arxiv: 2605.09004 · v1 · submitted 2026-05-09 · 🌌 astro-ph.CO

Recognition: 2 theorem links

· Lean Theorem

Separate Universe Super-Resolution Emulator

Dennis Fremstad , Julian Adamek , David F. Mota

Authors on Pith no claims yet

Pith reviewed 2026-05-12 01:46 UTC · model grok-4.3

classification 🌌 astro-ph.CO

keywords super-resolutionN-body simulationsgenerative adversarial networkspatial curvaturecosmological emulatorlarge-scale structurehalo abundance

0 comments

The pith

A conditional GAN generates high-resolution N-body simulations with spatial curvature from low-resolution inputs, recovering most missing power on large scales.

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

This paper develops a machine learning model to create super-resolution versions of N-body simulations that include non-vanishing spatial curvature. The model is a generative adversarial network conditioned on low-resolution fields along with parameters for curvature, matter density, sigma eight, Hubble constant, and redshift. It allows production of multiple high-resolution realizations at far lower computational cost than running full simulations. Evaluations show good agreement with true high-resolution runs on large scales but a remaining suppression of small-scale power and halo central densities at the ten percent level. The authors demonstrate its use by upscaling a light cone from a large curved simulation to produce a catalogue combining large and small scale features.

Core claim

Our model accurately reproduces large-scale statistics, robustly recovering most of the power that was missing from the low-resolution input, but exhibits a residual suppression of power on small scales of up to ∼10% at k∼1 h Mpc^{-1}. The abundance of halos around 10^{14} M_⊙ is affected at a similar level, and we find that the profiles of these halos have a lower central density. To show a production-scale use case, we apply our model to upscale the resolution of a light cone from a large-volume N-body simulation with spatial curvature, producing a first-of-its-kind catalogue that simultaneously captures geometric effects at large scales and accurate nonlinear structure at small scales.

What carries the argument

Conditional generative adversarial network for upscaling N-body simulation fields with injected noise for stochastic structure, taking cosmological parameters including Omega_k as input.

If this is right

Drastically reduces the computational cost of producing high-resolution simulations for modeling large-scale structure surveys.
Enables generation of ensembles of high-resolution realizations to account for stochastic small-scale structure.
Allows creation of light cone catalogues that include both curvature effects on large scales and nonlinear clustering on small scales.
Maintains fidelity on large scales while introducing up to 10% suppression in small-scale power and halo central densities.

Where Pith is reading between the lines

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

The residual small-scale discrepancies could be mitigated by expanding the training dataset or refining the model architecture.
This emulator might be combined with other simulation techniques to cover a wider range of cosmological models.
Further validation on simulations with parameters outside the training distribution would be needed before using it for precision cosmology.

Load-bearing premise

That the GAN trained on a finite set of simulations generalizes to produce unbiased small-scale structure and halo properties across the full range of input parameters and redshifts without systematic errors affecting downstream cosmological inference.

What would settle it

A direct comparison of the generated simulations' small-scale power spectrum and halo density profiles against an independent set of high-resolution N-body simulations with matching initial conditions and cosmological parameters.

Figures

Figures reproduced from arXiv: 2605.09004 by David F. Mota, Dennis Fremstad, Julian Adamek.

**Figure 1.** Figure 1: Projection of the Latin hypercube sampling into the different two-dimensional parameter ranges. We sample uniformly Ωk ∈ [−0.1, 0.1], h ∈ [0.674, 0.768], σ8 ∈ [0.76969, 0.85071], and Ωm ∈ [0.29555, 0.32666]. The different colours correspond to the 32 different training examples. and we assume three massless neutrino species with the standard value Neff = 3.046. The linear matter power spectrum at redshif… view at source ↗

**Figure 2.** Figure 2: Comparison of the matter power spectrum at z ≈ 0.817 of the LR input, the HR target, and the SR model prediction. The bottom panel shows the mean and standard deviation (bands) in the relative difference between the predicted and the target power spectrum of the whole test data-set. The vertical dashed line indicates the Nyquist wavenumber of the high-resolution initial data. To estimate the spectra, we pr… view at source ↗

**Figure 3.** Figure 3: Matter power spectrum of SR samples, along with their HR counterpart and the prediction from CLASS with the nonlinear model published by Mead et al. (2021) as a function of redshift. Here, we have used a finer sampling over redshift than what was used in the training data in order to also test the model’s ability to interpolate. The relative error with respect to the HR target is shown in the lower panel. … view at source ↗

**Figure 5.** Figure 5: The halo mass function at different redshifts. The upper panel shows the comparison between the produced SR, and the simulated HR halo mass functions. The lower panel shows the relative difference between the two. mass halos of ≲ 40%, while for intermediate-mass halos (around 5 × 1013 M⊙), the reduction reaches ∼ 10%. The increased error at small scales observed in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 7.** Figure 7: Slice through the density field of the low-resolution (left) and super-resolution (right) light cones. The top panels show the full 8080 h −1 Mpc slice where the observer is located at the centre, while the bottom panels show a zoomed-in region with side length 591.8 h −1 Mpc. The location of the zoomed-in regions are indicated by red boxes in the top panels. The light cone is apodised and truncated at the… view at source ↗

**Figure 8.** Figure 8: Angular power spectrum calculated from the SR light cone, the LR light cone and the prediction from CLASS, calculated at z = 1.1. Here, we have used a redshift bin with full width ∆z = 0.1. (Górski et al. 2005) and estimate the binned angular power spectrum with NaMaster1 (Alonso et al. 2019). In the bottom panel of [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

read the original abstract

We present a machine-learning model for generating super-resolution $N$-body simulations with non-vanishing spatial curvature, conditioned on a given low-resolution field, $\Omega_k$, $\Omega_\mathrm{m}$, $\sigma_8$, $h$, and redshift. By upscaling the resolution of $N$-body simulations, such models can drastically reduce the computational cost of producing high-resolution simulations suitable for modelling current and future surveys of large-scale structure. Our model is trained as a generative adversarial network, allowing injected noise to be interpreted as stochastic structure and enabling the generation of an ensemble of plausible high-resolution realisations. We evaluate the model performance by comparing key cosmological summary statistics in the generated simulations to their high-resolution counterparts. We find that the model accurately reproduces large-scale statistics, robustly recovering most of the power that was missing from the low-resolution input, but exhibits a residual suppression of power on small scales of up to $\sim 10\%$ at $k \sim 1\,h\,\mathrm{Mpc}^{-1}$. The abundance of halos around $10^{14}\,M_\odot$ is affected at a similar level, and we find that the profiles of these halos have a lower central density. Although the overall performance is decent, we anticipate that the fidelity of the generative model can be further increased with more and better training data, as well as through improvements in the model architecture and training process. To show a production-scale use case, we apply our model to upscale the resolution of a light cone from a large-volume $N$-body simulation with spatial curvature, producing a first-of-its-kind catalogue that simultaneously captures geometric effects at large scales and accurate nonlinear structure at small scales.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The GAN super-res emulator recovers large-scale power in curved cosmologies but leaves ~10% small-scale deficits and lower halo central densities.

read the letter

This paper builds a GAN that upsamples low-resolution N-body fields to high resolution while conditioning on spatial curvature, Omega_m, sigma_8, h, and redshift. The new element is applying this to separate-universe runs with non-zero Omega_k and then producing a curved light-cone catalog from a large-volume simulation. That catalog is the first of its kind and shows the method can handle geometric effects at large scales together with nonlinear structure at smaller scales. Large-scale statistics come out reasonably well, recovering most of the power missing from the low-res input. The adversarial training also lets them generate multiple realizations from the same low-res seed, which is practically useful for ensemble work. The authors are straightforward about the remaining problems: power is still suppressed by up to 10% at k ~ 1 h/Mpc, halo abundances around 10^14 solar masses are off at a similar level, and the halos themselves have lower central densities. Those residuals are presented as expected limitations rather than hidden. The main soft spot is the lack of detail on training-set size, architecture choices, validation splits, and how well the model generalizes across the full parameter range and redshifts. Without those, it is difficult to know how much the small-scale errors will propagate into downstream analyses. The work is aimed at groups that need many high-resolution curved simulations for next-generation surveys but cannot afford full high-res runs. A reader working on emulators or on curvature effects in large-scale structure will find the concrete example and the stated trade-offs useful. It is solid enough on its own terms to deserve a serious referee, mainly to check the training details and to see whether the small-scale biases can be reduced or at least quantified for specific science cases. I would send it to review.

Referee Report

2 major / 2 minor

Summary. The paper presents a GAN-based super-resolution emulator for N-body simulations that incorporates non-vanishing spatial curvature. Conditioned on a low-resolution field plus parameters (Ω_k, Ω_m, σ_8, h, redshift), the model generates high-resolution realizations and is evaluated on summary statistics including power spectra and halo properties. It reports robust recovery of large-scale power with a residual ~10% suppression at k~1 h Mpc^{-1}, similar effects on halo abundance near 10^{14} M_⊙, and lower central densities in halo profiles. A production-scale demonstration applies the model to upscale a light-cone catalogue from a large-volume curved simulation.

Significance. If the reported performance generalizes, the emulator offers a practical route to high-resolution simulations with curvature at reduced cost, which is valuable for modeling geometric effects in future large-scale structure surveys. The explicit conditioning on curvature and the GAN's ability to produce ensembles are strengths; the light-cone application demonstrates a concrete use case. The acknowledged small-scale residuals are framed transparently as improvable with more data or architecture changes.

major comments (2)

[§3] The manuscript provides limited information on training data volume, exact GAN architecture details (layers, loss terms, conditioning implementation), validation splits, and error propagation in the reported statistics. This weakens support for the central performance claims of large-scale recovery and quantified small-scale deficits (abstract and §4).
[§4.2] The evaluation of halo abundance and profiles (around 10^{14} M_⊙) reports ~10% level effects but lacks discussion of how these residuals would propagate into cosmological parameter biases or survey forecasts, which is load-bearing for the claimed utility in downstream analyses (§4.2).

minor comments (2)

Figure captions and text should specify the number of realizations in the ensemble averages and the precise k-range or mass bins used for the power spectrum and halo comparisons to improve reproducibility.
[Abstract] The abstract could briefly note the redshift and parameter ranges over which the model was trained and tested.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and recommendation of minor revision. We address the two major comments below, providing additional details where possible and clarifying the scope of the present work.

read point-by-point responses

Referee: [§3] The manuscript provides limited information on training data volume, exact GAN architecture details (layers, loss terms, conditioning implementation), validation splits, and error propagation in the reported statistics. This weakens support for the central performance claims of large-scale recovery and quantified small-scale deficits (abstract and §4).

Authors: We agree that the current manuscript is concise on these technical aspects. In the revised version we will expand the methods section to report the precise training data volume (including the number of low- and high-resolution simulation pairs and their box sizes), the full GAN architecture (number of layers, filter counts, and activation functions), the explicit loss terms and weighting, the implementation of conditioning on Ω_k, Ω_m, σ_8, h and redshift, the train/validation/test split ratios, and a quantitative discussion of statistical uncertainties on the reported power-spectrum and halo statistics derived from multiple realizations. revision: yes
Referee: [§4.2] The evaluation of halo abundance and profiles (around 10^{14} M_⊙) reports ~10% level effects but lacks discussion of how these residuals would propagate into cosmological parameter biases or survey forecasts, which is load-bearing for the claimed utility in downstream analyses (§4.2).

Authors: We acknowledge that a dedicated propagation study would strengthen the case for downstream use. However, a full forecast analysis lies outside the scope of this emulator-development paper, which is focused on demonstrating fidelity on summary statistics. In the revision we will add a concise paragraph in §4.2 that (i) notes the ~10 % residuals are comparable to those reported for other emulators at similar scales, (ii) qualitatively indicates that such suppression could induce percent-level biases in σ_8 or halo-mass-function-derived parameters, and (iii) states that users should perform their own validation for precision applications. We will also cite relevant literature on emulator error propagation. revision: partial

Circularity Check

0 steps flagged

No significant circularity; evaluation uses independent high-resolution benchmarks

full rationale

The paper describes a conditional GAN trained to upscale low-resolution N-body fields into high-resolution realizations, with performance assessed via direct comparison of power spectra, halo mass functions, and density profiles against separate high-resolution simulation outputs. No load-bearing step reduces by construction to fitted parameters or self-referential definitions; the model is not claimed to derive new physics but to emulate existing simulation results, and residuals are explicitly quantified against external truth data rather than being tautological. No self-citation chain, uniqueness theorem, or ansatz smuggling is invoked to justify the core architecture or results. The derivation chain is therefore self-contained through empirical validation on held-out simulation suites.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard assumptions of N-body gravity and GAN training; no new free parameters, axioms, or invented entities are introduced in the abstract.

pith-pipeline@v0.9.0 · 5606 in / 1077 out tokens · 31049 ms · 2026-05-12T01:46:58.402104+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Our model is trained as a generative adversarial network... conditioned on a given low-resolution field, Ω_k, Ω_m, σ_8, h, and redshift.
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We find that the model accurately reproduces large-scale statistics, robustly recovering most of the power... residual suppression of power on small scales of up to ∼10% at k∼1 h Mpc^{-1}.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 1 internal anchor

[1]

Abdul Karim, M. et al. 2025, Phys. Rev. D, 112, 083515

work page 2025
[2]

& Boschetti, R

Adamek, J. & Boschetti, R. 2026, JCAP, 03, 013

work page 2026
[3]

Aghanim, N. et al. 2020, Astron. Astrophys., 641, A6, [Erratum: Astron. Astro- phys. 652, C4 (2021)]

work page 2020
[4]

2019, Mon

Alonso, D., Sanchez, J., & Slosar, A. 2019, Mon. Not. Roy. Astron. Soc., 484, 4127

work page 2019
[5]

Wasserstein GAN

Arjovsky, M., Chintala, S., & Bottou, L. 2017 [arXiv:1701.07875]

work page Pith review arXiv 2017
[6]

Bacon, D. J. et al. 2020, Publ. Astron. Soc. Austral., 37, e007

work page 2020
[7]

2011, JCAP, 07, 034

Blas, D., Lesgourgues, J., & Tram, T. 2011, JCAP, 07, 034

work page 2011
[8]

Castander, F. J. et al. 2025, Astron. Astrophys., 697, A5

work page 2025
[9]

& Zaldarriaga, M

Chen, S.-F. & Zaldarriaga, M. 2025, JCAP, 08, 014

work page 2025
[10]

E., Richardson, M

Chisari, N. E., Richardson, M. L. A., Devriendt, J., et al. 2018, Mon. Not. Roy. Astron. Soc., 480, 3962 Di Dio, E., Montanari, F., Lesgourgues, J., & Durrer, R. 2013, JCAP, 11, 044 Di Dio, E., Montanari, F., Raccanelli, A., et al. 2016, JCAP, 06, 013

work page 2018
[11]

S., & White, S

Efstathiou, G., Davis, M., Frenk, C. S., & White, S. D. M. 1985, Astrophys. J. Suppl., 57, 241

work page 1985
[12]

Generative Adversarial Networks

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., et al. 2014 [arXiv:1406.2661] Górski, K. M., Hivon, E., Banday, A. J., et al. 2005, Astrophys. J., 622, 759

work page internal anchor Pith review arXiv 2014
[13]

Improved Training of Wasserstein GANs

Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V ., & Courville, A. 2017 [arXiv:1704.00028]

work page Pith review arXiv 2017
[14]

2019, Mon

Huang, H.-J., Eifler, T., Mandelbaum, R., & Dodelson, S. 2019, Mon. Not. Roy. Astron. Soc., 488, 1652 Ivezi´c, Ž. et al. 2019, Astrophys. J., 873, 111

work page 2019
[15]

2020, Mon

Lepori, F., Adamek, J., Durrer, R., Clarkson, C., & Coates, L. 2020, Mon. Not. Roy. Astron. Soc., 497, 2078

work page 2020
[16]

Levi, M. et al. 2013 [arXiv:1308.0847]

work page Pith review arXiv 2013
[17]

2000, Astrophys

Lewis, A., Challinor, A., & Lasenby, A. 2000, Astrophys. J., 538, 473

work page 2000
[18]

2014, Phys

Li, Y ., Hu, W., & Takada, M. 2014, Phys. Rev. D, 89, 083519

work page 2014
[19]

Li, Y ., Ni, Y ., Croft, R. A. C., et al. 2020 [arXiv:2010.06608]

work page arXiv 2020
[20]

2021, Mon

Mead, A., Brieden, S., Tröster, T., & Heymans, C. 2021, Mon. Not. Roy. Astron. Soc., 502, 1401

work page 2021
[21]

Mellier, Y . et al. 2025, Astron. Astrophys., 697, A1

work page 2025
[22]

Newburgh, L. B. et al. 2016, Proc. SPIE Int. Soc. Opt. Eng., 9906, 99065X

work page 2016
[23]

2021, Mon

Ni, Y ., Li, Y ., Lachance, P., et al. 2021, Mon. Not. Roy. Astron. Soc., 507, 1021

work page 2021
[24]

K., Charnock, T., Villaescusa-Navarro, F., & Wandelt, B

Ramanah, D. K., Charnock, T., Villaescusa-Navarro, F., & Wandelt, B. D. 2020, Mon. Not. Roy. Astron. Soc., 495, 4227

work page 2020
[25]

2024, Phys

Rouhiainen, A., Münchmeyer, M., Shiu, G., Gira, M., & Lee, K. 2024, Phys. Rev. D, 109, 123536

work page 2024
[26]

2023, Mon

Sipp, M., LaChance, P., Croft, R., Ni, Y ., & Di Matteo, T. 2023, Mon. Not. Roy. Astron. Soc., 525, 1807

work page 2023
[27]

2025, yt-project/yt_astro_analysis: yt_astro_analysis 1.1.4 Release

Smith, B., Turk, M., ZuHone, J., et al. 2025, yt-project/yt_astro_analysis: yt_astro_analysis 1.1.4 Release

work page 2025
[28]

2021, Mon

Springel, V ., Pakmor, R., Zier, O., & Reinecke, M. 2021, Mon. Not. Roy. Astron. Soc., 506, 2871

work page 2021
[29]

J., Smith, B

Turk, M. J., Smith, B. D., Oishi, J. S., et al. 2011, The Astrophysical Journal Supplement Series, 192, 9

work page 2011
[30]

2018, Pylians: Python libraries for the analysis of numer- ical simulations, Astrophysics Source Code Library, record ascl:1811.008

Villaescusa-Navarro, F. 2018, Pylians: Python libraries for the analysis of numer- ical simulations, Astrophysics Source Code Library, record ascl:1811.008

work page 2018
[31]

Villaescusa-Navarro, F. et al. 2020, Astrophys. J. Suppl., 250, 2

work page 2020
[32]

2015, Mon

Wagner, C., Schmidt, F., Chiang, C.-T., & Komatsu, E. 2015, Mon. Not. Roy. Astron. Soc., 448, L11

work page 2015
[33]

2024, Mon

Zhang, X., Lachance, P., Ni, Y ., et al. 2024, Mon. Not. Roy. Astron. Soc., 528, 281 Article number, page 10 of 10

work page 2024