When we model a physical system, we often start with its external behavior—how its outputs respond to its inputs over time. However, to control or simulate the system, we must construct an internal state-space representation. This process is known as realization theory.
A central challenge in realization theory is finding a minimal realization: a state-space model that captures the exact input-output behavior using the smallest possible number of internal state variables. In classical linear time-invariant (LTI) systems, Kalman's realization theory provides a complete answer: a system is minimal if and only if it is both controllable and observable, and any two minimal realizations of the same behavior are identical up to a unique linear change of coordinates (an isomorphism).
But what happens when the system is no longer invariant? Modern engineering increasingly relies on switched systems (which flip between different operating modes) and Linear Parameter-Varying (LPV) systems (whose dynamics change continuously based on external scheduling signals). In these regimes, a state variable might be highly visible in one mode but completely hidden in another.
A dedicated body of research has extended classical realization theory to these complex architectures, proving that despite the shifting dynamics, minimality can still be precisely characterized, and minimal representations remain fundamentally unique.
Redefining Reachability and Observability
To understand why a state-space model can be simplified, we must look at how inputs influence states (reachability) and how states influence outputs (observability). If a state cannot be influenced by any input, or if its variations never impact the output, it is redundant and can be stripped away.
For systems that switch between different linear or bilinear modes, evaluating these properties is challenging because the system's behavior depends on the history of its operational modes. As noted by Petreczky in Realization theory for linear switched systems, complex dynamical systems frequently contain redundant state variables that do not contribute to external behavior, creating unnecessary analytical complexity.
To resolve this, researchers adapted the geometric concepts of reachability and observability:
- Semi-Reachability and Span-Reachability: Instead of requiring a single input to drive the system to any state instantly, we look at the span of all states reachable across all possible switching sequences or parameter trajectories.
- Joint Conditions: For linear switched systems, Petreczky demonstrated in Realization theory for linear and bilinear switched systems: a formal power series approach that a realization achieves the lowest possible state-space dimension if and only if it is simultaneously observable and semi-reachable.
This geometric framework holds true across discrete and continuous time. For instance, in discrete-time linear switched systems, Petreczky et al. proved that the joint conditions of span-reachability and observability are both necessary and sufficient to establish minimality.
Similarly, for LPV state-space representations with affine dependence on their parameters, Petreczky et al. proved that state-minimality remains strictly equivalent to being observable and span-reachable. Under specific regularity certificates, this can simplify even further: Petreczky et al. showed that the minimality of an affine LPV system can be determined solely by evaluating its observability.
Classical LTI: Minimality <===> Controllability + Observability
Switched/LPV: Minimality <===> Span-Reachability + Observability
The Structural Invariance of Isomorphism
If two different engineers build minimal state-space models for the exact same switched or LPV behavior, will their models look completely different?
The core triumph of this research direction is proving that minimal realizations are unique up to an isomorphism. This means that while the specific state variables chosen by the engineers might differ on the surface, their models share the exact same underlying internal architecture and can be transformed into one another via a coordinate change.
In Realization theory for linear and bilinear hybrid systems, Petreczky proved that minimal hybrid realizations combining discrete automata and continuous dynamics are structurally unique up to a state-space isomorphism. This ensures that the minimal model's core architecture remains invariant regardless of how it was derived.
This structural invariance behaves in a remarkably rigid—and useful—way in parameter-varying settings. For LPV systems, one might expect that the transformation matrix needed to align two minimal representations would have to change over time alongside the scheduling variable. However, Petreczky et al. proved that all minimal affine LPV realizations of the same input-output function can be transformed into one another via a linear transformation that is completely independent of the scheduling variable. The coordinate change is static, even though the system itself is highly dynamic.
This mathematical guarantee extends well beyond linear structures:
- Behavioral Equivalence and Bisimulation: Working within a broader behavioral framework, van der Schaft demonstrated that for systems possessing a minimal canonical state representation, external behavioral equivalence is sufficient to guarantee bisimilarity, bridging the gap between external testing and deep internal structural equivalence.
- Nonlinear Nash Systems: For highly structured nonlinear systems governed by analytic semi-algebraic functions, Nêmcová et al. proved that minimal realizations are unique up to a local isomorphism, meaning they can be mapped to each other via a smooth, semi-algebraic change of coordinates within a local neighborhood.
Why It Works: Well-Posed System Identification
The fact that minimal realizations are unique up to an isomorphism is not just an abstract algebraic comfort; it is a mathematical prerequisite for robust data-driven engineering.
If you attempt to learn a system's parameters directly from input-output data (system identification), and there are infinitely many, structurally distinct minimal state-space representations that explain that data, the identification problem becomes ill-posed. The optimization algorithm can wander endlessly through equivalent but structurally incompatible models. Isomorphism bounds this search space to a single coordinate rotation.
However, when moving from deterministic models to real-world stochastic systems, minimality alone is no longer enough to guarantee a unique model. Random noise introduces extra degrees of freedom. To recover uniqueness, the model must be cast into innovation form, where the internal process noise is forced to align precisely with the external prediction error.
As shown by Rouphael et al., a stochastic LPV model must satisfy this innovation form condition to ensure that any two representations of the same data are mathematically identical under a coordinate change. Crucially, the authors proved that the minimality of these stochastic representations can be verified simply by checking their deterministic counterparts, bypassing complex stochastic proofs.
For generalized linear switched systems with inputs, Rouphael et al. similarly demonstrated that applying reachability and observability constraints to the stochastic component guarantees that the resulting innovation form is both minimal and unique up to an isomorphism.
[Input-Output Data]
│
▼ (System Identification)
[Innovation Form + Reachability/Observability Constraints]
│
▼
[Unique Minimal State-Space Model (Up to Isomorphism)]
This structural predictability also underpins model reduction and transformation. For instance, when converting an affine LPV system into a Linear Fractional Transformation (LFT) model for robust control synthesis, Petreczky et al. designed a transformation algorithm that explicitly preserves the original state-space minimality. Because minimality is preserved, the algorithm guarantees that the resulting fractional representation will lack redundant states, ensuring maximum computational efficiency.
Conversely, standard model reduction techniques like balanced truncation do not inherently preserve minimality. As noted by Petreczky et al. and Petreczky et al., truncating states in a switched system can inadvertently leave behind non-minimal, redundant representations. This highlights why realization theory remains essential: it provides the constructive algorithms needed to post-process these reduced models, stripping away hidden redundancies to reveal the true, minimal core of the system.
Key Foundational Work
Realization theory for linear switched systems (2004) — contributed 5 concept elements to this direction.
Realization theory for linear and bilinear switched systems: a formal power series approach (2005) — contributed 3 concept elements to this direction.
Realization theory for linear and bilinear hybrid systems (2005) — contributed 3 concept elements to this direction.