acceptodds
Under review as a conference paper at ICLR 2027

PolarNet: Single-Minima Neural Network for Modeling Lyapunov Functions

Abstract

Learning control strategies with provable stability guarantees continues to be a challenging problem. In this work, we examine a family of training-time behaviors exhibited by existing neural Lyapunov control methods under specific conditions, which can hinder the synthesis of a provably stable controller. We identify the root cause as the lack of neural network architectural guarantees on the learned Lyapunov function, and propose PolarNet, a network architecture that provably addresses these issues by structurally guarantee to have a single critical point. We provide theoretical guarantee regarding the properness and universality of PolarNet for modeling Lyapunov functions, and show that using it as a drop-in replacement in existing neural Lyapunov control methods can effectively circumvent particular difficulties in training. We conduct a set of numerical experiments to verify that PolarNet consistently maintains a single critical point and, when used as a drop-in replacement in existing neural Lyapunov control methods, successfully avoids training failures caused by the lack of architectural guarantees.

Primary area:
learning on time series and dynamical systems
Keywords:
Control Theory, Neural Network, Dynamic System, Lyapunov Stability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1tF5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses a subtle problem in training neural lyapunov functions - that the gradients of the parametric certificate don't converge to 0 near the origin (assuming the origin is the equilibrium being certified or the desired set point of the controller). To address this issue, this work proposes a construction of neural lyapunov functions that are guaranteed to have only a single minimum at the equilibrium. The proposed construction is guaranteed to have a single minimum that corresponds with the equilibrium. Control synthesis experiments are conducted on a toy example, where the failure case (that is, RoAs are poor using standard constructions) are compared with the PolarNet, which outperforms it.

Critical strengths
  • I think this paper addresses a subtle and interesting challenge in using neural networks for stability analysis and control. Using the PolarNet architecture potentially significantly improves the performance of neural lyapunov control style techniques.
  • The single experiment in the main body, despite being on only a single example, showcases the clear need for a PolarNet-style construction, and shows that the PolarNet outperforms it significantly.
  • The paper is generally written quite clearly, and is easy to follow.
Critical weaknesses
  • I feel the experimental slate is not systematic enough, especially compared to related literature in this space. The method needs to be evaluated on several more systems, particularly with higher dimensionalitiy than what is shown here - even examples with a modest increase in dimension (such as the caltech ducted fan) would at least somewhat indicate the general applicability of this proposed machinery.
  • There appears to be an error in the proof of Theorem 2. A simple counterexample: Ψ(x)=(x−1)3+1\Psi(x) = (x-1)^3 + 1 is smooth, invertible, and Ψ(0)=0\Psi(0) = 0. Yet Ψ(x)2\Psi(x)^2 has two critical points (at x=1 and x=0). Perhaps the statement could be rephrased to say that V=Ψ2V = \Psi^2 only has a single minimum at Ψ=0\Psi = 0? Also, I think that Theorem 3 is quite similar to the stated result in Khalil already - perhaps it should be restated as a corollary to that result.
  • I think the construction in Theorem 1 is quite similar to that given in Section 3 of "Asymptotic stability equals exponential stability, and ISS equals finite energy gain — if you twist your eyes" by Grune et al, 1999. If so, this weakens the novelty claims of this work.
Questions for authors
  • It looks like when the origin is the equilibrium, one can simply use the norm of the solution map to construct the Lyapunov function. Would it not be simpler to simply learn the closed-loop solution map along with the necessary constraints, and take its (squared) norm for the Lyapunov function?
4 comments

We thank the reviewer for the thoughtful comments and suggestions.

Q. Why not learn the closed-loop solution map and use its squared norm as the Lyapunov function?
The converse Lyapunov theorem does allow a Lyapunov function to be constructed from the closed-loop dynamics of a stable system. However, the construction is not simply the norm of the solution map. It typically requires integrating along state trajectories, which is computationally intractable. More importantly, it does not address the problem PolarNet solves. In controller synthesis, the closed-loop dynamics are being learned and are not known to be stable. The learned controller cannot easily be constrained to always produce stable dynamics with a valid, single-critical-point Lyapunov function. PolarNet enforces this property structurally.

In addition, we'd like to respond to the weaknesses mentioned but not covered in the questions:

W1. Higher-dimensional experiments.
We agree that scalability is important. We applied PolarNet to the Cart-Pole (4D) and 2D Quadrotor (6D) benchmarks of Wei et al. (2025), Sec. 5.3-5.4. All hyperparameters were unchanged, and only the Lyapunov network was replaced. Each experiment was repeated with 20 random seeds for both architectures. PolarNet's average RoA is 1.01×\times the baseline's on Cart-Pole and 1.10×\times on Quadrotor. PolarNet therefore scales at least as well as existing architectures. The gain is smaller here because spurious local minima rarely occur with the baseline on these systems (App. E).

Table: Parameter count, timing info and memory usage. Inference time reports the time of evaluating the controller with 1024 batch size for 10000 times. All experiments are conducted on the same device.

Method / ArchParameter CountWall-Clock Training TimeInference TimeVRAM Usage
Wei2023 / Baseline435525±325\pm3 min12±112\pm1 s1.61.6 GB (CPU)
Wei2023 / PolarNet144427±327\pm3 min12±112\pm1 s1.61.6 GB (CPU)
Wei2025 (2D system) / Baseline832033±533\pm5 s515±10515\pm10 ms188188 MB
Wei2025 (2D system) / PolarNet144464±564\pm5 s515±10515\pm10 ms148148 MB
Wei2025 (4D system) / Baseline8448197±10197\pm10 s540±10540\pm10 ms188188 MB
Wei2025 (4D system) / PolarNet1640349±10349\pm10 s540±10540\pm10 ms148148 MB
Wei2025 (6D system) / Baseline8576295±10295\pm10 s520±10520\pm10 ms190190 MB
Wei2025 (6D system) / PolarNet1836501±10501\pm10 s520±10520\pm10 ms149149 MB

W2. Theorem 2.
Thank you for your careful examination. By "invertible" we mean that Ψ\Psi is a C∞C^\infty diffeomorphism: it is invertible, and its Jacobian is nonsingular everywhere. This matches our proof and the construction of PolarNet, in which each coupling layer has a smooth inverse (Eq. 7). The revision states this explicitly.

W3. Theorem 3.
As suggested, it is now stated as a corollary.

W4. Relationship with the work of Grune et al, (1999).
Theorem 1 does have a construction similar to what's given by Grune et al. But this does not hinder the novelty of our work. Grune et al. considered the relationship between stability notions, and showed that UGAS and UGES, ISS and finite energy gain, can be made equivalent under appropriate coordinate changes. In contrast, our main contribution is that we discovered a failure mode in existing neural Lyapunov control methods, identified its root cause as lack of certain structural constraints, and proposed a network architecture that structurally resolves this issue. Theorem 1 serves as a mathematical tool that certifies universality of PolarNet, rather than the central conclusion drawn by our work.

Reviewer 1tF5, thanks for your review. Could you take a look at the authors’ response and comment on which of your concerns have or have not been resolved?

I thank the authors for their response, and I am happy to raise my score by 1.

Reviewer, thanks for your response. However, could you be more specific about which of your concerns have or have not been resolved? These are more important than the rating itself.

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

This paper introduces PolarNet, a neural network architecture for learning Lyapunov functions with a unique critical point at the origin. The proposed construction uses invertible neural networks to enforce PSD. The authors provide theoretical results regarding the representation capabilities of the architecture and demonstrate its integration into existing neural Lyapunov control methods. Numerical experiments on function approximation and nonlinear controller synthesis suggest improved training behavior and larger empirical regions of attraction compared to existing architectures.

Critical strengths
  • The paper identifies a relevant limitation of existing neural Lyapunov architectures. Enforcing a unique critical point by construction is a meaningful contribution, particularly since additional critical points can interfere with controller synthesis. The proposed architecture provides this property without introducing additional penalties or constraints during training.

  • The approach can be integrated into existing neural Lyapunov control frameworks without substantial modifications. The experiments demonstrate noticeable improvements in the empirical region of attraction across two different synthesis methods, providing initial evidence of its practical usefulness.

Critical weaknesses

The theoretical expressivity guarantees require further justification. Theorem 1 assumes a non-degenerate critical point at the origin, while Theorem 3 does not establish this property for the Lyapunov function whose existence it guarantees. Consequently, the claimed applicability to arbitrary asymptotically stable equilibria is not fully established by the presented arguments. Furthermore, the cited universality of affine coupling flows does not directly establish universality for the specific constrained, finite-depth architecture considered here.

The experimental evaluation is relatively limited. Controller synthesis is demonstrated on only two-dimensional systems, making it difficult to assess scalability and general applicability. More importantly, the reported regions of attraction are estimated through numerical simulations rather than formally certified. While these results are encouraging, they do not establish that the synthesized controllers satisfy the Lyapunov decrease condition throughout the claimed domains. Additional experiments on higher-dimensional systems and comparisons with capacity-matched architectures would strengthen the contribution.

Questions for authors

Can the authors clarify how Theorem 3 establishes the non-degeneracy assumption required by Theorem 1? Additionally, does the universality result remain valid under the architectural restrictions imposed on PolarNet, particularly the zero-crossing constraints?

Can the authors provide formal verification or stronger numerical evidence that the learned Lyapunov functions satisfy the decrease condition throughout the reported regions of attraction? In particular, how do the reported empirical regions compare with regions that can actually be certified?

3 comments

We thank the reviewer for the constructive comments and suggestions.

Q1(a). How does Theorem 3 establish the non-degeneracy assumption of Theorem 1?
Theorem 3 (Corollary 1 in the revision) does not guarantee non-degeneracy, and we do not claim that it does (line 057 and Remark 1 at line 225). Non-degeneracy is a mild, generic condition. Any degenerate critical point becomes non-degenerate under an arbitrarily small perturbation (Guillemin & Pollack, 2010, Ch. 1.7), so the condition is always satisfiable in practice.

Q1(b). Does universality hold under the zero-crossing constraint?
Yes. The zero-crossing constraint narrows the target class to origin-preserving diffeomorphisms. It does not change the approximation mechanism. The universality proof of Teshima et al. (2020) already reduces the problem to a zero-crossing case with an initial shift (their Sec. 4.2). Their construction requires only zero-crossing translation networks, which are exactly the kind used in PolarNet (see their Fig. 1 and Sec. 4.2).

Q2. Formal verification, and empirical vs. certified regions.
Formal verification checks only the decrease condition, and for neural Lyapunov functions this alone does not establish stability. In Fig. 4a, for example, an unconstrained VθV_\theta satisfies the decrease condition everywhere except at the two saddle points shown, yet its true RoA is much smaller than the region where the condition holds.
We verified the relevant experiment with the protocol of Wei et al. (2025), and the results confirm this point. The decrease condition holds on 98.90% of the training region for the baseline and on 99.96% for PolarNet. The two metrics are nearly identical, but the actual RoAs differ drastically. For PolarNet, the verified region closely matches the empirical RoA.

In addition, we'd like to respond to the weaknesses mentioned but not covered in the questions:

W1. Scalability and general applicability.
As suggested, we applied PolarNet to the Cart-Pole (4D) and 2D Quadrotor (6D) benchmarks of Wei et al. (2025), Sec. 5.3-5.4. All settings and hyperparameters were unchanged, and only the Lyapunov network was replaced. Each experiment was repeated with 20 random seeds for both architectures. PolarNet's average RoA is 1.01×\times the baseline's on Cart-Pole and 1.10×\times on Quadrotor. PolarNet therefore scales at least as well as existing architectures. The gain is smaller here because spurious local minima rarely occur with the baseline on these systems (App. E). Training and inference costs are reported below.

Table: Parameter count, timing info and memory usage. Inference time reports the time of evaluating the controller with 1024 batch size for 10000 times. All experiments are conducted on the same device.

Method / ArchParameter CountWall-Clock Training TimeInference TimeVRAM Usage
Wei2023 / Baseline435525±325\pm3 min12±112\pm1 s1.61.6 GB (CPU)
Wei2023 / PolarNet144427±327\pm3 min12±112\pm1 s1.61.6 GB (CPU)
Wei2025 (2D system) / Baseline832033±533\pm5 s515±10515\pm10 ms188188 MB
Wei2025 (2D system) / PolarNet144464±564\pm5 s515±10515\pm10 ms148148 MB
Wei2025 (4D system) / Baseline8448197±10197\pm10 s540±10540\pm10 ms188188 MB
Wei2025 (4D system) / PolarNet1640349±10349\pm10 s540±10540\pm10 ms148148 MB
Wei2025 (6D system) / Baseline8576295±10295\pm10 s520±10520\pm10 ms190190 MB
Wei2025 (6D system) / PolarNet1836501±10501\pm10 s520±10520\pm10 ms149149 MB

We thank the authors for their detailed response and additional experiments. The new 4D and 6D evaluations address our concerns regarding scalability. However, the theoretical concerns remain insufficiently resolved. In particular, the non-degeneracy assumption is not established for the relevant Lyapunov functions, and the universality claim under the imposed architectural constraints requires further justification. Can the authors provide a rigorous argument establishing these properties under the stated assumptions?

We thank the reviewer for helping us improve our manuscript. We have revised our network construction to provide a better theoretical guarantee with unchanged performance characteristics. We detail the clarification below:

Non-Degenerancy Property

We can rigirously show that the universality claim holds without the non-degenerancy property when the dimension of state space n≠4,5n\neq4,5. On the other hand, we can also formally show that non-degenerancy is true for almost every function.

For the case where n≠4,5n\neq4,5, we can refer to Proposition 1 from "Asymptotic stability equals exponential stability, and ISS equals finite energy gain--if you twist your eyes" by Grune et al, 1999: for any proper V:Rn→RV:\mathbb{R}^n\to\mathbb{R} where n≠4,5n\neq4,5, if VV is C1\mathcal{C}^1 and has non-vanishing gradient on Rn∖{0}\mathbb{R}^n\setminus\lbrace 0\rbrace (which is true for all Lyapunov functions), then there exists a homeomorphism T:Rn→RnT:\mathbb{R}^n\to\mathbb{R}^n where T(0)=0T(0) = 0 and V(T−1(x))=∥x∥2V(T^{-1}(x))=\|x\|^2. Additionally, TT can be chosen to be a diffeomorphism on Rn∖{0}\mathbb{R}^n\setminus\lbrace 0\rbrace. Our stated precondition (VV being C∞\mathcal{C}^{\infty}, hence also C1\mathcal{C}^1) is just a special case of this.

Furthermore, non-degenerancy is true for almost every function. And we can always get rid of degenerancy with an arbitrarily small perturbation. Formally ("Differential Topology" by Guillemin et al, 2010): Let X\mathcal{X} denote a manifold in Rn\mathbb{R}^n. For all functions f:X→Rf:\mathcal{X}\to\mathbb{R}, for almost every a∈Rna\in\mathbb{R}^n, the augmented function fa(x)=f(x)+⟨a,x⟩f_a(x)=f(x)+\langle a,x\rangle contains no degenerate critical point. The phrase "almost every" means that the values aa for which this assertion is false only constitute a set of measure zero. Similar results are also given in "Topics in Nonlinear Functional Analysis" by Nirenberg et al, 2001.

Architectural Constraints

Thank you for pointing out the inaccurary. We now revised the network construction, removed the zero-crossing constraint in sub-networks, and instead enforce zero-crossing property via the following construction: Ψ(x)=hcf(x)−hcf(0)\Psi(x)=h_{cf}(x)-h_{cf}(0), where hcfh_{cf} is an unconstrained coupling flow network fully compatible with the universality proof given by Teshima et al. Since hcf(0)h_{cf}(0) is a constant term, the proof still holds (shifted diffeomorphism is still a diffeomorphism).

We conducted all experiments with the revised construction, and the performance gain still holds (the results are nearly identical compared to previous construction). We have updated the results in the revised manuscript.

Reviewer_RjiMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an architecture for Lyapunov functions that has only one critical point at the origin by construction. The authors term this a "single-pole function". The architecture is designed as the square of an affine coupling flow architecture. It is then shown that the proposed architecture leads to less conservative estimates of the region of attraction.

Critical strengths

The problem is interesting. The use of affine coupling flow architecture is also somewhat interesting.

Critical weaknesses
  1. The paper does not prove that the constructed architecture actually leads to a Lyapunov function. No formal verification is used either.

  2. The examples are almost trivial to solve. Just look at (10), (31), and (32). Appropriate choices of uu can easily cancel the dynamics and globally stabilize the system. Hence, they are not valid benchmarks.

  3. In terms of regions of attraction/stabilization, the paper does not compare with the state of the art. The authors should look at tools that specialize in maximizing regions of attraction, including both sum-of-squares and neural network approaches for learning maximal Lyapunov functions.

  4. Local validity of the proposed architecture is not shown.

  5. The universality claim stated in the abstract is not clearly stated or proved in the paper.

  6. In terms of writing, the basic definitions of stability are incorrect. Definition 1 does not include Lyapunov stability as part of asymptotic stability. Definition 2 defines global asymptotic stability rather than asymptotic stability.

Questions for authors
  1. Was formal verification used to check the region of attraction?

  2. Is universality actually proved in the paper? In what sense? What assumptions do you need for the vector field?

  3. Can the proposed architecture be a valid Lyapunov function? Is there a proof?

3 comments

We thank the reviewer for the constructive feedback. We first clarify the problem our work addresses.

Scope of the work.
We study neural Lyapunov control: synthesizing a stabilizing controller uθu_\theta jointly with its Lyapunov function VθV_\theta. Existing methods model VθV_\theta with under-constrained architectures. We show why this can cause controller synthesis to fail, and we propose PolarNet, which removes this failure mode structurally. Our work concerns controller synthesis. It does not concern RoA estimation for given dynamics.

Here we detail our response to your questions:

Q1. Was formal verification used to check the region of attraction?
Yes. We verified the relevant experiment with the protocol of Wei et al. (2025). The decrease condition holds on 98.90% of the training region for the baseline and on 99.96% for PolarNet. However, the two models have nearly identical satisfaction rates and drastically different actual RoAs. A high satisfaction rate of the decrease condition alone therefore does not establish stability for neural Lyapunov functions. Fig. 4a illustrates this (see Sec. 2.2 for detailed discussion).

Q2. Is universality actually proved in the paper? In what sense? What assumptions do you need for the vector field?
Yes, in Sec. 3.2. Under a mild non-degeneracy condition, PolarNet can approximate a Lyapunov function to arbitrary precision for any controller synthesis task that requires an asymptotically stable equilibrium (line 300). The result combines three steps:

  • Theorem 1: every single pole function with a non-degenerate minimum equals ∥Ψ(x)∥2\|\Psi(x)\|^2 for some smooth, zero-crossing diffeomorphism Ψ\Psi.
  • Teshima et al. (2020): affine coupling flows are universal approximators of C2C^2 diffeomorphisms, and hence can approximate Ψ\Psi.
  • Corollary 1: (Theorem 3 before revision) every asymptotically stable equilibrium of a locally Lipschitz system admits a Lyapunov function that is a single pole function.

The only assumption on the vector field is local Lipschitz continuity (Sec. 2.1).

Q3. Can the proposed architecture be a valid Lyapunov function? Is there a proof?
Yes. By construction (Sec. 3.1, Theorem 2), PolarNet is a single pole function for any finite parameters. That is, it is C∞C^\infty, positive definite, and proper, and its only critical point is at the origin. All of these properties are consistent with a Lyapunov function. Corollary 1 further guarantees that every synthesis problem we consider admits a Lyapunov function in this class.

In addition, we'd like to respond to the weaknesses mentioned but not covered in the questions:

W1: "The paper does not prove that the constructed architecture actually leads to a Lyapunov function."
Response: This concern combines two separate guarantees. The existing neural Lyapunov control frameworks already prove that a properly converged VθV_\theta is a valid Lyapunov function, as shown in their own papers. PolarNet targets a different problem: these frameworks can fail to converge at all. PolarNet is a drop-in replacement for VθV_\theta that removes this failure mode, and the frameworks' existing guarantees still apply.

W2: The examples are almost trivial to solve and are not valid benchmarks.
Response: A human expert could indeed design such a uu by hand. Our task, however, is to synthesize uu automatically, without human supervision. The simplicity of these systems is the point: existing methods fail even on simple systems whenever the scenario in Sec. 2.2 occurs, which reveals a systematic failure mode. PolarNet resolves it. Note also that (31) is not a benchmark. It is the linearization of (10), used only for the LQR initialization step of Wei et al. (2023) (App. C).

W3: Lacking comparison to SOTA methods that specializes in learning maximal Lyapunov functions
Response: As stated in the above Scope of the work, our work considers controller synthesis, rather than RoA estimation. Methods concerning these two problems cannot be meaningfully compared against each other.

W4: Local validity of the architecture is not shown
Response: As we stated in the response to W1, existing neural Lyapunov control frameworks already prove that a properly converged VθV_\theta is a valid Lyapunov function on the training region. Our work resolves an issue where these frameworks fail to converge. After convergence, the existing guarantees of these works still apply.

W5: Proof of universality
Please refer to our response to Q2 above.

W6: Definitions of stability
Response: We thank the reviewer for the careful reading. In the revision, Definition 1 includes the stability requirement in addition to attractivity. Definition 2 is restricted to a neighborhood of the origin, which matches the rest of the manuscript, where only a bounded region X\mathcal{X} is considered.

Reviewer_RjiMOpenReview ↗

Thank you for the response

Thank you for the response. It appears that by verification, you meant testing whether the decrease conditions are satisfied. To guarantee a region of attraction under a given controller, we need 100% satisfaction of these conditions. Furthermore, the current examples are too simple. There are definitely existing works that address neural network-based controller synthesis and aim to maximize regions of attraction as well, some with formal verification (e.g., using SMT solvers and neural network verifiers). I encourage the authors to pursue comparisons with these works in future research. Unfortunately, for the current submission, I have decided to maintain my previous evaluation.

Thank you for clarifying the exact requirement of verification. In our previous response, we reported satisfaction of the condition over the entire valid state space, rather than RoA. We now re-verified the stability condition on the empirical RoA using the same protocol proposed by Wei et al, and found that it is indeed fully satisfied on the RoA. We apologize for the confusion caused by this.

Moreover, we'd like to highlight that our work proposes a method to resolve a common issue found in existing neural Lyapunov control frameworks. It is not about proposing another controller synthesis method. The frameworks involved in our experiments have been shown to produce verifiably stable controllers once they converge, and our work focus on resolving an issue where they fail to converge at all.

The experiments and examples in our work confirms our discovery and analysis of a particular failure mode in existing frameworks, and verified PolarNet's capability of preventing it. They serve as verifications (that the failure mode is real and is resolved by PolarNet), rather than a general performance benchmark. Being simple does not hinder their effectiveness for this purpose. We do not claim absolute performance gain in general cases, and our experiments are not designed to demonstrate that. For this reason, benchmarking between multiple methods for absolute performance gain is out of scope of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.