Stein Variational Inference for Gravitational Wave Likelihood Estimation

The planned upgrades to the gravitational wave detectors promises a vastly improved detection rate for binary inspirals; however, a full parameter estimation analysis of such a signal can take days or weeks. This presents a bottleneck in the performance of a gravitational wave analysis pipeline. Improvements in the speed and efficiency of parameter estimation would have many potential benefits, e.g. facilitating the use of more sophisticated---and computationally expensive---signal models. Current parameter estimation techniques (such as nested sampling) rely on stochastic sampling algorithms. We propose a new optimization based algorithm which relies on minimizing the Kullback---Leibler divergence between an approximation of, and a reference posterior. This procedure yields a transport map that can be used to produce an arbitrary sized empirical sampling. We test whether this indeed can lead to a faster, yet still accurate alternative.