acceptodds
Under review as a conference paper at ICLR 2027

Zubov-Net: Anchored Neural Networks for Learning Maximal Lyapunov Functions with Formal Guarantees

Abstract

Lyapunov functions are a central tool for certifying the stability of physical and engineered systems, yet constructing them remains a long-standing computational challenge. Neural networks can learn Lyapunov functions from data, but the learned functions rarely come with formal guarantees. We introduce Zubov-Net, a neural architecture whose outputs are guaranteed to be Lyapunov functions near the equilibrium and whose sublevel sets can cover any compact subset of the region of attraction (ROA). Zubov-Net multiplies a known local Lyapunov function, the anchor, by a learned positive correction. Because this design enforces the local Lyapunov conditions for every choice of network parameters, training only has to learn the global shape of the basin from trajectory data and a well-scaled residual of Zubov's equation. We prove a universal approximation theorem for anchored networks: they can approximate a maximal Lyapunov function arbitrarily well on each of its sublevel sets. Across a range of benchmarks with two to twenty state variables, the formally verified ROA estimates of Zubov-Net are consistently, and often substantially, larger than those of the baseline certificates.

Primary area:
learning on time series and dynamical systems
Keywords:
neural Lyapunov functions, neural network verification, region of attraction, physics-informed learning, Zubov equation, certified stability

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_UHC5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Zubov-Net multiplies a local Lyapunov function by a learned positive correction, guaranteeing local Lyapunov conditions, proving universal approximation of maximal Lyapunov functions, and certifying larger verified regions of attraction.

Critical strengths

Because Zubov-Net multiplies a known local Lyapunov function (the "anchor") by a positive correction, the output satisfies the strict Lyapunov conditions near the equilibrium for every parameter setting (Prop. 1). This removes a known weak point where small approximation errors at the origin break the certificate.

A universal approximation theorem (Thm. 1) shows that anchored networks can approximate a maximal Lyapunov function and its derivatives uniformly on compact subsets of the region of attraction. The union of their certified sublevel sets covers the whole constrained region.

Critical weaknesses

The training target (the Zubov PDE), the data and PDE loss terms, the counterexample-guided training loop and the α-β-CROWN verification all closely follow Liu et al. (2023; 2024b; 2025). The new part is mainly the multiplicative anchor, yet the main comparisons are against SOS certificates.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Response to reviewer UHC5

Thank you for your positive assessment of our paper and for your comment on its relationship to prior neural methods. We build directly on the Zubov PDE formulation, trajectory supervision, and PDE losses of Liu et al. (2023; 2025). Our implementation extends LyZNet (Liu et al., 2024b) and uses established counterexample-guided refinement and α-β-CROWN verification tools. Appendix E.1 now distinguishes this inherited framework from our theoretical and empirical contributions.

Our theoretical contribution is the anchored representation and its guarantees. Under Assumption 1, we prove that the multiplicative correction preserves positive definiteness and strict local decrease for every fixed parameter vector (Proposition 1). In Theorem 1, we show that invariant sublevel components of anchored networks exhaust the constrained ROA under the theorem's assumptions. We also simplify derivative verification through the architecture's exact factorization (Proposition 3) and allow certified sets to include constraint-boundary portions with strict inward flow (Proposition 2).

Systematic empirical evaluation is another main contribution. We apply one common recipe to the twelve 2D benchmarks and another to the four 3D benchmarks, obtaining certified ROAs larger than the SOS baselines across both suites (Tables 1 and 2). For higher-dimensional systems, we tailor training and architectures to their dynamics and sparsity. We emphasize SOS comparisons to extend the selected examples in Liu et al. (2025) into a systematic evaluation spanning common recipes and higher-dimensional systems. We retain several examples from that work, including the Van der Pol oscillators, two-machine swing system, and 20D Van der Pol network. On this network, our training procedure improves the estimated verified volume by 13% over the frozen neural baseline of Liu et al. (2024a), as reported in Table 4. The 5D aircraft comparison gives a further 42% gain, from 8,567 to 12,168, over the MLP baseline (Table 3). We match network size, trajectory data, and optimizer-update count while using losses suited to each representation (Appendix E.1).

Reviewer_aGLqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Zubov-Net, an anchored neural network architecture for learning maximal Lyapunov functions. The core idea is to start from a known local Lyapunov function (the anchor) and learn a multiplicative correction to approximate the global maximal Lyapunov function. This construction guarantees that the Lyapunov conditions are satisfied near the equilibrium for arbitrary parameters, while the correction term reshapes the sublevel sets away from the origin to approximate the maximal Lyapunov function. In addition, the paper proves a universal approximation theorem for anchored networks: they can uniformly approximate the maximal Lyapunov function and its derivatives on any compact subset of the constrained domain of attraction. For verification, the paper introduces viability-aware conditions that allow certified sublevel sets to intersect the constraint boundary, provided that the flow points strictly inward on that boundary, and leverages architectural properties to eliminate the tanh and exponential factors, thereby reducing verification cost. The paper conducts comprehensive experiments, comparing formally verified region-of-attraction estimates with SOS baselines on 22 benchmark systems and reporting gains ranging from a modest 1.04 to an extreme 2.03×10^4 (microgrid).

Critical strengths
  1. The anchored architecture is the core contribution of this paper. Its simple design solves the verification of Lyapunov conditions near the origin. Because the anchor function guarantees correctness near the origin, the network can focus on learning the global shape away from the origin.

  2. The paper conducts experiments on 22 benchmark systems, covering a wide dimensional range from 2 to 20 and including diverse application scenarios such as planar systems, three-dimensional systems, and power systems. More importantly, all Zubov-Net results are formally verified (using CROWN and dReal) rather than merely numerically screened. The comparison with SOS baselines is also systematic and fair—for many benchmarks, the authors re-verified the SOS certificates themselves rather than merely citing numbers from the literature. The empirical evaluation is comprehensive and rigorous.

Critical weaknesses
  1. The theoretical conclusion of Theorem 1 ("the union of certified sublevel sets covers the entire constrained domain of attraction") is a statement about representability, and the experiments do not test whether the trained networks approach that limit. A discussion of this gap, or a coverage diagnostic, would affect the practical significance of Theorem 1.

  2. The paper does not report the computational cost of training and verification. Computational efficiency is a key dimension of the method's practical utility. Proposition 3 is motivated by reducing verification cost, yet it does not quantify how much verification speedup is achieved.

Questions for authors
  1. How does the trajectory simulation procedure ensure coverage of the outer region of the domain of attraction, especially in high-dimensional settings?

  2. Could the authors add more discussion regarding computational costs?

  3. In the microgrid example (S5) in Section 6.2, the gain exceeds 2×10^4. The paper notes that sparse SOS optimizes a common decay margin rather than volume. How much would this gain shrink if SOS also optimized a volume objective (e.g., using bilinear or iterative methods)? Is there a comparison against volume-optimizing SOS?

1 comment
AuthorsOpenReview ↗

Response to reviewer aGLq

Thank you for recognizing the role of the anchored architecture and the formal verification in our experiments, and for the constructive comments.

1. Theorem 1 and the coverage achieved by training

Theorem 1 establishes representability: the anchored architecture can represent invariant sublevel components that exhaust the constrained ROA. We state immediately after the theorem that this does not guarantee that training finds those networks.

To address your question, Appendix D.2 now gives an achieved-coverage diagnostic for P3 using the existing frozen certificate. From the explicit flow, we derive the constrained basin area 19+log⁡(325/2)≈24.09067819+\log(325/2)\approx24.090678, accounting for transient exits from the proof box. The neural certificate has estimated area 23.847027±0.00065823.847027\pm0.000658, or 98.988698.9886\\%\pm0.0027 percentage points of that reference; the reported SOS certificate covers about 20.0820.08\\%. The reference is analytic, and the set-volume uncertainties are sampling standard errors. This diagnostic shows that training can achieve nearly complete constrained ROA coverage in this example.

2. Coverage of the outer region by trajectory simulation

We sample initial conditions throughout the proof box, including points potentially outside the unknown ROA. As clarified in Appendix D.1, the common recipes use Monte Carlo sampling to draw 100,000 points uniformly from the box. Converging trajectories can concentrate along a few slowly decaying directions. To improve coverage of the other directions, we supplement the box samples with up to 65,536 initial states sampled independently at different distances from the equilibrium, keeping only points inside the box. From trajectories that converge inside the box, we estimate U(x)U(x) by numerically integrating the running cost ω(x)=x⊤Qx\omega(x)=x^\top Qx from each retained state onward. We use these estimates in the data loss to train HθH_\theta. During continuation, sampled checks identify states where the candidate fails the required Lyapunov decrease condition. We simulate trajectories from those states and add labels from convergent runs to the training data. Random sampling provides a practical way to explore regions beyond the local core, including in higher dimensions, without a tensor-product grid; S5 and S6 additionally exploit two-state subsystem decompositions. We use scrambled Sobol points for candidate screening and volume estimation.

3. Training costs and the effect of Proposition 3

To answer your question, Appendix D.3 (Tables 6–8) now reports training and verification times for all examples, including the example with a homogeneous anchor. Anchored neural network verification uses CROWN on GPUs for all examples. Total verification time includes local-core certification, neural verification, and additional calls to enlarge the certified set or test another candidate.

To demonstrate that the simplified conditions proposed in Proposition 3 do save verification time, we have added a paired GPU runtime comparison on the frozen P1–P12 certificates, using the same networks, domains, inner and outer levels, local cores, and CROWN settings. The direct version keeps WθW_\theta and its full Lie derivative in the verification queries. Each version runs once on the same NVIDIA H100 NVL GPU in 64-bit arithmetic, with a 300 s neural verification limit and matched decrease margins. Both totals include the same CPU time for certifying the quadratic local core, measured once per case. Both formulations verify all twelve cases. Proposition 3 gives 1.53×–12.51× speedups in total verification time. Table 6 in Appendix D.3 now includes the direct comparison alongside the training times. The artifact includes the separate verification script and proof records for both formulations.

4. Microgrid comparison with volume-oriented SOS

Thank you for this comment. We originally chose the common decay-margin objective because it yields a single convex sparse semidefinite program. To address your question, we have now added a volume-oriented SOS comparison for S5. Keeping the same quartic degree, domain, and sparse decomposition, we alternate between the polynomial coefficients and SOS multipliers to enlarge a product of ellipsoids contained in the polynomial sublevel set. Starting from the original decay-margin certificate, the search stops after 16 iterations when the relative improvement in the common ellipsoid level falls below 1%. We formally verify the resulting certificate against the exact trigonometric dynamics, using the same certified quadratic core.

The estimated SOS volume increases from 0.00201 to 4.910 ± 0.005, while the Zubov-Net volume remains 40.719. The gain therefore decreases from approximately 2.03×1042.03\times10^4 to 8.29×. We now use this stronger SOS baseline in Section 6.2, Table 4, and the microgrid figure.

Reviewer_fESiOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Given a dynamic system \dot x = f(x), Zubov's equation is a partial differential equation (PDE) whose solution v(x) characterizes the domain of attraction (DoA) of the equilibrium of f(x) as its zub-one level set. It provides a more precise characterization of the DoA than Lyapunov functions, which are typically used to find underapproximations of the DoA. Both are hard to find, but this paper presents a method to learn a Zubov function using a neural network representation, coupled with a formal guarantee for the representation of the DoA. While others have worked on learning Lyapunov functions and Sums of Squares representation of such, the authors show, over a range of examples from 2 to 20 dimensions, how their Zubov-Nets provide better approximations of the DoA.

Critical strengths

The result presented in this paper important in the operation of embedded systems: a Zubov-Net provides a certificate that a control scheme works across a set of states.

The result is less conservative than other, competing methods, and scales to high dimensional systems.

The paper is well written and clear.

Critical weaknesses

I wouldn't call this a weakness, but I would be interested to see a comparison with neural certificates computed for Hamilton-Jacobi reachable sets, such as in Jingqi Li et al, RAL 2025.

Questions for authors

How does the method compare to other model checking methods that have been developed for continuous dynamics, like Hamilton-Jacobi reachability methods, which also now have their certified learning-based implementations? HJ allows the dynamic systems to have control inputs, but presumably could be used without control inputs and with a target set that is a neighborhood of the equilibrium.

2 comments
AuthorsOpenReview ↗

Response to reviewer fESi

Thank you for your positive assessment of our paper and for your comment on Hamilton--Jacobi (HJ) reachability. Appendix E.2 now explains how HJ reachability to a certified local ROA addresses the same attraction problem. Li et al. (2025) use a discrete-time Bellman formulation to learn a reach-avoid value function and control policy, then certify initial-state neighborhoods using trajectory bounds. Our method learns and verifies scalar certificates for fixed continuous-time dynamics. The approach proposed by Li et al. and its current implementation therefore do not offer a direct comparison with our approach.

A more natural comparison is between continuous-time HJ-based training of neural certificates and Zubov-based training, with common dynamics and verification requirements. For fixed dynamics and an accumulated running-cost objective, the stationary HJ equation reduces to the Lyapunov PDE ∇U⊤f=−ω\nabla U^\top f=-\omega, whose bounded transformation gives the Zubov equation. Thus the Zubov formulation already encodes this HJ condition, while adding a bounded transformation that facilitates training and verification, as illustrated in our paper. We identify the comparison of training formulations as an interesting direction for future work.

Reviewer_fESiOpenReview ↗

Thank you!

I appreciate your response with the comparison.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers for their positive assessments and constructive comments. We have addressed their questions through focused revisions and additional experiments:

  • Achieved ROA coverage. Appendix D.2 derives the exact constrained ROA for P3 and shows that the trained neural certificate covers approximately 98.99% of its area. This illustrates that training can approach the true ROA boundary, complementing the representability guarantee of Theorem 1.

  • Computational costs. Appendix D.3 reports recorded training and verification times for all examples. A matched GPU comparison on P1–P12 shows that Proposition 3 gives 1.53×–12.51× speedups in total verification time, with both formulations verifying all twelve cases.

  • Volume-oriented SOS comparison. For the microgrid S5, we optimize a volume surrogate within the same sparse quartic family and formally verify the resulting certificate against the exact trigonometric dynamics. Its estimated volume is 4.910 ± 0.005, compared with 40.719 for Zubov-Net, giving a gain of 8.29×. The revised paper uses this stronger SOS baseline.

  • Relation to prior work. Appendix E clarifies the theoretical and empirical contributions relative to prior neural Zubov methods, discusses the existing neural baseline comparisons, and explains the connection to Hamilton–Jacobi reachability.

For ease of review, revised text is highlighted in blue, with most additions placed in the appendices. The updated supplementary code includes the new experiment scripts, frozen models, and proof summaries. We address each reviewer's comments in the individual responses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.