This recording is in the process of being subtitled. We aim to have edited captions available within 2 weeks of publishing.
15 November
Jesús María Sanz-Serna
(University Carlos III of Madrid) Numerical integrators for the
Hamiltonian Monte Carlo method
The Hamiltonian Monte Carlo
(aka Hybrid Monte Carlo) method is an extremely popular Markov chain algorithm
to obtain samples from a target probability distribution. The computational
cost of the algorithm mostly originates from the need to numerically integrate,
at each step of the chain, a Hamiltonian system of ordinary differential
equations. It is therefore important to carry out those integrations as
efficiently as possible. While at present the Stormer/Verlet integrator is the
integrator of choice, it is possible to construct special integrators tailored
to this task. These alternative integrators, while being as easy to implement
as Stormer/Verlet, allow for very substantial reductions of the work required
to obtain the samples. The talk will focus on the numerical analysis principles
used to construct the new integrators.