Expressiveness and Structure Preservation in Learning Port-Hamiltonian Systems

Juan-Pablo Ortega, Daiying YinLecture Notes in Computer Science 2023 · DOI · OpenAlex

Abstract

A well-specified parametrization for single-input/single-output (SISO) linear port-Hamiltonian systems amenable to structure-preserving supervised learning is provided. The construction is based on controllable and observable normal form Hamiltonian representations for those systems, which reveal fundamental relationships between classical notions in control theory and crucial properties in the machine learning context, like structure-preservation and expressive power. The results in the paper suggest parametrizations of the estimation problem associated with these systems that amount, at least in the canonical case, to unique identification and prove that the parameter complexity necessary for the replication of the dynamics is only $O(n)$ and not $O(n^2)$, as suggested by the standard parametrization of these systems.

Introduction

In the burgeoning field of physics-informed machine learning, the primary objective is to develop models that do not merely fit data but also adhere to fundamental physical principles, such as the conservation of energy. Port-Hamiltonian (PH) systems offer a powerful mathematical framework for this task, as they describe dynamical systems through the lens of energy storage and exchange.

However, two significant hurdles have historically impeded the efficient "learning" of these systems from data. First, standard descriptions of PH systems are often redundant, utilizing $O(n^2)$ parameters, which leads to overfitting and optimization difficulties. Second, most existing methodologies assume full access to the internal state of the system. In practical engineering applications, we typically only have access to external inputs (e.g., voltage) and outputs (e.g., velocity), leaving the internal state hidden. This paper addresses these challenges by bridging the gap between classical control theory and modern machine learning.

Theoretical Framework and Methods

The authors propose a "minimal" representation of linear PH systems that preserves physical structure while focusing on input-output behavior. This is achieved through several key mathematical advancements:

  • System Morphisms: The research demonstrates that complex physical systems can be mapped to a "normal form" via mathematical morphisms without altering their observable input-output behavior.
  • Williamson’s Normal Form: Leveraging results from symplectic geometry, the authors simplify the Hamiltonian (energy) matrix, reducing its complexity.
  • Controllable and Observable Representations: The paper defines new canonical forms for Hamiltonian systems. Rather than learning a full, dense energy matrix $Q$, the model learns a vector $d$ of symplectic eigenvalues and an interconnection vector $v$.
  • Lie Groupoids: To address the "identification problem"—where multiple parameter sets might yield identical behavior—the authors utilize groupoid orbit spaces to define a space where parameters correspond to unique physical behaviors.

The $O(n)$ Complexity Breakthrough

The most significant insight of this work is the reduction of parameter complexity. For a linear Port-Hamiltonian system of dimension $n$, the authors prove that $O(n^2)$ parameters are unnecessary. Instead, $O(n)$ parameters are sufficient to fully describe the system's dynamics. This drastically simplifies the optimization landscape, making the training of these models more efficient and less prone to the "curse of dimensionality."

Furthermore, this approach ensures structure preservation by construction. Because the model is built using Hamiltonian normal forms, every iteration of the machine learning algorithm is guaranteed to represent a physically valid system. The model cannot "drift" into non-physical states that violate energy conservation laws.

Key Results

The central contribution of the paper is Theorem 2, which proves that any linear Port-Hamiltonian system can be represented by a controllable Hamiltonian representation of the same dimension. This theorem provides the theoretical justification for the $O(n)$ complexity reduction.

Additionally, the paper establishes that classical control theory concepts—specifically controllability (the ability to move a system to any state) and observability (the ability to infer states from outputs)—are directly linked to the expressive power of a machine learning model. By focusing on the "filter" (how the system transforms signals) rather than trying to uncover every hidden internal variable, the authors provide a more pragmatic tool for engineers.

Limitations and Future Work

While this paper provides a rigorous foundation for learning physical systems, it currently focuses on linear, SISO (Single-Input/Single-Output), and dissipation-less systems. Real-world physics is frequently non-linear and involves energy loss (dissipation).

The authors have indicated that future research will extend these results to Multi-Input/Multi-Output (MIMO) systems and further explore the use of group orbit spaces to solve the unique identification problem in more complex, dissipative scenarios. This work represents a vital step toward making physics-informed AI both mathematically sound and computationally efficient.

Summary updated

BibTeX
@inproceedings{ortega2023expressiveness,
  author = {Juan-Pablo Ortega and Daiying Yin},
  title = {{Expressiveness and Structure Preservation in Learning Port-Hamiltonian Systems}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2023},
  doi = {10.1007/978-3-031-38299-4\_33},
}

Loading related content...

Figure