Abstract :
Weak gravitational lensing is one of the premier probes of dark matter and dark energy. Upcoming Stage-IV surveys (Euclid, Rubin, Roman) are limited less by how much data they collect than by how well they can turn it into accurate cosmological constraints. Much of the information sits at small, non-linear scales, where the matter field is non-Gaussian and the standard two-point statistics miss a significant part of it.
We investigate whether the choice of mass-mapping method matters for the cosmological inference that follows. We then develop PnPMass, a deep-learning reconstruction trained once for any mask and noise level, with per-pixel uncertainties that carry a coverage guarantee. We then ask how much of the non-Gaussian information a hand-built summary statistic can hold against a neural compressor trained to be information-optimal. Finally, we quantify how far unmodelled baryonic feedback biases higher-order statistics as the survey area grows, and what cutting the contaminated scales away costs in constraining power.
