The impact of non-Gaussianity on the error covariance for observations of the Epoch of Reionization (EoR) 21-cm power spectrum

Recent simulations show the Epoch of Reionization (EoR) 21-cm signal to be inherently non-Gaussian whereby the error covariance matrix $\mathbf{C}_{ij}$ of the 21-cm power spectrum (PS) contains a trispectrum contribution that would be absent if the signal were Gaussian. Using the binned power spectrum and trispectrum from simulations, here we present a methodology for incorporating these with the baseline distribution and system noise to make error predictions for observations with any radio-interferometric array. Here we consider the upcoming SKA-Low. Non-Gaussianity enhances the errors introducing a positive deviation $\Delta$ relative to the Gaussian predictions. $\Delta$ increases with observation time $t_{\rm obs}$ and saturates as the errors approach the cosmic variance. Considering $t_{\rm obs}=1024$ hours where a $5 \sigma$ detection is possible at all redshifts $7 \le z \le 13$, in the absence of foregrounds we find that the deviations are important at small $k$ where we have $\Delta \sim 40-100 \%$ at $k~\sim 0.04 ~{\rm Mpc}^{-1}$ for some of the redshifts and also at intermediate $k \, (\sim 0.4 ~{\rm Mpc}^{-1})$ where we have $\Delta \sim 200 \%$ at $z=7$. Non-Gaussianity also introduces correlations between the errors in different $k$ bins, and we find both correlations and anticorrelations with the correlation coefficient value spanning $-0.4 \le r_{ij} \le 0.8$. Incorporating the foreground wedge, $\Delta$ continues to be important ($> 50\%$) at $z=7$. We conclude that non-Gaussianity makes a significant contribution to the errors and this is important in the context of the future instruments that aim to achieve high-sensitivity measurements of the EoR 21-cm PS.