Optimal multifrequency weighting for CMB lensing

TL;DR

This paper develops an optimal multi-frequency combination method to reduce biases in CMB lensing reconstruction caused by extragalactic foregrounds, improving accuracy while managing noise levels.

Contribution

It introduces a new optimal weighting scheme that balances bias reduction and noise increase, tailored for CMB lensing estimators and foreground mitigation.

Findings

Profile hardening reduces bias by 40% with 20% noise increase.

Joint deprojection further decreases bias to below statistical uncertainty.

Method achieves significant bias reduction with manageable noise trade-offs.

Abstract

Extragalactic foregrounds in Cosmic Microwave Background (CMB) temperature maps lead to significant biases in CMB lensing reconstruction if not properly accounted for. Combinations of multi-frequency data have been used to minimize the overall map variance (internal linear combination, or ILC), or specifically null a given foreground, but these are not tailored to CMB lensing. In this paper, we derive an optimal multi-frequency combination to jointly minimize CMB lensing noise and bias. We focus on the standard lensing quadratic estimator, as well as the "shear-only" and source-hardened estimators, whose responses to foregrounds differ. We show that an optimal multi-frequency combination is a compromise between the ILC and joint deprojection, which nulls the thermal Sunyaev-Zel'dovich (tSZ) and Cosmic Infrared Background (CIB) contributions. In particular, for a Simons Observatory-like experiment with $\ell_{\text{max},T}=3000$, we find that profile hardening alone (with the standard ILC) reduces the bias to the lensing power amplitude by $40\%$, at a $20\%$ cost in noise, while the bias to the cross-correlation with a LSST-like sample is reduced by nearly an order of magnitude at a $10\%$ noise cost, relative to the standard quadratic estimator. With a small amount of joint deprojection the bias to the profile hardened estimator can be further reduced to less than half the statistical uncertainty on the respective amplitudes, at a $20\%$ and $5\%$ noise cost for the auto- and cross-correlation respectively, relative to the profile hardened estimator with the standard ILC weights. Finally, we explore possible improvements with more aggressive masking and varying $\ell_{\text{max,}T}$.