Gravitational Lens Modeling with Genetic Algorithms and Particle Swarm Optimizers

TL;DR

This paper presents a novel approach to gravitational lens modeling using genetic algorithms and particle swarm optimizers, enabling thorough exploration of model parameters and automatic regularization for improved accuracy.

Contribution

It introduces a matrix-free, iterative modeling scheme combined with advanced global optimization techniques and an automatic L-curve method for regularization in gravitational lens analysis.

Findings

Genetic algorithms outperform particle swarm optimizers in parameter space exploration.

The method accurately determines source intensity and lens parameters.

Automatic regularization improves model fitting and degeneracy mapping.

Abstract

Strong gravitational lensing of an extended object is described by a mapping from source to image coordinates that is nonlinear and cannot generally be inverted analytically. Determining the structure of the source intensity distribution also requires a description of the blurring effect due to a point spread function. This initial study uses an iterative gravitational lens modeling scheme based on the semilinear method to determine the linear parameters (source intensity profile) of a strongly lensed system. Our 'matrix-free' approach avoids construction of the lens and blurring operators while retaining the least squares formulation of the problem. The parameters of an analytical lens model are found through nonlinear optimization by an advanced genetic algorithm (GA) and particle swarm optimizer (PSO). These global optimization routines are designed to explore the parameter space thoroughly, mapping model degeneracies in detail. We develop a novel method that determines the L-curve for each solution automatically, which represents the trade-off between the image chi-square and regularization effects, and allows an estimate of the optimally regularized solution for each lens parameter set. In the final step of the optimization procedure, the lens model with the lowest chi-square is used while the global optimizer solves for the source intensity distribution directly. This allows us to accurately determine the number of degrees of freedom in the problem to facilitate comparison between lens models and enforce positivity on the source profile. In practice we find that the GA conducts a more thorough search of the parameter space than the PSO.