Imagine a dynamical system that abruptly changes its behavior over time. A classic example is a networked control system where communication delays fluctuate dynamically, altering the system's governing equations (Performance Oriented Control over Networks - Switching Controllers and Switched Time Delay -). A more recent example from deep learning is the Mamba architecture, which dynamically alters its state-transition matrices based on input tokens (A Control-Theoretic View of Mamba on Stability and Robustness).
Even if every individual mode or configuration of such a system is perfectly stable on its own, rapid or adversarial switching between them can cause the system's state to diverge to infinity. To guarantee that a system remains stable under arbitrary switching, control theorists look for a Common Lyapunov Function (CLF). A CLF acts as a generalized energy function: if a single mathematical function decreases along the trajectories of all constituent subsystems, the entire switched system is guaranteed to be stable regardless of how frequently or unpredictably it transitions between modes.
The Geometric Trap of Quadratic Stability
Historically, the most common way to find a CLF was to restrict the search to quadratic forms, $V(x) = x^T P x$, where $P$ is a symmetric positive-definite matrix. This property, known as quadratic stability, ensures that a single quadratic function applies across an entire set of system matrices (On Padé approximations, quadratic stability and discretization of switched linear systems).
Geometrically, a quadratic Lyapunov function defines an ellipsoid in the state space. For the system to be certified as stable, this single ellipsoid must strictly shrink under the action of every single subsystem mode.
While computationally convenient because it reduces to solving Linear Matrix Inequalities (LMIs), this approach introduces severe conservatism. As explained by Pablo A. Parrilo et al. in Approximation of the joint spectral radius using sum of squares, many perfectly stable switched systems simply do not possess a common shrinking ellipsoid. Forcing a system to satisfy uniform quadratic matrix criteria often fails to validate systems that are stable under specific switching signals (Theoretical analysis of balanced truncation for linear switched systems). As Minkee Park argues in Nonquadratic Stability Condition of Continuous Fuzzy Systems, this requirement is frequently too stringent for complex nonlinear or fuzzy systems, creating an artificial bottleneck in stability verification.
Unlocking Flexibility with Sum of Squares Polynomials
To overcome the limitations of quadratic functions, researchers have shifted toward more flexible algebraic representations. Instead of forcing the energy landscape to be a rigid ellipsoid, what if it could be a more complex, higher-degree polynomial curve?
This is where Sum of Squares (SOS) programming becomes transformative. A polynomial $V(x)$ is an SOS if it can be written as a sum of squared polynomials, $V(x) = \sum f_i(x)^2$, which automatically guarantees that $V(x) \ge 0$ for all $x$. By relaxing the search from quadratic functions to higher-degree SOS polynomials, we can fit much more intricate energy landscapes to our switched systems.
The reason this works computationally is rooted in convex optimization. As shown by Amir Ali Ahmadi et al. in Sum of Squares Certificates for Stability of Planar, Homogeneous, and Switched Systems, verifying the stability of switched linear systems under arbitrary switching is equivalent to the existence of a common SOS Lyapunov function. This equivalence allows practitioners to use automated semidefinite programming to definitively certify stability.
Furthermore, the authors established a powerful universality result in Converse results on existence of sum of squares Lyapunov functions: if a switched linear system is asymptotically stable under arbitrary switching, a common homogeneous SOS polynomial Lyapunov function is guaranteed to exist, provided the polynomial degree is allowed to increase sufficiently. This eliminates the conservatism of the quadratic approach while maintaining a tractable computational path via semidefinite optimization.
Alternative Paths to Reduce Conservatism
Beyond SOS polynomials, other non-quadratic representations offer alternative ways to mitigate conservatism depending on the system's structure or switching constraints:
- Piecewise Quadratic Functions: When switching is not completely arbitrary but constrained by a minimum "dwell time" (the minimum time a system must remain in one mode before switching), the authors of Stability of switched linear systems under dwell time switching with piece wise quadratic functions demonstrate that piecewise quadratic functions can capture stability properties that simpler quadratic functions fail to represent.
- Symmetric System Exploitation: For specific classes of systems where the constituent matrices are symmetric, Chih‐Chung Chen et al. show in Performance oriented control over networks: Switching controllers and switched time delay that a structured common Lyapunov function can be derived directly to guarantee stability across any sequence of system switches, optimizing network resource costs without sacrificing safety.
Key Foundational Work
Performance Oriented Control over Networks - Switching Controllers and Switched Time Delay - (2006) — contributed 1 concept elements to this direction.
Nonquadratic Stability Condition of Continuous Fuzzy Systems (2003) — contributed 1 concept elements to this direction.