Revealing the information content of galaxy n-point functions with simulation-based inferenceConfirmed

by Beatriz Tucci (Max Planck Institute for Astrophysics)

Tuesday Dec 3, 2024, 11:00 a.m. → 12:30 p.m. America/Toronto

PI/4-405 - Bob Room (Perimeter Institute for Theoretical Physics)

Improving cosmological constraints from galaxy clustering presents several challenges, particularly in extracting information beyond the power spectrum due to the complexities involved in higher-order n-point function analysis. In this talk, I will introduce novel inference techniques that allow us to go beyond the state-of-the-art, not only by utilizing the galaxy trispectrum, a task that remains computationally infeasible with traditional methods, but also by accessing the full information encoded in the galaxy density field for the first time in cosmological analysis. I will present simulation-based inference (SBI), a powerful deep learning technique that enables cosmological inference directly from summary statistics in simulations, bypassing the need for explicit analytical likelihoods or covariance matrices. This is achieved using LEFTfield, a Lagrangian forward model based on the Effective Field Theory of Large Scale Structure (EFTofLSS) and the bias expansion, ensuring robustness on large scales. Furthermore, LEFTfield enables field-level Bayesian inference (FBI), where a field-level likelihood is used to directly analyze the full galaxy density field rather than relying on compressed statistics. I will conclude by exploring the question of how much cosmological information can be extracted at the field level through a comparison of σ8 constraints obtained from FBI, which directly uses the 3D galaxy density field, and those obtained from n-point functions via SBI.