Accurate parameter inference for the Light-cone Epoch of Reionization 21-cm signal

Abstract

The light-cone (LC) effect introduces line-of-sight (LoS) statistical inhomogeneity into the 21-cm signal. Consequently, the traditional power spectrum (PS) fails to capture the full two-point statistical information. The evolving power spectrum (ePS), $P_e(k, z)$, offers an alternative that accounts for this LoS evolution. We compare the statistical power of three different summary statistics: the standard cylindrical PS $P(k_\perp,k_\parallel)$, slice-wise PS $P_s(k, z)$ (3D PS for small bandwidth LC slices), and ePS $P_e(k, z)$. We first demonstrate that $P_e(k,z)$ successfully recovers the benchmark 3D PS of coeval simulations across most $k$ and $z$, whereas the slice-wise PS recovers only at large $k$. To efficiently perform parameter inference, we train artificial neural network (ANN) emulators on $500$ LC 21-cm signals. Our forecasts incorporate cosmic variance, estimated using $50$ statistically independent realizations of the signal, alongside SKA-Low system noise for integration times of $1000$ and $104$ hrs. We find that ePS outperforms its peers, yielding $3$ and $1.4$ times tighter constraints than $P(k_\perp,k_\parallel)$ and $P_s(k,z)$, respectively. Our results establish the ePS as an optimal summary statistic for interpreting forthcoming data.