Epoch of reionization parameter estimation with the 21-cm bispectrum

TL;DR

This paper demonstrates that incorporating the isosceles bispectrum alongside the power spectrum improves parameter inference accuracy for the 21-cm signal from the epoch of reionization, using mock SKA1-low observations.

Contribution

It introduces the first application of the isosceles bispectrum to MCMC inference for the 21-cm signal and shows its effectiveness in reducing bias and computational costs.

Findings

Using bispectrum reduces bias in parameter estimates.

Gaussian likelihood assumption does not significantly bias results.

Cosmic variance error can be interpolated to save computation.

Abstract

We present the first application of the isosceles bispectrum to MCMC parameter inference from the cosmic 21-cm signal. We extend the MCMC sampler 21cmMC to use the fast bispectrum code, BiFFT, when computing the likelihood. We create mock 1000h observations with SKA1-low, using PyObs21 to account for uv-sampling and thermal noise. Assuming the spin temperature is much higher than that of the CMB, we consider two different reionization histories for our mock observations: fiducial and late-reionization. For both models we find that bias on the inferred parameter means and 1-$\sigma$ credible intervals can be substantially reduced by using the isosceles bispectrum (calculated for a wide range of scales and triangle shapes) together with the power spectrum (as opposed to just using one of the statistics). We find that making the simplifying assumption of a Gaussian likelihood with a diagonal covariance matrix does not notably bias parameter constraints for the three-parameter reionization model and basic instrumental effects considered here. This is true even if we use extreme (unlikely) initial conditions which would be expected to amplify biases. We also find that using the cosmic variance error calculated with Monte-Carlo simulations using the fiducial model parameters whilst assuming the late-reionization model for the simulated data also does not strongly bias the inference. This implies we may be able to sparsely sample and interpolate the cosmic variance error over the parameter space, substantially reducing computational costs. All codes used in this work are publicly-available.