Orateur
Description
LISA is expected to observe tens of thousands of resolved Galactic Binaries (GBs), requiring data analysis pipelines to handle physically motivated populations and perform rigorous astrophysical inference. To prepare for future data streams, global fit pipelines must be tested against realistic scenarios. The most recent LISA Data Challenge, "Mojito," provides the ideal testbed, introducing full global fit tasks with unprecedented realism, including time-varying noise, realistic orbits, and detailed populations of interacting and detached GBs. In this work, we present the most recent advancements to the GPU-accelerated global fit pipeline, Erebor. More specifically, we focus on the analysis of the resolved population of GBs and demonstrate a methodology to perform rigorous astrophysical inference on the entire population of GBs.
To extract the resolved GBs, the pipeline now utilizes GPU-enhanced and computationally efficient waveform and likelihood computations using the recently developed TDI-on-the-fly approach (Littenberg & Cornish 2025). Furthermore, to navigate the vast parameter space associated with fitting tens of thousands of binaries, we integrate machine learning-enhanced proposals utilizing normalizing flows, which significantly accelerate the sampling convergence of the resolved GB block. Using this updated GB block, we present the results of preliminary runs on the Mojito data challenge and provide an outlook on future developments for the Erebor pipeline.
Crucially, this integration drastically accelerates sampling convergence, making it computationally tractable to perform joint population inference. This population inference must be intrinsic to the global fit, because traditional post-processing methods used in LVK data analysis fail due to the transdimensional nature of extracting an unknown number of sources. To address this, we present a novel astrophysical inference framework that directly operates within the global fit and performs a Bayesian model selection algorithm using Reversible Jump MCMC. This algorithm dynamically jumps between competing population synthesis models, altering the overarching astrophysical priors and the Poisson distributions governing the expected number of resolved systems. This approach rigorously quantifies and improves our understanding of LISA’s sensitivity to specific astrophysical processes and explores physical degeneracies between astrophysical hyperparameters.