15.5 Divergent Transitions
The Hamiltonian Monte Carlo algorithms (HMC and NUTS) simulate the trajectory of a fictitious particle representing parameter values when subject to a potential energy field, the value of which at a point is the negative log posterior density (up to a constant that does not depend on location). Random momentum is imparted independently in each direction, by drawing from a standard normal distribution. The Hamiltonian is defined to be the sum of the potential energy and kinetic energy of the system. The key feature of the Hamiltonian is that it is conserved along the trajectory the particle moves.
In Stan, we use the leapfrog algorithm to simulate the path of a particle along the trajectory defined by the initial random momentum and the potential energy field. This is done by alternating updates of the position based on the momentum and the momentum based on the position. The momentum updates involve the potential energy and are applied along the gradient. This is essentially a stepwise (discretized) first-order approximation of the trajectory. Leimkuhler and Reich (2004) provide details and error analysis for the leapfrog algorithm.
A divergence arises when the simulated Hamiltonian trajectory departs from the true trajectory as measured by departure of the Hamiltonian value from its initial value. When this divergence is too high,24 the simulation has gone off the rails and cannot be trusted. The positions along the simulated trajectory after the Hamiltonian diverges will never be selected as the next draw of the MCMC algorithm, potentially reducing Hamiltonian Monte Carlo to a simple random walk and biasing estimates by not being able to thoroughly explore the posterior distribution. Betancourt (2016a) provides details of the theory, computation, and practical implications of divergent transitions in Hamiltonian Monte Carlo.
The Stan interfaces report divergences as warnings and provide ways to access which iterations encountered divergences. ShinyStan provides visualizations that highlight the starting point of divergent transitions to diagnose where the divergences arise in parameter space. A common location is in the neck of the funnel in a centered parameterization, an example of which is provided in the user’s guide.
If the posterior is highly curved, very small step sizes are required for this gradient-based simulation of the Hamiltonian to be accurate. When the step size is too large (relative to the curvature), the simulation diverges from the true Hamiltonian. This definition is imprecise in the same way that stiffness for a differential equation is imprecise; both are defined by the way they cause traditional stepwise algorithms to diverge from where they should be.
The primary cause of divergent transitions in Euclidean HMC (other than bugs in the code) is highly varying posterior curvature, for which small step sizes are too inefficient in some regions and diverge in other regions. If the step size is too small, the sampler becomes inefficient and halts before making a U-turn (hits the maximum tree depth in NUTS); if the step size is too large, the Hamiltonian simulation diverges.
Diagnosing and Eliminating Divergences
In some cases, simply lowering the initial step size and increasing the target acceptance rate will keep the step size small enough that sampling can proceed. In other cases, a reparameterization is required so that the posterior curvature is more manageable; see the funnel example in the user’s guide for an example.
Before reparameterization, it may be helpful to plot the posterior draws, highlighting the divergent transitions to see where they arise. This is marked as a divergent transition in the interfaces; for example, ShinyStan and RStan have special plotting facilities to highlight where divergent transitions arise.
References
Betancourt, Michael. 2016a. “Diagnosing Suboptimal Cotangent Disintegrations in Hamiltonian Monte Carlo.” arXiv 1604.00695. https://arxiv.org/abs/1604.00695.
Leimkuhler, Benedict, and Sebastian Reich. 2004. Simulating Hamiltonian Dynamics. Cambridge: Cambridge University Press.
The current default threshold is a factor of \(10^3\), whereas when the leapfrog integrator is working properly, the divergences will be around \(10^{-7}\) and do not compound due to the symplectic nature of the leapfrog integrator.↩