Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Adding patch-level neural summaries to the power spectrum and bispectrum improves f_NL constraints by 30–45% at low k_max, capturing information beyond the 3-point function.

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

2026-08-04 09:53 UTC pith:ZQ5OXCND

load-bearing objection A credible Fisher study showing patch-level neural summaries add real f_NL information over P(k)+B(k), but the headline gains need error bars and a linearity check on the network derivative. the 4 major comments →

arxiv 2510.12715 v1 pith:ZQ5OXCND submitted 2025-10-14 astro-ph.CO

Hierarchical summaries for primordial non-Gaussianities

M.S. Cagliari , A. Bairagi , B. Wandelt This is my paper
classification astro-ph.CO
keywords primordial non-Gaussianitylocal f_NLFisher informationfield-level inferenceconvolutional neural networkpower spectrum and bispectrumhalo clusteringcosmological surveys
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 proposes a hierarchical estimator for measuring the amplitude of local primordial non-Gaussianity, f_NL, from galaxy surveys. The idea is to keep the large-scale information in the standard power spectrum and bispectrum, while a convolutional neural network called PatchNet compresses small cubic patches of the halo density field into one number each; the patch predictions are averaged and concatenated with the n-point statistics. Using Fisher forecasts on a large suite of N-body halo simulations, the paper claims this combined estimator always outperforms the standard statistics alone, with 30–45% improvements at k_max ~0.1 h/Mpc, and that the patch summary contains information beyond the bispectrum. If true, this offers a computationally cheap route toward sigma_fNL ~1, the threshold that would distinguish single-field from multi-field inflation, without the modeling burden of pushing standard statistics to small scales.

Core claim

The central claim is that a field-level neural summary computed on sub-volumes can be spliced onto standard large-scale statistics to recover nearly all available f_NL information. Concretely, the power spectrum and bispectrum of the full simulation box are measured up to a scale cut k_max, while PatchNet maps each 125^3 Mpc^3/h patch of the smoothed halo overdensity field to a predicted f_NL; the mean over 512 non-overlapping patches becomes a one-number summary. The Fisher information of the concatenated vector P(k)+B(k)+patches is computed with 5000 fiducial simulations for the covariance and finite-difference derivatives from f_NL = -50 and +50 simulation suites. The paper finds that the

What carries the argument

PatchNet: a convolutional neural network that takes a small cubic patch of the smoothed halo density field and outputs a single predicted f_NL value; averaging over all patches gives a field-level compression that is concatenated with the power spectrum and bispectrum. It extracts small-scale, higher-order information that the 2-point and 3-point functions miss, while the global n-point statistics retain the large-scale squeezed configurations that small patches cannot see. The hierarchy is motivated by the locality principle: late-time gravitational evolution produces local correlations, so the non-local primordial signal stays separable, making the combination of n-point statistics with a

Load-bearing premise

The forecast treats the neural network's output as a linear function of f_NL near zero, so the derivative measured from simulations at f_NL = +/- 50 is assumed to represent the true slope; this linearity is verified only for the power spectrum, not for PatchNet, and the quoted gains have no error bars.

What would settle it

Recompute the PatchNet Fisher derivative using several finite-difference offsets (e.g., f_NL = +/- 20, +/- 50, +/- 100) and verify that sigma_fNL and the 30–45% gain are stable; because the network is nonlinear, a strongly offset-dependent derivative would mean the reported gain is an artifact of the chosen delta f_NL.

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

If this is right

  • If the Fisher gains hold, adding PatchNet to a survey analysis at k_max about 0.1 h/Mpc delivers the f_NL constraining power of standard statistics at k_max about 0.5 h/Mpc, avoiding much of the small-scale modeling cost.
  • Because P(k)+patches beats P(k)+B(k), the hierarchical estimator captures information beyond the three-point function without explicitly computing higher-order correlation functions.
  • The largest 30–45% gains occur exactly where standard statistics are easiest to model, so the method improves the science return of Stage IV surveys at modest added computational cost.
  • The control experiment indicates that even when the halo mass function is fixed, the network extracts f_NL information from the density field, so the gain is not a trivial mass-function effect.
  • The method is most effective for higher-mass, highly biased tracers at k_max >= 0.2 h/Mpc, suggesting particular value for quasar-like samples used in f_NL searches.

Where Pith is reading between the lines

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

  • The paper's non-overlapping patch choice is deliberately conservative; if overlapping patches add 15–20% more information as the authors note, the realizable gains could exceed the reported 30–45%.
  • The reported gains assume no marginalization over cosmological or nuisance parameters; a testable next step is whether the relative gain persists after marginalizing, since network summaries can break degeneracies shared by n-point statistics.
  • The same hierarchical recipe could be retrained for equilateral or orthogonal PNG shapes, but the locality argument motivating the combination is shape-dependent, so whether the gains transfer is an open empirical question.
  • A direct comparison with other small-scale summaries, such as wavelet scattering transforms or marked power spectra, would clarify how much of the gain is specific to the learned network versus the general idea of adding local field-level information.

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 proposes a hybrid estimator for constraining local primordial non-Gaussianity f_NL, combining the halo power spectrum and bispectrum with a convolutional-neural-network summary (PatchNet) computed on sub-volumes of the halo density field. Using Quijote-PNG halo catalogs, the authors perform a Fisher forecast, estimating covariance from 5000 fiducial simulations with Hartlap correction and derivatives from finite differences between f_NL = ±50 simulations. They report that adding the patch summary always improves the expected f_NL constraint, with gains of 30–45% at low kmax (~0.1 h/Mpc), and that the patches capture information beyond the bispectrum. The paper studies the gains as a function of redshift, halo mass cut, and scale cut, and includes a control removing halo-mass-function information.

Significance. If the result holds, this is a useful and computationally efficient step toward extracting more f_NL information from Stage IV surveys, particularly in the regime where standard perturbative modeling is most reliable. The paper’s strengths include the use of standard Fisher machinery with Hartlap correction, the exploration of several redshifts and mass cuts, and a control experiment for the halo mass function. The main uncertainty is the derivative of a learned neural summary, on which the entire Fisher gain rests; this derivative is not validated for linearity or statistical convergence. The claim that patch-based compression 'always enhances' constraints and captures information beyond the bispectrum is therefore not yet fully supported.

major comments (4)
  1. [§3.1, Eq. (3)] The Fisher derivative for the PatchNet component is computed as a central finite difference with δf_NL = 50. The mean output of a CNN trained with an MSE loss is a learned regression function and is not guaranteed to be linear over f_NL ∈ [−50, 50]. If the response has curvature, the finite-difference slope differs from the local derivative at f_NL = 0, directly biasing σ_fNL and the headline 30–45% gains. Footnote 4 checks derivative stability with simulation count only for the power spectrum, not for the network. Please provide a linearity check (e.g., derivative estimates at multiple δf_NL, or a plot of mean network output versus f_NL) and/or bootstrap uncertainties on the derivative.
  2. [§2, §3.3] The paper does not state whether the 5000 fiducial simulations and the 500+500 f_NL = ±50 simulations used for the Fisher analysis are disjoint from the PatchNet training set. If any of these simulations were used in training, the covariance and derivative estimates would be optimistically biased. Please explicitly confirm the separation, or describe any cross-validation procedure that prevents leakage.
  3. [§4] The reported gains of 30–45% are presented without error bars or significance tests. The Fisher quantities are estimated from a finite number of simulations (5000 for covariance, 500 per f_NL sign for derivatives), so the point estimates have sampling noise. Reporting the uncertainty on σ_fNL and on the gain, or at least demonstrating that the gain is robust to the number of simulations used for the PatchNet derivative, is necessary to support the 'always enhances' claim.
  4. [§4, HMF control] The halo-mass-function removal control reports Pearson correlation coefficients (r ~ 0.78 without HMF vs r ~ 0.99 with HMF) but does not translate these into Fisher constraints. The conclusion that 'the network extracts information beyond the halo mass function' is supported only qualitatively. To connect this control to the central information-gain claim, please provide a Fisher forecast for the fixed-number-density configuration, or explicitly state that the correlation test is not a Fisher-information statement.
minor comments (4)
  1. [References] The reference 'Maas, A. L. 2013, in' is incomplete; the proceedings or volume information is missing.
  2. [Fig. 2 caption] The mass cut is written as 'M >3.2×M ⊙ h−1'; the exponent 10^13 is missing.
  3. [§4] The phrase 'when k_patch_max < k_P,B_max point again in the direction' is awkward and should be rephrased.
  4. [§3.3] The description of the patch construction states that patches are non-overlapping and then notes that overlapping patches increase information by 15–20%. It would be clearer to state explicitly that the reported Fisher results use the non-overlapping choice, which is already implied but could be made explicit for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the Fisher forecast is self-contained and the reported gains are measured, not forced by construction.

full rationale

The paper's derivation chain is a standard Fisher forecast: it constructs three summary statistics (power spectrum, bispectrum, and the PatchNet patch-average), estimates the covariance from 5000 fiducial Quijote-PNG simulations, estimates the fNL derivative from finite differences at fNL = ±50, and combines them through Eq. (1). The central quantitative claims—30–45% gains at low kmax and information beyond the bispectrum—are empirical outputs of this pipeline, not identities. PatchNet is trained on a separate 1000-simulation Latin hypercube set and evaluated on held-out ±50 simulations, so its derivative is a measured property of a learned estimator rather than a parameter fitted to the forecast data. The claim that adding patch compression 'always enhances' constraints is partly a monotonicity property of Fisher information, but the magnitude of the gains and the beyond-bispectrum comparison are nontrivial. Self-citations to Bairagi & Wandelt (2025) supply the architecture and motivation, but the fNL forecast is independently evaluated on Quijote-PNG simulations; no equation reduces to its inputs by construction. The only noteworthy limitation is that derivative stability with simulation count is checked for the power spectrum only (footnote 4), not for the network summary, and no linearity test for the finite-difference step is reported. That is a robustness/correctness concern, not evidence of circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No invented physical entities: PatchNet is an analysis tool adopted from prior work, not a new physical ingredient. Free parameters are the neural architecture/hyperparameters (untested by ablation) and the finite-difference step. The axioms list the domain assumptions the forecast depends on: the locality principle justifying the combination, linear response of the network summary, Gaussian Fisher likelihood, fixed-cosmology simulations, and the sufficiency of the mass-unweighted CIC-painted density field.

free parameters (2)
  • PatchNet architecture & training hyperparameters = 8→32→128 conv filters, 512→256 dense, LeakyReLU slope 0.5, n_b=32, n_sb=64, 512 patches/box of 16^3 voxels; optimizer/lr
    Carried over from Bairagi & Wandelt (2025); no ablation or tuning study in this paper, so sensitivity of the claimed 30-45% gains to these choices is untested.
  • finite-difference step δf_NL = 50
    Eq. (3) uses simulations at f_NL=±50 to estimate the derivative; stability is verified only against number of simulations (footnote 4), not against step size or linearity of the neural summary.
axioms (5)
  • domain assumption Locality principle: late-time gravitational evolution produces only local, causal effects and cannot generate the long-range correlations produced by primordial non-Gaussianity (Baumann & Green 2022); hence P+B+local-patches approximates a map-level optimal analysis.
    Invoked in §1 to justify the hierarchical combination and the 'nearly optimal' claim in the abstract; the paper does not derive it.
  • domain assumption The neural summary's expectation responds approximately linearly to f_NL on [-50, 50], so the finite-difference derivative at ±50 equals the local derivative at f_NL=0.
    Needed for Eq. (3); not validated for the network output (only P(k) Fisher stability vs sim count is checked, footnote 4).
  • standard math Fisher information with Gaussian likelihood, covariance evaluated at the fiducial model (f_NL=0) with Hartlap correction, adequately represents the estimator's constraining power.
    §3.1, Eqs. (1)-(2); standard practice in cosmological forecasts, but a Gaussian likelihood for the patch summary is assumed without verification.
  • domain assumption Quijote-PNG halo catalogs with all cosmological parameters fixed except f_NL faithfully represent the f_NL-dependence of the halo density field.
    §2 sets the simulation strategy; single-parameter forecasts ignore degeneracies, acknowledged in §5 as future work.
  • domain assumption The patch density field (halos CIC-painted on a 128^3 grid, no mass weighting) retains the small-scale f_NL information the paper claims to extract.
    §3.3; the HMF-removal control (r_w/o HMF = 0.78) partially supports that information survives without number-density information.

pith-pipeline@v1.3.0-alltime-deepseek · 11114 in / 21943 out tokens · 169576 ms · 2026-08-04T09:53:05.260524+00:00 · methodology

0 comments
read the original abstract

The advent of Stage IV galaxy redshift surveys such as DESI and Euclid marks the beginning of an era of precision cosmology, with one key objective being the detection of primordial non-Gaussianities (PNG), potential signatures of inflationary physics. In particular, constraining the amplitude of local-type PNG, parameterised by $f_{\rm NL}$, with $\sigma_{f_{\rm NL}} \sim 1$, would provide a critical test of single versus multi-field inflation scenarios. While current large-scale structure and cosmic microwave background analyses have achieved $\sigma_{f_{\rm NL}} \sim 5$-$9$, further improvements demand novel data compression strategies. We propose a hybrid estimator that hierarchically combines standard $2$-point and $3$-point statistics with a field-level neural summary, motivated by recent theoretical work that shows that such a combination is nearly optimal, disentangling primordial from late-time non-Gaussianity. We employ PatchNet, a convolutional neural network that extracts small-scale information from sub-volumes (patches) of the halo number density field while large-scale information is retained via the power spectrum and bispectrum. Using Quijote-PNG simulations, we evaluate the Fisher information of this combined estimator across various redshifts, halo mass cuts, and scale cuts. Our results demonstrate that the inclusion of patch-based field-level compression always enhances constraints on $f_{\rm NL}$, reaching gains of $30$-$45\%$ at low $k_{\rm max}$ ($\sim 0.1 \, h \, \text{Mpc}^{-1}$), and capturing information beyond the bispectrum. This approach offers a computationally efficient and scalable pathway to tighten the PNG constraints from forthcoming survey data.

Figures

Figures reproduced from arXiv: 2510.12715 by A. Bairagi, B. Wandelt, M.S. Cagliari.

Figure 1
Figure 1. Figure 1: Expected uncertainty for the different summary statistics at z = 1 for two halo-mass cuts as a function of the power spectrum and bispectrum kmax. In both cases, the PatchNet information (patches) improves the constraining power of the power spectrum and the power spectrum combined with bispectrum analyses. measurements. A first solution is to use effective field theory to increase the minimum scale we can… view at source ↗
Figure 2
Figure 2. Figure 2: Expected uncertainty for the different summary statistics as a function of the power spectrum and bispectrum kmax at different redshift snapshots. At all redshifts, we applied the same halo-mass cut, Mh > 3.2 × M⊙ h −1 . The patch information adds constraining power at every redshift. divide the 3-dimensional number density contrast into 163 non-overlapping cubes.5 This results in 512 patches for each simu… view at source ↗
Figure 3
Figure 3. Figure 3: Expected uncertainty for the different summary statistics at fixed number density over redshifts as a function of the power spectrum and bispectrum kmax. At each redshift, we selected a halo-mass cut that has in the fiducial cosmology the same number density of z = 3 (see [PITH_FULL_IMAGE:figures/full_fig_p006_3.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

39 extracted references · 5 linked inside Pith

  1. [1]

    N., Adshead, P., Ahmed, Z., et al

    Abazajian, K. N., Adshead, P., Ahmed, Z., et al. 2016, arXiv:1610.02743

    Pith/arXiv arXiv 2016

  2. [2]

    G., Aguilar, J., Ahlen, S., et al

    Adame, A. G., Aguilar, J., Ahlen, S., et al. 2025, JCAP, 2025, 028

    2025

  3. [3]

    R., et al

    Alvarez, M., Baldauf, T., Bond, J. R., et al. 2014, arXiv:1412.4671

    Pith/arXiv arXiv 2014

  4. [4]

    2024, arXiv:2412.11945

    Andrews, A., Jasche, J., Lavaux, G., et al. 2024, arXiv:2412.11945

    arXiv 2024

  5. [5]

    2023, MNRAS, 520, 5746

    Andrews, A., Jasche, J., Lavaux, G., & Schmidt, F. 2023, MNRAS, 520, 5746

    2023

  6. [6]

    & Wandelt, B

    Bairagi, A. & Wandelt, B. 2025, arXiv:2509.03165

    Pith/arXiv arXiv 2025

  7. [7]

    2025, arXiv:2503.13755

    Bairagi, A., Wandelt, B., & Villaescusa-Navarro, F. 2025, arXiv:2503.13755

    arXiv 2025

  8. [8]

    2011, in Theoretical Advanced Study Institute in El- ementary Particle Physics: Physics of the Large and the Small, 523–686

    Baumann, D. 2011, in Theoretical Advanced Study Institute in El- ementary Particle Physics: Physics of the Large and the Small, 523–686

    2011

  9. [9]

    & Green, D

    Baumann, D. & Green, D. 2022, JCAP, 2022, 061

    2022

  10. [10]

    M., Lewandowski, M., Mirbabayi, M., & Si- monović, M

    Cabass, G., Ivanov, M. M., Lewandowski, M., Mirbabayi, M., & Si- monović, M. 2023, Physics of the Dark Universe, 40, 101193

    2023

  11. [11]

    M., Philcox, O

    Cabass, G., Ivanov, M. M., Philcox, O. H. E., Simonović, M., & Zal- darriaga, M. 2022, Phys. Rev. D, 106, 043506

    2022

  12. [12]

    2017, JCAP, 2017, 003

    Cabass, G., Pajer, E., & Schmidt, F. 2017, JCAP, 2017, 003

    2017

  13. [13]

    S., Barberi-Squarotti, M., Pardede, K., Castorina, E., & D’Amico, G

    Cagliari, M. S., Barberi-Squarotti, M., Pardede, K., Castorina, E., & D’Amico, G. 2025, JCAP, 2025, 043

    2025

  14. [14]

    S., Castorina, E., Bonici, M., & Bianchi, D

    Cagliari, M. S., Castorina, E., Bonici, M., & Bianchi, D. 2024, JCAP, 2024, 036

    2024

  15. [15]

    2019, JCAP, 2019, 010

    Castorina, E., Hand, N., Seljak, U., et al. 2019, JCAP, 2019, 010

    2019

  16. [16]

    2025, JCAP, 2025, 029

    Chaussidon, E., Yèche, C., de Mattia, A., et al. 2025, JCAP, 2025, 029

    2025

  17. [17]

    Chen, X., Padmanabhan, N., & Eisenstein, D. J. 2025, JCAP, 2025, 055

    2025

  18. [18]

    R., Abel, T., & Banerjee, A

    Coulton, W. R., Abel, T., & Banerjee, A. 2024, MNRAS, 534, 1621

    2024

  19. [19]

    & Zaldarriaga, M

    Creminelli, P. & Zaldarriaga, M. 2004, JCAP, 2004, 006

    2004

  20. [20]

    P., Werner, M., Akeson, R., et al

    Crill, B. P., Werner, M., Akeson, R., et al. 2020, in Society of Photo- Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 11443, Space Telescopes and Instrumentation 2020: Optical, In- frared, and Millimeter Wave, ed. M. Lystrup & M. D. Perrin, 114430I

    2020

  21. [21]

    2008, Phys

    Dalal, N., Doré, O., Huterer, D., & Shirokov, A. 2008, Phys. Rev. D, 77, 123514 D’Amico, G., Lewandowski, M., Senatore, L., & Zhang, P. 2025, Phys. Rev. D, 111, 063514

    2008

  22. [22]

    S., & White, S

    Davis, M., Efstathiou, G., Frenk, C. S., & White, S. D. M. 1985, ApJ, 292, 371 DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016, arXiv e-prints, arXiv:1611.00036

    Pith/arXiv arXiv 1985

  23. [23]

    & Seljak, U

    Desjacques, V. & Seljak, U. 2010, Advances in Astronomy, 2010, 908640 Euclid Collaboration: Mellier, Y., Abdurro’uf, Acevedo Barroso, J. A., et al. 2025, A&A, 697, A1 Euclid Collaboration: Scaramella, R., Amiaux, J., Mellier, Y., et al. 2022, A&A, 662, A112

    2010

  24. [24]

    Giri, U., Münchmeyer, M., & Smith, K. M. 2023, Phys. Rev. D, 107, L061301

    2023

  25. [25]

    2020, JCAP, 2020, 040

    Hahn, C., Villaescusa-Navarro, F., Castorina, E., & Scoccimarro, R. 2020, JCAP, 2020, 040

    2020

  26. [26]

    2007, A&A, 464, 399 Article number, page 7 A&A proofs:manuscript no

    Hartlap, J., Simon, P., & Schneider, P. 2007, A&A, 464, 399 Article number, page 7 A&A proofs:manuscript no. main

    2007

  27. [27]

    2024, ApJ, 976, 109

    Jung, G., Ravenni, A., Liguori, M., et al. 2024, ApJ, 976, 109

    2024

  28. [28]

    2025, Phys

    Kvasiuk, Y., Münchmeyer, M., & Smith, K. 2025, Phys. Rev. D, 112, 023540

    2025

  29. [29]

    1989, Generalization and network design strategies, ed

    LeCun, Y. 1989, Generalization and network design strategies, ed. R. Pfeifer, Z. Schreter, F. Fogelman, & L. Steels (Elsevier) LSST Science Collaboration, Abell, P. A., Allison, J., et al. 2009, arXiv e-prints, arXiv:0912.0201

    Pith/arXiv arXiv 1989

  30. [30]

    2003, Journal of High Energy Physics, 2003, 013

    Maldacena, J. 2003, Journal of High Energy Physics, 2003, 013

    2003

  31. [31]

    2025, JCAP, 2025, 036

    Marinucci, M., Jung, G., Liguori, M., et al. 2025, JCAP, 2025, 036

    2025

  32. [32]

    Nagarajappa, C. G. & Ma, Y.-Z. 2024, MNRAS, 529, 3289

    2024

  33. [33]

    2024, JCAP, 2024, 021 Planck Collaboration, Akrami, Y., Arroja, F., et al

    Peron, M., Jung, G., Liguori, M., & Pietroni, M. 2024, JCAP, 2024, 021 Planck Collaboration, Akrami, Y., Arroja, F., et al. 2020, A&A, 641, A9

    2024

  34. [34]

    2015, Phys

    Scoccimarro, R. 2015, Phys. Rev. D, 92, 083532

    2015

  35. [35]

    & Zaldarriaga, M

    Senatore, L. & Zaldarriaga, M. 2012, Journal of High Energy Physics, 2012, 24

    2012

  36. [36]

    2008, JCAP, 2008, 031

    Slosar, A., Hirata, C., Seljak, U., Ho, S., & Padmanabhan, N. 2008, JCAP, 2008, 031

    2008

  37. [37]

    2008, MNRAS, 391, 1685

    Springel, V., Wang, J., Vogelsberger, M., et al. 2008, MNRAS, 391, 1685

    2008

  38. [38]

    Springel, V., White, S. D. M., Jenkins, A., et al. 2005, Nature, 435, 629 Villaescusa-Navarro,F.2018,Pylians:Pythonlibrariesfortheanalysis of numerical simulations, Astrophysics Source Code Library, record ascl:1811.008

    2005

  39. [39]

    2020, ApJS, 250, 2 Article number, page 8

    Villaescusa-Navarro, F., Hahn, C., Massara, E., et al. 2020, ApJS, 250, 2 Article number, page 8

    2020