Powered by Indico
v3.2.8
Weak gravitational lensing provides a powerful probe of the large-scale structure of the Universe, but traditional two-point statistics capture only part of the cosmological information encoded in the non-Gaussian features of the matter distribution. In this talk, we explore the use of high-order statistics to extract this additional information. We will first discuss recent advances in deep-learning-based mass map reconstruction, focusing on how neural networks can efficiently invert shear fields into high-fidelity convergence maps while mitigating noise and masking effects. To quantify the reliability of these reconstructions, we will present a framework based on conformal quantile prediction, which allows the calibration of pixel-wise uncertainties on the predicted mass maps in a statistically rigorous and distribution-free way. Building upon these uncertainty-aware maps, we introduce a new class of wavelet-based statistics, where the ℓ₁ norm of wavelet coefficients serves as a compact and robust descriptor of non-Gaussianity. This approach combines the interpretability and sparsity of wavelet representations with the sensitivity of higher-order moments, offering an efficient summary statistic that complements standard power spectrum analyses. We will present results from simulated weak lensing surveys demonstrating how these methods enhance cosmological parameter constraints and discuss prospects for their application to upcoming wide-field surveys such as Euclid and Rubin/LSST.