Impact of Galactic dust non-Gaussianity on searches for B-modes from inflation

TL;DR

This paper examines how non-Gaussian features of Galactic dust affect the analysis of primordial B-modes in the CMB, finding minimal impact on tensor-to-scalar ratio uncertainty for upcoming experiments but highlighting potential issues in goodness-of-fit assessments.

Contribution

It demonstrates that dust non-Gaussianity has negligible effect on B-mode detection uncertainty for future experiments, and provides guidance on avoiding biases in goodness-of-fit metrics.

Findings

Dust non-Gaussianity does not significantly affect $\sigma(r)$ for upcoming experiments.

Neglecting dust non-Gaussianity can bias goodness-of-fit metrics.

Using cleaned spectrum combinations mitigates issues in goodness-of-fit tests.

Abstract

A key challenge in the search for primordial B-modes is the presence of polarized Galactic foregrounds, especially thermal dust emission. Power-spectrum-based analysis methods generally assume the foregrounds to be Gaussian random fields when constructing a likelihood and computing the covariance matrix. In this paper, we investigate how non-Gaussianity in the dust field instead affects CMB and foreground parameter inference in the context of inflationary B-mode searches, capturing this effect via modifications to the dust power-spectrum covariance matrix. For upcoming experiments such as the Simons Observatory, we find no dependence of the tensor-to-scalar ratio uncertainty $\sigma(r)$ on the degree of dust non-Gaussianity or the nature of the dust covariance matrix. We provide an explanation of this result, noting that when frequency decorrelation is negligible, dust in mid-frequency channels is cleaned using high-frequency data in a way that is independent of the spatial statistics of dust. We show that our results hold also for non-zero levels of frequency decorrelation that are compatible with existing data. We find, however, that neglecting the impact of dust non-Gaussianity in the covariance matrix can lead to inaccuracies in goodness-of-fit metrics. Care must thus be taken when using such metrics to test B-mode spectra and models, although we show that any such problems can be mitigated by using only cleaned spectrum combinations when computing goodness-of-fit statistics.