[Paper Review] Foreground removal and 21 cm signal estimates: comparing different blind methods for the BINGO Telescope

Alessandro Marins, F. B. Abdalla+18 more|arXiv (Cornell University)|Sep 23, 2022

Blind Source Separation Techniques4 citations

TL;DR

This study evaluates three blind source separation methods—FastICA, GNILC, and GMCA—for foreground removal in the BINGO Telescope's 21 cm intensity mapping survey. Using simulated sky maps with 400 realizations and a foreground subspace dimension of three, all methods yield statistically equivalent 21 cm signal estimates, with FastICA showing the lowest computational cost, enabling efficient pipeline processing for cosmological power spectrum estimation in the 980–1260 MHz band.

ABSTRACT

The BINGO radiotelescope will observe hydrogen distribution using Intensity Mapping (IM) to analyze the Dark Energy paradigm through Baryon Acoustic Oscillations. The target signal is contaminated by unwanted signals and instrumental noise, making accurate estimations essential for characterizing the 21 cm signal. In this study, we evaluated the performance of three blind foreground-removing algorithms - FastICA, GNILC, and GMCA - on the BINGO pipeline. Each method used different approaches to estimate foreground contributions, and we also investigated how the number of simulations for debiasing affects estimation quality. Our findings indicate that using 50 or 400 simulations yields equivalent results at this stage of analysis. All algorithms produced statistically consistent estimates of the 21 cm signal. We used FastICA for estimating and debiasing the HI spectra from five years of observations, which yielded reliable results, although the first channel was affected by edge effects from the mixing matrix. The overall signal-to-noise ratio was 204, and the chi-squared value was 1.8.

Motivation & Objective

• To assess the performance of blind component separation algorithms in isolating the 21 cm signal from astrophysical foregrounds in the BINGO Telescope's intensity mapping survey.
• To determine the optimal dimension of the foreground subspace for accurate 21 cm signal estimation under realistic simulation conditions.
• To evaluate the impact of simulation count (50 vs. 400 realizations) and angular binning on signal reconstruction quality and computational efficiency.
• To compare the computational cost and accuracy of FastICA, GNILC, and GMCA in the current stage of the BINGO data pipeline.
• To validate the robustness of signal recovery across different frequency channels and sky regions, especially near the Galactic plane.

Proposed method

• Applied three blind source separation algorithms—FastICA, GNILC, and GMCA—to simulated sky maps containing 21 cm signal, foregrounds, and white instrumental noise.
• Used a fixed beam resolution of 40 arcminutes and a Gaussian beam, with five foreground components and no polarization leakage or sidelobe effects.
• Conducted signal reconstruction using 400 simulation realizations to estimate and subtract noise bias, with varying numbers of simulations (50 and 400) for comparison.
• Evaluated the angular power spectrum of the reconstructed 21 cm signal and compared it to the true signal and noise power spectra across four representative channels.
• Assessed algorithm performance by analyzing residuals and differences between reconstructions, particularly in the Galactic plane and masked regions.
• Tested the effect of reducing the number of multipoles in the angular power spectrum estimation to accelerate computation while preserving accuracy.

Experimental results

Research questions

• RQ1 How do FastICA, GNILC, and GMCA perform in recovering the 21 cm signal from foreground-contaminated maps in the BINGO Telescope's frequency range?
• RQ2 What is the optimal dimension of the foreground subspace that minimizes reconstruction error in the current BINGO simulation setup?
• RQ3 Does increasing the number of simulation realizations from 50 to 400 significantly improve the accuracy of 21 cm signal estimation?
• RQ4 Can the number of multipoles used in angular power spectrum estimation be reduced by half without degrading the signal recovery quality?
• RQ5 How do the algorithms compare in terms of computational cost and residual error, especially in regions with strong Galactic foregrounds?

Key findings

• All three blind source separation algorithms—FastICA, GNILC, and GMCA—produce statistically equivalent 21 cm signal estimates when using 400 simulation realizations and a foreground subspace dimension of three.
• The optimal foreground subspace dimension for best signal recovery is three, composed of non-physical templates, indicating that the foreground structure in the simulated data is effectively captured in this low-dimensional space.
• Using 50 or 400 simulation realizations yields statistically equivalent signal estimation performance, suggesting that 50 simulations may be sufficient for noise bias correction in this context.
• The number of multipoles in the angular power spectrum estimation can be reduced by half without significant loss in reconstruction quality, enabling faster computation.
• FastICA demonstrates the lowest computational cost among the three algorithms, making it the most efficient choice for the current stage of the BINGO pipeline.
• Residual analysis reveals that GMCA slightly underestimates Galactic foreground emissions, while FastICA shows a subtle but consistent difference in reconstruction quality in the Galactic plane, though both remain within acceptable statistical bounds.