at (RA, Dec) =
Using the predicted BHB probabilities, we select the (39.9583◦ , −34.4997◦ ) and (15.0183◦ , −33.7186◦ ), re-
BHB candidates from DES DR2 and use them for subse- spectively. For the Sgr stream, we first convert the
quent analysis to derive the density profile of the Milky Galactic coordinate system to the heliocentric spherical
Way. coordinate system defined by the orbit of the Sagittarius
dwarf galaxy (Majewski et al. 2003), with Sgr longitude-
4. DENSITY PROFILE like angle Λ and Sgr latitude-like angle B. Stars within
4.1. Calibration to Distance and Selection of BHBs the region |B| ≤ 20◦ are removed. For the LMC, we re-
move stars within the region Dec ≤ −50◦ , 50◦ ≤ RA ≤
We enforce a cut pBHB ≥ 0.5 to select the BHB stars.
150◦ . For NGC 300, we use a 0.5-deg radius cut centered
The distances of BHBs can be calibrated in a straight-
at (RA, Dec) = (13.7229◦ , −37.6844◦ ).
forward manner. We employ the relations between ab-
We show the radial distribution of the selected BHB
solute magnitude Mg and (g − r)0 color from Belokurov
stars in Figure 6, both before and after the removal of
& Koposov (2016),
these overdensities. In order of increasing heliocentric
distance, overdensities associated with the Sgr stream,
Mg = 0.398 − 0.392(g − r)0 + 2.729((g − r)0 )2 the LMC, and Sculptor are visible in the full sample. We
+ 29.1128((g − r)0 )3 + 113.569((g − r)0 )4 . (16) also present a view of the on-sky location of our BHB
candidates, before and after removal of overdensities,
See also a similar relation from Deason et al. (2011b). in Figure 7. We note that the removal of overdensities
We then use g0 − Mg to obtain the distance modulus, is not exhaustive, and there are numerous stars from
which is subsequently converted to heliocentric distance substructures which remain in the sample. We discuss
(d). these further in section 5.
To characterize the underlying density profile of the To establish an accurate density profile, it is important
halo itself, we need to remove stars associated with sub- to account for potential partial selection effects for the
structures that will likely result in overdensities. Known BHB sample, especially at the faint and bright extremes.
substructures in the Milky Way halo which are in the We assume that the BHB sample in DES is 100% com-
vicinity of our sample include Sculptor, Fornax, the plete in the magnitude range of 16 < g0 < 20.5. Since
Sgr stream, the Large Magellanic Cloud (LMC). Ac- the absolute magnitude of BHBs Mg is a function of g−r
cordingly, stars in regions close to these substructures color, we compute the maximum distance for a BHB at
are excluded from our density profile study. In addi- gmin = 16 in the color range of −0.35 < (g −r)0 < −0.05
tion, we also exclude the region near an external galaxy
8 Yu, Li, Speagle et al.

Figure 4. Bottom left: A corner plot depicting the 2D and 1-D marginal posterior probability distributions of the 14 parameters
of our photometric mixture model estimated with dynesty. All values were multiplied by a factor of 100 for visualization
purposes. The parameters model different aspects of the data, as noted in red at the bottom. The parameter BHB fraction*
describes the ratio of BHBs to non-outliers. Top right: Stars classified as BHBs, BSs, and outliers are shown in blue circles,
purple squares, and red stars, respectively. The predicted class is defined to be the one that shows the highest posterior-
marginalized probability, and the color is used to represent the probability (one of pBHB , pBS , poutlier ). The blue solid line and
purple solid line show the ridgelines µBHB , µBS , calculated from the best-fit parameters in Table 1. The shaded regions with
dashed lines on their boundaries along each ridgeline show the intrinsic scatters σint for each sequence. Middle right: The
predicted ratio of BHB (blue), BS (purple), and outliers (red) as a function of magnitude. Overall, our model has parameters
that are well-constrained, accurately trace the photometric distribution of stars, successfully identify photometric outliers, and
generate physically-sensible magnitude dependencies across each subgroup.

as the lower limit of our complete sample, dmin = 13.05 kpc. Similarly, We use the minimum distance for a BHB
 Mapping the Galactic Halo with BHBs 9
0.05 1.0
0.00
0.05
(i z)0

0.10
0.15 0.8
0.20
16 g0 17 17 g0 18 18 g0 19
0.25
0.05
0.00 0.6
0.05
(i z)0

0.10

pBHB
0.15
0.20 0.4
19 g0 19.5 19.5 g0 20 20 g0 20.25
0.25
0.05
0.00
0.05 0.2
(i z)0

0.10
0.15
0.20
20.25 g0 20.5 20.5 g0 20.75 20.75 g0 21
0.25 0.0
0.3 0.2 0.1 0.3 0.2 0.1 0.3 0.2 0.1
(g r)0 (g r)0 (g r)0
Figure 5. As Figure 3, but now highlighting the predicted BHB probability pBHB for individual stars as a function of g0
magnitude. The model easily classifies BHBs at brighter magnitudes (g0 < 19). The two sequences start to overlap at
19 < g0 < 20. At the faintest magnitudes, the changing relative fraction of BHB versus BS stars, in addition to larger
photometric uncertainties, leads to a more ambiguous classification.

at gmax = 20.5 in the same color range as the upper limit low we present the details of modeling the stellar den-
of our complete sample, dmax = 68.37 kpc. We consider sity profile based on the selection effects, and we model
our sample to be complete between these lower and up- each star using an inhomogeneous Poisson point process
per limits, and thus model the density profile only across (IPPP). This method has been successfully employed in
this region. We additionally exclude stars with Galacto- previous studies conducted by Bovy et al. (2012), Xue
centric distances R < 20 kpc, as we notice those stars are et al. (2015), and Han et al. (2022), among others.
likely highly incomplete, and do not model the density We define the Poisson intensity function in the Galac-
profile in this region. We obtain a total of 2103 BHBs af- tic coordinate system (l, b, d). The rate of finding a star
ter these selections, and we show histograms (blue) with can be written as
respect to heliocentric distance (d) and Galactocentric
distances (R) in the left and right panels of Figure 8
λ(l, b, d) = A × (R + Rsmooth )−α ×
respectively.
|J(l, b, d|R)| × Sdistance (d, R) × Ssky (l, b) (17)
4.2. Inhomogeneous Poisson Point Process Model
As the selected BHB candidates are influenced by where
1) the spatial coverage of the DES survey and 2) the
limits on Galactocentric distances (R), they are not a • R
p is the Galactocentric radius, defined by R =
true representation of the underlying stellar distribu- R02 + d2 − 2R0 d · cos(l) · cos(d) and R0 is the
tions. Hence when fitting the expected stellar distribu- distance from the Sun to the Galactic Center,
tion, we must consider these two selection effects. Be- taken to be 8.3 kpc (Gillessen et al. 2017).
 10 Yu, Li, Speagle et al.

Before After
RA RA
30° 0° 30° 0° 30

Galactic Latitude (degree)

60° 60° 40
-30° -30°
50

90° 90° 60
70
-60° -60°
80
80 100 80 100
0 20 40 st60ance (kpc) 0 20 40 st60ance (kpc)
Heliocentric di Heliocentric di

Figure 6. Radial distribution of BHB candidates (with pBHB ≥ 0.5 computed by our model) selected from DES DR2. Left: all
candidates. Right: after removing regions associated with known overdensities in Section 4. Color represents Galactic latitude b.
In order of increasing heliocentric distance, our selected BHBs correctly show overdensities associated with known substructures:
the Sgr stream, the LMC, and Sculptor. The removal is not exhaustive, and there are many other substructures left in the
sample (see Figure 13).

10 Before 10 After 80
0 0 70

heliocentric distance (kpc)

10 10 60
Declination ( )

20 20
50
30 30
40
40 40
30
50 50
60 60 20
70 70 10
50 0 50 50 0 50
Right Ascension ( ) Right Ascension ( )
Figure 7. On-sky distribution of BHB candidates, for the same samples as in Figure 6. Left: before removing substructures.
Right: after removing substructures. Color represents heliocentric distances. The removal of regions associated with LMC,
NGC300, Fornax, Sculptor, and Sgr are clearly shown.

• Rsmooth is a smoothing radius to allow for integra- • Sdistance (d, R) and Ssky (l, b) are the distance selec-
tion down to 0, and it is set to 1 kpc. tion function and spatial selection function based
on the DES footprint, respectively.

• |J(l, b, d|R)| is the Jacobian term to account

for coordinate transformation and |J(l, b, d|R)| =
• α is the slope for the power law. We do not look d2 cos(b)
for a power law with a breaking radius near 20
kpc as this is close to our lower limit in distance • A is a normalization coefficient.
(Medina et al. 2018; Watkins et al. 2009).
 Mapping the Galactic Halo with BHBs 11

350 500
Data Data
300 Best Fit Best Fit
400
250
200 300
N(BHB)

N(BHB)
150 200
100
100
50
0 0
20 30 40 50 60 70 20 30 40 50 60 70
d (kpc) R (kpc)
Figure 8. Histograms of final selected BHB candidates in DES DR2 versus simulated samples from the best-fit density profile.
Left: with heliocentric distance d. Right: with Galactic radius R. We see that our best-fit density profile accurately captures
the stellar density distribution, generating samples that match the data well from both heliocentric and Galactocentric points
of view.
We define our selection functions to be simple binary Table 2. Uniform prior range, samples that has the maxi-
indicators as a function of R, d, l, and b, with mum likelihood (best-fit), and uncertainties of the 2 param-
 eters in the density model.
1 if 20 kpc ≤ R and d
min < d < dmax
Sdistance (d, R) = Prior Best 68% Credible 95% Credible
0 else
Parameter Range Fit Interval Interval
(18)
α (2, 7) 4.28 (−0.07, 0.07) (−0.12,0.13)
and
log A (1,16) 10.97 (−0.25, 0.24) (−0.44, 0.47)

1

 if − 60◦ ≤ RA ≤ 100◦
Ssky (l, b) = S(RA, Dec) = and − 70◦ ≤ Dec ≤ 10◦


 over grids of l, b, d. To speed up the computation, we
0 else specify a grid of α values (Λ is a function of α) and
(19) pre-computed the integrals at each α. We then use
which approximately constrains us to within the cover- interpolation from scipy (Virtanen et al. 2020) on the
age of the DES footprint. Intuitively, λ · (dl · db · dd) α grid such that during the actual sampling process the
can be interpreted as the infinitesimal probability of a integral can be easily obtained by inputting α. We again
star existing in an infinitesimal volume region (dl, db, dd) assume uniform priors on the two parameters as listed
located at position (l, b, d). in Table 2. Similarly to Section 3, we sample from the
With the rate parameter λ, the log-likelihood of an posterior using the dynesty Nested Sampling package
IPPP can be written as with the default settings.
X
log L = log L(l, b, d|ϕ)
X
= −Λ + log λ(l, b, d|ϕ) (20)
4.3. Results
where ϕ represents the set of parameters α and A. Λ is We find α that maximizes the likelihood and its 95%
the integral of λ, and (2σ) credible interval to be 4.28+0.13
−0.12 and normalization
π
factor A to be 10.97+0.47
−0.44 . In Figure 9, the corner plot
Z dmax Z 2π Z 2
Λ=A d2 cos(b) (R + Rsmooth )−α × shows the 2D and 1D marginal posterior probability dis-
dmin 0 −π
2
tributions of the two parameters of our density profile es-
Sdistance (d, R) Ssky (l, b) db dl dd (21) timated with dynesty. We see that the power-law index
α and normalization coefficient A are well-constrained
As the integral Λ is analytically intractable, we ap- and positively correlated. This is expected given the
proximate it using the trapezoidal rule by integrating constant stellar count N (BHB).
12 Yu, Li, Speagle et al.

= 4.28+0.13
0.12 artificial contamination to our BHB sample. Although
a threshold of 0.5 appears to yield high accuracy for
bright stars (shown by testing on S5 × DES DR2 in
Figure 1), it is likely to overestimate or underestimate
the star count in the fainter range due to the increased
uncertainty. Thus, it is essential to investigate the sys-
tematic error as a result of applying different threshold
values.
A = 10.97+0.47
0.44
To address this, we generate 28 evenly-spaced (spac-
ing=0.25) probability threshold values ranging from
.0

pthreshold = 0.3 to 1.0. Under each threshold, we se-

12

lect BHB candidates and conduct the same analysis in

.5
11

Section 4 to estimate A and α. The results are presented

A

.0
11

in Figure 10, with the BHB count and estimate of the

power law index α displayed in the left and right panels,
.5
10

respectively. The normalization coefficient A is com-

.0
10

pletely degenerate with α (they collectively define the

5

.0
.5
.0
.5
.0
4.0

4.2

4.3

4.5

number of stars) and is thus omitted. As the threshold

10
10
11
11
12
A increases, the power-law index increases. This matches
our expectation since a stricter threshold cutoff will re-
Figure 9. A corner plot showing the marginal posterior sult in fewer BHBs selected in the fainter magnitude,
probability distributions of the two parameters of our density
corresponding to a steeper decline in the density, and
profile estimated with dynesty. We see that the power-law
index α and normalization coefficient A are well-constrained, equivalently, a larger power law index.
and they are correctly positively correlated since the total Notably, the plot reveals some small jumps in BHB
stellar count is constant. count and power-law index as the threshold becomes
stricter (at pthreshold ∼ 0.9). We hypothesize this is
4.4. Mock Data likely due to the local overdensities in the halo, so that
a specific threshold cutoff might accidentally remove a
As a verification, we generate mock data from our
group of stars. To visualize this, we calculate the aver-
best-fit parameters (that maximize the likelihood) and
age BHB probability with respect to g0 magnitude (as a
compare the predicted stellar distribution with the ac-
proxy for heliocentric distance) and Galactocentric dis-
tual distribution. To generate the predicted stellar dis-
tance R. After selecting stars with pBHB ≥ 0.3, which
tribution, we construct a grid of l, b, d values. For each
is the lowest threshold in Figure 10, we separate these
point (l, b, d), we compute the density function λ with
candidates into different groups based on both g0 band
the two selection functions S, and record the distances
photometry and Galactocentric distance R. Within each
d, R. Then, we use the trapezoidal rule over the grid
group, the average BHB probability is computed, shown
to compute the integral of densities λ. We normal-
as the value in each 2D bin in Figure 11. We see that the
ize this integral so it is equal to the total number of
average BHB probability generally decreases as distance
BHBs. After normalization, the previously computed
increases, or when the magnitude becomes larger. How-
density λ at different distances d, R will approximate
ever, it is worth noting that the trend is not smooth with
the predicted stellar count. In Figure 8, we compare the
respect to magnitude and distance. There are some bins
predicted stellar count (shown in red) to the observed
that have higher average BHB probabilities than their
distribution (shown in blue) for both the heliocentric
neighbours. A hard probability threshold cutoff may ac-
distances d (left panel) and Galactocentric distances R
cidentally remove these small overdensities, resulting in
(right panel). The distributions show good agreement,
the bumps seen in Figure 10.
indicating that our model captures the stellar density
The variations in the threshold value exert a non-
distribution of the data well.
negligible influence on the fitted parameters. This ex-
5. DISCUSSION periment assesses the systematic error and demonstrates
5.1. Systematic Error of Hard-Thresholding that depending on the method, the true power-law index
should fall between 4.2 to 4.5.
During the selection process of BHBs, we use a
Incorporating the probabilities as weights in the den-
hard probability threshold cutoff to select sources with
sity model fitting process could potentially enhance the
pBHB ≥ 0.5. Hard thresholding is likely to introduce
 Mapping the Galactic Halo with BHBs 13

2150 p=0.5
4.8 weighted
2100
4.7
2050 4.6
4.5
N(BHB)

2000
4.4
1950
4.3
1900 4.2
p=0.5
1850 P 4.1
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 P 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 weighted
Probability Threshold Probability Threshold
Figure 10. Left: BHB candidate count given different probability thresholds. Right: best-fit power-law index α for BHBs
samples corresponding to the differing probability thresholds p. The error bar shows 95% (2σ) credible interval.The normalization
coefficient A is completely degenerate with α and is thus omitted. The criterion we used for the previous section pBHB ≥ 0.5
is highlighted. As the selection becomes stricter, BHB count decreases and we observe a steeper descent. We note that the
trends in both panels are not smooth, with small bumps discernible, which is explained in Figure 11. The orange star shows the
best-fit α where we weigh the log-likelihood of each star by its BHB probability, instead of using the unweighted log-likelihood
of BHB candidates selected from hard-thresholding (see text for details).
20.0 To compare with N (BHB) from the thresholding ap-
19.5 0.9 proach, we sum up the BHB probabilities over all the
sources (that possibly include BHBs, BSs, and outliers),
19.0 0.8
Average Probability

shown as the orange star in the left panel of Figure

18.5 10. To compare with α, we use these BHB probabil-
0.7
g0 (mag)

ities to define a weighted log-likelihood function. The

18.0 0.6 best-fit α that maximizes this likelihood is shown in or-
17.5 ange in the right panel of Figure 10. The weighted log-
0.5 likelihood function multiplies the log-likelihood log λ of
17.0 each source by its BHB probability. Using this approach,
0.4 both the summed probability and power law index seem
16.5
to align with N(BHB) and α from the 0.5 threshold cut-
30 40 50 60 70 off. However, for the weighted log-likelihood approach,
Galactocentric Distance (kpc)
after we remove the log in the likelihood, multiplying by
Figure 11. Averaged BHB probability with respect to g0 the weight effectively exponentiates the likelihood value
band photometry and Galactocentric distance. Stars are se- of each star by its respective BHB probability, which
lected with pBHB ≥ 0.3 (same as the first threshold in Figure lacks a clear physical meaning. Therefore, we include it
10. Then the selected candidates are separated into different
for reference purposes but do not use it in our analysis.
2D bins based on g0 and Galactocentric distance R. Within
each bin, the average BHB probability is computed over the
stars inside that bin, shown by color. We observe that BHB 5.2. Sky Variation
probability generally decreases as distance increases, or when
the magnitude becomes larger. However, it is worth noting The stellar distribution can be influenced by the par-
that the trend is not smooth with respect to magnitude and ticular region of the sky under investigation. To un-
distance. Some bins appear to have higher BHB probabilities derstand the impact of different regions of the sky on
than their neighbours. A hard probability threshold cutoff the resulting density profile, we partition the DES foot-
may accidentally remove a group of stars in those bins, re- print into three distinct patches, as illustrated in the
sulting in the bumps we see in Figure 10. left panel of Figure 12. We apply the same analysis in
Section 4 to each patch and find the power law indexes
results. To this end, we explore another method where of those patches range from 4.05 to 5.12, presented in
we drop the threshold cutoff completely and use the the right panel of Figure 12. The variation of power
probabilities as weights in the log-likelihood for the sam- law indexes among different regions suggests that the
pling process instead. Milky Way halo is spatially inhomogeneous, and that it
14 Yu, Li, Speagle et al.

Figure 12. Left: We split DES coverage into 3 patches, shown by the dashed lines. Iso-latitude lines at b=−80◦ , −60◦ , −40◦
are shown in grey. Right: best-fit parameter α with 95% (2σ) credible interval of the BHB density profile within each patch.
The dashed line and shaded region show the best-fit parameter for the entire DES footprint. We see that α varies considerably
among different sky patches, which reflects the spatial inhomogeneity of the Milky Way halo and illustrates the importance of
considering the location and coverage of a survey when fitting a density profile.

is important to consider the location and coverage of a power law index is consistent with previous findings at
survey when fitting a density profile. In Section 5.5, we the same Galactocentric distances R.
further discuss the implication of the sky variation on a Our investigations on different threshold cutoffs and
flattened or triaxial halo density profile and suggestions different regions of the sky provide a good explanation
for future investigations. for the variability in the existing literature. Models with
different probability threshold cutoffs are analogous to
5.3. Search for Substructures the different methods used by other researchers, and we
It is worth highlighting that the removal of substruc- explain the variation for our model quantitatively by
tures is not exhaustive during the data processing phase. assessing the systematic error. The experiment on the
As an example, when we select stars at specific distances different sky patches illustrates the dependence of the
of stellar streams discovered Shipp et al. (2018), we are fitted density profile on the location and coverage of dif-
able to recover the presence of several previously dis- ferent surveys.
covered thinner streams, including Turranburra, Willka
Yaku, Chenab, and Elqui, shown in Figure 13. Our 5.5. Limitations and Future Studies
catalog thus assumes importance to complement the ex- There are a number of ways we can expand upon the
isting stellar stream catalog and proffer candidates for current study. Firstly, the analysis is conducted using
new streams. stars within a distance range of 20 kpc to 70 kpc due to
In addition, more diffused and massive streams like the conservative selection of high-probable stars. With
Palca/Cetus can be vaguely seen in Figure 6 and 7 future surveys with higher depth and better precision,
around RA ∼ 30◦ . Inevitably, these substructures may we can map out more distant regions of the halo.
also have some impact on the halo density measurement, Secondly, although our density model incorporates the
which may also impact the sky variation discussed in fundamental characteristics of the stellar halo, it is lim-
Section 5.2. However, given the large number of streams ited by assuming a perfect spheroidal shape centered at
in the survey area, it is not possible to remove all stream the Galactic center, with an identical decline in all di-
members. rections. However, the literature suggests that the true
shape (and orientation) of the halo is more complex.
5.4. Comparison with Existing Literature Several previous studies have claimed, for instance, that
We compare our estimated density profile power-law the halo is oblate (Olling & Merrifield 2000; Sesar et al.
index α with existing literature in Figure 14, including 2011; Deason et al. 2011; Bowden et al. 2015; Thomas
the measurements using photometric data from the DES et al. 2018), prolate (Helmi 2004; Banerjee & Jog 2011;
(thus the same sky coverage), but with different tracers Bowden et al. 2016; Fukushima et al. 2018), and triaxial
- RR Lyrae stars (Stringer et al. 2021) and main se- (Law & Majewski 2010; Deg & Widrow 2013; Iorio &
quence turn-off (MSTO) stars (Pieres et al. 2020). Our Belokurov 2019), although spherical distributions (Fell-
 Mapping the Galactic Halo with BHBs 15

Figure 13. Similar to Figure 7, but showing only BHB candidates within certain Galactocentric distances, as shown by the
range of each colorbar. Moving from the top left to the bottom right, we present examples of four streams: Turranburra, Willka
Yaku, Chenab, and Elqui, which are shown in red. Our sample displays overdensities that align closely with these previously
identified substructures. This alignment demonstrates the consistency of our catalog with prior discoveries and underscores its
potential utility in future searches for additional substructures.

hauer et al. 2006; Smith et al. 2009; Das et al. 2023; 5.5.1. Parametrization of the BHB/BS Mixture Model
Palau & Miralda-Escudé 2023) and models where the
halo flattening varies with Galactocentric radius or line In Section 3 we parametrize our model that the ratio
of sight (e.g., Vera-Ciro & Helmi 2013; Hernitschek et al. of BHB and BS is purely a function of g0 . However,
2018) have also been considered. Additionally, it has given that BHB and BS sequences represent two stellar
been proposed that the halo likely exhibits a misalign- populations with distinct evolutionary trajectories and
ment with respect to the galactic disk (e.g. Han et al. distributions in the MW, this model might be too sim-
2022), as a result of its accretion history (e.g., Prada plified.
et al. 2019; Dillamore et al. 2022). Furthermore, the First, the ratio of the two populations is not expected
halo density of the studied region may also be impacted to be constant as a function of (g − r)0 . For this, we ex-
by the wake induced by the infall of LMC (Belokurov periment with a parametric form involving both (g − r)0
et al. 2019; Conroy et al. 2021). Indeed, the variation in and g0 in the Appendix A. As detailed in the Appendix,
α seen in Figure 12 and Section 5.2 may already imply the resulting number of BHB stars as well as the power
that the density profile is more complicated than spher- law index α with the new model shows little difference
ical, as significantly different declines in stellar densities compared to the simpler model (i.e. without (g − r)0
are observed at similar Galactic latitudes. For future dependency).
work, it would be advantageous to consider more flexi- Moreover, BHB stars are intrinsically brighter than
ble models that can accommodate such complexities. BS stars, and thus locate much further than BS stars at
the same apparent magnitude. Hence we also expect to
16 Yu, Li, Speagle et al.

5.5 our work

Watkins_2009
Deason_2011
5.0 Xue_2015
Fukushima_2018
4.5 Lorio_2018
Medina_2018
4.0 Thomas_2018
Starkenburg_2019
Pieres_2020
3.5 Stringer_2021
Han_2022
3.0

2.5
BHB
2.0 RRL
BHB,RRL,K-giant
MSTO
1.5
0 20 40 60 80 100
R (kpc)
Figure 14. A comparison between our power-law index α (depicted in red) and values reported in existing literature (Watkins
et al. 2009; Deason et al. 2011; Xue et al. 2015; Fukushima et al. 2018; Iorio et al. 2018; Medina et al. 2018; Thomas et al.
2018; Starkenburg et al. 2019; Pieres et al. 2020; Stringer et al. 2021; Han et al. 2022). The α values are plotted as a function
of Galactocentric radius R, with shaded regions representing 68% (1σ) credible interval, taken from corresponding literature
when applicable. The lines are categorized based on the types of stars being analyzed. Within the distance range covered by
our sample, our best-fit α closely matches the findings of previous studies.

see a wide range of observed BHB to BS ratios across BHB and BS sequences through precise griz photom-
different lines of sight, in particular, as a function of etry. We identify ∼ 2100 highly probable BHB candi-
Galactic latitude where BS is likely dominated by the dates in the Southern Hemisphere, and investigate the
disk stars. However, since our model considers the sky as stellar halo within a distance range of 20 to 70 kpc. After
a whole, averaging out these variations, the differences in excluding stars in the area associated with major known
ratios along various lines of sight become less significant. substructures, we observe a smooth decline in the stellar
For future research, it would be beneficial to model them density, with a power law index α = 4.28+0.13 −0.12 , consis-
as separate populations and parameterize densities of tent with existing literature values. By drawing connec-
BHB and BS as functions of both distance and location tions to our assessment of systematic error in threshold
on the sky (in Galactic latitude and longitude). cutoffs and sky locations, we argue that the variations
in current power law indexes in the literature can be
6. CONCLUSIONS largely associated with (a) different methodologies used
In this study, we have developed a mixture model that to derive the density profiles, and (b) the inherent spa-
predicts the probability of a star being a BHB based on tial inhomogeneity of the halo. We provide the entire
its g0 band magnitude, (g − r)0 and (i − z)0 photometry catalog (which contains 46031 sources) with computed
using DES DR2. Our study demonstrates that, even in pBHB using our model in Appendix B. We hope this
the absence of the u−band, we can distinguish between
 Mapping the Galactic Halo with BHBs 17

catalog will be useful for future research on the Galactic This paper made use of the Whole Sky Database
halo. (wsdb) created by Sergey Koposov and maintained at
To gain a more comprehensive understanding of the the Institute of Astronomy, Cambridge by Sergey Ko-
stellar halo, future studies should focus on using more posov, Vasily Belokurov and Wyn Evans with financial
flexible models and probing more distant stars in the support from the Science & Technology Facilities Coun-
halo. Such a BHB sample would not only help us study cil (STFC) and the European Research Council (ERC).
the stellar density profile of the Milky Way halo, but This project used public archival data from the Dark
would also help identify old substructures such as stellar Energy Survey (DES). Funding for the DES Projects
streams. The latter will contribute to our understanding has been provided by the U.S. Department of Energy,
of the Milky Way’s accretion history. the U.S. National Science Foundation, the Ministry of
With forthcoming photometric surveys, like the Ru- Science and Education of Spain, the Science and Tech-
bin Observatory’s Legacy Survey of Space and Time nology Facilities Council of the United Kingdom, the
(LSST), we anticipate not only extending our reach to Higher Education Funding Council for England, the Na-
greater distances but also expanding our observational tional Center for Supercomputing Applications at the
scope across a broader expanse of the sky. This will University of Illinois at Urbana-Champaign, the Kavli
enable us to investigate the density profile of the halo Institute of Cosmological Physics at the University of
out to the virial radius of the MW and explore potential Chicago, the Center for Cosmology and Astro-Particle
spatial inhomogeneity more comprehensively. Overall, Physics at the Ohio State University, the Mitchell Insti-
this study provides insights into the properties of the tute for Fundamental Physics and Astronomy at Texas
stellar halo and sets the stage for future investigations A&M University, Financiadora de Estudos e Projetos,
that aim to unravel the complex formation and evolution Fundação Carlos Chagas Filho de Amparo à Pesquisa
processes of our Galaxy. do Estado do Rio de Janeiro, Conselho Nacional de
Desenvolvimento Cientı́fico e Tecnológico and the Min-
istério da Ciência, Tecnologia e Inovação, the Deutsche
ACKNOWLEDGMENTS Forschungsgemeinschaft, and the Collaborating Insti-
The authors would like to thank the Summer Un- tutions in the Dark Energy Survey. The Collaborat-
dergraduate Data Science (SUDS) Opportunities Pro- ing Institutions are Argonne National Laboratory, the
gram at the Data Science Institute at the Univer- University of California at Santa Cruz, the University
sity of Toronto for providing the funding and oppor- of Cambridge, Centro de Investigaciones Energéticas,
tunity to enable this project. T.S.L. acknowledges fi- Medioambientales y Tecnológicas-Madrid, the Univer-
nancial support from Natural Sciences and Engineering sity of Chicago, University College London, the DES-
Research Council of Canada (NSERC) through grant Brazil Consortium, the University of Edinburgh, the
RGPIN-2022-04794. S.K. acknowledges support from Eidgenössische Technische Hochschule (ETH) Zürich,
Science & Technology Facilities Council (STFC) (grant Fermi National Accelerator Laboratory, the Univer-
ST/Y001001/1). G.L. acknowledges FAPESP (proc. sity of Illinois at Urbana-Champaign, the Institut de
2021/10429-0). G.M.E. acknowledges financial support Ciències de l’Espai (IEEC/CSIC), the Institut de Fı́sica
from NSERC through grant RGPIN-2020-04554. d’Altes Energies, Lawrence Berkeley National Labo-
ratory, the Ludwig-Maximilians Universität München
The authors would like to thank Alex Drlica-Wagner,
and the associated Excellence Cluster Universe, the
Andrew Pace, Adriano Pieres for their helpful com-
University of Michigan, the National Optical Astron-
ments. This paper includes data obtained with the
omy Observatory, the University of Nottingham, The
Anglo-Australian Telescope in Australia. We acknowl-
Ohio State University, the OzDES Membership Con-
edge the traditional owners of the land on which the
sortium, the University of Pennsylvania, the Univer-
AAT stands, the Gamilaraay people, and pay our re-
sity of Portsmouth, SLAC National Accelerator Labora-
spects to elders past and present.
tory, Stanford University, the University of Sussex, and
For the purpose of open access, the author has applied
Texas A&M University. Based in part on observations
a Creative Commons Attribution (CC BY) licence to
at Cerro Tololo Inter-American Observatory, National
any Author Accepted Manuscript version arising from
Optical Astronomy Observatory, which is operated by
this submission.
the Association of Universities for Research in Astron-
This research has made use of the SIMBAD database,
omy (AURA) under a cooperative agreement with the
operated at CDS, Strasbourg, France (Wenger et al.
National Science Foundation.
2000). This research has made use of NASA’s Astro-
physics Data System Bibliographic Services.
18 Yu, Li, Speagle et al.

The Legacy Surveys consist of three individual and The Legacy Survey team makes use of data products
complementary projects: the Dark Energy Camera from the Near-Earth Object Wide-field Infrared Sur-
Legacy Survey (DECaLS; NOAO Proposal ID # 2014B- vey Explorer (NEOWISE), which is a project of the Jet
0404; PIs: David Schlegel and Arjun Dey), the Beijing- Propulsion Laboratory/California Institute of Technol-
Arizona Sky Survey (BASS; NOAO Proposal ID # ogy. NEOWISE is funded by the National Aeronautics
2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the and Space Administration.
Mayall z-band Legacy Survey (MzLS; NOAO Proposal The Legacy Surveys imaging of the DESI footprint
ID # 2016A-0453; PI: Arjun Dey). DECaLS, BASS is supported by the Director, Office of Science, Office
and MzLS together include data obtained, respectively, of High Energy Physics of the U.S. Department of En-
at the Blanco telescope, Cerro Tololo Inter-American ergy under Contract No. DE-AC02-05CH1123, by the
Observatory, National Optical Astronomy Observatory National Energy Research Scientific Computing Center,
(NOAO); the Bok telescope, Steward Observatory, Uni- a DOE Office of Science User Facility under the same
versity of Arizona; and the Mayall telescope, Kitt contract; and by the U.S. National Science Foundation,
Peak National Observatory, NOAO. The Legacy Surveys Division of Astronomical Sciences under Contract No.
project is honored to be permitted to conduct astronom- AST-0950945 to NOAO.
ical research on Iolkam Du’ag (Kitt Peak), a mountain Software: numpy (van der Walt et al. 2011), scipy
with particular significance to the Tohono O’odham Na- (Virtanen et al. 2020), matplotlib (Hunter 2007),
tion. astropy (Astropy Collaboration et al. 2013, 2018, 2022),
NOAO is operated by the Association of Universities dynesty (Speagle 2020; Koposov et al. 2023),
for Research in Astronomy (AURA) under a cooperative
agreement with the National Science Foundation.

APPENDIX

A. BHB RATIO AS A FUNCTION OF BOTH We conduct the same sampling process to find the pa-
COLOR AND MAGNITUDE rameters that maximize the likelihood, shown in Figure
We explore a parametric form of BHB ratio defined 15. We also plot the probability of a source being each
with (g − r)0 and g0 . The statistical model for this of the group, shown in Figure 16. We then find the num-
parametric form is defined similarly as the previous (see ber of BHBs following the same processing discussed in
Section 3.1), except that we describe the probability of Section 4. The number of BHB is found to be 2111, and
having a source belong to the BHB group using a poly- the power law index α is 4.27+0.13
−0.13 , shown in Figure 17
nomial function of g0 magnitude and (g − r)0 color (vs and Figure 18, respectively. Both of them show little
equation 13, which is defined solely on g0 ). We keep difference from the previous fit.
the probability of having a source belong to the out-
lier group as a linear function of g0 . This gives us five
parameters:

fBHB (g0 ) = (1 − fout (g0 )) × [d0 + d1 × g0

+ d2 × (g − r)0 + d3 × g0 × (g − r)0 + d4 × (g − r)20 ]
(A1) B. BHB CATALOG
Table 3 shows a compiled catalog from Section 2.3,
with the probability of being in the BS class determined
which contains 46031 sources, with pBHB computed by
via:
our mixture model described in Section 3 and heliocen-
fBS (g0 ) = 1 − fBHB (g0 ) − fout (g0 ) (A2)
tric distance d computed assuming the sources are BHBs
Hence we have a total of 17 unique parameters in our using Equation 16. We also provide Coadd Object ID,
model to parametrize the distribution of BHB, BS, and RA, Dec, and griz photometry, obtained from DES DR2
outlier sources. for each source.

REFERENCES
Abbott, T. M. C., Abdalla, F. B., Allam, S., et al. 2018, Abbott, T. M. C., Adamów, M., Aguena, M., et al. 2021,
ApJS, 239, 18, doi: 10.3847/1538-4365/aae9f0 ApJS, 255, 20, doi: 10.3847/1538-4365/ac00b3
 Mapping the Galactic Halo with BHBs 19

Figure 15. Same as Figure 4, but with a more complex parametric form of the BHB ratio. Bottom left: A corner plot depicting
the 2D and 1-D marginal posterior probability distributions of the 17 parameters of our photometric mixture model estimated
with dynesty. All values are multiplied by a factor of 100 for visualization purposes. The parameters di s are defined according
to section A. Top right: The predicted class probabilities for the stars used to derive the model in color-color space. BHB (blue)
and BS (purple) sequences are shown as a solid line with the best-fit scatter as a light-shaded region with dashed lines defining
its boundary. Stars classified as BHBs, BSs, and outliers based on their posterior-marginalized class probabilities are shown in
blue circles, purple squares, and red stars respectively. The parameters are well-constrained, accurately trace the photometric
distribution of stars, and successfully identify photometric outliers.

Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., Astropy Collaboration, Price-Whelan, A. M., Sipőcz, B. M.,
et al. 2018, AJ, 156, 123, doi: 10.3847/1538-3881/aabc4f
et al. 2013, A&A, 558, A33,
Astropy Collaboration, Price-Whelan, A. M., Lim, P. L.,
doi: 10.1051/0004-6361/201322068
et al. 2022, ApJ, 935, 167, doi: 10.3847/1538-4357/ac7c74
20 Yu, Li, Speagle et al.
BHB BS Outliers

0.6 0.420
0.5 0.415
0.5
0.4 0.410
0.4 0.405

f(BHB)

f(out)
f(BS)
0.3 0.400
0.3
0.2 0.395
0.2 0.390
0.1
0.1 0.385
0.0 0.380
0.00 0.00 0.00
0.05 0.05 0.05
0.10 0.10 0.10
16 0.15 16 0.15 16 0.15
17 0.20 17 0.20 17 0.20
r)0

r)0

r)0
18 0.25 18 0.25 18 0.25
(g

(g

(g
g0 19 0.30 g0 19 0.30 g0 19 0.30
20 0.35 20 0.35 20 0.35
21 21 21

Figure 16. Same as bottom right of Figure 4, but with a more complex parametric form of the BHB ratio. The predicted
probability of a source in the BHB group, the BS group, and the outlier group, parametrized by equations defined in section A.
Our model generates physically-sensible magnitude and color dependencies across each subgroup.

350 500
Data Data
300 Best Fit Best Fit
400
250 N=2111

200 300
N(BHB)

N(BHB)

150 200
100
100
50
0 0
20 30 40 50 60 70 20 30 40 50 60 70
d (kpc) R (kpc)
Figure 17. Same as Figure 8, but with a more complex parametric form of the BHB ratio. Histograms of final selected BHB
candidates in DES DR2 versus simulated samples from the best-fit density profile. Left: with heliocentric distance d. Right:
with Galactic radius R. The number of BHBs is 2111, which is very close to 2103 in the previous fit.

Banerjee, A., & Jog, C. J. 2011, ApJL, 732, L8, Dark Energy Survey Collaboration, Abbott, T., Abdalla,
doi: 10.1088/2041-8205/732/1/L8 F. B., et al. 2016, MNRAS, 460, 1270,
Belokurov, V., Deason, A. J., Erkal, D., et al. 2019, doi: 10.1093/mnras/stw641
MNRAS, 488, L47, doi: 10.1093/mnrasl/slz101 Das, M., Ianjamasimanana, R., McGaugh, S. S.,
Belokurov, V., & Koposov, S. E. 2016, MNRAS, 456, 602, Schombert, J., & Dwarakanath, K. S. 2023, ApJL, 946,
doi: 10.1093/mnras/stv2688 L8, doi: 10.3847/2041-8213/acc10e
Bovy, J., Rix, H.-W., Liu, C., et al. 2012, The Astrophysical Deason, A. J., Belokurov, V., & Evans, N. W. 2011a,
Journal, 753, 148, doi: 10.1088/0004-637X/753/2/148 MNRAS, 416, 2903,
Bowden, A., Belokurov, V., & Evans, N. W. 2015, MNRAS, doi: 10.1111/j.1365-2966.2011.19237.x
449, 1391, doi: 10.1093/mnras/stv285 —. 2011b, MNRAS, 416, 2903,
Bowden, A., Evans, N. W., & Williams, A. A. 2016, doi: 10.1111/j.1365-2966.2011.19237.x
MNRAS, 460, 329, doi: 10.1093/mnras/stw994 Deason, A. J., Belokurov, V., & Evans, N. W. 2011,
Conroy, C., Naidu, R. P., Garavito-Camargo, N., et al. Monthly Notices of the Royal Astronomical Society, 416,
2021, Nature, 592, 534, doi: 10.1038/s41586-021-03385-7 2903, doi: 10.1111/j.1365-2966.2011.19237.x
Conroy, C., Bonaca, A., Cargile, P., et al. 2019, ApJ, 883, Deason, A. J., Belokurov, V., & Koposov, S. E. 2018, ApJ,
107, doi: 10.3847/1538-4357/ab38b8 852, 118, doi: 10.3847/1538-4357/aa9d19
 Mapping the Galactic Halo with BHBs 21

Table 3. A catalog with BHB probability pBHB computed from the mixture model described in Section 3 using griz photometry
from DES DR2. Distance is the heliocentric distance, computed assuming the sources are BHBs. A total of 46031 sources are
included. Only the first ten lines are shown. The complete table is available online in a machine readable format.
Coadd Object ID RA Dec g0 r0 i0 z0 distance pBHB
994953737 345.969096 -42.549333 20.89 21.08 21.26 21.35 118.9 0.00
999753919 346.598197 -42.483844 19.32 19.54 19.73 19.78 56.18 0.95
999777071 346.818066 -42.739140 19.98 20.14 20.3 20.33 79.45 0.97
995168807 348.283024 -43.375280 17.19 17.50 17.74 17.84 17.53 0.98
998800330 347.896323 -43.028549 17.66 17.89 18.09 18.19 25.66 0.00
1001308502 349.326354 -43.156899 17.41 17.73 17.93 18.07 18.91 0.00
998795299 348.005177 -42.971495 20.24 20.36 20.50 20.60 90.63 0.00
998786086 347.917693 -42.875529 18.55 18.69 18.85 18.93 41.12 0.00
998773023 347.836941 -42.732092 17.52 17.66 17.82 17.89 25.72 0.00
1001283704 348.961490 -42.916494 18.60 18.65 18.75 18.80 43.20 0.00

= 4.27+0.13
0.13
Feroz, F., Hobson, M. P., & Bridges, M. 2009, MNRAS,
398, 1601, doi: 10.1111/j.1365-2966.2009.14548.x
Flaugher, B., Diehl, H. T., Honscheid, K., et al. 2015, AJ,
150, 150, doi: 10.1088/0004-6256/150/5/150
Fukushima, T., Chiba, M., Homma, D., et al. 2018,
Publications of the Astronomical Society of Japan,
Volume 70, Issue 4, id.69, 70, 69,
doi: 10.1093/pasj/psy060
A = 10.94+0.46
0.46 Fukushima, T., Chiba, M., Tanaka, M., et al. 2019,
Publications of the Astronomical Society of Japan, 71,
.0 .5 .0 .5 .0

72, doi: 10.1093/pasj/psz052

10 10 11 11 12

Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al.

A

2023, A&A, 674, A1, doi: 10.1051/0004-6361/202243940

Gerhard, O. 2012, Proceedings of the International
Astronomical Union, 8, 211,
doi: 10.1017/S174392131300481X
5
0
5
0

.0
.5
.0
.5
.0
4.0
4.2
4.3
4.5

Gillessen, S., Plewa, P. M., Eisenhauer, F., et al. 2017, ApJ,

A
10
10
11
11
12

837, 30, doi: 10.3847/1538-4357/aa5c41

Han, J. J., Conroy, C., Johnson, B. D., et al. 2022, The
Figure 18. Same as Figure 9, but with a more complex
parametric form of the BHB ratio. A corner plot showing Astronomical Journal, 164, 249,
the marginal posterior probability distributions of the two doi: 10.3847/1538-3881/ac97e9
parameters of our density profile estimated with dynesty. Helmi, A. 2004, MNRAS, 351, 643,
We see that the power-law index α = 4.27+0.13
−0.13 , very close doi: 10.1111/j.1365-2966.2004.07812.x
to the previous fit α = 4.28+0.13
−0.12 .
Helmi, A. 2008, The Astronomy and Astrophysics Review,
15, 145, doi: 10.1007/s00159-008-0009-6
Deg, N., & Widrow, L. 2013, MNRAS, 428, 912, Helmi, A. 2020, ARA&A, 58, 205,
doi: 10.1093/mnras/sts089 doi: 10.1146/annurev-astro-032620-021917
Dey, A., Schlegel, D. J., Lang, D., et al. 2019, AJ, 157, 168, Hernitschek, N., Cohen, J. G., Rix, H.-W., et al. 2018, The
doi: 10.3847/1538-3881/ab089d Astrophysical Journal, 859, 31,
doi: 10.3847/1538-4357/aabfbb
Dillamore, A. M., Belokurov, V., Font, A. S., & McCarthy,
Higson, E., Handley, W., Hobson, M., & Lasenby, A. 2019,
I. G. 2022, MNRAS, 513, 1867,
Statistics and Computing, 29, 891,
doi: 10.1093/mnras/stac1038
doi: 10.1007/s11222-018-9844-0
Fellhauer, M., Belokurov, V., Evans, N. W., et al. 2006, Hunter, J. D. 2007, Computing in Science and Engineering,
ApJ, 651, 167, doi: 10.1086/507128 9, 90, doi: 10.1109/MCSE.2007.55
22 Yu, Li, Speagle et al.

Iorio, G., & Belokurov, V. 2019, MNRAS, 482, 3868, Skilling, J. 2004, in American Institute of Physics
doi: 10.1093/mnras/sty2806 Conference Series, Vol. 735, Bayesian Inference and
Iorio, G., Belokurov, V., Erkal, D., et al. 2018, Monthly Maximum Entropy Methods in Science and Engineering:
Notices of the Royal Astronomical Society, 474, 2142, 24th International Workshop on Bayesian Inference and
doi: 10.1093/mnras/stx2819 Maximum Entropy Methods in Science and Engineering,
Koposov, S., Speagle, J., Barbary, K., et al. 2023, ed. R. Fischer, R. Preuss, & U. V. Toussaint, 395–405,
joshspeagle/dynesty: v2.1.2, Zenodo,
doi: 10.1063/1.1835238
doi: 10.5281/zenodo.7995596
Skilling, J. 2006, Bayesian Analysis, 1, 833 ,
Law, D. R., & Majewski, S. R. 2010, ApJ, 718, 1128,
doi: 10.1088/0004-637X/718/2/1128 doi: 10.1214/06-BA127
Li, T., & S5 Collaboration. 2021, Southern Stellar Stream Smith, M. C., Evans, N. W., & An, J. H. 2009, ApJ, 698,
Spectroscopic Survey: The First Public Data Release, 1110, doi: 10.1088/0004-637X/698/2/1110
Data Release 1, Zenodo, doi: 10.5281/zenodo.4695135 Speagle, J. S. 2020, MNRAS, 493, 3132,
Li, T. S., Koposov, S. E., Zucker, D. B., et al. 2019, doi: 10.1093/mnras/staa278
MNRAS, 490, 3508, doi: 10.1093/mnras/stz2731 Starkenburg, E., Youakim, K., Martin, N., et al. 2019,
Majewski, S. R., Skrutskie, M. F., Weinberg, M. D., & Monthly Notices of the Royal Astronomical Society, 490,
Ostheimer, J. C. 2003, ApJ, 599, 1082, 5757, doi: 10.1093/mnras/stz2935
doi: 10.1086/379504
Stringer, K. M., Drlica-Wagner, A., Macri, L., et al. 2021,
Medina, G. E., Muñoz, R. R., Vivas, A. K., et al. 2018, The
ApJ, 911, 109, doi: 10.3847/1538-4357/abe873
Astrophysical Journal, 855, 43,
Thomas, G. F., McConnachie, A. W., Ibata, R. A., et al.
doi: 10.3847/1538-4357/aaad02
Olling, R. P., & Merrifield, M. R. 2000, MNRAS, 311, 361, 2018, Monthly Notices of the Royal Astronomical
doi: 10.1046/j.1365-8711.2000.03053.x Society, 481, 5223, doi: 10.1093/mnras/sty2604
Palau, C. G., & Miralda-Escudé, J. 2023, MNRAS, 524, van der Walt, S., Colbert, S. C., & Varoquaux, G. 2011,
2124, doi: 10.1093/mnras/stad1930 Computing in Science and Engineering, 13, 22,
Pieres, A., Girardi, L., Balbinot, E., et al. 2020, MNRAS, doi: 10.1109/MCSE.2011.37
497, 1547, doi: 10.1093/mnras/staa1980 Vera-Ciro, C., & Helmi, A. 2013, ApJL, 773, L4,
Prada, J., Forero-Romero, J. E., Grand, R. J. J., Pakmor, doi: 10.1088/2041-8205/773/1/L4
R., & Springel, V. 2019, MNRAS, 490, 4877,
Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020,
doi: 10.1093/mnras/stz2873
Nature Methods, 17, 261, doi: 10.1038/s41592-019-0686-2
Preston, G. W., Shectman, S. A., & Beers, T. C. 1991,
Watkins, L. L., Evans, N. W., Belokurov, V., et al. 2009,
ApJ, 375, 121, doi: 10.1086/170175
Monthly Notices of the Royal Astronomical Society, 398,
Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ,
500, 525, doi: 10.1086/305772 1757, doi: 10.1111/j.1365-2966.2009.15242.x
Searle, L., & Zinn, R. 1978, The Astrophysical Journal, Wenger, M., Ochsenbein, F., Egret, D., et al. 2000, A&AS,
225, 357, doi: 10.1086/156499 143, 9, doi: 10.1051/aas:2000332
Sesar, B., Jurić, M., & Ivezić, Ž. 2011, ApJ, 731, 4, Xue, X.-X., Rix, H.-W., Ma, Z., et al. 2015, The
doi: 10.1088/0004-637X/731/1/4 Astrophysical Journal, 809, 144,
Sharma, S., Bland-Hawthorn, J., Silk, J., & Boehm, C. doi: 10.1088/0004-637X/809/2/144
2023, MNRAS, 521, 4074, doi: 10.1093/mnras/stad721 Yanny, B., Newberg, H. J., Kent, S., et al. 2000, ApJ, 540,
Shipp, N., Drlica-Wagner, A., Balbinot, E., et al. 2018, 825, doi: 10.1086/309386
ApJ, 862, 114, doi: 10.3847/1538-4357/aacdab