Skip to content
DraconiticResearch
Search Sources API
194,161 studies31 sourcesUpdated 22 min ago
PreprintarXiv licencearXiv

Local Algebraic Stability of Wasserstein Gradient Descent-Ascent and Mean Field Langevin Dynamics without Spectral Gap

Abstract

We study the local stability of Wasserstein gradient descent-ascent (GDA) for entropy regularized two-player zero-sum games on noncompact domains. When one or both equilibrium measures fail to satisfy a Poincaré inequality, the spectral gap arguments underlying local exponential convergence are unavailable. we show that Wasserstein GDA converges algebraically to the mixed Nash equilibrium in chi-squared divergence, provided the initial relative density is a small perturbation in L2 with bounded algebraic moments. The proof has two main ingredients: a cancellation in a negative Sobolev energy estimate, and a bootstrap argument that keeps the moments bounded uniformly in time. For polynomial-tailed equilibria, we construct perturbations supported far out in the tails that survive over diffusive time scales, which gives algebraic lower bounds on the worst-case relaxation rate. We also establish analogous local algebraic stability and lower bound results for mean-field Langevin dynamics, when the equilibrium fails to satisfy a spectral gap.

Related

  1. A CONSTRAINED WASSERSTEIN GRADIENT FLOW FOR MEAN-FIELD LANGEVIN DYNAMICS
    Mohamed Alfaki Aboubacrine Assadek
    HALPreprintCC BY 4.0Full textAnalysis & PDEsMathematical physicsHALarXiv
  2. Theoretical guarantees for stochastic gradient Langevin dynamics
    Daniel Paulin, Peter A. Whalley
    arXivPreprintarXiv licenceFull textNumerical analysis & scientific computingProbabilityarXiv
  3. On the computational cost of Stochastic Gradient Langevin Dynamics
    Mateusz B. Majka, Tigran Nagapetyan, Łukasz Szpruch +2
    arXivPreprintarXiv licenceFull textNumerical analysis & scientific computingProbabilityarXiv
  4. Statistical and Privacy Guarantees for Deep M-Regression via Stochastic Gradient Langevin Dynamics
    Yeheng Ge, Jian Huang, Shuyuan Wu
    arXivPreprintarXiv licenceFull textStatisticsarXiv
  5. Initialization Stability and Convergence of Projected Langevin Dynamics
    Qi Feng, Nathan Sauldubois, Xin Zhang
    arXivPreprintCC BY 4.0Full textOptimisation & controlProbabilityarXiv
BibTeX
@misc{an2026local,
  title = {{Local Algebraic Stability of Wasserstein Gradient Descent-Ascent and Mean Field Langevin Dynamics without Spectral Gap}},
  author = {Jing An and Lihan Wang},
  howpublished = {arXiv},
  year = {2026},
  month = oct,
  doi = {10.48550/arxiv.2610.35528},
  eprint = {2610.35528},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35528},
}
RIS
TY  - UNPB
TI  - Local Algebraic Stability of Wasserstein Gradient Descent-Ascent and Mean Field Langevin Dynamics without Spectral Gap
AU  - Jing An
AU  - Lihan Wang
PY  - 2026
DA  - 2026/10/26
PB  - arXiv
DO  - 10.48550/arxiv.2610.35528
UR  - https://arxiv.org/abs/2610.35528
AB  - We study the local stability of Wasserstein gradient descent-ascent (GDA) for entropy regularized two-player zero-sum games on noncompact domains. When one or both equilibrium measures fail to satisfy a Poincaré inequality, the spectral gap arguments underlying local exponential convergence are unavailable. we show that Wasserstein GDA converges algebraically to the mixed Nash equilibrium in chi-squared divergence, provided the initial relative density is a small perturbation in L^2 with bounded algebraic moments. The proof has two main ingredients: a cancellation in a negative Sobolev energy estimate, and a bootstrap argument that keeps the moments bounded uniformly in time. For polynomial-tailed equilibria, we construct perturbations supported far out in the tails that survive over diffusive time scales, which gives algebraic lower bounds on the worst-case relaxation rate. We also establish analogous local algebraic stability and lower bound results for mean-field Langevin dynamics, when the equilibrium fails to satisfy a spectral gap.
ER  -