Pith. sign in

REVIEW 3 major objections 5 minor 62 references

A single neural network can recover weak-lensing shear from full images, absorbing detection, deblending, and calibration into one trained step.

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 · grok-4.5

2026-07-14 14:56 UTC pith:OXXY3QPJ

load-bearing objection Solid, carefully scoped demo that amortized NPE can absorb blending, variable PSFs, stars and detector junk into well-calibrated constant-shear posteriors when the simulator is trusted; the fidelity limit is already priced in by the authors. the 3 major comments →

arxiv 2607.09867 v1 pith:OXXY3QPJ submitted 2026-07-10 astro-ph.IM astro-ph.COcs.CVstat.AP

Neural Posterior Estimation for Inferring Weak Lensing Shear

Tim White , Dingrui Tao , Camille Avestruz , Jeffrey Regier , the LSST Dark Energy Science Collaboration This is my paper
classification astro-ph.IM astro-ph.COcs.CVstat.AP
keywords weak gravitational lensingshear estimationneural posterior estimationsimulation-based inferenceconvolutional neural networksastronomy image processingBayesian statisticsobservational cosmology
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.

Weak-lensing surveys measure the tiny distortions that intervening mass imprints on distant galaxy shapes. The usual pipeline detects galaxies, measures ellipticities, then applies an elaborate calibration to undo noise, blending, and selection biases. This paper shows that the entire chain can be replaced by neural posterior estimation: a residual network is trained on simulated multiband images and directly outputs a probability distribution over the two shear components. Because every anticipated observational effect is present in the training images, the network absorbs those effects without needing explicit masks or analytic corrections. On constant-shear simulations that progressively add blending, spatially varying point-spread functions, stars, cosmic rays and bad columns, the resulting posteriors are accurate and well calibrated. The method therefore becomes practical precisely as simulators become faithful enough to capture the systematics that will dominate Stage-IV surveys.

Core claim

When a residual network is trained by neural posterior estimation on simulated constant-shear images, the resulting mean-field Gaussian posteriors for both shear components remain accurate and well calibrated even after blended galaxies, spatially varying PSFs, stars (including saturated bleed trails), cosmic rays and bad CCD columns are introduced, provided the training and test images are drawn from the same simulator.

What carries the argument

Neural posterior estimation with a mean-field Gaussian variational family: a deep residual network maps a 3 imes2048 imes2048 multiband image to four numbers (means and variances of the two shear components) that approximate the true posterior after all nuisance variables have been marginalized.

Load-bearing premise

Every observational effect that will matter on real survey data must already be present and correctly modeled in the training simulator; otherwise the network will face distribution shift it was never trained to handle.

What would settle it

Train the same network on the paper’s highest-fidelity setting, then evaluate it on images that introduce a new, unsimulated effect (for example correlated coadd noise or a spatially varying shear field) and check whether the empirical coverage of the reported credible intervals collapses below the nominal levels.

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

If this is right

  • Shear estimation cost becomes a single network forward pass once training is finished, amortizing inference across billions of galaxies.
  • Uncertainty quantification is obtained for free: the network returns full posterior distributions rather than calibrated point estimates.
  • As higher-fidelity simulators become available, the same architecture can be retrained to absorb additional systematics without redesigning analytic corrections.
  • The approach naturally extends to the field-level inference of spatially varying shear and convergence maps once those fields can be simulated.

Where Pith is reading between the lines

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

  • The negative multiplicative bias observed at high shear amplitudes is likely a prior-volume effect of the Gaussian training prior; a uniform prior may remove it without architectural change.
  • Because the network never receives an explicit PSF model, any residual PSF knowledge it uses must be learned from the stars and galaxies themselves, suggesting that the same weights could serve double duty as an implicit PSF estimator.
  • If the method survives a controlled distribution-shift test, it would supply a practical route to joint inference of shear and photometric redshifts from the same image pixels.

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 proposes neural posterior estimation (NPE) as an amortized alternative to multistage weak-lensing shear pipelines. A residual network is trained to map simulated 3 imes2048 imes2048 riz coadds to a mean-field Gaussian variational posterior over the two constant shear components, thereby folding detection, deblending, measurement and calibration into a single forward pass. Five simulation settings of increasing complexity (fixed then spatially varying PSF, stars, cosmic rays/bad columns, reduced galaxy density) are generated with descwl-shear-sims; ten independent train/val/test splits are used per setting. Coverage diagnostics, RMSE, Pearson r and multiplicative/additive bias are reported, with AnaCal as a point-estimate baseline on the first two settings. The central claim is that NPE yields accurate, well-calibrated shear posteriors when training and evaluation images are drawn from the same simulator that already includes the listed observational effects.

Significance. If the scoped claim holds, NPE offers a practical route to shear inference that amortizes the cost of handling blends, variable PSFs, stars and detector artifacts once a high-fidelity simulator exists. The experimental design is unusually careful for an ML-for-astronomy methods paper: ten independent replicates, explicit coverage plots that track nominal levels, standard shear metrics with standard deviations, and a direct AnaCal comparison where both methods can be run. Public code and a modular simulator fork further strengthen reproducibility. The work is a useful controlled sensitivity study that complements the authors’ related field-level NPE efforts and will be of interest to the Stage-IV shear community as simulation fidelity improves.

major comments (3)
  1. §5 and Table 2: multiplicative biases for NPE are systematically negative (m ≈ −0.01 to −0.03) and grow in magnitude once stars and artifacts are introduced, while AnaCal’s m is closer to zero on the two settings where it is run. The text attributes this to the Gaussian prior (σ = 0.015) undersampling high-amplitude shear, but no quantitative test (e.g., uniform prior, reweighting, or restricted high-|γ| evaluation) is provided. Because multiplicative bias is the primary figure of merit for Stage-IV shear requirements, the claim of “accurate” posteriors needs either a demonstrated mitigation or a clearer statement that the residual m is an accepted limitation of the present prior/architecture.
  2. §4.2 and Table 2: AnaCal is applied only to settings 1–2 because of difficulties replicating its star/artifact masking. The paper therefore never shows that NPE remains competitive once both methods face the same stellar contamination and detector defects. A short additional experiment—either a simplified AnaCal masking pipeline or a public Metadetection run on the same test sets—would make the comparative claim load-bearing rather than provisional.
  3. §6.1 correctly flags the simulator simplifications (independent Gaussian coadd noise, constant shear, zero convergence, single-redshift galaxies). The abstract and conclusion, however, still present NPE as “viable … as simulation fidelity improves.” The manuscript would be stronger if it quantified, even approximately, how sensitive the present networks are to one of these violations (e.g., a small correlated-noise or multi-redshift test set) rather than leaving all robustness questions to future work.
minor comments (5)
  1. §3.1, Eq. (3): the mean-field Gaussian family is asserted to be “sufficient,” yet no diagnostic of residual γ1–γ2 dependence is shown. A brief scatter of posterior means or a note that the true prior is already independent would close the point.
  2. Appendix B / Table B1: AnaCal timing for setting 2 is dominated by per-image GalSim PSF construction. A one-sentence caveat that this overhead is implementation-specific (and can be reduced by cell-wise PSF approximation) would prevent over-reading the speed comparison.
  3. Figures 2–7: the left-hand 90 % credible-interval panels use green/orange for cover/non-cover; a color-blind-safe palette or hatching would improve accessibility.
  4. §2.1: the CatSim magnitude cut (i < 27) and the resulting raw density of 240 arcmin−2 are stated, but the effective number density after blending is never quoted; a single sentence citing Sanchez et al. (2021) numbers would help readers compare with other shear papers.
  5. Typographical: “W eak” in the title line of the draft header; “therizbands” missing spaces in Table 1 caption; “descwl-shear-simspackage” missing space in several places.

Circularity Check

0 steps flagged

No significant circularity: NPE is trained and scored against ground-truth shear injected by the same simulator; claims do not reduce to inputs by construction.

full rationale

The paper's central results (accurate, well-calibrated NPE posteriors for constant shear under blended galaxies, variable PSFs, stars, and artifacts) are obtained by training a residual network to minimize the expected KL from the true posterior (Eq. 4) on simulated images whose shear is known by construction, then evaluating posterior means, RMSE, Pearson r, multiplicative bias, and empirical coverage on independent held-out draws from the identical generative process (Sections 4–5, Table 2, Figs. 2–7). This is ordinary simulation-based inference validation, not a circular derivation: the network does not recover a parameter that was fitted to the same quantity, nor is any reported metric definitionally equal to an input. The mean-field Gaussian variational family and ResNet architecture are design choices, not uniqueness theorems. Self-citations to related NPE work (Patel et al. 2025 on source detection; White et al. 2026+ on tomographic maps) appear only as complementary context in Section 6.2 and are not load-bearing premises for the accuracy claims. AnaCal supplies an independent external baseline on settings 1–2. The paper itself scopes the claim to the no-distribution-shift regime and flags the remaining simulator simplifications (Section 6.1), so the derivation chain is self-contained against the reported benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central empirical claim rests on the fidelity of the descwl-shear-sims generative model, on the mean-field Gaussian variational family, and on a set of architectural and preprocessing choices. No new physical entities are postulated; free parameters are the usual neural-network weights and a handful of simulation and training hyperparameters.

free parameters (5)
  • Neural-network weights η
    Trained by minimizing the expected KL divergence (Eq. 4) on 3 000 images per split; the entire mapping from image to posterior is encoded in these weights.
  • Shear prior σ = 0.015
    Gaussian prior width chosen to match expected LSST patch amplitudes (Section 2.1); affects the distribution of training labels and may contribute to the observed negative multiplicative bias at high |γ|.
  • Pixel-value clamp at 99th percentile
    Ad-hoc preprocessing to limit dynamic range of saturated stars (Section 4.1).
  • Median imputation for missing pixels
    Ad-hoc handling of cosmic-ray and bad-column pixels (Section 4.1).
  • ResNet channel depths and residual-block strides
    Architectural hyperparameters listed in Appendix A; chosen by the authors rather than derived.
axioms (4)
  • domain assumption Training and evaluation images are i.i.d. draws from the same implicit generative model p(x|γ,z_nuisance) defined by descwl-shear-sims.
    Stated throughout Sections 2 and 6.1; the paper never claims robustness to distribution shift.
  • ad hoc to paper Mean-field independent Gaussians are a sufficiently expressive variational family for the marginal posteriors of the two constant shear components.
    Section 3.1; justified by the claim that the optimal mean-field factors recover the true marginals when the family is rich enough.
  • domain assumption Pixel noise in the coadd coordinate system is independent and Gaussian; all galaxies share a single redshift and a single constant shear; convergence is zero.
    Section 2 and 6.1; limitations of the current simulator.
  • standard math The expected KL objective (Eq. 4) is minimized by stochastic gradient descent on finite simulated batches.
    Standard NPE / amortized VI theory.

pith-pipeline@v1.1.0-grok45 · 20918 in / 2884 out tokens · 21739 ms · 2026-07-14T14:56:10.411352+00:00 · methodology

0 comments
read the original abstract

The prevailing approach to inferring weak gravitational lensing shear from images involves detecting galaxies, estimating their ellipticities, and calibrating these estimates to correct for image noise, selection bias, and model misspecification. Characterizing the statistical model and assumptions underlying this pipeline is challenging, which makes it difficult to propagate uncertainty through its various stages. As an alternative, we propose to infer shear using neural posterior estimation (NPE), a type of simulation-based inference. We train a deep neural network to map a simulated multiband image to a variational distribution over the underlying shear field, thereby folding galaxy detection, deblending, measurement, and calibration into a single implicit inference step. Once trained, the network accounts for all features present in the simulated images, including potential sources of bias. In experiments on simulated constant-shear images with increasingly complex observational effects, NPE produces accurate and well-calibrated posterior approximations for both shear components in the presence of blended galaxies, spatially varying point spread functions, stars, and detector artifacts. These results demonstrate that NPE can be a viable shear estimation method in settings where all anticipated features and artifacts can be simulated, a requirement that will become increasingly feasible as simulation fidelity improves in the coming decades.

Figures

Figures reproduced from arXiv: 2607.09867 by Camille Avestruz, Dingrui Tao, Jeffrey Regier, The LSST Dark Energy Science Collaboration, Tim White.

Figure 1
Figure 1. Figure 1: 2048×2048-pixel images simulated with the descwl-shear-sims package. Here, we display only the r band of each image. The top-left panel displays an image from setting 2, which has 240 galaxies per square arcminute, a spatially varying PSF, no stars, and no image artifacts (Section 2.2). The top-right panel displays an image from setting 3, which adds stars (Section 2.3). The bottom-left panel displays an i… view at source ↗
Figure 2
Figure 2. Figure 2: NPE calibration plots for setting 1. Left: True γ1 and γ2 versus 90% credible intervals inferred by NPE for a randomly selected 1,000-image test set. Green intervals cover the true shear (dotted line), while orange intervals do not. Right: Nominal versus empirical coverage of credible intervals for γ1 and γ2, computed by pooling across the ten test sets [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: NPE calibration plots for setting 2. Left: True γ1 and γ2 versus 90% credible intervals inferred by NPE for a randomly selected 1,000-image test set. Green intervals cover the true shear (dotted line), while orange intervals do not. Right: Nominal versus empirical coverage of credible intervals for γ1 and γ2, computed by pooling across the ten test sets [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: AnaCal scatterplots for settings 1 and 2. True γ1 and γ2 versus AnaCal point estimates for a randomly selected 1,000-image test set from setting 1 (left) and setting 2 (right). 5.4. Setting 4 Based on all three metrics reported in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: NPE calibration plots for setting 3. Left: True γ1 and γ2 versus 90% credible intervals inferred by NPE for a randomly selected 1,000-image test set. Green intervals cover the true shear (dotted line), while orange intervals do not. Right: Nominal versus empirical coverage of credible intervals for γ1 and γ2, computed by pooling across the ten test sets [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: NPE calibration plots for setting 4. Left: True γ1 and γ2 versus 90% credible intervals inferred by NPE for a randomly selected 1,000-image test set. Green intervals cover the true shear (dotted line), while orange intervals do not. Right: Nominal versus empirical coverage of credible intervals for γ1 and γ2, computed by pooling across the ten test sets [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: NPE calibration plots for setting 5. Left: True γ1 and γ2 versus 90% credible intervals inferred by NPE for a randomly selected 1,000-image test set. Green intervals cover the true shear (dotted line), while orange intervals do not. Right: Nominal versus empirical coverage of credible intervals for γ1 and γ2, computed by pooling across the ten test sets. information to the network about the underlying shea… 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

62 extracted references · 11 canonical work pages · 3 internal anchors

  1. [1]

    2021a, The Astrophysical Journal Supplement Series, 253, 31, doi: 10.3847/1538-4365/abd62c

    Abolfathi, B., Alonso, D., Armstrong, R., et al. 2021a, The Astrophysical Journal Supplement Series, 253, 31, doi: 10.3847/1538-4365/abd62c

    doi:10.3847/1538-4365/abd62c

  2. [2]

    2021b, arXiv:2101.04855, doi: 10.48550/arXiv.2101.04855

    Abolfathi, B., Armstrong, R., Awan, H., et al. 2021b, arXiv:2101.04855, doi: 10.48550/arXiv.2101.04855

    Pith/arXiv arXiv doi:10.48550/arxiv.2101.04855

  3. [3]

    2019, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    Akeson, R., Armus, L., Bachelet, E., et al. 2019, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    Pith/arXiv arXiv doi:10.48550/arxiv.1902.05569 2019

  4. [4]

    Forward Amortized Inference for Likelihood-Free Variational Marginalization

    Ambrogioni, L., G¨ u¸ cl¨ u, U., Berezutskaya, J., et al. 2019, in The 22nd International Conference on Artificial Intelligence and Statistics, PMLR, 777–786, doi: 10.48550/arXiv.1805.11542

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1805.11542 2019

  5. [5]

    2024, arXiv:2407.01771, doi: 10.48550/arXiv.2407.01771

    Armstrong, R., Sheldon, E., Huff, E., et al. 2024, arXiv:2407.01771, doi: 10.48550/arXiv.2407.01771

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.01771 2024

  6. [6]

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

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

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

  7. [7]

    M., & Armstrong, R

    Bernstein, G. M., & Armstrong, R. 2014, Monthly Notices of the Royal Astronomical Society, 438, 1880, doi: 10.1093/mnras/stt2326

    doi:10.1093/mnras/stt2326 2014

  8. [8]

    M., Armstrong, R., Krawiec, C., & March, M

    Bernstein, G. M., Armstrong, R., Krawiec, C., & March, M. C. 2016, Monthly Notices of the Royal Astronomical Society, 459, 4467, doi: 10.1093/mnras/stw879

    doi:10.1093/mnras/stw879 2016

  9. [9]

    Cannon, P., Ward, D., & Schmon, S. M. 2022, arXiv:2209.01845, doi: 10.48550/arXiv.2209.01845

    Pith/arXiv arXiv doi:10.48550/arxiv.2209.01845 2022

  10. [10]

    Chen, T., Bansal, V., & Scott, J. G. 2025, arXiv:2507.17030, doi: 10.48550/arXiv.2507.17030

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2507.17030 2025

  11. [11]

    2024, Astronomy & Astrophysics, 691, A319, doi: 10.1051/0004-6361/202450617

    Congedo, G., Miller, L., Taylor, A., et al. 2024, Astronomy & Astrophysics, 691, A319, doi: 10.1051/0004-6361/202450617

    doi:10.1051/0004-6361/202450617 2024

  12. [12]

    J., Angeli, G

    Connolly, A. J., Angeli, G. Z., Chandrasekharan, S., et al. 2014, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 9150, Modeling, Systems Engineering, and Project Management for Astronomy VI, doi: 10.1117/12.2054953

    doi:10.1117/12.2054953 2014

  13. [13]

    R., Gelman, A., & Rubin, D

    Cook, S. R., Gelman, A., & Rubin, D. B. 2006, Journal of Computational and Graphical Statistics, 15, 675, doi: 10.1198/106186006X136976

    doi:10.1198/106186006x136976 2006

  14. [14]

    2017, Gravitational Lensing (Cambridge University Press), doi: 10.1017/9781316424254

    Dodelson, S. 2017, Gravitational Lensing (Cambridge University Press), doi: 10.1017/9781316424254

    work page doi:10.1017/9781316424254 2017

  15. [15]

    2019, arXiv:1904.01174, doi: 10.48550/arXiv.1904.01174

    Dore, O., Hirata, C., Wang, Y., et al. 2019, arXiv:1904.01174, doi: 10.48550/arXiv.1904.01174

    Pith/arXiv arXiv doi:10.48550/arxiv.1904.01174 2019

  16. [16]

    2026, The Astronomical Journal, 171, 112, doi: 10.3847/1538-3881/ae30df Fenech Conti, I., Herbonnet, R., Hoekstra, H., et al

    Duan, Y., Li, X., Avestruz, C., Regier, J., & LSST Dark Energy Science Collaboration. 2026, The Astronomical Journal, 171, 112, doi: 10.3847/1538-3881/ae30df Fenech Conti, I., Herbonnet, R., Hoekstra, H., et al. 2017, Monthly Notices of the Royal Astronomical Society, 467, 1627, doi: 10.1093/mnras/stx200

    work page doi:10.3847/1538-3881/ae30df 2026

  17. [17]

    T., Kelly, R., Drovandi, C., & Warne, D

    Frazier, D. T., Kelly, R., Drovandi, C., & Warne, D. J. 2024, arXiv:2411.12068, doi: 10.48550/arXiv.2411.12068

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.12068 2024

  18. [18]

    2016, Monthly Notices of the Royal Astronomical Society, 457, 3522, doi: 10.1093/mnras/stw174

    Gurvich, A., & Mandelbaum, R. 2016, Monthly Notices of the Royal Astronomical Society, 457, 3522, doi: 10.1093/mnras/stw174

    work page doi:10.1093/mnras/stw174 2016

  19. [19]

    2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770–778, doi: 10.48550/arXiv.1512.03385

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770–778, doi: 10.48550/arXiv.1512.03385

    Pith/arXiv arXiv doi:10.48550/arxiv.1512.03385 2016

  20. [20]

    2022, Transactions on Machine Learning Research, doi: 10.48550/arXiv.2110.06581

    Hermans, J., Delaunoy, A., Rozet, F., et al. 2022, Transactions on Machine Learning Research, doi: 10.48550/arXiv.2110.06581

    Pith/arXiv arXiv doi:10.48550/arxiv.2110.06581 2022

  21. [21]

    2012, Monthly Notices of the Royal Astronomical Society, 421, 381, doi: 10.1111/j.1365-2966.2011.20312.x 12White et al

    Heymans, C., Rowe, B., Hoekstra, H., et al. 2012, Monthly Notices of the Royal Astronomical Society, 421, 381, doi: 10.1111/j.1365-2966.2011.20312.x 12White et al

    doi:10.1111/j.1365-2966.2011.20312.x 2012

  22. [22]

    2006, Monthly Notices of the Royal Astronomical Society, 368, 1323, doi: 10.1111/j.1365-2966.2006.10198.x

    Heymans, C., Van Waerbeke, L., Bacon, D., et al. 2006, Monthly Notices of the Royal Astronomical Society, 368, 1323, doi: 10.1111/j.1365-2966.2006.10198.x

    doi:10.1111/j.1365-2966.2006.10198.x 2006

  23. [23]

    M., Hirata, C

    Huff, E. M., Hirata, C. M., Mandelbaum, R., et al. 2014, Monthly Notices of the Royal Astronomical Society, 440, 1296, doi: 10.1093/mnras/stu144

    work page doi:10.1093/mnras/stu144 2014

  24. [24]

    2006, Monthly Notices of the Royal Astronomical Society, 366, 101, doi: 10.1111/j.1365-2966.2005.09782.x Ivezi´ c,ˇZ., Kahn, S

    Huterer, D., Takada, M., Bernstein, G., & Jain, B. 2006, Monthly Notices of the Royal Astronomical Society, 366, 101, doi: 10.1111/j.1365-2966.2005.09782.x Ivezi´ c,ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, The Astrophysical Journal, 873, 111, doi: 10.3847/1538-4357/ab042c

    doi:10.1111/j.1365-2966.2005.09782.x 2006

  25. [25]

    2006, Journal of Cosmology and Astroparticle Physics, 2006, 001, doi: 10.1088/1475-7516/2006/02/001

    Jain, B., Jarvis, M., & Bernstein, G. 2006, Journal of Cosmology and Astroparticle Physics, 2006, 001, doi: 10.1088/1475-7516/2006/02/001

    work page doi:10.1088/1475-7516/2006/02/001 2006

  26. [26]

    2016, Monthly Notices of the Royal Astronomical Society, 460, 2245, doi: 10.1093/mnras/stw990

    Jarvis, M., Sheldon, E., Zuntz, J., et al. 2016, Monthly Notices of the Royal Astronomical Society, 460, 2245, doi: 10.1093/mnras/stw990

    doi:10.1093/mnras/stw990 2016

  27. [27]

    P., Warne, D

    Kelly, R. P., Warne, D. J., Frazier, D. T., et al. 2025, arXiv:2503.12315, doi: 10.48550/arXiv.2503.12315

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.12315 2025

  28. [28]

    2015, Reports on Progress in Physics, 78, 086901, doi: 10.1088/0034-4885/78/8/086901

    Kilbinger, M. 2015, Reports on Progress in Physics, 78, 086901, doi: 10.1088/0034-4885/78/8/086901

    doi:10.1088/0034-4885/78/8/086901 2015

  29. [29]

    2011, arXiv:1110.3193, doi: 10.48550/arXiv.1110.3193

    Laureijs, R., Amiaux, J., Arduini, S., et al. 2011, arXiv:1110.3193, doi: 10.48550/arXiv.1110.3193

    Pith/arXiv arXiv doi:10.48550/arxiv.1110.3193 2011

  30. [30]

    2023, Monthly Notices of the Royal Astronomical Society, 521, 4904, doi: 10.1093/mnras/stad890

    Li, X., & Mandelbaum, R. 2023, Monthly Notices of the Royal Astronomical Society, 521, 4904, doi: 10.1093/mnras/stad890

    doi:10.1093/mnras/stad890 2023

  31. [31]

    2024, Monthly Notices of the Royal Astronomical Society, 527, 10388, doi: 10.1093/mnras/stad3895

    Li, X., Mandelbaum, R., Jarvis, M., et al. 2024, Monthly Notices of the Royal Astronomical Society, 527, 10388, doi: 10.1093/mnras/stad3895

    doi:10.1093/mnras/stad3895 2024

  32. [32]

    2025, Monthly Notices of the Royal Astronomical Society, 536, 3663, doi: 10.1093/mnras/stae2764

    Li, X., Mandelbaum, R., & LSST Dark Energy Science Collaboration. 2025, Monthly Notices of the Royal Astronomical Society, 536, 3663, doi: 10.1093/mnras/stae2764

    doi:10.1093/mnras/stae2764 2025

  33. [33]

    I., Starck, J.-L., & Kilbinger, M

    Liaudat, T. I., Starck, J.-L., & Kilbinger, M. 2023, Frontiers in Astronomy and Space Sciences, 10, 1158213, doi: 10.3389/fspas.2023.1158213

    doi:10.3389/fspas.2023.1158213 2023

  34. [34]

    2015, Journal of Instrumentation, 10, C05017, doi: 10.1088/1748-0221/10/05/C05017

    Mandelbaum, R. 2015, Journal of Instrumentation, 10, C05017, doi: 10.1088/1748-0221/10/05/C05017

    doi:10.1088/1748-0221/10/05/c05017 2015

  35. [35]

    2018, Annual Review of Astronomy and Astrophysics, 56, 393, doi: 10.1146/annurev-astro-081817-051928

    Mandelbaum, R. 2018, Annual Review of Astronomy and Astrophysics, 56, 393, doi: 10.1146/annurev-astro-081817-051928

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

  36. [36]

    H., et al

    Mandelbaum, R., Jarvis, M., Lupton, R. H., et al. 2023, The Open Journal of Astrophysics, 6, doi: 10.21105/astro.2209.09253

    Pith/arXiv arXiv doi:10.21105/astro.2209.09253 2023

  37. [37]

    M., Seljak, U., et al

    Mandelbaum, R., Hirata, C. M., Seljak, U., et al. 2005, Monthly Notices of the Royal Astronomical Society, 361, 1287, doi: 10.1111/j.1365-2966.2005.09282.x

    doi:10.1111/j.1365-2966.2005.09282.x 2005

  38. [38]

    2015, MNRAS, 450, 2963, doi: 10.1093/mnras/stv781

    Mandelbaum, R., Rowe, B., Armstrong, R., et al. 2015, MNRAS, 450, 2963, doi: 10.1093/mnras/stv781

    doi:10.1093/mnras/stv781 2015

  39. [39]

    2018, arXiv:1809.01669, doi: 10.48550/arXiv.1809.01669

    Mandelbaum, R., Eifler, T., Hloˇ zek, R., et al. 2018, arXiv:1809.01669, doi: 10.48550/arXiv.1809.01669

    Pith/arXiv arXiv doi:10.48550/arxiv.1809.01669 2018

  40. [40]

    2013, Monthly Notices of the Royal Astronomical Society, 429, 661, doi: 10.1093/mnras/sts371

    Massey, R., Hoekstra, H., Kitching, T., et al. 2013, Monthly Notices of the Royal Astronomical Society, 429, 661, doi: 10.1093/mnras/sts371

    doi:10.1093/mnras/sts371 2013

  41. [41]

    2021, Nature Reviews Physics, 3, 712, doi: 10.1038/s42254-021-00353-y

    Gruen, D. 2021, Nature Reviews Physics, 3, 712, doi: 10.1038/s42254-021-00353-y

    doi:10.1038/s42254-021-00353-y 2021

  42. [42]

    2012, Monthly Notices of the Royal Astronomical Society, 424, 2757, doi: 10.1111/j.1365-2966.2012.21381.x

    Melchior, P., & Viola, M. 2012, Monthly Notices of the Royal Astronomical Society, 424, 2757, doi: 10.1111/j.1365-2966.2012.21381.x

    doi:10.1111/j.1365-2966.2012.21381.x 2012

  43. [43]

    Differentiable Forward Modeling for Efficient and Accurate Shear Inference

    Mendoza, I., Guinot, A., Becker, M. R., et al. 2026, arXiv:2604.22048, doi: 10.48550/arXiv.2604.22048

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.22048 2026

  44. [44]

    2025, Monthly Notices of the Royal Astronomical Society, 544, 3799, doi: 10.1093/mnras/staf1833

    OpenUniverse, LSST Dark Energy Science Collaboration, Roman HLIS Project Infrastructure, et al. 2025, Monthly Notices of the Royal Astronomical Society, 544, 3799, doi: 10.1093/mnras/staf1833

    doi:10.1093/mnras/staf1833 2025

  45. [45]

    2025, The Astronomical Journal, 170, 155, doi: 10.3847/1538-3881/adef32

    Patel, A., Zhang, T., Avestruz, C., Regier, J., & LSST Dark Energy Science Collaboration. 2025, The Astronomical Journal, 170, 155, doi: 10.3847/1538-3881/adef32

    work page doi:10.3847/1538-3881/adef32 2025

  46. [46]

    2026, in Encyclopedia of Astrophysics (First Edition), 1st edn., ed

    Prat, J., & Bacon, D. 2026, in Encyclopedia of Astrophysics (First Edition), 1st edn., ed. I. Mandel (Oxford: Elsevier), 508–537, doi: 10.1016/B978-0-443-21439-4.00071-7

    doi:10.1016/b978-0-443-21439-4.00071-7 2026

  47. [47]

    P., & Burchatd, P

    Sanchez, J., Mendoza, I., Kirkby, D. P., & Burchatd, P. R. 2021, Journal of Cosmology and Astroparticle Physics, doi: 10.1088/1475-7516/2021/07/043

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

  48. [48]

    Schmitt, M., B¨ urkner, P.-C., K¨ othe, U., & Radev, S. T. 2024, in Pattern Recognition, ed. U. K¨ othe & C. Rother (Cham: Springer Nature Switzerland), 541–557, doi: 10.1007/978-3-031-54605-1 35

    work page doi:10.1007/978-3-031-54605-1 2024

  49. [49]

    D., Hogg, D

    Schneider, M. D., Hogg, D. W., Marshall, P. J., et al. 2015, The Astrophysical Journal, 807, 87, doi: 10.1088/0004-637X/807/1/87

    doi:10.1088/0004-637x/807/1/87 2015

  50. [50]

    S., Becker, M

    Sheldon, E. S., Becker, M. R., Jarvis, M., Armstrong, R., & LSST Dark Energy Science Collaboration. 2023, arXiv:2303.03947, doi: 10.48550/arXiv.2303.03947

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.03947 2023

  51. [51]

    S., Becker, M

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

    doi:10.3847/1538-4357/abb595 2020

  52. [52]

    S., & Huff, E

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

    doi:10.3847/1538-4357/aa704b 2017

  53. [53]

    2015, arXiv:1503.03757, doi: 10.48550/arXiv.1503.03757

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

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

  54. [54]

    J., Courteau, S., Cuillandre, J.-C., et al

    Stone, C. J., Courteau, S., Cuillandre, J.-C., et al. 2023, Monthly Notices of the Royal Astronomical Society, 525, 6377, doi: 10.1093/mnras/stad2477

    doi:10.1093/mnras/stad2477 2023

  55. [55]

    2018, arXiv:1804.06788, doi: 10.48550/arXiv.1804.06788 NPE for Inferring Weak Lensing Shear13

    Gelman, A. 2018, arXiv:1804.06788, doi: 10.48550/arXiv.1804.06788 NPE for Inferring Weak Lensing Shear13

    Pith/arXiv arXiv doi:10.48550/arxiv.1804.06788 2018

  56. [56]

    2026+, In internal review with the Dark Energy Science Collaboration

    White, T., Chandrashekaran, S., Avestruz, C., & Regier, J. 2026+, In internal review with the Dark Energy Science Collaboration

    2026

  57. [57]

    2018, in Proceedings of the European Conference on Computer Vision (ECCV), 3–19, doi: 10.48550/arXiv.1803.08494

    Wu, Y., & He, K. 2018, in Proceedings of the European Conference on Computer Vision (ECCV), 3–19, doi: 10.48550/arXiv.1803.08494

    Pith/arXiv arXiv doi:10.48550/arxiv.1803.08494 2018

  58. [58]

    2025, Monthly Notices of the Royal Astronomical Society, 543, 4156, doi: 10.1093/mnras/staf1661

    Yamamoto, M., Becker, M., Sheldon, E., et al. 2025, Monthly Notices of the Royal Astronomical Society, 543, 4156, doi: 10.1093/mnras/staf1661

    doi:10.1093/mnras/staf1661 2025

  59. [59]

    2019, IEEE Transactions on Pattern Analysis and Machine Intelligence, 41, 2008, doi: 10.1109/TPAMI.2018.2889774

    Zhang, C., B¨ utepage, J., Kjellstr¨ om, H., & Mandt, S. 2019, IEEE Transactions on Pattern Analysis and Machine Intelligence, 41, 2008, doi: 10.1109/TPAMI.2018.2889774

    doi:10.1109/tpami.2018.2889774 2019

  60. [60]

    2025, Astronomy & Astrophysics, 702, A166, doi: 10.1051/0004-6361/202556133

    Zhang, S., Li, S.-S., & Hoekstra, H. 2025, Astronomy & Astrophysics, 702, A166, doi: 10.1051/0004-6361/202556133

    work page doi:10.1051/0004-6361/202556133 2025

  61. [61]

    2022, Monthly Notices of the Royal Astronomical Society, 510, 1978, doi: 10.1093/mnras/stab3584

    Zhang, T., Mandelbaum, R., & LSST Dark Energy Science Collaboration. 2022, Monthly Notices of the Royal Astronomical Society, 510, 1978, doi: 10.1093/mnras/stab3584

    doi:10.1093/mnras/stab3584 2022

  62. [62]

    2023, Monthly Notices of the Royal Astronomical Society, 520, 2328, doi: 10.1093/mnras/stac3350 14White et al

    Zhang, T., Almoubayyed, H., Mandelbaum, R., et al. 2023, Monthly Notices of the Royal Astronomical Society, 520, 2328, doi: 10.1093/mnras/stac3350 14White et al. APPENDIX A.NEURAL NETWORK ARCHITECTURE Figure A1 illustrates how our neural networkf η maps an imagexto variational parametersϕ. The network’s inputs and outputs (outlined in black) are connected...

    doi:10.1093/mnras/stac3350 2023