Neural networks are increasingly deployed to control physical systems, from autonomous quadrotors to maritime vessels. However, their "black-box" nature presents a significant hurdle for safety-critical applications. Traditional control theory relies on rigorous mathematical proofs of stability, while deep learning typically relies on empirical testing.
A growing body of research bridges this gap by treating neural networks not as opaque functions, but as structured nonlinearities that can be "wrapped" in formal certificates. By combining Lyapunov stability theory with robust control tools, researchers can now prove that a neural controller will never drive a system into an unstable state.
Bounding the Nonlinearity
The primary challenge in analyzing a neural network (NN) within a feedback loop is its nonlinearity. Activation functions like ReLU or Tanh are difficult to track through multiple layers. To solve this, researchers use sector conditions and slope restrictions.
As established by Johan Suykens and colleagues in their foundational $NL_q$ theory, we can model activation functions as operators that stay within a specific "sector" or envelope. For example, a Tanh function always stays between two lines passing through the origin. By bounding the output of the NN relative to its input, the complex network begins to look like a more manageable linear system with bounded uncertainty.
He Yin, Peter Seiler, and Murat Arcak refined this in Stability Analysis using Quadratic Constraints for Systems with Neural Network Controllers by incorporating the rate of change (slope) of these functions. This extra information allows for much tighter bounds, reducing the "conservatism" of the proof—meaning we can certify larger safe operating regions without being overly cautious.
From Constraints to Certificates
Once the neural network is bounded, the next step is to prove that the entire closed-loop system—the physical plant plus the NN controller—is stable. This is typically done using Lyapunov functions, which act as a mathematical "energy" measure of the system. If we can prove this energy always decreases over time, the system is guaranteed to converge to a stable equilibrium.
The search for these Lyapunov functions is often formulated as Linear Matrix Inequalities (LMIs). In Offset-free setpoint tracking using neural network controllers, Patricia Pauli, Johannes Köhler, and Frank Allgöwer demonstrate how to use LMIs to calculate a Region of Attraction (ROA). The ROA is a "safe zone": if the system starts within this ellipsoidal region, the NN controller is mathematically guaranteed to bring it to the target setpoint with zero steady-state error.
This approach extends beyond simple controllers. Shengze Cai and Jian Chu applied similar sector-bounding techniques to Neural Observers, which estimate hidden system states, ensuring that the estimation error also converges to zero.
Scaling to High Dimensions
While LMIs work well for small systems, they struggle to scale to the high-dimensional neural networks used in modern robotics. Recent breakthroughs have shifted from general optimization solvers to specialized neural network verification tools.
In Lyapunov-stable Neural Control for State and Output Feedback, Lujie Yang, Huan Zhang, and Russ Tedrake leveraged linear bound propagation (specifically the $\alpha,\beta$-CROWN framework). Instead of solving a massive matrix inequality, they propagate bounds through the network layers to rigorously check if the Lyapunov derivative is negative. This allowed them to provide the first formal stability certificates for complex systems like a quadrotor using raw lidar data for output feedback.
Similarly, Mohamed Serry and Jun Liu introduced a framework in Safe Domains of Attraction for Discrete-Time Nonlinear Systems that uses physics-informed neural networks to solve the Zubov equation. By verifying these solutions with bound propagation, they can characterize much larger and more complex safe regions than traditional methods.
Practical Impact
These theoretical unifications mean that neural networks are no longer restricted to low-stakes tasks. For instance, Daniel Frank and Michael Muehlebach have applied these performance guarantees to Recurrent Neural Networks identifying ship motion, ensuring that the model remains stable even when predicting complex maritime dynamics.
By treating the neural network as a constrained component within a larger dynamical system, this research direction provides the mathematical "glue" needed to use deep learning in the real world with the same confidence we have in classical engineering.