Résumé / Abstract Seminaire_GReCO

"A Hamiltonian Monte Carlo for Bayesian Inference of Supermassive Black Hole Binaries"

Ed Porter
Laboratoire AstroParticule et Comologie, Université Paris Diderot (Paris, France)

Markov Chain Monte Carlo algorithms have become a popular tool for Bayesian inference in GW astronomy. While successful, these methods are random walk algorithms, with indeterminate convergence timescales. The Hamiltonian Monte Carlo algorithm works by eliminating this randomness from the problem. By equating the black hole source parameters to state space variables, and by introducing a set of canonical momenta, one then solves Hamilton’s equations of motion to evolve trajectories in state space. While superior to standard MCMC methods, the Hamiltonian Monte Carlo is not very widely used. This stems from the fact that each trajectory in state space requires multiple calculations of the gradient of the log likelihood. In most physical problems, there is no closed-form solution for the likelihood, meaning that both it and its gradients need to be calculated numerically. This leads to a bottleneck that renders the algorithm too computationally expensive for practical use. In this presentation, we present a solution to this problem and demonstrate how a Hamiltonian Monte Carlo can be used for parameter estimation for supermassive black hole binaries.

lundi 17 février 2014 - 11:00
Salle des séminaires Évry Schatzman,
Institut d'Astrophysique de Paris

Page web du séminaire / Seminar's webpage