Super sample covariance and the volume scaling of galaxy survey covariance matrices

TL;DR

This paper investigates methods to accurately include super sample covariance in galaxy survey covariance matrices, proposes volume-scaling techniques to reduce computational costs, and demonstrates a practical approach to achieve high accuracy with smaller simulations.

Contribution

It introduces a new method for correcting mode effects, compares perturbative expansions for SSC, and demonstrates volume scaling of covariance matrices with high accuracy for galaxy surveys.

Findings

Perturbative methods yield similar SSC results.

Volume scaling can achieve 3% accuracy in covariance matrices.

A 512x volume increase reduces computational costs significantly.

Abstract

Super sample covariance (SSC) is important when estimating covariance matrices using a set of mock catalogues for galaxy surveys. If the underlying cosmological simulations do not include the variation in background parameters appropriate for the simulation sizes, then the scatter between mocks will be missing the SSC component. The coupling between large and small modes due to non-linear structure growth makes this pernicious on small scales. We compare different methods for generating ensembles of mocks with SSC built in to the covariance, and contrast against methods where the SSC component is computed and added to the covariance separately. We find that several perturbative expansions, developed to derive background fluctuations, give similar results. We then consider scaling covariance matrices calculated for simulations of different volumes to improve the accuracy of the covariance matrix for a given computational time. On large scales, we find that the primary limitation is from the discrete number of modes contributing to the measured power spectrum, and we propose a new method for correcting this effect. Correct implementation of SSC and the effect of discrete mode numbers allows covariance matrices created from mocks to be scaled between volumes, potentially leading to a significant saving on computational resources when producing covariance matrices. We argue that a sub-percent match is difficult to achieve because of the effects of modes on scales between the box sizes, which cannot be easily included. Even so, when working in real space and cubic boxes, we show that a 3% match in the dark matter power spectrum covariance is achievable on scales of interest for current surveys scaling the simulation volume by 512x, costing a small fraction of the computational time of running full-sized simulations.