Equivalence of the field-level inference and conventional analyses on large scales

Abstract

We study a simple setup with dark matter halos in real space, with the amplitude of the linear density field $A$ as the only free cosmological parameter. We show that Eulerian perturbation theory is adequate for describing this system on large scales, compute the leading $n$-point functions and perform a joint power spectrum, bispectrum and trispectrum analysis. Beyond the bispectrum which is crucial for breaking the degeneracy between $A$ and the linear bias, we find that addition of the trispectrum reduces the error on $A$ by only $20-30\%$. Our results for the joint analysis are in good agreement with recent field-level analyses in the same setup. This implies that the field-level inference on large scales does not get significant information from large displacements beyond those in Eulerian kernels or higher-order $n$-point functions beyond the trispectrum. We provide further evidence for this showing that the dependence of the error bars on the maximum wavenumbers used in the analysis is the same in the two approaches. Our results are in disagreement with some of the recent joint power spectrum and bispectrum analyses using likelihood-free inference based on perturbative forward modeling. We discuss a possible origin of this discrepancy and highlight the importance of resolving it in order to have the optimal results in cosmological analyses based on perturbation theory.