Mitigating point-source contamination in CMB polarization: a Generalized Point Spread Function fitting approach

Abstract

Observations of Cosmic Microwave Background (CMB) B-mode polarization provide a way to probe primordial gravitational waves and test inflationary predictions. Extragalactic point sources become a major source of contamination after foreground cleaning and can bias estimates of the tensor-to-scalar ratio $r$ at the $10^{-3}$ level. We introduce Generalized Point Spread Function Fitting (GPSF), a method for removing point-source contamination in polarization maps. GPSF uses the full pixel-domain covariance, including off-diagonal terms, and models overlapping sources. This allows accurate flux estimation under realistic conditions, particularly for small-aperture telescopes with large beams that are more susceptible to source blending. We test GPSF on simulated sky maps, apply foreground cleaning using the Needlet Internal Linear Combination (NILC) method, and compare its performance with standard masking and inpainting. The results show GPSF reduces point-source contamination without significantly affecting the background signal, as seen in both the maps and their power spectra. For the constraint on $r$, GPSF reduces the bias from $1.67 \times 10^{-3}$ to $2.9 \times 10^{-4}$, with only a 2% increase in standard deviation. Compared to inpainting and masking, GPSF yields lower bias while maintaining comparable variance. This suggests that it may serve as a promising method for future CMB experiments targeting measurements of $r \sim 10^{-3}$.

Figures (13)

• Figure 1: Orthogonal projection of the northern sky patch and its corresponding masks in Equatorial coordinates. The patch is centered at $(RA = 197^\circ, Dec = 55^\circ)$, extending from $RA = 150^\circ$ to $RA = 250^\circ$ and from $Dec = 25^\circ$ to $Dec = 70^\circ$. The background represents the polarization intensity map, $P = \sqrt{Q^2 + U^2}$, observed by Planck at 353 GHz.
• Figure 2: B-mode polarization maps for a representative point-source region at 30 GHz. The leftmost panel shows the No PS case, which serves as the reference for all comparisons. The upper row displays the reconstructed maps for the PS unmitigated, GPSF, and Inpainting cases, with the dashed line separating the reference from the reconstructed maps. The lower row shows the corresponding difference maps relative to the No PS case.
• Figure 3: Same as Figure \ref{['fig:freq30_maps']} but for 155 GHz.
• Figure 4: Noise-debiased $B$-mode power spectra (left) and their relative residuals and relative standard deviations (right) for different point-source mitigation methods. The upper row shows results at 30 GHz, and the lower row at 155 GHz. In each row, the left panel presents the noise-debiased B-mode power spectrum with standard deviations estimated from 200 realizations (accounting for CMB and noise), while the right panel displays the relative residuals and relative standard deviations. The No PS serves as the reference. The relative residual is defined as the difference between the mean power spectrum of each method and that of the reference, normalized by the reference mean. The relative standard deviation is computed as the standard deviation of each method divided by its own mean.
• Figure 5: NILC B-mode polarization maps in a representative sky region. The leftmost panel shows the fiducial (input) CMB realization used as the reference. The upper row displays the reconstructed maps for the No PS, PS unmitigated, GPSF, and Inpainting cases, with the dashed line separating the reference from the reconstructed maps. The lower row shows the corresponding difference maps relative to the fiducial CMB.