Efficient Bayesian Sampling with Langevin Birth-Death Dynamics
By: Alex Leviyev, Francesco Iacovelli, Aaron Zimmerman
Potential Business Impact:
Helps find hidden signals in noisy data.
Bayesian inference plays a central role in scientific and engineering applications by enabling principled reasoning under uncertainty. However, sampling from generic probability distributions remains a computationally demanding task. This difficulty is compounded when the distributions are ill-conditioned, multi-modal, or supported on topologically non-Euclidean spaces. Motivated by challenges in gravitational wave parameter estimation, we propose simulating a Langevin diffusion augmented with a birth-death process. The dynamics are rescaled with a simple preconditioner, and generalized to apply to the product spaces of a hypercube and hypertorus. Our method is first-order and embarrassingly parallel with respect to model evaluations, making it well-suited for algorithmic differentiation and modern hardware accelerators. We validate the algorithm on a suite of toy problems and successfully apply it to recover the parameters of GW150914 -- the first observed binary black hole merger. This approach addresses key limitations of traditional sampling methods, and introduces a template that can be used to design robust samplers in the future.
Similar Papers
An Approximate Maximum Likelihood Estimator for Discretely Observed Linear Birth-and-Death Processes
Computation
Quickly estimates cell growth from messy data.
An Approximate Maximum Likelihood Estimator for Discretely Observed Linear Birth-and-Death Processes
Computation
Quickly estimates cell growth from messy data.
Gaussian approximations for fast Bayesian inference of partially observed branching processes with applications to epidemiology
Methodology
Makes tracking disease spread much faster.