Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Learning Correlated Astrophysical Foregrounds with Denoising Diffusion Probabilistic Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A denoising diffusion model trained on simulated CIB-tSZ patches reproduces the foregrounds' two-, three-, and four-point statistics, generating new correlated map pairs in seconds.

desk verdict Useful new DDPM application for CIB+tSZ foregrounds, but the validation may be leaking: the train/validation patch split is not shown to be spatially disjoint, and that weakens the headline fidelity claim. read the letter →

arxiv 2506.09036 v3 pith:C6DWP2JE submitted 2025-06-10 astro-ph.CO

Karthik Prabhu , Srinivasan Raghunathan , Ethan B. Anderes , Lloyd E. Knox This is my paper
classification astro-ph.CO
keywords cosmicinfraredbackgroundthermalSunyaev-Zel'dovicheffectdenoisingdiffusionprobabilisticmodelsCMBforegroundsnon-GaussianstatisticsbispectrumtrispectrumMinkowskifunctionals
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

The paper sets out to show that a denoising diffusion probabilistic model (DDPM), a neural network trained to reverse a gradual noising process, can learn the joint distribution of two correlated extragalactic foregrounds that contaminate small-scale cosmic microwave background (CMB) measurements: the cosmic infrared background (CIB) and the thermal Sunyaev-Zel'dovich (tSZ) effect. Using 150 GHz patches from the Agora simulation suite as training data, the authors claim the model generates new, physically correlated CIB-tSZ map pairs in seconds that match the simulations' power spectra, bispectra, trispectra, pixel histograms, and Minkowski functionals. If correct, this turns a process that takes thousands of CPU-hours per N-body realization into a fast sampler, making it practical to build large ensembles for covariance estimation, bias tests, and forward-modeling in CMB lensing and small-scale temperature analyses. The paper also shows a multi-frequency extension that learns spectral energy distributions across bands.

What carries the argument

The engine is the denoising diffusion probabilistic model (DDPM), in which a forward process adds Gaussian noise to a $256\times256\times2$ patch (CIB and tSZ channels) over 1000 steps and a U-Net with self-attention learns the reverse denoising operation, approximating the score function $\nabla \log p(x)$ of the joint data distribution. The reverse process is trained with a velocity-prediction objective and does not directly optimize any of the summary statistics used for evaluation. A post-training global variance rescaling, multiplying each generated patch by the ratio of target to generated standard deviation, compensates for the model's tendency to underproduce rare bright pixels and is credited with substantially improving agreement across all statistics.

What would settle it

Train the identical DDPM on an independent correlated foreground simulation, or compare 150 GHz DDPM-generated CIB and tSZ statistics directly to observed maps from a current small-scale CMB experiment after matching masks, beams, and filters. If the generated bispectra, trispectra, or stacked tSZ profiles deviate from the target beyond the sample variance shown in the paper, the claim that the model has learned the true joint foreground distribution is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a DDPM trained on paired CIB and tSZ patches can reproduce the non-Gaussian joint statistics of the training simulation well enough to be used as a fast generative surrogate. Concretely, for the auto- and cross-power spectra the generated maps agree with Agora within about 25% of sample variance, the collapsed equilateral bispectrum and trispectrum of the summed CIB+tSZ+noise fields match the simulation, and the one-point pixel distributions and Minkowski functionals are recovered better than Gaussian realizations matched to the same power spectra. A single global rescaling of generated pixel intensities, by factors 1.0328 for CIB and 1.1425 for tSZ, corrects a systematic underproduction of rare extreme-value pixels and restores the power-spectrum amplitudes. These results are demonstrated against hold-out Agora patches, not against observed sky data.

Load-bearing premise

The load-bearing premise is that the Agora simulation, which supplies all training examples, is a faithful enough model of the true CIB and tSZ sky at 150 GHz; the paper validates the DDPM only against Agora itself, so any mismatch between the simulation's cosmological and astrophysical assumptions and the real sky would be inherited by the generated samples.

Editorial extensions

If this is right

  • Large ensembles of correlated CIB-tSZ patches can be generated in seconds, making it practical to compute covariance matrices and bias tests that currently require thousands of expensive N-body realizations.
  • The trained DDPM can serve as a non-Gaussian, physically motivated prior or as a forward simulator inside Bayesian inference pipelines, replacing Gaussian foreground assumptions in CMB analyses.
  • The multi-frequency result indicates that the framework can learn spectral energy distributions and inter-frequency decorrelations, which is directly relevant to component separation in multi-band CMB surveys.
  • The model's documented weakness, underproducing rare high-intensity pixels, is partially corrected by a simple global rescaling, but remains a limitation for analyses that depend on extreme-value statistics.
  • Scaling to the full survey area is not yet demonstrated: training on 6°x6° patches works, while larger patches degrade, so extending the method to pipeline-sized sky coverage requires further work.

Reading between the lines

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

  • If the joint distribution is captured as claimed, the same pipeline should extend to the other correlated components in the simulation, such as kSZ, lensing convergence, and radio galaxies, yielding a single fast joint sky simulator; the paper does not test this.
  • Conditioning the generative model on cosmological and astrophysical parameters, for example $\sigma_8$ or neutrino mass, would turn it into a foreground emulator able to marginalize over modeling uncertainties, an extension the authors mention but do not implement.
  • The success of a single global rescaling in fixing most of the power-spectrum deficit suggests the rare-event underproduction is a low-dimensional systematic of diffusion models on this data; testing a tail-aware loss or importance sampling could remove the need for post-hoc rescaling.
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.

Referee Report

4 major / 5 minor

Summary. This paper applies a denoising diffusion probabilistic model (DDPM) to jointly generate cosmic infrared background (CIB) and thermal Sunyaev-Zel'dovich (tSZ) maps at 150 GHz, using paired 6°×6° patches extracted from the Agora simulation after point-source and cluster masking and low-pass filtering. The authors compare DDPM samples with Agora “hold-out” patches through visual inspection, stacked tSZ profiles, auto- and cross-power spectra, pixel-intensity histograms, Minkowski functionals, and collapsed equilateral bispectra and trispectra. After applying a global per-channel variance rescaling to compensate for underproduction of extreme pixels, they report agreement within sample variance for two-point statistics and good reproduction of non-Gaussian statistics, while noting residual discrepancies in distribution tails and at low multipoles of the trispectrum. An extension generates CIB maps at 95, 150, and 857 GHz and checks inter-frequency correlations.

Significance. If substantiated, the result would provide a fast generative model for correlated extragalactic foregrounds, a useful complement to expensive N-body-based simulations and relevant to CMB lensing and small-scale temperature analyses. The paper is transparent about its main limitations—the underrepresentation of rare high-intensity pixels, the need for post-hoc rescaling, the dependence on Agora's assumptions, and the difficulty of scaling to larger sky areas—and it makes code available. The multi-frequency CIB demonstration is a useful proof of concept. However, the significance of the headline claim is conditional on the validation being truly independent of the training data and on the two-point agreement not being enforced by the calibration of the rescaling factor; both points require attention before the abstract's 2/3/4-point claim can be taken at face value.

major comments (4)
  1. [Sec. 2.1 / Appendix A] The data preparation as described does not rule out train/validation leakage. The text states that 6°×6° patches are extracted from a full-sky HEALPix map, randomly rotated/flipped, and then split 80/20, but it does not state that the extraction is a non-overlapping tiling or that augmentation is applied after the split. Since ~4,000 augmented patches of 36 deg² each cover ~3.5 times the full sky, either the base patches overlap or augmented copies are used; in either case validation patches may share sky area or exact pixel patterns with training patches. All of the distribution-level comparisons in Sec. 4 use these “hold-out patches” (Sec. 3.2), so the reported 2/3/4-point agreement could be partly due to memorization rather than generalization. Please specify the patch sampling scheme, perform the split on base patches before augmentation (or otherwise ensure spatial disjointness), and re-run the validation comparisons; this is load-bearing for the central empirical claim.
  2. [Sec. 3.2 / Sec. 4.3] The post-training rescaling factors 1.0328 (CIB) and 1.1425 (tSZ) are computed as the ratio of standard deviations of all Agora samples to all generated samples. Because the same Agora samples are then used as the validation target for the power-spectrum comparison in Sec. 4.3, the reported agreement of the two-point amplitude after rescaling is guaranteed at zero lag (total variance) by construction and is therefore not an independent test of the model's two-point fidelity. The higher-order statistics are not directly optimized and remain informative, but the abstract's “faithfully reproduce … 2-point” claim should be qualified, and the paper should report the pre-rescaling power spectra or calibrate the rescaling on a separate subset.
  3. [Sec. 4.6 / Fig. 7 / Sec. 6] The paper states that “minor differences are present between the trispectrum of Agora and the DDPM generated samples at low multipoles” and that masking extreme-value pixels does not noticeably change the discrepancy. This is in tension with the abstract's claim of faithful reproduction of the 4-point correlation functions and with the unqualified statement in the Conclusions. Please quantify the low-multipole discrepancy (e.g., significance in units of standard error at each ℓ_c) and, based on that, either soften the headline claim or explain why the difference is ignorable for the intended applications.
  4. [Sec. 2.1 / Sec. 4.2] The stacked tSZ profile validation is ambiguous because the maps used for training and validation have massive clusters masked and inpainted with Gaussian random values (Sec. 2.1). If the 5–20σ and ≥20σ stacking is performed on the same masked maps, the high-SNR bins will be dominated by the inpainting rather than astrophysical cluster signal; if it is performed on the original unmasked maps, the DDPM was not trained on those clusters. The text should specify which maps enter the stacking, and the 19% discrepancy for ≥20σ should be interpreted accordingly.
minor comments (5)
  1. [Abstract / Appendix B] The phrase “learn the spectral energy distributions (SEDs) across different bands” is stronger than what is demonstrated; Appendix B models CIB maps at three frequencies and compares correlation coefficients, not SEDs. Please reword or add a proper SED test.
  2. [Sec. B / Fig. 9] The sentence “The impact of this masking is only marginal and well within the errors shown in Fig. 8” should refer to the left panel of Fig. 9, since Fig. 8 displays example maps rather than errors.
  3. [Sec. 4.3] The phrase “within 25% of the sample variance” is ambiguous; please state whether residuals are below 0.25σ(C_ℓ) per bin and over which multipole range this holds.
  4. [Sec. 2.1] The conversion from HEALPix full-sky maps to flat-sky 6° patches is not described (projection scheme, pixelization, treatment of map edges); please specify this, since distortion and boundary effects differ across the sky.
  5. [Sec. 5.3] The phrase “resolved dusty galaxies (CIB)” is imprecise at the 1.4′ pixel scale used here; consider “compact dusty sources” or similar.

Circularity Check

1 steps flagged · score 3.0 of 10

Power-spectrum validation is partly by construction through a global variance-rescaling fit, but the 3- and 4-point statistics remain non-circular.

  1. fitted input called prediction [Section 3.2, 'Summary Statistics'; applied in Section 4.3, 'Power Spectra comparison']
    "To mitigate this, we apply a post-training rescaling of the pixel intensities. Specifically, we multiply each DDPM sample by a single global factor: the ratio of the standard deviation of all the Agora samples to that of all the generated samples. For the CIB maps, we find a rescaling factor of 1.0328, while for the tSZ maps, we obtain 1.1425. … After applying variance rescaling as described in Section 3.2 to the DDPM outputs, to match the true variance, we find that the generated power spectra match the ground truth within 25% of the sample variance across all multipoles."

    The global rescaling factor is fit to the ratio of the standard deviations of the Agora and DDPM samples, so the total variance of the DDPM maps is made equal to the Agora variance by construction. The angular power spectrum integrates to the map variance, so the overall amplitude of the reported C_l agreement is not an independent prediction but a restatement of this fitted normalization. The shape of C_l and all higher-order statistics remain non-forced, so this is partial, not full, circularity.

full rationale

The central DDPM training objective is the ELBO, and the paper explicitly notes that 'the objective does not directly incorporate any of the summary statistics described below' (Section 3.2). Thus the reported matches of the bispectrum, trispectrum, pixel histograms, and Minkowski functionals are genuine, non-circular validations that are not encoded in the loss. The only load-bearing by-construction element is the post-training variance rescaling: fitting a single factor to the ratio of standard deviations forces the total variance (and hence the integrated power-spectrum amplitude) to match, so the Section 4.3 power-spectrum comparison is partly manufactured. This does not reduce the higher-order claims, which remain independent. The possible train/validation overlap is a validation-integrity concern, not a circularity, and the paper's acknowledged reliance on Agora as ground truth (Section 5.2) is an external-fidelity limitation, not a circular reduction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the fidelity of the Agora training simulation and on two post-hoc rescaling parameters fitted to the target data. No new physical entities are introduced.

free parameters (2)
  • CIB variance rescaling factor = 1.0328
    Global multiplier applied to all DDPM CIB samples after training, fitted to the ratio of the standard deviation of Agora samples to generated samples. Restores power-spectrum amplitude (Section 3.2).
  • tSZ variance rescaling factor = 1.1425
    Global multiplier applied to all DDPM tSZ samples after training, fitted to the ratio of standard deviations. Restores power-spectrum amplitude and improves higher-order statistics (Section 3.2).
assumptions (3)
  • domain assumption Agora simulation is a faithful representation of the true CIB and tSZ sky at 150 GHz.
    The DDPM is trained exclusively on Agora patches and validated only against Agora; no real-data validation is provided. The authors acknowledge this in Section 5.2 as a limitation.
  • domain assumption Patches of 6x6 degrees, after masking, low-pass filtering, and inpainting, are representative of full-sky foreground statistics.
    The model is trained on processed patches; scaling to larger areas is not demonstrated and Section 5.2 reports degraded quality for 10-degree patches.
  • domain assumption Gaussian inpainting of masked clusters and point sources does not bias the learned statistics.
    Masked pixels are replaced with Gaussian random values with map-wide mean/std; the paper does not test how this affects the learned joint distribution (Section 2.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Correlated Astrophysical Foregrounds with Denoising Diffusion Probabilistic Models." pith.science (2026). https://pith.science/paper/C6DWP2JE

@misc{pith2026250609036,
  author       = {Pith},
  title        = {Pith review of: Learning Correlated Astrophysical Foregrounds with Denoising Diffusion Probabilistic Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6DWP2JE}},
  note         = {Machine review of arXiv:2506.09036}
}
abstract

Extragalactic foregrounds -- most notably the Cosmic Infrared Background (CIB) and the thermal Sunyaev-Zel'dovich (tSZ) effect -- exhibit complex, non-Gaussian structure and correlations that can bias analyses of small-scale cosmic microwave background (CMB) temperature anisotropies. These foregrounds can introduce mode coupling at small-scales (multipoles $\ell \geq 3000$) that mimic true lensing signals, thereby complicating analyses such as CMB lensing reconstruction. We present a novel approach to learn their full joint distribution using Denoising Diffusion Probabilistic Models (DDPMs) trained on paired CIB-tSZ patches at 150 GHz, from the Agora suite of extragalactic sky simulations. While simulations like Agora, which are based on N-body calculations, can take thousands of CPU hours, DDPM can synthesize realistic CIB-tSZ patches that faithfully reproduce both auto- and cross-spectral statistics of the 2-point, 3-point, and 4-point correlation functions, in a matter of seconds. We further demonstrate matching pixel-value histograms and Minkowski functionals, confirming that conventional non-Gaussian benchmarks are also satisfied. This framework provides a powerful generative tool for forward-modeling correlated extragalactic foregrounds in current and future CMB analyses. Although we mainly demonstrate the joint modeling of tSZ and CIB at a single frequency, we also include examples of its extension to multiple frequencies, showing that the framework can learn the spectral energy distributions (SEDs) across different bands. While establishing DDPMs as a promising tool for addressing foreground contamination in next-generation CMB surveys, we also outline remaining challenges to their practical deployment in analysis pipelines, such as scaling to larger sky areas and reliance on the underlying cosmological and astrophysical assumptions in the simulations used for training.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 27 canonical work pages

  1. [1]

    Kaplinghat, L

    M. Kaplinghat, L. Knox, and Y.-S. Song, Phys. Rev. Lett.91, 241301 (2003)

    work page 2003

  2. [2]

    Horowitz and U

    B. Horowitz and U. Seljak, MNRAS469, 394 (2017), arXiv:1609.01850 [astro-ph.CO] . – 17 – 1000 2000 3000 4000 Multipole 𝓁 −1.00 −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 ∆C𝓁/σC𝓁 95x95 95x150 95x857 150x150 150x857 857x857 1000 2000 3000 4000 Multipole 𝓁 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.05 Correlation Coefficient Agora DDPM Figure 9.Left panel:Stan...

    arXiv 2017

  3. [3]

    Bocquet, S

    S. Bocquet, S. Grandis, L. E. Bleem, others, (SPT, and DES Collaborations), Physical Review D110, 083510 (2024), arXiv:2401.02075 [astro-ph.CO]

    arXiv 2024

  4. [4]

    M. S. Madhavacheril, F. J. Qu, B. D. Sherwin,et al., arXiv e-prints , arXiv:2304.05203 (2023), arXiv:2304.05203 [astro-ph.CO]

    arXiv 2023

  5. [5]

    F. Ge, M. Millea, E. Camphuis, C. Daley, N. Huang, Y. Omori, W. Quan,et al.(SPT-3G Collaboration), Phys. Rev. D111, 083534 (2025)

    work page 2025

  6. [6]

    F. J. Qu, B. D. Sherwin, M. S. Madhavacheril, D. Han, K. T. Crowley,et al., ApJ962, 112 (2024), arXiv:2304.05202 [astro-ph.CO]

    arXiv 2024

  7. [7]

    C. L. Reichardt, S. Patil,et al., ApJ908, 199 (2021), arXiv:2002.06197 [astro-ph.CO]

    arXiv 2021

  8. [8]

    Improved constraints on reionisation from CMB observations: A parameterisation of the kSZ effect

    A. Gorce, S. Ili´ c, M. Douspis, D. Aubert, and M. Langer, A&A640, A90 (2020), arXiv:2004.06616 [astro-ph.CO]

    work page Pith review arXiv 2020

Show all 63 references
  1. [9]

    Raghunathan, P

    S. Raghunathan, P. A. R. Ade, A. J. Anderson,et al.(SPT-3G and SPTpol Collaboration), Phys. Rev. Lett.133, 121004 (2024)

    2024

  2. [10]

    L. E. Bleem, B. Stalder, T. de Haan,et al., ApJS216, 27 (2015), arXiv:1409.0850

    2015 arXiv

  3. [11]

    Planck Collaboration, P. A. R. Ade, N. Aghanim, M. Arnaud,et al., A&A594, A24 (2016), arXiv:1502.01597

    2016 arXiv

  4. [12]

    L. E. Bleem, S. Bocquet, B. Stalder,et al., ApJS247, 25 (2020), arXiv:1910.04121 [astro-ph.CO]

    2020 arXiv

  5. [13]

    Hilton, C

    M. Hilton, C. Sif´ on, S. Naess,et al., ApJS253, 3 (2021), arXiv:2009.11043 [astro-ph.CO]

    2021 arXiv

  6. [14]

    Kornoelje, L

    K. Kornoelje, L. E. Bleem, E. S. Rykoff,et al., arXiv e-prints , arXiv:2503.17271 (2025), arXiv:2503.17271 [astro-ph.CO]

    2025 arXiv

  7. [15]

    Schaan, S

    E. Schaan, S. Ferraro, S. Amodeo,et al., Physical Review D103, 063513 (2021), arXiv:2009.05557 [astro-ph.CO] . – 18 – 0.0 0.5 1.0 1.5 Bispectra S3 [µK3] a = CIB b = tSZSaaa 3 [104] Gaussian Agora DDPM −10 −8 −6 −4 −2 0 Sbbb 3 [104] Noise level: S4-Ultra deep S4-Wide SPT-3G 100...

    2021 arXiv

  8. [16]

    Bigwood, A

    L. Bigwood, A. Amon, A. Schneider,et al., MNRAS534, 655 (2024), arXiv:2404.06098 [astro-ph.CO]

    2024 arXiv

  9. [17]

    Hadzhiyska, S

    B. Hadzhiyska, S. Ferraro, B. Ried Guachalla,et al., arXiv e-prints , arXiv:2407.07152 (2024), arXiv:2407.07152 [astro-ph.CO]

    2024 arXiv

  10. [18]

    Ried Guachalla, E

    B. Ried Guachalla, E. Schaan, B. Hadzhiyska,et al., arXiv e-prints , arXiv:2503.19870 (2025), arXiv:2503.19870 [astro-ph.GA]

    2025

  11. [19]

    M. P. Viero, C. L. Reichardt, B. A. Benson,et al., The Astrophysical Journal881, 96 (2019)

    2019

  12. [20]

    Raghunathan and Y

    S. Raghunathan and Y. Omori, ApJ954, 83 (2023), arXiv:2304.09166 [astro-ph.CO]

    2023 arXiv

  13. [21]

    Sehgal, S

    N. Sehgal, S. Aiola, Y. Akrami,et al., inBulletin of the American Astronomical Society, Vol. 51 (2019) p. 6, arXiv:1906.10134 [astro-ph.CO]

    2019 arXiv

  14. [22]

    Stein, M

    G. Stein, M. A. Alvarez, J. R. Bond, A. v. Engelen, and N. Battaglia, Journal of Cosmology – 19 – −0.5 0.0 0.5 1.0 1.5 2.0 Trispectra S4 [µK4] a = CIB b = tSZSaaaa 4 [107] Gaussian Agora DDPM 1000 2000 3000 4000 5000 Multipole 𝓁c 0 1 2 3 4 5 6 7 8Trispectra S4 [µK4] Sbbbb 4 [1...

    2020

  15. [23]

    Omori, Monthly Notices of the Royal Astronomical Society530, 5030 (2024), https://academic.oup.com/mnras/article-pdf/530/4/5030/57527656/stae1031.pdf

    Y. Omori, Monthly Notices of the Royal Astronomical Society530, 5030 (2024), https://academic.oup.com/mnras/article-pdf/530/4/5030/57527656/stae1031.pdf

    2024

  16. [24]

    D. Han, N. Sehgal, and F. Villaescusa-Navarro, Phys. Rev. D104, 123521 (2021), arXiv:2105.11444 [astro-ph.CO]

    2021 arXiv

  17. [25]

    Mebratu and W

    M. Mebratu and W. L. K. Wu, arXiv e-prints , arXiv:2505.21220 (2025), arXiv:2505.21220 [astro-ph.CO]

    2025 arXiv

  18. [26]

    Fl¨ oss, W

    T. Fl¨ oss, W. R. Coulton, A. J. Duivenvoorden, F. Villaescusa-Navarro, and B. D. Wandelt, MNRAS533, 423 (2024), https://academic.oup.com/mnras/article-pdf/533/1/423/58786991/stae1818.pdf . – 20 –

    2024

  19. [27]

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

    2025 arXiv

  20. [28]

    Mudur and D

    N. Mudur and D. P. Finkbeiner, arXiv e-prints , arXiv:2211.12444 (2022), arXiv:2211.12444 [astro-ph.CO]

    2022 arXiv

  21. [29]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, inProceedings of the 32nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 37, edited by F. Bach and D. Blei (PMLR, Lille, France, 2015) pp. 2256–2265

    2015

  22. [30]

    J. Ho, A. Jain, and P. Abbeel, inAdvances in neural information processing systems, Vol. 33 (2020) p. 6840–6851

    2020

  23. [31]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, CoRR abs/2011.13456(2020), 2011.13456

    2020 arXiv

  24. [32]

    Klypin, G

    A. Klypin, G. Yepes, S. Gottl¨ ober, F. Prada, and S. Heß, Monthly Notices of the Royal Astronomical Society457, 4340 (2016), https://academic.oup.com/mnras/article-pdf/457/4/4340/18515365/stw248.pdf

    2016

  25. [33]

    van Engelen, S

    A. van Engelen, S. Bhattacharya, N. Sehgal, G. P. Holder, O. Zahn, and D. Nagai, ApJ786, 13 (2014), arXiv:1310.7023 [astro-ph.CO]

    2014 arXiv

  26. [34]

    M. S. Madhavacheril and J. C. Hill, Physical Review D98, 023534 (2018), arXiv:1802.08230 [astro-ph.CO]

    2018 arXiv

  27. [35]

    Sailer, E

    N. Sailer, E. Schaan, and S. Ferraro, Physical Review D102, 063517 (2020), arXiv:2007.04325 [astro-ph.CO]

    2020 arXiv

  28. [36]

    MacCrann, B

    N. MacCrann, B. D. Sherwin, F. J. Qu, T. Namikawa, M. S. Madhavacheril,et al., arXiv e-prints , arXiv:2304.05196 (2023), arXiv:2304.05196 [astro-ph.CO]

    2023 arXiv

  29. [37]

    Doohan, M

    M. Doohan, M. Millea, S. Raghunathan, F. Ge, L. Knox, K. Prabhu, C. L. Reichardt, and W. L. K. Wu, JCAP (2025), arXiv:2502.20801 [astro-ph.CO]

    2025 arXiv

  30. [38]

    Fabbian, A

    G. Fabbian, A. Lewis, and D. Beck, JCAP2019, 057 (2019), arXiv:1906.08760 [astro-ph.CO]

    2019 arXiv

  31. [39]

    Omori, E

    Y. Omori, E. J. Baxter, C. Chang,et al., Physical Review D107, 023529 (2023), arXiv:2203.12439 [astro-ph.CO]

    2023 arXiv

  32. [40]

    J. A. Sobrin, A. J. Anderson, A. N. Bender,et al., ApJS258, 42 (2022), arXiv:2106.11202 [astro-ph.IM]

    2022 arXiv

  33. [41]

    I. G. McCarthy, J. Schaye, S. Bird, and A. M. C. Le Brun, MNRAS465, 2936 (2017), arXiv:1603.02702 [astro-ph.CO]

    2017 arXiv

  34. [42]

    A. J. Mead, T. Tr¨ oster, C. Heymans, L. Van Waerbeke, and I. G. McCarthy, A&A641, A130 (2020), arXiv:2005.00009 [astro-ph.CO]

    2020 arXiv

  35. [43]

    Behroozi, R

    P. Behroozi, R. H. Wechsler, A. P. Hearin, and C. Conroy, Monthly Notices of the Royal Astronomical Society488, 3143–3194 (2019)

    2019

  36. [44]

    K. M. G´ orski, E. Hivon, A. J. Banday, B. D. Wandelt, F. K. Hansen, M. Reinecke, and M. Bartelmann, ApJ622, 759 (2005), arXiv:astro-ph/0409513 [astro-ph]

    2005 arXiv

  37. [45]

    Abazajian, G

    K. Abazajian, G. E. Addison, P. Adshead,et al., ApJ926, 54 (2022), arXiv:2008.12619 [astro-ph.CO]

    2022 arXiv

  38. [46]

    Raghunathan, ApJ928, 16 (2022), arXiv:2112.07656 [astro-ph.CO]

    S. Raghunathan, ApJ928, 16 (2022), arXiv:2112.07656 [astro-ph.CO]

    2022 arXiv

  39. [47]

    denoising-diffusion-pytorch,

    P. Wang, “denoising-diffusion-pytorch,” https://github.com/lucidrains/denoising-diffusion-pytorch(2022)

    2022

  40. [48]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox, inMedical image computing and computer-assisted intervention – MICCAI 2015(Springer International Publishing, Cham, 2015) p. 234–241. – 21 –

    2015

  41. [49]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, inAdvances in Neural Information Processing Systems, Vol. 30 (2017)

    2017

  42. [50]

    Stamatelopoulos and T

    S. Stamatelopoulos and T. P. Sapsis, Computer Methods in Applied Mechanics and Engineering435, 117589 (2025)

    2025

  43. [51]

    Schmalzing, M

    J. Schmalzing, M. Kerscher, and T. Buchert, inDark Matter in the Universe, edited by S. Bonometto, J. R. Primack, and A. Provenzale (1996) p. 281, arXiv:astro-ph/9508154 [astro-ph]

    1996 arXiv

  44. [52]

    A. M. P. Boelens and H. A. Tchelepi, SoftwareX16, 100823 (2021)

    2021

  45. [53]

    J. J. Lee, J. R. Bond, P. Motloch, A. van Engelen, and G. Stein, Monthly Notices of the Royal Astronomical Society529, 2543 (2024), https://academic.oup.com/mnras/article-pdf/529/3/2543/57035729/stae605.pdf

    2024

  46. [54]

    J. C. Hill, E. J. Baxter, A. Lidz, J. P. Greco, and B. Jain, Physical Review D97, 083501 (2018), arXiv:1706.03753 [astro-ph.CO]

    2018 arXiv

  47. [55]

    J. R. Bond, B. J. Carr, and C. J. Hogan, ApJ367, 420 (1991)

    1991

  48. [56]

    Millea and U

    M. Millea and U. Seljak, arXiv:2112.09354 [astro-ph, stat] (2021), arxiv:2112.09354 [astro-ph, stat]

    2021 arXiv

  49. [57]

    Z. Wang, Y. Jiang, H. Zheng, P. Wang, P. He, Z. Wang, W. Chen, and M. Zhou, arXiv e-prints , arXiv:2304.12526 (2023), arXiv:2304.12526 [cs.CV]

    2023 arXiv

  50. [58]

    Dhariwal and A

    P. Dhariwal and A. Nichol, arXiv e-prints , arXiv:2105.05233 (2021), arXiv:2105.05233 [cs.LG]

    2021 arXiv

  51. [59]

    Grover, M

    A. Grover, M. Dhar, and S. Ermon, arXiv e-prints , arXiv:1705.08868 (2017), arXiv:1705.08868 [cs.LG]

    2017 arXiv

  52. [60]

    Thorne, J

    B. Thorne, J. Dunkley, D. Alonso, and S. Næss, MNRAS469, 2821 (2017), arXiv:1608.02841 [astro-ph.CO]

    2017 arXiv

  53. [61]

    Zhang and Y

    Q. Zhang and Y. Chen, arXiv e-prints , arXiv:2110.07579 (2021), arXiv:2110.07579 [cs.LG]

    2021 arXiv

  54. [62]

    Salimans and J

    T. Salimans and J. Ho, inInternational Conference on Learning Representations(2022)

    2022

  55. [63]

    D. P. Kingma, arXiv preprint arXiv:1412.6980 (2014). – 22 –

    2014 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.