Speaker
Description
CMB observations show that early density fluctuations were nearly Gaussian. The gravitational evolution of matter formed large structures, and nonlinear evolution introduced non-Gaussianity, making it challenging to analyze. In this project, we employ the scattering transform, a powerful statistical tool that shares ideas with convolutional neural networks (CNNs) but requires no training or tuning. The scattering transform generates a compact set of coefficients—scattering coefficients—that capture non-Gaussian information through hierarchical, interpretable summary statistics. This method is particularly suited for fields with localized structures and hierarchical clustering, like cosmological density fields. We apply the scattering transform to weak lensing convergence maps from the CosmoGridV1 simulations, using Gaussian-enveloped, harmonic kernels. These compact, stable descriptors serve as robust statistics, allowing us to make tight cosmological parameter forecasts with the Fisher information matrix, which is 14 times tighter than that of the power spectrum (Cheng et al. 2020). Additionally, we develop a deep-learning emulator trained across cosmologies, incorporating noise, to predict scattering coefficients and employ MCMC to constrain cosmological parameters. Further, a tomographic analysis using auto and cross correlations further tightens these constraints.
| Abstract title | Cosmological Constraints with Weak Lensing Scattering Transform |
|---|