acceptodds
Under review as a conference paper at ICLR 2027

Ambient integration rules for Riemannian manifold Hamiltonian Monte Carlo

Abstract

Riemannian manifold Hamiltonian Monte Carlo (RMHMC) adapts the metric of Hamiltonian Monte Carlo to the local geometry of the posterior, but every integration step requires the solution of an implicit system that contains derivatives of the metric. We consider target posterior distributions of hierarchical models, where the joint is given by a diffeomorphic map , where is a standard multivariate normal distributed random variable. We use the fact that conditioning on restricts to a submanifold with chart to derive integrators of RMHMC that take paths along secants through ambient space. To do this we consider variational integration methods and rephrase their kinetic term in ambient coordinates which makes them invariant under reparamterisation of the manifold. We show that each implicit integration method allows for an analogous secant version. We show that within our framework the solution of the implicit RATTLE integrator can be derived as a special case and is identical to the Generalized Leapfrog integrator in ambient space. We compare eight solvers for four implicit integrators, including the different solvers for RATTLE, on four test problems, including Neal's funnel, a supernova dust emission model and a hierarchical Gaussian process. Our experiments show that ambient secant formulations are superior in both robustness to the chosen step size as well as solver cost, where we observe that ambient formulations can allow for a 10x larger step-size. We then compare the resulting implicit MCMC schemes to the explicit LMC scheme and find that when matched for trajectory length, secant based methods can be competitive and can even improve over LMC especially for higher dimensional problems, even when using an efficient implementation of LMC that makes use of the geometric structure.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
RMHMC, implicit midpoint, RATTLE, generalized leapfrog, geometric integration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_WpEkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper, the authors proposed a secant version of integrator for several numerical integrators utilized in Riemannian Hamiltonian Monte Carlo methods. The approach is motivated by applying variational integration method. The new integrator scheme are tested against three experiments for eight different solver. The empirical studies produce effective outcome.

Critical strengths

The problem under consideration is an important one. The idea proposed in the paper is novel.

Critical weaknesses

There is no solid theoretical results for analysis of validity and complexity of the proposed new method. The numerical experiments are limited, their outcome are not as conclusive as suggested in abstract and conclusion.

Questions for authors

Q1, The term hierarchical model is used several time, but it is not really explained. Q2. The map ϕ(q)\phi(q) and ψq\psi_q have not been properly defined. Related the chart map is given in a unconventional way without explanation. Q3. Line 156 the sentence starts wit "With this setup..." is not very well explained, including why UAU^A is the representation of UU with constraint. Q4, please confirm that the variable uu in (15) also take the value of q˙A{\dot q}^A when applied to equation (9). Q5. The parameters set for the experiments seem to be limited and arbitrary, please explain the logic behind the set up.

1 comment

Thank you for replying to my comments and questions. Really appreciate the addition to Appendix B. I would like to adjust my score.

Reviewer_rKzjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

RMHMC is a standard technique for solving probabilistic problems by adapting the dynamics to the local geometry of the target distribution, but finding this is expensive and thus is typically approximated. This approach uses secants over a diffeomorphic transform to form an ambient space for integrators. The paper shows that various integrators can be placed into the invariant ambient space, and in particular, that RATTLE is a special case of the generalized leapfrog integrator. Results demonstrate that ambient secant reformulation can be faster and more robust. Interestingly, it shows that the gap with explicit LMC can be closed.

Critical strengths
  • RMHMC is a tricky, but important, problem with many different solvers, tricks and integrators. This approach is conceptually simple and clean - use secants in the ambient space where varying curvature is no longer a dominant issue and then step sizes do not need to be carefully chosen to deal with highly curved regions of the manifold.
  • The claim that RATTLE is a special case of generalized leapfrog is also a useful contribution.
  • The results show ESS/sec is higher and can outperform LMC.
Critical weaknesses
  • Experimental results could be more clearly presented with more comprehensive comparison between "standard" RMHMC and the proposed ambient secant approach. This is done in Fig 1, but not really through the remainder of the paper.
Questions for authors
  • AI results and proofs are nice, but some deeper discussion and analysis would be helpful - when would this work well? What are the key limitations? Should we replace LMC with ambient secant RMHMC?
  • Can you show perhaps a high level conceptual figure or graphic showing how your ambient secant approach conceptually works over the manifold?
Reviewer_tLTzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The present submission tackles a difficult problem, which is the problem of sampling from general distributions with varying curvature. For this kind of distribution, many existing methods experience significant slowdown. Here, the focus is on Hamiltonian Monte Carlo, where a varying Riemannian metric is introduced in order to deal with the issue of varying curvature.

Section 2 briefly explains the main idea of Hamiltonian Monte Carlo (by now, a very well-known concept in the field of MCMC sampling), as well as some standard integrators (generalised leapfrog, LMC, and RATTLE).

Section 3 explains rather clearly the idea of ambient integrators and introduces more integrators which are finally compared in Section 4.

In Section 4, stability and rate of convergence, as well as complexity of single iterations for each solver are compared on four "difficult" target distributions. These experiments are quite detailed and show the advantage of the solver proposed by the authors.

I have to say I am not an expert on HMC and did find some parts of this work too dense and specialised. I do think the authors have really tried to make their presentation accessible, but some parts are just too difficult for non-specialists.

Critical strengths
  • the paper clearly explains its contribution and where this contribution stands in comparison to existing litterature

  • the numerical experiments are detailed and carefully designed and organised

  • the results of these experiments are discussed in a very detailed way, leading up to the final conclusion

  • the paper deals with a clearly identified practical difficulty

Critical weaknesses
  • the conclusions are only experimental and not supported by mathematical reasoning

  • the paper is quite difficult to follow for non-experts

Questions for authors

clarification request : it was hard for me to follow Section 3. I would like to ask you to introduce the methods discussed in this section using a "definition" format that clearly separates them from one another

Comments on the paper

2 comments
Area_Chair_K7gqOpenReview ↗

Please engage in the discussion

Reviewers: if you have not responded to the authors' reply, please read it and engage in the discussion. Thanks!

We would like to thank all reviewers for their feedback that allowed us to improve our work. We also appreciate the following engangement in the discussions. Independently of the outcome of this process, we believe the article is now in a significantly better shape. Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.