ERGODICITY OF THE UNDERDAMPED MEAN-FIELD LANGEVIN DYNAMICS

被引：1

作者：

Kazeykina, Anna ^{[1
]}

Ren, Zhenjie ^{[2
]}

Tan, Xiaolu ^{[3
]}

Yang, Junjian ^{[4
]}

机构：

[1] Univ Paris Saclay, Fac Sci Orsay, Dept Math, Paris, France

[2] PSL, Univ Paris Dauphine, CEREMADE, Paris, France

[3] Chinese Univ Hong Kong, Dept Math, Hong Kong, Peoples R China

[4] Vienna Univ Technol, Fak Math & Geoinformat, FAM, Vienna, Austria

来源：

ANNALS OF APPLIED PROBABILITY | 2024年 / 34卷 / 03期

关键词：

Underdamped mean-field Langevin dynamics; ergodicity; coupling; GAN; EXPONENTIAL CONVERGENCE; MOLECULAR-DYNAMICS; KINETIC-EQUATIONS; CONTRACTION RATES; TIME-REVERSAL; HYPOCOERCIVITY; PROPAGATION; EQUILIBRIUM; COUPLINGS; BEHAVIOR;

D O I：

10.1214/23-AAP2036

中图分类号：

O21 [概率论与数理统计]; C8 [统计学];

学科分类号：

020208 ; 070103 ; 0714 ;

摘要：

We study the long time behavior of an underdamped mean -field Langevin (MFL) equation, and provide a general convergence as well as an exponential convergence rate result under different conditions. The results on the MFL equation can be applied to study the convergence of the Hamiltonian gradient descent algorithm for the overparametrized optimization. We then provide some numerical examples of the algorithm to train a generative adversarial network (GAN).

引用

页码：3181 / 3226

页数：46

共 50 条

[1] Optimization by linear kinetic equations and mean-field Langevin dynamics
Pareschi, Lorenzo
MATHEMATICAL MODELS & METHODS IN APPLIED SCIENCES, 2024, 34 (12): : 2191 - 2216
[2] Mean-field Langevin dynamics and energy landscape of neural networks
Hu, Kaitong
Ren, Zhenjie
Siska, David
Szpruch, Lukasz
ANNALES DE L INSTITUT HENRI POINCARE-PROBABILITES ET STATISTIQUES, 2021, 57 (04): : 2043 - 2065
[3] Viscoelastic active diffusion governed by nonequilibrium fractional Langevin equations: Underdamped dynamics and ergodicity breaking
Joo, Sungmin
Jeon, Jae-Hyung
CHAOS SOLITONS & FRACTALS, 2023, 177
[4] On weak ergodicity breaking in mean-field spin glasses
Folena, Giampaolo
Zamponi, Francesco
SCIPOST PHYSICS, 2023, 15 (03):
[5] Feature learning via mean-field Langevin dynamics: classifying sparse parities and beyond
Suzuki, Taiji
Wu, Denny
Oko, Kazusato
Nitanda, Atsushi
ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 36 (NEURIPS 2023), 2023,
[6] Uncertainty relations for underdamped Langevin dynamics
Tan Van Vu
Hasegawa, Yoshihiko
PHYSICAL REVIEW E, 2019, 100 (03):
[7] Heat fluctuations for underdamped Langevin dynamics
Rosinberg, Martin Luc
Tarjus, Gilles
Munakata, Toyonori
EPL, 2016, 113 (01)
[8] Gradient Estimates and Exponential Ergodicity for Mean-Field SDEs with Jumps
Song, Yulin
JOURNAL OF THEORETICAL PROBABILITY, 2020, 33 (01) : 201 - 238
[9] Strong ergodicity breaking in aging of mean-field spin glasses
Bernaschi, Massimo
Billoire, Alain
Maiorano, Andrea
Parisi, Giorgio
Ricci-Tersenghi, Federico
PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA, 2020, 117 (30) : 17522 - 17527
[10] Gradient Estimates and Exponential Ergodicity for Mean-Field SDEs with Jumps
Yulin Song
Journal of Theoretical Probability, 2020, 33 : 201 - 238

← 1 2 3 4 5 →