$\texttt{matryoshka}$ II: Accelerating Effective Field Theory Analyses of the Galaxy Power Spectrum

TL;DR

This paper introduces the EFTEMU, a neural network emulator that accelerates Effective Field Theory of Large Scale Structure analyses of galaxy power spectra, achieving high accuracy and speed in parameter inference.

Contribution

The paper presents a new neural network emulator, EFTEMU, that significantly speeds up EFTofLSS analyses while maintaining high accuracy, enabling efficient cosmological parameter inference.

Findings

Prediction accuracy better than 1% on scales up to 0.19 h/Mpc.

Recovers true cosmology within 1σ across multiple redshifts.

Speeds up inference by three orders of magnitude.

Abstract

In this paper we present an extension to the $\texttt{matryoshka}$ suite of neural-network-based emulators. The new editions have been developed to accelerate EFTofLSS analyses of galaxy power spectrum multipoles in redshift space. They are collectively referred to as the $\texttt{EFTEMU}$. We test the $\texttt{EFTEMU}$ at the power spectrum level and achieve a prediction accuracy of better than 1\% with BOSS-like bias parameters and counterterms on scales $0.001\ h\ \mathrm{Mpc}^{-1} \leq k \leq 0.19\ h\ \mathrm{Mpc}^{-1}$. We also run a series of mock full shape analyses to test the performance of the $\texttt{EFTEMU}$ when carrying out parameter inference. Through these mock analyses we verify that the $\texttt{EFTEMU}$ recovers the true cosmology within $1\sigma$ at several redshifts ($z=[0.38,0.51,0.61]$), and with several noise levels (the most stringent of which is Gaussian covariance associated with a volume of $5000^3 \ \mathrm{Mpc}^3 \ h^{-3}$). We compare the mock inference results from the $\texttt{EFTEMU}$ to those obtained with a fully analytic EFTofLSS model and again find no significant bias, whilst speeding up the inference by three orders of magnitude. The $\texttt{EFTEMU}$ is publicly available as part of the $\texttt{matryoshka}$ $\texttt{Python}$ package.