Pith. sign in

REVIEW 2 major objections 4 minor 49 references

Applying standard reconstruction before a CNN shifts the best input scale down to about forty to one hundred Mpc and improves accuracy.

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

2026-08-01 22:07 UTC pith:2IMFP4LA

load-bearing objection Solid, useful measurement of an input-scale shift after standard reconstruction, but the scan conflates field-of-view with resolution, so the design guidance is strong while the physical interpretation is only plausible. the 2 major comments →

arxiv 2607.15850 v1 pith:2IMFP4LA submitted 2026-07-17 astro-ph.CO

Standard Reconstruction Shifts the Optimal Input Scale for CNN-Based Density-Field Reconstruction

Koichiro Nakashima , Kiyotomo Ichiki , Atsushi J. Nishizawa This is my paper
classification astro-ph.CO
keywords density-field reconstructionconvolutional neural networksstandard reconstructionZel'dovich approximationlarge-scale structureinput scalebaryon acoustic oscillationshybrid methods
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 asks what physical scale of input a convolutional network needs to reconstruct the early (z=10) dark-matter density from the late-time (z=0) density, and whether that scale changes if a standard first-order reconstruction step is applied first. The authors find that without reconstruction the best input cube is large, about 150-200 h^-1 Mpc, because the network must infer long-wavelength bulk flows; after standard reconstruction removes those flows, the preferred cube shrinks to about 38-114 h^-1 Mpc, and the single small-cube hybrid (stdrec.+CNN) is more accurate than both a single large-cube CNN and a dual-input CNN that combines two scales. If this is right, analytic reconstruction and neural networks naturally divide the work: perturbative methods handle coherent large-scale displacements, CNNs handle residual quasi-linear and non-linear structure, and a single small input cube suffices. The result also explains why the roughly 60 h^-1 Mpc receptive scale adopted in earlier hybrid studies works well.

Core claim

Using fixed 39^3-voxel input cubes drawn from dark-matter-only N-body simulations and varying their physical side length from ~38 to ~380 h^-1 Mpc, the paper claims that the optimal input scale for a pointwise CNN reconstructing the z=10 density from z=0 shifts from ~150-200 h^-1 Mpc when applied directly to the evolved field to ~38-114 h^-1 Mpc when a standard first-order (Zel'dovich-based) reconstruction is applied first. The single-input hybrid (stdrec.+CNN) is reported to outperform the single-input CNN and the dual-input CNN at every statistic examined: normalized loss, residual field, one-point PDF and KL divergence, and Fourier-space correlation. Adding a second, large-scale input bra

What carries the argument

The central object is standard first-order reconstruction (the Zel'dovich/LGS scheme), which estimates the large-scale displacement field from a smoothed density field and shifts mass tracers back by that displacement, approximately undoing bulk flows whose variance is dominated by modes with wavelengths of several hundred Mpc. The paper couples this to a pointwise CNN architecture with a limited receptive field (39^3-voxel input, 3D convolutions) and a controlled scale scan: keeping the voxel count fixed at 39^3 while varying the physical cell size, so larger L_sub gives more spatial context but coarser resolution, and vice versa. The interaction of the analytic removal of non-local displac

Load-bearing premise

The comparison varies L_sub by changing the physical cell size while holding the voxel count fixed, so the preference for smaller L_sub after reconstruction could partly reflect the higher input resolution rather than a reduced need for large-scale context; the paper smooths target fields to a common 3 h^-1 Mpc scale to mitigate this, but cannot fully eliminate the effect.

What would settle it

A controlled test that keeps the physical field of view fixed while varying resolution - for example, an L_sub=76 h^-1 Mpc cube with 39^3 voxels versus the same L_sub with 79^3 voxels - would show whether the post-reconstruction gain comes from scale or resolution. Additionally, evaluating stdrec.+CNN on several independent test realizations would check whether its reported advantage over the dual-input CNN holds across cosmic variance rather than on a single realization.

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

If this is right

  • A hybrid pipeline that runs standard reconstruction before a CNN is more accurate than either a CNN-only or a dual-scale CNN, and it needs only one small input cube, a computational saving.
  • The preferred post-reconstruction input range of roughly 38-114 h^-1 Mpc brackets the ~60 h^-1 Mpc effective receptive scale adopted in earlier hybrid studies, providing an independent justification for that design choice.
  • Once standard reconstruction is applied, the large-scale branch of a dual-input CNN becomes redundant because the large-scale information has already been incorporated analytically.
  • The shift in optimal scale implies a division of labour: perturbative reconstruction handles coherent long-wavelength displacements, while CNN handles localized residual evolution.
  • The preferred small input cube has finer effective resolution (cell sizes ~1-3 h^-1 Mpc), which preserves more quasi-linear cosmic-web structure for the CNN to model.

Where Pith is reading between the lines

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

  • If the scale shift reflects a reduced need for long-wavelength context rather than resolution, then the optimal input size should track the amplitude of the displacement field: repeating this scan at intermediate redshifts (z=1-2), where bulk flows are weaker, would be a direct testable extension.
  • The results suggest that applying standard reconstruction before training could make the input-scale choice less sensitive to survey volume, which matters for realistic survey geometries with limited contiguous volume; this is an inference about applicability, not a claim the paper makes.
  • A cleaner separation of physical scale from resolution would compare configurations with the same field of view but different voxel sizes (e.g., 39^3 at 76 Mpc versus 79^3 at 76 Mpc); the common 3 h^-1 Mpc target smoothing mitigates but may not completely remove this degeneracy.

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

2 major / 4 minor

Summary. This paper studies how the optimal physical size of a 39^3-voxel CNN input for predicting the z=10 dark-matter density from z=0 LSS changes when the input is first processed by standard (Zel'dovich) reconstruction. Using Indra N-body simulations, the authors compare a single-input CNN, a dual-input CNN, and a single-input CNN applied after standard reconstruction ("stdrec.+CNN"), scanning L_sub ~38-380 h^-1 Mpc. They report that the optimal L_sub shifts from ~150-200 h^-1 Mpc for the CNN-only model to ~38-114 h^-1 Mpc for stdrec.+CNN, with the latter outperforming both CNN-only baselines on normalized loss, PDF/KL divergence, residual maps, and Fourier-space correlation. They interpret this as evidence for a physically motivated separation of scales: standard reconstruction handles coherent long-wavelength displacements, while the CNN models residual quasi-linear and non-linear evolution on smaller scales.

Significance. If correct, the result has practical value for hybrid reconstruction design and provides an independent explanation for the ~60 h^-1 Mpc effective receptive scale adopted by Shallue & Eisenstein (2023). The paper's strengths include measuring the optimum directly from simulation output rather than fitting it, checking multiple independent summary statistics, and using public simulation and reconstruction software. However, two issues prevent the central physical interpretation from being fully supported: the L_sub scan varies field of view and resolution simultaneously, and the final comparisons rest on a single test realization with no cross-realization error bars.

major comments (2)
  1. [Section 2.3 / Table 2 and Section 4.1] L_sub and input resolution are perfectly confounded: at fixed 39^3 voxels, L_sub=38 h^-1 Mpc has 0.97 h^-1 Mpc cells while L_sub=380 has 9.76 h^-1 Mpc cells. The 3 h^-1 Mpc smoothing of the z=10 target does not equalize the CNN input resolution; for large L_sub the input Nyquist frequency is below that kernel. The paper concedes in Section 2.3 that this may not completely eliminate resolution effects. Hence the post-reconstruction preference for small L_sub could simply reflect the usefulness of finer voxels once standard reconstruction has removed large-scale coherent displacements, rather than a reduced need for long-wavelength context. To support the physical claim of Section 4.1, the authors need a control that varies field of view independently of resolution (e.g., fixed cell size with variable voxel count, or a scale-decomposition analysis).
  2. [Section 2.3; Figures 1, 4, 5] The final comparison is based on one independent test realization; the error bars in Fig. 1 are iteration-to-iteration scatter of the same realization, not cosmic variance. Therefore the claims that "stdrec.+CNN consistently outperforms" and that the optimum lies in a narrow 38-114 h^-1 Mpc range are not supported by a realization-level uncertainty estimate. Neighboring L_sub values (e.g., 76 vs 114) show small differences in several panels; multiple test realizations or a bootstrap over subvolumes are needed to establish the ranking and the preferred range robustly.
minor comments (4)
  1. [Section 2.3] The sentence "We reconstruct the dark matter density at z=10 from the density field at z=0." appears twice consecutively; remove the duplicate.
  2. [Section 2.4 / Eq. (15)] Because the batch size changes during training and the normalized loss uses the batch variance, the normalization in Eq. (15) is not constant across iterations. Stating this explicitly would help reproducibility.
  3. [Section 5] The preferred range 38-114 h^-1 Mpc is a synthesis of different metric-specific optima (e.g., Fourier correlation favors 38-76, loss/KL favor 76-114). The conclusions should state these metric-dependent ranges to avoid over-smoothing the distinction.
  4. [Section 2.1] The standard reconstruction smoothing scale R=20 h^-1 Mpc is fixed. The optimal post-reconstruction CNN input scale may depend on R, since R sets which displacements are removed; this should be acknowledged as a limitation of the quantitative range, even if the qualitative shift is robust.

Circularity Check

0 steps flagged

No derivation-to-fit circularity; optimal scale is measured, not fitted. Minor self-citation in baselines is the only reason the score is not 0.

full rationale

I find no circular derivation. The central quantity, the preferred input scale L_sub, is obtained as the empirical argmin of separately measured loss, KL-divergence, and Fourier-correlation curves (Figs 1, 4, 5), not by substituting Eq. 13 into a target formula. Changing L_sub at fixed 39^3 voxels does change l_cell^sub, but this is a resolution-versus-context degeneracy that the paper explicitly flags: Section 4.1 states 'The preferred L_sub should therefore not be interpreted as a pure measurement of the required physical scales,' and Section 2.3 concedes that common Gaussian smoothing 'may not completely eliminate the effects of the differing grid resolutions.' That is a validity caveat about the physical interpretation of the shift, not a circular reduction: the optimum is still measured, not imposed by construction. The only self-citation is the use of the authors' own Nakashima et al. (2025) baselines, with the paper stating 'The results for "CNN only" and "Dual-input CNN" are taken from Nakashima et al. (2025), while the results for "stdrec.+CNN" are obtained in this work.' That supplies comparison numbers from a separate, published prior study rather than deriving the present result from itself, and the new stdrec.+CNN outputs are trained and evaluated independently in this paper. The consistency with Shallue & Eisenstein (2023)'s ~60 h^-1 Mpc receptive scale is a post-hoc cross-check, not an input to the scan. No equation reduces to a fitted value, no uniqueness theorem is imported, and no ansatz is smuggled in via citation. Score 2 reflects the minor self-citation in the baselines; the central derivation is self-contained.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper is an empirical comparison, so its conclusions rest less on derivation assumptions and more on evaluation choices: smoothing scales, grid-resolution handling, and single-realization testing. The central scale shift is a measured quantity rather than a fitted formula, so there are no invented entities and no strong circularity burden.

free parameters (3)
  • standard reconstruction smoothing scale R = 20 h^-1 Mpc
    Chosen in Sec. 2.3 without a scan. This scale sets which displacements are removed before the CNN sees the field and could directly influence the measured optimal L_sub.
  • target Gaussian smoothing scale = 3 h^-1 Mpc
    Chosen in Sec. 2.3 to make targets comparable across parent-grid resolutions. It suppresses fluctuations below this scale and enters every metric.
  • CNN hyperparameters (batch schedule, learning rate, iterations) = from Mao et al. (2021)
    Fixed, not tuned here (Sec. 2.4). The measured optimal scale could depend on network capacity and training length, though all models share the same protocol.
axioms (4)
  • domain assumption The Zel'dovich/first-order LPT displacement relation (Eq. 3) is an adequate description of the large-scale bulk flow that standard reconstruction removes.
    Invoked in Sec. 2.1 and Sec. 4.2. If the displacement estimate is biased, the scale shift and hybrid advantage could be artifacts of an imperfect reconstruction rather than a physical separation of scales.
  • domain assumption The 3 h^-1 Mpc-smoothed z=10 target provides a comparable ground truth across all L_sub configurations despite differing parent-grid resolutions.
    Sec. 2.3 states smoothing may not completely eliminate resolution differences. The cross-configuration comparison relies on this assumption.
  • domain assumption Dark-matter-only N-body simulations with WMAP7 cosmology and 1024^3 particles in a 1 h^-1 Gpc box faithfully represent the relevant density evolution for this idealized comparison.
    Dataset description in Sec. 2.3. Sec. 5 explicitly lists redshift-space distortions, tracer bias, and shot noise as future work, so the central claim is scoped to this idealized setup.
  • domain assumption A single independent test realization is sufficient to rank methods and locate the optimum.
    Sec. 2.3 evaluates on one independent realization; no standard errors are reported for the final field-level ranking in Figs. 2-5.

pith-pipeline@v1.3.0-alltime-deepseek · 14890 in / 13046 out tokens · 127385 ms · 2026-08-01T22:07:13.783164+00:00 · methodology

0 comments
read the original abstract

We investigate convolutional neural network (CNN) methods for reconstructing the high-redshift density field from late-time large-scale structure, focusing on how the physical scale of the CNN input changes when standard first-order reconstruction is applied beforehand. Using dark-matter-only $N$-body simulations, we compare three approaches: a single-input CNN, a dual-input CNN combining two physical scales, and a single-input CNN applied to the density field after standard reconstruction. We vary the physical side length of the input sub-box over $L_\mathrm{sub}\sim38$-$380~h^{-1}\mathrm{Mpc}$ while keeping its numerical size fixed at $39^3$ voxels, allowing us to examine the trade-off between spatial context and resolution. For the CNN applied directly to the evolved density field, the reconstruction performs best at $L_\mathrm{sub}\sim150$-$200~h^{-1}\mathrm{Mpc}$. After standard reconstruction, however, the preferred scale shifts to $L_\mathrm{sub}\sim38$-$114~h^{-1}\mathrm{Mpc}$. The single-input CNN after standard reconstruction consistently outperforms both the single- and dual-input CNNs without standard reconstruction according to the normalized loss, density probability distribution, Kullback-Leibler divergence, residual field, and Fourier-space correlation. These results indicate that coherent large-scale displacements are more efficiently recovered by perturbative reconstruction, while the CNN is better suited to modelling the remaining quasi-linear and non-linear evolution on smaller scales. The preferred post-reconstruction input range includes the effective receptive scale of approximately $60~h^{-1}\mathrm{Mpc}$ adopted in previous hybrid reconstruction studies. Our findings therefore support a physically motivated separation of scales between analytic and data-driven reconstruction and demonstrate the advantage of combining the two approaches.

Figures

Figures reproduced from arXiv: 2607.15850 by Atsushi J. Nishizawa, Kiyotomo Ichiki, Koichiro Nakashima.

Figure 1
Figure 1. Figure 1: Left: Training and validation losses, defined by equation (15), for three representative models: the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc, corresponding to the setup of Mao et al. (2021); the dual-input CNN with 𝐿sub ∼ {76, 380} ℎ −1 Mpc, taken from Nakashima et al. (2025); and the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc applied after standard reconstruction. Training and validation curves are disting… view at source ↗
Figure 2
Figure 2. Figure 2: Slices through the residual field 𝛿recon − 𝛿target, with a slice thickness of 1.95 ℎ −1 Mpc, for three reconstruction methods. From left to right, the panels show the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc, the dual-input CNN with 𝐿sub ∼ {76, 380} ℎ −1 Mpc, and the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc applied after standard reconstruction [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two-dimensional histograms of 𝛿target(𝑧 = 10) versus 𝛿recon for the test data set. Left: the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc; Centre: the dual-input CNN with 𝐿sub ∼ {76, 380} ℎ −1 Mpc; Right: the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc applied after standard reconstruction. The colour bar represents log10 𝑁, where 𝑁 is the number of data points in each histogram bin. standard reconstruction befor… view at source ↗
Figure 4
Figure 4. Figure 4: Left: Probability distribution functions (PDFs) of the density fluctuation 𝛿. Solid and dotted lines denote the PDFs of the target 𝛿target(𝑧 = 10) and input 𝛿input(𝑧 = 0) density fields, respectively. The PDFs of the CNN output 𝛿recon are shown for three representative setups: the single-input CNN with 𝐿sub ∼ 76 ℎ −1 Mpc, the dual-input CNN with 𝐿sub ∼ {76, 380} ℎ −1 Mpc, and the single-input CNN with 𝐿sub… view at source ↗
Figure 5
Figure 5. Figure 5: Correlation coefficient between the reconstructed and target density fields, as defined in equation (17). Left : Scale-dependent correlations for the different reconstruction methods. “No reconstruction” denotes the direct comparison between the 𝑧 = 0 and 𝑧 = 10 density fields, while “stdrec. only” denotes the density field obtained using standard reconstruction without a subsequent CNN. Right : Correlatio… 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

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

  1. [1]

    , keywords =

    An iterative reconstruction of cosmological initial density fields. , keywords =. doi:10.1093/mnras/sty1203 , archivePrefix =. 1804.04738 , primaryClass =

    Pith/arXiv arXiv

  2. [2]

    , keywords =

    Measuring the cosmological density field twice: A novel test of dark energy using the CMB quadrupole. , keywords =. doi:10.1103/PhysRevD.105.063507 , archivePrefix =. 2202.11332 , primaryClass =

    Pith/arXiv arXiv

  3. [3]

    , keywords =

    Baryon acoustic oscillations reconstruction using convolutional neural networks. , keywords =. doi:10.1093/mnras/staa3741 , archivePrefix =. 2002.10218 , primaryClass =

    Pith/arXiv arXiv 2002

  4. [4]

    , keywords =

    Indra: a public computationally accessible suite of cosmological N-body simulations. , keywords =. doi:10.1093/mnras/stab1823 , archivePrefix =. 2101.03631 , primaryClass =

    Pith/arXiv arXiv

  5. [5]

    , keywords =

    The cosmological simulation code GADGET-2. , keywords =. doi:10.1111/j.1365-2966.2005.09655.x , archivePrefix =. astro-ph/0505010 , primaryClass =

    arXiv 2005

  6. [6]

    , keywords =

    Seven-year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Interpretation. , keywords =. doi:10.1088/0067-0049/192/2/18 , archivePrefix =. 1001.4538 , primaryClass =

    Pith/arXiv arXiv

  7. [7]

    , title =

    Nair, Vinod and Hinton, Geoffrey E. , title =. Proceedings of the 27th International Conference on International Conference on Machine Learning , pages =. 2010 , isbn =

    2010

  8. [8]

    , title =

    Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E. , title =. Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 1 , pages =. 2012 , publisher =

    2012

  9. [9]

    arXiv e-prints , keywords =

    Adam: A Method for Stochastic Optimization. arXiv e-prints , keywords =. doi:10.48550/arXiv.1412.6980 , archivePrefix =. 1412.6980 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980

  10. [10]

    Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , pages =

    Understanding the difficulty of training deep feedforward neural networks , author =. Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , pages =. 2010 , editor =

    2010

  11. [11]

    arXiv e-prints , keywords =

    PyTorch: An Imperative Style, High-Performance Deep Learning Library. arXiv e-prints , keywords =. doi:10.48550/arXiv.1912.01703 , archivePrefix =. 1912.01703 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.01703 1912

  12. [12]

    , keywords =

    Reconstructing cosmological initial conditions from late-time structure with convolutional neural networks. , keywords =. doi:10.1093/mnras/stad528 , archivePrefix =. 2207.12511 , primaryClass =

    Pith/arXiv arXiv

  13. [13]

    , keywords =

    Getting around cosmic variance. , keywords =. doi:10.1103/PhysRevD.56.4511 , archivePrefix =. astro-ph/9703118 , primaryClass =

    Pith/arXiv arXiv

  14. [14]

    , keywords =

    Polarization signal of distant clusters and reconstruction of primordial potential fluctuations. , keywords =. doi:10.1103/PhysRevD.62.123004 , archivePrefix =. astro-ph/0009222 , primaryClass =

    Pith/arXiv arXiv

  15. [15]

    , keywords =

    Observational probes of cosmic acceleration. , keywords =. doi:10.1016/j.physrep.2013.05.001 , archivePrefix =. 1201.2434 , primaryClass =

    Pith/arXiv arXiv 2013

  16. [16]

    , keywords =

    Analysis of the Kamionkowski-Loeb method of reducing cosmic variance with CMB polarization. , keywords =. doi:10.1103/PhysRevD.70.063504 , archivePrefix =. astro-ph/0402173 , primaryClass =

    Pith/arXiv arXiv

  17. [17]

    , keywords =

    Non-Gaussianity from inflation: theory and observations. , keywords =. doi:10.1016/j.physrep.2004.08.022 , archivePrefix =. astro-ph/0406398 , primaryClass =

    Pith/arXiv arXiv 2004

  18. [18]

    , keywords =

    Tracing Large-Scale Fluctuations Back in Time. , keywords =. doi:10.1086/171360 , adsurl =

    doi:10.1086/171360

  19. [19]

    , year = 1970, month = mar, volume =

    Gravitational instability: An approximate theory for large density perturbations. , year = 1970, month = mar, volume =

    1970

  20. [20]

    , keywords =

    Improving Cosmological Distance Measurements by Reconstruction of the Baryon Acoustic Peak. , keywords =. doi:10.1086/518712 , archivePrefix =. astro-ph/0604362 , primaryClass =

    Pith/arXiv arXiv

  21. [21]

    , keywords =

    Nonlinear Structure Formation and the Acoustic Scale. , keywords =. doi:10.1086/589921 , archivePrefix =. 0805.0117 , primaryClass =

    Pith/arXiv arXiv

  22. [22]

    , keywords =

    Reconstructing baryon oscillations. , keywords =. doi:10.1103/PhysRevD.80.123501 , archivePrefix =. 0909.1802 , primaryClass =

    Pith/arXiv arXiv

  23. [23]

    , keywords =

    Reconstructing baryon oscillations: A Lagrangian theory perspective. , keywords =. doi:10.1103/PhysRevD.79.063523 , archivePrefix =. 0812.2905 , primaryClass =

    Pith/arXiv arXiv

  24. [24]

    , keywords =

    Eulerian BAO reconstructions and N -point statistics. , keywords =. doi:10.1103/PhysRevD.92.123522 , archivePrefix =. 1508.06972 , primaryClass =

    Pith/arXiv arXiv

  25. [25]

    Methods and application to the Sloan Digital Sky Survey

    A 2 per cent distance to z = 0.35 by reconstructing baryon acoustic oscillations - I. Methods and application to the Sloan Digital Sky Survey. , keywords =. doi:10.1111/j.1365-2966.2012.21888.x , archivePrefix =. 1202.0090 , primaryClass =

    arXiv 2012

  26. [26]

    , keywords =

    Measuring D _ A and H at z=0.35 from the SDSS DR7 LRGs using baryon acoustic oscillations. , keywords =. doi:10.1093/mnras/stt379 , archivePrefix =. 1206.6732 , primaryClass =

    Pith/arXiv arXiv

  27. [27]

    , keywords =

    Measuring the 2D baryon acoustic oscillation signal of galaxies in WiggleZ: cosmological constraints. , keywords =. doi:10.1093/mnras/stw2725 , archivePrefix =. 1611.08040 , primaryClass =

    Pith/arXiv arXiv

  28. [28]

    , year = 2015, month = may, volume =

    Deep learning. , year = 2015, month = may, volume =. doi:10.1038/nature14539 , adsurl =

    doi:10.1038/nature14539 2015

  29. [29]

    Deep Learning , author=

  30. [30]

    arXiv e-prints , keywords =

    Initial Conditions from Galaxies: Machine-Learning Subgrid Correction to Standard Reconstruction. arXiv e-prints , keywords =. doi:10.48550/arXiv.2504.01092 , archivePrefix =. 2504.01092 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.01092

  31. [31]

    , keywords =

    Effective cosmic density field reconstruction with convolutional neural network. , keywords =. doi:10.1093/mnras/stad1868 , archivePrefix =. 2306.10538 , primaryClass =

    Pith/arXiv arXiv

  32. [32]

    Primordial density and BAO reconstruction

    Primordial density and BAO reconstruction. arXiv e-prints , keywords =. doi:10.48550/arXiv.1609.07041 , archivePrefix =. 1609.07041 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1609.07041

  33. [33]

    , keywords =

    Nonlinear reconstruction. , keywords =. doi:10.1103/PhysRevD.96.123502 , archivePrefix =. 1611.09638 , primaryClass =

    Pith/arXiv arXiv

  34. [34]

    , keywords =

    Increasing Fisher information by Potential Isobaric Reconstruction. , keywords =. doi:10.1093/mnras/stx774 , archivePrefix =. 1611.10013 , primaryClass =

    Pith/arXiv arXiv

  35. [35]

    , keywords =

    Isobaric Reconstruction of the Baryonic Acoustic Oscillation. , keywords =. doi:10.3847/2041-8213/aa738c , archivePrefix =. 1703.09742 , primaryClass =

    Pith/arXiv arXiv 2041

  36. [36]

    , keywords =

    Halo Nonlinear Reconstruction. , keywords =. doi:10.3847/1538-4357/aa89e7 , archivePrefix =. 1703.08301 , primaryClass =

    Pith/arXiv arXiv

  37. [37]

    , keywords =

    Understanding the Reconstruction of the Biased Tracer. , keywords =. doi:10.3847/1538-4357/aaf231 , archivePrefix =. 1807.06381 , primaryClass =

    Pith/arXiv arXiv

  38. [38]

    , keywords =

    Nonlinear reconstruction of redshift space distortions. , keywords =. doi:10.1103/PhysRevD.97.043502 , archivePrefix =. 1711.03218 , primaryClass =

    Pith/arXiv arXiv

  39. [39]

    , keywords =

    Iterative initial condition reconstruction. , keywords =. doi:10.1103/PhysRevD.96.023505 , archivePrefix =. 1704.06634 , primaryClass =

    Pith/arXiv arXiv

  40. [40]

    , keywords =

    Towards optimal extraction of cosmological information from nonlinear data. , keywords =. doi:10.1088/1475-7516/2017/12/009 , archivePrefix =. 1706.06645 , primaryClass =

    Pith/arXiv arXiv 2017

  41. [41]

    , keywords =

    New method for initial density reconstruction. , keywords =. doi:10.1103/PhysRevD.97.023505 , archivePrefix =. 1709.06350 , primaryClass =

    Pith/arXiv arXiv

  42. [42]

    , keywords =

    Iterative removal of redshift-space distortions from galaxy clustering. , keywords =. doi:10.1093/mnras/staa2136 , archivePrefix =. 1912.03392 , primaryClass =

    Pith/arXiv arXiv 1912

  43. [43]

    , keywords =

    Searching optimal scales for reconstructing cosmological initial conditions using convolutional neural networks. , keywords =. doi:10.1093/mnras/staf1802 , archivePrefix =. 2505.10636 , primaryClass =

    Pith/arXiv arXiv

  44. [44]

    , keywords =

    nbodykit: An Open-source, Massively Parallel Toolkit for Large-scale Structure. , keywords =. doi:10.3847/1538-3881/aadae0 , archivePrefix =. 1712.05834 , primaryClass =

    Pith/arXiv arXiv

  45. [45]

    arXiv e-prints , keywords =

    Field-Level Inference from Galaxies: BAO Reconstruction. arXiv e-prints , keywords =. doi:10.48550/arXiv.2603.15732 , archivePrefix =. 2603.15732 , primaryClass =

    doi:10.48550/arxiv.2603.15732

  46. [46]

    , keywords =

    Neural network reconstruction of non-Gaussian initial conditions from dark matter halos. , keywords =. doi:10.1088/1475-7516/2025/08/030 , archivePrefix =. 2502.11846 , primaryClass =

    Pith/arXiv arXiv 2025

  47. [47]

    , keywords =

    Cosmological reconstruction from galaxy light: neural network based light-matter connection. , keywords =. doi:10.1088/1475-7516/2018/10/028 , archivePrefix =. 1805.02247 , primaryClass =

    Pith/arXiv arXiv 2018

  48. [48]

    , title =

    Jasche, Jens and Wandelt, Benjamin D. , title =. Monthly Notices of the Royal Astronomical Society , volume =. 2013 , month =. doi:10.1093/mnras/stt449 , url =

    doi:10.1093/mnras/stt449 2013

  49. [49]

    arXiv e-prints , keywords =

    CosmicRIM : Reconstructing Early Universe by Combining Differentiable Simulations with Recurrent Inference Machines. arXiv e-prints , keywords =. doi:10.48550/arXiv.2104.12864 , archivePrefix =. 2104.12864 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2104.12864