Speaker
Description
Weak gravitational lensing is a primary probe for Stage IV cosmology, but fully exploiting upcoming surveys such as Euclid requires going beyond the standard power spectrum. Higher-order statistics (HOS) can capture the non-Gaussian information in tomographic data, yet two major challenges remain: their sensitivity to complex astrophysical systematics, and their reliance on computationally expensive simulations without independent analytical validation.
In this talk, we present a unified framework to optimize tomographic weak lensing analyses with HOS for Euclid. Using simulation-based inference with both analytical and neural network summaries, we show that while avoiding baryonic biases requires conservative scale cuts, HOS still yield significantly tighter cosmological constraints than standard two-point statistics on these “safe” scales. Moreover, the interpretable starlet $\ell_1$-norm achieves near-optimal performance, providing a simple and powerful alternative to black-box neural methods.
To push beyond these conservative cuts, we revisit tomographic nulling via the Bernardeau–Nishimichi–Taruya (BNT) transform, which isolates redshift contributions to enable physically motivated scale cuts. While BNT is widely thought to degrade constraining power, we show that this limitation is not intrinsic. Using optimal neural summaries, we demonstrate that the transform is in fact lossless, and the apparent degradation arises from an incomplete treatment of cross-correlations, providing a prescription for its use in realistic analyses.
Finally, we introduce a complementary, theory-driven approach based on Large Deviation Theory to model HOS directly, delivering a crucial analytical cross-check and reducing reliance on simulations. Together, these results establish a robust and near-optimal strategy for extracting cosmological information from Euclid data, with preliminary applications in ongoing preparations for Data Release 1.
| Other topic / keywords: | Weak gravitational lensing; cosmological inference |
|---|