Abstract
This paper presents a rigorous theoretical analysis of model reduction for linear switched systems (LSS) using balanced truncation. We establish a guaranteed bound on the approximation error in the $L_2$ norm for continuous-time systems and the $l_2$ norm for discrete-time systems. Furthermore, we demonstrate that the performance of this reduction method depends solely on the system's input-output map, ensuring realization invariance. Our findings confirm that balanced truncation preserves quadratic stability and provides a system-theoretic interpretation of Grammians and their singular values in the context of switching dynamics.
Introduction
Linear Switched Systems (LSS) are a versatile class of models used to describe processes that transition between different linear modes of operation. Common applications include robotics (e.g., legged locomotion), power electronics, and automotive control. However, as the complexity of these systems increases, the state dimension often becomes computationally prohibitive for real-time control and simulation.
Model reduction techniques, particularly balanced truncation, offer a path toward simplifying these models while maintaining high fidelity. While balanced truncation is well-understood for standard linear time-invariant (LTI) systems, its application to LSS has historically lacked the same level of theoretical rigor. This paper bridges that gap by providing the first comprehensive "safety net" for LSS model reduction.
The Model Reduction Problem
The goal of model reduction is to find a lower-dimensional system that approximates the input-output behavior of a high-dimensional original system. In the context of LSS, this involves identifying states that contribute minimally to the system's dynamics—specifically those that are difficult to reach (low controllability) and difficult to observe (low observability).
Prior to this work, several questions remained unanswered:
- Error Quantification: Can we provide a strict mathematical bound on the loss of accuracy?
- Realization Invariance: Is the reduction quality dependent on the specific mathematical representation (state-space realization) used?
- Stability Preservation: Does the reduced-order model inherit the stability properties of the original system?
Methodology: Balanced Truncation via LMIs
The authors utilize Linear Matrix Inequalities (LMIs) to compute the Controllability Grammian ($P$) and the Observability Grammian ($Q$). These matrices represent the energy required to reach certain states and the energy generated by those states at the output, respectively.
The reduction process follows a structured three-step approach:
- Solving LMIs: Identify $P$ and $Q$ such that the system's energy remains bounded under any valid switching sequence.
- Balancing Transformation: Apply a state-space transformation (isomorphism) to reach a "balanced" realization where $P = Q = \Sigma$, where $\Sigma$ is a diagonal matrix of singular values ($\sigma_1, \sigma_2, \dots, \sigma_n$).
- Truncation: Discard the states associated with the smallest singular values, which represent the least "energetic" components of the system.
Key Theoretical Insights
The Error Bound
The most significant contribution of this research is the proof of a global error bound. For a system $\Sigma$ and its reduced version $\hat{\Sigma}$, the approximation error is bounded by: $$\|Y^\Sigma - Y^{\hat{\Sigma}}\|_2 \le 2 \sum_{k=r+1}^n \sigma_k$$ This result mirrors the famous error bound for standard LTI systems, providing engineers with a reliable metric to predict the performance of the reduced model.
Realization Invariance
The paper proves that the singular values are fundamental properties of the system's input-output map. This means that regardless of how the internal equations are written, the balanced truncation algorithm will yield a consistent level of accuracy, provided the underlying physical behavior is the same.
Stability and "Nice" Grammians
The authors distinguish between quadratic stability and "strong stability." They prove that if the original LSS is quadratically stable—meaning a single Lyapunov function exists for all switching modes—the reduced system is guaranteed to remain quadratically stable. For discrete-time systems, they introduce "nice" Grammians that provide a deeper connection to the reachability and controllability of the system.
Results and Discussion
The theoretical framework established here confirms that LSS behave predictably under balanced truncation. The relationship between the largest singular value and the Hankel norm provides a physical interpretation of these values as the maximum energy the system can produce.
However, the authors note some limitations. The current method may produce models that are not "minimal," meaning an even smaller representation might exist that captures the same behavior. Additionally, the proven error bounds are often conservative; in practice, the actual error is frequently much lower than the theoretical maximum.
Conclusion
By establishing rigorous error bounds and proving realization invariance, this work provides the necessary theoretical foundation for applying balanced truncation to linear switched systems. These results ensure that model reduction is not just an empirical heuristic but a mathematically sound tool for managing the complexity of modern switched dynamical systems.
Summary updated