Unbiased constrained sampling with Self-Concordant Barrier Hamiltonian Monte Carlo

In this paper, we propose Barrier Hamiltonian Monte Carlo (BHMC), a version of the HMC algorithm which aims at sampling from a Gibbs distribution $\pi$ on a manifold $\mathrm{M}$, endowed with a Hessian metric $\mathfrak{g}$ derived from a self-concordant barrier. Our method relies on Hamiltonian dynamics which comprises $\mathfrak{g}$. Therefore, it incorporates the constraints defining $\mathrm{M}$ and is able to exploit its underlying geometry. However, the corresponding Hamiltonian dynamics is defined via non separable Ordinary Differential Equations (ODEs) in contrast to the Euclidean case. It implies unavoidable bias in existing generalization of HMC to Riemannian manifolds. In this paper, we propose a new filter step, called"involution checking step", to address this problem. This step is implemented in two versions of BHMC, coined continuous BHMC (c-BHMC) and numerical BHMC (n-BHMC) respectively. Our main results establish that these two new algorithms generate reversible Markov chains with respect to $\pi$ and do not suffer from any bias in comparison to previous implementations. Our conclusions are supported by numerical experiments where we consider target distributions defined on polytopes.

Paper

Similar papers

Peer review

Reviewer SWNi6/10 · confidence 2/52023-06-25

Summary

This paper propose a computational framework for sampling via Hamiltonian Monte Carlo via self-concordant barrier functions for the purposes of constrained sampling on suitably nice Riemannian manifolds, where the Riemannian metric defines the barrier function. They provide computational and theoretical guarantees on their procedure, which hinges on an "involution checking step", making it more theoretically sound than other papers on this matter, in particular this makes the procedure unbiased. The authors supplement their theoretical findings with computational experiments in both real-world and synthetic data.

Strengths

Strong theoretical contributions relative to prior work --- kudos to the authors for fixing the discrepancies in Kook et al (2022a), and proposing their solution (note, I didn't read the proofs in this paper either, only skimmed, but the results appear reasonable. I thought the background section, etc, contained all the relevant information, and the related works section also provided ample context.

Weaknesses

I think the presentation is a bit suboptimal, namely the paper is very dense and hard to read in some parts. Otherwise, I really enjoyed this paper. I think the technical contributions appears to be somewhat minimal, though surely it took a lot of work to justify the "involution checking step", but again, not an expert in this area. I also understand that Kook et al. (2022b) provided some theoretical understanding of the algorithm from the previous paper --- would the authors comment a bit more on that? Do the theoretical results in that paper suffer from the same problems as in this one, or is everything remedied and this paper is concurrent with that one?

Questions

- What is this [x,y] or x : y notation in lines 101-102 (and other places)? I was not able to parse this... - I was also largely confused and unconvinced by the numerical experiments... In Table 1 it appears what the authors propose is about "as good" as CRHMC (as in, 50/50 chance it is better performing). I also cannot understand Fig 2 at all --- what does the y-axis mean, and what is the ground truth value the algorithms are trying to estimate? Is the goal to simply be as close to MALA as possible? Seems odd, but I'm far from an expert

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Yes

Reviewer iZM76/10 · confidence 3/52023-07-07

Summary

The paper proposes a new algorithm for sampling on a constrained space via a Hamiltonian Monte Carlo algorithm, which uses a Riemannian Manifold HMC algorithm, with the Hessian metric given by a self-concordant barrier. This generalizes and improves upon existing work in constrained sampling, while removing a source of bias by adding an “involution checking step”.

Strengths

The intuition for the algorithm, including the choice of self-concordant barrier, is well-informed and intuitive, although it has appeared in some form in Kook et al. (2022). Thus, I will discuss many of the strengths and weaknesses in relation to the prior work of Kook et al. (2022). This paper possesses two generalizations of Kook et al. Firstly, the primary problem setting in Kook et al. (Eq 11 of that paper) is less general than that in this paper. The proof for this method is also cleaner, and purports to avoid some technical issues present in the proof Kook et al. (I did not carefully verify this point). Furthermore, it avoids the problems of bias presented in Kook et al., which is carefully justified by a Theorem. The method is clearly unbiased when evaluated on the hypercube and simplex, by a margin that is usually many standard deviations. The broad details of the proof seem to be correct and well-written, although I did not check every detail.

Weaknesses

No rigorous non-asymptotic theory was presented, although this is also true of Kook et al., likely because of the complexity of these algorithms. It is difficult to draw firm conclusions from the real-world instances in the experiments, since the method does not show improvement in all/a clear majority of cases. The notation in this paper is a bit cumbersome, particularly in the Appendix. Nonetheless this is understandable given the complexity of the algorithm, and the authors did a good job in reducing the notational burden within the main text. To summarize, it was difficult to review all the details/proofs in this manuscript, given its length and technical nature. Nonetheless, I feel like it presents an important development within constrained sampling and merits publication. This informs my final score.

Questions

How difficult is it to establish any non-asymptotic convergence results in this context? What would be the expected rates, and would they differ (in the order of dependencies) with Kook et al.? As noted under “weaknesses”, the experiments on “real-world data” do not seem fully convincing. Can the authors elaborate further on why this would be the case, and if there was extensive tuning of the parameters involved? It seems from the Appendix that the parameters of both algorithms were not extensively tuned, which then begs the question whether this is really a fair comparison. It is curious that a (constrained) Gaussian distribution is chosen for the synthetic data, when Kook et al. consider a uniform distribution instead. Their choice allows for some simple tests of uniformity. Does considering a uniform distribution make any appreciable difference in the output? The discussion in Appendix J is quite useful and I think some more of these details could be also summarized in the main text. It is my personal preference, but I believe text in the paper should not be highlighted. Use boldface or perhaps underlining to emphasize the text in Ll. 192-193 and in Algorithm 1. L. 91 “motivates to” -> “motivates us to” L. 170 “the definition domain” -> “the domain” L. 213 “enables” -> “enables one/enables us” The authors should be careful about the formatting of some equations in the Appendix, since they exceed the column width.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

None beyond those raised in earlier sections.

Reviewer prii6/10 · confidence 4/52023-07-09

Summary

This paper introduces a "involution checking step" in Riemannian Hamiltonian Monte Carlo methods that ensures time reversibility of the resulting Markov chain. The numerical experiments demonstrate the competitiveness of the proposed algorithm with regard to the method proposed in Kook et al. (2022a). --- The rebuttal addressed my quetions. I have raised the rating to 6.

Strengths

1. The algorithms are novel. 2. The algorithms and analyses take practical numerical integrators into consideration. 3. Appendix J points out clearly the theoretical gaps in Kook (2022a).

Weaknesses

1. **Comparison with existing work.** The most relevant previous work seems to be Kook et al. (2022a) and Kook et al. (2022b). The second one appeared later and focused on theoretical analyses, but this submission only discusses about the gaps in the analysis of the first one. 2. **Gap between theory and practice.** The algorithm that actually ensures time reversibility is Algorithm 3, but the numerical experiments are conducted with Algorithm 1. 3. **Imprecise statement about existing work.** In Ln. 304, it is claimed that the convex constraint set considered by Kook et al. (2022a) takes a specific form. But that is simply a special case of the class of constraint sets in Kook et al. (2022a). 4. **Presentation.** The presentation involves many terminologies without explanations. The paper may not be readable to a layperson. 1. There are some highlights with unclear meanings. 2. What do the vertical axes mean in Figure 2? 5. **Typos.** 1. Ln. 115: Concordance -> self-concordace 2. Ln. 118: polynomial -> polynomial-time 3. What does the colon mean in (4)? 4. Ln. 287: a involution -> an involution 5. Some words in the reference list are not properly capitalized.

Questions

1. How does this work compare with Kook et al. (2022b)? 1. Please address the gap between theory and practice in the weaknesses block. 1. What do the highlights in the submitted files mean?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

This is a paper on theoretically rigorous algorithm design. The assumptions are explicitly stated.

Reviewer nmMp7/10 · confidence 3/52023-07-10

Summary

The authors resolve an existing bias with naive manifold versions of HMC by adding an involution checking step, and therefore leading the new HMC algorithm to be unbiased. I'm not super familiar with this problem, and I would like the authors to explain some of the technical details, as the proofs are quite involved for a conference review.

Strengths

The authors proposes a solution to an existing bias in manifold HMC, both justifying theoretically and proposing convincing empirical simulations.

Weaknesses

The technical details are quite involved and I believe could be explained better.

Questions

While I have many questions, I am still open to raising my score once these questions are adequately addressed. 1. When you say there's no guarantee that $G_h$ has a unique solution, do you mean that (1) there are cases with multiple solutions, or (2) there are cases with no solutions, or both? If possible, can you also provide an intuitive example where this failure occurs? 2. Can the authors explain to me the main intuitions behind why the naive implementations of Kook et al. (2022a) leads to a biased stationary measure? In particular, why does the Metropolis accept-reject step not fix this bias? 3. To add to the previous question, on an intuitive level, how does this involution check fix this issue? 4. It appears that this is not a problem when we can simulate the ideal Hamiltonian dynamics, is this related to the fact that the domain of flow is the entire cotangent bundle $T^* M$? 5. The proofs are quite involved. Can you explain to me on a high level (a) how does the proof work, and (b) where do the complicated calculations mostly arise from?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer nmMp2023-08-12

Response

Hi Everyone, Thank you for the response. Let me start by saying I think the 1D example is incredibly convincing, and this situation finally made sense for me for the first time. The rest of my questions are also addressed, albeit I would hope the authors consider rewriting some portions of the paper, and maybe add some more intuitive content in the extra page (if accepted). I will raise my score to 7.

Reviewer prii2023-08-14

Thanks

The rebuttal addressed my quetions. I have raised the rating to 6.

Reviewer iZM72023-08-20

Response

I thank the authors for their extremely detailed response. I appreciate the explanation on why a non-uniform example was chosen and I believe that this work makes an important contribution in terms of identifying a true gap in Kook et al. For me, one major drawback of this paper continues to be the lack of non-asymptotic convergence results, which would be really necessary to make this paper fully self-contained. As a result, I will maintain my current score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC