Minimality and identifiability of discrete-time SARX systems

Mihály Petreczky, Laurent Bako, S. Lecœuche, K.m.d. MotchonarXiv 2020 · DOI · arXiv

Abstract

The paper studies the problem of minimality and identifiability for Switched Auto-Regressive eXogenous (abbreviated by SARX) systems. We propose formal definitions of the concepts of identifiability and minimality for SARX models. Based on these formalizations, we derive conditions for minimality and identifiability of SARX systems. In particular, we show that polynomially parametrized SARX systems are generically identifiable.

Introduction

In the field of control theory and system identification, building mathematical models from input-output data is a foundational task. Switched Auto-Regressive eXogenous (SARX) systems have emerged as a powerful tool for modeling "hybrid" behaviors—systems that switch between different linear modes, such as a vehicle shifting gears or a thermostat cycling between heating and cooling.

Despite their widespread use, the theoretical foundations of SARX systems remained underdeveloped. Specifically, two critical questions lacked rigorous answers:

  1. Minimality: Is a given model the simplest possible representation of the observed behavior, or does it contain redundant parameters?
  2. Identifiability: Can the model's parameters be uniquely recovered from perfect data, or could multiple parameter sets yield identical outputs?

This paper addresses these gaps by establishing formal definitions and providing checkable conditions for both minimality and identifiability.

Theoretical Framework: Bridging SARX and LSS

To analyze SARX systems, the authors bridge the gap between input-output models and state-space models, specifically Linear Switched Systems (LSS). By transforming a SARX system into an "associated LSS," the researchers leverage powerful tools from state-space theory to prove properties about the SARX structure.

Formalizing Minimality

A key contribution of the paper is the introduction of "Strong Minimality." The authors demonstrate a counter-intuitive property of hybrid systems: the minimality of a SARX system is not merely the sum of the minimality of its individual subsystems. In fact, a SARX system can be minimal even if every one of its constituent modes, when viewed in isolation, is redundant. This insight highlights the unique complexity of switched systems compared to traditional linear models.

Identifiability and Genericity

Identifiability requires the mapping from parameters to the system to be injective (one-to-one). The paper proves that for polynomially parametrized SARX systems, identifiability is a "generic" property. This means that if a model class contains at least one identifiable system, then almost all systems within that class are identifiable. In practical terms, non-identifiable models are rare mathematical exceptions rather than the rule.

Methodology and Algorithmic Verification

Beyond theoretical proofs, the authors provide a roadmap for verifying these properties in complex systems. They utilize Gröbner bases—a sophisticated tool from computational algebraic geometry—to develop an automated procedure (Procedure 1). This procedure allows researchers to identify subsets of parameters that are guaranteed to be identifiable.

While the paper is primarily theoretical, it validates these concepts using illustrative examples, such as a discrete-time model of a spark ignition engine intake manifold. This example demonstrates how physical systems are mapped into the SARX framework and how the proposed theorems apply to real-world dynamics.

Key Results

  • Theorems 1 & 2: Establish the necessary and sufficient conditions for checking if a Single-Input Single-Output (SISO) SARX system is minimal and identifiable.
  • Theorem 3 (Genericity): Confirms that for any polynomial parametrization, identifiability is almost always guaranteed if the model structure is well-posed.
  • Computational Implementation: The authors successfully implemented their Gröbner basis-based procedure, proving that abstract algebraic properties can be verified algorithmically.

Limitations and Future Directions

While this work provides a rigorous foundation, several challenges remain for future research:

  • Computational Complexity: Calculating Gröbner bases is computationally intensive. For high-dimensional systems with a large number of parameters, symbolic approaches may face scalability issues.
  • MIMO Systems: While the general theory is broad, the most explicit, easy-to-check conditions are currently focused on SISO systems. Extending these to Multi-Input Multi-Output (MIMO) systems is a natural next step.
  • Practical Identifiability: This study focuses on structural identifiability (the mathematical properties of the model). Future work must address how sensor noise and finite data affect the practical ability to estimate these parameters in real-world environments.

Conclusion

By providing the first formal definitions of minimality and identifiability for SARX systems, this paper transforms a popular modeling heuristic into a mathematically sound framework. The proof of generic identifiability and the introduction of algebraic verification tools offer researchers the confidence that their identification algorithms are well-posed and their models are non-redundant.

Summary updated

Abstract

The paper studies the problem of minimality and identifiability for Switched Auto-Regressive eXogenous (abbreviated by SARX) systems. We propose formal definitions of the concepts of identifiability and minimality for SARX models. Based on these formalizations, we derive conditions for minimality and identifiability of SARX systems. In particular, we show that polynomially parametrized SARX systems are generically identifiable.
BibTeX
@misc{petreczky2020minimality,
  author = {Mihály Petreczky and Laurent Bako and S. Lecœuche and K.M.D. Motchon},
  title = {{Minimality and identifiability of discrete-time SARX systems}},
  year = {2020},
  eprint = {2002.01818},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2002.01818},
}
Figure