In control theory and formal verification, ensuring that a system operates safely and stably is paramount. To guarantee that an autonomous vehicle never crashes or a chemical reactor never overheats, engineers construct mathematical proofs using certificates—such as Lyapunov functions for stability or Control Barrier Certificates (CBCs) for safety.
Finding these certificates for nonlinear systems is notoriously difficult. The most successful modern framework for automating this search relies on Sum-of-Squares (SOS) programming. SOS optimization translates the intractable problem of proving a function is globally non-negative into a convex Semidefinite Program (SDP), which can be solved efficiently using standard numerical toolboxes.
However, this powerful pipeline suffers from a fundamental restriction: the polynomial dynamics constraint. As highlighted by Han Wang et al., standard computational frameworks for safety verification require both the system dynamics and the safety sets to be expressed strictly as polynomial functions to remain computationally tractable.
Why the Algebraic Link Breaks
To understand why SOS programming demands polynomials, we must look at how it evaluates a system's behavior over time. Consider a continuous-time nonlinear system governed by the differential equation:
$$\dot{x} = f(x)$$
To prove that a candidate barrier certificate $V(x)$ certifies safety, we must show that its value never decreases along the system's trajectories when approaching an unsafe boundary. Mathematically, this requires evaluating the directional derivative (or Lie derivative) of $V(x)$ with respect to the vector field $f(x)$:
$$\dot{V}(x) = \nabla V(x) \cdot f(x) \leq 0$$
If we choose our certificate $V(x)$ to be a polynomial, its gradient $\nabla V(x)$ is also a polynomial. If the underlying system dynamics $f(x)$ are also polynomial, then their dot product $\dot{V}(x)$ remains a polynomial. At this point, real algebraic geometry tools (such as Putinar’s Positivstellensatz) can be deployed to guarantee that $-\dot{V}(x)$ is non-negative by decomposing it into a sum of squared polynomials.
But physical reality is rarely purely algebraic. Real-world systems routinely feature transcendental functions ($\sin\theta$, $\cos\theta$), exponential growth ($e^t$), logarithmic relationships, or non-smooth switches. As V. Cerone et al. point out, polynomial representations cannot natively handle these transcendental terms.
The moment $f(x)$ contains a non-polynomial nonlinearity like $\sin(x)$, the derivative $\dot{V}(x) = \nabla V(x) \cdot \sin(x)$ ceases to be a polynomial. The algebraic link breaks, and the standard SOS-to-SDP translation completely collapses.
The Hidden Costs of Polynomial Approximations
Faced with non-polynomial systems, researchers traditionally resort to two main workarounds, both of which introduce severe trade-offs.
1. Taylor Expansions and Preprocessing
The most common workaround is to replace transcendental functions with their polynomial approximations. As noted by Angeliki Kamoutsi et al. and Hanrui Zhao et al., non-polynomial terms are frequently preprocessed using Taylor series expansions to force them into an algebraic mold.
The danger here is conservatism and error. He Yin et al. warn that approximating non-polynomial physical systems introduces unmodeled truncation errors. If the approximation errs even slightly, the resulting safety certificate may validate a model that is safe on paper but dangerously unstable in reality.
2. State-Space Expansion via Auxiliary Variables
Another common algebraic trick is to introduce auxiliary variables to "lift" the system into a higher-dimensional polynomial space. For example, a state variable representing an angle $\theta$ inside a trigonometric function can be replaced by two new variables, $s = \sin\theta$ and $c = \cos\theta$, subject to the polynomial constraint $s^2 + c^2 = 1$.
While mathematically elegant, Yixuan Wang et al. emphasize that this strategy drastically expands the state space. Because the number of coefficients in an SOS polynomial grows exponentially with the number of variables, adding auxiliary states rapidly triggers the curse of dimensionality, rendering the underlying SDP optimization computationally impossible for complex systems.
Transcending the Polynomial Constraint
To bypass these limitations, a major push in contemporary control research focuses on developing verification methods that can handle non-polynomial dynamics directly without losing formal guarantees.
┌────────────────────────────────────────────────────────────────────────┐
│ Non-Polynomial Dynamics │
│ (sin θ, e^x, ln x, etc.) │
└───────────────────────────────────┬────────────────────────────────────┘
│
┌──────────────────────────┼──────────────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌──────────────────┐ ┌───────────────────┐
│ Infinite-Dim │ │ Neural Barrier │ │ Data-Driven & │
│ SOS Kernels │ │ Certificates │ │ Density Functions │
├─────────────────┤ ├──────────────────┤ ├───────────────────┤
│ Maps dynamics │ │ Uses neural net │ │ Bypasses explicit │
│ to reproducing │ │ expressivity; │ │ vector fields via │
│ kernel Hilbert │ │ verified via │ │ trajectories or │
│ spaces. │ │ formal logic. │ │ dual formulations.│
└─────────────────┘ └──────────────────┘ └───────────────────┘
Infinite-Dimensional Sum-of-Squares
Rather than forcing the system into a finite polynomial space, Eloïse Berthier et al. introduce an infinite-dimensional SOS framework. By leveraging kernel methods and Reproducing Kernel Hilbert Spaces (RKHS), this approach allows optimization over much broader, non-polynomial function spaces. This enables the direct analysis of smooth, non-algebraic dynamics with significantly better computational efficiency and zero truncation error.
Neural Barrier Certificates
Another alternative is to abandon purely algebraic certificates altogether. Andrea Peruffo et al. demonstrate that neural networks can be trained to act as barrier certificates. Because neural networks are universal approximators, they possess the flexibility to match the complex, transcendental contours of non-polynomial systems. The resulting neural certificate is then verified using formal logic and satisfiability modulo theories (SMT) solvers to maintain absolute safety guarantees.
Data-Driven and Dual Formulations
Other paradigms sidestep the polynomial vector field restriction by changing how the system is modeled. Instead of searching for standard barrier functions, Jian Zheng et al. utilize density functions—a dual formulation of Lyapunov safety that tracks the flow of trajectories. Concurrently, Arash Bahari Kordabad demonstrates that model-free reinforcement learning techniques can be used to verify stability and synthesize storage functions for complex systems containing logarithmic or exponential functions, bypassing the need for an explicit algebraic model entirely.
Key Foundational Work
Residual Generator Design for Non-Linear, Polynomial Systems - A Gröbner Basis Approach (2000) — contributed 2 concept elements to this direction.