4 610 universes—exactly that many cosmological models were needed by astrophysicists from Marseille and Barcelona to show that the familiar Gaussian mask no longer works for next-generation data. Future sky surveys like Euclid and Rubin, using gravitational lensing, promise dark matter maps with unprecedented precision. These maps are built from tiny distortions in the shapes of distant galaxies—weak lensing that is sensitive to the entire large-scale structure of the Universe. From them, we can measure the matter density Ωm and the clustering amplitude S8, and through them, probe the nature of dark energy and the accelerated expansion of the cosmos. But to extract these parameters, you need statistical inference. The classic approach—explicit likelihood inference (ELI)—places a “Gaussian mask” over the data: it assumes all uncertainties follow a normal distribution, and the link to cosmology is captured by a simple emulator trained on numerical simulations. The alternative—likelihood-free inference (LFI)—removes the mask: it learns the shape of the distribution directly from examples using Machine Learning, and doesn’t require Gaussian dogmas. Until now, with these two methods giving different results, nobody knew which was right.
To settle the debate, the scientists built an ideal testbed. Inside the GLASS simulation pipeline, they generated 4 610 independent universes with different combinations of Ωm and σ8—all within the ΛCDM model, where Friedmann equations dictate expansion and gravitational lensing describes light bending. Each universe is a full celestial sphere, divided into pixels with a resolution of about 7 arcminutes. Realistic galaxy shape noise was added, matching the final Euclid release (30 galaxies per square arcminute), and a 100 square degree patch—a typical size for early surveys—was cut out. For each universe, they measured two types of statistics: classic shear two-point correlation functions (shear-2PCFs) and a novel convolutional neural network (CNN) estimator trained directly on the maps. Next, the data was compressed down to two numbers: either with the linear MOPED method, optimal for Gaussian fields, or with a nonlinear neural network. Then two inference procedures were run. ELI: trained a Gaussian process emulator on 250 universes, estimated the covariance matrix from 20 400 realizations, and applied a Gaussian likelihood. LFI: trained a neural density estimator on 3 687 universes to learn the form of the likelihood without any assumptions. Both methods then went through the same Bayesian inference with Markov chains—a fair duel on the same turf.
The very first comparison revealed a drama. For ELI with MOPED compression, the Gaussian mask cracked: constraints on Ωm ended up twice as wide as those from LFI, and the posterior distribution shifted away from the truth. The cause was a poorly emulated second compression component that carried most of the matter-density information, along with mild non-Gaussianity. When the scientists removed this component or fed the same statistical model into LFI (simulating the emulator and Gaussian noise), the discrepancy vanished—both methods agreed. With nonlinear neural-network compression, the Gaussian mask shattered completely: the compressed data distribution was no longer normal, and ELI produced errors 2.2 times larger in S8 and 1.3 times larger in Ωm, while LFI maintained or even improved accuracy. The clash became even more striking when comparing correlation functions with CNN. In an ideal Gaussian world, two-point statistics should capture all the information, yet ELI, constrained by its mask, suggested that CNN was two times better at measuring Ωm but 1.6 times worse at S8. LFI, freed from the mask, upheld theoretical expectations: CNN and compressed correlations gave consistent results with only a 30% difference. The TARP calibration tests, normally applied only to LFI, emerged as the main diagnostic: they clearly showed where ELI loses calibration and where it doesn’t. Traditional Gaussianity checks (Kolmogorov-Smirnov test, χ²) turned out to be too weak to spot dangerous deviations.
The takeaway: in the era of high-precision surveys, the Gaussian mask can no longer be the default standard. This work proves that likelihood-free inference (LFI) is not just a fancy toy but an essential tool for honest data analysis. At the same time, it vindicates ELI by showing that with careful emulation and no non-Gaussianity, it performs just as well. The main practical message: any future cosmological analysis must be accompanied by rigorous calibration tests of posterior distributions—TARP and PIT histograms—otherwise we risk mistaking methodological artifacts for new physical effects. Moreover, the study underscores that the choice of data compression method (linear or nonlinear) is critical for ELI but almost irrelevant for LFI—giving researchers the freedom to design optimal summary statistics.
The next steps are clear. First, replace Gaussian simulations with full N-body calculations, where nonlinear structure growth generates genuine non-Gaussianity—there the gap between ELI and LFI will become even more dramatic. Second, move from a single redshift to tomographic slices, which will exponentially increase the data dimensionality and the burden on both methods. Third, incorporate systematic effects: baryonic physics, intrinsic galaxy alignments, source clustering. All this will demand even more flexible neural estimators and new ways to verify their reliability. At stake is our ability to distinguish Einstein’s dark energy from modified gravity.
The results will directly impact the data analysis for Euclid and the Rubin Observatory’s Legacy Survey of Space and Time. Consortia are already planning to integrate LFI into main pipelines, and calibration procedures will become mandatory practice.
Developing more accurate deep-learning-based emulators and building a reference simulation suite for testing inference methods. In parallel, adapting TARP diagnostics for non-Gaussian likelihoods within the ELI framework.
This work directly addresses fundamental questions: what is dark matter made of, what is the nature of dark energy, and is general relativity correct on cosmological scales? Without proper statistical inference, we can neither confirm ΛCDM nor discover new particles or fields.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study compares two approaches to statistical inference in weak lensing: classical explicit inference (ELI) with a Gaussian likelihood and likelihood-free inference (LFI) with neural density estimation. Using simulated Euclid data, it shows that ELI yields biased estimates of Ωm and σ8, while LFI remains accurate even when Gaussian assumptions are violated.
Where its strength is
The authors built a controlled testbed of 4610 universes, allowing direct measurement of each method's systematic errors. The strength lies in the honest comparison on realistic simulations and the demonstration that abandoning Gaussian simplifications is critical for future surveys.
Where it could go next
Worth trying
Why this may matter
The shift to likelihood-free inference could become the standard for next-generation cosmological surveys, eliminating hidden systematics that currently limit the precision of dark energy and dark matter measurements.
Nearby works in our archive
The closest work in the archive sits at 0.66 by meaning — the usual level for our corpus.
as written by the authors arXiv:2607.17942v1
Simulation-based inference (SBI) has become a major tool for extracting cosmological information from weak-lensing (WL) surveys, particularly from non-Gaussian observables. We compare its two main paradigms: explicit likelihood inference (ELI), based on a Gaussian likelihood built from an emulator and covariance matrix, and likelihood-free inference (LFI), which learns the likelihood directly from simulations using neural density estimators. Using Gaussian random field mocks representative of the non-tomographic final Euclid data release, we analyse shear two-point correlation functions (shear-2PCFs), compressed with linear or non-linear methods, together with a fundamentally different map-level convolutional neural network (CNN) statistic, focusing on $$\Omega_{\rm m}$$ and $$S_8$$. We deploy posterior calibration diagnostics developed for LFI, including the test of accuracy with random points (TARP), showing that ELI becomes strongly miscalibrated under emulation inaccuracies or likelihood non-Gaussianity, whereas LFI remains well calibrated. These effects drive substantial disagreement between ELI and LFI, which largely vanishes once addressed. We further show that the compression scheme can significantly degrade ELI while leaving LFI largely unaffected. Although shear-2PCFs should capture all the information in Gaussian fields, finite compression and non-Gaussian likelihoods cause ELI constraints to differ by up to a factor of two from those inferred with the CNN, while the discrepancy drops to $$\approx 30\%$$ for LFI, underscoring the robustness of the deep-learning probe. Overall, our results indicate that in our simple setup, which neglects systematic biases, LFI provides a more robust and better-calibrated framework, while highlighting accurate non-Gaussian likelihood modelling and posterior calibration diagnostics as essential for future ELI analyses.
To accumulate the 20 400 independent realizations used for covariance estimation, a real telescope would have to observe the same patch of sky for 56 000 years — five times longer than agriculture has existed.