Step size for stable Hamiltonian trajectories

tags
Hamiltonian Monte Carlo

Let us consider the following Hamiltonian (corresponds to sampling from a univariate Gaussian with standard deviation σ):

H(q,p)=q2/2σ+p2/2

A leapfrog step with step size ϵ is a linear mapping between (q(t),p(t)) and (q(t+ϵ),p(t+ϵ)) with the following transition matrix:

(1ϵ2/2σ2ϵϵ/σ2+ϵ3/4σ41ϵ2/2σ2)

Whether this leads to a stable trajectory or diverges depends on the magnitude of thge eigenvalues:

Missing \begin{equation*} or extra \end{equation*}

When ϵ/σ>2 these eigenvalues are real and one will have absolute value greated than one. Trajectories computed with ϵ<2σ are thus stable. For multi-dimensional problems, the stability will be determined by the width of the distribution in the most constrained region. Stability for general quadratic hamiltonian K(p)=pTM1p

References

  • Neal Radford, MCMC using Hamiltonian dynamics (p 22)

Links to this note