Technical Summary: PolarNet: Single-Minima Neural Network for Modeling Lyapunov Functions

1. Problem Statement

Learning control strategies with provable stability guarantees using neural networks remains a significant challenge. While recent "neural Lyapunov control" methods have shown promise by jointly learning a controller $u_\theta$ and a Lyapunov function $V_\theta$ to satisfy stability criteria, these approaches suffer from a specific training-time failure mode.

The core issue identified is that existing neural network architectures, while capable of modeling positive-definite functions, do not structurally guarantee that the learned Lyapunov function possesses exactly one critical point (the equilibrium at the origin). For a valid Lyapunov function $V_\theta$ in an asymptotically stable system, the gradient must satisfy $\nabla V_\theta(x) \neq 0$ for all $x \neq 0$. However, standard flexible architectures (e.g., MLPs) can learn functions with spurious local minima or saddle points during training.

This architectural deficiency leads to a specific failure scenario:

1. The network learns a $V_\theta$ with multiple local minima (e.g., at points $P$ and $Q$).
2. The controller $u_\theta$ is trained to minimize the empirical risk of $\dot{V}_\theta \geq 0$. If the system dynamics allow the state to settle at a spurious equilibrium (where $\dot{x}=0$ and $\nabla V_\theta=0$), the loss term vanishes.
3. Consequently, the training process stalls, and the synthesized controller fails to ensure global asymptotic stability, even though the empirical loss suggests convergence.

Existing mitigation strategies, such as adding loss margins or actively searching for counterexamples, alleviate but do not fundamentally resolve this issue, as they rely on optimization heuristics rather than structural guarantees.

2. Methodology: PolarNet

To address this, the authors propose PolarNet, a neural network architecture designed to structurally enforce the "single-critical-point" property.

2.1 Theoretical Foundation

The method relies on a key mathematical insight: any valid Lyapunov function (specifically, a "single pole function") for an asymptotically stable equilibrium can be represented as the squared Euclidean norm of a smoothly invertible mapping.

• Definition 3 (Single Pole Function): A function $V(x)$ is a single pole function if it is $C^\infty$ smooth, positive definite ($V(0)=0, V(x)>0$ for $x \neq 0$), has exactly one critical point at $x=0$, and is proper (sublevel sets are compact).
• Theorem 1: For any single pole function $V(x)$ with a non-degenerate critical point at the origin, there exists a $C^\infty$ smooth, invertible, zero-crossing vector-valued function $\Psi(x)$ such that $V(x) = \|\Psi(x)\|^2$.
• Theorem 2: If $\Psi(x)$ is a $C^\infty$ smooth, zero-crossing, and invertible mapping, then $V_\theta(x) = \|\Psi(x)\|^2$ is guaranteed to be a single pole function, regardless of the specific parameter values (provided they are finite). This ensures $\nabla V_\theta(x) \neq 0$ for all $x \neq 0$.
• Theorem 3: For any asymptotically stable equilibrium of a locally Lipschitz dynamic system, there exists a Lyapunov function that is also a single pole function. This establishes that PolarNet is theoretically capable of approximating the necessary Lyapunov function for any such system.

2.2 Network Architecture

PolarNet implements the mapping $\Psi(x)$ using Affine Coupling Flows, a type of Invertible Neural Network (INN).

• Structure: The network consists of $L$ layers where each layer $h^{(l)}$ partitions the input $y$ into two parts ($y_{\leq k}$ and $y_{>k}$). One part is passed through unchanged, while the other is transformed via an affine transformation scaled and shifted by neural networks $f_s$ and $f_t$ acting on the first part.
• Invertibility: The architecture is designed to be mathematically invertible by construction.
• Zero-Crossing Constraint: To ensure $\Psi(0)=0$ (required for $V(0)=0$), the translation functions $f_t$ in each layer are implemented as Multi-Layer Perceptrons (MLPs) without bias terms and using zero-crossing activation functions (tanh).
• Smoothness: All activation functions are smooth (tanh), ensuring the resulting $V_\theta$ is $C^\infty$ smooth.

The final Lyapunov function is computed as $V_\theta(x) = \|\Psi(x)\|^2$.

3. Key Contributions

1. Identification of a Failure Mode: The paper identifies and characterizes a specific class of training-time behaviors in neural Lyapunov control where the lack of structural guarantees leads to spurious equilibria and training stagnation.
2. PolarNet Architecture: The proposal of a novel network architecture that structurally enforces the single-critical-point property, ensuring that the learned function is a valid candidate for a Lyapunov function (satisfying $\nabla V \neq 0$ for $x \neq 0$) by design.
3. Theoretical Guarantees:
   • Proof that PolarNet always produces a single pole function.
   • Proof of universality: PolarNet can approximate any single pole function (and thus any valid Lyapunov function for asymptotically stable systems) to arbitrary precision.
4. Drop-in Replacement: Demonstration that PolarNet can be used as a direct replacement for existing network backbones in neural Lyapunov control frameworks without altering the training objective or algorithm.

4. Experimental Results

The authors conducted numerical experiments to validate the theoretical claims:

• Function Approximation (Single Pole Property):

  • The authors trained PolarNet and baseline architectures (MLP, Lyapunov Net, Wei et al.) to fit scalar fields with multiple local minima.
  • Result: Baseline models successfully fitted the multi-minima functions, thereby violating the single-critical-point constraint. PolarNet, by contrast, failed to fit the multi-minima shapes (as expected by its structural constraints) but successfully maintained the single-critical-point property in all cases. This confirmed that PolarNet cannot produce spurious local minima.

• Controller Synthesis:

  • A challenging nonlinear control problem (a 2D system with unknown dynamics) was used to synthesize a stabilizing controller.
  • Baseline: Using a standard architecture (as in [21]), the training failed to stabilize the system; the learned Lyapunov function developed multiple local minima, causing the controller to get stuck at spurious equilibria.
  • PolarNet: Replacing the Lyapunov network with PolarNet allowed the training to successfully converge. The resulting controller stabilized the system across the entire state space, and the learned Lyapunov function maintained a unique minimum at the origin.

5. Significance and Claims

The paper claims that PolarNet offers a principled and practical solution to a previously overlooked failure mode in neural Lyapunov control. By shifting the burden of stability guarantees from the optimization process to the network architecture itself, PolarNet ensures that the learned Lyapunov function adheres to the necessary mathematical conditions for stability.

The authors position PolarNet as the first architecture to structurally enforce the single-critical-point property. They emphasize that this approach is distinct from heuristic training strategies (like counterexample generation) because it fundamentally prevents the network from learning invalid Lyapunov shapes.

6. Limitations

The authors explicitly note the following limitations:

• System Class: The theoretical guarantees assume the dynamic system is locally Lipschitz. For non-Lipschitz systems, PolarNet may still be applied, but the guarantee that it can represent a valid Lyapunov function (Theorem 3) does not hold.
• Scope: The method is designed for continuous-time systems and Lyapunov stability. It does not currently address discrete-time systems or other stability definitions.
• Task Specificity: PolarNet is optimized for synthesizing controllers for a single desired equilibrium. It is not designed for tasks requiring the identification of multiple stable equilibria, where its structural constraints might limit expressiveness.