Differentiable Forward Modeling for Efficient and Accurate Shear Inference
Ismael Mendoza ⋆, Axel Guinot, Matthew R. Becker, Camille Avestruz, Jean-Eric Campagne, Natalia Porqueres, Michael Schneider, Eleni Tsaprazi, The LSST Dark Energy Science Collaboration · 2026
Read Differentiable Forward Modeling for Efficient and Accurate Shear Inference by Ismael Mendoza ⋆, Axel Guinot, Matthew R. Becker, Camille Avestruz, Jean-Eric Campagne, Natalia Porqueres, Michael Schneider, Eleni Tsaprazi, The LSST Dark Energy Science Collaboration (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-22048.txt.
Introduction
Mendoza and colleagues present a version of the Bayesian shear method of Schneider and colleagues, with known backgrounds. As an application, the authors infer the cosmic shear result from mock images of lone exponential galaxies. In this simple case, absolute scale bias is below 1 times 10 to the power minus 3 at 3 sigma when galaxy property distributions are known. It is below 1.3 times 10 to the power minus 3 at 3 sigma when inferred alongside shear. Their final galaxy-fitting MCMC produces 300 effective samples in under 1 second per galaxy using a single A100 GPU.
Catalog record
| CREATOR | Ismael Mendoza ⋆, Axel Guinot, Matthew R. Becker, Camille Avestruz, Jean-Eric Campagne, Natalia Porqueres, Michael Schneider, Eleni Tsaprazi, The LSST Dark Energy Science Collaboration |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61300 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.22048
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Differentiable Forward Modeling for Efficient and Accurate Shear Inference Ismael Mendoza ⋆, Axel Guinot, Matthew R. Becker, Camille Avestruz, Jean-Eric Campagne, Natalia Porqueres, Michael Schneider, Eleni Tsaprazi, The LSST Dark Energy Science Collaboration Abstract Forthcoming Stage-IV dark energy optical surveys, such as LSST, have the ambitious goal of measuring cosmological parameters at sub-percent precision. Realizing their full scientific potential requires very precise measurement of the cosmic shear signal and control of corresponding systematics. In this work, we present a modern implementation of the Bayesian shear inference framework in Schneider et al. (2014), in the case that the PSF and sky background are known. This framework automatically propagates the pixel-noise measurement error from each galaxy into the final shear estimate, and thus requires no external calibration to handle noise bias. As a first application of this new implementation, we infer the cosmic shear posterior from simulated images consisting of isolated exponential galaxies with LSST-like levels of shape and pixel noise. In this simplified scenario, we estimate the absolute multiplicative bias | m | |m| of our approach to be below 0.9 × 10 − 3 [ 3 σ ] 0.9\times 10^{-3}\,[3\sigma] when the intrinsic distribution of galaxy properties is known, and below 1.3 × 10 − 3 [ 3 σ ] 1.3\times 10^{-3}\,[3\sigma] when these distributions are inferred alongside shear. Both results are within the LSST requirement of | m | < 2 × 10 − 3 |m|<2\times 10^{-3} . Additionally, we make progress towards the algorithm’s computational feasibility in the context of modern wide-field surveys, where billions of galaxies must be processed, by leveraging differentiable forward models of galaxies, gradient-based samplers, and GPUs. Our final galaxy-fitting MCMC produces 300 300 effective samples of galaxy properties in 0.45 0.45 seconds per galaxy using a single A100 GPU. In the future, we seek to generalize our algorithm to handle selection, detection, and model shear biases so it can be applied to real survey data. ⋆ Corresponding author: ][email protected] 1 Introduction The gravitational potential generated by structure in our universe distorts the light from background source galaxies, altering their observed shapes due to weak gravitational lensing (WL) – an effect called cosmic shear . Cosmic shear traces the distribution of dark matter and the nature of dark energy, making WL one of the most powerful cosmological probes ( 40 ; 45 ; 69 ) . Cosmic shear measurements have been successfully used in several Stage-III optical cosmological surveys, including KiDS ( 2 ; 89 ) , DES ( 22 ; 75 ) , and HSC ( 54 ; 20 ) to constrain dark energy cosmological parameters to about a percent level. Modern Stage-IV optical surveys such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) ( 42 ) , Euclid ( 25 ) , and the Roman Space Telescope ( 84 ) have the objective of constraining dark energy to a sub-percent, which could pave the way for beyond \Lambda CDM physics. Precise and accurate measurements of cosmic shear are critical to maximize the scientific return of these Stage-IV surveys. Shear measurement is challenging: cosmic shear is about a 2\% change in a galaxy’s axis ratio on an average cosmological line of sight ( 7 ; 59 ) . LSST, for instance, has the ambitious goal of performing shear measurement with statistical errors below 1 part in 1000 of this 2\% ( 59 ) . Several shear measurement algorithms have been proposed over the years, but only a few have shown the potential to meet this challenging requirement. Currently, the dominant paradigm calibrates the ensemble shear estimator directly on the survey images by simulating the effect of a known shear, so-called self-calibration . This is the motivation for the Metacalibration ( 38 ; 81 ) algorithm, and its extension Metadetection ( 80 ; 34 ; 33 ; 79 ) . Another related approach is Anacal ( 52 ; 53 ) , which analytically calibrates the shear response using auto-differentiation and a suitable basis of FPFS shapelets ( 51 ) . Both Metadetection and Anacal have been shown to meet the LSST shear bias requirements, and Metacalibration and Metadetection have already been successfully applied to real data in DES ( 92 ; 26 ; 90 ) . Despite the successes of self-calibration methods, they will face new challenges in the context of Stage-IV surveys. For example, the coupled effects of blending in both shear and redshift calibration were shown in DES weak lensing analyses to be an important contribution to the mean multiplicative bias ( 56 ) . This effect is likely to worsen with the increased depth of surveys like LSST ( 73 ; 60 ) . Another set of systematics gaining relevancy are chromatic biases ( 61 ; 23 ; 15 ; 44 ; 4 ) . These arise because the Point Spread Function (PSF), a wavelength-dependent quantity, is modeled using stars which typically have different SEDs than galaxies. Currently, there is no Metacalibration extension accounting for these effects; Anacal cannot account for the aforementioned blending effect, but PSF-level corrections were recently implemented for chromatic biases and tested on Roman simulations ( 4 ) . It is likely that these methods will require external calibration from simulations as demonstrated in 56 . Separately, both Metacalibration and Anacal require injecting additional noise to images as part of their procedure, significantly reducing the signal-to-noise ratio of measurements, and increasing systematic uncertainty. However, extensions of these methods that use deep field images are in development and might reduce the impact of this effect ( 91 ; 67 ) . Given some of the current limitations in self-calibration methods, exploring other shear inference methods might still be fruitful. Another set of algorithms for shear measurement use a forward model of galaxy images that includes relevant systematics. In principle, this approach reduces the bias on the shear estimator at the expense of increased model complexity and estimated uncertainty. A recent implementation of this idea is LensMC ( 24 ) , which samples galaxy ellipticities with an optimized MCMC and aggregates them across multiple galaxies to produce a final shear measurement. This algorithm has been tested on Euclid-like simulations and achieved reasonable shear measurement accuracy. One downside of LensMC is that the prescription to aggregate shapes into shear is an ad-hoc weighting and average scheme, and does not utilize the entire set of galaxy properties. Another related method is JIF ( 13 ) which used a physically-informed prior and selection corrections on the likelihood to obtain posterior samples of galaxy properties from images in the LSST DESC DC2 simulation ( 55 ) . This algorithm is statistically rigorous and established good probabilistic calibration for many of the galaxies in this simulation. However, they have not demonstrated that their samples can be used to estimate shear. Both approaches might also be limited by their runtime in a Stage-IV survey: LensMC was estimated to process one galaxy in 5 seconds whereas JIF takes a few minutes. In this work we propose a new shear inference algorithm based on the Bayesian hierarchical framework from 74 . This framework infers shear from MCMC samples of galaxy properties obtained via model fitting based on a