Tensor-based homogeneous polynomial dynamical system analysis from data

Xin Mao, Anqi Dong, Ziqin He, Yidan Mei, Shenghan Mei, Can ChenarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces a novel framework for analyzing complex, higher-order dynamical systems (HPDSs) by leveraging tensor decomposition techniques. The authors develop efficient methods for identifying system parameters and analyzing crucial properties like controllability and observability, significantly reducing computational complexity for large-scale systems.

What Problem Does This Paper Address?

Many real-world systems, from ecological networks to social dynamics, exhibit complex interactions that are best modeled by polynomial dynamical systems, particularly homogeneous polynomial dynamical systems (HPDSs). However, analyzing these systems is notoriously difficult due to their inherent nonlinearity and the sheer scale of parameters involved, especially in large systems. Traditional methods struggle with computational complexity and memory demands. This paper aims to bridge this gap by providing a more efficient and scalable approach to HPDS analysis.

Methods

The core of this work lies in representing HPDSs as tensors and then applying advanced tensor decomposition techniques: Tensor Train Decomposition (TTD) and Hierarchical Tucker Decomposition (HTD). These methods allow for a compact, low-rank representation of the system's dynamic tensor, drastically reducing the number of parameters.

Key methodological contributions include:

  • Memory-Efficient System Identification: The authors generalize the "fundamental lemma" to directly estimate system parameters from time-series data within the TTD and HTD frameworks. This bypasses the need to reconstruct the full, high-dimensional tensor.
  • Efficient Controllability and Observability Analysis: They derive necessary and sufficient algebraic conditions for determining these critical system properties using the factor matrices and tensors from TTD and HTD. This avoids the computationally intensive calculations associated with full tensor representations.
  • Complexity Analysis: Rigorous analysis demonstrates significant reductions in both memory and computational complexity compared to traditional full-tensor approaches.

Datasets

Not discussed in the paper. The paper focuses on a theoretical framework and validates it through numerical experiments using randomly generated tensors.

Key Insights and Ideas

  • Tensors as a Natural Representation for HPDSs: The paper solidifies the idea that HPDSs can be elegantly and equivalently represented by tensors, opening the door to powerful tensor algebra tools for system analysis.
  • TTD and HTD as Game-Changers for Scale: The choice of TTD and HTD is crucial. They offer a practical way to handle the "curse of dimensionality" that plagues high-order, high-dimensional systems, making analysis feasible where it was previously intractable.
  • Algebraic Conditions for Controllability/Observability: Moving beyond complex geometric approaches (like Lie brackets), the paper provides concrete, computable algebraic conditions based on tensor decompositions. This is a significant practical advancement.
  • Trade-offs Between TTD and HTD: The paper implicitly highlights that the choice between TTD and HTD can depend on the underlying structure of the dynamic tensor (e.g., chain-like vs. hierarchical), offering flexibility in application.

Key Results

  • Significant Memory Reduction: Numerical experiments show that TTD and HTD representations drastically reduce the number of model parameters compared to the full tensor representation, especially for tensors with inherent low-rank structures. For instance, Figure 3 illustrates orders-of-magnitude reductions in parameters.
  • Substantial Computational Speedup: Figure 4 demonstrates that TTD and HTD-based methods significantly outperform the full tensor representation in terms of computation time for constructing controllability matrices. The speedup is particularly pronounced as the tensor order and dimension increase.
  • Generalization of Existing Results: The framework extends existing controllability and observability conditions for HPDSs to more general cases (e.g., almost symmetric tensors) and provides efficient computational means to verify them.

Limitations and Future Work

  • Limitations:
    • The paper primarily focuses on autonomous and input-output HPDSs with linear inputs. Extending to more complex nonlinear inputs or systems with more intricate structures might require further development.
  • The analysis of controllability for even-degree systems is based on "strong controllability," and the paper notes that global controllability for such systems remains an open problem in general.
  • The current framework assumes that state derivatives are not directly measurable and are approximated.
  • Future Work:
    • Stochastic and Hybrid Systems: Extending the TTD/HTD framework to analyze stochastic and hybrid dynamical systems.
  • Integration with Machine Learning: Combining data-driven tensor-based methods with machine learning algorithms to enhance scalability, adaptability, and real-time decision-making.
  • Robustness and Uncertainty: Investigating the robustness and efficiency of these approaches in high-dimensional and uncertain environments.
  • Real-World Applications: Applying the framework to complex, higher-order networked systems in fields like biology, neuroscience, and engineering.

Summary updated

Abstract

Numerous complex real-world systems, such as those in biological, ecological, and social networks, exhibit higher-order interactions that are often modeled using polynomial dynamical systems or homogeneous polynomial dynamical systems (HPDSs). However, identifying system parameters and analyzing key system-theoretic properties remain challenging due to their inherent nonlinearity and complexity, particularly for large-scale systems. To address these challenges, we develop an innovative computational framework in this article that leverages advanced tensor decomposition techniques, namely tensor train and hierarchical Tucker decompositions, to facilitate efficient identification and analysis of HPDSs that can be equivalently represented by tensors. Specifically, we introduce memory-efficient system identification techniques for directly estimating system parameters represented through tensor decompositions from time-series data. Additionally, we develop necessary and sufficient conditions for determining controllability and observability using the tensor decomposition-based representations of HPDSs, accompanied by detailed complexity analyses that demonstrate significant reductions in computational demands. The effectiveness and efficiency of our framework are validated through numerical examples.
BibTeX
@misc{mao2025tensor,
  author = {Xin Mao and Anqi Dong and Ziqin He and Yidan Mei and Shenghan Mei and Can Chen},
  title = {{Tensor-based homogeneous polynomial dynamical system analysis from data}},
  year = {2025},
  eprint = {2503.17774},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.17774},
}
Figure