A compact group lens modeled with GIGA-Lens: Enhanced inference for complex systems

TL;DR

This paper enhances the GIGA-Lens code with GPU-accelerated Bayesian inference to efficiently model complex multi-galaxy lens systems, demonstrated on a compact group lens with promising results for large astronomical surveys.

Contribution

It introduces an improved, fast inference method combining image and pixel data with annealing sampling for complex lens modeling, capable of handling many parameters with limited priors.

Findings

Successfully modeled a lens system with 29 free parameters

Measured redshifts for the group and arcs accurately

Predicted additional sources at different redshifts

Abstract

In the era of large-scale astronomical surveys, fast modeling of strong lens systems has become increasingly vital. While significant progress has been made for galaxy-scale lenses, the development of automated methods for modeling larger systems, such as groups and clusters, is not as extensive. Our study aims to extend the capabilities of the GIGA-Lens code, enhancing its efficiency in modeling multi-galaxy strong lens systems. We focus on demonstrating the potential of GPU-accelerated Bayesian inference in handling complex lensing scenarios with a high number of free parameters. We employ an improved inference approach that combines image position and pixelated data with an annealing sampling technique to obtain the posterior distribution of complex models. This method allows us to overcome the challenge of limited prior information, a high number of parameters, and memory usage. Our process is exemplified through the analysis of the compact group lens system DES J0248-3955, for which we present VLT/X-shooter spectra. We measure a redshift of $z = 0.69 \pm 0.04$ for the group, and $z = 1.2722 \pm 0.0005$ for one of the extended arcs. Our enhanced method successfully constrained a lens model with 29 free parameters and lax priors in a remarkably short time. The mass of the lens is well described by a single dark-matter halo with a velocity dispersion of $\sigma_v = (690 \pm 30) \, km \, s^{-1}$. The model predicts the presence of a second source at the same redshift and a third source at approximately $z \sim 2.7$. Our study demonstrates the effectiveness of our lens modeling technique for dealing with a complex system in a short time using ground-based data. This presents considerable potential within the context of large surveys such as LSST.