acceptodds
Under review as a conference paper at ICLR 2027

Stability of Underdamped Langevin Dynamics in Distributional Minimax Optimization

Abstract

Underdamped Langevin methods have received increasing attention in mean-field optimization, motivated by their convergence and discretization properties. In this paper, we show that for distributional minimax optimization, coupling the two players can destabilize underdamped Langevin dynamics. For a bilinear game, we derive an exact friction–coupling threshold separating exponentially stable and unstable underdamped dynamics, while the overdamped counterpart remains stable at every coupling strength. We further show that this instability persists for a bounded smooth interaction. To avoid simultaneous cross-player coupling, we introduce an anchored scheme that freezes the current strategies while running underdamped Langevin dynamics toward their Gibbs responses, and updates the strategies between these runs. For a class of convex-concave functionals of probability laws in the mean-field setting, we establish last-iterate convergence of the proposed scheme, with an decay of the Nikaido–Isoda error. Lastly, we provide a finite-particle, discrete-time implementation with a last-iterate convergence guarantee. More broadly, our results provide a principled framework for understanding and stabilizing underdamped dynamics in distributional minimax optimization.

Primary area:
learning theory
Keywords:
Underdamped Langevin dynamics, min-max optimization, mean-field, stability, particle algorithms

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_L5UMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors extend the mean-field underdamped Langevin dynamics to the distributional minimax optimization setting. They show (1) while the overdamped (mean-field) system converges, the underdamped system does not generally converge, (2) they extend prior analysis using an anchored best-response framework to the underdamped setting, and (3) they provide analysis for the finite-particle, discrete-time setting.

Critical strengths
  1. The problem setting is novel, and the paper is generally well-written.
Critical weaknesses
  1. The work as written, feels kind of incremental; it seems the main ideas for the algorithm is a suitable combination of Fu & Wilson, (2024) and Kim et al., (2024). The authors should highlight the technical novelties and/or the theoretical innovations compared to prior work.
  2. One place that seems to improve upon prior work is that the authors seem to provide propagation of chaos/quantify the bias of the particle-discretization in Theorem 4 in the appendix (this seems to be novel for the anchored best-response type algorithm), but the rates suffer from the curse of dimensionality.
Questions for authors
  1. The problem in eq (1) is strictly a subcase of the Kim et al. (2024); they allow arbitrary reference measures. Can the authors also analyze this setting? Can the authors discuss where their analysis requires the Gaussian reference measures is used? This seems to be a limitation of their work compared to this prior work.
  2. In Section 5.4, why do we need to take uniformly selected particles at outer steps? This seems to be the cause of the curse of dimensionality issue. Also, I'm curious why there is not the usual propagation of chaos term, e.g. see the first term in Corollary 1, [N24].

Other comments:

  • The authors do not motivate why we should study the underdamped dynamics, instead of just using the overdamped dynamics.

[N24] Nitanda. Improved Particle Approximation Error for Mean Field Neural Networks. NeurIPS 2024.

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

This paper studies the stability of underdamped mean-field Langevin dynamics in distributional minimax optimization. For bilinear games, the authors derive an exact friction–coupling threshold separating stability from instability and show that instability can also arise under bounded smooth interactions. They then analyze an underdamped anchored best-response scheme that freezes the strategies during finite-time Gibbs-response sampling. The paper establishes an O(1/K) Nikaido–Isoda error rate. Numerical experiments illustrate the stability transition and compare different Langevin samplers.

Critical strengths
  1. Meaningful research question. I think investigating whether momentum-based Langevin dynamics remain stable when two distributional players evolve simultaneously is a worthwhile problem. The contrast between underdamped instability and overdamped stability highlights an important challenge caused by cross-player coupling. This is a positive factor in my overall assessment.
  2. Relatively complete theoretical results. The theoretical results in this paper is comprehensive, which is quite important for theory papers. This provides a coherent account of both the instability mechanism and a possible remedy.
Critical weaknesses
  1. Novelty could be clearer. Anchored best-response methods for distributional minimax optimization and underdamped Langevin sampling have both been investigated previously. While the coupling-dependent stability analysis is valuable, it is not sufficiently clear which parts of the proposed algorithm and convergence proofs go beyond adapting existing frameworks. A sharper comparison with closely related papers would better establish the paper’s distinct contribution. This uncertainty limits the strength of my recommendation.
  2. Limited numerical evaluation. The current experiments are too simple as practical support. They provide limited evidence for the method’s advantages in higher-dimensional or more realistic minimax applications. This is not fatal for a theoretical paper, but it reduces the support for the broader motivation.
Questions for authors
  1. Clarifying novelty. Could the authors explicitly distinguish the proposed underdamped anchored scheme and its analysis from existing work? In particular, which technical steps or guarantees are fundamentally new?
  2. Broader numerical evidence. Could the authors add experiments on higher-dimensional games or more realistic applications (e.g., a mean-field GAN objective)?
Reviewer_NbUHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers the convergence of underdampened Langevin dynamics in minimax optimization. The paper derives a sufficient condition for stability of the dynamics and shows that without this condition, under some special cases the dynamics are not stable. The paper then constructs a continuous-time underdampened Langevin algorithm and analyzes its convergence properties. Supporting simulations are shown.

Critical strengths

-The results seem non-trivial, answer relevant problems, and are clearly presented

-The authors generally write the paper clearly and format everything in an understandable way (a few criticisms towards this are below).

Critical weaknesses

-The notation of the paper is quite heavy making it harder to follow in places. This however is partially due to the nature of the problem this paper studies.

-The presentation of the paper could also be improved. It would be nice if the authors first present the Theorem and then the supporting Lemmas; the writing seems to be done in opposite order

-Another weakness is about the related work and positioning in the literature. What is the motivation to study this topic, sampling in mI find it a bit surprising that these results have not been established yet in the literature, although I am not experienced in this particular topic. More on this is in my questions.

Questions for authors

Have there been previous works in convergence of Langevin dynamics in distributional minimax optimization? what are the previous results in this space, and how does the main convergence result Thm 3 improve on these prior results?

Comments on the paper

1 comment
Area_Chair_8fabOpenReview ↗

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.