Analyzing large-scale polarization data for next-generation CMB experiments

Current and near-future precision measurements give us the unique opportunity to refine the standard model of cosmology. One of the next milestones is the experimental proof of cosmic inflation,which would resolve outstanding problems related to the standard model and shed light on the origin of cosmic structure. Upcoming cosmic microwave background (CMB) experiments aim at a first detection of the curl-like polarization B-modes, a signal that is sourced by cosmological gravitational waves and considered a “smoking gun” of inflation. A key question, not only relevant for B-mode searches but for precision cosmology in general, is how to find the best analysis strategy when systematic effects match or dominate the signal of interest. In this Thesis, we present three original works that address different facets of this question, and apply them to three complementary problems in cosmological data analysis. The main focus of this Thesis is the robust measurement of cosmological B-modes in the presence of Galactic foreground emission, considered to be a major source of systematic bias. We developed, tested, and validated the power-spectrum-based B-mode analysis pipeline for the Simons Observatory, an upcoming experiment that is expected to set new constraints within the next five years. The second work aims at constraining the optical depth to reionization from current large-scale CMB polarization data, which are known to contain non-Gaussian instrumental systematic effects that are difficult to write down in a likelihood model. We developed a novel likelihood-free estimator based on neural networks (NNs) and applied it to real Planck data to retrieve the optical depth directly from maps. This represents, to our awareness, the first cosmological parameter estimation on CMB polarization maps that is performed entirely by NNs. In the third work, we address the difficulty to efficiently marginalize over many astrophysical “nuisance parameters” that commonly arise in two-point analyses of the cosmic large-scale structure (LSS). We present a novel analytical likelihood approximation based on Laplace’s method and apply it to real LSS data, achieving a speedup of a factor 3-5 compared to the standard “brute-force” approach. The rising significance of systematic effects, exemplified by the search for cosmological B-modes, requires efficient problem-tailored analysis methods that are easy to interpret and to employ. It is essential that such methods are robust, meaning that we can reliably exclude potential sources of bias. We anticipate that novel methods, such as hybrid analysis pipelines or NN-based estimators, will prove highly beneficial to this endeavor.