Edward Porter (APC, Paris) : DeepHMC: a Hamiltonian Monte Carlo algorithm for fast BNS inference

The Hamiltonian Monte Carlo (HMC) algorithm is a non-random walker sampler that converges D times faster than
standard MCMC-type samplers (where D is the dimensionality of the parameter space). This is achieved by replacing
random jump proposals in parameter space by simulated Hamiltonian trajectories. While more efficient, the HMC
suffers from a serious computational bottleneck which restricts more common usage: at each point in the Hamiltonian
trajectory, the calculation of the D-dimensional gradients of the target distribution are required. This is expensive for
GW inference as it requires multple waveform generation per step. DeepHMC avoids this bottleneck by using a gradient
model based on a deep neural network to replace the waveform calculations. We will show that when applied to GW170817,
we can reduce the time to extract 10,000 statistically independent samples from weeks, as seen with current LVK algorithms,
to approximately two hours.