The Shear-to-Cosmology Paradigm I. Hybrid Field-Level and Simulation-Based Framework for Weak Lensing Surveys

Chen Su; Huanyuan Shan; Jiacheng Ding; Ji Yao; Le Zhang

arxiv: 2511.22851 · v2 · pith:7X5MDKRJnew · submitted 2025-11-28 · 🌌 astro-ph.CO

The Shear-to-Cosmology Paradigm I. Hybrid Field-Level and Simulation-Based Framework for Weak Lensing Surveys

Jiacheng Ding , Chen Su , Ji Yao , Le Zhang , Huanyuan Shan This is my paper

Pith reviewed 2026-05-21 18:27 UTC · model grok-4.3

classification 🌌 astro-ph.CO

keywords weak lensingshear fieldscosmological inferencemachine learningsimulation-based inferencefield-level inferencePCA denoisingnon-Gaussian information

0 comments

The pith

Direct shear-field inference with machine learning doubles the cosmological constraining power of weak lensing surveys compared to convergence-based methods.

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

The paper develops a hybrid machine-learning framework that combines field-level inference with simulation-based inference to map observed shear fields directly to cosmological parameters. This bypasses the conventional step of reconstructing convergence maps from shear, which can discard non-Gaussian information. On CSST-like mock catalogs the shear-based pipeline yields roughly twice the figure of merit in cosmological constraints relative to convergence-based inference. Adding a training-free PCA denoising step and ML compression further raises performance by 36.4 percent over standard shear two-point statistics. A reader cares because next-generation surveys will deliver enormous shear datasets, and any method that extracts more information from the same data tightens constraints on dark energy and other parameters without requiring new observations.

Core claim

The central claim is that a hybrid FLI-SBI network trained on shear fields, preceded by blind PCA denoising, extracts richer non-Gaussian information and produces tighter cosmological posteriors than either convergence reconstruction or conventional two-point shear statistics, delivering approximately twice the FoM of convergence-based inference and a 36.4 percent FoM gain over standard shear two-point statistics on CSST-like mocks.

What carries the argument

Hybrid field-level and simulation-based inference network that ingests denoised shear fields to produce compressed features whose posteriors are modeled directly via simulation-based inference.

If this is right

Cosmological parameters are inferred directly from shear fields, removing information loss associated with convergence reconstruction.
Non-Gaussian features in the shear field contribute measurably more constraining power than standard two-point statistics alone.
Blind PCA denoising mitigates shape noise while preserving cosmological signal for downstream inference.
The resulting framework scales to the data volumes expected from Stage-IV weak-lensing surveys.

Where Pith is reading between the lines

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

The same direct-shear pipeline could be retrained on mocks tailored to Euclid or LSST to check whether comparable gains appear for those surveys.
If the performance advantage persists on real data, traditional summary-statistic pipelines might be supplemented or replaced by field-level ML methods for upcoming analyses.
The approach opens a route to joint inference with other lensing or galaxy-clustering probes that also produce shear-like fields.

Load-bearing premise

The CSST-like mock catalogs used for testing accurately reproduce the statistical properties, noise characteristics, non-Gaussian features, and intrinsic alignments present in real weak-lensing observations.

What would settle it

Running the trained pipeline on real weak-lensing survey data from an existing catalog and finding that the reported FoM gains over convergence-based or two-point methods do not appear.

Figures

Figures reproduced from arXiv: 2511.22851 by Chen Su, Huanyuan Shan, Jiacheng Ding, Ji Yao, Le Zhang.

**Figure 1.** Figure 1: Schematic overview of the proposed framework, consisting of: 1. a field-level inference (FLI) network (inside the dashed box) and 2. a simulation-based inference (SBI) module (outside the dashed box). The field-level shear input is a six-dimensional tensor with shape (B, Nγ, Nz, Nblk, H, W) = (B, 2, 16, 4, 64, 64), where B is the batch size, Nγ the shear components, Nz the number of photo-z slices, Nblk th… view at source ↗

**Figure 2.** Figure 2: Preparation of an input sky block for training the FLI network. At a resolution of 0.1 deg/pixel, based on the original 12.8×12.8 deg2 blocks, we generate 6.4×6.4 deg2 inputs via two augmentations: (1) random cropping (boosting spatial diversity by 642 ), (2) random masking with overlapping circular masks to mimic survey masks. The corresponding convergence field can then be reconstructed from the prepa… view at source ↗

**Figure 3.** Figure 3: CSST-like photo-z distribution. The red curve shows the photo-z distribution with photo-z uncertainties σ(z) = σz(1 + z), σz = 0.05. Using this z-dependent kernel, the true-z distribution (marked curve) is obtained by deconvolution. A mock true-z galaxy catalog is generated assuming a galaxy number density of 26 gal/arcmin2 and intrinsic shape noise σe = 0.288. Incorporating photo-z uncertainties, the mo… view at source ↗

**Figure 4.** Figure 4: Illustration of the sky-region partitioning into training, validation, and test sets. For each cosmology, the dataset contains 108 sky blocks, grouped in fours within the same declination region. Groups enclosed by red, blue, and white dashed boxes indicate the training, test, and validation sets, respectively, with no overlapping sky regions. Each block is a 128 × 128 mesh spanning 12.8 × 12.8 deg2 and i… view at source ↗

**Figure 5.** Figure 5: Pipeline of shear measurement in each photo-z bin including 4 steps to enhance the shear signal quality. When PCA processing is applied, the denoised shear field is denoted as γ PCA; otherwise, the resulting field is denoted as γ Avg . information retention and overfitting prevention, expressed as γMsm(θ) = 1 Nsm N Xsm n=1 γ sm(θ). (10) (c) Denoising: Conventional denoising algorithms (like smoothing and… view at source ↗

**Figure 6.** Figure 6: Visualization of shear denoising. The bottom panels show shear fields from the pipeline in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: Evolution of the loss (batch size = 32) for different input data: the shear maps γ PCA (with denoising) and γ Avg (without denoising), the corresponding convergence maps κ PCA and κ Avg via KS algorithm. In neural network training, the validation set is excluded from parameter optimization and serves solely to trigger automatic learning rate reduction based on variations in its loss. Consequently, the e… view at source ↗

**Figure 8.** Figure 8: Comparison of PIT validation results between field-level inference (FLI) and ML-derived simulation-based inference (SBI), performed using four types of input data: (γ PCA, γ Avg , κ PCA, κ Avg). Compared with FLI, the PIT distribution of SBI is closer to a uniform distribution, indicating that the SBI posteriors are better calibrated and more consistent with the true parameters. shear 2PCF (ˆξ+−). The cons… view at source ↗

**Figure 9.** Figure 9: Posteriors inferred via SBI using three types of summary statistics over the sky area of 655.4 deg2 , with true cosmological parameters Ωtrue m = 0.3062 and σ true 8 = 0.7615. The left panel shows the posterior using the 2PCF (ˆξ+−) of the PCA-denoised shear field (γ PCA). The middle panel uses ML-derived features from the convergence field (κ PCA), reconstructed from γ PCA via the KS method. The right pan… view at source ↗

**Figure 10.** Figure 10: Comparison of SBI posteriors obtained from γ Avg and γ PCA over 655.4 deg2 , highlighting the impact of PCA denoising on two types of summary statistics: the 2PCFs (ˆξ+−) and ML-derived features (xML). Based on FoM/deg2 , PCA denoising improves the constraining power by 25.7% when using ML-derived features. features, PCA denoising results in a 25.7% increase in constraining power. This demonstrates that t… view at source ↗

read the original abstract

Precise cosmological inference from next-generation weak lensing surveys requires extracting non-Gaussian information beyond standard two-point statistics. We present a hybrid machine-learning (ML) framework that integrates field-level inference (FLI) with simulation-based inference (SBI) to map observed shear fields directly to cosmological parameters, eliminating the need for convergence reconstruction. The FLI network extracts rich non-Gaussian information from the shear field to produce informative features, which are then used by SBI to model the resulting complex posteriors. To mitigate noise from intrinsic galaxy shapes, we develop a blind, training-free, PCA-based shear denoising method. Tests on CSST-like mock catalogs reveal significant performance gains. The shear-based inference achieves approximately twice the cosmological constraining power in Figure of Merit (FoM) compared to the conventional convergence-based approach. Moreover, the combination of PCA denoising and ML compression can deliver a 36.4% improvement in FoM over standard shear two-point statistics. This work establishes a scalable and robust pathway for cosmological inference, unlocking the full potential of Stage-IV weak-lensing surveys.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

The paper's hybrid FLI-plus-SBI pipeline on raw shear fields with blind PCA denoising reports solid FoM gains on mocks, but those gains rest on how faithfully the CSST-like simulations match real non-Gaussian shear statistics and intrinsic alignments.

read the letter

The one thing to take away is that this work shows a practical way to skip convergence reconstruction and pull cosmological parameters straight from shear maps by combining field-level ML feature extraction with simulation-based inference, plus a simple PCA step to handle shape noise without any training data. On their CSST-like mocks this setup roughly doubles the figure of merit relative to standard convergence methods and adds another 36 percent over plain shear two-point statistics. That numerical improvement is the clearest result they present. The blind, training-free denoising is a sensible addition that avoids introducing extra model assumptions at the noise-cleaning stage. The overall architecture is laid out clearly enough that someone could try to reproduce the pipeline on similar mocks. The central limitation is the exclusive reliance on those mocks for all the reported gains. If the simulations do not fully capture the joint distribution of shear, shape noise, intrinsic alignments, and higher-order non-Gaussian features that will appear in real Stage-IV data, the network could be fitting simulation artifacts rather than cosmology. The abstract gives no details on validation checks, training stability, or how they quantified uncertainty from the ML components, so it is hard to judge how robust the FoM numbers actually are. This paper is mainly for people already working on weak-lensing analysis pipelines for surveys like CSST or Euclid who want concrete ideas for moving past two-point statistics. A reader who needs a new methodological starting point with numerical benchmarks on mocks will get something usable here. It is worth sending to a serious referee because the claims are specific and can be checked directly against the mocks and code they presumably supply, even if more work on real-data realism will be required.

Referee Report

3 major / 2 minor

Summary. The paper introduces a hybrid machine-learning framework that combines field-level inference (FLI) with simulation-based inference (SBI) to map observed weak-lensing shear fields directly to cosmological parameters, avoiding convergence reconstruction. A blind, training-free PCA-based denoising method is developed to mitigate shape noise. On CSST-like mock catalogs, the shear-based approach is reported to achieve roughly twice the Figure of Merit (FoM) of the conventional convergence-based method, while the combination of PCA denoising and ML compression yields a 36.4% FoM improvement over standard shear two-point statistics.

Significance. If validated, the hybrid FLI+SBI pipeline and PCA denoising could meaningfully increase the cosmological information extracted from Stage-IV weak-lensing surveys by capturing non-Gaussian features in the shear field. The approach is presented as scalable and could reduce reliance on convergence maps, which is a practical advantage for future surveys.

major comments (3)

[Abstract / Results] Abstract and Results section: the headline FoM gains (2× versus convergence; 36.4% versus shear 2pt) are demonstrated exclusively on CSST-like mocks, yet the manuscript provides no quantitative details on validation procedures, covariance estimation, training stability of the FLI network, or tests for simulation-specific artifacts. These omissions are load-bearing because any mismatch between mock and real non-Gaussian shear statistics or intrinsic-alignment modeling would directly inflate the reported information gains.
[Method] Method section (FLI network and SBI step): the free parameters listed in the axiom ledger (PCA component count, FLI architecture, training hyperparameters) are tuned on the same class of mocks used for evaluation. Without an explicit cross-validation or held-out simulation suite that varies the underlying cosmology and noise model independently, it is unclear whether the network is learning cosmology or simulation-specific features.
[Results / Comparison] Comparison to convergence-based baseline: the claim that shear-based inference doubles the FoM assumes an otherwise identical analysis pipeline. The manuscript does not specify whether the convergence maps are reconstructed with the same denoising, the same mask, or the same SBI posterior model; any difference in these choices would undermine the direct comparison.

minor comments (2)

[Method] Notation for the PCA denoising procedure is introduced without an explicit equation showing how the principal components are selected or how the reconstruction threshold is chosen; a short equation or pseudocode block would improve reproducibility.
[Abstract] The abstract states “approximately twice” the FoM; the exact numerical values and their uncertainties should be reported in the main text or a table for precision.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their detailed and constructive comments on our manuscript. We address each of the major comments below and outline the revisions we plan to make to improve the clarity and robustness of our results.

read point-by-point responses

Referee: [Abstract / Results] Abstract and Results section: the headline FoM gains (2× versus convergence; 36.4% versus shear 2pt) are demonstrated exclusively on CSST-like mocks, yet the manuscript provides no quantitative details on validation procedures, covariance estimation, training stability of the FLI network, or tests for simulation-specific artifacts. These omissions are load-bearing because any mismatch between mock and real non-Gaussian shear statistics or intrinsic-alignment modeling would directly inflate the reported information gains.

Authors: We agree that additional quantitative details on the validation procedures are necessary to support the reported FoM gains. In the revised manuscript, we will add a new subsection in the Results section detailing the validation procedures, including the use of k-fold cross-validation on the mock catalogs, covariance estimation via jackknife resampling, assessment of training stability through multiple independent training runs with different random seeds, and tests for simulation-specific artifacts by comparing results across different mock generation pipelines. These analyses confirm the robustness of our findings, and we will include the corresponding quantitative metrics. revision: yes
Referee: [Method] Method section (FLI network and SBI step): the free parameters listed in the axiom ledger (PCA component count, FLI architecture, training hyperparameters) are tuned on the same class of mocks used for evaluation. Without an explicit cross-validation or held-out simulation suite that varies the underlying cosmology and noise model independently, it is unclear whether the network is learning cosmology or simulation-specific features.

Authors: We acknowledge the importance of demonstrating that the network learns cosmological features rather than simulation-specific ones. The hyperparameters were selected using a separate validation set drawn from the same mock suite but with different random realizations. To strengthen this, we will incorporate results from a held-out simulation suite in the revised version, where we vary the cosmology and noise properties independently. We will report the performance on this held-out set to show generalization. revision: yes
Referee: [Results / Comparison] Comparison to convergence-based baseline: the claim that shear-based inference doubles the FoM assumes an otherwise identical analysis pipeline. The manuscript does not specify whether the convergence maps are reconstructed with the same denoising, the same mask, or the same SBI posterior model; any difference in these choices would undermine the direct comparison.

Authors: We clarify that the convergence-based baseline uses the same mask as the shear analysis and employs the standard Kaiser-Squires reconstruction without the PCA denoising step, as the denoising method is tailored to the shear field. The SBI posterior modeling is identical for both approaches. We will add explicit details in the revised manuscript, including a table comparing the pipeline components for the shear and convergence cases, to ensure the comparison is transparent and fair. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical gains demonstrated on external mocks without self-referential reduction

full rationale

The paper proposes a hybrid FLI+SBI framework and reports empirical FoM improvements on CSST-like mock catalogs. The PCA denoising is explicitly training-free, and the central performance claims (2x FoM vs. convergence; 36.4% gain over shear 2pt) are presented as simulation results rather than first-principles derivations. No equations reduce to self-definition, no fitted parameters are relabeled as independent predictions, and no load-bearing uniqueness theorems or ansatzes are imported via self-citation. The framework is self-contained against the provided simulation benchmarks, satisfying the default expectation of no circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the fidelity of the mock catalogs and on the effectiveness of the ML components, which involve multiple tuned parameters whose values are not reported.

free parameters (2)

PCA component count and selection
Number of components retained for denoising is chosen to balance noise removal and signal preservation on the mocks.
FLI network architecture and training hyperparameters
Network depth, width, learning rate, and regularization parameters are fitted during development on the training mocks.

axioms (1)

domain assumption Mock catalogs faithfully reproduce the statistical properties of real weak-lensing observations including shape noise and intrinsic alignments.
Performance claims are demonstrated exclusively on these mocks.

pith-pipeline@v0.9.0 · 5733 in / 1467 out tokens · 67349 ms · 2026-05-21T18:27:03.041953+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

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

hybrid machine-learning (ML) framework that integrates field-level inference (FLI) with simulation-based inference (SBI) to map observed shear fields directly to cosmological parameters

What do these tags mean?

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

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages · 1 internal anchor

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...

work page
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...

work page
[3]

Cosmology with cosmic shear observations: a review

thebibliography [1] 20pt to REFERENCES 6pt =0pt \@twocolumntrue 12pt -12pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key o...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/0034-4885/78/8/086901 2017

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...

work page

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...

work page

[3] [3]

Cosmology with cosmic shear observations: a review

thebibliography [1] 20pt to REFERENCES 6pt =0pt \@twocolumntrue 12pt -12pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key o...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/0034-4885/78/8/086901 2017