Institute for Advanced Study/Princeton University Early Universe/Cosmology Lunch Discussion

Searching for signatures of non-Gaussianity in primordial fluctuations is one of the foremost quests in modern cosmology. In my talk, I will present a novel, non-perturbative formalism to constrain the local type primordial non-Gaussianity (PNG) parameter f_NL, tapping into information from deeply non-linear scales. Our formalism defines an auxiliary field, \pi , which is an estimate of local small-scale power spectrum of a late-time cosmological density field. The local PNG signature appears as a unique 1/k2 term in the bias of \pi field with respect to the density field. The auto-spectrum and cross-spectrum of the field with the large-scale density combines information about local PNG contained in the squeezed bispectrum and collapsed trispectrum in a computationally amenable and consistent way, while avoiding the need to model complicated covariances of these higher N-point functions. We subsequently generalize our work and train a neural network to estimate the local amplitude of fluctuations. The \pi field returned by the trained neural network is sensitive to PNG signature contained in the soft-limits of higher N-point functions. We apply our method to simulations and show that constraints on f_NL are several times better than those obtained from a standard matter+halo based analysis. Our method is robust to systematics as symmetry arguments imply that the signature cannot be faked by any other physics.