Improving constraints on primordial non-Gaussianity using neural network based reconstruction

Abstract

We study the use of U-Nets in reconstructing the linear dark matter density field and its consequences for constraining cosmological parameters, in particular primordial non-Gaussianity. Our network is able to reconstruct the initial conditions of redshift z = 0 density fields from N-body simulations with 90% accuracy out to k ≤ 0.4 h/Mpc, competitive with state-of-the-art reconstruction algorithms at a fraction of the computational cost. We study the information content of the reconstructed z = 0 density field with a Fisher analysis using the QUIJOTE simulation suite, including non-Gaussian initial conditions. Combining the pre- and post-reconstructed power spectrum and bispectrum data up to k max = 0.52 h/Mpc, we find significant improvements in all parameters. Most notably, we find a factor 3.65 (local), 3.54 (equilateral), and 2.90 (orthogonal) improvement on the marginalized errors of f NL as compared to only using the pre-reconstructed data. We show that these improvements can be attributed to a combination of reduced data covariance and parameter degeneracy. The results constitute an important step towards a more optimal inference of primordial non-Gaussianity from non-linear scales.