- The paper establishes that partial realizations exist under finite Hankel matrix rank conditions and are unique up to isomorphism.
- It introduces two constructive algorithms—one via Hankel column space and a Kalman-Ho-like factorization—for practical system realization.
- The formal power series framework unifies hybrid system theory and extends classical realization methods to improve model identification and reduction.
Introduction and Motivation
The paper "Partial-realization theory and algorithms for linear switched systems: A formal power series approach" (1010.5160) systematically develops partial-realization theory for linear switched systems (LSSs) using a formal power series framework. The motivation is twofold: to extend the reach of realization and system identification theory into the hybrid systems domain, and to leverage rational formal power series for algorithmic and theoretical advancements in identification and model reduction tasks. The authors focus primarily on LSSs, which are a canonical class of hybrid systems characterized by continuous-time linear dynamics in each discrete mode with arbitrary switching, and provide generalizations that are applicable to broader hybrid classes.
The partial-realization problem is to determine, for a given family of input-output maps Φ (with Markov parameters known up to some finite order), whether there exists a finite-dimensional LSS that captures this finite behavior. Specifically, the aim is
- To characterize the conditions under which such a partial realization exists, and
- To establish minimality criteria and algorithms for synthesis.
The problem is motivated by practical limitations: in system identification, only finite data (such as a finite number of Markov parameters or moments) can be collected, and thus, the classical realization problem (which considers infinite data) is inapplicable.
The paper rigorously relates the realization problem for LSSs to the corresponding setting for rational families of formal power series. The authors construct a mapping from the input-output behavior of an LSS to a family of formal power series whose coefficients encode generalized Markov parameters indexed by mode sequences and input indices. The crucial tool is the Hankel matrix associated with this family, whose finite or infinite rank properties become decisive for realization theorems.
The central insight is that
- The existence of a partial realization is equivalent to certain rank conditions on finite subblocks of the Hankel matrix.
- The minimal order of a partial realization corresponds to the rank of these finite subblocks.
Main Theoretical Results
The authors provide a full characterization:
- Existence and Uniqueness: Given a family Φ of input-output maps with Markov parameters up to order $2N+1$, if the upper-left (N,N),(N+1,N),(N,N+1) Hankel submatrices all have the same finite rank n, there exists an n-dimensional $2N+1$-partial realization, and all such realizations are isomorphic.
- Minimality: This partial realization is minimal in the sense that any other with the same matching order has state-space dimension at least n.
- Extension to Complete Realizations: If the rank of the finite submatrix matches the rank of the full Hankel matrix, the partial realization is a minimal complete realization of Φ.
These results mirror the classical Ho-Kalman framework for linear systems but are fully generalized to the hybrid/LSS context.
Algorithms
Two constructive synthesis algorithms are presented:
- Realization via Hankel Column Space: This algorithm constructs an LSS realization by identifying a basis in the column space of the finite Hankel matrix (composed of Markov parameters up to order $2N+1$). System matrices are defined to induce the correct shifts among columns, and input/output maps are recovered by canonical projections.
- Kalman-Ho-Like Factorization Algorithm: This numerically-oriented variant directly factorizes the finite Hankel submatrix (typically via SVD or similar), decomposing it into observable and reachable factors, from which system realization matrices are extracted. This approach is closer to modern subspace identification algorithms and is implementable in practice, as demonstrated by the authors.
Both approaches are proven to produce isomorphic (i.e., equivalent under similarity) realizations when the underlying rank conditions hold.
Numerical Example
A detailed numerical case is provided for a specific two-mode, single-input, single-output LSS. The Markov parameters are computed, and the realization procedures are explicitly illustrated. The example shows:
- For N large enough so that the finite Hankel submatrix has full rank, the exact minimal realization is recovered.
- For smaller N, the resulting partial realization matches the Markov parameters up to the order considered, but not the entire behavior, providing a concrete illustration of the theory.
Implications
Theoretical Implications
- Unified Framework: The formal power series approach unifies the partial realization theory for linear, bilinear, switched, and more general hybrid systems.
- Bridges System and Language Theory: The methodology exploits deep connections between system theory (especially reachability/observability) and formal language theory (e.g., the algebra of power series).
- Generalization Potential: Results announced for piecewise-affine and other classes suggest broad applicability beyond the LSSs.
Practical Implications
- System Identification: The results underpin subspace-oriented identification algorithms for LSSs, extending robust identification tools from linear systems to the hybrid context.
- Model Reduction: The partial realization framework can serve as a foundation for moment-matching model reduction techniques for LSSs, crucial for high-dimensional applications.
- Computational Algorithms: The presented numerical algorithms can be implemented directly for moderate problem sizes, with further optimization possible for large-scale or structured settings.
Contradictory/Strong Claims
A strong---and well-understood---claim is that, under the finite rank conditions, the partial realization constructed is unique (up to isomorphism) and minimal, paralleling classic linear theory in a hybrid/switching systems context.
Directions for Future Work
Multiple avenues for further development are identified:
- Extension to Broader Hybrid Classes: Generalizing to piecewise-affine systems with guards and nontrivial discrete dynamics.
- Data-Driven Algorithms: Integrating noise robustness, finite sampling, and statistical estimation procedures.
- Connections to Geometric System Theory: Investigating properties of system moduli spaces, metrics, and geometric identification methods for LSSs.
- Numerical Methods and Complexity: Improving scalability and robustness of algorithms (e.g., low-rank approximations, specialized SVDs for sparse hybrid Hankel matrices).
Conclusion
The paper (1010.5160) makes substantial progress in the systematic and algorithmic theory of partial realization for linear switched systems using a rational formal power series approach. By extending foundational tools from classic realization theory, it provides not only structural results but also concrete algorithms of interest for system identification and model reduction in switched and hybrid systems. This work also lays the groundwork for further generalizations, both theoretical and methodological, in hybrid systems analysis and data-driven modeling.