{"id":"7d35e598-9a5b-4f8d-b988-bf6d3739962c","arxiv_id":"2505.00763","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"JFlow uses equivariant continuous normalizing flows and kernel-density-estimated projected likelihoods to perform model-independent spherical Jeans analysis, recovering dark matter mass densities in a mock dwarf spheroidal galaxy within about 2 sigma.","lead":"This paper introduces JFlow, an unsupervised machine learning method that infers dark matter density in dwarf spheroidal galaxies from star positions and line-of-sight velocities without assuming a specific density profile. It validates the approach on mock data, which matters because parametric assumptions can bias dark matter estimates used in direct and indirect detection experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'model-independent' claim is undercut by the implicit Gaussian velocity assumption; a non-Gaussian mock test is needed before the claim can stand.","rationale":"The reader identified the β input as the weakest assumption; I agree that β is a real limitation and is already demonstrated in Fig. 2. However, the more hidden and equally load-bearing assumption is that the velocity distribution is exactly Gaussian. The paper presents this as a convenience for variance estimation, but the training objective fits the full projected distribution, not just the projected variance. The standard Jeans analysis only needs second moments, so a non-Gaussian velocity distribution with the same dispersions should produce the same mass profile. If JFlow's KL-based training does not preserve the second moments under misspecification of the velocity shape, then the method is not model-independent even when β is known. This is not a criticism of the math of the paper, but of the scope of the claim: the paper's proof-of-concept uses one mock, and the velocity distribution of that mock is not varied. The proposed test directly targets this gap and would settle whether the Gaussian assumption is benign or central. The verdict stays CONDITIONAL because the method is promising and the derivation is sound, but the 'model-independent' language needs to be tightened and the Gaussian-velocity dependence needs to be tested or explicitly acknowledged.","tokens_in":14839,"tokens_out":16513,"duration_ms":185210,"concrete_test":"Generate a spherical mock with the same stellar density and β=0 as NonplumCoreIso, but draw the 3D velocities from a two-component Gaussian mixture (or an Eddington-inverted isotropic distribution function) engineered to have exactly the same σ_r^2(r) and σ_t^2(r) as the original mock. Feed 1,000 projected stars to the JFlow pipeline with β=0 fixed. If the recovered v_r^2(r) and ρ(r) fall inside the quoted 2σ bands, the Gaussianity concern is refuted; if they shift systematically outside the bands, the 'model-independent' claim must be qualified to 'Gaussian-velocity Jeans analysis.'","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing weak point is the implicit Gaussian velocity distribution, not just the acknowledged β input. Equations (10)–(11) model the 3D velocity distribution at each radius as exactly Gaussian, with only the dispersion profile free. The training objective (12) is a full-distribution KL divergence between smoothed projected densities, so asymptotically JFlow returns the KL projection of the true 6D phase-space density onto the Gaussian-velocity family. The spherical Jeans equations (2)–(3) constrain only second moments and are agnostic to the velocity distribution shape; nothing guarantees that this KL projection preserves σ_r^2(r). If the true local velocity distribution is non-Gaussian with the same σ_r^2(r) and β(r), the fitted σ_r^2(r) may be biased, and that bias propagates directly into M(r) and ρ(r) through Eq. (3). The paper validates only on the NonplumCoreIso mock, and no test with a deliberately non-Gaussian velocity distribution is reported; the Gaussian assumption is not listed among the stated limitations. This is load-bearing because the headline 'model-independent' claim would fail if the inferred mass depends on the unstated Gaussianity assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces JFlow, an unsupervised machine-learning method for spherical Jeans analysis of dwarf spheroidal galaxies. It models the stellar number density n(r) with equivariant continuous normalizing flows and the radial velocity dispersion v2_r(r) with a conditional Gaussian flow, then substitutes the fitted functions into the spherical Jeans equation (Eq. 3) to estimate the enclosed mass M(r) and dark matter density ρ(r), given an input anisotropy profile β(r). Training uses only projected positions and line-of-sight velocities; the loss (Eq. 12) is the Kullback-Leibler divergence between kernel-smoothed projected densities of the data and the model. A proof-of-concept on the NonplumCoreIso Gaia Challenge mock with 1,000 tracers recovers the stellar density, radial velocity dispersion, enclosed mass, and mass density within 2σ over most of the radial range, and a comparison with parametric fitting shows wider credible intervals for JFlow, as expected for a free-form model.","tokens_in":15018,"tokens_out":9207,"duration_ms":100949,"significance":"If robust, JFlow would be a valuable free-form alternative to parametric Jeans analyses, enabling exploration of non-standard dark matter profiles without imposing analytic forms. The KDE-based training on projected data is a novel and asymptotically consistent approach, and the use of equivariant flows guarantees spherical symmetry by construction. The explicit treatment of the mass-anisotropy degeneracy via input β, and the demonstration that the enclosed-mass estimate near the Wolf radius is insensitive to β mismatch, are strengths. However, the 'model-independent' claim is overstated: the method also implicitly assumes a Gaussian velocity distribution at each radius, and it requires β as an input; the Gaussian assumption is not listed among the limitations and is not stress-tested in the paper.","major_comments":[{"comment":"The velocity model in Eqs. (10)–(11) is a position-dependent Gaussian with diagonal covariance, and the training loss (Eq. 12) is a full-distribution KL divergence between smoothed projected densities. In the asymptotic limit, JFlow therefore converges to the KL projection of the true phase-space density onto this Gaussian-velocity family. The spherical Jeans equation constrains only the second velocity moments, and there is no general guarantee that this KL projection preserves σ_r^2(r) when the true local velocity distribution is non-Gaussian. The paper validates the method only on the NonplumCoreIso mock and reports no test with deliberately non-Gaussian velocities; the Gaussian assumption is not listed among the limitations. Since the inferred M(r) and ρ(r) via Eq. (3) depend directly on σ_r^2(r), this is load-bearing for the central 'model-independent' claim. I request a non-Gaussian mock test (e.g., a two-component Gaussian or a distribution with non-zero excess kurtosis but identical σ_r^2(r) and β) and a quantification of the resulting bias; if bias appears, either relax the Gaussian velocity model or state the Gaussian assumption explicitly as a limitation in the abstract and conclusions.","section":"II (Eqs. 10–11) and III (Eq. 12)"},{"comment":"The statement that 'the optimal solution of the training in the asymptotic limit is f⊥∗Kh = f̂⊥∗Kh, and f̂⊥ will converge to f⊥ in the case of Gaussian kernels' is only true if the true projected density f⊥ is representable by the model family. Because the velocity model is Gaussian, this representability is not guaranteed for arbitrary dSph data; the general convergence target is the KL projection of f⊥ onto the family, not f⊥ itself. Please qualify this sentence accordingly, e.g., 'converges to the closest density within the model family.'","section":"III (near Eq. 12)"}],"minor_comments":[{"comment":"There is a typo: 'resplectively' should be 'respectively'.","section":"II, Eq. (2)"},{"comment":"The relation between the cusp parameter c and the inner logarithmic slope of the stellar density, namely d log n/d log r ≈ −3c/(c+1) at small r, is used implicitly in the iterative training of c in Appendix B but is never written out in the main text. Stating this explicitly would improve clarity.","section":"II, Eq. (9)"},{"comment":"Please state whether the NonplumCoreIso mock's intrinsic velocity distribution is Gaussian. This is directly relevant to the robustness test requested in Major Comment 1 and to interpreting the validation results.","section":"III and Appendix A"},{"comment":"The paper does not mention whether the code and trained models will be released. Given the complexity of the method, a code-availability statement would aid reproducibility.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The core methodology is sound and the proof-of-concept is convincing on the Gaussian-velocity mock, but the unstated Gaussian velocity assumption is a genuine threat to the 'model-independent' claim. The authors should be asked to add a non-Gaussian mock test; if the method is shown to be biased, the claim should be softened and the assumption listed as a limitation. The paper is otherwise within the scope of the journal and has moderate novelty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this: JFlow is the first to do spherical Jeans analysis with equivariant continuous normalizing flows and a KDE-based likelihood-free loss. The math is straightforward, the loss is asymptotically consistent, and on the one Gaia Challenge mock it recovers the stellar density, velocity dispersion, enclosed mass, and dark matter density within 2σ. It deserves a serious referee.\n\nWhat's good: The paper doesn't oversell the hard parts. It clearly states that β(r) is an external input because line-of-sight data can't lift the mass-anisotropy degeneracy. The training trick—smoothing both data and model with the same kernel so the KL loss is well-defined—is sensible. They also compare against parametric fitting and show the free-form model gives wider, more honest uncertainties. The small-radius and outer-radius caveats are flagged in the text.\n\nThree things bother me. First, the velocity model at every radius is exactly Gaussian (Eq. 10–11), and this is not listed as a limitation. The loss is a full-distribution KL divergence, so if the true 3D velocity distribution is non-Gaussian—which is realistic—the fitted dispersion could be biased because the model can't represent the shape. A single test on a non-Gaussian mock would settle this; they don't provide it. This is a fair stress-test concern, not a fatal flaw, but it undercuts the 'fully model-independent' phrase. Second, the validation is one dataset with 1000 stars and 20 bootstrap resamples; that's thin. Third, no code is released; for a method paper, that hurts reproducibility.\n\nWho's this for: people doing dSph dark matter inference or ML-based phase-space density estimation. It's a proof of concept, not the final tool—real dSphs have contamination, binary stars, and non-sphericity. But as a step toward free-form Jeans analysis, it's useful.\n\nSend it to peer review. Ask for a non-Gaussian mock test, a second mock (e.g., different anisotropy), and code release. The central idea is sound; the claims need calibration.","headline":"JFlow is a clean, honest proof-of-concept for free-form spherical Jeans analysis, but its 'model-independent' claim needs an explicit caveat about the Gaussian velocity assumption and a non-Gaussian stress test.","tokens_in":15606,"tokens_out":3961,"would_cite":true,"duration_ms":42805,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"JFlow replaces parametric stellar and dark-matter profiles with an equivariant continuous normalizing flow, recovering spherical Jeans-equation mass profiles from projected positions and line-of-sight velocities given an anisotropy input.","keywords":["spherical Jeans equation","dwarf spheroidal galaxies","dark matter density","normalizing flows","equivariant flows","velocity anisotropy","mass-anisotropy degeneracy","unsupervised machine learning"],"falsifier":"Take a mock spherical system with a known cored or cuspy dark-matter profile and a deliberately strong tangential anisotropy, supply JFlow with the true $\\beta(r)$, and check whether the recovered $\\rho_{\\rm DM}(r)$ matches the input over the well-sampled radial range; the paper already reports one $2.5\\sigma$ deviation at $r\\simeq0.4$ kpc, so repeating the test with larger tracer samples would reveal whether that deviation is finite-sample noise or a systematic bias that would invalidate the model-independent claim.","tokens_in":14612,"feed_emoji":"🌌","tokens_out":10716,"duration_ms":101920,"temperature":0.7,"pith_summary":"JFlow is an unsupervised machine-learning method for solving the spherical Jeans equation—the equilibrium condition linking stellar motions to enclosed mass—without assuming analytic forms for the stellar density, velocity dispersion, or dark-matter halo. From projected stellar positions and line-of-sight velocities, it trains an equivariant continuous normalizing flow to learn the spherically symmetric stellar number density $n(r)$ and a conditional flow to learn the radial velocity dispersion $\\overline{v_r^2}(r)$. Inserting these into the Jeans equation yields the enclosed dark-matter mass $M(r)$, provided the velocity-anisotropy profile $\\beta(r)$ is supplied as input. On a mock spherical dwarf-galaxy system with only 1,000 tracer stars, the method recovers $n(r)$, $\\overline{v_r^2}$, $M(r)$, and the dark-matter density within about $2\\sigma$ over the sampled radial range. A sympathetic reader would care because this replaces profile-choice bias in dwarf dark-matter inference, letting halo shapes beyond standard cusps and cores be tested directly.","feed_headline":"No-model flows recover dark matter from 1,000 stars","feed_subtitle":"An equivariant neural flow learns stellar density and velocity dispersion, then solves the spherical Jeans equation.","key_machinery":"The load-bearing object is an equivariant continuous normalizing flow: a neural ODE $\\frac{d\\vec r_t}{dt} = \\hat r \\tanh(|\\vec r_t|/r_s) F_r(|\\vec r_t|,t;\\theta)$ that evolves only the radial coordinate of an initially Gaussian sample, so the learned 3D stellar density is spherically symmetric by construction. A power-law warp $T_{\\rm pow}: |\\vec r|\\to|\\vec r|^{c+1}$ is composed with the cored flow to allow cuspy profiles, handling both flat-core and power-law stellar distributions. A companion conditional linear flow multiplies velocities by a diagonal matrix $L(r;\\theta)$ whose diagonal entries are neural-network models of the radial and tangential velocity dispersions, with the tangential components slaved to the input $\\beta(r)$. Training is done without an explicit projected likelihood: generated 3D samples are projected to observable coordinates, smoothed by a Gaussian kernel, and compared with kernel-smeared data through a Kullback-Leibler loss, which is differentiable so the flow parameters can be optimized. The spherical Jeans equation then is the bridge from these learned density and dispersion functions to $M(r)$ and $\\rho_{\\rm DM}(r)$.","core_discovery":"The paper's central claim is that the three ingredients of spherical Jeans analysis—stellar number density, radial velocity dispersion, and the resulting enclosed mass—can be estimated from projected sky positions and line-of-sight velocities alone, with no parametric model for the stellar or dark-matter distributions. The flow is rotationally equivariant by construction: the base Gaussian is spherically symmetric and the learned ODE only moves points along the radial direction, so the modeled density is guaranteed spherical. A separate position-dependent linear flow parameterizes the velocity variance, and both models are trained by a likelihood-free loss that compares kernel-smoothed generated samples with kernel-smeared data in the projected observable space. With $\\beta(r)$ given, the Jeans equation $M(r) = -\\frac{r\\overline{v_r^2}}{G}\\left(\\frac{d\\ln(n\\overline{v_r^2})}{d\\ln r} + 2\\beta\\right)$ converts the learned functions into the dark-matter mass profile. On a mock cuspy-stellar, cored-dark-matter halo with 1,000 tracers, JFlow's estimates agree with the true profiles within about $2\\sigma$, and the paper argues this carries over to other spherical systems and to testing nonstandard halo physics.","pith_inferences":["A decisive practical extension would be to feed JFlow anisotropy profiles derived from independent data—proper motions, higher-order velocity moments, or multi-component modelling—because the paper itself shows that an incorrect $\\beta$ biases the mass and density except near the special radius.","The reported $2.5\\sigma$ deviation in mass density at $r\\simeq0.4$ kpc is a candidate systematic tied to the KDE bandwidth and the curvature of $\\overline{v_r^2}$ near the training-data boundary; a testable prediction is that increasing sample size or reducing bandwidth should shrink this deviation if it is statistical.","The same architecture could be used to estimate dark-matter density slopes for indirect-detection $J$-factors, where the inner density profile matters most and where free-form estimates would quantify the model uncertainty that parametric fits understate."],"forward_implications":["Dark-matter density profiles of dispersion-supported dwarf galaxies can be inferred from a few hundred to a thousand member stars without committing to a halo model, improving the statistical reach of small spectroscopic samples.","Because no functional form is imposed, inferred halos can show cores, cusps, or other departures from standard profiles, making JFlow a direct test of self-interacting and wave dark-matter predictions.","Enclosed mass near the radius where the Jeans expression is nearly independent of $\\beta$ remains reliable even when the input anisotropy is wrong, giving a robust anchor for mass estimates.","The likelihood-free KDE training strategy does not rely on analytic Abel projection, so the same machinery can be extended to non-spherical systems and to observing conditions where projection is complicated, as the paper anticipates.","Parametric fits that happen to include the true profile give narrower credible intervals, but JFlow avoids the bias that appears when the true profile lies outside the assumed family; with correct $\\beta$, it estimates the mass profile without this bias."],"supporting_citations":[{"why":"Supplies the normalizing-flow change-of-variables density estimation machinery on which JFlow's density model is built.","marker":"[15]"},{"why":"Provides the neural-ODE formulation that makes the continuous normalizing flow a trainable coordinate transformation.","marker":"[38]"},{"why":"Gives the free-form continuous dynamics training recipe used to fit the radial flow without parametric densities.","marker":"[39]"},{"why":"Shows how equivariant transformations preserve the symmetry of a base distribution, guaranteeing the learned stellar density is spherical.","marker":"[40]"},{"why":"Supplies the mock spherical dwarf system with known stellar and dark-matter profiles used to validate the method.","marker":"[54]"},{"why":"Identifies the radius where Jeans enclosed-mass estimates are nearly independent of anisotropy, explaining the robustness seen under a wrong $\\beta$.","marker":"[56]"}],"fun_headline_variants":["Neural flow solves Jeans equation with no model assumptions","Equivariant flow recovers dark matter from sparse tracers","JFlow: model-free dark matter from stellar kinematics","Unsupervised flow cracks spherical Jeans analysis","Flow-based mass profiles from 1,000 stars without models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The velocity-anisotropy profile $\\beta(r)$ is provided as an external input and is never constrained by the data, so an incorrect $\\beta$ propagates directly into biased dark-matter mass and density estimates except near the radius where the enclosed mass is nearly $\\beta$-independent.","fun_headline_variants_meta":{"raw":{"variants":["Neural flow solves Jeans equation with no model assumptions","Equivariant flow recovers dark matter from sparse tracers","JFlow: model-free dark matter from stellar kinematics","Unsupervised flow cracks spherical Jeans analysis","Flow-based mass profiles from 1,000 stars without models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1416,"prompt_tokens":970,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":366}},"tokens_in":586,"tokens_out":446,"duration_ms":4991,"temperature":1.0,"reasoning_tokens":366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:35:54.934160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a mock spherical system with a known cored or cuspy dark-matter profile and a deliberately strong tangential anisotropy, supply JFlow with the true $\\beta(r)$, and check whether the recovered $\\rho_{\\rm DM}(r)$ matches the input over the well-sampled radial range; the paper already reports one $2.5\\sigma$ deviation at $r\\simeq0.4$ kpc, so repeating the test with larger tracer samples would reveal whether that deviation is finite-sample noise or a systematic bias that would invalidate the model-independent claim.","supporting_citations":[{"cited_title":"K¨ ohler, L","cited_arxiv_id":null,"evidence_quote":"Shows how equivariant transformations preserve the symmetry of a base distribution, guaranteeing the learned stellar density is spherical."},{"cited_title":"Reed, Spherical & triaxial, in 3rd Gaia Challenge (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the mock spherical dwarf system with known stellar and dark-matter profiles used to validate the method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Identifies the radius where Jeans enclosed-mass estimates are nearly independent of anisotropy, explaining the robustness seen under a wrong $\\beta$."}],"review_version":1}