Pith. sign in

REVIEW 3 major objections 5 minor 69 references

Nearest-neighbour distance maps turn halo catalogues into continuous images that, combined with kNN-CDF statistics in a hybrid neural network, recover Omega_m and sigma_8 from the Quijote suite with R^2 of 0.80 and 0.93 respectively.

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 21:47 UTC pith:S3M6OSOL

load-bearing objection Useful new field-level representation and a sensible hybrid architecture, but the headline accuracy claims hinge on an unspecified train/test split and an overstatement of 'state-of-the-art'. the 3 major comments →

arxiv 2511.13393 v2 pith:S3M6OSOL submitted 2025-11-17 astro-ph.CO

Learning Cosmology from Nearest Neighbour Statistics

Atrideb Chatterjee , Arka Banerjee , Francisco Villaescusa-Navarro , Tom Abel This is my paper
classification astro-ph.CO PACS 98.80.-k95.75.Pq
keywords nearest neighbour statisticskNN-CDFdistance transformcosmological parameter inferenceOmega_msigma_8Quijote simulationshybrid neural network
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 introduces a new way to represent sparse galaxy or halo catalogues for machine learning: turning each point's location into a smooth map of distances to its k-th nearest neighbour. The authors train a hybrid neural network that takes both these nearest-neighbour distance maps and the existing kNN cumulative distribution functions as inputs, and show that this combination recovers the cosmological parameters Omega_m and sigma_8 from halo positions alone with far better accuracy than either input alone. They report R^2 scores of 0.80 for Omega_m and 0.93 for sigma_8, with relative errors near 15% and 3%, and note that their framework is 5–10 times more computationally efficient than point-cloud or graph-based methods. If correct, this offers a practical, cheap way to extract cosmological constraints from the huge halo and galaxy catalogues expected from upcoming surveys.

Core claim

A discrete catalogue of halo positions can be converted into a continuous field by assigning to every spatial point the distance to its k-th nearest halo, producing a set of 2D 'NN distance maps'. These maps preserve spatial information that the kNN-CDF summary statistics throw away, yet the maps alone are insufficient: a ResNet trained on them fails to constrain sigma_8. The central result is that a hybrid network combining the maps (via a ResNet-18 backbone) with the kNN-CDFs (via a fully connected network) performs best, achieving R^2 = 0.80 and 0.93 for Omega_m and sigma_8, respectively. This is the paper's main claim: that the joint use of nearest-neighbour distance maps and kNN-CDFs yi

What carries the argument

The key object is the nearest-neighbour distance map: for each pixel of a 2D slice through the simulation box, compute the distance to the 1st, 2nd, 3rd, or 4th nearest halo, producing a smooth image that encodes local clustering geometry (the 1NN map's critical points correspond to faces, edges, and nodes of the Voronoi tessellation of the halos). The second piece is the hybrid neural network architecture: a ResNet-18 processes the distance maps, its 512-dimensional feature vector is concatenated with the 200-bin (4 x 50) kNN-CDF vector, and an MLP regresses the posterior means and standard deviations of Omega_m and sigma_8. The maps carry spatial information; the CDFs carry global clusteri

Load-bearing premise

The reported accuracy figures assume the training, validation, and test sets are split simulation-by-simulation, so no map from a simulation used in training appears in the test set—but the paper only says 'We split the available input data into training (80%), validation (10%), and testing (10%)' and never states that the split is at the simulation level.

What would settle it

Retrain the hybrid Map+CDF model on the Quijote halo catalogues with a strict per-simulation split (e.g., leave out entire simulations for testing) and compare the test R^2 and relative errors to the paper's reported values. If the test R^2 for sigma_8 drops substantially below 0.93 or the relative error rises above 3%, the quoted metrics are inflated by data leakage. Alternatively, inspect the original split indices to check whether any test map shares a simulation ID with a training map.

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

If this is right

  • If the reported accuracy holds, upcoming surveys like Euclid, DESI, and LSST can extract Omega_m and sigma_8 from dense halo catalogues using standard CNNs on distance maps, avoiding the memory and compute costs of point-cloud networks.
  • The hybrid approach recovers sigma_8 (R^2 = 0.93) far better than either the maps alone (R^2 = 0.00) or the CDFs alone (R^2 = 0.59), indicating that the combination of local spatial fields and global statistics is a promising general recipe for likelihood-free inference.
  • A proof-of-concept with dark matter particles instead of halos yields even tighter constraints (R^2 = 0.98 for sigma_8), suggesting the method benefits from more complete tracer samples and may be applicable to galaxy catalogs with higher number densities.
  • The framework's 5–24x speedup over point-cloud/GNN baselines (about 2 hours on a single A100 GPU versus 2 days) could make field-level simulation-based inference tractable at survey scale.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the NN distance maps and kNN-CDFs are complementary views of the same underlying field—one local differential, the other global cumulative—and that this complementarity, rather than the specific architecture, is what drives the gain. A testable extension would be to feed the same two inputs into a simple concatenated MLP and check whether the improvement persists wit
  • An editorial concern: the reported numbers are only trustworthy if the 80/10/10 train/validation/test split is done per simulation, so no halo field from a training simulation appears in the test set. The paper does not state this explicitly; a map-level split would leak correlated data and inflate all metrics, especially for sigma_8.
  • The authors mention future work on including halo velocities and masses. A natural extension would be to concatenate scalar halo features (mass, velocity dispersion) as extra map channels, which could break the bias–sigma_8 degeneracy they suspect limits the maps-only model.
  • The method's reliance on a fixed number of query points (10^4 per slice) and 10 random slices per simulation suggests a sensitivity to map resolution and slice selection; an ablation varying these would clarify how robust the constraints are.

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

3 major / 5 minor

Summary. The paper introduces "nearest-neighbour (NN) distance maps," a field-level representation of discrete halo catalogues in which each pixel stores the distance to the k-th nearest halo, and combines these maps with kNN-CDF summary statistics as inputs to a neural network. Using 2,000 Quijote Latin Hypercube simulations and 10^5 most massive halos per simulation, the authors train three models: a CDF-only fully connected network, a ResNet-18 map-only model, and a hybrid map+CDF model, to predict Omega_m and sigma_8. The hybrid model is reported as best: R^2 = 0.80 and 0.93 and relative errors 15.3% and 3.2% for Omega_m and sigma_8, respectively (Table 1, Fig. 4). The paper also compares against the two-point correlation function and several recent Quijote-based machine-learning methods, claiming state-of-the-art accuracy and 5-10x computational efficiency.

Significance. The NN distance map idea is a natural and potentially useful extension of kNN-CDFs: it converts sparse halo catalogues into dense, CNN-friendly fields while preserving spatial information, and it avoids the memory and compute costs of point-cloud/graph methods. The paper uses the public Quijote suite and makes direct comparisons to several existing methods, which is helpful for the community. The specific claims, however, rest on the integrity of the train/test split and on the stability of the reported metrics. If the split is indeed simulation-level and the results are reproducible across seeds, the hybrid framework is a valuable, computationally cheap alternative for parameter inference from halo catalogues. As written, the central quantitative claims are not yet fully supported.

major comments (3)
  1. [Section 3.3 and Section 2.1] The train/validation/test split is not specified as being performed at the simulation level. Section 2.1 generates 20,000 maps from 2,000 simulations, with 10 maps per simulation. If the split described in Section 3.3 is a random map-level split, maps from the same simulation and the same cosmology will appear in both training and test sets. Because the 10 slices of a simulation share the same halo field and large-scale modes, the network can memorize realization-specific features rather than learning the Omega_m-sigma_8 dependence, which would inflate every metric in Table 1 and Fig. 4. The authors must state explicitly whether the split is by simulation, and if not, re-run with a simulation-level split. Even with a simulation-level split, the 10 maps per simulation are highly correlated; the validation metrics should either be computed per simulation or the effective number of independ
  2. [Section 4, Table 1] The reported metrics are based on a single training run with no repeated-seed or bootstrap error bars. For Omega_m, the difference between CDF-only and Map+CDF is small (R^2 = 0.79 vs 0.80; MSE = 2.9e-3 vs 2.6e-3). Without multiple training runs, it is not possible to assess whether the hybrid improvement is significant. Similarly, the Map-only sigma_8 result (R^2 = 0.00) could be a seed-dependent pathology. Please provide mean and standard deviation over at least 5 random seeds, or otherwise demonstrate that the rankings in Table 1 are stable.
  3. [Section 5 and Abstract] The abstract's claim of 'state-of-the-art, if not the best, accuracy' is not supported by the paper's own comparison. In Section 5, Cuesta-Lazaro & Mishra-Sharma (2023) is quoted as achieving ~5% relative error on Omega_m, while the present hybrid model reports 15.3%. The comparisons also use different halo selections (10^5 halos here versus 5,000 or 8,192 in the cited works), which is not an apples-to-apples comparison. Either temper the claim to 'competitive accuracy' or provide a matched comparison under identical halo selections and training setups.
minor comments (5)
  1. [Section 2.1, step 4] The wording 'we randomly select 10 maps corresponding to each NN distance' is ambiguous. If each map is a 4-channel stack of 1NN-4NN distances from the same slice, the total of 20,000 maps is consistent with 2,000 simulations x 10 slices. Please clarify this in the text.
  2. [Figure 1 caption] The caption says the actual maps are 'produced with 10^2 random query points' in a 100x100 grid, but a 100x100 grid has 10^4 query points. Please correct this typo.
  3. [Section 3.4 and Table 1] The chi^2 values are interpreted as 'close to 1 indicates calibrated errors,' but some entries are far from 1 (e.g., CDF-only Omega_m chi^2 = 0.65; Appendix A Map-only Omega_m chi^2 = 2.89). The statement in Section 5 that 'our study does not suffer from any such issues' needs more support than a single chi^2 value; a calibration plot or coverage test would be appropriate.
  4. [Section 5] The paper says the method is '5-10 times more computationally efficient than some of the existing point-cloud-based ML methods,' but later quotes a 24x speedup relative to Chatterjee & Villaescusa-Navarro (2025). Please make the efficiency comparison consistent and specify which methods the 5-10x factor refers to.
  5. [Reproducibility] Please report the random seed(s) used for the slice selection, the train/validation/test split, and the neural network initialization. No code or data products are provided, so these details are important for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the reported constraints are empirical supervised-learning results on held-out Quijote data, not derivations that reduce to their inputs.

full rationale

The paper's central quantitative claims (hybrid Map+CDF gives R^2 = 0.80 and 0.93 and relative errors ~15% and ~3% for Omega_m and sigma_8) are obtained by training neural networks on Quijote halo catalogues and evaluating on a held-out split (Sec. 3.3, Table 1, Fig. 4). This is a direct empirical measurement of generalization performance, not a quantity that is equal to its inputs by construction. The new NN distance maps are introduced as input features, and no equation in the paper defines the inferred parameters in terms of the training labels or vice versa. The kNN-CDF formalism is cited from prior work by the same group (Banerjee & Abel 2021a,b), but the present results do not depend on those cited theorems being true: the CDFs are used as network inputs and the accuracy is measured, not derived from the citations. Self-citations appear as background motivation and method context, not as load-bearing justifications for the headline accuracy. One non-circularity concern is that Sec. 3.3 states only 'We split the available input data into training (80%), validation (10%), and testing (10%)' without specifying whether the split is per simulation; since Sec. 2.1 generates 20,000 maps from 2,000 simulations, a map-level split could cause leakage and inflate the metrics. This is a statistical validity issue rather than a circular step, and without evidence it cannot support a circularity finding. The comparison with prior work is also explicit and admits where other methods perform better (e.g., Cuesta-Lazaro & Mishra-Sharma's ~5% Omega_m error vs the present 15.3%). Overall, no derivation in the paper is equivalent to its own input by definition, so the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claim is an empirical performance claim, so its ledger is dominated by modeling and training choices rather than physical entities. The network weights/hyperparameters and map/CDF construction parameters are free parameters fitted or selected to maximize test performance. The paper introduces no new physical entities; its new object is a data representation (NN distance maps). Axioms are mostly domain assumptions about the faithfulness of Quijote simulations and the sufficiency of the 2D slice representation; these are not proven in the text. No external real-data validation is provided, so the claim is confined to the simulation suite.

free parameters (3)
  • Neural network weights and hyperparameters (ResNet-18 backbone + inference MLP) = Optimized via OPTUNA over >100 trials; final layer sizes/learning rates/weight decay not tabulated
    The central result is the trained network's generalization performance; all network weights and hyperparameters are fitted to the Quijote training set.
  • Map-generation choices (10 random 2D slices per simulation, 100x100 query grid, 10 Mpc/h slice spacing, k=1..4, 50 CDF b
    These choices are hand-selected and directly affect the input representation and the reported metrics; no sensitivity study is provided.
  • Halo sample selection (10^5 most massive halos per simulation)
    The tracer population is chosen by hand; the effect of this mass cut on the learned Ωm-σ8 mapping is not tested.
axioms (6)
  • standard math kNN-CDFs capture combinations of all N-point functions of the underlying clustering (Banerjee & Abel 2021b).
    Invoked in Section 2 to motivate the information content of the CDFs; the paper does not re-derive this relation and relies on the cited work.
  • domain assumption Quijote Latin Hypercube halo catalogs accurately encode the cosmological parameter dependence of halo clustering at z=0 within the quoted parameter ranges.
    Section 2 uses these simulations as ground truth for training and testing; no convergence or resolution validation is presented.
  • domain assumption Ten 2D NN-distance slices per simulation retain sufficient 3D clustering information for inference.
    Section 2.1 selects 10 random 100x100 slices per simulation; the Map-only model is explicitly 2D, and the authors caution about comparing with 3D statistics (Section 4).
  • domain assumption The 10^5 most massive halos form a sufficient tracer sample; discarding lower-mass halos does not bias the learned Ωm-σ8 mapping.
    Section 2 states halos are downsampled to the 10^5 most massive; the effect of this cut on inference is not tested.
  • standard math The moment-matching loss (Eq. 4) yields calibrated posterior mean and standard deviation estimates (Jeffrey & Wandelt 2020).
    Section 3.2 adopts the loss without independent derivation; reported χ2 values (0.65–1.19) suggest only approximate calibration.
  • ad hoc to paper Random selection of 10 slices per simulation yields representative, non-redundant training samples.
    Section 2.1 step 4; no test of map independence or spatial correlation between slices is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 14582 in / 14194 out tokens · 134730 ms · 2026-08-03T21:47:53.371946+00:00 · methodology

0 comments
read the original abstract

Extracting cosmological parameters from galaxy/halo catalogues with sub-percent level accuracy is an important aspect of modern cosmology, especially in view of ongoing and upcoming surveys such as Euclid, DESI, and LSST. While traditional two-point statistics have been known to be suboptimal for this task, recently proposed k-Nearest Neighbour (kNN) based summary statistics have demonstrated tighter constraining power. Building on the kNN statistics, we introduce a new field-level representation of discrete halo catalogues - NN distance maps. We employ this technique on the halo catalogues obtained from Quijote N-body simulation suites. By combining these maps with kNN-based summary statistics, we train a hybrid neural network to infer cosmological parameters, showing that the resulting constraints achieve state-of-the-art, if not the best, accuracy. In addition, our hybrid framework is 5-10 times more computationally efficient than some of the existing point-cloud-based ML methods.

Figures

Figures reproduced from arXiv: 2511.13393 by Arka Banerjee, Atrideb Chatterjee, Francisco Villaescusa-Navarro, Tom Abel.

Figure 1
Figure 1. Figure 1: A 2D slice of the 1st (left) and 4th (right) Nearest neighbour distance maps for one of the simulations in the Quijote simulation suites used in this study. Each pixel is coloured by the distance from the pixel to the nearest data point on the left panel, and by the distance to the 4th nearest neighbour data point in the right panel. As can be seen, this converts the discrete dataset into a smooth, continu… view at source ↗
Figure 2
Figure 2. Figure 2: The CDF (left panel) and Peaked CDF (right panel) for 1NN (orange), 2NN (red), 3NN (magenta), and 4NN (blue) corresponding to one of the Quijote simulations in this study. of the learned features, i.e., the number of output channels from the ResNet (as shown in the figure) is kept to 512. These fea￾tures are then concatenated with the 1D summary statistics vec￾tor (of length 192, corresponding to the conca… view at source ↗
Figure 3
Figure 3. Figure 3: Hybrid Network in this study. The NN distance maps are used as input to the ResNet block. The output of the ResNet is then concatenated with the NN CDFs, and the merged input then passes through the inference blocks (containing several Linear, ReLU, and dropout layers) to predict the mean and standard deviation of the inferred cosmological parameters. The values in brackets show the dimension of the tensor… view at source ↗
Figure 4
Figure 4. Figure 4: The performance of different models when trained to predict likelihood-free inference on both the values of Ωm (left column) and σ8 (right column) in 3 scenarios: 1) top row: CDF-only 2)Middle row: Map-Only 3) bottom panel: Map+CDF. The values for different validation metrics are mentioned in the legend. As can be seen, the Map-only scenario (middle panel) is performing worse than the CDF-only (top panel) … view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between ξ(r)-Only and CDF-Only. As shown, CDF-Only performs much better compared to ξ(r)-Only as expected from (Banerjee & Abel 2021b). Article number, page 10 [PITH_FULL_IMAGE:figures/full_fig_p010_5.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

69 extracted references · 28 linked inside Pith

  1. [1]

    E., et al

    Abazajian, K., Abdulghafour, A., Addison, G. E., et al. 2022, arXiv e-prints, arXiv:2203.08024

    Pith/arXiv arXiv 2022

  2. [2]

    2020, Phys

    Ajani, V ., Peel, A., Pettorino, V ., et al. 2020, Phys. Rev. D, 102, 103531

    2020

  3. [3]

    2019, in The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2623–2631

    Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. 2019, in The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2623–2631

    2019

  4. [4]

    2022, arXiv e-prints, arXiv:2211.12346

    Anagnostidis, S., Thomsen, A., Kacprzak, T., et al. 2022, arXiv e-prints, arXiv:2211.12346

    Pith/arXiv arXiv 2022

  5. [5]

    Anbajagane, D. et al. 2023, Mon. Not. Roy. Astron. Soc., 526, 5530

    2023

  6. [6]

    F., Rocha, A

    Azevedo, B. F., Rocha, A. M. A., & Pereira, A. I. 2024, Machine Learning, 113, 4055

    2024

  7. [7]

    & Wandelt, B

    Bairagi, A. & Wandelt, B. 2025, arXiv e-prints, arXiv:2509.03165

    Pith/arXiv arXiv 2025

  8. [8]

    & Abel, T

    Banerjee, A. & Abel, T. 2022, Monthly Notices of the Royal Astronomical So- ciety, 519, 4856

    2022

  9. [9]

    2022, Mon

    Banerjee, A., Kokron, N., & Abel, T. 2022, Mon. Not. Roy. Astron. Soc., 511, 2765

    2022

  10. [10]

    E., Villaescusa-Navarro, F., Massara, E., et al

    Bayer, A. E., Villaescusa-Navarro, F., Massara, E., et al. 2021, Detecting neutrino mass by combining matter clustering, halos, and voids

    2021

  11. [11]

    2014, arXiv e-prints, arXiv:1403.5237

    Benitez, N., Dupke, R., Moles, M., et al. 2014, arXiv e-prints, arXiv:1403.5237

    Pith/arXiv arXiv 2014

  12. [12]

    2025, Monthly Notices of the Royal Astronomical Society, 538, 2204

    Chand, E., Banerjee, A., Foreman, S., & Villaescusa-Navarro, F. 2025, Monthly Notices of the Royal Astronomical Society, 538, 2204

    2025

  13. [13]

    & Villaescusa-Navarro, F

    Chatterjee, A. & Villaescusa-Navarro, F. 2025, ApJ, 985, 132

    2025

  14. [14]

    Chen, X., Yang, Q., Wu, J., Li, H., & Tan, K. C. 2023, arXiv e-prints, arXiv:2305.16594

    Pith/arXiv arXiv 2023

  15. [15]

    Colas, T., d’Amico, G., Senatore, L., Zhang, P., & Beutler, F. 2020, J. Cosmology Astropart. Phys., 2020, 001

    2020

  16. [16]

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

    Coulton, W. R., Abel, T., & Banerjee, A. 2024, Mon. Not. Roy. Astron. Soc., 534, 1621

    2024

  17. [17]

    & Mishra-Sharma, S

    Cuesta-Lazaro, C. & Mishra-Sharma, S. 2023, arXiv e-prints, arXiv:2311.17141 d’Amico, G., Gleyzes, J., Kokron, N., et al. 2020, J. Cosmology Astropart. Phys., 2020, 005

    Pith/arXiv arXiv 2023

  18. [18]

    J., et al

    Dattilo, A., Vanderburg, A., Shallue, C. J., et al. 2019, AJ, 157, 169 de Santi, N. S. M., Shao, H., Villaescusa-Navarro, F., et al. 2023, ApJ, 952, 69

    2019

  19. [19]

    Demiss, B. A. & Elsaigh, W. A. 2024, Engineering Research Express, 6, 032102 DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016, arXiv e-prints, arXiv:1611.00036

    Pith/arXiv arXiv 2024

  20. [20]

    2022, arXiv e-prints, arXiv:2204.07646 Euclid Collaboration: Castro, T., Fumagalli, A., Angulo, R

    Eickenberg, M., Allys, E., Moradinezhad Dizgah, A., et al. 2022, arXiv e-prints, arXiv:2204.07646 Euclid Collaboration: Castro, T., Fumagalli, A., Angulo, R. E., et al. 2022, arXiv e-prints, arXiv:2208.02174

    Pith/arXiv arXiv 2022

  21. [21]

    2019, Phys

    Fluri, J., Kacprzak, T., Lucchi, A., et al. 2019, Phys. Rev. D, 100, 063514

    2019

  22. [22]

    2020, MNRAS, 498, 464

    Friedrich, O., Uhlemann, C., Villaescusa-Navarro, F., et al. 2020, MNRAS, 498, 464

    2020

  23. [23]

    2025, arXiv e-prints, arXiv:2502.07713

    Gangopadhyay, K., Banerjee, A., & Abel, T. 2025, arXiv e-prints, arXiv:2502.07713

    arXiv 2025

  24. [24]

    2019, MNRAS, 484, 282

    Gillet, N., Mesinger, A., Greig, B., Liu, A., & Ucci, G. 2019, MNRAS, 484, 282

    2019

  25. [25]

    & Smith, K

    Giri, U. & Smith, K. M. 2020, arXiv e-prints, arXiv:2010.07193

    Pith/arXiv arXiv 2020

  26. [26]

    2021, arXiv e-prints, arXiv:2104.03976

    Gualdi, D., Gil-Marin, H., & Verde, L. 2021, arXiv e-prints, arXiv:2104.03976

    Pith/arXiv arXiv 2021

  27. [27]

    2020, arXiv e-prints, arXiv:2009.02290

    Gualdi, D., Novell, S., Gil-Marín, H., & Verde, L. 2020, arXiv e-prints, arXiv:2009.02290

    Pith/arXiv arXiv 2020

  28. [28]

    Gupta, K. R. & Banerjee, A. 2024, Mon. Not. Roy. Astron. Soc., 531, 4619

    2024

  29. [29]

    2020, Jour- nal of Cosmology and Astroparticle Physics, 2020, 040–040

    Hahn, C., Villaescusa-Navarro, F., Castorina, E., & Scoccimarro, R. 2020, Jour- nal of Cosmology and Astroparticle Physics, 2020, 040–040

    2020

  30. [30]

    S., Habaebi, M

    Halbouni, A., Gunawan, T. S., Habaebi, M. H., et al. 2022, IEEE Access, 10, 99837 Harnois-Déraps, J., Martinet, N., Castro, T., et al. 2021, MNRAS, 506, 1623

    2022

  31. [31]

    2022, ApJ, 937, 83

    Hassan, S., Villaescusa-Navarro, F., Wandelt, B., et al. 2022, ApJ, 937, 83

    2022

  32. [32]

    2016, in 2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR, 1

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in 2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR, 1

    2016

  33. [33]

    J., Chartier, N., et al

    Ho, M., Bartlett, D. J., Chartier, N., et al. 2024, arXiv e-prints, arXiv:2402.05137

    Pith/arXiv arXiv 2024

  34. [34]

    Hortua, H. J. 2021, arXiv e-prints, arXiv:2112.11865

    Pith/arXiv arXiv 2021

  35. [35]

    2025, arXiv e-prints, arXiv:2507.03707 Ivezi´c, Ž., Kahn, S

    Huang, N., Stiskalek, R., Lee, J.-Y ., et al. 2025, arXiv e-prints, arXiv:2507.03707 Ivezi´c, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111

    arXiv 2025

  36. [36]

    2020, arXiv e-prints, arXiv:2009.08459

    Jeffrey, N., Alsing, J., & Lanusse, F. 2020, arXiv e-prints, arXiv:2009.08459

    Pith/arXiv arXiv 2020

  37. [37]

    & Wandelt, B

    Jeffrey, N. & Wandelt, B. D. 2020, arXiv e-prints, arXiv:2011.05991

    Pith/arXiv arXiv 2020

  38. [38]

    2006, IEEE Transactions on Visualiza- tion and Computer Graphics, 12, 581 Article number, page 7 A&A proofs:manuscript no

    Jones, M., Baerentzen, J., & Sramek, M. 2006, IEEE Transactions on Visualiza- tion and Computer Graphics, 12, 581 Article number, page 7 A&A proofs:manuscript no. main

    2006

  39. [39]

    Kingma, D. P. & Ba, J. 2014, CoRR, abs/1412.6980

    Pith/arXiv arXiv 2014

  40. [40]

    Krause, E. et al. 2025, Astrophys. J., 990, 99

    2025

  41. [41]

    Y ., Abrams, N., Andrews, J., et al

    Lam, C. Y ., Abrams, N., Andrews, J., et al. 2023, arXiv e-prints, arXiv:2306.12514

    Pith/arXiv arXiv 2023

  42. [42]

    2011, arXiv e-prints, arXiv:1110.3193

    Laureijs, R., Amiaux, J., Arduini, S., et al. 2011, arXiv e-prints, arXiv:1110.3193

    Pith/arXiv arXiv 2011

  43. [43]

    & Villaescusa-Navarro, F

    Lee, J.-Y . & Villaescusa-Navarro, F. 2025, arXiv e-prints, arXiv:2505.23904

    Pith/arXiv arXiv 2025

  44. [44]

    2022, arXiv e-prints, arXiv:2204.02945 Lucas Makinen, T., Heavens, A., Porqueres, N., et al

    Liu, W., Jiang, A., & Fang, W. 2022, arXiv e-prints, arXiv:2204.02945 Lucas Makinen, T., Heavens, A., Porqueres, N., et al. 2025, J. Cosmology As- tropart. Phys., 2025, 095

    Pith/arXiv arXiv 2022

  45. [45]

    L., Charnock, T., Lemos, P., et al

    Makinen, T. L., Charnock, T., Lemos, P., et al. 2022, The Open Journal of Astro- physics, 5, 18

    2022

  46. [46]

    A., Liu, J., Zorrilla Matilla, J

    Marques, G. A., Liu, J., Zorrilla Matilla, J. M., et al. 2019, J. Cosmology As- tropart. Phys., 2019, 019 Mobina Hosseini, S. & Soleimanpour Salmasi, B. 2025, arXiv e-prints, arXiv:2508.05842

    Pith/arXiv arXiv 2019

  47. [47]

    2021, arXiv e-prints, arXiv:2111.12088

    Naidoo, K., Massara, E., & Lahav, O. 2021, arXiv e-prints, arXiv:2111.12088

    Pith/arXiv arXiv 2021

  48. [48]

    J., Yuan, S., & Garrison, L

    Ntampaka, M., Eisenstein, D. J., Yuan, S., & Garrison, L. H. 2020, ApJ, 889, 151

    2020

  49. [49]

    2019, PyTorch: An Imperative Style, High- Performance Deep Learning Library

    Paszke, A., Gross, S., Massa, F., et al. 2019, PyTorch: An Imperative Style, High- Performance Deep Learning Library

    2019

  50. [50]

    D., Laureijs, R., Stagnaro, L., et al

    Racca, G. D., Laureijs, R., Stagnaro, L., et al. 2016, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, V ol. 9904, Space Tele- scopes and Instrumentation 2016: Optical, Infrared, and Millimeter Wave, ed. H. A. MacEwen, G. G. Fazio, M. Lystrup, N. Batalha, N. Siegler, & E. C. Tong, 99040O

    2016

  51. [51]

    2017, arXiv e-prints, arXiv:1711.02033

    Ravanbakhsh, S., Oliva, J., Fromenteau, S., et al. 2017, arXiv e-prints, arXiv:1711.02033

    Pith/arXiv arXiv 2017

  52. [52]

    2023, arXiv e-prints, arXiv:2311.01588

    Roncoli, A., ´Ciprijanovi´c, A., V oetberg, M., Villaescusa-Navarro, F., & Nord, B. 2023, arXiv e-prints, arXiv:2311.01588

    Pith/arXiv arXiv 2023

  53. [53]

    2021, arXiv e-prints, arXiv:2102.01696

    Samushia, L., Slepian, Z., & Villaescusa-Navarro, F. 2021, arXiv e-prints, arXiv:2102.01696

    Pith/arXiv arXiv 2021

  54. [54]

    1996, Proc

    Schmalzing, J., Kerscher, M., & Buchert, T. 1996, Proc. Int. Sch. Phys. Fermi, 132, 281

    1996

  55. [55]

    2022, arXiv e-prints, arXiv:2209.06843

    Shao, H., Villaescusa-Navarro, F., Villanueva-Domingo, P., et al. 2022, arXiv e-prints, arXiv:2209.06843

    Pith/arXiv arXiv 2022

  56. [56]

    2019, in 2019 Asia-Pacific signal and information processing association annual summit and conference (AP- SIPA ASC), IEEE, 939–944

    Shi, X., Wang, T., Wang, L., Liu, H., & Yan, N. 2019, in 2019 Asia-Pacific signal and information processing association annual summit and conference (AP- SIPA ASC), IEEE, 939–944

    2019

  57. [57]

    Siraj, M. S. & Ahad, M. 2020, in 2020 Joint 9th International Conference on

    2020

  58. [58]

    Informatics, Electronics & Vision (ICIEV) and 2020 4th International Con- ference on Imaging, Vision & Pattern Recognition (icIVPR), IEEE, 1–7

    2020

  59. [59]

    2020, MNRAS, 495, 4006

    Uhlemann, C., Friedrich, O., Villaescusa-Navarro, F., Banerjee, A., & Codis, S. 2020, MNRAS, 495, 4006

    2020

  60. [60]

    & Dvorkin, C

    Valogiannis, G. & Dvorkin, C. 2021, arXiv e-prints, arXiv:2108.07821

    Pith/arXiv arXiv 2021

  61. [61]

    2022, ApJS, 259, 61

    Villaescusa-Navarro, F., Genel, S., Anglés-Alcázar, D., et al. 2022, ApJS, 259, 61

    2022

  62. [62]

    2020, The Astrophysical Journal Supplement Series, 250, 2

    Villaescusa-Navarro, F., Hahn, C., Massara, E., et al. 2020, The Astrophysical Journal Supplement Series, 250, 2

    2020

  63. [63]

    Villaescusa-Navarro, F. et al. 2022, Astrophys. J., 929, 132

    2022

  64. [64]

    & Villaescusa-Navarro, F

    Villanueva-Domingo, P. & Villaescusa-Navarro, F. 2022, ApJ, 937, 115

    2022

  65. [65]

    2021, arXiv e-prints, arXiv:2111.14874

    Villanueva-Domingo, P., Villaescusa-Navarro, F., Genel, S., et al. 2021, arXiv e-prints, arXiv:2111.14874

    Pith/arXiv arXiv 2021

  66. [66]

    2022, Mon

    Wang, Y ., Banerjee, A., & Abel, T. 2022, Mon. Not. Roy. Astron. Soc., 514, 3828

    2022

  67. [67]

    2023, Monthly Notices of the Royal Astronom- ical Society, 522, 3935

    Yuan, S., Zamora, A., & Abel, T. 2023, Monthly Notices of the Royal Astronom- ical Society, 522, 3935

    2023

  68. [68]

    2022, Ieee Access, 10, 47361

    Zeng, C., Ma, C., Wang, K., & Cui, Z. 2022, Ieee Access, 10, 47361

    2022

  69. [69]

    2025, Astrophys

    Zhou, Z., Cisewski-Kehe, J., Fang, K., & Banerjee, A. 2025, Astrophys. J., 979, 194 Article number, page 8 Atrideb Chatterjee et al.: Learning Cosmology from Nearest Neighbour Statistics 0.1 0.2 0.3 0.4 0.5 Truth 0.1 0.2 0.3 0.4 0.5Prediction CDF-Only m R2=0.79 =17.3% 2 = 0.65 MSE = 2.90e-03 0.6 0.7 0.8 0.9 1.0 Truth 0.6 0.7 0.8 0.9 1.0Prediction CDF-Only...

    2025