Exploring the posterior surface of the large scale structure reconstruction

Abstract

The large scale structure (LSS) of the universe is generated by the linear density gaussian modes, which are evolved into the observed nonlinear LSS. The posterior surface of the modes is convex in the linear regime, leading to a unique global maximum (MAP), but this is no longer guaranteed in the nonlinear regime. In this paper we investigate the nature of posterior surface using the recently developed MAP reconstruction method, with a simplified but realistic N-body simulation as the forward model. The reconstruction method uses optimization with analytic gradients from back-propagation through the simulation. For low noise cases we recover the initial conditions well into the nonlinear regime ($k\sim 1$ h/Mpc) nearly perfectly. We show that the large scale modes can be recovered more precisely than the linear expectation, which we argue is a consequence of nonlinear mode coupling. For noise levels achievable with current and planned LSS surveys the reconstruction cannot recover very small scales due to noise. We see some evidence of non-convexity, specially for smaller scales where the non-injective nature of the mappings: several very different initial conditions leading to the same near perfect final data reconstruction. We investigate the nature of these phenomena further using a 1-d toy gravity model, where many well separated local maximas are found to have identical data likelihood but differ in the prior. We also show that in 1-d the prior favors some solutions over the true solution, though no clear evidence of these in 3-d. Our main conclusion is that on very small scales and for a very low noise the posterior surface is multi-modal and the global maximum may be unreachable with standard methods, while for realistic noise levels in the context of the current and next generation LSS surveys MAP optimization method is likely to be nearly optimal.

Version accepted for JCAP

References in corpus (3)

• Improving Cosmological Distance Measurements by Reconstruction of the Baryon Acoustic Peak
• ELUCID - Exploring the Local Universe with reConstructed Initial Density field I: Hamiltonian Markov Chain Monte Carlo Method with Particle Mesh Dynamics
• Bayesian reconstruction of the cosmological large-scale structure: methodology, inverse algorithms and numerical optimization

Cited by in corpus (21)

• Cosmological Parameters from the BOSS Galaxy Power Spectrum
• Learning to Predict the Cosmological Structure Formation
• Cosmological Parameters and Neutrino Masses from the Final Planck and Full-Shape BOSS Data
• Reconstructing large-scale structure with neutral hydrogen surveys
• ENCORE: An $\mathcal{O}(N_{\mathrm{g}}^2)$ Estimator for Galaxy $N$-Point Correlation Functions
• TARDIS Paper I: A Constrained Reconstruction Approach to Modeling the z~2.5 Cosmic Web Probed by Lyman-alpha Forest Tomography
• A fast particle-mesh simulation of non-linear cosmological structure formation with massive neutrinos
• The Reconstructed Power Spectrum in the Zeldovich Approximation
• Consistency tests of field level inference with the EFT likelihood
• Machine Learning for Observational Cosmology
• Baryon acoustic oscillations reconstruction using convolutional neural networks
• Reconstructing Cosmological Initial Conditions from Late-Time Structure with Convolutional Neural Networks
• Neural physical engines for inferring the halo mass distribution function
• Joint velocity and density reconstruction of the Universe with nonlinear differentiable forward modeling
• Understanding the Reconstruction of the Biased Tracer
• Biased tracer reconstruction with halo mass information
• On the primordial information available to galaxy redshift surveys
• Modeling iterative reconstruction and displacement field in the large scale structure
• Learning the Universe: Learning to Optimize Cosmic Initial Conditions with Non-Differentiable Structure Formation Models
• Fast Sampling of Cosmological Initial Conditions with Gaussian Neural Posterior Estimation
• On the Connection between Field-Level Inference and $n$-point Correlation Functions