Discovering gravitational waveform distortions from lensing: a deep dive into GW231123

Abstract: Gravitational waves (GWs) are unique messengers as they travel through the Universe without alteration except for gravitational lensing. Their long wavelengths make them susceptible to diffraction by cosmic structures, providing an unprecedented opportunity to map dark matter substructures. Identifying lensed events requires the analysis of thousands to millions of simulated events to reach high statistical significances. This is computationally prohibitive with standard GW parameter estimation methods. We build on top of state-of-the-art neural posterior algorithms to accelerate the lensed inference from CPU days to minutes with DINGO-lensing. We showcase its capabilities by reanalyzing GW231123, the most promising lensed candidate so far, and find that its statistical significance cannot exceed 4$σ$. We observe that 8% of GW231123-like nonlensed simulations favor lensing, which could be explained by the self-similarity of short-duration signals. Still, 58% of GW231123-like lensed simulations have larger support for lensing, showing that higher detection statistics are possible. Although GW231123 exposes the challenges of claiming the first GW lensing detection, our deep-learning methods have demonstrated to be powerful enough to enable the upcoming discovery of lensed GWs.