Fr\'echet Vectors as sensitive tools for blind tests of CMB anomalies

TL;DR

This paper introduces Fréchet Vectors as sensitive tools for blind detection of anomalies in CMB data, demonstrating their effectiveness in simulations and real Planck data to test Gaussianity and isotropy.

Contribution

It presents a novel application of Fréchet Vectors for analyzing CMB anisotropies, improving null tests of Gaussianity and isotropy at small scales.

Findings

FVs successfully detect mock Cold Spot anomalies in simulations.

Planck data shows FVs reject Gaussianity and isotropy at small scales with high significance.

Anisotropic noise simulations reduce some deviations, but significant anomalies remain.

Abstract

Cosmological data collected on a sphere, such as CMB anisotropies, are typically represented by the spherical harmonic coefficients, denoted as $a_{\ell m}$. The angular power spectrum, or $C_\ell$, serves as the fundamental estimator of the variance in this data. Alternatively, spherical data and their variance can also be characterized using Multipole Vectors (MVs) and the Fr\'echet variance. The vectors that minimize this variance, known as Fr\'echet Vectors (FVs), define the center of mass of points on a compact space, and are excellent indicators of statistical correlations between different multipoles. We demonstrate this using both simulations and real data. Through simulations, we show that FVs enable a blind detection and reconstruction of the location associated with a mock Cold Spot anomaly introduced in an otherwise isotropic sky. Applying these tools to the 2018 Planck maps, we implement several improvements on previous null tests of Gaussianity and statistical isotropy, down to arc-minute scales. Planck's MVs appear consistent with these hypotheses at scales $2 \leq\ell \leq 1500$ when the common mask is applied, whereas the same test using the FVs rejects them with significances between 5.3 and 8.2$\sigma$. The inclusion of anisotropic noise simulations render the FVs marginally consistent ($\geq 2\sigma$) with the null hypotheses at the same scales, but still rejects them at $3.5-3.7\sigma$ when we consider scales above $\ell=1500$, where the signal-to-noise is small. Limitations of the noise and/or foregrounds modeling may account for these deviations from the null hypothesis.