Is the NUTS algorithm correct?

2020/05/04 by J. M. Sanz-Serna, Sanz-Serna, J. M.
Computer Science · Mathematics · #Computation (stat.CO) #FOS: Computer and information sciences #FOS: Mathematics #Numerical Analysis (math.NA) #cs.NA #math.NA #stat.CO

paper · pdf · doi:10.48550/arxiv.2005.01336

Some statements in the paper are misleading. It is possible to think of NUTS at not being a slice/Gibbs sampler and, with an alternative interpretation, it may be be possible to prove that the algorithm is correct. In addition the experiment reported in Figure 2 should have had many initial states drawn from the target rather than using a single value

arxiv created 2020/05/07 · arxiv updated 2020/05/08

Abstract

This paper is devoted to investigate whether the popular No U-turn (NUTS) sampling algorithm is correct, i.e. whether the target probability distribution is exactly conserved by the algorithm. It turns out that one of the Gibbs substeps used in the algorithm cannot always be guaranteed to be correct.

Related

• A Theoretical Comparison of No-U-Turn Sampler Variants: Necessary and Sufficient Convergence Conditions and Mixing Time Analysis under Gaussian Targets
• Mixing of the No-U-Turn Sampler and the Geometry of Gaussian Concentration
• Incorporating Local Step-Size Adaptivity into the No-U-Turn Sampler using Gibbs Self Tuning
• SpreadNUTS -- Moderate Dynamic Extension of Paths for No-U-Turn Sampling & Partitioning Visited Regions
• The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo