Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Normalizing flows trained to reproduce weak-lensing convergence maps match the mean and variance of summary statistics, but systematically underestimate the off-diagonal elements of the covariance matrix by up to 25 percent unless mitigatio

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

2026-08-03 07:16 UTC pith:LSMA3G5F

load-bearing objection Useful qualitative warning about NF covariances, but the mitigation claims are undermined by an area mismatch and missing sampling errors. the 4 major comments →

arxiv 2601.20669 v2 pith:LSMA3G5F submitted 2026-01-28 astro-ph.CO

Replicating weak-lensing summary-statistic covariances with normalizing flows

Joaquin Armijo , Leander Thiele , Jia Liu This is my paper
classification astro-ph.CO
keywords weak lensingconvergence mapsnormalizing flowsgenerative modelssummary statisticscovariance estimationcosmic shearMinkowski functionals
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.

This paper tests whether normalizing-flow generative models can reproduce not just the mean and variance but the full covariance structure of weak-lensing summary statistics. Using a suite of 954 simulated convergence maps, the authors train flows at two levels: on the binned data vectors themselves and on full pixel maps. They find that means and variances match the simulations to within about a percent, but off-diagonal covariance elements come out systematically low by roughly 5 to 25 percent depending on the statistic and map size. They show that enlarging the network, augmenting the training data by cropping and rotating, and adding pixel noise all improve covariance recovery, bringing power-spectrum off-diagonals to about 5 percent. The conclusion is a warning: a generative model that passes mean and variance checks can still misestimate covariances, which would bias cosmological error bars.

Core claim

The paper's central claim is that the fidelity of a generative model cannot be judged by means and variances alone. When normalizing flows are trained on the data vectors of the convergence power spectrum, the one-point PDF, and the Minkowski functionals, the recovered means and standard deviations agree with the simulations at the percent level, yet the off-diagonal terms of the covariance matrix—defined through the ratio rho' = sqrt(C_NF/C_Sim)—are underestimated by roughly 5% for the power spectrum and PDF and by about 10% for the Minkowski functionals. Training on full convergence maps makes the deficit worse, growing with map resolution up to differences of 25% or more on 256x256 maps.

What carries the argument

The central machinery is the normalizing flow itself, used in two regimes: a neural spline flow that learns the joint distribution of binned summary-statistic vectors, and a multiscale flow that learns the pixel distribution of whole convergence maps by transforming Gaussian latent noise through invertible coupling layers. The diagnostic that carries the argument is the ratio rho' = sqrt(C_NF/C_Sim) of NF-generated to simulation covariances, taken elementwise for off-diagonal entries; it isolates covariance deficits without being masked by variance differences. Mitigation is tested through three levers: network depth (more squeeze-and-split multiscale levels), data augmentation (random 5x5 d

Load-bearing premise

The load-bearing assumption is that the covariance measured from the 954 simulated light cones is a noiseless ground truth; if its sampling uncertainty is comparable to the quoted 5-25% deficits, the central quantitative claim is not established.

What would settle it

Compute the sampling error on the reference covariance, for example by bootstrap or jackknife resampling the 954 simulations, and check whether the ratio rho' - 1 departs from zero by more than the resulting error bars. If the ground-truth covariance noise alone can account for the reported offsets, the claim of systematic underestimation collapses.

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

If this is right

  • Users of generative synthetic maps must validate off-diagonal covariances before using them for cosmological inference, since means and variances alone are insufficient.
  • If unmitigated, underestimated covariances would shrink error bars and produce overconfident parameter contours in weak-lensing analyses.
  • Injecting survey-like pixel noise during training is an effective regularization that improves covariance recovery, suggesting noise should be treated as part of the training strategy.
  • The failure worsens with map size for fixed network capacity, so resolution pushes the need for deeper multiscale architectures.
  • The non-monotonic accuracy-versus-training-size curve implies that more training data does not monotonically improve covariance fidelity; small training sets can mask the deficit through added stochasticity.

Where Pith is reading between the lines

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

  • The same diagnostic likely applies to other generative models such as diffusion models and GANs; any generative model that matches marginals may still misestimate joint covariances, so covariances should be checked before relying on synthetic mocks.
  • The crop-based augmentation changes the effective survey window from 10x10 to 5x5 degrees, which may alter large-scale modes; a cleaner test would compare against a full-resolution augmented sample to separate augmentation effects from resolution trade-offs.
  • The reported deficits could be partly attributable to finite-sample noise in the ground-truth covariance; propagating the sampling uncertainty of the reference covariance would sharpen or weaken the quantitative claims.
  • If the 25% deficit is generic, it motivates developing flow architectures that are explicitly trained to match second-order statistics, or post-hoc covariance rescaling.

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 / 4 minor

Summary. The paper trains normalizing flows (an NSF on summary-statistic vectors and a multiscale NF on convergence maps) on the 954-map SLICS suite and assesses how well the generated ensembles reproduce the mean, standard deviation, and covariance of the angular power spectrum, the convergence PDF, and the Minkowski functionals. It reports percent-level accuracy for means and variances, but finds that off-diagonal covariances are underestimated unless mitigation strategies are applied. The claimed mitigation strategies are larger network size, data augmentation, and training with added Gaussian pixel noise, which the paper reports improves power-spectrum covariance recovery to O(5%). The central cautionary message is that generative-model users must validate off-diagonal covariance fidelity before using synthetic maps for inference.

Significance. If established, this is a useful and timely diagnostic for the weak-lensing generative-model literature: agreement on means and variances is not sufficient to guarantee covariance fidelity, and off-diagonal covariance errors can bias inferred contours. The paper has clear strengths: it uses the public SLICS suite, tests several summary statistics, uses exact-likelihood normalizing flows, and explicitly explores network size, augmentation, and noise. However, the quantitative claims are not yet rigorous because the mitigation comparison uses an inconsistent sky-area reference, the noisy-field comparison is asymmetric, and no sampling-error bars are attached to the central covariance ratios. The qualitative conclusion is plausible, but the specific 5-25% numbers require a matched and noise-budgeted comparison.

major comments (4)
  1. [§V.C, Figs. 7 and 9] The mitigation comparison is not matched in sky area. The SLICS reference covariance is computed from the original 10×10 deg² maps, while the text states that the NF-generated maps in this section have angular size 5×5 deg² (from cropped training images). For a fixed angular-mode binning, the sample covariance of C_ell scales roughly as 1/f_sky, so the 5×5 target distribution has a substantially larger covariance than the 10×10 reference. The reported improvement to ~5% off-diagonal accuracy may therefore be an artifact of the cropping-induced covariance increase rather than of augmentation or noise. A matched 5×5 SLICS covariance (e.g., computed from four crops per 10×10 map) must be used as the reference, or the generated maps must be compared at 10×10.
  2. [§V.A, definition of ρ′ and Figs. 3, 5, 7, 8, 9] All covariance ratio heatmaps and the Figure 9 accuracy curve are reported without sampling-error bars on either C_Sim (954 SLICS realizations) or C_NF (~4000 generated maps). The finite-sample error on a diagonal covariance element is roughly sqrt(2/(954-1)) ≈ 4.6%, and off-diagonal errors are generally larger. The quoted 5% mitigation gain is comparable to this noise level. Without bootstrap/jackknife uncertainties or an analytic error estimate, the central quantitative deficits (5-25%, or up to 75% in §V.B) are not established.
  3. [§V.C, Figs. 6-8] The noisy-field training comparison is not apples-to-apples. Gaussian pixel noise with σκ=0.008 is added to the generated maps, but no statement is made that the same noise was added to the SLICS maps before computing the reference summary statistics. If the reference remains noiseless, the extra stochasticity in the generated ensemble can inflate variances and covariances, artificially improving the ratio ρ′. The paper must either noise the ground truth identically or quantify and subtract the noise contribution before claiming that noise training improves covariance recovery.
  4. [Abstract vs §V.B and §V.C] The paper's headline numbers are internally inconsistent. §V.B states that for the largest maps (256×256), variance terms are at least 25% smaller and discrepancies reach up to 75%, while the abstract quotes off-diagonal underestimation 'up to ~25%'. The mitigation gain is quoted as ~5% in §V.C and Conclusions, but as O(10%) in the abstract/full text. These discrepancies need resolution because the abstract's quantitative claim is the paper's central message.
minor comments (4)
  1. [Section VI] There is a typo: 'we to study the replication' should be 'we study the replication'.
  2. [§V.C] Typos: 'spatial the correlations' and 'understimated/overstimated' should be corrected. The sentence 'In this regime, the stochasticity in the generated maps arising can introduce an additional scatter...' is grammatically unclear and should be rewritten.
  3. [Figure 9] The definition of 'off-diagonal accuracy' (averaging over off-diagonal elements) should be stated in the main text, not only in the caption, with the exact multipole bins and weighting specified. Adding markers for the six training-set sizes would improve readability.
  4. [General] No code or data availability statement is provided. Given the reproducibility emphasis of the paper, a link to trained models or generation scripts would strengthen the contribution.

Circularity Check

0 steps flagged

No circular derivation: the NF fidelity test is an internal benchmark against the same SLICS suite, not a fitted parameter renamed as a prediction.

full rationale

The paper's central claim is that normalizing flows trained on SLICS convergence maps reproduce means and variances well while underestimating covariance off-diagonals, and that augmentation/noise improve recovery. This is an internal fidelity test: the flow is trained on SLICS data and the generated summary statistics are compared to the same SLICS suite as ground truth. No parameter is fitted to a subset and then 'predicted' on a closely related quantity; the reported ratios rho' = sqrt(C_NF_ij / C_Sim_ij) are direct comparisons, not quantities determined by construction. The mitigation strategies (network size, data augmentation, pixel noise) are interventions on the training procedure, and their effect is measured against the same SLICS covariance; even if the comparison is imperfect (e.g., the 5x5 deg^2 augmented maps are compared to the full 10x10 deg^2 SLICS covariance in Fig. 9, and finite-sample noise on the 954-realization ground truth is not propagated), these are methodological/statistical concerns, not circularity. The paper explicitly flags limitations: 'Some of these values might be expected... covariance [is] more sensitive to sampling noise' (Sec. V.A) and 'we emphasize how the limited information used for training can compromise the accuracy of the off-diagonal terms' (Sec. VI). Self-citations appear (e.g., [42] for MF+S8 constraints, [38] for PDF constraints), but they are background context and not load-bearing for the NF covariance results. No uniqueness theorem, ansatz-by-citation, or renaming of a known result as new is used to force the conclusions. Therefore the derivation chain is self-contained for what it claims; the central results do not reduce to their inputs by definition.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper's quantitative claims rest on simulation fidelity, chosen augmentation/noise hyperparameters, and an unquantified ground-truth covariance. No new physical entities are introduced.

free parameters (5)
  • Pixel noise standard deviation σκ = 0.008
    Chosen by hand to mimic LSST-like survey; drives the claimed 5% covariance improvement in §V.C.
  • Data augmentation crop size = 5×5 deg²
    Random crops from 10×10 deg² SLICS maps; changes large-scale modes and the ℓ-range used for comparison.
  • NF network parameter counts = MS1 518K / MS2 665K / MS3 893K
    Architecture sizes chosen by hand; covariance accuracy depends on them.
  • Angular smoothing kernel = 2 arcmin
    Applied before computing PDF and MFs in Figure 6; affects the resulting covariance values.
  • Number of generated maps for evaluation = 4000
    Chosen for the noisy-map analysis; the finite sample size affects the estimated covariance noise.
axioms (4)
  • domain assumption SLICS N-body light-cone maps are a valid ground truth for weak-lensing convergence statistics
    Entered in §III; all comparisons treat SLICS estimates as truth.
  • domain assumption Rotation by 90°/180° and cropping preserve the target summary-statistic distributions
    Assumed in §V.C for data augmentation; relies on statistical isotropy and scale separability.
  • ad hoc to paper Gaussian pixel noise with σκ=0.008 approximates LSST survey noise and acts as a useful regularizer
    Introduced in §V.C; not justified by a detailed survey noise model beyond an order-of-magnitude choice.
  • domain assumption Normalizing flows with the described architectures converge to the target map distribution given enough training samples
    Standard ML background assumption invoked throughout §IV-V; the paper's diagnostics depend on it.

pith-pipeline@v1.3.0-alltime-deepseek · 137 in / 12094 out tokens · 168782 ms · 2026-08-03T07:16:29.370582+00:00 · methodology

0 comments
read the original abstract

We explore the ability of normalizing flow (NF) generative models to reproduce weak-lensing summary statistics when trained on a set of cosmological simulations. Our analysis focuses on how accurately NF models recover the mean, standard deviation, and covariance of key statistics derived from convergence ($\kappa$) maps: The angular power spectrum $C_{\ell}$, probability density function, and Minkowski functionals of weak lensing convergence $\kappa$-maps. We test two scenarios for training: (1) on the data vectors and (2) on the full $\kappa$-maps. In both cases, the NF models reproduce the mean and variance of the target statistics within percent-level accuracy. However, the accuracy of the off-diagonal elements of the covariance matrix is underestimated by up to $\sim25\%$. We study several mitigation strategies and find that data augmentation and training with noisy fields help improve covariance recovery to $\mathcal{O}(5\%)$ on power spectrum statistics. Our study demonstrates that while the means and variances of weak lensing statistics can be well modeled by NF, covariances can be significantly underestimated if mitigation strategies are not applied. We present this test as a rigorous diagnostic of generative-model fidelity.

Figures

Figures reproduced from arXiv: 2601.20669 by Jia Liu, Joaquin Armijo, Leander Thiele.

Figure 1
Figure 1. Figure 1: FIG. 1. Diagram of multiscale NF network used for learning [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Replication of summary statistics using NF network and SLICS simulations (ground truth) dataset. We calculate the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Ratio of the covariance matrix of the summary statistic showed in Fig. 2, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Angular Power spectrum [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Ratio of covariance for [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Same as Figure 2 but with NF generated maps using the multiscale architecture. We calculate the summary statistics [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Same as in 7 (noisy data maps), but for PDF and MF covariance. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Accuracy of the off-diagonal covariance terms of the [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Machine-learning applications for weak-lensing cosmology

    astro-ph.CO 2026-05 unverdicted novelty 2.0

    Machine learning techniques can mitigate limitations in traditional weak-lensing analyses and enhance extraction of cosmological information from galaxy imaging surveys.

Reference graph

Works this paper leans on

60 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bartelmann and P

    M. Bartelmann and P. Schneider, Physics Reports340, 291 (2001)

    2001

  2. [2]

    Kilbinger, Reports on Progress in Physics78, 086901 (2015)

    M. Kilbinger, Reports on Progress in Physics78, 086901 (2015)

    2015

  3. [3]

    Mandelbaum, Annual Review of Astronomy and As- trophysics56, 393 (2018)

    R. Mandelbaum, Annual Review of Astronomy and As- trophysics56, 393 (2018)

    2018

  4. [4]

    Huterer, General Relativity and Gravitation42, 2177–2195 (2010)

    D. Huterer, General Relativity and Gravitation42, 2177–2195 (2010)

    2010

  5. [5]

    Huterer and D

    D. Huterer and D. L. Shafer, Reports on Progress in Physics81, 016901 (2018), arXiv:1709.01091 [astro- ph.CO]

    Pith/arXiv arXiv 2018

  6. [6]

    A. e. a. Amon, Physical Review D (2022)

    2022

  7. [7]

    Heymans, T

    C. Heymans, T. Tr¨ oster, M. Asgari, C. Blake, H. Hilde- brandt, B. Joachimi, K. Kuijken, C.-A. Lin, A. G. S´ anchez, J. L. van den Busch, A. H. Wright, A. Amon, M. Bilicki, J. de Jong, M. Crocce, A. Dvornik, T. Erben, M. C. Fortuna, F. Getman, B. Giblin, K. Glazebrook, H. Hoekstra, S. Joudaki, A. Kannawadi, F. K¨ ohlinger, C. Lidman, L. Miller, N. R. Nap...

    Pith/arXiv arXiv 2021

  8. [8]

    C. e. a. Hikage, Publications of the Astronomical Society of Japan71, 43 (2019)

    2019

  9. [9]

    R. e. a. Laureijs, Euclid Definition Study Report, Tech. Rep. 2011-12 (ESA/SRE, 2011)

    2011

  10. [10]

    L. S. Collaboration, LSST Science Book, Version 2.0, Tech. Rep. (Rubin Observatory, 2009) arXiv:0912.0201

    Pith/arXiv arXiv 2009

  11. [11]

    Kaiser and G

    N. Kaiser and G. Squires, The Astrophysical Journal 404, 441 (1993)

    1993

  12. [12]

    R. e. a. Takahashi, The Astrophysical Journal850, 24 (2017)

    2017

  13. [13]

    Vale and M

    C. Vale and M. White, ApJ592, 699 (2003), arXiv:astro- ph/0303555 [astro-ph]

    arXiv 2003

  14. [14]

    Pires, J

    S. Pires, J. L. Starck, A. Amara, R. Teyssier, A. R´ efr´ egier, and J. Fadili, MNRAS395, 1265 (2009), arXiv:0804.4068 [astro-ph]

    Pith/arXiv arXiv 2009

  15. [15]

    N. e. a. Jeffrey, Monthly Notices of the Royal Astronom- ical Society505, 4626 (2021)

    2021

  16. [16]

    Massey, H

    R. Massey, H. Hoekstra, T. Kitching, J. Rhodes, M. Cropper, J. Amiaux, D. Harvey, Y. Mellier, M. Meneghetti, L. Miller, S. Paulin-Henriksson, S. Pires, R. Scaramella, and T. Schrabback, MNRAS429, 661 (2013), arXiv:1210.7690 [astro-ph.CO]

    Pith/arXiv arXiv 2013

  17. [17]

    Harnois-D´ erapset al., Monthly Notices of the Royal Astronomical Society481, 1337 (2018)

    J. Harnois-D´ erapset al., Monthly Notices of the Royal Astronomical Society481, 1337 (2018)

    2018

  18. [18]

    Tr¨ oster, C

    T. Tr¨ oster, C. Ferguson, J. Harnois-D´ eraps, and I. G. McCarthy, MNRAS487, L24 (2019), arXiv:1903.12173 [astro-ph.CO]

    Pith/arXiv arXiv 2019

  19. [19]

    A. C. Rodr ´ ıguez, T. Kacprzak, A. Lucchi, A. Amara, R. Sgier, J. Fluri, T. Hofmann, and A. R´ efr´ egier, Com- putational Astrophysics and Cosmology5, 4 (2018), arXiv:1801.09070 [astro-ph.CO]

    Pith/arXiv arXiv 2018

  20. [20]

    S. e. a. Ravanbakhsh, Proceedings of the 34th Interna- tional Conference on Machine Learning (2017)

    2017

  21. [21]

    Mustafa, D

    M. Mustafa, D. Bard, W. Bhimji, Z. Luki´ c, R. Al-Rfou, and J. M. Kratochvil, Computational Astrophysics and Cosmology6, 1 (2019), arXiv:1706.02390 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  22. [22]

    Fluri, T

    J. Fluri, T. Kacprzak, A. Lucchi, A. Amara, et al., Phys- ical Review D100, 063514 (2019)

    2019

  23. [23]

    Dai and U

    B. Dai and U. Seljak, in Machine Learning for Astrophysics (2023) p. 10, arXiv:2306.04689 [astro-ph.CO]

    Pith/arXiv arXiv 2023

  24. [24]

    Whitney, T

    J. Whitney, T. Liaudat, M. Price, M. Mars, and J. D. McEwen, arXiv e-prints , arXiv:2406.15424 (2024), arXiv:2406.15424 [astro-ph.CO]

    Pith/arXiv arXiv 2024

  25. [25]

    J. J. Whitney, T. I. Liaudat, M. A. Price, M. Mars, and J. D. McEwen, arXiv e-prints , arXiv:2410.24197 (2024), arXiv:2410.24197 [astro-ph.CO]

    arXiv 2024

  26. [26]

    S. S. Boruah, M. Jacob, and B. Jain, Phys. Rev. D111, 083542 (2025), arXiv:2502.04158 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  27. [27]

    Gupta, J

    A. Gupta, J. M. Z. Matilla, D. Hsu, and Z. Haiman, Phys. Rev. D97, 103515 (2018)

    2018

  28. [28]

    Perraudin, S

    N. Perraudin, S. Marcon, A. Lucchi, and T. Kacprzak, arXiv e-prints , arXiv:2004.08139 (2020), arXiv:2004.08139 [astro-ph.CO]

    Pith/arXiv arXiv 2004

  29. [29]

    F. e. a. Villaescusa-Navarro, The Astrophysical Journal 915, 71 (2021)

    2021

  30. [30]

    Dai and U

    B. Dai and U. Seljak, MNRAS516, 2363 (2022), arXiv:2202.05282 [astro-ph.CO]

    Pith/arXiv arXiv 2022

  31. [31]

    Amon et al., Physical Review D105, 023520 (2022)

    A. Amon et al., Physical Review D105, 023520 (2022)

    2022

  32. [32]

    Asgari et al., Astronomy & Astrophysics645, A104 (2021)

    M. Asgari et al., Astronomy & Astrophysics645, A104 (2021)

    2021

  33. [33]

    have provided precise measurements of cosmic struc- ture. A. Summary statistics Extracting cosmological information from weak- lensing convergence maps relies on summary statistics that quantify the spatial structure of the projected mat- ter density field. Althoughκis a line-of-sight projection of the underlying matter distribution, it becomes signifi- c...

    2000

  34. [34]

    Hikage et al., Publications of the Astronomical Society of Japan71, 43 (2019)

    C. Hikage et al., Publications of the Astronomical Society of Japan71, 43 (2019)

    2019

  35. [35]

    Munshi and B

    D. Munshi and B. Jain, MNRAS318, 109 (2000), arXiv:astro-ph/9911502 [astro-ph]

    Pith/arXiv arXiv 2000

  36. [36]

    Petri, Physical Review D94, 063534 (2016)

    A. Petri, Physical Review D94, 063534 (2016)

    2016

  37. [37]

    Gatti, B

    M. Gatti, B. Jain, C. Chang, M. Raveri, D. Z¨ urcher, L. Secco, L. Whiteway, N. Jeffrey, C. Doux, T. Kacprzak, D. Bacon, P. Fosalba, A. Alarcon, A. Amon, K. Bech- tol, M. Becker, G. Bernstein, J. Blazek, A. Campos, A. Choi, C. Davis, J. Derose, S. Dodelson, F. Elsner, J. Elvin-Poole, S. Everett, A. Ferte, D. Gruen, I. Har- rison, D. Huterer, M. Jarvis, E....

    Pith/arXiv arXiv 2022

  38. [38]

    Uhlemann, O

    C. Uhlemann, O. Friedrich, F. Villaescusa-Navarro, A. Banerjee, and S. Codis, MNRAS495, 4006 (2020), arXiv:1911.11158 [astro-ph.CO]

    Pith/arXiv arXiv 2020

  39. [39]

    Thiele, G

    L. Thiele, G. A. Marques, J. Liu, and M. Shirasaki, Phys. Rev. D108, 123526 (2023)

    2023

  40. [40]

    Castiblanco, C

    L. Castiblanco, C. Uhlemann, J. Harnois-D´ eraps, and 12 A. Barthelemy, The Open Journal of Astrophysics7, 59 (2024), arXiv:2405.09651 [astro-ph.CO]

    arXiv 2024

  41. [41]

    J. M. Kratochvil, E. A. Lim, and S. Wang, Physical Re- view D85, 103513 (2012)

    2012

  42. [42]

    Munshi et al., Monthly Notices of the Royal Astro- nomical Society419, 536 (2012)

    D. Munshi et al., Monthly Notices of the Royal Astro- nomical Society419, 536 (2012)

    2012

  43. [43]

    Armijo, G

    J. Armijo, G. A. Marques, C. P. Novaes, L. Thiele, J. A. Cowell, D. Grand´ on, M. Shi- rasaki, and J. Liu, Monthly Notices of the Royal Astronomical Society , staf257 (2025), https://academic.oup.com/mnras/advance-article- pdf/doi/10.1093/mnras/staf257/61870752/staf257.pdf

    work page doi:10.1093/mnras/staf257/61870752/staf257.pdf 2025

  44. [44]

    G. A. Marques, J. Liu, M. Shirasaki, L. Thiele, D. Grand´ on, K. M. Huffenberger, S. Cheng, J. Harnois- D´ eraps, K. Osato, and W. R. Coulton, MNRAS528, 4513 (2024), arXiv:2308.10866 [astro-ph.CO]

    Pith/arXiv arXiv 2024

  45. [45]

    Cheng, G

    S. Cheng, G. A. Marques, D. Grand´ on, L. Thiele, M. Shirasaki, B. M´ enard, and J. Liu, arXiv e-prints , arXiv:2404.16085 (2024), arXiv:2404.16085 [astro- ph.CO]

    Pith/arXiv arXiv 2024

  46. [46]

    C. P. Novaes, L. Thiele, J. Armijo, S. Cheng, J. A. Cow- ell, G. A. Marques, E. G. M. Ferreira, M. Shirasaki, K. Osato, and J. Liu, Phys. Rev. D111, 083510 (2025), arXiv:2409.01301 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  47. [47]

    Durkan, A

    C. Durkan, A. Bekasov, I. Murray, and G. Papa- makarios, arXiv e-prints , arXiv:1906.04032 (2019), arXiv:1906.04032 [stat.ML]

    Pith/arXiv arXiv 1906

  48. [48]

    Papamakarios, T

    G. Papamakarios, T. Pavlakou, and I. Murray, arXiv e-prints , arXiv:1705.07057 (2017), arXiv:1705.07057 [stat.ML]

    Pith/arXiv arXiv 2017

  49. [49]

    L. Dinh, J. Sohl-Dickstein, and S. Bengio, arXiv e-prints , arXiv:1605.08803 (2016), arXiv:1605.08803 [cs.LG]

    Pith/arXiv arXiv 2016

  50. [50]

    D. P. Kingma and P. Dhariwal, arXiv e-prints , arXiv:1807.03039 (2018), arXiv:1807.03039 [stat.ML]

    Pith/arXiv arXiv 2018

  51. [51]

    Ribli, B

    D. Ribli, B. ´A. Pataki, J. M. Zorrilla Matilla, D. Hsu, Z. Haiman, and I. Csabai, MNRAS490, 1843 (2019), arXiv:1902.03663 [astro-ph.CO]

    Pith/arXiv arXiv 2019

  52. [52]

    A. C. Pope and I. Szapudi, MNRAS389, 766 (2008), arXiv:0711.2509 [astro-ph]

    Pith/arXiv arXiv 2008

  53. [53]

    Petri, Z

    A. Petri, Z. Haiman, and M. May, Phys. Rev. D93, 063524 (2016), arXiv:1601.06792 [astro-ph.CO]

    Pith/arXiv arXiv 2016

  54. [54]

    D. W. Pearson and L. Samushia, Monthly No- tices of the Royal Astronomical Society457, 993 (2016), https://academic.oup.com/mnras/article- pdf/457/1/993/18169038/stw062.pdf

    2016

  55. [55]

    Shirasaki, N

    M. Shirasaki, N. Yoshida, and S. Ikeda, Phys. Rev. D 100, 043527 (2019)

    2019

  56. [56]

    Zhong, M

    K. Zhong, M. Gatti, and B. Jain, Phys. Rev. D110, 043535 (2024), arXiv:2403.01368 [astro-ph.CO]

    Pith/arXiv arXiv 2024

  57. [57]

    S. S. Boruah, P. Fiedorowicz, R. Garcia, W. R. Coulton, E. Rozo, and G. Fabbian, arXiv e-prints , arXiv:2406.05867 (2024), arXiv:2406.05867 [astro- ph.CO]

    Pith/arXiv arXiv 2024

  58. [58]

    Shirasaki and S

    M. Shirasaki and S. Ikeda, The Open Journal of Astro- physics7, 42 (2024), arXiv:2310.17141 [astro-ph.CO]

    Pith/arXiv arXiv 2024

  59. [59]

    S. D. Aoyama, K. Osato, and M. Shirasaki, arXiv e-prints , arXiv:2505.00345 (2025), arXiv:2505.00345 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  60. [60]

    Shirasaki, K

    M. Shirasaki, K. Moriwaki, T. Oogi, N. Yoshida, S. Ikeda, and T. Nishimichi, MNRAS504, 1825 (2021), arXiv:1911.12890 [astro-ph.CO]

    Pith/arXiv arXiv 2021