Variations of cosmic large-scale structure covariance matrices across parameter space

TL;DR

This paper develops an analytical method using third-order Eulerian perturbation theory to model how cosmic large-scale structure covariance matrices vary with cosmological parameters, reducing reliance on costly simulations.

Contribution

It introduces a basis-based formalism to efficiently approximate covariance matrix variations across parameter space using perturbation theory, enabling optimized sampling for simulations.

Findings

Method accurately reproduces covariance variations within errors for multipoles up to 1300.

Formalism captures expected degeneracies and scaling with cosmological parameters.

Proposes an economical sampling strategy to minimize interpolation errors in parameter space.

Abstract

The likelihood function for cosmological parameters, given by e.g. weak lensing shear measurements, depends on contributions to the covariance induced by the nonlinear evolution of the cosmic web. As nonlinear clustering to date has only been described by numerical $N$-body simulations in a reliable and sufficiently precise way, the necessary computational costs for estimating those covariances at different points in parameter space are tremendous. In this work we describe the change of the matter covariance and of the weak lensing covariance matrix as a function of cosmological parameters by constructing a suitable basis, where we model the contribution to the covariance from nonlinear structure formation using Eulerian perturbation theory at third order. We show that our formalism is capable of dealing with large matrices and reproduces expected degeneracies and scaling with cosmological parameters in a reliable way. Comparing our analytical results to numerical simulations we find that the method describes the variation of the covariance matrix found in the SUNGLASS weak lensing simulation pipeline within the errors at one-loop and tree-level for the spectrum and the trispectrum, respectively, for multipoles up to $\ell\leq 1300$. We show that it is possible to optimize the sampling of parameter space where numerical simulations should be carried out by minimising interpolation errors and propose a corresponding method to distribute points in parameter space in an economical way.