Modeling Latent Non-Linear Dynamical System over Time Series

Ren Fujiwara, Yasuko Matsubara, Yasushi Sakurai

Introduction

Discovering the governing mathematical laws of dynamical systems from observed time series is a fundamental challenge in science and engineering. While sparse identification methods like SINDy have shown promise, they often falter when faced with high measurement noise or when the underlying dynamics occur in a latent space not directly observed. This paper presents LaNoLem (Latent Non-Linear equation modeling), a framework designed to robustly identify non-linear governing equations by mapping noisy observations into a latent state space.

Methodology

LaNoLem addresses the limitations of existing methods by explicitly modeling the system as a latent non-linear dynamical system. The framework consists of two primary components: an observation model and a transition model.

Overview of the LaNoLem framework, illustrating the decomposition of original data into latent states and the sparse non-linear dynamics governing those states.

As shown in the figure above, the observed data $\mathbf{X}$ is modeled as a linear projection of latent states $\mathbf{S}$ via an observation matrix $\mathbf{C}$ and an offset $\mathbf{u}$. The dynamics of these latent states are governed by a sparse non-linear equation: $$\mathbf{s}(t+1) \sim \mathbf{A}\mathbf{s}(t) + \mathbf{F}\phi(\mathbf{s}(t), d_\phi) + \mathbf{b}$$ where $\mathbf{A}$ represents linear dynamics, $\mathbf{F}$ captures non-linear interactions through a library of basis functions $\phi$, and $\mathbf{b}$ is an offset.

To optimize this model, the authors employ an alternating minimization strategy. An Inference Step uses an Extended Kalman Filter (EKF) and smoother to estimate latent states, while a Learning Step utilizes Generalized Gradient Descent with elastic-net regularization to identify sparse coefficients. The model complexity is automatically determined using the Minimum Description Length (MDL) principle.

Experimental Evaluation

The framework was evaluated using the dysts database, comprising 71 chaotic systems (e.g., Lorenz, Rossler) under varying noise ratios (5%, 25%, and 50%). Additionally, real-world Google Trends data for seasonal search terms was used to test interpretability.

Results and Discussion

LaNoLem demonstrates significant improvements over baseline methods such as MIOSR, SSR, and STLSQ, particularly in high-noise environments.

Performance comparison of LaNoLem against baselines across 71 chaotic systems. (a) Coefficient error and (b) Prediction error at different noise levels.

As illustrated in the performance plots, LaNoLem consistently achieves the lowest prediction and coefficient errors. While other methods degrade rapidly as noise increases to 50%, LaNoLem maintains stability by effectively filtering noise through its latent state representation. In real-world applications, the model successfully identified seasonal latent factors in search data, grouping winter and summer activities into opposing weights on a single latent dimension without prior domain knowledge.

Conclusion

LaNoLem provides a robust framework for discovering the hidden mathematical structures of non-linear systems from noisy time series. By integrating state-space modeling with sparse regression, it bridges the gap between denoising and equation discovery. Future work includes expanding the library of basis functions beyond polynomials and improving initialization strategies to avoid local optima.

Summary updated

BibTeX
@inproceedings{fujiwara2025modeling,
  author = {Ren Fujiwara and Yasuko Matsubara and Yasushi Sakurai},
  title = {{Modeling Latent Non-Linear Dynamical System over Time Series}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i11.33269},
}
Figure