The epoch of reionization (EoR) 21-cm signal is expected to become highly non-Gaussian as
reionization progresses. This severely affects the error-covariance of the EoR 21-cm power
spectrum that is important for predicting the prospects of a detection with ongoing and future
experiments. Most earlier works have assumed that the EoR 21-cm signal is a Gaussian random
field where (1) the error-variance depends only on the power spectrum and the number of
Fourier modes in the particular k bin, and (2) the errors in the different k bins are uncorrelated.
Here, we use an ensemble of simulated 21-cm maps to analyse the error-covariance at various
stages of reionization. We find that even at the very early stages of reionization (x¯H I ∼ 0.9),
the error-variance significantly exceeds the Gaussian predictions at small length-scales (k >
0.5 Mpc−1) while they are consistent at larger scales. The errors in most k bins (both large and
small scales) are however found to be correlated. Considering the later stages (x¯H I = 0.15), the
error-variance shows an excess in all k bins within k ≥ 0.1 Mpc−1, and it is around 200 times
larger than the Gaussian prediction at k ∼ 1 Mpc−1. The errors in the different k bins are all also
highly correlated, barring the two smallest k bins that are anti-correlated with the other bins.
Our results imply that the predictions for different 21-cm experiments based on the Gaussian
assumption underestimate the errors, and it is necessary to incorporate the non-Gaussianity
for more realistic predictions.
Key words: methods: statistical