A Bayesian method for combining theoretical and simulated covariance matrices for large-scale structure surveys

TL;DR

This paper introduces a Bayesian method that combines theoretical and simulated covariance matrices, reducing the number of simulations needed for large-scale structure surveys by incorporating prior information and propagating uncertainties.

Contribution

It presents a novel Bayesian framework that integrates theoretical models and simulations for covariance estimation, allowing fewer simulations and improved uncertainty quantification.

Findings

The method reduces the number of simulations needed for accurate covariance estimation.

Weakly informative priors can significantly decrease simulation requirements.

The correlation matrix of summary statistics is crucial for covariance modeling.

Abstract

Accurate and precise covariance matrices will be important in enabling planned cosmological surveys to detect new physics. Standard methods imply either the need for many N-body simulations in order to obtain an accurate estimate, or a precise theoretical model. We combine these approaches by constructing a likelihood function conditioned on simulated and theoretical covariances, consistently propagating noise from the finite number of simulations and uncertainty in the theoretical model itself using an informative Inverse-Wishart prior. Unlike standard methods, our approach allows the required number of simulations to be less than the number of summary statistics. We recover the linear 'shrinkage' covariance estimator in the context of a Bayesian data model, and test our marginal likelihood on simulated mock power spectrum estimates. We conduct a thorough investigation into the impact of prior confidence in different choices of covariance models on the quality of model fits and parameter variances. In a simplified setting we find that the number of simulations required can be reduced if one is willing to accept a mild degradation in the quality of model fits, finding that even weakly informative priors can help to reduce the simulation requirements. We identify the correlation matrix of the summary statistics as a key quantity requiring careful modelling. Our approach can be easily generalized to any covariance model or set of summary statistics, and elucidates the role of hybrid estimators in cosmological inference.