A lensed radio jet at milli-arcsecond resolution I: Bayesian comparison of parametric lens models

We investigate the mass structure of a strong lens galaxy at $z=0.35$, taking advantage of the milli-arcsecond angular resolution of very long baseline interferometric (VLBI) observations. In the first analysis of its kind at this resolution, we jointly infer the lens model parameters and the pixellated radio source surface brightness. We consider several lens models of increasing complexity, starting from a simple elliptical power-law density profile. We extend this model to include angular multipole structures, a separate stellar mass component, additional nearby field galaxies, and/or a generic external potential. We compare these models using their relative Bayesian log-evidence (Bayes factor). We find strong evidence for angular structure in the lens; our best model is comprised of a power-law profile plus multipole perturbations and external potential, with a Bayes factor of $+14984$ relative to the elliptical power-law model. It is noteworthy that the elliptical power-law mass distribution is a remarkably good fit on its own, with additional model complexity correcting the deflection angles only at the $\sim 5$ mas level. We also consider the effects of added complexity in the lens model on time-delay cosmography and flux-ratio analyses. We find that an overly simplistic power-law ellipsoid lens model can bias the measurement of $H_0$ by $\sim 3$ per cent and mimic flux ratio anomalies of $\sim 8$ per cent. Our results demonstrate the power of high-resolution VLBI observations to provide strong constraints on the inner density profiles of lens galaxies.