acceptodds
Under review as a conference paper at ICLR 2027

Reflected Proximal Dynamics for Accelerated High-Accuracy Log-Concave Sampling

Abstract

We study sampling from on , where is -strongly convex and -smooth with condition number . For this problem, proximal sampler preserves the target exactly and recovers overdamped Langevin dynamics in the small-step limit, but its Gaussian resampling step discards directional information. Underdamped Langevin dynamics retains such information through momentum to accelerate exploration, motivating a proximal construction that combines target preservation with directional persistence. We introduce *reflected proximal dynamics*, which combines proximal sampling with an exact reflection of the auxiliary variable, conditional kinetic motion, and controlled refreshment. We develop two realizations: a flux-resampled underdamped half-turn and a persistent-momentum quarter-turn. From a warm start with bounded order-two Rényi divergence, both methods achieve total variation error using expected oracle queries. We further remove the supplied warm-start assumption by combining either sampler with existing posterior samplers in a coarse-scale proximal initialization stage and optimizing the proximal step size. Accounting for both initialization and sampling yields expected oracle queries. These results identify reflected proximal samplers as a promising framework for accelerated high-accuracy log-concave sampling.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Sampling, Markov Chain Monte Carlo, Reflected Proximal Dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HNMdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces reflected proximal dynamics for accelerated high-accuracy log-concave sampling. By combining exact Gaussian reflection, conditional kinetic transport, and controlled refreshment, reflected proximal dynamics accelerate the outer chain. From a constant-D2D_2 warm start, both realizations use O~(κ1/2d1/4)\tilde O( κ^{1/2} d^{1/4}) expected oracle queries, while the persistent construction makes only a polylogarithmic expected number of approximate-RGO calls. Including localization and warm-up yields the fully charged O~(κ6/7d1/4)\tilde O (κ^{6/7} d^{1/4}) bound. Similar results can be found in a concurrent work Chen et al. (2026b).

Critical strengths

Excellent conditional number and dimension dependence.

Critical weaknesses
  1. All theoretical results are obtained in a log-concave setting.
  2. Can numerical experiments support your theoretical findings?
Questions for authors
  1. Can you give more details on extending the macroscopic estimates from strong convexity to functional inequalities?
  2. Can you provide some numerical tests to confirm your theoretical findings?
Reviewer_JhAgOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies high-accuracy sampling from smooth, strongly log-concave distributions through reflected proximal dynamics. It develops two kinetic realizations, provides warm-start convergence guarantees, and extends the analysis to include initialization costs. Overall, this is a substantial, carefully developed theoretical contribution with particularly clear writing and extensive supporting intuition.

Critical strengths
  1. Well-motivated algorithmic design. The paper combines the target-preserving structure of proximal sampling with directional persistence. Reflection, conditional kinetic motion, and occasional refreshment form a coherent construction, with clear explanations of their respective roles.

  2. A unified framework with two concrete realizations. The flux-resampled half-turn and persistent-momentum quarter-turn illustrate different ways to incorporate kinetic motion into the framework. This gives the work substance beyond an abstract algorithmic proposal.

  3. Explicit theoretical guarantees. Under strong convexity and smoothness assumptions, the paper establishes high-accuracy total variation guarantees and makes the dependence on dimension, condition number, and target accuracy explicit. In particular, it obtains an expected oracle-query bound of O~(κd1/4)\widetilde{O}(\sqrt{\kappa}d^{1/4}) from a constant-D2D_2 warm start.

  4. Careful treatment of initialization and oracle costs. The analysis extends beyond a supplied warm start to include localization, warm-up, and final sampling. It also accounts for internal work associated with approximate conditional sampling, optimization, and event simulation, making the complexity statements more informative.

  5. Clear writing and substantial intuition. This is a technically demanding paper, but the authors make considerable efforts to support the reader through intuitive explanations, illustrations, and a one-dimensional example. These explanations help connect the abstract framework to the concrete algorithmic operations.

  6. Detailed discussion and a highly polished presentation. The treatment of algorithmic choices, assumptions, implementation, and initialization is thoughtful and thorough. The manuscript feels carefully developed, with a high degree of completeness and attention to detail.

Critical weaknesses
  1. The practical significance of reducing conditional-sampling calls remains insufficiently demonstrated. At the chosen proximal scale, both an approximate RGO call and the reference-point optimization used by the persistent version have O~(1)\widetilde{O}(1) oracle cost, while the overall query bounds remain of the same order. It is therefore unclear whether the constants, logarithmic factors, or rejection costs of conditional sampling create a substantial practical bottleneck. Although these operations are included in the theoretical accounting, a small reproducible experiment with concrete parameter choices would help establish whether the proposed changes yield observable gains after accounting for optimization and event-simulation overhead.

  2. The incremental advantage over existing methods with comparable rates could be communicated more directly. The paper acknowledges that concurrent PBPS achieves the same warm-start query complexity. However, for a reader outside this specific area, it remains difficult to identify precisely which limitations of existing approaches the new framework resolves, and how the contributions of the new dynamics and initialization strategy should be distinguished. A more concentrated comparison under matched assumptions and initialization conditions would make the framework's advantages easier to assess.

Questions for authors

Given the same asymptotic query complexity, when do fewer RGO calls yield practical savings over the flux half-turn or PBPS? Could a small reproducible experiment demonstrate this, including optimization and event-simulation overhead?

In Section 5.3, what quantitative control on the Gaussian reference ensures that replacing the exact conditional reference retains the accelerated convergence rate, given that invariance alone does not imply fast mixing?

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

The paper studies high-accuracy sampling from smooth, strongly log-concave distributions using reflected proximal dynamics. It introduces two constructions: a flux-resampled half-turn and a persistent quarter-turn. From an initial distribution with bounded order-two Rényi divergence, both achieve O~(κ d1/4)\widetilde O(\sqrt{\kappa}\,d^{1/4}) expected oracle queries. The persistent construction also requires only polylogarithmically many approximate-RGO calls in expectation. The paper accounts for localization and warm-up to remove the need for a supplied warm start.

Critical strengths
  1. The conditional constructions require substantial technical work: they must preserve the target distribution while retaining enough information across iterations to support acceleration. The analysis addresses invariance, contraction, and implementation costs. The persistent quarter-turn, together with its Gaussian reference construction, gives a provable reduction in RGO calls. Including initialization costs also makes the complexity guarantees more informative.

  2. The convergence analysis of the persistent variant is another strength. It tracks how information moves between the auxiliary variable, conditional position, and retained momentum, and establishes contraction through a resolvent argument. This provides additional technical content beyond the half-turn analysis shared with PBPS.

Critical weaknesses
  1. My main concern is the contribution relative to concurrent PBPS. The outer reflection and half-turn acceleration mechanism, combined with occasional independent refreshment, also appears in PBPS, which achieves the same warm-start O~(κ d1/4)\widetilde O(\sqrt{\kappa}\,d^{1/4}) oracle complexity. The overlap extends to the flux analysis: both papers use covariance-deficit-weighted half-turn estimates and a modified L2L^2 energy argument. The central acceleration mechanism and the resulting dimension and condition-number exponents therefore do not distinguish this submission from PBPS.

  2. The flux construction uses different conditional dynamics, but the paper does not establish a stronger complexity guarantee or demonstrate a practical benefit from this change. The persistent construction provides a clear reduction in RGO calls, although the leading polynomial dependence of the total oracle complexity remains unchanged. I would like to understand when this reduction translates into a meaningful computational advantage after the cost of reference construction and other operations is included. Without experiments, the practical merits of either construction remain uncertain.

  3. My concern about presentation is mainly one of positioning. The paper should explain more precisely what is shared with PBPS, what requires new ideas, and what additional benefit those ideas deliver. The initialization analysis and reduced RGO count are relevant contributions, but their significance relative to PBPS needs a stronger case.

Given the substantial overlap and the limited evidence of additional advantages, I lean toward Weak Reject.

Questions for authors
  1. What theoretical or practical advantage does the flux-resampled half-turn offer over PBPS, given their matching leading-order complexity bounds?

  2. What computational benefit does the persistent construction’s reduction in RGO calls provide once reference construction and other costs are included? Which parts of this reduction require the persistent quarter-turn, and could a similar reference replacement be used in PBPS?

  3. Which parts of the initialization guarantee depend on the proposed conditional dynamics? Could the same initialization strategy be combined with PBPS as the final sampling stage to obtain a comparable bound?

  4. Could the authors provide focused synthetic comparisons against PBPS under matched initialization and accuracy criteria, reporting both wall-clock time and total oracle evaluations? Such experiments would help establish whether the proposed constructions offer practical advantages.

Comments on the paper

1 comment
Area_Chair_2Y96OpenReview ↗

Reviewer-Author Discussion Phase is Ending

Dear Reviewers,

Thank you for your service in the review process. Now that we are approaching the deadline of the author-reviewer discussion period. Please take time to read authors' responses and let them know if it addresses your concerns. Thank you.

Best,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.