Weak-lensing shear estimates with general adaptive moments, and studies of bias by pixellation, PSF distortions, and noise

TL;DR

This paper investigates the biases in weak lensing shear estimation caused by pixellation, PSF distortions, and noise, proposing a Bayesian approach with adaptive moments to improve accuracy.

Contribution

It introduces a general adaptive moments (GLAM) method for shear estimation, analyzing bias sources and proposing an optimized estimator and Bayesian framework to mitigate them.

Findings

Moment-based shear estimates are prone to underfitting bias due to residuals.

Pixellation fundamentally limits shear estimation accuracy.

Bayesian approach can address noise and prior biases in shear inference.

Abstract

In weak gravitational lensing, weighted quadrupole moments of the brightness profile in galaxy images are a common way to estimate gravitational shear. We employ general adaptive moments (GLAM) to study causes of shear bias on a fundamental level and for a practical definition of an image ellipticity. The GLAM ellipticity has useful properties for any chosen weight profile: the weighted ellipticity is identical to that of isophotes of elliptical images, and in absence of noise and pixellation it is always an unbiased estimator of reduced shear. We show that moment-based techniques, adaptive or unweighted, are similar to a model-based approach in the sense that they can be seen as imperfect fit of an elliptical profile to the image. Due to residuals in the fit, moment-based estimates of ellipticities are prone to underfitting bias when inferred from observed images. The estimation is fundamentally limited mainly by pixellation which destroys information on the original, pre-seeing image. We give an optimized estimator for the pre-seeing GLAM ellipticity and quantify its bias for noise-free images. To deal with pixel noise, we consider a Bayesian approach where the posterior of the GLAM ellipticity can be inconsistent with the true ellipticity if we do not properly account for our ignorance about fit residuals. This underfitting bias is S/N-independent but changes with the pre-seeing brightness profile and the correlation or heterogeneity of pixel noise over the post-seeing image. Furthermore, when inferring a constant ellipticity or, more relevantly, constant shear from a source sample with a distribution of intrinsic properties (sizes, centroid positions, intrinsic shapes), an additional, now noise-dependent bias arises towards low S/N if incorrect priors for the intrinsic properties are used. We discuss the origin of this prior bias.