Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks

arXiv:2606.14174 · nlin.CD, cond-mat.stat-mech · Submitted 2026-06-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Nonlinear Sciences. The day in nonlinear science: chaos, patterns, networks and adaptation.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks".

Mira: Here's the gist of the paper. Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks introduces a generalized Lyapunov function (GLF) as a unifying framework for nonequilibrium potentials and develops a data-free physics-informed neural network (PINN) approach to construct these functions by embedding physical constraints directly into the loss function.

Kai: The resulting framework successfully recovers known analytical landscapes for various nonlinear systems, demonstrating its versatility in constructing potential descriptions for complex non-gradient dynamics where closed-form expressions are unavailable.

Mira: Generalized Lyapunov Function Definition The generalized Lyapunov function (GLF) is defined as a scalar potential that is monotonically non-increasing along deterministic trajectories, satisfying the Lyapunov inequality where the definition is given by phi˙ = ∇ϕ · f(x) ≤ 0.

Kai: This definition unifies several familiar scalar objects, including conventional Lyapunov functions, Freidlin–Wentzell quasi-potentials, and potentials obtained from the Ao decomposition.

Title and authors: Mira: So, looking at the title, "Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks," it suggests this isn't just about one specific system; they aim for a general method applicable across different nonlinear dynamics >

Kai: That’s right. It’s trying to provide a universal framework, meaning it should work regardless of whether we are dealing with linear systems or these more complex, non-gradient flows that have recurrent structures >

Lev: I wonder how much data this actually needs from the start since they claim it's data-free for construction, though they still need some physical constraints embedded in the loss function to guide the learning process >

Mira: Well, their summary says they tackle those complex invariant structures where traditional methods fail to build global potential descriptions for systems with recurrent dynamics >

Kai: That’s a big deal. Imagine trying to visualize stability and organization in something like a limit cycle or linked cycles when you have no closed-form equation for the potential that describes it >

Lev: It sounds like they are showing how this AI approach can map out regions of attraction and neutral dynamics just by enforcing the dynamical rules mathematically >

Mira: That’s what I mean. The paper demonstrates that these generalized Lyapunov functions can identify invariant sets as low- or constant-potential structures, which lets us visualize those important features in the flow >

The paper's summary: Kai: So, to recap the main thrust of this work on "Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks," it’s about using a PINN to build a scalar potential that obeys the Lyapunov inequality >

Mira: They show they can recover known analytical landscapes for systems like two-dimensional linear systems and the Hopf normal form, which is a significant check on how well the method works >

Lev: The paper also shows it produces coherent potential landscapes for things like the van der Pol oscillator and a three-dimensional Hopf-link system with two linked limit cycles without needing any known closed-form GLFs >

Kai: So what this means practically is that they’ve got a practical way to get a potential description for nonlinear non-gradient dynamics even when an explicit analytical GLF just isn't available >

Mira: And they specifically show it works for things like the asymmetric competitive Lotka–Volterra system, which is quite complex and has no simple potential form >

Lev: That capability means we can start mapping out the global organization of these systems, even when we don't have the explicit equations of motion solved analytically >

Kai: It really seems to suggest that GLFs give us a practical landscape framework for looking at how these complex invariant structures are organized globally >

The paper's improvements: Mira: One key improvement they highlight is that the PINN loss function contains three specific terms, not just one simple constraint, which helps guide the network toward a physically sensible result >

Kai: Specifically, they use the Lyapunov inequality term to penalize any points where the learned potential actually increases along the flow trajectory >

Lev: Then there’s that anchor term fixing a constant value at some point x zero, and finally, this weak divergence-scale compatibility condition which acts like a selection rule >

Mira: That compatibility condition is what allows them to select one physically meaningful representative from the whole nonunique family of admissible GLFs that the network could otherwise generate >

Kai: So they’ve essentially built a way to enforce stability without actually having to solve all the equations of motion for every single point in space, which is a huge computational saving >

Conclusion: Mira: Looking at the conclusion of "Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks," they wrap up by stating this framework provides a practical landscape description for nonlinear non-gradient dynamics with increasingly complex invariant structures >

Kai: It seems the main implication is that we can use this data-free PINN approach to investigate the global organization of these systems, and it’s complementary to symbolic-learning methods rather than competing with them >

Lev: For real hardware running this, the overhead lands in how well that weak divergence-scale compatibility condition truly selects the best potential out of all those possibilities >

Mira: And for me, I think the most important thing is that we have a way to construct these potentials directly as a continuous neural scalar field subject to dynamical constraints rather than relying on pre-defined mathematical structures >

Kai: So we’ve seen how this paper tackles the construction of generalized Lyapunov functions, and it shows that for nonlinear flows, even when closed-form expressions are missing, we can still build a visualization of their stability and organization using physics embedded directly into the AI training process >

Z. C. Tu, *

Beijing Normal University

nlin.CD, cond-mat.stat-mech

Submitted: 2026-06-12

Updated: 2026-08-18

Comments: 8 pages, 5 figures

Journal ref: Commun. Theor. Phys. 78 (2026) 125601

DOI: 10.1088/1572-9494/aeaad5

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks introduces a generalized Lyapunov function (GLF) as a unifying…

Key concepts

Generalized Lyapunov Function (GLF)
A GLF is a scalar potential that does not need to be strictly positive definite or uniquely normalized. It must be monotonically non-increasing along system trajectories, satisfying the condition phi dot f less than or equal to zero. This broad definition unifies several different types of potentials used in dynamical systems analysis.
Physics-Informed Neural Network (PINN) Construction
Instead of training on trajectory data, this method uses a neural network to represent the potential function. The loss function is designed with physical constraints—such as penalizing violations of the Lyapunov inequality—to force the network to learn a physically meaningful potential directly from dynamical rules.
Lyapunov Inequality Loss Term
This term in the PINN loss function enforces the core requirement of a GLF: phi dot f less than or equal to zero. By penalizing points where this condition is violated, the neural network learns a potential that guarantees stability along the system's deterministic paths without needing to solve complex differential equations.
Weak Divergence-Scale Compatibility Term
This term acts as a selection rule among the many mathematically possible GLFs. It helps choose a specific, physically relevant representative from the nonunique family of functions that satisfy the basic physical constraints, guiding the learning process toward a meaningful potential landscape.

Terminology

Summary

Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks introduces a generalized Lyapunov function (GLF) as a unifying framework for nonequilibrium potentials and develops a data-free physics-informed neural network (PINN) approach to construct these functions by embedding physical constraints directly into the loss function. p. 0 The resulting framework successfully recovers known analytical landscapes for various nonlinear systems, demonstrating its versatility in constructing potential descriptions for complex non-gradient dynamics where closed-form expressions are unavailable. p. 0

Generalized Lyapunov Function Definition

The generalized Lyapunov function (GLF) is defined as a scalar potential that is monotonically non-increasing along deterministic trajectories, satisfying the Lyapunov inequality where the definition is given by phi˙ = ∇ϕ · f(x) ≤ 0. This definition unifies several familiar scalar objects, including conventional Lyapunov functions, Freidlin–Wentzell quasi-potentials, and potentials obtained from the Ao decomposition. p. 1 The definition is intentionally broader than the conventional one because it does not require positive definiteness or a unique normalization. p. 1

PINN Construction Methodology

The GLF is represented by a neural network phiΘ(x), and the construction method is data-free, meaning no trajectory data or prescribed potential values are required. The loss function contains three terms:

  1. The Lyapunov inequality is imposed through LLyap = ReLU ∇ϕΘ · ff, which penalizes points where the learned potential increases along the flow. p. 1

  2. An anchor term Lzero = [ϕΘ(x0)]2 fixes the additive constant of the potential to zero by setting it at some point x0. p. 1

  3. A weak divergence-scale compatibility term Ldiv = ∇2ϕΘ + ∇ · f1 + ∇ · f2 selects a physically meaningful representative from the nonunique family of admissible GLFs. p. 0

Validation Across Diverse Systems

The framework was demonstrated on diverse systems, including linear dynamics, the Hopf normal form, the van der Pol oscillator, a three-dimensional Hopf-link flow, and symmetric and asymmetric competitive Lotka–Volterra systems. The method recovers known analytical landscapes for two-dimensional linear systems and the Hopf normal form. p. 1 For the van der Pol oscillator, the learned landscape places the limit cycle inside a low-potential valley, consistent with qualitative properties of a complete Lyapunov function. p. 2

Conclusion and Future Directions

The approach suggests that GLFs provide a practical landscape framework for investigating the global organization of nonlinear non-gradient systems with increasingly complex invariant structures. The authors contrast their PINN method with symbolic-learning methods, noting that their objective is complementary rather than competing, as they construct the GLF directly as a continuous neural scalar field subject to dynamical constraints. p. 5 Future work should investigate the mathematical mechanism by which the PINN loss selects a dynamically meaningful representative from the nonunique family of admissible GLFs. p. 5 The authors also suggest extending this method to chaotic attractors and high-dimensional systems. p. 6

Acknowledgments

The author is grateful to Ping Ao and Haiwen Liu for drawing his attention to the problem of constructing potential landscapes. This work was supported by the National Natural Science Foundation of China (Grant Nos. p. 6 12475032 and 11975050). p. 6 Some of the computational resources were kindly provided by Jing Wu. p. 6 The author used ChatGPT for language polishing, coding, and debugging assistance. p. 6 All scientific content, numerical methods, and interpretations of the results were checked and finalized by the author. p. 6

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Data-free construction of global potential landscapes for non-gradient systems: The PINN framework allows for constructing a global scalar landscape even when analytical expressions are unavailable, which enables the AI system to visualize stability and organization in complex, recurrent structures like limit cycles or linked cycles.

  2. Identification of invariant sets as low- or constant-potential structures: The learned landscapes can reveal that invariant sets are low- or constant-potential structures, allowing the AI system to map out regions of attraction and neutral dynamics for nonlinear flows without needing explicit solutions.

  3. Robust representation of non-conservative dynamics: The framework handles systems where the scalar potential obtained from the Helmholtz decomposition is generally not a suitable global landscape function, enabling analysis of dissipative and conservative components that conventional methods fail to capture globally.

  4. Construction of GLFs from physical constraints: The system enforces the Lyapunov inequality as a loss term, ensuring that the learned potential satisfies phi dot f less than or equal to zero along trajectories, which certifies stability without solving equations of motion.

  5. Selection of physically meaningful potentials: The inclusion of the weak divergence-scale compatibility condition acts as a weak selection rule, helping to select a physically meaningful representative from the highly nonunique family of admissible GLFs.

Sources

Related papers