Lyapunov Neural Networks

• Lyapunov neural networks are frameworks that integrate Lyapunov functions into network design to ensure stability and convergence.
• The models employ varied methodologies including recurrent relaxations, ICNN-based parameterizations, and PINN approaches to certify regions of attraction.
• These techniques are applied in safe control, associative memory, and dynamical diagnostics to enhance performance and provide formal stability guarantees.

Searching arXiv for recent and foundational papers on Lyapunov neural networks and related methods. Use the arXiv tool to search for "Lyapunov neural network control Zubov PINN". Lyapunov neural networks are neural-network models, learning frameworks, or verification pipelines in which a Lyapunov function, energy function, or Lyapunov exponent is not merely an analytical afterthought but a primary organizing object for the network’s dynamics, training rule, or stability certificate. In the literature, the term spans several technically distinct constructions: recurrent networks whose neural state relaxes toward attractors under a decreasing energy; feedforward or recurrent architectures whose weights evolve under Lyapunov-governed plasticity; neural approximators of control Lyapunov functions and regions of attraction; physics-informed neural networks trained on Lyapunov or Zubov partial differential equations; and training methods that use Lyapunov exponents to characterize contraction, chaos, or operation at the edge of chaos (Stern et al., 2020, Chang et al., 2020, Mittra, 2024).

1. Historical roots and the expansion of the concept

The classical antecedent is the Hopfield network. In that setting, the network state is a vector of binary neuron activities, the recurrent connectivity is symmetric, and an explicit energy function decreases whenever neurons update. Because the energy is bounded below and decreases monotonically, trajectories settle into local minima, which are fixed points of the dynamics. The stored memories are precisely these attractors, so associative recall is implemented as dynamical relaxation toward a stable state correlated with the initial condition (Stern et al., 2020).

This attractor-based interpretation gave Lyapunov methods a computational role rather than a purely analytical one. The network state moves over an “activity landscape,” and convergence follows from monotonic descent on a scalar function. That picture remains foundational because it identifies a direct link between computation, stability, and geometry of phase space: memories correspond to basins of attraction, and retrieval is the downhill flow into a basin minimum (Stern et al., 2020).

A later development extends the same organizing idea to representation learning and classification. In the Krotov–Hopfield framework, the architecture is a three-layer feedforward network with an input layer, a hidden layer, and an output perceptron. The hidden-layer weights are not learned by standard backpropagation; instead, they follow a biologically plausible, Hebbian-like local plasticity rule with normalization constraints. Here the Lyapunov function is defined on the connectivity weights rather than on the neural state. The dynamics therefore do not descend to a memory minimum; they monotonically change the Lyapunov function until a local maximum is reached, yielding a stable connectivity configuration that increases class separability in the hidden representation (Stern et al., 2020).

This historical trajectory broadens the meaning of “Lyapunov neural network.” It no longer denotes only a recurrent memory network. It also includes neural systems in which stability is imposed on weight dynamics, certificate networks in control, and optimization procedures analyzed through Lyapunov exponents. A common misconception is to equate the field solely with neural networks that output a Lyapunov certificate. The published literature is materially broader: it includes memory retrieval, classification, adaptive control, region-of-attraction estimation, symbolic discovery, and chaos-aware training (Stern et al., 2020, Benati et al., 15 Jun 2025).

2. Core mathematical structures and architectural patterns

A recurring design objective is to guarantee positive definiteness, a unique equilibrium minimum, or bounded level-set geometry by construction rather than by penalty terms alone. One early construction defines a neural Lyapunov candidate as

$v_\theta(x)=\phi_\theta(x)^\top \phi_\theta(x),$

with the network structured so that $\phi_\theta(x)=0$ if and only if $x=0$. This is enforced through layerwise weight matrices with trivial nullspaces and activations such as $\tanh$ and leaky ReLU that are Lipschitz and have trivial nullspaces. Under these constraints, the resulting candidate is positive-definite and locally Lipschitz (Richards et al., 2018).

A different construction, “Lyapunov-Net,” hard-codes positive definiteness through

$$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$

Because $V_\theta(0)=0$ and $V_\theta(x)\ge \bar{\alpha}\|x\|>0$ for $x\neq 0$, training can focus primarily on the orbital derivative condition. In that formulation, the empirical risk is essentially a single-term penalty on violations of $DV_\theta(x)\cdot f(x)<0$, which reduces hyper-parameter burden relative to multi-term objectives that separately enforce positivity and decay (Gaby et al., 2021).

High-dimensional approximation motivates compositional constructions. For nonlinear ODEs admitting a compositional Lyapunov function

$V(x)=\sum_{i=1}^s \widehat V_i(z_i),$

with the state partitioned into low-dimensional blocks, a deep network can be organized so that each subnetwork approximates one subsystem contribution. Under this assumption, and under a nonlinear small-gain condition, the number of neurons needed for fixed approximation accuracy grows only polynomially in the full state dimension. The same work also considers an unknown subsystem decomposition and introduces a lower linear layer that learns a coordinate transform before the upper nonlinear layer approximates the compositional structure (Grüne, 2020).

Several later architectures add stronger geometric bias. One ICNN-based neural Lyapunov parameterization is

$\phi_\theta(x)=0$0

where $\phi_\theta(x)=0$1 is a continuously differentiable invertible map, $\phi_\theta(x)=0$2 is an input-convex neural network, $\phi_\theta(x)=0$3, and $\phi_\theta(x)=0$4 is a smoothed ReLU-type function. The shift by $\phi_\theta(x)=0$5 and the quadratic term bias the model toward positive semidefiniteness, while the ICNN encourages convexity and helps avoid unwanted local optima (Feng et al., 2024).

For piecewise-linear verification, another architecture builds the Lyapunov function from monotone units over half-spaces in different directions. The resulting network satisfies non-negativity, a unique global minimum, star-convex nested level sets, and radial unboundedness by construction, while remaining exactly representable in mixed-integer linear programming form (Wang et al., 2024). Parameter-dependent settings introduce yet another pattern:

$\phi_\theta(x)=0$6

which makes positivity automatic for nonlinear parameter-varying systems with scheduling parameters and input constraints (Niloy et al., 17 Mar 2026).

Across these designs, the architectural question is not merely approximation power. It is how to embed Lyapunov structure directly into the hypothesis class so that positivity, uniqueness of the equilibrium minimum, or favorable level-set geometry are not left entirely to optimization. This suggests that, in Lyapunov neural networks, representational bias is part of the certificate itself rather than only a regularization choice.

3. Learning, verification, and region-of-attraction synthesis

A central branch of the literature learns Lyapunov functions jointly with controllers and then verifies the Lyapunov inequalities over a domain. In continuous-time control, the basic conditions are

$\phi_\theta(x)=0$7

with $\phi_\theta(x)=0$8 the closed-loop vector field. “Neural Lyapunov Control” parameterizes both the feedback controller and a differentiable neural Lyapunov candidate, optimizes a Lyapunov risk over sampled states, and then uses a falsifier based on delta-complete SMT solving. If the negation of the Lyapunov conditions is unsatisfiable, the controlled nonlinear system is provably stable on the verified domain; in the reported benchmarks, the learned neural Lyapunov controllers obtain ROAs 300% to 600% larger than LQR (Chang et al., 2020).

A formally sound synthesis variant makes the Lyapunov function itself a feedforward neural network with polynomial activations and no additive biases, so that $\phi_\theta(x)=0$9 holds structurally. A CEGIS loop alternates a numerical learner with an SMT verifier over nonlinear real arithmetic. The verifier searches for

$x=0$0

and counterexamples are reincorporated into training. This line of work emphasizes formal soundness rather than merely numerical plausibility (Abate et al., 2020).

Unknown-system settings add a learned model of the plant. A two-network framework first approximates the unknown nonlinear dynamics with a shallow feedforward network, then learns a neural Lyapunov function and controller using the learned dynamics, and finally verifies the candidate with dReal. The method derives an explicit error bound linking the learned dynamics to the true system so that negativity of the Lie derivative can transfer from the approximation to the actual plant under stated Lipschitz and approximation conditions (Zhou et al., 2022).

Discrete-time systems require a different verification logic. One approach introduces an $x=0$1-stability notion outside a small $x=0$2 ball around the origin, represents both the controller and Lyapunov function with ReLU networks, and verifies the positivity and one-step decrease conditions with MILP. The same framework computes verified invariant sublevel sets, which serve as ROA certificates. On path tracking, the reported method outperforms recent neural Lyapunov control baselines by an order of magnitude in both running time and the size of the region of attraction, and on cartpole and PVTOL it is reported as the first automated approach to return a provably stable controller (Wu et al., 2023).

A second major branch casts Lyapunov theory as a PDE problem. Zubov-based PINN methods replace inequality-only training by residual minimization for

$x=0$3

with $x=0$4 and $x=0$5 near the boundary of the domain of attraction. This makes the learned function directly informative about the domain of attraction rather than only local stability. One paper provides viscosity-solution theory, approximation-error bounds, and SMT-verifiable sufficient conditions for learned neural Lyapunov functions; on a two-machine power system, the neural Lyapunov function yielded a certified ROA covering about 82.5% of the domain of interest, versus only about 18.5% for the SOS approach (Liu et al., 2023). A related PINN framework aims at nearly maximal Lyapunov functions whose sublevel sets approach the true ROA boundary; in a reversed Van der Pol example, a three-hidden-layer, width-10 network with PINN+data achieved a verified level $x=0$6 and a verified volume of 96.31% (Liu et al., 2023).

Interpretability motivates symbolic post-processing. CoNSAL first trains a neural Lyapunov function, then uses symbolic regression to distill it into an analytical expression, and uses root finding on the symbolic candidate for falsification and refinement. The reported systems include a 2-D inverted pendulum, Van der Pol oscillator, 3-D trig dynamics, a rotating wheel pendulum, a 6-D nonlinear system, and a 6-D 3-bus power system (Feng et al., 2024).

Taken together, these verification-oriented approaches show that Lyapunov neural networks are not only approximators. They are often embedded in synthesis loops whose output is a controller, an ROA certificate, or an explicit analytical Lyapunov function.

4. Lyapunov exponents, spectra, and the analysis of training dynamics

A separate but increasingly important usage of the term concerns Lyapunov exponents rather than Lyapunov functions. In this line of work, neural-network training itself is treated as a dynamical system in parameter space. For standard gradient descent,

$x=0$7

the separation between nearby parameter trajectories is modeled as

$x=0$8

where the sign of the largest Lyapunov exponent $x=0$9 indicates contraction, neutrality, or chaos (Mittra, 2024).

A controlled XOR study compared hidden-layer activation functions by estimating the maximal Lyapunov exponent with a hybrid of Wolf’s algorithm and Kantz’s algorithm. The reported table was: Sigmoid, Lyapunov exponent $\tanh$0, final loss $\tanh$1; Linear, Lyapunov exponent $\tanh$2, final loss $\tanh$3; ReLU, Lyapunov exponent $\tanh$4, final loss $\tanh$5. The empirical claim is that more negative Lyapunov exponents correspond to faster convergence and lower final loss, and that the learning rate can move training between stable and chaotic regimes (Mittra, 2024).

“Lyapunov Learning” goes further by regularizing a neural network to operate at the edge of chaos. There the network is treated as a recurrently induced dynamical system, the Jacobian product is differentiated through a QR-stabilized Lyapunov estimate, and the loss includes a term of the form

$\tanh$6

with $\tanh$7 the largest Lyapunov exponent. The aim is not to maximize chaos but to keep the maximum exponent near zero. In a Lorenz regime-shift experiment, the reported loss ratio was $\tanh$8, described as about 96% improvement over the vanilla model in post-shift prediction error; the reported comparator ratios were Dropout $\tanh$9, L2 $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$0, L1 $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$1, and Lyapunov $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$2 (Benati et al., 15 Jun 2025).

For recurrent neural networks, the analysis can be extended from the largest exponent to the full Lyapunov spectrum. In chaotic firing-rate networks, the full spectrum yields the entropy rate

$$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$3

and the Kaplan–Yorke attractor dimension

$$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$4

The reported results show extensive chaos with a size-invariant Lyapunov spectrum, attractor dimensions much smaller than the phase-space dimension, point-symmetry of the continuous-time spectrum around approximately $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$5, and suppression of both entropy rate and attractor dimension by fluctuating input. The same paper connects Lyapunov spectra to error propagation and vanishing or exploding gradients in trained RNNs (Engelken et al., 2020).

Local finite-time analysis appears in fractional-order complex-valued BAM networks. There, model-based Lyapunov theory proves global Mittag–Leffler synchronization, while data-driven proxies estimate local instability by micro-ensemble FTLE slopes and kNN prediction-error growth. The reported empirical comparison favors the FTLE micro-ensemble as the clearest local instability signature (Muruganantham et al., 18 Dec 2025).

These exponent-based studies modify the conceptual scope of Lyapunov neural networks. Stability is no longer only a certificate on state trajectories of a controlled system; it also becomes a diagnostic on training trajectories, representation dynamics, and susceptibility to regime shifts.

5. Domains of application

The application spectrum is unusually broad. At one extreme are neural systems in which Lyapunov structure is the computation itself. Hopfield-style networks implement associative memory through convergence to attractors, while Krotov–Hopfield networks use Lyapunov-governed synaptic plasticity to learn hidden representations that support classification, with supervised training only at the output perceptron (Stern et al., 2020).

Safe learning and nonlinear control form the largest application area. A discrete-time closed-loop formulation uses a Lyapunov neural network to learn the shape of the largest safe region from state transitions alone, without assuming a specific analytical model structure; the reported inverted-pendulum example shows a certified safe set that adapts to a non-ellipsoidal ROA better than quadratic or SOS candidates (Richards et al., 2018). Continuous-time neural Lyapunov control has been demonstrated on the inverted pendulum, path following, Caltech ducted fan, and 2-link balancing, with reported larger ROAs than LQR and SOS/SDP (Chang et al., 2020). Unknown nonlinear systems add a learned plant model to the same stabilization-and-certification loop (Zhou et al., 2022), and discrete-time neural Lyapunov control extends the paradigm to path tracking, cartpole, and PVTOL with sound MILP verification (Wu et al., 2023).

Region-of-attraction estimation in higher dimension is another prominent use case. PINN-Zubov methods report certified ROA estimates for the reversed Van der Pol equation, a two-machine power system, a 10-dimensional example, and a 20-dimensional networked Van der Pol system, with compositional verification used to scale beyond monolithic SOS/SDP synthesis (Liu et al., 2023). Parameter-dependent synthesis has been extended to nonlinear parameter-varying systems, including an inverted pendulum with one scheduling parameter and a quadrotor with three scheduling parameters; the reported PGD verification on $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$6 adversarial samples gave about 1% violation rate at tolerance $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$7, and trajectory verification showed 100% convergence over $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$8 sampled initial points in the reported experiments (Niloy et al., 17 Mar 2026).

Adaptive and stochastic control yield another interpretation of the term. In Lyapunov-based deep neural networks for stochastic nonlinear systems, multiple DNNs approximate unknown drift and diffusion-related terms, online weight-update laws are derived from a Lyapunov generator analysis, and the tracking error is shown to be uniformly ultimately bounded in probability. In the reported 5-dimensional stochastic benchmark, the controller achieved an RMS tracking error norm of 0.533 in the nominal case (Akbari et al., 2024). For Euler–Lagrange systems, a Lyapunov-based physics-informed DNN controller incorporates the skew-symmetry property of $$V_\theta(x)=|\phi_\theta(x)-\phi_\theta(0)|+\bar{\alpha}\|x\|.$$9 into the adaptation laws; on a two-link planar revolute robot, the proposed SS-LbPINN reported a 19.87% improvement in overall function approximation error relative to a physics-informed baseline without the skew-symmetric prediction error, while tracking performance remained nearly identical (Hart et al., 23 Oct 2025).

The framework has also migrated into graph learning. Lyapunov Stable Graph Neural Flow introduces a learned Lyapunov function and a projection mechanism that maps the graph neural flow into a stable space, yielding integer-order exponential stability or fractional-order Mittag–Leffler stability. The paper reports that the Lyapunov-stable variants substantially outperform base graph neural flows under attacks such as PGD, TDGIA, and MetaGIA; one representative example reports around a 24% average gain on Cora under PGD in inductive learning when F-GRAND is converted to FL-GRAND (Chu et al., 13 Mar 2026).

6. Limitations, controversies, and open directions

The field’s central tension is between expressiveness, verification strength, and scalability. Neural parameterizations can represent Lyapunov functions or controllers that are inaccessible to low-degree polynomial templates, but formal verification of large neural certificates can be expensive or limited to relatively small networks. This issue is stated explicitly in symbolic-distillation work, which motivates analytical extraction partly to avoid expensive neural-network verification (Feng et al., 2024). It also appears in SMT- and MILP-based control synthesis, where verification is a major computational bottleneck (Zhou et al., 2022, Wu et al., 2023).

Another limitation is that many demonstrations remain low-dimensional, bounded-domain, or highly structured. Lyapunov Learning explicitly notes full Jacobian construction costs of $V_\theta(0)=0$0 and QR decomposition costs of $V_\theta(0)=0$1, and reports only low-dimensional, noise-free chaotic systems (Benati et al., 15 Jun 2025). The complex-valued BAM study demonstrates the data-driven proxies on a small numerical example rather than large-scale systems (Muruganantham et al., 18 Dec 2025). The unknown-dynamics control framework is also demonstrated on low-dimensional benchmarks and depends on Lipschitz and approximation-error estimates together with a known linearization (Zhou et al., 2022).

Not all verification claims have the same status. SMT- and MILP-backed methods aim at formal guarantees. By contrast, CoNSAL’s symbolic root-finding stage is described as a numerically efficient screening procedure for closed-form candidates, not a full formal proof (Feng et al., 2024). Discrete-time methods often prove $V_\theta(0)=0$2-stability outside a small neighborhood rather than exact global asymptotic stability on an unbounded region (Wu et al., 2023). PINN-based PDE approaches are more expressive than SOS/SDP but rely on non-convex optimization and then on separate verification stages to recover formal guarantees (Liu et al., 2023).

There is also a conceptual controversy about what should count as a Lyapunov neural network. Some works use the phrase for networks that output a candidate $V_\theta(0)=0$3; others use it for controllers whose adaptive update laws are derived from a Lyapunov analysis; still others use Lyapunov exponents as training diagnostics or regularizers. This suggests that the field is best understood as a family of Lyapunov-structured neural methods rather than a single architecture class.

A plausible implication is that future progress will continue along three converging directions. One is stronger architectural priors, so that positivity, monotonicity, or invariant geometry are built into the model class. Another is scalable verification, whether through compositional certificates, PDE characterizations, or symbolic distillation. The third is dynamical diagnostics beyond simple loss curves, especially Lyapunov exponents, finite-time spectra, and edge-of-chaos criteria, which turn stability analysis into a design signal for learning itself (Gaby et al., 2021, Liu et al., 2023, Benati et al., 15 Jun 2025).