Developing statistical inference tools for future observations of the cosmic dawn and the epoch of reionization

Abstract

The epoch of reionization (EoR) is one of the least known periods in the history of the Universe. The next-generation telescopes (e.g. the SKA, HERA) will have the capability to directly probe the distribution of neutral hydrogen from this era through the redshifted 21-cm line. Once such observations are successful in detecting the 21-cm signal from the EoR, one would then aim to constrain the astrophysical parameters of the EoR accurately by doing statistical analysis of the observed data. One way of achieving this is through Bayesian inference of various signal statistics. To draw inference in a Bayesian framework, one would need to compare the observed signal statistic with a model statistic while performing random walks in the multi-dimensional astrophysical parameter space. In case of the EoR 21-cm signal, this implies simulating the signal in a large cosmological volume with reasonably high precision for each step of the random walker in this parameter space. The conventional simulations of the EoR (radiative transfer or semi-numerical) takes a large amount of computer memory ( 1 TB in RAM) and a significant amount of computation time to simulate the signal in a reasonable volume for one such set of parameters (i.e. one step of the random walker). The requirement of computing time will be proportionately higher when one has to take of the order of millions of random steps in the parameter space. To circumvent this problem, we have developed an artificial neural network (ANN) based EoR 21-cm signal statistics (presently for power spectrum and bispectrum) emulator EmuPBk12, which is orders of magnitude faster than a semi-numerical simulation. The EmuPBk was trained over 1000 semi-numerically simulated (using Region-Yuga code) EoR 21-cm power spectra and bispectra. Our tests show that it has a reasonable capability of predicting the EoR 21-cm power spectra and bispectra. Further, using these emulated power spectra and bispectra of the signal, our MCMC analysis based Bayesian inference shows that one will be able to put tighter constraints on the reionization parameters using the bispectra compared to the power spectra. This is due to the fact that the 21-cm signal from the EoR is highly non-Gaussian and power spectra do not capture this non-Gaussianity. However, the bispectra is sensitive to such non-Gaussianities which are dependent on the time evolving topology of the signal.