Pith. sign in

REVIEW 4 major objections 5 minor 40 references

A residual neural network trained on the IllustrisTNG300 simulation corrects the classical Projected Mass Estimator's systematic overestimate, reducing the simulated mass ratio from ~1.3–1.5 to 1.02 and yielding a Milky Way mass of 1.144 ×

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 23:39 UTC pith:VOLLHCHO

load-bearing objection Core PME bias result and MLP calibration are solid on TNG, but the headline Milky Way mass is taken from an extrapolative regime the authors themselves flag, and the EAGLE validation is inconsistently reported. the 4 major comments →

arxiv 2607.15339 v1 pith:VOLLHCHO submitted 2026-07-16 astro-ph.GA astro-ph.COgr-qc

Tighter Dark Matter Constraints from the Projected Mass Method: A Neural Network Enhanced Method for Galaxy Groups and Clusters

Yinbo Huang , David Benisty This is my paper
classification astro-ph.GA astro-ph.COgr-qc
keywords dark mattergalaxy groupsgalaxy clustersmass estimationsatellite kinematicsneural networkProjected Mass EstimatorIllustrisTNG
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the classical Projected Mass Estimator, widely used to weigh galaxies and groups from satellite motions, systematically overestimates halo masses by roughly 30–45% because it assumes isotropic orbits while real satellites in ΛCDM halos are radially biased (mean squared eccentricity ≈ 0.57). To fix this, the authors train a residual neural network on thousands of halos from the IllustrisTNG300 simulation to predict the logarithmic offset between the estimated and true mass. On test halos the correction brings the median mass ratio from 1.30 (2D) and 1.46 (3D) to 1.02 and cuts logarithmic scatter from roughly 0.3 dex to 0.13 dex. Applied to the Local Universe, the method yields a Milky Way mass of 1.144^{+0.399}_{-0.296} × 10^12 M_sun, an M81 mass of 2.42^{+0.67}_{-0.52} × 10^12 M_sun, and an NGC 5128 mass of 3.91^{+1.3}_{-0.97} × 10^12 M_sun, consistent with independent dynamical modeling. If correct, many prior PME-based halo masses in the literature are biased high, and the N-segmented network provides a practical recalibration for sparse tracer samples.

Core claim

The central claim is that the systematic bias of the Projected Mass Estimator originates from its fixed isotropic coefficient C = 16/π, and that this bias can be learned and removed by a residual multi-layer perceptron trained on cosmological simulations. The network takes four summary features — the turnaround radius R0 derived from the PME mass, the raw PME mass, the line-of-sight velocity dispersion σ_v, and the mean line-of-sight velocity — and predicts the logarithmic residual Δ = log10(M_true/M_PME). Because models are trained separately for each tracer count N from 5 to 50, the correction adapts to the noise regime of sparse samples. On independent test halos the corrected mass ratio

What carries the argument

The central object is the Projected Mass Estimator, M_PME = (16/π G N) Σ v_los² R, whose coefficient C encodes projection geometry and orbital anisotropy; the paper measures ⟨e²⟩ ≈ 0.57 in TNG300, i.e., radial bias, rendering C = 16/π too large. The correction machinery is a residual MLP with three residual blocks, layer normalization, and GELU activations, trained to predict the logarithmic mass residual using four features (log R0, log M_PME, σ_v, |v_los|). The N-segmented training strategy — independent models per tracer number — is what lets the network account for the small-sample noise that a single universal recalibration (C_new ≈ 3.97) cannot remove.

Load-bearing premise

The calibration rests on the assumption that the N most massive bound subhalos in IllustrisTNG300 stand in for the N brightest observed satellites, with the same radial anisotropy (⟨e²⟩ ≈ 0.57), so once faint satellites without simulated counterparts are included, the correction becomes an extrapolation.

What would settle it

Apply the trained N=20 model to a galaxy group or cluster whose mass is independently known from weak lensing or X-ray hydrostatics; if the corrected PME differs from the independent mass by more than the 0.13 dex scatter quoted here, the simulation-based calibration fails to transfer to real systems.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Existing halo masses derived with the classical PME in the isotropic limit are likely systematically high by ~30–45%; applying a factor ≈0.78 correction would remove the median bias.
  • Satellite-based mass measurements of nearby groups can reach ~0.13 dex precision, competitive with more expensive dynamical modeling, even with as few as 5–10 tracers.
  • Mass-to-light ratios and dark-matter-dominated fractions for galaxy groups and clusters should be re-evaluated using the corrected masses.
  • The cross-simulation validation suggests the learned correction captures orbital-anisotropy effects common across ΛCDM hydrodynamical simulations, so the network may transfer to other simulation-based mass estimators.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reliance on abundance matching (most massive subhalos ↔ brightest satellites) is the most fragile link; the paper's own Milky Way test shows drift once faint satellites like Hydrus I enter, suggesting the method should be restricted to the brightest tracers or retrained with explicit luminosity–mass scatter.
  • A direct testable extension would be to train the same residual-learning architecture on other estimators (virial theorem, caustics) to see whether the learned corrections are estimator-specific or encode the same radial-anisotropy bias.
  • The feature-importance result — that σ_v dominates while anisotropy descriptors add little — implies the network is effectively learning a velocity-dispersion-based scaling; this could be distilled into an analytic correction formula for easier adoption.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a neural-network (MLP) correction to the classical Projected Mass Estimator (PME) for galaxy-group and cluster masses. Separate residual networks are trained for each tracer multiplicity N on IllustrisTNG300 simulated halos, predicting Δ = log10(M_true/M_PME) from four features (PME mass, turnaround radius, line-of-sight velocity dispersion, mean line-of-sight velocity). On TNG test halos, the classical PME is found to overestimate masses by ~30–46% with RMSE ~0.29–0.32 dex, while the MLP-corrected estimates are consistent with the 1:1 relation with RMSE ~0.13 dex. A simple Bayesian rescaling of the PME coefficient gives s=0.78, C_new≈3.97. The authors report cross-validation on the EAGLE simulation and apply the method to the Milky Way, M81, and NGC 5128, quoting masses of 1.144, 2.42, and 3.91 × 10^12 M_sun respectively, with tighter error bars than previous work.

Significance. The TNG-only results (Figs. 1, 4, 5) are internally consistent and provide a credible demonstration that a residual MLP trained on a large cosmological simulation can remove the systematic bias of the PME and reduce scatter for simulated satellite systems. The N-segmented training strategy is a sensible way to handle the strong N-dependence of the estimator's noise properties, and the independent test-set evaluation is appropriate. If the cross-simulation validation is made reliable and the application-domain caveats are properly handled, the method could be a practically useful recalibration of a widely used estimator. The paper is honest in several places about the training/application mismatch, but the abstract and conclusion do not fully carry those caveats through to the headline numbers.

major comments (4)
  1. [§III 'EAGLE Cross Validation'; §V; Fig. 8] The EAGLE validation metrics are reported inconsistently. §III states the MLP reduces the mean bias from 0.141 dex to 0.133 dex and the RMSE from 0.164 dex to 0.145 dex. §V states the reduction is from 0.194 dex to 0.087 dex in bias and from 0.415 dex to 0.340 dex in RMSE. Fig. 8's caption implies log10(1.23)=0.090 and log10(0.89)=−0.051 for the raw and corrected bias. These are not round-off differences. Since the EAGLE cross-check is the main evidence that the correction is not a TNG artifact, the manuscript must provide one consistent set of validation metrics and explain the discrepancies.
  2. [§IV.A; Appendix Table III; Abstract] The abstract's headline Milky Way mass, M_MW = 1.144^{+0.399}_{-0.296} × 10^12 M_sun, is the N=20 entry in Table III (faintest satellite m_V=14.8). But §IV.A explicitly states that once ultra-faint satellites such as Hydrus I enter at N=11, the MLP correction is 'increasingly forced into an extrapolative regime' and that the model is most informative for the brightest subset. The N=5 and N=10 estimates, which lie closer to the training domain, are 0.944 and 0.818 × 10^12 M_sun. The headline value is therefore taken from the regime the authors themselves identify as unreliable. The abstract and conclusion should report the N=5–10 range or give a quantitative argument for trusting N=20 despite the stated domain mismatch.
  3. [§III, selection of tracers; §IV applications] The training set uses, for each halo, the N most massive bound subhalos, while the observational applications use the N brightest satellites. The paper justifies this by abundance-matching monotonicity, but no scatter or completeness modeling is presented. At the satellite masses relevant here, the luminosity–subhalo-mass relation has significant scatter, and a magnitude-limited sample can be systematically different from a mass-selected sample. The real-data inference is therefore conditional on an unquantified mapping. A validation experiment on TNG300 with luminosity- or stellar-mass-selected tracers and a realistic apparent-magnitude cut would directly test this assumption and should be added or explicitly argued to be unnecessary.
  4. [Fig. 9; Abstract] The quoted 1σ uncertainties on the real-system masses are the calibration scatter from the TNG300 test set, as indicated by the 'ML-corrected calibration scatter' label in Fig. 9. The outer band in Fig. 9 is a ±10% input-perturbation test. Neither term accounts for the selection mapping systematics discussed in the previous comment or for the choice of N. The abstract's statement that the method gives 'a tighter constraint' is therefore not fully supported for real data as it stands; the error bars should either include a systematic component from the simulation–observation domain shift or be explicitly labeled as conditional.
minor comments (5)
  1. [Fig. 6] The axis label reads 'log10 (MTME)' — a typo for 'MPME'.
  2. [§I, Eq. (9)] The derivation of the isotropic coefficient C=16/π is abbreviated: starting from a delta-function DF on a single orbit, one must average over orbital eccentricity to recover the isotropic result, but that average is not shown. Please clarify the missing step or state the assumed eccentricity distribution.
  3. [Abstract] The phrase 'dark matter rate prediction' is unclear; presumably the authors mean mass-to-light ratios or dark matter content. Please rephrase.
  4. [References [10] and [11]] References [10] and [11] appear to be the same paper (Di Cintio et al. 2012) with slightly different spelling of an author name. One duplicate should be removed or the intended distinct references listed.
  5. [Fig. 5 and Appendix Table III] Fig. 5 shows N from 10 to 50 on the x-axis, while the abstract and Table III include N=5. Please indicate whether N=5 models were trained and evaluated, and show that point in Fig. 5 or explain its omission.

Circularity Check

0 steps flagged

No significant circularity: supervised calibration with held-out TNG test and external EAGLE validation; MW faint-satellite extrapolation is a flagged limitation, not a circular step.

full rationale

The claimed improvement over the classical PME is a supervised calibration, not a first-principles derivation. The MLP is trained on TNG300 to minimize MSE on Δ = log10(M_true/M_PME), and all reported test metrics are computed on an independent 20% holdout ('All reported performance metrics are computed on an independent test set that is not used during training'), so the reduction from M_proj/M_true ≈ 1.30 to 1.02 is a genuine out-of-sample evaluation rather than a fit recycled as a prediction. The EAGLE cross-simulation validation provides an external benchmark with no retuning, further breaking any self-referential loop. The applied Local-Universe masses are forward predictions of the trained model; the quoted error bars are the calibration scatter of the test set and are explicitly labeled as such ('should therefore be interpreted as a calibration uncertainty'), so they are not presented as independent measurements. The paper itself flags the main weakness: once faint satellites such as Hydrus I enter the Milky Way sample at N = 11, 'the MLP correction is increasingly forced into an extrapolative regime' — this is a domain-mismatch limitation affecting the headline Milky Way mass, but it is not circularity because the model output is not defined in terms of the target mass or fitted to the Milky Way. There is a minor self-citation (Wagner et al. [32], co-authored by Benisty) for the M81 catalog and comparison value, but the method's validity does not reduce to that citation; it rests on the TNG test-set and EAGLE results. The internal inconsistency between the EAGLE numbers in §III and §V is a reproducibility concern, not an instance of a result being equivalent to its input by construction. No quoted step meets the bar of exhibiting a specific reduction of a claim to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper's own theoretical contribution (§I) re-derives the standard PME coefficient C = 16/π under isotropy; everything quantitative beyond that is calibrated to TNG300. The free parameters comprise the global rescale s = 0.780, the measured mean eccentricity ⟨e²⟩ = 0.57, and the MLP weights for each N-segmented model (not released). The load-bearing domain assumptions are (a) that bound TNG300 subhalos are faithful tracers of M200c, (b) that selecting the N most massive subhalos in simulation matches selecting the N brightest satellites in observations, and (c) that TNG orbital anisotropy is representative of real groups. No new entities (particles, forces, dimensions) are postulated.

free parameters (4)
  • s — global PME rescaling factor = 0.780 ± 0.001 (C_new = 3.974 ± 0.006)
    Bayesian posterior fit to TNG300 (Eq. 11–12); the paper's own baseline for absorbing the PME's median bias.
  • MLP weights (N-segmented models) = not released
    Trained on TNG300 to predict Δ = log10(M_true/M_PME) from four features; the core of the correction, unavailable for independent reproduction.
  • ⟨e²⟩ — mean squared orbital eccentricity of satellites = 0.57 ± 0.02 (stacked), 0.58 ± 0.02 (largest halo)
    Measured from TNG300 subhalos (§II.B); used to argue C ≈ 3.9, implying radially biased orbits.
  • Network hyperparameters = 3 residual blocks × 2 FC layers; widths and learning rate not stated
    Chosen without reported ablation; affect the correction's behavior in the low-N regime.
axioms (4)
  • standard math Jeans' theorem and steady-state spherical symmetry justify the phase-space average used to derive the PME coefficient (§I, Eq. 6).
    Classical derivation reproduced from Bahcall & Tremaine (1981) and Heisler et al. (1985); not new content.
  • domain assumption Bound subhalos (ϵ<0) within the turnaround radius R0 of isolated TNG300 halos trace M200c in the same way real satellites trace group masses.
    Underlies both the measured PME bias factor (Fig. 1) and the training labels; central to the whole calibration.
  • domain assumption The N most massive subhalos in simulation correspond to the N brightest satellites in observations (abundance-matching monotonicity).
    Stated in §III; this mapping legitimizes the MW/M81/NGC5128 applications, and the authors concede it breaks down for faint MW satellites (§IV.A).
  • domain assumption TNG300's radially biased satellite orbits (⟨e²⟩ ≈ 0.57) are representative of the real Local Volume.
    The magnitude and direction of the learned correction inherit this; the EAGLE check is the only external probe and its reported numbers are inconsistent.

pith-pipeline@v1.3.0-alltime-deepseek · 15423 in / 23157 out tokens · 213543 ms · 2026-08-01T23:39:04.865616+00:00 · methodology

0 comments
read the original abstract

Measuring the total mass of the Milky Way and nearby galaxy groups is difficult because classical dynamical estimators rely on assumptions about satellite orbital geometry that are rarely satisfied in practice, and because only a handful of satellite galaxies are typically available as kinematic tracers. We present a new framework that corrects the well-known Projected Mass Estimator (PME) using a residual neural network trained on thousands of simulated galaxy groups from the IllustrisTNG cosmological simulation. Separate networks are trained for each satellite sample size, from as few as 5 satellites up to 50, so that the correction automatically accounts for the statistical noise that dominates when only a small number of tracers is available. In tests on simulated halos, the classical PME systematically overestimates halo masses by factors of $M_{\rm proj}/M_{\rm true} = 1.30^{+0.72}_{-0.62}$ (using the 2D distance) and $1.46^{+0.97}_{-0.72}$ (using the 3D distance), with RMSE of 0.29 and 0.32 dex respectively. The neural-network correction reduces this to $M_{\rm proj}/M_{\rm true} = 1.02^{+0.30}_{-0.26}$ with an RMSE of 0.13 dex. Applied to the Milky Way, the method yields a total mass of $M_{\rm MW} = 1.144^{+0.399}_{-0.296}\times10^{12}\,M_\odot$, with estimates based on the brightest 5-10 satellites favoring a somewhat lower range of $(0.8$-$0.95)\times10^{12}\,M_\odot$. The modified PME gives a tighter constraint on the virial masses and the dark matter rate prediction in galaxy groups and clusters.

Figures

Figures reproduced from arXiv: 2607.15339 by David Benisty, Yinbo Huang.

Figure 1
Figure 1. Figure 1: FIG. 1. Comparison between the true halo mass [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Distribution of subhalos around the selected host halo [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. The eccentricity distribution of the bound satellite [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Comparison of mass estimation performance for the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Permutation importance analysis for the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Pearson correlation matrix for the test-set outputs of [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Comparison of mass estimation performance on the EAGLE test set. The MLP-corrected estimates show a significant [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Milky Way mass estimates as a function of the num [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Total mass estimates for the M81 group using our [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11. Halo mass estimates of NGC 5128 derived from [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12. Mass-to-light ratio in the [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 32 linked inside Pith

  1. [1]

    J. F. Navarro, C. S. Frenk, and S. D. M. White, Astro- phys. J.462, 563 (1996), arXiv:astro-ph/9508025

    Pith/arXiv arXiv 1996

  2. [2]

    J. F. Navarro, C. S. Frenk, and S. D. M. White, Astro- phys. J.490, 493 (1997), arXiv:astro-ph/9611107

    Pith/arXiv arXiv 1997

  3. [3]

    Salucci, Astron

    P. Salucci, Astron. Astrophys. Rev.27, 2 (2019), arXiv:1811.08843 [astro-ph.GA]

    Pith/arXiv arXiv 2019

  4. [4]

    R. B. Tully, Astron. J.149, 54 (2015), arXiv:1411.1511 [astro-ph.GA]

    Pith/arXiv arXiv 2015

  5. [5]

    E. F. Bell and R. S. de Jong, Astrophys. J.550, 212 (2001), arXiv:astro-ph/0011493

    Pith/arXiv arXiv 2001

  6. [6]

    J. N. Bahcall and S. Tremaine, Astrophys. J.244, 805 (1981)

    1981

  7. [7]

    Heisler, S

    J. Heisler, S. Tremaine, and J. N. Bahcall, Astrophys. J. 298, 8 (1985)

    1985

  8. [8]

    L. L. Watkins, N. W. Evans, and J. H. An, Mon. Not. R. Astron. Soc.406, 264 (2010), arXiv:1002.4565 [astro- 11 ph.GA]

    Pith/arXiv arXiv 2010

  9. [9]

    Makarov, D

    D. Makarov, D. Makarov, L. Makarova, and N. Libeskind, Astron. Astrophys.698, A178 (2025), arXiv:2505.06642 [astro-ph.GA]

    Pith/arXiv arXiv 2025

  10. [11]

    Di Cintio, A

    A. Di Cintio, A. Knebe, N. I. Libeskind, Y. Hoffman, G. Yepes, and S. Gottlöber, Mon. Not. R. Astron. Soc. 423, 1883 (2012), arXiv:1204.0005 [astro-ph.CO]

    Pith/arXiv arXiv 2012

  11. [12]

    Pillepichet al., Mon

    A. Pillepichet al., Mon. Not. Roy. Astron. Soc.490, 3196 (2019), arXiv:1902.05553 [astro-ph.GA]

    Pith/arXiv arXiv 2019

  12. [13]

    V. F. Calderon and A. A. Berlind, Mon. Not. Roy. As- tron. Soc.490, 2367 (2019), arXiv:1902.02680 [astro- ph.GA]

    Pith/arXiv arXiv 2019

  13. [14]

    T. M. Callingham, M. Cautun, A. J. Deason, C. S. Frenk, W. Wang,et al., Mon. Not. R. Astron. Soc.484, 5453 (2019), arXiv:1808.10456

    Pith/arXiv arXiv 2019

  14. [15]

    Makarov, D

    D. Makarov, D. Makarov, K. Kozyrev, and N. Libe- skind, Universe11, 144 (2025), arXiv:2503.12612 [astro- ph.GA]

    Pith/arXiv arXiv 2025

  15. [16]

    Vogelsberger, S

    M. Vogelsberger, S. Genel, V. Springel, P. Torrey, D. Si- jacki, D. Xu, G. Snyder, S. Bird, D. Nelson, and L. Hern- quist, Nature (London)509, 177 (2014), arXiv:1405.1418 [astro-ph.CO]

    Pith/arXiv arXiv 2014

  16. [17]

    Nelson, A

    D. Nelson, A. Pillepich, S. Genel, M. Vogelsberger, V. Springel, P. Torrey, V. Rodriguez-Gomez, D. Si- jacki, G. F. Snyder, B. Griffen, F. Marinacci, L. Blecha, L. Sales, D. Xu, and L. Hernquist, Astronomy and Com- puting13, 12 (2015), arXiv:1504.00362 [astro-ph.CO]

    Pith/arXiv arXiv 2015

  17. [18]

    Genel, P

    D.Nelson, A.Pillepich, V.Springel, R.Pakmor, R.Wein- berger, S. Genel, P. Torrey, M. Vogelsberger, F. Mari- nacci, and L. Hernquist, Mon. Not. Roy. Astron. Soc. 490, 3234 (2019), arXiv:1902.05554 [astro-ph.GA]

    Pith/arXiv arXiv 2019

  18. [19]

    Springel, S

    V. Springel, S. D. M. White, G. Tormen, and G. Kauff- mann, Mon. Not. R. Astron. Soc.328, 726 (2001), arXiv:astro-ph/0012055 [astro-ph]

    Pith/arXiv arXiv 2001

  19. [20]

    Sandage, Astrophys

    A. Sandage, Astrophys. J.307, 1 (1986)

    1986

  20. [21]

    Barber, E

    C. Barber, E. Starkenburg, J. Navarro, A. McConnachie, and A. Fattahi, Mon. Not. Roy. Astron. Soc.437, 959 (2014), arXiv:1310.0466 [astro-ph.GA]

    Pith/arXiv arXiv 2014

  21. [22]

    K. L. Lange, R. J. A. Little, and J. M. G. Taylor, Journal of the American Statistical Association84, 881 (1989)

    1989

  22. [23]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville,Deep Learn- ing("", 2016)

    2016

  23. [24]

    S. B. Green, M. Ntampaka, D. Nagai, L. Lovisari, K. Dolag, D. Eckert, and J. A. ZuHone, Astrophys. J. 884, 33 (2019), arXiv:1908.02765 [astro-ph.CO]

    Pith/arXiv arXiv 2019

  24. [25]

    K. He, X. Zhang, S. Ren, and J. Sun, inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)(2016) pp. 770–778

    2016

  25. [26]

    D. P. Kingma and J. Ba, arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  26. [27]

    P. S. Behroozi, C. Conroy, and R. H. Wechsler, Astro- phys. J.717, 379 (2010), arXiv:1001.0015 [astro-ph.CO]

    Pith/arXiv arXiv 2010

  27. [28]

    Schaye, R

    J. Schaye, R. A. Crain, R. G. Bower, M. Furlong, et al., Mon. Not. R. Astron. Soc.446, 521 (2015), arXiv:1407.7040

    Pith/arXiv arXiv 2015

  28. [29]

    Pillepich, V

    A. Pillepich, V. Springel, D. Nelson, S. Genel, J. Naiman, R. Pakmor, L. Hernquist, P. Torrey, M. Vogelsberger, R. Weinberger, and F. Marinacci, Mon. Not. R. Astron. Soc.473, 4077 (2018), arXiv:1703.02970 [astro-ph.GA]

    Pith/arXiv arXiv 2018

  29. [30]

    R. A. Crain, J. Schaye, R. G. Bower,et al., Mon. Not. R. Astron. Soc.450, 1937 (2015), arXiv:1501.01963

    Pith/arXiv arXiv 1937

  30. [31]

    S. E. Koposov, M. G. Walker, V. Belokurov, A. R. Casey, A. Geringer-Sameth, D. Mackey, G. Da Costa, D. Erkal, P. Jethwa, M. Mateo, E. W. Olszewski, and J. I. Bailey, Mon. Not. R. Astron. Soc.479, 5343 (2018), arXiv:1804.06430 [astro-ph.GA]

    Pith/arXiv arXiv 2018

  31. [32]

    Wagner, D

    J. Wagner, D. Benisty, and I. D. Karachentsev, Astron. Astrophys.706, A92 (2026), arXiv:2510.24840 [astro- ph.CO]

    arXiv 2026

  32. [33]

    Antoine, A

    D. Antoine, A. C. Seth, J. Strader, D. J. Sand, K. Voggel, A. K. Hughes, D. A. Forbes, N. Caldwell, and D. Crno- jević, Astronomy & Astrophysics685, A132 (2024)

    2024

  33. [34]

    I. D. Karachentsev and K. N. Telikova, Astron. Nachr. 339, 615 (2018), arXiv:1810.06326 [astro-ph.GA]

    Pith/arXiv arXiv 2018

  34. [35]

    I. D. Karachentsev and O. G. Kashibadze, Astron. Nachr. 342, 999 (2021), arXiv:2109.00336 [astro-ph.GA]

    Pith/arXiv arXiv 2021

  35. [36]

    Z. Li, Q. D. Wang, and S. Hameed, Mon. Not. R. Astron. Soc.376, 960 (2007), arXiv:astro-ph/0701487 [astro-ph]

    Pith/arXiv arXiv 2007

  36. [37]

    G. L. H. Harris, Publ. Astron. Soc. Aust.27, 475 (2010), arXiv:1004.4907 [astro-ph.GA]

    Pith/arXiv arXiv 2010

  37. [38]

    Posti and A

    L. Posti and A. Helmi, Astron. Astrophys.621, A56 (2019), arXiv:1805.01408

    Pith/arXiv arXiv 2019

  38. [39]

    Li, Y.-Z

    Z.-Z. Li, Y.-Z. Qian, J. Han, T. S. Li, W. Wang, and Y. P. Jing, Astrophys. J.894, 10 (2020), arXiv:2001.09136

    Pith/arXiv arXiv 2020

  39. [40]

    Birdet al., Mon

    S. Birdet al., Mon. Not. R. Astron. Soc.516, 731 (2022), arXiv:2205.02860

    Pith/arXiv arXiv 2022

  40. [41]

    Benisty, E

    D. Benisty, E. Vasiliev, N. W. Evans, A.-C. Davis, O. V. Hartl, and L. E. Strigari, Astrophys. J. Lett.928, L5 (2022), arXiv:2202.00033 [astro-ph.GA]. Appendix A: Source Data for Local-Universe Mass Estimates For clarity and reproducibility, this appendix lists the numerical source data underlying the Local-Universe ap- plications discussed in Chapter IV....

    Pith/arXiv arXiv 2022