Neural Lyapunov Models for Stability Certification

• Neural Lyapunov Models are machine learning frameworks that train neural networks to represent Lyapunov functions, ensuring rigorous stability certification for nonlinear systems.
• They integrate physics-informed losses, PDE constraints, and verification tools (e.g., SMT solvers) to enforce Lyapunov conditions over complex dynamical domains.
• Empirical results demonstrate enhanced region of attraction estimates and efficient controller synthesis in applications like robotics, power systems, and safety-critical control.

Neural Lyapunov Models are a class of machine learning frameworks in which neural networks are trained to represent Lyapunov functions—scalar-valued certificates that provide rigorous guarantees of stability for nonlinear dynamical systems. Such models enable data-driven or physics-informed approaches to the verification of stability, the construction of regions of attraction (ROA), and the synthesis of controllers in otherwise analytically intractable systems. The core principle is to encode Lyapunov conditions—typically expressed as nonlinear partial differential equations (PDEs) or algebraic inequalities—directly into the parameterization, optimization, and validation of neural networks. This paradigm has been extensively developed for both autonomous and controlled systems, deterministic and stochastic models, and is applied in fields ranging from nonlinear control to power systems and deep learning.

1. Problem Formulation and Lyapunov Characterization

The underlying mathematical setting considers a continuous-time, nonlinear autonomous system,

$$\dot{x} = f(x), \quad f: \mathbb{R}^n \to \mathbb{R}^n, \quad f \;\text{locally Lipschitz}, \quad \text{eq.}\; x^*=0,$$

and seeks to establish uniform asymptotic stability of the equilibrium by constructing a Lyapunov function $V : D(A) \to [0, \infty)$ with the following properties:

• $V(x^*) = 0$, $V(x) > 0$ for $x \in D(A) \setminus \{x^*\}$
• $\dot{V}(x) = \nabla V(x) \cdot f(x) = -\Phi(x) < 0$ for $x \in D(A) \setminus \{x^*\}$ and some continuous, strictly positive $\Phi$
• $V(x) \to \infty$ as $x \to \partial D(A)$, where $D(A)$ denotes the domain of attraction.

Maximal Lyapunov function refers to a function whose sublevel sets $\{ x : V(x) \le c \}$ can be made arbitrarily close (in the Hausdorff metric) to $D(A)$; in effect, such a function precisely characterizes the true region of attraction (Liu et al., 2023).

Zubov's theorem provides a PDE characterization of the domain of attraction using a bounded, monotonic transformation $W(x) \in [0,1)$ of $V(x)$ solving,

$$\nabla W(x)^\top f(x) = -\Psi(x)(1-W(x)), \quad W(x^*)=0, \quad W(x) \to 1 \text{ as } x\to \partial D(A),$$

with $\Psi > 0$ (Liu et al., 2023, Liu et al., 2023).

For control-affine or stochastic systems, analogous constructions constrain a neural Lyapunov function or a control Lyapunov function (CLF) to provide certified decrease under the closed-loop flow or sample paths (Chang et al., 2020, Lawrence et al., 2021, McCutcheon et al., 19 Mar 2025).

2. Neural Parameterizations and Model Construction

Neural Lyapunov Models instantiate the Lyapunov function $V(x; \theta)$ as a feed-forward neural network, with various architectural choices enforcing positive definiteness and smoothness:

• Feed-forward Networks (tanh/softplus/ReLU): $V(x) = NN_\theta(x)$ or $V(x) = |NN_\theta(x)| + \epsilon$, typically with $L$ layers and width $w$ (Liu et al., 2023, McCutcheon et al., 19 Mar 2025).
• Structured/Compositional Networks: Partitioned to reflect system decomposability, e.g.,

$$V_\theta(x) = \sum_{i=1}^s W_i(z_i; \theta_i), \quad z_i \in \mathbb{R}^{d_i},$$

to exploit subsystem or network structure (Grüne, 2020).

• Hybrid Quadratic–Neural Networks: $V(x) = x^\top P x + N_\theta(x)$, capturing local quadratic behavior near the equilibrium and global higher-order terms (Barreau et al., 2024).
• Input-Convex Neural Networks (ICNNs): Convex by architecture, supporting global stability verification (Lawrence et al., 2021, Mukherjee et al., 2022).
• PLNet-based Models: $\displaystyle H(x) = \frac{1}{2} \|g(x)\|^2$ with $g$ bi-Lipschitz, providing quadratic sandwich bounds and ensuring uniqueness and non-degeneracy of the Lyapunov function's minima (Cheng et al., 2024).

Network architectures are often further constructed to ensure unique global minima, radial unboundedness, and to enforce $V(0) = 0$ either by design or by architectural choice (Wang et al., 2024, Gaby et al., 2021, Richards et al., 2018).

3. Physics-Informed Losses, PDEs, and Optimization Schemes

Training of neural Lyapunov models is governed by composite loss functions encoding the Lyapunov decrease and positive definiteness conditions:

• PDE Residual Loss: For the Zubov or classical Lyapunov equations, the main loss enforces,

$$L_r(\theta) = \frac{1}{N_r} \sum_{x_i} | \nabla_x W_\theta(x_i) \cdot f(x_i) + \Psi(x_i) (1 - W_\theta(x_i)) |^2$$

(Liu et al., 2023, Liu et al., 2023).

• Boundary/Positivity Loss: Penalties for $V(0) \ne 0$, enforcing $V(x)\ge0$ and $V(0)=0$ via direct regularization (McCutcheon et al., 19 Mar 2025, Gaby et al., 2021).
• Data-driven/Hybrid Loss: Combine physics-based PDE residuals with trajectory-matching or supervised data when available (Liu et al., 2023, McCutcheon et al., 19 Mar 2025, Barreau et al., 2024).

Optimization employs stochastic gradient descent (Adam), possibly with primal-dual steps to enforce constrained decay ($\dot V < 0$) via Lagrange multipliers, and may include projection (e.g., for $P \succ 0$ in quadratic terms) or bi-level routines for certified min–max optimization (Barreau et al., 2024, Wang et al., 2024).

For controlled dynamical systems, policy and Lyapunov function parameters are learned jointly to optimize a task-reward objective regularized by the Lyapunov risk (Mukherjee et al., 2022, McCutcheon et al., 19 Mar 2025, Chang et al., 2020).

4. Verification: SMT Solvers, MILPs, and Counterexample-driven Synthesis

Formal and approximate soundness of neural Lyapunov models is pursued via automated verification frameworks. Satisfiability modulo theories (SMT) solvers (e.g., dReal) or mixed-integer linear programs (MILPs) are used to verify the Lyapunov decrease and positivity over the whole domain or to falsify candidates:

• Local ROI Certification: Linearization-based invariance conditions, e.g., $f(x) = Ax + g(x)$, verify quadratic Lyapunov invariance on sublevel sets (Liu et al., 2023, Liu et al., 2023).
• Global ROA Certification: Verified via checking that $\{\nabla V(x) \cdot f(x) < 0\}$ on sublevel boundaries of the learned Lyapunov network (Abate et al., 2020, Liu et al., 2023, Wang et al., 2024).
• Counterexample-Guided Loop: The learner alternates between neural network parameter updates (minimizing the empirical Lyapunov risk on collected samples) and symbolic falsification (querying the global domain for violations of the Lyapunov conditions); newly found counterexamples augment the training set (Abate et al., 2020, Chang et al., 2020, Huang et al., 2020).
• ROA Enlargement: For discrete-time and control settings, the certified region of attraction is systematically expanded (e.g., via level-set or inscribed cube maximization, solved as an MILP) (Wang et al., 2024, Mittal et al., 2020).

This process yields provably certified neural Lyapunov functions and state-space regions for safety-critical control and verification tasks.

5. Theoretical Guarantees and Approximation Properties

Existence and uniqueness of maximal Lyapunov functions and viscosity solutions to (Zubov-)PDEs are guaranteed under mild regularity assumptions (local Lipschitz continuity, positive-definite dissipation) (Liu et al., 2023, Liu et al., 2023). For compositional architectures, the number of neurons required to achieve $\varepsilon$-accurate Lyapunov approximation scales

$N = O(n^{d_{\max}+1} \varepsilon^{-d_{\max}})$

where $d_{\max}$ is the largest subsystem dimension, resulting in polynomial scaling with the system state dimension under block-decomposability (small-gain property) (Grüne, 2020).

For general models, density results prove that the specific neural architectures (e.g., RePU, monotonic ReLU, hybrid quadratic-MLP) form universal approximators for the class of smooth Lyapunov functions, allowing arbitrary accuracy within the compact region (Gaby et al., 2021, Barreau et al., 2024). Error bounds for neural-PDE PINN models guarantee that, under suitable empirical risk minimization, the neural function converges uniformly to the true PDE solution (Liu et al., 2023).

Termination and soundness of the counterexample-driven loops are ensured for $\delta$-complete solvers and bounded domains, yielding certificates of asymptotic stability (Abate et al., 2020).

6. Empirical Performance, Practical Implementation, and Applications

Neural Lyapunov models have been numerically demonstrated across a range of paradigmatic systems:

• ODE Benchmarks: Van der Pol oscillator, double integrator, unicycle, networked microgrids, damped pendula (Liu et al., 2023, Liu et al., 2023, Huang et al., 2020, Cheng et al., 2024).
• High-dimensional Nonlinear Systems: Up to $n=10$ decoupled/coupled or $n=30$ synthetic state dimensions, with favorable scaling for compositional/generic architectures (Grüne, 2020, Gaby et al., 2021).
• Region of Attraction (ROA) Recovery: Neural models reliably yield certified regions that are 2–3$\times$ larger (in volume) than classical quadratic/SOS certificates, e.g., 95% of the true ROA in Van der Pol, compared to 60% for standard polynomials (Liu et al., 2023, Liu et al., 2023, Huang et al., 2020).
• Sample Efficiency and Training Time: PINN+data and structured architectures achieve accurate certification with modest sample counts (e.g., 90,000 points in 2D), typically requiring several minutes of computation for low to moderate dimensions (Liu et al., 2023, Liu et al., 2023).
• Control, MPC, and RL: Jointly optimized Lyapunov functions and controllers reliably stabilize inverted pendula, path-following vehicles, and robotic manipulators, with sample efficiency and stability regions exceeding RL/SDP-based baselines (McCutcheon et al., 19 Mar 2025, Chang et al., 2020, Mittal et al., 2020, Mukherjee et al., 2022).

7. Strengths, Limitations, and Future Directions

Strengths

• Neural Lyapunov Models can learn nearly-maximal Lyapunov functions, reducing conservativeness in ROA and stability certification (Liu et al., 2023, Liu et al., 2023).
• PINN-based and compositional architectures enable structured, scalable training without explicit trajectory rollouts (Grüne, 2020).
• Certified, SMT-verified neural Lyapunov functions provide formal safety, crucial for safety-critical systems and learning-based controllers (Abate et al., 2020, Wang et al., 2024).
• Flexible to model structure (ODE, controlled, stochastic) and data availability (physics-informed or data-driven) (Lawrence et al., 2021, Liu et al., 2023, McCutcheon et al., 19 Mar 2025).

Limitations

• High-dimensional scalability remains a challenge due to curse of dimensionality in sampling and verification—SMT/MILP bottlenecks can dominate as system size grows (Liu et al., 2023, Wang et al., 2024).
• Accurate approximation near equilibria and global certification in complex or stiff systems require careful design and may degrade as $n$ increases (Grüne, 2020).
• Most methods assume $f(x)$ is known or accurately learned; model uncertainty can compromise certification (Huang et al., 2020, McCutcheon et al., 19 Mar 2025).

Research Directions

• Compositional and hierarchical PINN architectures to exploit sparsity and modularity for high-dimensional problems (Liu et al., 2023).
• Rigorous approximation and finite-sample error bounds for neural PDE solvers (Liu et al., 2023, Barreau et al., 2024).
• Integration with neural policy iteration, adaptive safe RL, and online controller synthesis (McCutcheon et al., 19 Mar 2025, Chang et al., 2020).
• Extensions to barrier functions, passivity certificates, input-to-state stability, and stochastic control with formal guarantees (Cheng et al., 2024, Lawrence et al., 2021, McCutcheon et al., 19 Mar 2025).

Neural Lyapunov Models unify machine learning with nonlinear stability theory, providing both computational expressiveness and, when paired with formal verification, rigorous safety guarantees across broad classes of dynamical systems.