Abstract:We develop an innovative learning framework that incorporate the noise structure to infer the governing equations from observation of trajectory data generated by stochastic dynamics. Our approach can proficiently captures both the noise and the drift terms. Furthermore, it can also accommodate a wide range of noise types, including correlated and state-dependent variations. Moreover, our method demonstrates scalability to high-dimensional systems. Through extensive numerical experiments, we showcase the exceptional performance of our learning algorithm in accurately reconstructing the underlying stochastic dynamics.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is to infer the governing equations generated by stochastic dynamics from the observed trajectory data, especially to accurately estimate the drift and diffusion terms in stochastic differential equations (SDEs) from data with noise structures. Specifically: 1. **Introducing a noise - guided learning framework**: The paper proposes an innovative learning framework that incorporates the noise structure into the learning process to infer the governing equations from the observed trajectory data. This method can efficiently capture the drift and noise terms and can handle multiple types of noise, including correlated noise and state - dependent noise. 2. **Scalability for high - dimensional systems**: This method demonstrates scalability for high - dimensional systems and verifies its superior performance in accurately reconstructing the underlying stochastic dynamics through extensive numerical experiments. 3. **Diversity of noise structures**: The method in the paper is not limited to a specific type of noise but can adapt to different noise structures, thus increasing the flexibility and applicability of the model. 4. **Challenges in practical applications**: In practical applications, usually only the observation data at discrete time points can be obtained. Therefore, the paper also explores how to effectively perform parameter estimation and model learning in this situation. ### Formula Representation The key formulas involved in the paper are as follows: - SDE model form: \[ dx_t = f(x_t)dt+\sigma(x_t)dw_t, \quad x_t, w_t\in\mathbb{R}^d \] where \(f:\mathbb{R}^d\rightarrow\mathbb{R}^d\) is the drift term, \(\sigma:\mathbb{R}^d\rightarrow\mathbb{R}^{d\times d}\) is the diffusion coefficient matrix, and \(w_t\) is an independent standard Brownian motion. - Covariance matrix: \[ \Sigma(x)=\sigma(x)\sigma(x)^T \] - Loss function: \[ E_H(\tilde{f})=\mathbb{E}\left[\frac{1}{2}\int_0^T\langle\tilde{f}(x_t),\Sigma^{\dagger}(x_t)\tilde{f}(x_t)\rangle dt - 2\langle\tilde{f}(x_t),\Sigma^{\dagger}(x_t)dx_t\rangle\right] \] where \(\Sigma^{\dagger}\) is the pseudo - inverse of \(\Sigma\). - Loss function for discrete data: \[ E_{L,M,H}(\tilde{f})=\frac{1}{2TM}\sum_{l = 1}^{L - 1}\sum_{m = 1}^{M}\left(\langle\tilde{f}(x_m^l),\Sigma^{-1}(x_m^l)\tilde{f}(x_m^l)\rangle(t_{l+1}-t_l)-2\langle\tilde{f}(x_m^l),\Sigma^{-1}(x_m^l)(x_m^{l+1}-x_m^l)\rangle\right) \] Through these formulas, the paper proposes a systematic framework to handle the learning problem of stochastic dynamic systems with noise and demonstrates its effectiveness and robustness in multiple application scenarios.

Noise Guided Structural Learning from Observing Stochastic Dynamics

Learning Stochastic Dynamics from Data

Non-Markovian Reduced Models to Unravel Transitions in Non-equilibrium Systems

An end-to-end deep learning approach for extracting stochastic dynamical systems with $α$-stable Lévy noise

Learning noise-induced transitions by multi-scaling reservoir computing

Learning Collective Behaviors from Observation

Noise-aware neural network for stochastic dynamics simulation

Deep Active Learning with Noise Stability

Joint Learning of Full-structure Noise in Hierarchical Bayesian Regression Models

Learning Dynamics from Noisy Measurements using Deep Learning with a Runge-Kutta Constraint

Data-driven learning and control of nonlinear system dynamics

Noise Helps Optimization Escape from Saddle Points in the Neural Dynamics.

Data-driven approach for extracting the most probable exit trajectory of stochastic dynamical systems with non-Gaussian Lévy noise

A General Framework for Analyzing Stochastic Dynamics in Learning Algorithms

How to Learn and Generalize From Three Minutes of Data: Physics-Constrained and Uncertainty-Aware Neural Stochastic Differential Equations

Weak Collocation Regression for Inferring Stochastic Dynamics with Lévy Noise

One Noise to Rule Them All: Learning a Unified Model of Spatially-Varying Noise Patterns

Neural Structure Learning with Stochastic Differential Equations

Learning Control-Oriented Dynamical Structure from Data

Learning interpretable dynamics of stochastic complex systems from experimental data