26th March 2021, 3:00 pm – 4:00 pm
The efficiency of the Hamiltonian Monte Carlo method, in all its variants, depends crucially on the integrator used to simulate the Hamiltonian dynamics. The Stormer/Verlet/leapfrog algorithm is currently the integrator of choice but better alternatives exist. In the talk I will discuss the probabilistic issues that play a key role in choosing or designing integrators for HMC sampling. I will also show that, after taking those issues into account, it is possible to reduce the computational cost of HMC by a very substantial factor.