Abstract:We explain the general theories involved in solving an infinite-dimensional stochastic differential equation (ISDE) for interacting Brownian motions in infinite dimensions related to random matrices. Typical examples are the stochastic dynamics of infinite particle systems with logarithmic interaction potentials such as the sine, Airy, Bessel, and also for the Ginibre interacting Brownian motions. The first three are infinite-dimensional stochastic dynamics in one-dimensional space related to random matrices called Gaussian ensembles. They are the stationary distributions of interacting Brownian motions and given by the limit point processes of the distributions of eigenvalues of these random matrices. The sine, Airy, and Bessel point processes and interacting Brownian motions are thought to be geometrically and dynamically universal as the limits of bulk, soft edge, and hard edge scaling. The Ginibre point process is a rotation- and translation-invariant point process on R 2 \mathbb {R}^2 , and an equilibrium state of the Ginibre interacting Brownian motions. It is the bulk limit of the distributions of eigenvalues of non-Hermitian Gaussian random matrices. When the interacting Brownian motions constitute a one-dimensional system interacting with each other through the logarithmic potential with inverse temperature β = 2 \beta = 2 , an algebraic construction is known in which the stochastic dynamics are defined by the space-time correlation function. The approach based on the stochastic analysis (called the analytic approach) can be applied to an extremely wide class. If we apply the analytic approach to this system, we see that these two constructions give the same stochastic dynamics. From the algebraic construction, despite being an infinite interacting particle system, it is possible to represent and calculate various quantities such as moments by the correlation functions. We can thus obtain quantitative information. From the analytic construction, it is possible to represent the dynamics as a solution of an ISDE. We can obtain qualitative information such as semi-martingale properties, continuity, and non-collision properties of each particle, and the strong Markov property of the infinite particle system as a whole. Ginibre interacting Brownian motions constitute a two-dimensional infinite particle system related to non-Hermitian Gaussian random matrices. It has a logarithmic interaction potential with β = 2 \beta = 2 , but no algebraic configurations are known.The present result is the only construction.

Stochastic weight matrix dynamics during learning and Dyson Brownian motion

Online Stochastic Gradient Descent Learns Linear Dynamical Systems from A Single Trajectory

Machine learning in and out of equilibrium

Matrix Product Belief Propagation for reweighted stochastic dynamics over graphs

Stochastic geometry and dynamics of infinitely many particle systems—random matrices and interacting Brownian motions in infinite dimensions

Graph Neural Stochastic Differential Equations for Learning Brownian Dynamics

Learning stochastic dynamical systems with neural networks mimicking the Euler-Maruyama scheme

A General Framework for Analyzing Stochastic Dynamics in Learning Algorithms

Rigorous dynamical mean field theory for stochastic gradient descent methods

Learning Stochastic Dynamical Systems As an Implicit Regularization with Graph Neural Networks

Stochastic Learning of Computational Resource Usage as Graph Structured Multimarginal Schrödinger Bridge

Learning effective stochastic differential equations from microscopic simulations: Linking stochastic numerics to deep learning.

A quantum inspired approach to learning dynamical laws from data -- block-sparsity and gauge-mediated weight sharing

Tensor-Networks-based Learning of Probabilistic Cellular Automata Dynamics

On the weight dynamics of learning networks

An Algorithm for Computing Stochastically Stable Distributions with Applications to Multiagent Learning in Repeated Games

Learning the Non-Equilibrium Dynamics of Brownian Movies

Generalized Glauber dynamics for inference in biology

Learning Markovian dynamics with spectral maps

Learning effective stochastic differential equations from microscopic simulations: combining stochastic numerics and deep learning

Stochastic Modified Flows, Mean-Field Limits and Dynamics of Stochastic Gradient Descent