Exploring local fNL estimators based on the binned bispectrum

TL;DR

This paper compares neural network and chi-squared estimators for measuring the local fNL parameter from CMB bispectrum data, demonstrating their robustness and efficiency using WMAP simulations and real data.

Contribution

It introduces a neural network-based estimator for local fNL that converges faster than traditional methods and evaluates the impact of inpainting and linear terms on estimator performance.

Findings

Neural network estimator converges faster than chi-squared methods.

Both estimators are robust to inpainting in simulations.

Real data analysis constrains fNL between -3 and 83 at 95% confidence.

Abstract

We explore different estimators of the local non-linear coupling parameter, fNL, based on the binned bispectrum presented in Bucher et al. Using simulations of Wilkinson Microwave Anisotropy Probe (WMAP)-7yr data, we compare the performance of a regression neural network with a \chi^2-minimization and study the dependence of the results on the presence of the linear term in the analysis and on the use of inpainting for masked regions. Both methods obtain similar results and are robust to the use of inpainting, but the neural network estimator converges considerably faster. We also examine the performance of a simplified \chi^2 estimator that assumes a diagonal matrix and has the linear term subtracted, which considerably reduces the computational time; in this case inpainting is found to be crucial. The estimators are also applied to real WMAP-7yr data, yielding constraints at 95% confidence level of -3< fNL <83.