Lyapunov-Based Stability Analysis

• Lyapunov-Based Stability is a method that constructs a scalar function to certify convergence and stability of equilibria in dynamical systems.
• Modern approaches extend classical quadratic functions to neural and data-driven formulations for complex nonlinear, high-dimensional, and hybrid systems.
• Lyapunov methods enable estimation of regions of attraction and guide robust controller synthesis across stochastic, hybrid, and quantum domains.

Lyapunov-based stability is a foundational methodology in the analysis and synthesis of dynamical, control, and hybrid systems, offering a systematic approach to certifying qualitative behaviors such as convergence, boundedness, and resilience to perturbations. The core idea is to construct a scalar functional—the Lyapunov function—whose monotonicity along system trajectories certifies stability properties of equilibrium points, invariant sets, or even regions in state space. Modern developments extend classical quadratic and energy-based Lyapunov constructions to learned, data-driven, and non-quadratic forms, enabling analysis and design for nonlinear, high-dimensional, interconnected, and hybrid systems in both deterministic and stochastic settings.

1. Fundamental Theory of Lyapunov-Based Stability

Lyapunov’s direct (second) method states that for a system $\dot x = f(x)$ with $f(0) = 0$, the existence of a continuous, positive definite function $V(x)$ with $V(0)=0$, $V(x)>0$ for $x\neq0$, and $\dot V(x) = \nabla V(x)^\top f(x) < 0$ for $x\neq 0$ guarantees that $x=0$ is an asymptotically stable equilibrium. The sublevel sets $\mathcal{S}_d = \{ x: V(x)\leq d \}$ define forward-invariant regions—so-called regions of attraction (RoA)—that lower bound the set of initial conditions drawn to the origin (Huang et al., 2020).

These principles extend to various forms: quadratic Lyapunov functions ($f(0) = 0$0), non-quadratic analytic or polynomial forms, and, increasingly, parameterizations via neural networks or data-driven methods to remove limitations of traditional analytic structures. The classical Lyapunov method generalizes to input-to-state, fixed-time, stochastic, and hybrid systems by tailoring the properties required of $f(0) = 0$1 and its increments or derivatives in the context of system specifics.

2. Neural and Data-Driven Lyapunov Function Constructions

Classical constructions are often too conservative, especially for nonlinear or large-scale systems. Neural Lyapunov methods replace the analytic ansatz with a small neural network parameterization $f(0) = 0$2, trained to enforce positivity and decrease requirements empirically over a sampled design ball, and certified globally over the design region via counterexample-guided retraining (e.g., SMT solver search for falsifiers) (Huang et al., 2020). The loss function typically combines penalties for $f(0) = 0$3, $f(0) = 0$4, and $f(0) = 0$5, with margins enforced to ensure strict Lyapunov conditions.

Physics-informed approaches such as LyapInf infer a quadratic Lyapunov function $f(0) = 0$6 by fitting to measured or simulated trajectory data, minimizing the residual of the Zubov partial differential equation (PDE) which encodes the Lyapunov decrease property without explicit knowledge of $f(0) = 0$7 (Koike et al., 11 Nov 2025). Deep Lyapunov Function frameworks employ neural networks, training $f(0) = 0$8 to satisfy positivity and negative definite Lie derivative constraints over sampled state-space domains (Mehrjou et al., 2019).

The data-driven paradigm, including off-policy reinforcement learning with Lyapunov penalties, extends to stability-constrained optimal control where policies are trained to guarantee expected (in mean or almost-sure sense) decrease in a learned Lyapunov function, even off-policy, ensuring certifiable stability in closed-loop implementations (Gill et al., 11 Sep 2025, Yao et al., 2024).

3. Stability Region and Region-of-Attraction Estimation

The region of attraction (RoA) is the set of initial states whose trajectories are guaranteed to converge to the desired equilibrium. Lyapunov methods provide inner approximations to the RoA via level sets of $f(0) = 0$9 satisfying the requisite decrease conditions. Neural Lyapunov architectures, as in (Huang et al., 2020), can yield non-ellipsoidal (non-quadratic) RoA approximations, which are quantitatively demonstrated to be substantially less conservative than classical quadratic regions in networked microgrid applications: up to 55% larger certified radius in low-dimensional cases and over 30% in large systems.

In these settings, the largest certified region is extracted by minimizing $V(x)$0 on the boundary of the design ball (solving for $V(x)$1), typically via Lagrange multipliers and advanced optimization solvers. For data-driven quadratic $V(x)$2, Monte Carlo sampling identifies the maximal level set for which $V(x)$3, thus certifying global convergence for all initial conditions within this ellipsoid (Koike et al., 11 Nov 2025).

4. Generalizations: Stochastic, Hybrid, and Quantum Domains

Lyapunov-based stability generalizes to stochastic systems via Lyapunov functions (or measures) that satisfy expected decrease conditions in mean or almost-sure sense. In quantum control, Lyapunov theory employs positive-definite operators on Hilbert space, with convergence properties dictated by commutation relations with the system Hamiltonian; invariant sets and equivalence classes account for global phase invariance (Jamalinia et al., 2021).

Hybrid and delay systems leverage generalized Lyapunov functionals or function-maxima (Razumikhin, Krasovskii), with sufficient conditions incorporating the full memory arc or supremum over memory intervals, ensuring decay in both continuous flows and discrete jumps (Liu et al., 2015). For systems exhibiting Zeno behavior, categorical Lyapunov morphisms unify continuous and discrete stability criteria, providing universal Lyapunov-like theorems in a coalgebraic framework (Moeller et al., 6 Apr 2026).

5. Extensions to Control Synthesis and Reinforcement Learning

Lyapunov theory not only certifies stability but guides synthesis of robust controllers, observers, and distributed policies. In the context of reinforcement learning, Lyapunov-based constraints are embedded into actor-critic updates, ensuring mean-cost stability in distributed control by enforcing expected decrease of a global Lyapunov function aggregated from subsystem value functions or critics (Yao et al., 2024). In off-policy Lyapunov RL, such constraints are integrated with mainstream policy optimization architectures (SAC, PPO) by penalizing violation of decrease conditions, conferring formal stability certificates to data-driven policies even under data reuse (Gill et al., 11 Sep 2025).

Lyapunov-based approaches also enable trajectory tracking under feedback linearization with unbounded perturbations by guaranteeing invariance of a contracting tube surrounding the nonconstant reference trajectory (Titze et al., 3 Dec 2025). For robust control in uncertain or time-varying settings, Lyapunov redesign remains a cornerstone due to its constructive and computationally adaptive nature.

6. Conservativeness and Impossibility Certificates

Lyapunov-based analysis is inherently sufficient but not necessary. The existence of a quadratic (or poly-quadratic) Lyapunov function is strictly stronger than global uniform asymptotic stability—there exist GUAS systems with no quadratic certificate. Recent advances provide dual LMI-based theorems of alternatives: if the Lyapunov LMI is infeasible, one can construct a certificate of nonexistence for a poly-quadratic Lyapunov function, clarifying whether infeasibility stems from numerical issues or true absence of a Lyapunov certificate (Meijer et al., 2023). This duality extends to analysis of controller and observer synthesizability and directs practitioners toward richer Lyapunov function classes (e.g., higher-degree polynomial, piecewise, or convex sum-of-squares).

7. Practical Applications and Quantitative Impact

Lyapunov-based stability certification is indispensable across domains from power systems to robotics, quantum control, and stochastic processes. In power system transient stability, neural Lyapunov methods certified robustly larger security regions than quadratic benchmarks, directly translating to improved disturbance tolerance (Huang et al., 2020). In quantum systems, Lyapunov operator design enables almost global stabilization of target pure states under drift and control, with convergence properties sensitive to operator commutation (Jamalinia et al., 2021). In reinforcement learning and distributed control, Lyapunov-embedded algorithms guarantee closed-loop stability even in high-dimensional, interconnected, model-free scenarios (Yao et al., 2024, Gill et al., 11 Sep 2025). The versatility and extensibility of Lyapunov-based stability to multidomain, hybrid, and data-driven systems ensure its continued dominance as the primary method for rigorous stability certification in complex dynamical systems.