COMET: Clustering Observables Modelled by Emulated perturbation Theory

In this paper we present COMET, a Gaussian process emulator of the galaxy power spectrum multipoles in redshift-space. The model predictions are based on one-loop perturbation theory and we consider two alternative descriptions of redshift-space distortions: one that performs a full expansion of the real- to redshift-space mapping, as in recent effective field theory models, and another that preserves the non-perturbative impact of small-scale velocities by means of an effective damping function. The outputs of COMET can be obtained at arbitrary redshifts (up to $z \sim 3$), for arbitrary fiducial background cosmologies, and for a large parameter space that covers the shape parameters $ω_c$, $ω_b$, and $n_s$, as well as the evolution parameters $h$, $A_s$, $Ω_K$, $w_0$, and $w_a$. This flexibility does not impair COMET's accuracy, since we exploit an exact degeneracy between the evolution parameters that allows us to train the emulator on a significantly reduced parameter space. While the predictions are sped up by at least two orders of magnitude, validation tests reveal an accuracy of $0.1\,\%$ for the monopole and quadrupole ($0.3\,\%$ for the hexadecapole), or alternatively, better than $0.25\,σ$ for all three multipoles in comparison to statistical uncertainties expected for the Euclid survey with a tenfold increase in volume. We show that these differences translate into shifts in mean posterior values that are at most of the same size, meaning that COMET can be used with the same confidence as the exact underlying models. COMET is a publicly available Python package that also provides the tree-level bispectrum multipoles in redshift-space and Gaussian covariance matrices.