Bias-Free Shear Estimation using Artificial Neural Networks

TL;DR

This paper introduces a neural network-based method to nearly eliminate bias in shear estimation for weak lensing surveys, improving accuracy over traditional methods and demonstrating competitive results on simulated data.

Contribution

The authors develop a neural network approach for shear estimation that adapts bias correction to individual galaxy properties, outperforming existing pipelines.

Findings

Neural networks significantly reduce shear bias in simulations.

Circularization of the PSF improves measurement accuracy.

Results are competitive with top methods in GREAT08.

Abstract

Bias due to imperfect shear calibration is the biggest obstacle when constraints on cosmological parameters are to be extracted from large area weak lensing surveys such as Pan-STARRS-3pi, DES or future satellite missions like Euclid. We demonstrate that bias present in existing shear measurement pipelines (e.g. KSB) can be almost entirely removed by means of neural networks. In this way, bias correction can depend on the properties of the individual galaxy instead on being a single global value. We present a procedure to train neural networks for shear estimation and apply this to subsets of simulated GREAT08 RealNoise data. We also show that circularization of the PSF before measuring the shear reduces the scatter related to the PSF anisotropy correction and thus leads to improved measurements, particularly on low and medium signal-to-noise data. Our results are competitive with the best performers in the GREAT08 competition, especially for the medium and higher signal-to-noise sets. Expressed in terms of the quality parameter defined by GREAT08 we achieve a Q = 40, 140 and 1300 without and 50, 200 and 1300 with circularization for low, medium and high signal-to-noise data sets, respectively.