SHINE: Shear inference environment - Generative forward modelling for radio weak gravitational lensing

arXiv:2611.14301 · astro-ph.IM · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "SHINE: Shear inference environment - Generative forward modelling for radio weak gravitational lensing".

Jocelyn: Here's the gist of the paper. Next-generation radio interferometers will enable weak gravitational lensing studies using radio data, which not only complement optical observations but also probe higher redshifts The proposed method is a cosmic shear inference method based on forward generative modelling, which avoids explicit shape measurements This approach obtains unbiased shear estimates with sub-percent accuracy when applied to parametric galaxy simulations The authors compare their method to state-of-the-art shear estimators, demonstrating a reduction in uncertainty of up to 3 times and no apparent multiplicative bias They also show that applying this parametric model to COSMOS galaxies with realistic morphologies introduces a bias in the shear estimate over 10σ Finally, by incorporating a generative forward model based on an autoencoder architecture trained on real galaxy images, they can mitigate this bias to below the 2σ level.

Paper summary: Vera: So, this paper is called "SHINE: Shear inference environment - Generative forward modelling for radio weak gravitational lensing." It’s about how we can actually figure out cosmic shear using radio data from next-generation telescopes like the SKA. The main idea is that traditional ways of measuring shear don't work well with radio images because they are in Fourier space, which makes things tricky.

Jocelyn: Right, so the core problem they’re tackling here is that we can use weak gravitational lensing from radio observations to look at the universe at higher redshifts than optical surveys allow. But you can’t just measure shear directly on those radio images because they are delocalized, and deconvolution methods like CLEAN introduce errors when you try to get back to the image domain.

Subrahmanyan: And what this paper proposes is a way around that by using a generative forward model instead of trying to measure the galaxy shapes directly. They simulate how galaxy images look, apply the shear transformation, and then use the radio interferometer's response to create simulated visibilities in Fourier space where they define their likelihood function.

Vera: So, what’s the payoff here? They use Monte Carlo sampling algorithms to estimate the posterior distribution of cosmic shear based on galaxy observations that are all assumed to share a common shear. And when they test this on parametric galaxy simulations, they get unbiased shear estimates with sub-percent accuracy.

Jocelyn: That sounds really promising for getting some clean measurements out of these complex radio data sets. But they also compared their method against existing state-of-the-art estimators and found a reduction in uncertainty of up to three times, and no apparent multiplicative bias in their results.

Subrahmanyan: That reduction in uncertainty is significant, especially when you think about the cosmological constraints we're trying to put on structure formation. But they also found something important when they applied this parametric model to COSMOS galaxies with realistic morphologies. They saw a bias in the shear estimate over ten sigma.

Vera: Ten sigma is a big deal, that means it’s not just a small statistical fluctuation; it’s a real systematic shift in what the measurement suggests about the shear of those specific galaxies. But they tackled that by incorporating an autoencoder architecture trained on real galaxy images into their generative forward model.

Jocelyn: And by doing that, they were able to reduce that bias down to below two sigma. So they took a method designed for simulations and made it robust enough for real data observations, which is what we need.

Subrahmanyan: It shows the power of combining the theoretical modeling with generative techniques to handle the complexities of interferometric data and realistic galaxy shapes in a way that yields reliable cosmological constraints.

Vera: So, to wrap up this paper on "SHINE: Shear inference environment - Generative forward modelling for radio weak gravitational lensing," they've shown that we can infer cosmic shear from radio observations without having to measure the exact shape of every single galaxy pixel by pixel. The authors did this by creating a forward generative model that simulates the images and then sampling the shear in Fourier space to get an unbiased estimate, which improves on previous methods substantially when applied to simulations.

Jocelyn: It seems like it really opens up a pathway for using next-generation radio arrays to probe weak lensing at higher redshifts where optical surveys struggle. The authors found that while their method works well on parametric simulations, applying it to real galaxies with actual morphologies still introduces a bias over ten sigma, which they then managed to fix by adding an autoencoder based generative model.

Subrahmanyan: This work suggests that the next major hurdle in radio weak lensing is not just getting the data from the telescope, but developing inference methods that can correctly handle the complex physics of image formation and shear distortion in Fourier space. It gives us a concrete tool for testing cosmological models against these new datasets.

Conclusion: Vera: So, this paper is about something called SHINE, which stands for Shear inference environment - Generative forward modelling for radio weak gravitational lensing.

Jocelyn: It’s really about how to actually measure cosmic shear using radio data from big telescopes like the SKA without having to measure the exact shapes of every single galaxy directly.

Subrahmanyan: The main idea is moving away from explicit shape measurements because that just doesn't work well in Fourier space, where radio observations live.

Vera: Right, so they build a generative model that simulates what those galaxy images look like and then applies the shear transformation to it to see how the radio data should look.

Jocelyn: And they use that model, along with the actual noise from the radio instrument, to predict what we’d see in Fourier space.

Subrahmanyan: They use Monte Carlo sampling algorithms to estimate the final cosmic shear based on those simulated observations that all share a common shear.

Vera: When they tested this on fake galaxy simulations, they got these unbiased shear estimates with less than one percent error. That’s pretty solid for testing their method against existing estimators.

Jocelyn: But then they tried it on real COSMOS galaxies, and the bias in their estimate got bigger—over ten sigma.

Subrahmanyan: That shows the model works well on idealized shapes, but when you introduce realistic galaxy forms, you get a systematic shift that's way too big to ignore.

Vera: So what they did next was add an autoencoder architecture trained on real galaxy images into their generative model.

Jocelyn: They used that autoencoder to learn the messy details of real galaxies so the model could correct those biases down to below two sigma.

Subrahmanyan: That means they’ve built a framework that can handle both the theoretical setup and the complicated reality of observing actual galaxies in radio data.

Vera: It suggests we can use these radio surveys to get better constraints on how structure formed in the universe at higher redshifts than we could before.

Jocelyn: The next step is seeing if this approach scales up to a real survey like SKA, where you’re collecting millions of galaxy measurements over a huge area.

Ezequiel Centofanti, Benjamin Remy, Emma Ayçoberry, François Lanusse, Jean-Luc Starck, Samuel Farrens

AIM, CEA, CNRS, University of Paris-Saclay University of Paris NSF-Simons AI Institute for the Sky Department of Astronomy and Astrophysics University of Chicago Institutes of Computer Science and Astrophysics Foundation for Research and Technology Hellas Flatiron Institute

astro-ph.IM

Submitted: 2026-11-10

Updated: 2026-11-10

Code: https://github.com/b-remy/probabilistic-shear

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: Next-generation radio interferometers will enable weak gravitational lensing studies using radio data…

Key concepts

Forward Generative Modelling
This technique simulates the entire process of observing a galaxy in the radio interferometer. It starts with a known galaxy image, applies the shear transformation mathematically, and then accounts for how that distorted image appears as measurable radio visibilities (data) by including the specific instrumental response of the radio telescope. This allows researchers to infer cosmic shear directly from data without needing to measure shapes manually.
Fourier Domain Observation
Radio interferometers acquire data in the Fourier domain, meaning they measure spatial frequencies rather than direct image pixels. Because galaxy images are delocalized in this space, measuring the physical shape (ellipticity) of a galaxy becomes mathematically complex and non-trivial. The proposed method leverages this domain by defining the likelihood function directly within these Fourier measurements.
Autoencoder Architecture
This is a type of deep learning model trained on real galaxy images to learn their underlying structure. The paper uses an autoencoder to create a generative forward model that can accurately simulate realistic galaxy images. By training this model on actual data, the researchers can generate synthetic observations that are morphologically faithful, which helps in mitigating biases when applying the inference method to real galaxies.

Terminology

Summary

Next-generation radio interferometers will enable weak gravitational lensing studies using radio data, which not only complement optical observations but also probe higher redshifts The proposed method is a cosmic shear inference method based on forward generative modelling, which avoids explicit shape measurements This approach obtains unbiased shear estimates with sub-percent accuracy when applied to parametric galaxy simulations The authors compare their method to state-of-the-art shear estimators, demonstrating a reduction in uncertainty of up to 3 times and no apparent multiplicative bias They also show that applying this parametric model to COSMOS galaxies with realistic morphologies introduces a bias in the shear estimate over 10σ Finally, by incorporating a generative forward model based on an autoencoder architecture trained on real galaxy images, they can mitigate this bias to below the 2σ level. p. 0

Where the paper says it

We then show that applying this parametric model to COSMOS galaxies with realistic morphologies introduces a bias in the shear estimate over 10σ. Finally, we demonstrate that, by incorporating a generative forward model based on an autoencoder architecture trained on real galaxy images, we can mitigate this bias to below the 2σ level.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated shear estimation for radio data can be performed without explicit shape measurements by simulating visibilities of each galaxy individually with a forward model that includes cosmic shear and instrumental response.

  2. The system can utilize two generative models, a parametric Spergel surface brightness profile and an autoencoder based generative model, to handle different morphological complexities in galaxy observations.

  3. The system can employ differentiable sampling methods such as Microcanonical Langevin Monte Carlo (MCLMC) for efficient estimation of the posterior distribution of cosmic shear in high-dimensional parameter spaces.

  4. The AI system can be trained on real galaxy images from surveys like COSMOS to build a deep neural network generative model, allowing it to generate realistic surface brightness profiles and mitigate model bias when applied to complex observed morphologies.

  5. The system can incorporate a normalising flow, such as the Real NVP flow, within the forward model to enable sampling in the latent space of the autoencoder for generating new galaxy images from a base distribution.

  6. The system can perform point estimation of cosmic shear using MAP optimization, which is up to 30 times faster than full posterior estimation while remaining consistent with sampled posterior distributions.

Abstract

Next-generation radio interferometers, such as SKA, will observe the radio sky with unprecedented sensitivity and resolution enabling weak gravitational lensing studies using radio data, which not only complement optical observations but also probe higher redshifts. However, traditional shear estimation methods, based on measuring the ellipticity of observed galaxies, are not directly applicable to radio interferometric data. Interferometric observations are acquired in the Fourier domain, where galaxy images are delocalised, making shape measurements non-trivial. Furthermore, deconvolution methods such as CLEAN can introduce systematic errors when converting observations to the image domain, biasing shear estimates. To address these challenges, we propose a cosmic shear inference method based on forward generative modelling, which avoids explicit shape measurements. Our forward model simulates galaxy images, applies the shear transformation and, by including the instrumental response of the radio interferometer, simulates the visibilities of observed galaxies, with the likelihood explicitly defined in Fourier space. Using Monte Carlo sampling algorithms, we estimate the posterior distribution of the cosmic shear conditioned on a set of galaxy observations that share a common shear. We obtain unbiased shear estimates with sub-percent accuracy when applied to parametric galaxy simulations. We compare our method to state-of-the-art shear estimators, demonstrating a reduction in uncertainty of up to 3 times and no apparent multiplicative bias. We then show that applying this parametric model to COSMOS galaxies with realistic morphologies introduces a bias in the shear estimate over 10σ. Finally, we demonstrate that, by incorporating a generative forward model based on an autoencoder architecture trained on real galaxy images, we can mitigate this bias to below the 2σ level.

Sources

Related papers