🔭 astrophysics

Replicating weak-lensing summary-statistic covariances with normalizing flows

This paper demonstrates that while normalizing flow models can accurately reproduce the mean and variance of weak-lensing summary statistics, they tend to underestimate off-diagonal covariance elements by up to 25% unless mitigation strategies like data augmentation and training with noisy fields are applied to improve recovery to within 5%.

Original authors: Joaquin Armijo, Leander Thiele, Jia Liu

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joaquin Armijo, Leander Thiele, Jia Liu

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

Imagine the universe as a giant, invisible ocean of matter, mostly made of a mysterious substance called "dark matter" that we can't see directly. To map this hidden ocean, astronomers play a game of cosmic billiards. They look at distant galaxies, which act like tiny marbles, and watch how their shapes get slightly squished or stretched as their light travels through the gravity of the dark matter ocean. This stretching is called "weak gravitational lensing." By measuring the squishiness of millions of these galaxy marbles, scientists can reconstruct a 3D map of the universe's hidden structure.

However, the universe is messy. The dark matter isn't spread out smoothly like a calm lake; it clumps together in a chaotic, lumpy way, especially on small scales. To understand this chaos, scientists need to run massive computer simulations to create "mock universes" and compare them to real telescope data. But running these simulations is like trying to bake a billion cakes at once—it takes so much computing power that it becomes a major bottleneck. This is where a new type of artificial intelligence called "Normalizing Flows" enters the story. Think of these flows as a magical, reversible machine that learns the recipe for a universe. Once trained, it can instantly bake new, fake universes that look just like the real ones, saving scientists from waiting years for a computer to finish its work. But before we trust this magic machine to help us solve the mysteries of the universe, we have to make sure it's actually baking the right kind of cake, not just a pretty-looking fake.

This paper puts that magical machine to the test. The researchers, Joaquin Armijo, Leander Thiele, and Jia Liu, trained a Normalizing Flow model on a set of high-quality cosmological simulations called SLICS. Their goal was to see if the AI could perfectly recreate the "summary statistics" of the universe—basically, the key numbers that describe the shape and structure of the dark matter maps. They focused on three specific ways to measure the universe: the "angular power spectrum" (which measures how much clumping happens at different sizes), the "probability density function" (which counts how often certain clumpiness levels appear), and "Minkowski functionals" (which describe the shape and connectivity of the cosmic web, like how many holes or loops exist).

The team found that the AI is surprisingly good at getting the basics right. When they asked the model to generate fake maps, the average size of the clumps and the average amount of variation (the "mean" and "variance") were reproduced with incredible accuracy, often within 1% of the real simulations. It was as if the AI learned the average recipe perfectly. However, the story gets trickier when looking at the "covariance," which is a fancy word for how different parts of the map are connected to each other. Imagine if you knew the average height of a forest, but you didn't know if the tall trees tended to grow in clusters or scattered randomly. The paper found that while the AI got the average height right, it struggled to learn the clustering patterns. In many cases, the AI underestimated the connections between different parts of the map by up to 25%. It was like the AI knew the forest had trees, but it thought the trees were more spread out and less connected than they actually were.

The researchers didn't just stop at finding the problem; they tried to fix it. They tested three different strategies to help the AI understand the universe better. First, they made the AI's "brain" (the neural network) bigger and more complex. Second, they used "data augmentation," which is like taking a few pictures of the universe, cutting them into smaller pieces, and rotating them to create more training examples. Third, and most importantly, they added "noise" to the training data, simulating the kind of static or fuzziness that real telescopes see. They discovered that adding this noise was the secret sauce. When the AI was trained on noisy maps, its ability to recreate the complex connections (the covariance) improved dramatically, getting the error down to about 3% for the power spectrum.

The paper concludes that while Normalizing Flows are a powerful tool for generating fake universes, they aren't perfect out of the box. If scientists use these AI-generated maps without fixing the underestimation of connections, they might end up with results that look too confident, leading them to believe they know the universe's secrets better than they actually do. The study suggests that by using techniques like adding noise and expanding the training data, we can make these AI models much more reliable. This is a crucial step forward, ensuring that when we use these fast, AI-generated maps to study the dark energy and the fate of the universe, we aren't just looking at a pretty illusion, but a faithful representation of the cosmic reality.

Technical Summary: Replicating Weak-Lensing Summary-Statistic Covariances with Normalizing Flows

Problem Statement
Weak gravitational lensing is a primary probe for constraining cosmological parameters (e.g., $\Omega_m$ , $\sigma_8$ , dark energy properties) by analyzing the statistical distribution of matter. However, analyzing weak-lensing data requires accurate covariance matrices for summary statistics (such as the angular power spectrum, probability density functions, and Minkowski functionals) to quantify uncertainties. Generating the large ensembles of cosmological simulations necessary to estimate these covariances empirically is computationally expensive, creating a bottleneck for analyses of current and future surveys (e.g., Euclid, LSST).

While machine learning generative models (VAEs, GANs, Normalizing Flows) offer a path to generating synthetic convergence ( $\kappa$ ) maps at negligible computational cost, their validity depends on whether they accurately reproduce not just the mean and variance of summary statistics, but also the covariance structure (specifically off-diagonal elements). A model that matches means and variances but fails to capture the joint distribution of statistic bins may produce overconfident cosmological constraints. This paper investigates the fidelity of Normalizing Flow (NF) models in replicating these covariance structures.

Methodology
The authors utilize the SLICS suite of high-resolution N-body simulations ( $N_{sim} = 954$ light-cones) to train and test NF models. The study evaluates two distinct training scenarios:

Data Vector Training: Training NFs directly on 1D summary statistic vectors (Angular Power Spectrum $C_\ell$ , $\kappa$ -PDF, and Minkowski Functionals) derived from the simulations.
Map-Level Training: Training NFs on the full 2D convergence maps ( $\kappa$ -maps) to learn the pixel distribution directly.

Two NF architectures are employed:

Neural Spline Flows (NSF): Used for the 1D data vectors. These employ rational-quadratic splines parameterized by neural networks to model complex distributions.
Multiscale Normalizing Flows (MS): Used for 2D $\kappa$ -maps. This architecture utilizes a hierarchy of squeeze-and-split operations (inspired by RealNVP) to process images at multiple spatial resolutions, transforming the image into a Gaussian latent space and vice versa.

The authors test the models across different resolutions ( $64\times64$ , $128\times128$ , $256\times256$ pixels) and investigate mitigation strategies for covariance underestimation, including:

Increasing network depth (adding multiscale levels).
Data augmentation (cropping and rotating $10\times10$ deg $^2$ maps into $5\times5$ deg $^2$ patches).
Training with added Gaussian pixel noise (simulating LSST-like survey conditions).

Key Results

Mean and Variance Recovery: In both training scenarios, NF models successfully reproduce the mean and standard deviation (variance) of the summary statistics ( $C_\ell$ , PDF, MFs) with percent-level accuracy.
Covariance Underestimation: A significant limitation is identified in the recovery of off-diagonal covariance elements.
- When trained on data vectors, the covariance is underestimated by approximately 5% for $C_\ell$ and PDF, but up to 13% for Minkowski Functionals.
- When trained on full maps using a simple multiscale architecture, the underestimation is severe: variance is underestimated by ~25% and off-diagonal covariance terms by up to 75% for larger maps ( $256\times256$ ).
Mitigation Strategies:
- Network Size: Increasing the depth of the multiscale architecture reduces the discrepancy but does not fully resolve it (reducing variance/covariance errors to ~18%).
- Data Augmentation: Cropping and rotating maps to increase the effective training set size improves covariance recovery, reducing off-diagonal errors to ~16%.
- Noise Injection: Adding Gaussian noise to the training maps is the most effective strategy. It reduces the off-diagonal covariance underestimation for the power spectrum to ~3%. However, for non-Gaussian statistics (PDF and MFs), the improvement is limited (errors remain at 13–30%) because the injected noise is Gaussian and does not capture the non-Gaussian structure driving those statistics.
Training Set Size: The study finds that approximately 200 simulations are sufficient to recover the covariance structure with ~92% accuracy. Accuracy plateaus near 95% as the training set approaches the full size of the simulation suite.

Significance and Claims
The paper positions the accurate recovery of covariance matrices as a rigorous diagnostic for generative model fidelity, arguing that matching means and variances is insufficient for cosmological inference. The authors claim that while NFs are powerful tools for generating high-fidelity synthetic maps and emulating summary statistics, they suffer from a systematic tendency to underestimate off-diagonal covariances, particularly when applied to high-dimensional data with limited training samples.

The study concludes that without specific mitigation strategies—specifically data augmentation and training with noisy fields—NF-generated maps may lead to overconfident cosmological posteriors due to underestimated uncertainties. The authors present these findings not as a failure of the method, but as a necessary diagnostic step before deploying generative models as fast emulators or components of simulation-based inference pipelines for stage-IV weak lensing surveys. They suggest that future applications, particularly those involving diffusion models or complex non-Gaussian statistics, must account for these scaling limitations and noise characteristics.

Technical Summary: Replicating Weak-Lensing Summary-Statistic Covariances with Normalizing Flows

More like this