Speaker
Description
Foreground contamination remains one of the central challenges in 21 cm intensity mapping, and as experiments become more sensitive, our analysis methods need to keep pace. I'll present a Bayesian forward-modelling framework for jointly separating foregrounds and the 21 cm signal in single-dish data cubes, using Gibbs sampling and Gaussian Constrained Realisations (GCR).
The key challenge we address is scalability: our model has over 2 million free parameters, yet we can draw samples from the full joint posterior in under 30 seconds per iteration on a single CPU core. This is made tractable by ensuring each component — foreground PCA amplitudes, Hi Fourier modes, and their covariances — has a Gaussian conditional distribution, allowing us to solve for the posterior peak directly rather than evaluating an expensive likelihood function.
I'll show results on simulated MeerKLASS-like data, demonstrating recovery of the Hi power spectrum to within 2σ, comparable to the standard transfer function correction approach. Crucially, the framework also handles RFI flagging naturally: rather than requiring explicit inpainting, the GCR steps fill flagged channels with statistically consistent signal realisations as a byproduct of the sampling. We plan to extend this to real MeerKLASS data, including per-antenna systematics and beam effects.
| Consent on use of personal information: Abstract Submission | Yes, I ACCEPT |
|---|