Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

This paper claims that a neural network whose architecture hard-codes the 90-degree rotation and mirror symmetry of galaxy shapes, calibrated by an analytic response calculation, can measure weak-lensing shear with multiplicative bias below

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 05:44 UTC pith:UAV6DFUZ

load-bearing objection A credible proof-of-concept that D4-equivariant CNNs plus analytic gradient calibration can meet Stage-IV shear-bias requirements on isolated galaxies; the isotropy caveat and a few fixable presentation issues don't undercut it. the 2 major comments →

arxiv 2603.19046 v2 pith:UAV6DFUZ submitted 2026-03-19 astro-ph.IM astro-ph.CO

D₄CNNtimesAnaCal: Physics-Informed Machine Learning for Accurate and Precise Weak Lensing Shear Estimation

Shurui Lin , Xiangchong Li , Ji Li , Shengcao Cao , Xin Liu , Yu-Xiong Wang This is my paper
classification astro-ph.IM astro-ph.CO
keywords weak gravitational lensingshear estimationequivariant neural networksanalytic calibrationshape noisegalaxy shape measurementcosmic sheardeep learning
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.

The paper argues that the two things that make machine-learning shear estimators untrustworthy—uncontrolled symmetries and heuristic calibration—can be removed by construction. It builds a convolutional network whose architecture hard-codes the D4 symmetry (90-degree rotations and mirror flips) of galaxy ellipticity, and calibrates the output with the Analytic Calibration framework, which computes the shear response analytically from backpropagated gradients. On Stage-IV survey-style isolated-galaxy simulations, calibrated multiplicative biases satisfy |m| < 10^-3 across noise, PSF, and selection variations, with most measurements near 10^-4, while shape noise drops about 10% in the high-noise regime. If this holds in real data, machine-learning shape measurement can combine the flexibility of deep networks with the rigor of analytic self-calibration.

Core claim

The central claim is that if the shape-measurement network is exactly equivariant under the dihedral group D4 (90-degree rotations and mirrors) and is smooth, then all even-order terms in the ellipticity–shear response vanish, so after a single linear analytic calibration the estimator is unbiased to third order in shear. The paper demonstrates that the combination—a D4-equivariant CNN with smooth activations calibrated via backpropagated gradients using analytic pixel-level shear-response formulas—achieves multiplicative bias |m| < 10^-3 (most at ~10^-4), additive bias at the ~10^-5 level, and about 10% lower shape noise than the moment-based Fourier Power Function Shapelets estimator in th

What carries the argument

The D4-equivariant CNN (D4CNN). The network runs the same CNN backbone over all eight images in the D4 orbit (four rotations plus mirrors), maps the features back to the reference frame, and combines them with signed weights to form spin-2 features; an odd tanh MLP preserves the sign, guaranteeing that the predicted ellipticity transforms exactly as a spin-2 quantity. The Analytic Calibration machinery supplies the complementary piece: because the network is smooth, the Jacobian of the output with respect to the input image is obtained by backpropagation, and the pixel-level shear response of the re-smoothed image is computed analytically in Fourier space; their contraction gives the respons

Load-bearing premise

The guarantee that even-order shear-response terms vanish assumes that the ensemble of intrinsic galaxy orientations is statistically isotropic (D4-symmetric); if real galaxies have any preferred orientation, second-order terms reappear as additive bias no matter how symmetric the network is.

What would settle it

Measure the shear response and residual additive bias on a sample in which galaxy orientations are not uniformly distributed—for example, a field with coherent intrinsic alignments or a detector with a preferred pixel-grid orientation. Under the paper's claim, the second-order term in the response should become measurable and the additive bias should scale with the orientation anisotropy; if the additive bias stays at the ~10^-5 level despite a strongly anisotropic orientation distribution, the D4 assumption is not the limiting factor.

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

If this is right

  • Machine-learning shear estimators can meet the sub-0.2% multiplicative-bias requirement of Stage-IV surveys across a wide range of noise, PSF, and selection conditions.
  • Calibration no longer requires re-rendering sheared simulations for every data set: the response is computed from the network's own gradients plus analytic pixel-response formulas, making ML shear estimation fast (about 1 ms per galaxy).
  • The ~10% shape-noise reduction in the high-noise regime gives a ~20% effective gain in galaxy number density, increasing the statistical power of cosmic-shear surveys or allowing fainter galaxies to be used.
  • Because the framework only requires a smooth, appropriately symmetric model, it applies beyond CNNs to any differentiable architecture, including deeper residual networks or vision transformers.
  • Extension to blended sources and multi-band data is the stated next step, with blending currently the dominant source of bias in stamp-based estimators.

Where Pith is reading between the lines

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

  • If the result transfers to blended scenes, the same analytic-response idea may remove the dominant blend-induced bias in current pipelines, since neighbor gradients can be propagated through the response matrix without re-rendering images.
  • The factor-of-eight parameter reduction from hard-coded symmetry explains why training on only 10,000 postage stamps suffices; this suggests the method may transfer to a new instrument or PSF with a small recalibration set, which could be tested by fine-tuning on a single field.
  • A direct empirical prediction of the theory is that residual bias after calibration is cubic in shear; verifying the 1:64 ratio between bias at shear 0.01 and 0.04 would isolate the claimed third-order behavior.
  • The shape-noise reduction with no change in training target suggests the network acts as a learned denoiser for ellipticity; a testable extension is whether the gain grows with network depth, as the paper anticipates with residual or transformer backbones.

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 / 6 minor

Summary. The paper presents D4CNN×AnaCal, a machine-learning shear estimator that hard-codes D4 (90-degree rotation and mirror) equivariance in a continuously differentiable CNN and calibrates its output analytically using the AnaCal framework. The shear response is computed as the contraction of the network's backpropagated gradient with an analytic pixel-level shear response, giving a linear calibration without image re-renderings. The authors claim that, under D4 symmetry, the estimator is unbiased to third order in shear, and they validate this on LSST-like single-band simulations of isolated galaxies with Gaussian noise and known PSF. They report multiplicative biases satisfying |m|<1e-3 (abstract says 2e-3 in one version) across noise, PSF, and magnitude-cut variations, and a ~10% reduction in shape noise relative to FPFS at high noise, equivalent to a ~20% effective number-density gain. Ablations show that equivariance suppresses quadratic response terms and that smooth activations reduce scatter.

Significance. If the results hold, this is an important advance: it combines a physically motivated symmetry with analytic, gradient-based self-calibration, avoiding the need for calibration image re-renderings while retaining ML flexibility. The empirical program is strong: large paired simulations (40 million galaxies per setup), bootstrap error estimates, tests across noise/PSF/magnitude, and a clean ablation that isolates the effects of D4 equivariance and smooth activations. The planned public code release will aid reproducibility. The main caveat is that the theoretical third-order-unbiased claim rests on an unstated statistical-isotropy assumption, which needs to be made explicit and tested.

major comments (2)
  1. [2.1.1, Eqs. (7)–(8)] The claim that 'D4 symmetry of ellipticity would force the expectation of all even-order terms to be zero' is cited to prior work rather than proven. More importantly, D4 equivariance of the network alone does not imply <∂²e/∂γ²>=0 for an arbitrary selected ensemble. For a fixed galaxy the quadratic response is generally nonzero; D4 symmetry only relates the coefficient of a galaxy to that of its 90-degree-rotated counterpart. The cancellation therefore requires the ensemble of intrinsic orientations to be D4-symmetric (statistical isotropy). The simulations enforce this via random rotations and orthogonal pairs (Sec. 2.2.2), but the assumption is not stated as a premise of Eq. (8). Real observations with intrinsic alignments or orientation-dependent selection can violate it; then second-order terms contribute to additive bias c that the linear calibration of Eq. (15) does not remove. Pl
  2. [3.2 / Fig. 6] The ~10% shape-noise improvement is a headline quantitative claim, but Figure 6 plots no error bars. The text says the uncertainty is at the ~1e-4 level, but the reader cannot verify the significance of the red-vs-blue differences from the figure. Please add error bars or state that they are smaller than the markers, and report the measured values with uncertainties for both ellipticity components at each noise level.
minor comments (6)
  1. [Abstract / Sec. 5] The abstract quotes 'all measurements satisfying |m|<2×10^{-3}' while the full-text abstract and Section 5 quote |m|<10^{-3}. These are a factor of two apart; harmonize and specify whether the quoted bound is the worst-case or typical value.
  2. [Sec. 3.1.1] The text says 'we generate 20 million paired galaxies,' but Sec. 2.2.2 defines 40 million galaxies consisting of 20 million orthogonal pairs. Clarify whether 'paired galaxies' means pairs or individual galaxies.
  3. [Sec. 3.2 / Fig. 6 caption] State explicitly that the ML model in this figure is trained with FPFS ellipticity labels; this explains why the low-noise shape-noise values coincide with FPFS.
  4. [Sec. 5, 'High efficiency' bullet] The bullet quotes '0.7 ms per galaxy' but the following comparison uses '1–1.5 ms per galaxy.' Reconcile the two timing estimates.
  5. [Figs. 2, 3, 7, 8] The activation is referred to as both 'GeLU' and 'GELU'; the non-equivariant model is called both 'CNN' and 'Non-Eq.' Standardize the notation and define model labels once.
  6. [Sec. 2.3.2] The statement that hard-coded symmetry reduces 'the effective number of free parameters by a factor of eight' is heuristic; the orbit-averaging construction does not literally reduce the parameter count eightfold. Please clarify or soften.

Circularity Check

0 steps flagged

No significant circularity: reported biases are measured against independently injected shears; AnaCal and the D4-symmetry theorem are prior published, externally validated results, and the central derivation does not reduce to its inputs by construction.

full rationale

The derivation chain is: Taylor-expand the model ellipticity response (Eq. 6), invoke the D4-symmetry result to drop even-order terms (Eqs. 7-8), form the linearly calibrated estimator gammahat = <e>/<R> (Eqs. 10-15), and validate on simulations with known injected shear using the standard m and c definitions (Eqs. 22-23). The reported biases are measured against independently injected shears; they are not fitted predictions. The same response used for calibration enters the m definition, but that is a standard residual-bias test, not a by-construction equality. The central D4 theorem is cited to X. Li et al. 2024, a coauthor paper, but it is a published, externally validated mathematical result with stated assumptions, and the present paper independently confirms its content via the non-equivariant CNN ablation (Table 2 and Figure 8 show the expected quadratic term when D4 symmetry is absent). Similarly, the AnaCal pixel-shear-response formulas (Eqs. 13-14) come from prior published work with independent HSC/LSST validation, not from the present target result. The main caveat is that the second-order cancellation requires an ensemble D4-symmetric (statistically isotropic) distribution of intrinsic orientations; the simulations enforce this explicitly, but the paper does not qualify the theoretical guarantee for real anisotropic ensembles. That is a correctness/assumption gap, not circularity. There is also a minor internal inconsistency between the abstract's |m| < 2e-3 and Section 5's |m| < 1e-3. Overall, no step reduces a prediction to its inputs by definition or by a self-citation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No ad hoc numerical constants are fitted to make the central claim: the reported m and c biases are measured on injected shears, not tuned. The framework rests on several imported or assumed pieces — the D4-symmetric shape ensemble, the analytic pixel shear-response formula from prior AnaCal work, the adequacy of the trained gradient as the true response, and the fidelity of the idealized simulations. No new physical entities are introduced.

axioms (5)
  • domain assumption The intrinsic shape distribution of source galaxies is D4-symmetric (invariant under 90-degree rotations and mirror), so ensemble-averaged even-order terms in the shear response vanish (Eqs 7-8).
    The theoretical claim of third-order-only bias relies on this; in simulations it is enforced by random rotations and orthogonal pairs, but for real data it assumes statistical isotropy.
  • domain assumption The analytic pixel-level shear response of the resmoothed image (Eqs 13-14) is correct for the discrete, pixelated, finite stamps used.
    This is imported from X. Li 2025 and is load-bearing: the calibration response R is the product of this analytic derivative and the network gradient.
  • domain assumption The trained D4CNN is a sufficiently accurate and smooth approximation of the true ellipticity mapping that its backpropagated gradient equals the true response.
    The calibration uses the model gradient; if the model is poorly trained, the response is wrong. The paper validates empirically with injected shears, but this is not guaranteed by construction.
  • domain assumption Fourier-based resmoothing/re-noising in Eq 3 removes noise bias and PSF anisotropy (standard AnaCal result).
    Imported from prior AnaCal papers; underpins the definition of the input f_h and the response formula.
  • domain assumption The DC1-based xlens simulations faithfully represent LSST i-band coadds for isolated galaxies with Gaussian noise and known PSF.
    All empirical claims are demonstrated on these simulations; the paper explicitly limits its scope to this idealized setting.

pith-pipeline@v1.3.0-alltime-deepseek · 19209 in / 19380 out tokens · 183691 ms · 2026-08-04T05:44:11.010957+00:00 · methodology

0 comments
read the original abstract

Traditional weak gravitational lensing shear estimators are carefully calibrated but struggle to fully capture realistic galaxy morphologies, point-spread-function (PSF) effects, blending, and noise in deep surveys, while blindly trained machine learning (ML) models can introduce significant calibration biases. Here we construct a fully D$_4$-equivariant deep neural network for galaxy shape measurement whose architecture enforces symmetry under 90$^{\circ}$ rotations and mirror transformations, and adopt the Analytical Calibration framework (AnaCal) to calibrate the model using its backpropagated gradients. For isolated galaxies in LSST-like single-band simulations, we demonstrate that our approach achieves $\sim$10% lower shape noise than the traditional moment-based Fourier Power Function Shapelets estimator in the high-noise regime, equivalent to a $\sim$20% gain in effective galaxy number density, while simultaneously achieving multiplicative biases consistent with zero across a wide range of noise levels, PSF sizes and ellipticities, and magnitude selection cuts, with all measurements satisfying $|m| {<} 2 \times 10^{-3}$ (i.e., within the 0.2% LSST requirement) and most at the ${\sim}10^{-4}$ level. We demonstrate this framework on isolated single-band galaxy images with Gaussian noise and known PSF, establishing a rigorous, physics-informed foundation for future extensions of ML-based shear estimation to blended sources and multi-band observations in Stage-IV surveys. All codes and data products will be made publicly available upon acceptance.

Figures

Figures reproduced from arXiv: 2603.19046 by Ji Li, Shengcao Cao, Shurui Lin, Xiangchong Li, Xin Liu, Yu-Xiong Wang.

Figure 1
Figure 1. Figure 1: Workflow of the analytical shear calibration for an ML–based shear estimator. Starting from a galaxy image, the ML model produces a raw shape estimate and corresponding model gradient via backpropagation. In parallel, the shear response of the resmoothed image is computed analytically at the pixel level using the expressions in Equation 13. The contraction of the model gradient with the pixel shear respons… view at source ↗
Figure 2
Figure 2. Figure 2: Example architecture of the D4-equivariant shape measurement model (D4CNN). The input galaxy image is transformed over the full D4 orbit (four rotations and their mirrored counterparts), and we forward the CNN eight times on each of the eight transformed variants to produce a set of features. These features are then mapped back to the original reference frame and combined through a weighted average to cons… view at source ↗
Figure 3
Figure 3. Figure 3: Transformation of measured shape under rotation for our D4CNN. The same galaxy profile with fixed ellipticity (e = 0.4) but different orientations (−90◦ to +90◦ between the x-axis and the major axis) is fed into the model. The predicted e1 and e2 components (blue and orange points) are compared against theoretical sine/cosine functions (dashed and dash-dotted curves) scaled to match the measured am￾plitude… view at source ↗
Figure 4
Figure 4. Figure 4: Shear estimation results for galaxies with mi < 25.3 under variations in image noise level (top row), PSF ellipticity (middle row), and PSF full width at half maximum (bottom row). The left panels show the multiplicative bias m1 (blue) and m2 (red), while the right panels show the additive bias c1 (blue) and c2 (red). Data points are horizontally slightly shifted from the original value to avoid overlap. P… view at source ↗
Figure 5
Figure 5. Figure 5: Shear estimation results for galaxies with different i-band magnitude cuts. The shaded gray region in the multiplicative-bias panels denotes the LSST requirement of |m| < 0.002. Left panels show the multiplicative bias m1 (blue) and m2 (red), while the right panels show the additive bias c1 (blue) and c2 (red). Points indicate the mean bias estimates, with error bars corresponding to 1σ (deep color) and 3σ… view at source ↗
Figure 6
Figure 6. Figure 6: Shape noise comparison between our CNN model (red squares) and FPFS (blue circles) as a function of image noise level. The highest noise level corresponds to the standard deviation expected for LSST 10-year coadded i-band images. While both methods achieve comparable shape noise at low noise levels, the CNN model shows a ∼10% reduction in shape noise in the high-noise regime, demonstrating the improved noi… view at source ↗
Figure 7
Figure 7. Figure 7: Transformation of measured shape under rotation for our Non-Eq model as in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Non-linear response of the three models to applied shear. Galaxies with identical circular profiles but different input shear g are fed into the models. The linear response of the measured ellipticity is fitted and subtracted, and the residual ei −linear fit is plotted as a function of the input shear. The top and bottom rows correspond to the two ellipticity components e1 and e2, respectively. The D4CNN (… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Slay the Shear: A Unified Statistical Framework for Weak Gravitational Lensing Shear Estimation

    astro-ph.CO 2026-06 unverdicted novelty 6.0

    Unified framework proves the score function yields the minimum-variance unbiased shear estimator and that response-weighted inverse-variance weights minimize shape noise independent of galaxy shape distributions, with...

Reference graph

Works this paper leans on

49 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [1]

    2018, PASJ, 70, S4, doi: 10.1093/pasj/psx066 Astropy Collaboration, Robitaille, T

    Aihara, H., Arimoto, N., Armstrong, R., et al. 2018, PASJ, 70, S4, doi: 10.1093/pasj/psx066 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068 Astropy Collaboration, Price-Whelan, A. M., Sipőcz, B. M., et al. 2018, AJ, 156, 123, doi: 10.3847/1538-3881/aabc4f Astropy Collaboration, Price-...

    doi:10.1093/pasj/psx066 2018

  2. [2]

    2001, Physics Reports, 340, 291, doi: 10.1016/S0370-1573(00)00082-X 16Lin et al

    Bartelmann, M., & Schneider, P. 2001, Physics Reports, 340, 291, doi: 10.1016/S0370-1573(00)00082-X 16Lin et al

    doi:10.1016/s0370-1573(00)00082-x 2001

  3. [3]

    2025, doi: 10.48550/ARXIV.2505.00093

    Berlfein, F., Mandelbaum, R., Li, X., et al. 2025, doi: 10.48550/ARXIV.2505.00093

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.00093 2025

  4. [4]

    M., & Jarvis, M

    Bernstein, G. M., & Jarvis, M. 2002, The Astronomical Journal, 123, 583, doi: 10.1086/338085

    doi:10.1086/338085 2002

  5. [5]

    S., & Welling, M

    Cohen, T. S., & Welling, M. 2016, arXiv e-prints, arXiv:1602.07576, doi: 10.48550/arXiv.1602.07576

    Pith/arXiv arXiv doi:10.48550/arxiv.1602.07576 2016

  6. [6]

    A., Schneider, M

    Dawson, W. A., Schneider, M. D., Tyson, J. A., & Jee, M. J. 2016, ApJ, 816, 11, doi: 10.3847/0004-637X/816/1/11

    doi:10.3847/0004-637x/816/1/11 2016

  7. [7]

    W., & Dambre, J

    Dieleman, S., Willett, K. W., & Dambre, J. 2015, MNRAS, 450, 1441, doi: 10.1093/mnras/stv632

    doi:10.1093/mnras/stv632 2015

  8. [8]

    2020, arXiv e-prints, arXiv:2010.11929

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. 2020, arXiv e-prints, arXiv:2010.11929. https://arxiv.org/abs/2010.11929 Euclid Collaboration, Mellier, Y., Abdurro’uf, et al. 2024, doi: 10.48550/ARXIV.2405.13491

    Pith/arXiv arXiv 2020

  9. [9]

    2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  10. [10]

    2016, arXiv e-prints, arXiv:1606.08415, doi: 10.48550/arXiv.1606.08415

    Hendrycks, D., & Gimpel, K. 2016, arXiv e-prints, arXiv:1606.08415, doi: 10.48550/arXiv.1606.08415

    Pith/arXiv arXiv doi:10.48550/arxiv.1606.08415 2016

  11. [11]

    Huber, P. J. 1964, The Annals of Mathematical Statistics, 35, 73 , doi: 10.1214/aoms/1177703732

    arXiv 1964

  12. [12]

    2017, arXiv e-prints, arXiv:1702.02600, doi: 10.48550/arXiv.1702.02600 Ivezić, Ž., Kahn, S

    Huff, E., & Mandelbaum, R. 2017, arXiv e-prints, arXiv:1702.02600, doi: 10.48550/arXiv.1702.02600 Ivezić, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042c

    Pith/arXiv arXiv doi:10.48550/arxiv.1702.02600 2017

  13. [13]

    2000, The Astrophysical Journal, 537, 555, doi: 10.1086/309041

    Kaiser, N. 2000, The Astrophysical Journal, 537, 555, doi: 10.1086/309041

    doi:10.1086/309041 2000

  14. [14]

    E., Kevrekidis, I

    Karniadakis, G. E., Kevrekidis, I. G., Lu, L., et al. 2021, Nature Reviews Physics, 3, 422, doi: 10.1038/s42254-021-00314-5

    doi:10.1038/s42254-021-00314-5 2021

  15. [15]

    2015, Reports on Progress in Physics, 78, 086901, doi: 10.1088/0034-4885/78/8/086901 Lei Ba, J., Kiros, J

    Kilbinger, M. 2015, Reports on Progress in Physics, 78, 086901, doi: 10.1088/0034-4885/78/8/086901 Lei Ba, J., Kiros, J. R., & Hinton, G. E. 2016, arXiv e-prints, arXiv:1607.06450, doi: 10.48550/arXiv.1607.06450

    Pith/arXiv arXiv doi:10.48550/arxiv.1607.06450 2015

  16. [16]

    2025, arXiv e-prints, arXiv:2506.16607, doi: 10.48550/arXiv.2506.16607

    Li, X. 2025, arXiv e-prints, arXiv:2506.16607, doi: 10.48550/arXiv.2506.16607

    doi:10.48550/arxiv.2506.16607 2025

  17. [17]

    2018, MNRAS, 481, 4445, doi: 10.1093/mnras/sty2548

    Li, X., Katayama, N., Oguri, M., & More, S. 2018, MNRAS, 481, 4445, doi: 10.1093/mnras/sty2548

    doi:10.1093/mnras/sty2548 2018

  18. [18]

    2022, MNRAS, 511, 4850, doi: 10.1093/mnras/stac342

    Li, X., Li, Y., & Massey, R. 2022, MNRAS, 511, 4850, doi: 10.1093/mnras/stac342

    doi:10.1093/mnras/stac342 2022

  19. [19]

    2023, MNRAS, 521, 4904, doi: 10.1093/mnras/stad890

    Li, X., & Mandelbaum, R. 2023, MNRAS, 521, 4904, doi: 10.1093/mnras/stad890

    doi:10.1093/mnras/stad890 2023

  20. [20]

    2024, MNRAS, 527, 10388, doi: 10.1093/mnras/stad3895

    Li, X., Mandelbaum, R., Jarvis, M., et al. 2024, MNRAS, 527, 10388, doi: 10.1093/mnras/stad3895

    doi:10.1093/mnras/stad3895 2024

  21. [21]

    2025, MNRAS, 536, 3663, doi: 10.1093/mnras/stae2764

    Li, X., Mandelbaum, R., & The LSST Dark Energy Science Collaboration. 2025, MNRAS, 536, 3663, doi: 10.1093/mnras/stae2764

    doi:10.1093/mnras/stae2764 2025

  22. [22]

    2019, Decoupled Weight Decay Regularization, https://arxiv.org/abs/1711.05101

    Loshchilov, I., & Hutter, F. 2019, Decoupled Weight Decay Regularization, https://arxiv.org/abs/1711.05101

    Pith/arXiv arXiv 2019

  23. [23]

    2018, ARA&A, 56, 393, doi: 10.1146/annurev-astro-081817-051928

    Mandelbaum, R. 2018, ARA&A, 56, 393, doi: 10.1146/annurev-astro-081817-051928

    doi:10.1146/annurev-astro-081817-051928 2018

  24. [24]

    2014, ApJS, 212, 5, doi: 10.1088/0067-0049/212/1/5

    Mandelbaum, R., Rowe, B., Bosch, J., et al. 2014, ApJS, 212, 5, doi: 10.1088/0067-0049/212/1/5

    work page doi:10.1088/0067-0049/212/1/5 2014

  25. [25]

    2007, MNRAS, 376, 13, doi: 10.1111/j.1365-2966.2006.11315.x

    Massey, R., Heymans, C., Bergé, J., et al. 2007, MNRAS, 376, 13, doi: 10.1111/j.1365-2966.2006.11315.x

    arXiv 2007

  26. [26]

    1999, ARA&A, 37, 127, doi: 10.1146/annurev.astro.37.1.127

    Mellier, Y. 1999, ARA&A, 37, 127, doi: 10.1146/annurev.astro.37.1.127

    doi:10.1146/annurev.astro.37.1.127 1999

  27. [27]

    J., et al

    Merz, G., Liu, Y., Burke, C. J., et al. 2023, MNRAS, 526, 1122, doi: 10.1093/mnras/stad2785

    doi:10.1093/mnras/stad2785 2023

  28. [28]

    2025, The Open Journal of Astrophysics, 8, 40, doi: 10.33232/001c.136809

    Merz, G., Liu, X., Schmidt, S., et al. 2025, The Open Journal of Astrophysics, 8, 40, doi: 10.33232/001c.136809

    doi:10.33232/001c.136809 2025

  29. [29]

    2023, CoRR, abs/2311.01500, doi: 10.48550/ARXIV.2311.01500

    Pandya, S., Patel, P., O, F., & Blazek, J. 2023, CoRR, abs/2311.01500, doi: 10.48550/ARXIV.2311.01500

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.01500 2023

  30. [30]

    2022, Proceedings of the AAAI Conference on Artificial Intelligence, 36, 2071, doi: 10.1609/aaai.v36i2.20103

    Paul, S., & Chen, P.-Y. 2022, Proceedings of the AAAI Conference on Artificial Intelligence, 36, 2071, doi: 10.1609/aaai.v36i2.20103

    doi:10.1609/aaai.v36i2.20103 2022

  31. [31]

    2020a, A&A, 643, A158, doi: 10.1051/0004-6361/202038658

    Pujol, A., Bobin, J., Sureau, F., Guinot, A., & Kilbinger, M. 2020a, A&A, 643, A158, doi: 10.1051/0004-6361/202038658

    doi:10.1051/0004-6361/202038658

  32. [32]

    2019, A&A, 621, A2, doi: 10.1051/0004-6361/201833740

    Pujol, A., Kilbinger, M., Sureau, F., & Bobin, J. 2019, A&A, 621, A2, doi: 10.1051/0004-6361/201833740

    doi:10.1051/0004-6361/201833740 2019

  33. [33]

    2020b, A&A, 641, A164, doi: 10.1051/0004-6361/202038657

    Pujol, A., Sureau, F., Bobin, J., et al. 2020b, A&A, 641, A164, doi: 10.1051/0004-6361/202038657

    doi:10.1051/0004-6361/202038657

  34. [34]

    2019, MNRAS, 489, 4847, doi: 10.1093/mnras/stz2374

    Ribli, D., Dobos, L., & Csabai, I. 2019, MNRAS, 489, 4847, doi: 10.1093/mnras/stz2374

    doi:10.1093/mnras/stz2374 2019

  35. [35]

    Á., Zorrilla Matilla, J

    Ribli, D., Pataki, B. Á., Zorrilla Matilla, J. M., et al. 2019, MNRAS, 490, 1843, doi: 10.1093/mnras/stz2610

    doi:10.1093/mnras/stz2610 2019

  36. [36]

    2015, GalSim: The modular galaxy image simulation toolkit, https://arxiv.org/abs/1407.7676

    Rowe, B., Jarvis, M., Mandelbaum, R., et al. 2015, GalSim: The modular galaxy image simulation toolkit, https://arxiv.org/abs/1407.7676

    Pith/arXiv arXiv 2015

  37. [37]

    E., Hinton, G

    Rumelhart, D. E., Hinton, G. E., & Williams, R. J. 1986, Nature, 323, 533, doi: 10.1038/323533a0

    doi:10.1038/323533a0 1986

  38. [38]

    Collaboration, L. D. E. S. 2021, JCAP, 2021, 043, doi: 10.1088/1475-7516/2021/07/043

    doi:10.1088/1475-7516/2021/07/043 2021

  39. [39]

    Scaife, A. M. M., & Porter, F. 2021, MNRAS, 503, 2369, doi: 10.1093/mnras/stab530

    doi:10.1093/mnras/stab530 2021

  40. [40]

    S., Becker, M

    Sheldon, E. S., Becker, M. R., Jarvis, M., & Armstrong, R. 2023, The Open Journal of Astrophysics, 6, doi: 10.21105/astro.2303.03947

    Pith/arXiv arXiv 2023

  41. [41]

    S., Becker, M

    Sheldon, E. S., Becker, M. R., MacCrann, N., & Jarvis, M. 2020, ApJ, 902, 138, doi: 10.3847/1538-4357/abb595

    doi:10.3847/1538-4357/abb595 2020

  42. [42]

    S., & Huff, E

    Sheldon, E. S., & Huff, E. M. 2017, ApJ, 841, 24, doi: 10.3847/1538-4357/aa704b

    doi:10.3847/1538-4357/aa704b 2017

  43. [43]

    2015, arXiv e-prints, arXiv:1503.03757, doi: 10.48550/arXiv.1503.03757

    Spergel, D., Gehrels, N., Baltay, C., et al. 2015, arXiv e-prints, arXiv:1503.03757, doi: 10.48550/arXiv.1503.03757

    Pith/arXiv arXiv doi:10.48550/arxiv.1503.03757 2015

  44. [44]

    M., Ofek, E

    Springer, O. M., Ofek, E. O., Weiss, Y., & Merten, J. 2020, MNRAS, 491, 5301, doi: 10.1093/mnras/stz2991 D4CNN×AnaCal17 Sánchez, J., Walter, C. W., Awan, H., et al. 2020, Monthly Notices of the Royal Astronomical Society, 497, 210–228, doi: 10.1093/mnras/staa1957

    doi:10.1093/mnras/stz2991 2020

  45. [45]

    2019, A&A, 621, A36, doi: 10.1051/0004-6361/201833775 The LSST Dark Energy Science Collaboration, Mandelbaum, R., Eifler, T., et al

    Tewes, M., Kuntzer, T., Nakajima, R., et al. 2019, A&A, 621, A36, doi: 10.1051/0004-6361/201833775 The LSST Dark Energy Science Collaboration, Mandelbaum, R., Eifler, T., et al. 2018, arXiv e-prints, arXiv:1809.01669, doi: 10.48550/arXiv.1809.01669

    Pith/arXiv arXiv doi:10.48550/arxiv.1809.01669 2019

  46. [46]

    A., Jarvis, M., et al

    Yamamoto, M., Troxel, M. A., Jarvis, M., et al. 2023, MNRAS, 519, 4241, doi: 10.1093/mnras/stac2644

    doi:10.1093/mnras/stac2644 2023

  47. [47]

    S., & Becker, M

    Zhang, Z., Sheldon, E. S., & Becker, M. R. 2023, The Open Journal of Astrophysics, 6, 16, doi: 10.21105/astro.2206.07683

    Pith/arXiv arXiv 2023

  48. [48]

    2024, A&A, 683, A209, doi: 10.1051/0004-6361/202345903

    Zhang, Z., Shan, H., Li, N., et al. 2024, A&A, 683, A209, doi: 10.1051/0004-6361/202345903

    doi:10.1051/0004-6361/202345903 2024

  49. [49]

    2018, MNRAS, 481, 1149, doi: 10.1093/mnras/sty2219

    Zuntz, J., Sheldon, E., Samuroff, S., et al. 2018, MNRAS, 481, 1149, doi: 10.1093/mnras/sty2219

    work page doi:10.1093/mnras/sty2219 2018