🔭 astrophysics

Differentiable Forward Modeling for Efficient and Accurate Shear Inference

This paper presents a modern, GPU-accelerated implementation of a Bayesian shear inference framework that uses differentiable forward modeling to achieve the high-precision, low-bias galaxy shape measurements required for upcoming Stage-IV dark energy surveys like LSST.

Original authors: Ismael Mendoza, Axel Guinot, Matthew R. Becker, Camille Avestruz, Jean-Eric Campagne, Natalia Porqueres, Michael Schneider, Eleni Tsaprazi

Published 2026-04-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ismael Mendoza, Axel Guinot, Matthew R. Becker, Camille Avestruz, Jean-Eric Campagne, Natalia Porqueres, Michael Schneider, Eleni Tsaprazi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Cosmic Detective: Solving the Mystery of Dark Energy

Imagine you are looking at a massive, crowded ballroom through a very thick, wavy piece of glass. In this ballroom, thousands of dancers (which represent galaxies) are moving around. You want to know if the floor itself is slightly tilted or if there is a subtle, invisible force pushing everyone in one direction.

The problem? The wavy glass (which represents gravity and dark energy) distorts the dancers' shapes. A dancer who is actually standing perfectly straight might look like they are leaning to the left. If you can figure out exactly how much every single dancer is being "tilted" by the glass, you can work backward to understand the invisible force causing the distortion.

In science, this "tilt" is called cosmic shear, and the invisible force is dark energy.

The Problem: The "Blurry Photo" Dilemma

Current astronomers face a massive headache. To measure dark energy, they need to look at billions of galaxies. But galaxies are tiny, faint, and often look like blurry smudges.

Existing methods are like trying to guess a dancer's posture by looking at a low-quality, grainy photo. If you guess wrong, your entire understanding of the universe's expansion is off. Most current tools try to "fix" the blurriness by adding even more artificial noise to the photos to see how they react—which is like intentionally making a photo even blurrier just to see if you can still recognize a face. It’s slow, it’s clunky, and it’s prone to errors.

The Solution: The "Smart Digital Twin"

The researchers in this paper have built a new, high-tech way to solve this. Instead of just guessing, they use something called Differentiable Forward Modeling.

Think of it like this: Instead of just looking at a blurry photo and guessing, the researchers have created a perfect digital simulation of a dancer.

They take a "digital dancer" in a computer.
They apply the "wavy glass" effect to that digital dancer.
They compare the digital version to the real, blurry photo.
Because their computer model is "differentiable" (which is a fancy way of saying the computer can "feel" its way toward the truth), the computer can instantly adjust the digital dancer's pose—tilting them, stretching them, or brightening them—until the digital version matches the real photo perfectly.

The Secret Sauce: The Supercharged Engine

To do this for billions of galaxies, you can't use a standard laptop; you'd be waiting until the end of the universe. The researchers used two "superpowers" to make this fast:

GPUs (Graphics Processing Units): These are the same chips used for high-end video games. They are incredibly good at doing thousands of tiny calculations at once.
Gradient-Based Sampling: Instead of a person randomly guessing "Is the dancer leaning 5 degrees? How about 6?" the computer uses math to "slide" down a hill toward the correct answer, finding the truth much faster.

The Result: A Gold Standard

The team tested their "Digital Twin" method on a massive simulated dataset. They found that their method was incredibly accurate—meeting the strict requirements for the next generation of giant telescopes (like the LSST).

Even when they didn't know the "rules" of the dancers (like how bright they usually are or how much they usually wobble), the system was smart enough to learn those rules at the same time it was measuring the tilt.

Why does this matter?

We are currently in a race to understand why the universe is expanding faster and faster. If we get the math wrong, we get the history of the universe wrong. This paper provides a faster, smarter, and more accurate "lens" that will allow astronomers to peer through the cosmic blur and finally see the invisible hand of dark energy clearly.

Technical Summary: Differentiable Forward Modeling for Efficient and Accurate Shear Inference

1. Problem Statement

The primary objective of upcoming Stage-IV dark energy surveys (e.g., LSST, Euclid, Roman) is to constrain cosmological parameters with sub-percent precision. This requires measuring cosmic shear—the subtle distortion of galaxy shapes due to weak gravitational lensing—with extreme accuracy.

Current state-of-the-art methods, such as Metacalibration and Anacal, rely on "self-calibration" (applying artificial shears to images to measure response). However, these methods face significant challenges in the Stage-IV era, including:

Systematic Biases: Blending (overlapping galaxies) and chromatic effects (wavelength-dependent PSF) can introduce significant multiplicative biases.
Signal-to-Noise Reduction: Self-calibration often requires injecting additional noise into images, which degrades measurement precision.
Computational Scalability: Existing Bayesian forward-modeling approaches (like LensMC or JIF) are often too slow to process the billions of galaxies expected in modern wide-field surveys.

2. Methodology

The authors propose a modern implementation of a Hierarchical Bayesian Shear Inference framework. The core innovation lies in combining statistical rigor with high-performance computational techniques.

A. Statistical Framework:
The method uses a two-phase Bayesian approach:

Image Sampling Phase: For each individual galaxy image, the algorithm performs MCMC sampling to infer an "interim posterior" of galaxy properties (flux, size, centroid, and shape/ellipticity) using a differentiable parametric model.
Shear Sampling Phase: A second MCMC chain aggregates these individual samples across the entire dataset to infer the global cosmic shear ( $g$ ) and the hyperparameters ( $\alpha$ ) that describe the intrinsic distribution of galaxy properties (e.g., the shape noise $\sigma_e$ ).

B. Computational Innovations:
To overcome the "speed bottleneck" of traditional Bayesian methods, the authors leverage:

Differentiable Forward Modeling: Using JAX-GalSim (a JAX-based reimplementation of GalSim), the galaxy models are fully differentiable.
Gradient-Based Samplers: They utilize the No-U-Turn Sampler (NUTS) via the BlackJAX library, which uses gradients to navigate high-dimensional parameter spaces more efficiently than random-walk MCMC.
GPU Acceleration: The entire pipeline is vectorized and executed on NVIDIA A100 GPUs, allowing thousands of galaxy chains to be processed in parallel.

3. Key Contributions

New Formalism: A derivation of the shear inference likelihood specifically optimized for cases where the PSF and sky background are known.
High-Efficiency Implementation: A pipeline capable of producing 300 effective samples of galaxy properties in just 0.45 seconds per galaxy on a single A100 GPU.
Joint Inference Capability: The ability to simultaneously infer the cosmic shear and the underlying distribution of galaxy properties (hyperparameters), which helps mitigate selection biases.
Open Source Tools: The release of JAX-GalSim, a differentiable version of the industry-standard GalSim.

4. Results

The algorithm was tested on a simulated dataset of 320,000 isolated exponential galaxies with LSST-like noise levels. The authors used the "shape noise cancellation" trick to rigorously measure multiplicative bias ( $m$ ).

Multiplicative Bias ( $m$ ):
- When the intrinsic galaxy property distribution is known: $|m| < 0.9 \times 10^{-3}$ ( $3\sigma$ ).
- When the distribution is inferred alongside shear (all-free): $|m| < 1.3 \times 10^{-3}$ ( $3\sigma$ ).
Compliance: Both results are well within the LSST requirement of $|m| < 2 \times 10^{-3}$ .
Convergence: Individual galaxy chains showed excellent convergence ( $\bar{R} < 1.01$ ).
Hyperparameter Accuracy: The method successfully recovered the true hyperparameters of the galaxy distributions (flux, size, and shape noise) within their $2\sigma$ posterior contours.

5. Significance and Future Work

This work demonstrates that Bayesian forward modeling is no longer computationally unfeasible for large-scale surveys. By moving away from self-calibration and toward differentiable modeling, the community gains a framework that is inherently more capable of handling complex systematics like blending and chromaticity.

Future Directions:

Handling Selection/Detection Bias: Generalizing the model to account for the fact that certain galaxy shapes are easier to detect than others.
Mitigating Blending: Extending the model to jointly fit multiple overlapping sources.
Model Bias Reduction: Moving beyond simple parametric models (like Sersic profiles) toward more flexible, machine-learning-derived galaxy models to avoid errors caused by "model mismatch."