Identification of Piecewise Linear Models of Complex Dynamical Systems

Ronald L. Westra, Mihály Petreczky, Ralf L. M. PeetersarXiv 2011 · OpenReview

Abstract

The paper addresses the realization and identification problem for a subclass of piecewise-affine hybrid systems. The paper provides necessary and sufficient conditions for existence of a realization, a characterization of minimality, and an identification algorithm for this subclass of hybrid systems. The considered system class and the identification problem are motivated by applications in systems biology.

Introduction

Biological signaling networks, such as gene-protein interactions, often exhibit complex behaviors that defy simple linear modeling. These systems frequently operate in a "hybrid" fashion, maintaining linear dynamics around an equilibrium before abruptly switching modes due to internal state changes or external stimuli. This paper establishes a theoretical framework for identifying Piecewise Linear (PWL) models to describe these networks. While any PWL system can theoretically be reduced to a single-state model with "jumps," the authors argue that imposing constraints on the number of discrete states is vital for maintaining biological interpretability.

Theoretical Framework

The authors focus on a subclass of Piecewise-Affine Hybrid Systems (PWL) that allow the continuous state to "jump" to a new starting point upon mode switching.

Realization and Minimality

A central contribution is the concept of Linear Systems with State Jumps (LSSJ). The authors demonstrate that a multi-mode PWL system can be mathematically represented as a single linear system that undergoes periodic state resets. Using a modified Hankel-Matrix approach—analogous to the classical Kalman-Ho algorithm—they derive necessary and sufficient conditions for the existence of a PWL realization based on observed Markov parameters.

Interestingly, the authors uncover a "minimality paradox": any PWL system can be transformed into a minimal system with only one discrete state. However, such a reduction is often undesirable in biology, as it obscures the distinct functional modes (e.g., phases of a cell cycle). To address this, they introduce $K-N$ Realization constraints, ensuring the model preserves $K$ discrete modes and $N$ continuous dimensions.

Identification of piecewise linear dynamics from simulated data, showing distinct mode transitions.

Identification Algorithm

For systems with full state observation, the authors propose an iterative identification algorithm. This method alternates between estimating the active mode (weights) and calculating the linear parameters of those modes using least squares. The paper also proves that PWL systems are functionally equivalent to switched Auto-Regressive (AR) models, bridging hybrid systems theory with time-series analysis.

Experimental Results

The identification algorithm was validated across three distinct scenarios:

  1. Artificial Systems: The algorithm successfully recovered parameters from randomly generated PWL systems, demonstrating robustness even at a 5% Signal-to-Noise Ratio (SNR).
  2. Lorenz Attractor: To test approximation capabilities on non-linear systems, the algorithm was applied to the chaotic Lorenz attractor. It successfully partitioned the "butterfly" attractor into two primary lobes and a transition region using just three subsystems.
  3. Yeast Cell Cycle: Using the Tyson-Novak model, the algorithm automatically identified two discrete modes corresponding to the S phase (DNA synthesis) and M phase (mitosis). The reconstruction was robust to noise levels up to 10%.
Partitioning the chaotic Lorenz attractor into discrete linear modes (2-mode vs 3-mode representations).

Discussion and Future Work

While the proposed framework is powerful, it currently assumes full observation of all system variables—a luxury rarely available in experimental biology where many components remain "hidden." Additionally, the iterative algorithm can occasionally converge to local minima. Future research will focus on extending these methods to partial observation models and establishing formal convergence proofs for the identification process.

Summary updated

Abstract

The paper addresses the realization and identification problem or a subclass of piecewise-affine hybrid systems. The paper provides necessary and sufficient conditions for existence of a realization, a characterization of minimality, and an identification algorithm for this subclass of hybrid systems. The considered system class and the identification problem are motivated by applications in systems biology.
BibTeX
@misc{westra2011identification,
  author = {Ronald L. Westra and Mihály Petreczky and Ralf L. M. Peeters},
  title = {{Identification of Piecewise Linear Models of Complex Dynamical Systems}},
  year = {2011},
  eprint = {1103.4756},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1103.4756},
}
Figure