Q. Why not embed discrete parameters so that the resulting surrogate density function is smooth?

A. This is only possible in very special settings. Let’s say we have a target distribution \pi(\theta, n), where ‘\theta’ is continuous and ‘n’ is discrete. To construct a surrogate smooth density, we would need to somehow smoothly interpolate a collection of functions f_n(\theta) = \pi(\theta, n) for n = 1, 2, …. It is not clear to us how we can achieve this in a general and tractable way.

Q. How to generalize the algorithm to a more complex parameter space?

A. We provide a clear solution to dealing with a discontinuous target density defined on a continuous parameter space. We agree, however, that there remains the question of whether and how a more complex parameter space can be embedded into a continuous space. This certainly deserves a further investigation. For example, a binary tree can be embedded in to an interval [0,1] through a dyadic expansion of a real number.

Q. Physical intuition of discontinuous Hamiltonian dynamics is not clear from a theory of differential measure-valued equation and selection principle.

A. Hamiltonian dynamics with a discontinuous potential energy has long been used by physicists as a natural model for some physical phenomena (also known as “impulsive systems”). The main difference from a smooth system is that a gradient become a “delta function” at the discontinuity, causing an instantaneous “push” toward the direction of lower potential energy. A theory of differential measure-valued equation / inclusion and selection principle is only a mathematical formalization of such physical systems.

Q. (A special case of) DHMC looks like taking multiple Gibbs steps?

A. The crucial difference from Metropolis-within-Gibbs is the presence of momentum in DHMC, which helps guide a Markov chain toward a high density region.

The effect of momentum is evident in the Jolly-Seber example of Section 5.1, where DHMC shows 60-fold efficiency improvement over a sampler “NUTS-Gibbs” based on conditional updates. Also, a direct comparison of DHMC and Metropolis-within-Gibbs can be found in Section S4.1 where DHMC, thanks to the momentum, is about 7 times more efficient than Metropolis-within-Gibbs (with optimal proposal variances).

Q. Unlike HMC, DHMC does not seem to use structural information about the parameter space and local information about the target density?

It does. After all, other than the use of Laplace momentum and discontinuity in the target density, DHMC is based on the same principle as HMC — simulating Hamiltonian dynamics to generate a proposal.

The confusion is perhaps due to the fact that the coordinate-wise integrator of DHMC does not require gradients. The gradient of the log density — which may be a “delta” function at discontinuities — plays a clear role if you look at Hamilton’s equations Eq (10) corresponding to a Laplace momentum. It’s just that, thanks to a property of a Laplace momentum and conservation of energy principle, we can approximate the exact dynamics without ever computing the gradient. This is in fact a remarkable property of a Laplace momentum and our coordinate-wise integrator.

]]>In the end, HMC takes structural information about the parameter space and local information about the target density and turns it into an algorithm. I didn’t think this algorithm did either. Dinh et al’s paper (ref’d within) is a better attempt at this, but still is only a step towards the goal of “if your discrete parameter has structure x, here’s how you design an algorithm that takes that into account”. This paper is more “here’s a magic transformation that gives some ground on this problem”, which is less generalisable.

]]>