Constraining the population of dark matter halo shapes using hierarchical inference with extragalactic stellar streams

David Chemaly; Denis Erkal; Elisabeth Sola; GyuChul Meyong; Hanyuan Zhang; Sergey Koposov; Vasily Belokurov

REVIEW 2 major objections 6 minor 1 references

Constraining the population of dark matter halo shapes using hierarchical inference with extragalactic stellar streams

T0 review · 2 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Dark matter halo flattening can be inferred from ensembles of faint extragalactic stellar streams seen only as two-dimensional tracks, by hierarchically combining weak individual posteriors.

desk verdict Useful proof-of-concept for population-level halo shape inference from photometric streams, but the printed hierarchical estimator is wrong as written; needs a fix and a re-check before trusting Fig. 8. read the letter →

arxiv 2601.15373 v3 pith:ZKE5BT3Q submitted 2026-01-21 astro-ph.GA

David Chemaly , Elisabeth Sola , Vasily Belokurov , Sergey Koposov , GyuChul Meyong , HanYuan Zhang , Denis Erkal This is my paper

classification astro-ph.GA

keywords stellarstreamsdarkmatterhaloshapesflatteningpopulationinferencehierarchicalbayesianextragalacticphotometricstreamtracksparticle-spraymodeling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Single faint extragalactic stellar streams, seen only as two-dimensional tracks on the sky, constrain a dark matter halo's flattening only weakly and can even favor the wrong oblate or prolate shape. This paper tries to establish that the population distribution of halo flattening can nevertheless be recovered by hierarchically combining dozens of such weak single-stream posteriors, and demonstrates it on 35 mock streams for oblate, spherical, and prolate populations. If true, halo shape — a key diagnostic of galaxy formation physics and dark matter self-interactions — becomes measurable at population level with photometry alone, no kinematics, at a computational cost that grows only linearly with sample size. The practical payoff is that upcoming wide-field imaging surveys, which will find many extragalactic streams, could yield the first population-scale halo-shape constraints outside the Milky Way.

What carries the argument

Two pieces carry the argument. First, a fast particle-spray stream generator: particles are injected at the tidal Lagrange points of an axisymmetric dark-matter halo and integrated as massless tracers, making thousands of fits tractable. Second, hierarchical Bayesian reweighting: individual stream fits produce posterior samples for the flattening q; the population evidence for a proposed mean and width is the sample average of the population-prior density evaluated at those q samples, divided by the uniform individual prior. A parametrisation maps the length of a zero-mean unit-variance Gaussian vector to q, yielding an exactly uniform prior on the flattening range and avoiding angular-perio

What would settle it

Generate mock streams from triaxial halos (or axisymmetric halos with radially varying flattening), add correlated or larger-than-2% noise, and rerun the same 35-stream hierarchical pipeline; if the input population shape is not recovered within 2 sigma, the population claim depends on the closed-loop setup. A cheaper check: run the pipeline with the progenitor's on-track position free and compare posterior widths.

Watch

Extended reading notes

Core claim

The paper's central claim is that ensembles of purely photometric stream tracks carry sufficient information to constrain the halo flattening distribution: for each of three mock populations (mean flattening 0.8, 1.0, 1.2 with spread 0.1, 35 streams each), hierarchical reweighting of per-stream flattening posteriors recovers the input population parameters, with ground truth within 2 sigma. Individual posteriors are broad and multiply modal, so the insight is that the population combination removes those modes and supplies tight, accurate constraints. The authors state that this is a proof of concept relying on closed-loop mocks and that real data will inevitably deviate from model assumptio

Load-bearing premise

The demonstration is closed-loop: the same axisymmetric halo forward model, 2% Gaussian radial noise, and truncated-Gaussian population family generate the mocks and are assumed in the fits, so any real-world mismatch (triaxial halos, baryonic disks, different noise structure, unknown on-track position) could bias the population inference.

Editorial extensions

If this is right

A sample of roughly 35 observable streams suffices to separate oblate, spherical, and prolate halo populations at the 2-sigma level in closed-loop tests.
The method uses only projected tracks with about 2% radial noise, so it can be applied to galaxies with no measured kinematics or distances along the stream.
The computational cost scales linearly with sample size, making population inference practical as stream catalogs grow by orders of magnitude in upcoming wide-field surveys.
Projection-induced oblate/prolate multi-modalities seen in single streams largely wash out under aggregation, but leave a residual pull toward rounder shapes and inflated population width.
The inferred projected halo orientation can be compared with the baryonic disk orientation in the same galaxies, enabling population-level studies of halo-disk misalignment as a future extension.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the closed-loop result survives real-data mismatch, a direct comparison of inferred population flattening to dark-matter-only and self-interacting-dark-matter simulations becomes a clean test of dark matter microphysics; the paper's axisymmetric mocks bracket what such a comparison could look like.
The 2%-noise and fixed on-track coordinate assumptions likely widen or bias posteriors when relaxed; a calibration step on more realistic mocks would be needed before trusting absolute population widths from real data.
The appendix's weak trends between stream morphology and fit quality suggest that selecting streams by shape (for example, curved or multi-loop tracks) may boost constraining power per galaxy more than simply adding more streams — a testable selection-strategy prediction.
Because individual fits show projection-induced degeneracies that population inference resolves, a natural extension is a mixture population model that asks how many streams are required to detect a secondary halo-shape component.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short take: this paper deserves a referee, but not a clean acceptance. The idea is good—combining many weak individual stream fits into a population constraint on halo flattening is a natural and timely move, and the mock tests are honestly presented. The individual fits are convincingly honest: wide posteriors, multi-modalities, prior-dominated degeneracies are all documented. StreaMAX seems like a real practical gain. The authors also state plainly that the validation is closed-loop and that model mismatch can bias real-data inference. That is the right attitude.

The problem is the math behind the population stage. Eq. (24) is the correct importance-sampling integral: an expectation over the posterior of pi(theta|alpha)/pi(theta). The Monte Carlo estimator in Eq. (25) then multiplies each sample by p(theta_ni|d_n), the posterior density itself. As written, that converges to the second moment of the posterior weighted by the ratio, not the marginal likelihood. The same extra factor appears in Eq. (32), the working expression, and via that expression it feeds directly into the population posteriors in Fig. 8. If the implementation follows the printed equations, the results are biased in a way the closed-loop tests cannot reveal, because the same (wrong) estimator is used to generate and to fit. The paper gives no effective sample size or comparison against a direct estimator, so the severity is unknown. This is a fixable defect if it is just a typo, but the authors need to correct the derivation and confirm the implementation.

Other soft spots are minor relative to this. The validation is entirely closed-loop, as the authors concede; triaxiality, baryons, unknown on-track position (y0 fixed) and non-2% noise are all deferred. That is acceptable for a proof-of-concept, but it means the headline claim about Euclid/Rubin readiness is a promise, not a demonstrated capability. The fitting code is 'available upon request' rather than in a public repo; for a methods paper that is a real weakness.

Bottom line: I would send this to peer review, because the concept is valuable and the weaknesses are largely correctable. But I would ask the referee to verify the estimator numerically on a toy problem before accepting. If the estimator is as printed, the central result is not supported.

Referee Report

2 major / 6 minor

Summary. The paper presents a hierarchical Bayesian framework to constrain the population distribution of dark matter halo flattening from projected (2D) stellar stream tracks in external galaxies, with no kinematic data. Streams are forward-modelled with a new JAX-accelerated particle-spray package (StreaMAX), individual streams are fitted with nested sampling to obtain posteriors on the flattening q and nuisance parameters, and the individual posteriors are then combined via a reweighting estimator to infer hyperparameters (mu_pop, sigma_pop) of a truncated Gaussian population distribution. The method is validated on three mock populations of 35 streams each (oblate, spherical, prolate); the true hyperparameters are recovered within about 2σ, with a modest bias toward spherical shapes. The authors frame this as a proof of concept for population-level inference of halo shapes from photometry alone, with applicability to future Euclid/LSST data.

Significance. If the central estimator is correct, this is a timely and potentially important proof-of-concept: it demonstrates that ensembles of low-information, purely photometric stream tracks can, in principle, constrain population-level halo morphology. The StreaMAX package is a valuable technical contribution, offering large speedups over existing stream-generation tools, and the authors are transparent about the closed-loop nature of their validation. However, the hierarchical reweighting estimator as printed is statistically invalid, and this threatens the primary quantitative claim (Fig. 8). The paper also contains an unverified modelling assumption about fixing the progenitor's on-track position. The significance of the work depends on whether the estimator is corrected and the analysis repeated.

major comments (2)

[Section 4, Eqs. (25) and (32)] The printed Monte Carlo estimator is not a valid approximation to the integral in Eq. (24). If {θ_ni} are samples from the posterior p(θ_n|d_n), the correct estimator is (1/K_n) Σ_i π(θ_ni|α)/π(θ_ni). The extra factor p(θ_ni|d_n) in Eq. (25) changes the target integral to ∫ p(θ|d)^2 [π(θ|α)/π(θ)] dθ, not the hierarchical likelihood. Since Eq. (32) is stated as the working expression, the population posteriors in Fig. 8 are formally biased unless the implementation omits that factor. The paper does not report an effective sample size or a comparison against a direct (correct) estimator, so the magnitude of the bias is unknown. This must be fixed: either correct the equations and re-run the population analysis, or, if the implementation correctly uses posterior weights, explain this and correct the printed formula.
[Section 2.3, constraint justification] The constraints y0=0, x0>0, z0>0, vy0>0 are justified by the statement 'From our tests, these assumptions do not affect the recovered posteriors of the halo parameters,' but no such test is presented. The mock data are generated under the same constraints, so the closed-loop recovery in Section 4 cannot validate this claim. If, for real streams, fixing the projected progenitor position y0=0 introduces a bias, the population inference would be biased in a way not captured by the mock tests. The authors should either show the test (e.g., fitting a mock stream with y0≠0) or explicitly soften the claim and discuss the expected impact.

minor comments (6)

[Throughout] There are numerous typographical errors and awkward formulations: 'thirtenn' (Table 1 caption), 'paramters' (Section 3.2), 'seperate' (Section 4), 'hierchical' (Section 4), 'constrains' for 'constraints' (Section 5). A careful proofread is needed.
[Section 3.1, Eq. (18)] The text says 'In the main text of a results paper one should specify whether a hard rejection or an explicit penalty of the form (18) is used.' This meta-comment is out of place; the authors should simply state their own choice and remove the self-referential remark.
[Section 4, Eq. (32)] Even if the erroneous density factor were removed, the paper does not describe how p(q_ni|d_n) is computed (e.g., KDE, histogram, or nested-sampling weights). In nested sampling, posterior samples come with weights, not densities. The authors should specify the exact estimator used so the analysis is reproducible.
[Section 2.2 and 2.3] The wording 'we choose the XY-plane' (Section 2.2) and 'the observer is always fixed at +z-axis' (Section 2.3) is slightly confusing; clarify that the projection plane is the XY plane and the line of sight is along z.
[Section 4, population generation] In the description of the mock populations, state explicitly that the true q values are drawn from a Gaussian truncated to [0.5,1.5], and that the same truncation is used in the fitted population model. This is implied but not clearly stated.
[Abstract and Section 5] The abstract and conclusions claim that 'ensembles of purely photometric streams carry sufficient information to constrain dark matter halo shapes.' This is only demonstrated under the closed-loop, axisymmetric-NFW, fixed-y0 assumptions. The authors do acknowledge this in Section 5(iv), but the abstract could be more cautious.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the population-level inference is a genuine closed-loop mock recovery rather than a construction that reduces to its inputs.

full rationale

The central result (Fig. 8) is a simulation-recovery test, not a derivation from assumed population parameters. Individual fits use Eq. (15) to build posteriors from mock tracks, and the hierarchical posterior in Eqs. (24)- (32) reweights posterior samples using the population prior; the hyperparameters (mu_pop, sigma_pop) are sampled by nested sampling rather than set to the generating values. The paper even reports a bias toward spherical shapes and inflated sigma for the oblate/prolate cases, which is the opposite of what a forced or circular fit would produce. The self-citations to Sola et al. (2025) / STRRINGS are used to set N=35, binning, and the 2% noise level, not as the mathematical justification of the hierarchical method, so they are not load-bearing for the inference. Section 5(iv) explicitly acknowledges that all tests are closed loop: mocks and fits share the same forward model and noise, and the population prior and likelihood families match. This is a real validation limitation that can overstate how well the method will transfer to real extragalactic streams, but it is not a definitional equivalence between outputs and inputs. The printed Monte-Carlo estimator in Eq. (25) appears to contain an extra posterior-density factor and is therefore a numerical-correctness risk, but that does not make the population posterior identical to an input by construction. No circular step is established.

Assumptions & free parameters 10 free parameters · 7 assumptions · 0 invented entities

The pipeline fits 13 free parameters per stream plus two population hyperparameters; the q prior is imposed exactly uniform via the Maxwellian probability integral transform (Eqs. 6-11). The main domain assumptions are the particle-spray approximation, the axisymmetric NFW idealization (no triaxiality, no baryons), the exactly-matched noise model, the implicit selection conditioning, and the overlap condition for the two-stage reweighting. One ad hoc modeling choice fixes y0 and imposes sign constraints to suppress projection modalities. No invented physical entities.

free parameters (10)

log10(M_halo/M_sun) = ≈12.46 (Fig. 3 truth)
Per-stream halo mass, uniform prior U(11,14); nuisance parameter for the q inference.
R_s (kpc) = ≈18.76 (Fig. 3 truth)
Halo scale radius, uniform prior U(10,25); nuisance parameter.
(x_hat, y_hat, z_hat) orientation vector = ≈(0.04, 0.14, 1.02) (Fig. 3 truth)
Gaussian N(0,1), z≥0; flattening q is a deterministic function of |r| via Eq. 11, so q is not independently free.
log10(m_prog/M_sun) = ≈8.40 (Fig. 3 truth)
Progenitor Plummer mass, uniform prior U(7,9); nuisance parameter.
r_s,prog (kpc) = ≈1.76 (Fig. 3 truth)
Progenitor Plummer scale radius, uniform prior U(1,5); nuisance parameter.
(x0, z0) projected final positions = ≈(71.5, 23.7) (Fig. 3 truth)
Gaussian N(0,150), truncated positive; y0 is fixed to 0 by hand.
(vx0, vy0, vz0) final velocities = ≈(204, 372, 28) km/s (Fig. 3 truth)
Gaussian N(0,250), vy0>0; nuisance parameters.
t_int (Gyr) = ≈2.93 (Fig. 3 truth)
Integration time, uniform prior U(1,4); trades off against halo mass under the no-kinematics degeneracy.
mu_pop = truths 0.8 / 1.0 / 1.2
Population mean of flattening; the actual target hyperparameter, hyperprior U(0.5,1.5).
sigma_pop = truth 0.1 for all three populations
Population scatter of flattening; target hyperparameter, hyperprior U(0,1).

assumptions (7)

domain assumption Particle-spray generation (massless debris released at L1/L2 with a Plummer dispersion prescription) reproduces observable stream tracks with sufficient fidelity for shape inference.
Adopted in Section 2.1 following Kupper et al. (2012), Gibbons et al. (2014), Fardal et al. (2015); fidelity is asserted, not validated against N-body in this paper; width and density information are intentionally discarded.
domain assumption The host halo is exactly an axisymmetric NFW potential; the flattening q fully describes its shape, with no triaxiality, baryonic disk, or radial-profile variation.
Section 2.3; acknowledged in Sections 3.4 and 5(iii) as an idealization that introduces degeneracies in reality.
domain assumption The likelihood noise model (sigma_i = 0.02 r_i^data) equals exactly the noise injected into the mock data.
Section 3.1; the bookkeeping is exact by construction, removing any noise-model mismatch from the validation; the 2% level is asserted to derive from STRRINGS.
domain assumption Selection cuts (angular coverage > pi/2, arc length > 100 kpc, radial extent [10,500] kpc, >=100 particles per bin) can be applied to mock data without explicit selection terms in the likelihood, beyond hard rejection of tracks shorter than the data.
Sections 2.3 and 3.1; selection is on data observables so conditioning is treated as ignorable, but the joint selection-plus-coverage effect is not modeled.
domain assumption Individual posteriors on q cover the population-relevant region, and pi(q|alpha) lies inside the individual prior pi(q), so the two-stage reweighting has adequate support overlap.
Section 4, labeled 'a critical condition'; no effective-sample-size or importance-weight diagnostics are reported for the overlap.
ad hoc to paper Fixing y0=0 and imposing sign constraints (x0>0, z0>0, vy0>0) removes projection degeneracies without biasing the recovered halo-parameter posteriors.
Section 2.3; the authors write 'This assumes we know the progenitor's position in one dimension, an assumption that does not always hold,' and assert the constraints do not affect halo posteriors.
standard math Probability integral transform: Maxwell-distributed radius maps to exactly uniform q in [0.5,1.5] with an isotropic orientation prior restricted to one hemisphere.
Section 2.3, Eqs. (6)-(11); mathematically correct, though the text mislabels the orientation prior as 'anisotropic' (it is isotropic on the hemisphere z>=0).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Constraining the population of dark matter halo shapes using hierarchical inference with extragalactic stellar streams." pith.science (2026). https://pith.science/paper/ZKE5BT3Q

@misc{pith2026260115373,
  author       = {Pith},
  title        = {Pith review of: Constraining the population of dark matter halo shapes using hierarchical inference with extragalactic stellar streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKE5BT3Q}},
  note         = {Machine review of arXiv:2601.15373}
}

read the original abstract

Stellar streams, the debris of tidally disrupted satellites, trace their host's gravitational potential and thus probe dark matter halo structure. While six-dimensional phase-space data of Galactic streams enable precise dark matter halo modelling in the Milky Way, streams around external galaxies are typically available only as low surface brightness features without kinematics (i.e. two-dimensional photometric data), providing only weak constraints when considered individually. We present a hierarchical Bayesian framework that infers the population distribution of halo flattening using only projected stream tracks. Streams are forward-modelled in StreaMAX, a new JAX-accelerated particle-spray package that achieves orders of magnitude faster stream generation when compared to traditional methods. For each stream we fit an axisymmetric dark matter halo model and obtain a posterior on the flattening. These posteriors are then combined through hierarchical reweighting to constrain the population distribution. Using mock data, we show that individual fits recover the correct flattening with modest precision and exhibit projection-induced multi-modalities. Nevertheless, aggregating these fits yields accurate and confident constraints on the underlying population distribution of dark matter halo morphologies, clearly distinguishing between oblate, spherical, and prolate populations. The total computational cost scales linearly with sample size. Our results demonstrate that ensembles of purely photometric streams carry sufficient information to constrain dark matter halo shapes in external galaxies at the population level. With the forthcoming samples from Euclid and Rubin/LSST, this approach offers a practical path to population-level inferences of halo morphology without any kinematic measurements.

Figures

Figures reproduced from arXiv: 2601.15373 by the authors.

**Figure 1.** Time required to model a tidal stream as a function of the number of particles, using identical integration settings across Gala and StreaMAX on both CPU and GPU. CPU tests were performed on a 44-core node, and GPU tests use an NVIDIA L4. StreaMAX outperforms the conventional package, with the GPU version delivering the fastest and flattest scaling with particle number. particles are born at different times. A parti… view at source ↗

**Figure 2.** Schematic representation showing the coordinate system of the axisymmetric NFW potential (grey grid), the halo flattened axis 𝒓, and the progenitor’s present-day phase-space location (black circle). The origin is fixed and taken as the centre of the DM halo. The direction of the halo symmetry axis is then given by the unit vector 𝒏ˆ = 𝒓 𝑟 . (4) To remove the trivial reflection degeneracy of an axisymmetric potentia… view at source ↗

**Figure 3.** Corner plot for a single projected stream fit (13 parameters). Posteriors are shown in blue and the ground truth in red. The mock data include 2% Gaussian radial noise. Overall, the posteriors recover the truth, though several degeneracies remain (see Section 3.3). Gaussian noise to the radius of the mock data based on STRRINGS (Sola et al. 2025). Therefore, for our Gaussian likelihood to be correct and completely u… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Summary of the 1000 samples drawn uniformly at random from the posterior (blue) compared to the ground truth (red) for a single stream. Shading indicates the 1, 2, and 3𝜎 drawned sample envelopes. Left: Comparison between the samples to the mock data on the 𝑋𝑌 plane. M…

**Figure 5.** Figure 5: Joint posterior distributions showing the degeneracy of the mass of the halo (log10 (𝑀 [𝑀⊙ ] )) as a function of the mass of the progenitor (log10 (𝑚 [𝑀⊙ ] )), the velocity (log10 (v [km/s] )) and integration time. Ground truth is shown is red. Increasing the mass of t…

**Figure 6.** Figure 6: Zoom on the correlated posterior of the orientation from one mock stream track projected onto the 𝑋𝑌-plane (the setting is the same than [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Examples of best-fit streams projected in the XY plane (top) and their posteriors for the flattening 𝑞 (bottom). From left to right: oblate, spherical, and prolate cases, with the ground truth in red. In all three cases, the fits recover the mock tracks and the posteri…

**Figure 8.** Figure 8: Posteriors on the parameters of the underlying population distribution of flattening for three distinct dark matter halo morphologies: oblate (left), spherical (middle) and prolate (right). Blue dashed vertical lines indicate the 16th, 50th, and 84th percentiles, corre…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 linked inside Pith

[1]

Abbott T. M. C., et al., 2018, ApJS, 239, 18 Allgood B., Flores R. A., Primack J. R., Kravtsov A. V., Wechsler R. H., Faltenbacher A., Bullock J. S., 2006, MNRAS, 367, 1781 Amorisco N. C., 2015, MNRAS, 450, 575 Amorisco N. C., Martinez-Delgado D., Schedler J., 2015, arXiv e-prints, p. arXiv:1504.03697 Atkinson A. M., Abraham R. G., Ferguson A. M. N., 2013...

arXiv 2018

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Abbott T. M. C., et al., 2018, ApJS, 239, 18 Allgood B., Flores R. A., Primack J. R., Kravtsov A. V., Wechsler R. H., Faltenbacher A., Bullock J. S., 2006, MNRAS, 367, 1781 Amorisco N. C., 2015, MNRAS, 450, 575 Amorisco N. C., Martinez-Delgado D., Schedler J., 2015, arXiv e-prints, p. arXiv:1504.03697 Atkinson A. M., Abraham R. G., Ferguson A. M. N., 2013...

arXiv 2018