Foreground removal in HI 21 cm intensity mapping under frequency-dependent beam distortions

TL;DR

This work addresses foreground subtraction for HI 21 cm intensity mapping in the presence of frequency-dependent beam distortions. It evaluates SDecGMCA, a beam-aware extension of DecGMCA that jointly performs spherical deconvolution and blind source separation, against model-fitting and traditional BSS methods using realistic simulations for MeerKAT/SKA-Mid. The study finds that SDecGMCA provides the most robust HI signal recovery across evolving and oscillating beams, suppresses the spurious spectral peak around $k_ u \sim 0.3$, and achieves $\\lesssim 5\%$ accuracy for $10<\\ell<200$, though beam inversion instability limits performance beyond $\\ell \sim 200$. Masking strategies further improve reconstruction, and the results position SDecGMCA as a promising technique for upcoming full-sky 21 cm intensity mapping surveys, with ongoing work to optimize hyperparameters and incorporate masking directly into the framework.

Abstract

Neutral hydrogen (HI) intensity mapping with single-dish experiments is a powerful approach for probing cosmology in the post-reionization epoch. However, the presence of bright foregrounds over four orders of magnitude stronger than the HI signal makes its extraction highly challenging. While all methods perform well when assuming a Gaussian beam degraded to the worst resolution, most of them degrade significantly in the presence of a more realistic beam model. In this work, we investigate the performance of SDecGMCA. This method extends DecGMCA to spherical data, combining sparse component separation with beam deconvolution. Our goal is to evaluate this method in comparison with established foreground removal techniques, assessing its ability to recover the cosmological HI signal from single-dish intensity mapping observations under varying beam conditions. We use simulated HI signal and foregrounds, covering the frequency ranges relevant to MeerKAT and SKA-Mid. The foreground removal techniques tested fall into two main categories: model-fitting methods (polynomial and parametric) and blind source separation methods (PCA, ICA, GMCA, and SDecGMCA). Their effectiveness is evaluated based on the recovery of the HI angular and frequency power spectra under progressively more realistic beam conditions. While all methods perform adequately under a uniform degraded beam, SDecGMCA remains robust when frequency-dependent beam distortions are introduced. In the oscillating beam case, SDecGMCA suppresses the spurious spectral peak at $k_ν\sim 0.3$ and achieves $\lesssim 5\%$ accuracy at intermediate angular scales ($10 < \ell < 200$), outperforming other methods. Beam inversion, however, remains intrinsically unstable beyond $\ell \sim 200$, setting a practical limit on the method.

Figures (12)

• Figure 1: Angular power spectrum of the reconstructed HI 21 cm signal using different methods, each represented by a distinct color. The bottom panel displays the ratio of the reconstructed to the true HI power spectrum for each method. From left to right, the panels correspond to different beam models: Gaussian degraded, Gaussian evolving, and Gaussian oscillating.
• Figure 2: Equivalent to Fig. \ref{['fig:angular_powspec']}, but showing the frequency power spectrum.
• Figure 3: Total MSE of the SDecGMCA reconstruction as a function of the number of sources $n_s$. The MSE is computed with respect to the ground truth by combining both the frequency and angular power spectra. The curve highlights the trade-off between underfitting at low $n_s$ and overfitting at high $n_s$, with the minimum occurring around $n_s = 5$ (marked in red), which provides the most balanced reconstruction.
• Figure 4: Reconstructed HI frequency (left) and angular (right) power spectra obtained with SDecGMCA for different choices of the number of sources ($n_s$, shown in different colors). The black line shows the input HI signal for reference, and the lower panels display the ratio between the reconstruction and the true signal.
• Figure 5: Filtering lines of sight prior to computing the power spectrum of the reconstructed HI signal. Left: Brightness temperature of the reconstructed HI signal as a function of frequency for all flagged lines of sight that deviate from the expected spectral flatness. Right: Sky map showing the location of the flagged lines of sight, illustrating the regions excluded by the constructed mask. The percentage of lines of sight that are masked is indicated at the top of the map.